chore: updated config and calibration files

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

.cache/calibration/aloha_default/left_follower.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2015,
+        3058,
+        3061,
+        1071,
+        1071,
+        2035,
+        2152,
+        2029,
+        2499
+    ],
+    "end_pos": [
+        -1008,
+        -1963,
+        -1966,
+        2141,
+        2143,
+        -971,
+        3043,
+        -1077,
+        3144
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/left_leader.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -1024
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2035,
+        3024,
+        3019,
+        979,
+        981,
+        1982,
+        2166,
+        2124,
+        1968
+    ],
+    "end_pos": [
+        -990,
+        -2017,
+        -2015,
+        2078,
+        2076,
+        -1030,
+        3117,
+        -1016,
+        2556
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/right_follower.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2056,
+        2895,
+        2896,
+        1191,
+        1190,
+        2018,
+        2051,
+        2056,
+        2509
+    ],
+    "end_pos": [
+        -1040,
+        -2004,
+        -2006,
+        2126,
+        2127,
+        -1010,
+        3050,
+        -1117,
+        3143
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/right_leader.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2068,
+        3034,
+        3030,
+        1038,
+        1041,
+        1991,
+        1948,
+        2090,
+        1985
+    ],
+    "end_pos": [
+        -1025,
+        -2014,
+        -2015,
+        2058,
+        2060,
+        -955,
+        3091,
+        -940,
+        2576
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/so100/main_follower.json

@@ -0,0 +1 @@
+{"homing_offset": [-2090, 3086, -952, -2089, 2063, -2472], "drive_mode": [0, 1, 0, 0, 1, 0], "start_pos": [2077, 3156, 908, 2047, 2047, 2048], "end_pos": [3114, -2062, 1976, 3113, -1039, 3496], "calib_mode": ["DEGREE", "DEGREE", "DEGREE", "DEGREE", "DEGREE", "LINEAR"], "motor_names": ["shoulder_pan", "shoulder_lift", "elbow_flex", "wrist_flex", "wrist_roll", "gripper"]}

.cache/calibration/so100/main_leader.json

@@ -0,0 +1 @@
+{"homing_offset": [-2038, 2985, -1088, -2115, 2060, -1962], "drive_mode": [0, 1, 0, 0, 1, 0], "start_pos": [2052, 3087, 960, 2047, 2011, 2047], "end_pos": [3062, -1961, 2112, 3139, -1036, 2986], "calib_mode": ["DEGREE", "DEGREE", "DEGREE", "DEGREE", "DEGREE", "LINEAR"], "motor_names": ["shoulder_pan", "shoulder_lift", "elbow_flex", "wrist_flex", "wrist_roll", "gripper"]}

.dockerignore

@@ -0,0 +1,160 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Misc
+.git
+tmp
+wandb
+data
+outputs
+.vscode
+rl
+media
+
+
+# Logging
+logs
+
+# HPC
+nautilus/*.yaml
+*.key
+
+# Slurm
+sbatch*.sh
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+!tests/artifacts
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Ignore .cache except calibration
+.cache/*
+!.cache/calibration/
+!.cache/calibration/**
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/

.gitattributes

@@ -0,0 +1,20 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+*.memmap filter=lfs diff=lfs merge=lfs -text
+*.stl filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.json !text !filter !merge !diff

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,68 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: "\U0001F41B Bug Report"
+description: Submit a bug report to help us improve LeRobot
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to submit a bug report! 🐛
+        If this is not a bug related to the LeRobot library directly, but instead a general question about your code or the library specifically please use our [discord](https://discord.gg/s3KuuzsPFb).
+
+  - type: textarea
+    id: system-info
+    attributes:
+      label: System Info
+      description: If needed, you can share your lerobot configuration with us by running `python -m lerobot.scripts.display_sys_info` and copy-pasting its outputs below
+      render: Shell
+      placeholder: lerobot version, OS, python version, numpy version, torch version, and lerobot's configuration
+    validations:
+      required: true
+
+  - type: checkboxes
+    id: information-scripts-examples
+    attributes:
+      label: Information
+      description: 'The problem arises when using:'
+      options:
+        - label: "One of the scripts in the examples/ folder of LeRobot"
+        - label: "My own task or dataset (give details below)"
+
+  - type: textarea
+    id: reproduction
+    validations:
+      required: true
+    attributes:
+      label: Reproduction
+      description: |
+        If needed, provide a simple code sample that reproduces the problem you ran into. It can be a Colab link or just a code snippet.
+        Sharing error messages or stack traces could be useful as well!
+        Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting
+        Try to avoid screenshots, as they are hard to read and don't allow copy-and-pasting.
+
+      placeholder: |
+        Steps to reproduce the behavior:
+
+          1.
+          2.
+          3.
+
+  - type: textarea
+    id: expected-behavior
+    validations:
+      required: true
+    attributes:
+      label: Expected behavior
+      description: "A clear and concise description of what you would expect to happen."

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,34 @@
+## What this does
+Explain what this PR does. Feel free to tag your PR with the appropriate label(s).
+
+Examples:
+|  Title               | Label           |
+|----------------------|-----------------|
+| Fixes #[issue]       | (🐛 Bug)        |
+| Adds new dataset     | (🗃️ Dataset)    |
+| Optimizes something  | (⚡️ Performance) |
+
+## How it was tested
+Explain/show how you tested your changes.
+
+Examples:
+- Added `test_something` in `tests/test_stuff.py`.
+- Added `new_feature` and checked that training converges with policy X on dataset/environment Y.
+- Optimized `some_function`, it now runs X times faster than previously.
+
+## How to checkout & try? (for the reviewer)
+Provide a simple way for the reviewer to try out your changes.
+
+Examples:
+```bash
+pytest -sx tests/test_stuff.py::test_something
+```
+```bash
+python lerobot/scripts/train.py --some.option=true
+```
+
+## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR
+**Note**: Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
+members/contributors who may be interested in your PR. Try to avoid tagging more than 3 people.
+
+**Note**: Before submitting this PR, please read the [contributor guideline](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md#submitting-a-pull-request-pr).

.github/workflows/build-docker-images.yml

@@ -0,0 +1,135 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Inspired by
+# https://github.com/huggingface/peft/blob/main/.github/workflows/build_docker_images.yml
+name: Builds
+
+on:
+  workflow_dispatch:
+  workflow_call:
+  schedule:
+    - cron: "0 1 * * *"
+
+permissions: {}
+
+env:
+  PYTHON_VERSION: "3.10"
+
+jobs:
+  latest-cpu:
+    name: CPU
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Install Git LFS
+        run: |
+          sudo apt-get update
+          sudo apt-get install git-lfs
+          git lfs install
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+        with:
+          cache-binary: false
+
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          lfs: true
+          persist-credentials: false
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push CPU
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-cpu/Dockerfile
+          push: true
+          tags: huggingface/lerobot-cpu
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
+
+
+  latest-cuda:
+    name: GPU
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Install Git LFS
+        run: |
+          sudo apt-get update
+          sudo apt-get install git-lfs
+          git lfs install
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+        with:
+          cache-binary: false
+
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          lfs: true
+          persist-credentials: false
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-gpu/Dockerfile
+          push: true
+          tags: huggingface/lerobot-gpu
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
+
+
+  latest-cuda-dev:
+    name: GPU Dev
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+        with:
+          cache-binary: false
+
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU dev
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-gpu-dev/Dockerfile
+          push: true
+          tags: huggingface/lerobot-gpu:dev
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}

.github/workflows/nightly-tests.yml

@@ -0,0 +1,93 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Inspired by
+# https://github.com/huggingface/peft/blob/main/.github/workflows/nightly.yml
+name: Nightly
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "0 2 * * *"
+
+permissions: {}
+
+# env:
+  # SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
+jobs:
+  run_all_tests_cpu:
+    name: CPU
+    strategy:
+      fail-fast: false
+    runs-on:
+      group: aws-general-8-plus
+    container:
+      image: huggingface/lerobot-cpu:latest
+      options: --shm-size "16gb"
+      credentials:
+        username: ${{ secrets.DOCKERHUB_USERNAME }}
+        password: ${{ secrets.DOCKERHUB_PASSWORD }}
+    defaults:
+      run:
+        shell: bash
+        working-directory: /lerobot
+    steps:
+      - name: Tests
+        run: pytest -v --cov=./lerobot --disable-warnings tests
+
+      - name: Tests end-to-end
+        run: make test-end-to-end
+
+
+  run_all_tests_single_gpu:
+    name: GPU
+    strategy:
+      fail-fast: false
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0"
+      TEST_TYPE: "single_gpu"
+    container:
+      image: huggingface/lerobot-gpu:latest
+      options: --gpus all --shm-size "16gb"
+      credentials:
+        username: ${{ secrets.DOCKERHUB_USERNAME }}
+        password: ${{ secrets.DOCKERHUB_PASSWORD }}
+    defaults:
+      run:
+        shell: bash
+        working-directory: /lerobot
+    steps:
+      - name: Nvidia-smi
+        run: nvidia-smi
+
+      - name: Test
+        run: pytest -v --cov=./lerobot --cov-report=xml --disable-warnings tests
+      #   TODO(aliberts): Link with HF Codecov account
+      # - name: Upload coverage reports to Codecov with GitHub Action
+      #   uses: codecov/codecov-action@v4
+      #   with:
+      #     files: ./coverage.xml
+      #     verbose: true
+      - name: Tests end-to-end
+        env:
+          DEVICE: cuda
+        run: make test-end-to-end
+
+    #   - name: Generate Report
+    #     if: always()
+    #     run: |
+    #       pip install slack_sdk tabulate
+    #       python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY

.github/workflows/quality.yml

@@ -0,0 +1,72 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: Quality
+
+on:
+  workflow_dispatch:
+  workflow_call:
+  pull_request:
+  push:
+    branches:
+      - main
+
+permissions: {}
+
+env:
+  PYTHON_VERSION: "3.10"
+
+jobs:
+  style:
+    name: Style
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+
+      - name: Get Ruff Version from pre-commit-config.yaml
+        id: get-ruff-version
+        run: |
+          RUFF_VERSION=$(awk '/repo: https:\/\/github.com\/astral-sh\/ruff-pre-commit/{flag=1;next}/rev:/{if(flag){print $2;exit}}' .pre-commit-config.yaml)
+          echo "ruff_version=${RUFF_VERSION}" >> $GITHUB_OUTPUT
+
+      - name: Install Ruff
+        env:
+          RUFF_VERSION: ${{ steps.get-ruff-version.outputs.ruff_version }}
+        run: python -m pip install "ruff==${RUFF_VERSION}"
+
+      - name: Ruff check
+        run: ruff check --output-format=github
+
+      - name: Ruff format
+        run: ruff format --diff
+
+  typos:
+    name: Typos
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+
+      - name: typos-action
+        uses: crate-ci/typos@v1.29.10

.github/workflows/test-docker-build.yml

@@ -0,0 +1,82 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Inspired by
+# https://github.com/huggingface/peft/blob/main/.github/workflows/test-docker-build.yml
+name: Test Dockerfiles
+
+on:
+  pull_request:
+    paths:
+      # Run only when DockerFile files are modified
+      - "docker/**"
+
+permissions: {}
+
+env:
+  PYTHON_VERSION: "3.10"
+
+jobs:
+  get_changed_files:
+    name: Detect modified Dockerfiles
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+
+      - name: Get changed files
+        id: changed-files
+        uses: tj-actions/changed-files@3f54ebb830831fc121d3263c1857cfbdc310cdb9 #v42
+        with:
+          files: docker/**
+          json: "true"
+
+      - name: Run step if only the files listed above change  # zizmor: ignore[template-injection]
+        if: steps.changed-files.outputs.any_changed == 'true'
+        id: set-matrix
+        run: |
+          echo "matrix=${{ steps.changed-files.outputs.all_changed_files}}" >> $GITHUB_OUTPUT
+
+  build_modified_dockerfiles:
+    name: Build modified Docker images
+    needs: get_changed_files
+    runs-on:
+      group: aws-general-8-plus
+    if: needs.get_changed_files.outputs.matrix != ''
+    strategy:
+      fail-fast: false
+      matrix:
+        docker-file: ${{ fromJson(needs.get_changed_files.outputs.matrix) }}
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+        with:
+          cache-binary: false
+
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+
+      - name: Build Docker image
+        uses: docker/build-push-action@v5
+        with:
+          file: ${{ matrix.docker-file }}
+          context: .
+          push: False
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}

.github/workflows/test.yml

@@ -1,118 +1,150 @@
-name: Test
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: Tests
 
 on:
   pull_request:
-    branches:
-      - main
-    types: [opened, synchronize, reopened, labeled]
+    paths:
+      - "lerobot/**"
+      - "tests/**"
+      - "examples/**"
+      - ".github/**"
+      - "pyproject.toml"
+      - ".pre-commit-config.yaml"
+      - "Makefile"
+      - ".cache/**"
   push:
     branches:
       - main
+    paths:
+      - "lerobot/**"
+      - "tests/**"
+      - "examples/**"
+      - ".github/**"
+      - "pyproject.toml"
+      - ".pre-commit-config.yaml"
+      - "Makefile"
+      - ".cache/**"
+
+permissions: {}
+
+env:
+  UV_VERSION: "0.6.0"
 
 jobs:
-  test:
-    if: |
-      ${{ github.event_name == 'pull_request' && contains(github.event.pull_request.labels.*.name, 'CI') }} ||
-      ${{ github.event_name == 'push' }}
+  pytest:
+    name: Pytest
     runs-on: ubuntu-latest
     env:
-      POETRY_VERSION: 1.8.1
+      MUJOCO_GL: egl
     steps:
-      #----------------------------------------------
-      #       check-out repo and set-up python
-      #----------------------------------------------
-      - name: Check out repository
-        uses: actions/checkout@v4
-      - name: Set up python
-        id: setup-python
-        uses: actions/setup-python@v5
+      - uses: actions/checkout@v4
         with:
-          python-version: '3.10'
-      #----------------------------------------------
-      #         install & configure poetry
-      #----------------------------------------------
-      - name: Load cached Poetry installation
-        id: restore-poetry-cache
-        uses: actions/cache/restore@v3
-        with:
-          path: ~/.local  # the path depends on the OS
-          key: poetry-${{ env.POETRY_VERSION }}  # increment to reset cache
-      - name: Install Poetry
-        if: steps.restore-poetry-cache.outputs.cache-hit != 'true'
-        uses: snok/install-poetry@v1
-        with:
-          version: ${{ env.POETRY_VERSION }}
-          virtualenvs-create: true
-          installer-parallel: true
-      - name: Save cached Poetry installation
-        if: |
-          steps.restore-poetry-cache.outputs.cache-hit != 'true' &&
-          github.ref_name == 'main'
-        id: save-poetry-cache
-        uses: actions/cache/save@v3
-        with:
-          path: ~/.local  # the path depends on the OS
-          key: poetry-${{ env.POETRY_VERSION }}  # increment to reset cache
-      - name: Configure Poetry
-        run: poetry config virtualenvs.in-project true
-      #----------------------------------------------
-      #           install dependencies
-      #----------------------------------------------
-      - name: Load cached venv
-        id: restore-dependencies-cache
-        uses: actions/cache/restore@v3
-        with:
-          path: .venv
-          key: venv-${{ steps.setup-python.outputs.python-version }}-${{ env.POETRY_VERSION }}-${{ hashFiles('**/poetry.lock') }}
-      - name: Install dependencies
-        if: steps.restore-dependencies-cache.outputs.cache-hit != 'true'
+          lfs: true  # Ensure LFS files are pulled
+          persist-credentials: false
+
+      - name: Install apt dependencies
+      # portaudio19-dev is needed to install pyaudio
         run: |
-          poetry install --no-interaction --no-root
-          git clone https://github.com/real-stanford/diffusion_policy
-          cp -r diffusion_policy/diffusion_policy $(poetry env info -p)/lib/python3.10/site-packages/
-      - name: Save cached venv
-        if: |
-            steps.restore-dependencies-cache.outputs.cache-hit != 'true' &&
-            github.ref_name == 'main'
-        id: save-dependencies-cache
-        uses: actions/cache/save@v3
+          sudo apt-get update && \
+          sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
+
+      - name: Install uv and python
+        uses: astral-sh/setup-uv@v5
         with:
-          path: .venv
-          key: venv-${{ steps.setup-python.outputs.python-version }}-${{ env.POETRY_VERSION }}-${{ hashFiles('**/poetry.lock') }}
-      #----------------------------------------------
-      #             install project
-      #----------------------------------------------
-      - name: Install project
-        run: poetry install --no-interaction
-      #----------------------------------------------
-      #               run tests
-      #----------------------------------------------
-      - name: Run tests
+          enable-cache: true
+          version: ${{ env.UV_VERSION }}
+          python-version: "3.10"
+
+      - name: Install lerobot (all extras)
+        run: uv sync --all-extras
+
+      - name: Test with pytest
         run: |
-          source .venv/bin/activate
-          pytest tests
-      - name: Test train pusht end-to-end
+          uv run pytest tests -v --cov=./lerobot --durations=0 \
+            -W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
+            -W ignore::UserWarning:torch.utils.data.dataloader:558 \
+            -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
+            && rm -rf tests/outputs outputs
+
+  pytest-minimal:
+    name: Pytest (minimal install)
+    runs-on: ubuntu-latest
+    env:
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
+          persist-credentials: false
+
+      - name: Install apt dependencies
+        run: sudo apt-get update && sudo apt-get install -y ffmpeg
+
+      - name: Install uv and python
+        uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+          version: ${{ env.UV_VERSION }}
+          python-version: "3.10"
+
+      - name: Install lerobot
+        run: uv sync --extra "test"
+
+      - name: Test with pytest
         run: |
-          source .venv/bin/activate
-          python lerobot/scripts/train.py \
-            hydra.job.name=pusht \
-            env=pusht \
-            wandb.enable=False \
-            offline_steps=1 \
-            online_steps=0 \
-            device=cpu
-      # TODO(rcadene, aliberts): Add end-to-end test of eval checkpoint post training
-      # - name: Test eval pusht end-to-end
-      #   run: |
-      #     source .venv/bin/activate
-      #     python lerobot/scripts/eval.py
-      #       hydra.job.name=pusht \
-      #       env=pusht \
-      #       wandb.enable=False \
-      #       eval_episodes=1 \
-      #       device=cpu
-      #----------------------------------------------
-      #              cleanup
-      #----------------------------------------------
-      - name: Cleanup
-        run: rm -rf diffusion_policy data
+          uv run pytest tests -v --cov=./lerobot --durations=0 \
+            -W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
+            -W ignore::UserWarning:torch.utils.data.dataloader:558 \
+            -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
+            && rm -rf tests/outputs outputs
+
+  end-to-end:
+    name: End-to-end
+    runs-on: ubuntu-latest
+    env:
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
+          persist-credentials: false
+
+      - name: Install apt dependencies
+      # portaudio19-dev is needed to install pyaudio
+        run: |
+          sudo apt-get update && \
+          sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
+
+      - name: Install uv and python
+        uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+          version: ${{ env.UV_VERSION }}
+          python-version: "3.10"
+
+      - name: Install lerobot (all extras)
+        run: |
+          uv venv
+          uv sync --all-extras
+
+      - name: venv
+        run: |
+          echo "PYTHON_PATH=${{ github.workspace }}/.venv/bin/python" >> $GITHUB_ENV
+
+      - name: Test end-to-end
+        run: |
+          make test-end-to-end \
+            && rm -rf outputs

.github/workflows/trufflehog.yml

@@ -0,0 +1,35 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+on:
+  push:
+
+name: Secret Leaks
+
+permissions: {}
+
+jobs:
+  trufflehog:
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+      with:
+        fetch-depth: 0
+        persist-credentials: false
+
+    - name: Secret Scanning
+      uses: trufflesecurity/trufflehog@main
+      with:
+        extra_args: --only-verified

.gitignore

@@ -1,19 +1,39 @@
-# Custom
-diffusion_policy
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 # Logging
 logs
 tmp
 wandb
+
+# Data
 data
 outputs
+
+# Apple
+.DS_Store
+
+# VS Code
 .vscode
-rl
 
 # HPC
 nautilus/*.yaml
 *.key
 
+# Slurm
+sbatch*.sh
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]
@@ -43,6 +63,10 @@ share/python-wheels/
 *.egg
 MANIFEST
 
+# uv/poetry lock files
+poetry.lock
+uv.lock
+
 # PyInstaller
 #  Usually these files are written by a python script from a template
 #  before PyInstaller builds the exe, so as to inject date/other infos into it.
@@ -54,12 +78,12 @@ pip-log.txt
 pip-delete-this-directory.txt
 
 # Unit test / coverage reports
+!tests/artifacts
 htmlcov/
 .tox/
 .nox/
 .coverage
 .coverage.*
-.cache
 nosetests.xml
 coverage.xml
 *.cover
@@ -67,6 +91,11 @@ coverage.xml
 .hypothesis/
 .pytest_cache/
 
+# Ignore .cache except calibration
+.cache/*
+!.cache/calibration/
+!.cache/calibration/**
+
 # Translations
 *.mo
 *.pot
@@ -88,6 +117,7 @@ instance/
 docs/_build/
 
 # PyBuilder
+.pybuilder/
 target/
 
 # Jupyter Notebook
@@ -100,13 +130,6 @@ ipython_config.py
 # pyenv
 .python-version
 
-# pipenv
-#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
-#   However, in case of collaboration, if having platform-specific dependencies or dependencies
-#   having no cross-platform support, pipenv may install dependencies that don't work, or not
-#   install all needed dependencies.
-#Pipfile.lock
-
 # PEP 582; used by e.g. github.com/David-OConnor/pyflow
 __pypackages__/
 
@@ -117,6 +140,14 @@ celerybeat.pid
 # SageMath parsed files
 *.sage.py
 
+# Environments
+.env
+.venv
+env/
+venv/
+env.bak/
+venv.bak/
+
 # Spyder project settings
 .spyderproject
 .spyproject
@@ -134,3 +165,9 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/

.pre-commit-config.yaml

@@ -1,9 +1,31 @@
-exclude: ^(data/|tests/|diffusion_policy/)
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+exclude: "tests/artifacts/.*\\.safetensors$"
 default_language_version:
     python: python3.10
 repos:
+  ##### Meta #####
+  - repo: meta
+    hooks:
+      - id: check-useless-excludes
+      - id: check-hooks-apply
+
+
+  ##### Style / Misc. #####
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.5.0
+    rev: v5.0.0
     hooks:
       - id: check-added-large-files
       - id: debug-statements
@@ -13,21 +35,40 @@ repos:
       - id: check-toml
       - id: end-of-file-fixer
       - id: trailing-whitespace
+
+  - repo: https://github.com/adhtruong/mirrors-typos
+    rev: v1.31.1
+    hooks:
+      - id: typos
+        args: [--force-exclude]
+
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.15.1
+    rev: v3.19.1
     hooks:
     -   id: pyupgrade
+
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.2.2
+    rev: v0.11.4
     hooks:
       - id: ruff
         args: [--fix]
       - id: ruff-format
-  - repo: https://github.com/python-poetry/poetry
-    rev: 1.8.0
+
+
+  ##### Security #####
+  - repo: https://github.com/gitleaks/gitleaks
+    rev: v8.24.2
     hooks:
-      - id: poetry-check
-      - id: poetry-lock
-        args:
-          - "--check"
-          - "--no-update"
+      - id: gitleaks
+
+  - repo: https://github.com/woodruffw/zizmor-pre-commit
+    rev: v1.5.2
+    hooks:
+      - id: zizmor
+
+  - repo: https://github.com/PyCQA/bandit
+    rev: 1.8.3
+    hooks:
+    - id: bandit
+      args: ["-c", "pyproject.toml"]
+      additional_dependencies: ["bandit[toml]"]

CODE_OF_CONDUCT.md

@@ -0,0 +1,133 @@
+
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our
+community a harassment-free experience for everyone, regardless of age, body
+size, visible or invisible disability, ethnicity, sex characteristics, gender
+identity and expression, level of experience, education, socio-economic status,
+nationality, personal appearance, race, caste, color, religion, or sexual
+identity and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming,
+diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment for our
+community include:
+
+* Demonstrating empathy and kindness toward other people
+* Being respectful of differing opinions, viewpoints, and experiences
+* Giving and gracefully accepting constructive feedback
+* Accepting responsibility and apologizing to those affected by our mistakes,
+  and learning from the experience
+* Focusing on what is best not just for us as individuals, but for the overall
+  community
+
+Examples of unacceptable behavior include:
+
+* The use of sexualized language or imagery, and sexual attention or advances of
+  any kind
+* Trolling, insulting or derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or email address,
+  without their explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of
+acceptable behavior and will take appropriate and fair corrective action in
+response to any behavior that they deem inappropriate, threatening, offensive,
+or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject
+comments, commits, code, wiki edits, issues, and other contributions that are
+not aligned to this Code of Conduct, and will communicate reasons for moderation
+decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when
+an individual is officially representing the community in public spaces.
+Examples of representing our community include using an official email address,
+posting via an official social media account, or acting as an appointed
+representative at an online or offline event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported to the community leaders responsible for enforcement at
+[feedback@huggingface.co](mailto:feedback@huggingface.co).
+All complaints will be reviewed and investigated promptly and fairly.
+
+All community leaders are obligated to respect the privacy and security of the
+reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining
+the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed
+unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing
+clarity around the nature of the violation and an explanation of why the
+behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series of
+actions.
+
+**Consequence**: A warning with consequences for continued behavior. No
+interaction with the people involved, including unsolicited interaction with
+those enforcing the Code of Conduct, for a specified period of time. This
+includes avoiding interactions in community spaces as well as external channels
+like social media. Violating these terms may lead to a temporary or permanent
+ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including
+sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public
+communication with the community for a specified period of time. No public or
+private interaction with the people involved, including unsolicited interaction
+with those enforcing the Code of Conduct, is allowed during this period.
+Violating these terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community
+standards, including sustained inappropriate behavior, harassment of an
+individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within the
+community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage],
+version 2.1, available at
+[https://www.contributor-covenant.org/version/2/1/code_of_conduct.html][v2.1].
+
+Community Impact Guidelines were inspired by
+[Mozilla's code of conduct enforcement ladder][Mozilla CoC].
+
+For answers to common questions about this code of conduct, see the FAQ at
+[https://www.contributor-covenant.org/faq][FAQ]. Translations are available at
+[https://www.contributor-covenant.org/translations][translations].
+
+[homepage]: https://www.contributor-covenant.org
+[v2.1]: https://www.contributor-covenant.org/version/2/1/code_of_conduct.html
+[Mozilla CoC]: https://github.com/mozilla/diversity
+[FAQ]: https://www.contributor-covenant.org/faq
+[translations]: https://www.contributor-covenant.org/translations

CONTRIBUTING.md

@@ -0,0 +1,308 @@
+# How to contribute to 🤗 LeRobot?
+
+Everyone is welcome to contribute, and we value everybody's contribution. Code
+is thus not the only way to help the community. Answering questions, helping
+others, reaching out and improving the documentations are immensely valuable to
+the community.
+
+It also helps us if you spread the word: reference the library from blog posts
+on the awesome projects it made possible, shout out on Twitter when it has
+helped you, or simply ⭐️ the repo to say "thank you".
+
+Whichever way you choose to contribute, please be mindful to respect our
+[code of conduct](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md).
+
+## You can contribute in so many ways!
+
+Some of the ways you can contribute to 🤗 LeRobot:
+* Fixing outstanding issues with the existing code.
+* Implementing new models, datasets or simulation environments.
+* Contributing to the examples or to the documentation.
+* Submitting issues related to bugs or desired new features.
+
+Following the guides below, feel free to open issues and PRs and to coordinate your efforts with the community on our [Discord Channel](https://discord.gg/VjFz58wn3R). For specific inquiries, reach out to [Remi Cadene](mailto:remi.cadene@huggingface.co).
+
+If you are not sure how to contribute or want to know the next features we working on, look on this project page: [LeRobot TODO](https://github.com/orgs/huggingface/projects/46)
+
+## Submitting a new issue or feature request
+
+Do your best to follow these guidelines when submitting an issue or a feature
+request. It will make it easier for us to come back to you quickly and with good
+feedback.
+
+### Did you find a bug?
+
+The 🤗 LeRobot library is robust and reliable thanks to the users who notify us of
+the problems they encounter. So thank you for reporting an issue.
+
+First, we would really appreciate it if you could **make sure the bug was not
+already reported** (use the search bar on Github under Issues).
+
+Did not find it? :( So we can act quickly on it, please follow these steps:
+
+* Include your **OS type and version**, the versions of **Python** and **PyTorch**.
+* A short, self-contained, code snippet that allows us to reproduce the bug in
+  less than 30s.
+* The full traceback if an exception is raised.
+* Attach any other additional information, like screenshots, you think may help.
+
+### Do you want a new feature?
+
+A good feature request addresses the following points:
+
+1. Motivation first:
+* Is it related to a problem/frustration with the library? If so, please explain
+  why. Providing a code snippet that demonstrates the problem is best.
+* Is it related to something you would need for a project? We'd love to hear
+  about it!
+* Is it something you worked on and think could benefit the community?
+  Awesome! Tell us what problem it solved for you.
+2. Write a *paragraph* describing the feature.
+3. Provide a **code snippet** that demonstrates its future use.
+4. In case this is related to a paper, please attach a link.
+5. Attach any additional information (drawings, screenshots, etc.) you think may help.
+
+If your issue is well written we're already 80% of the way there by the time you
+post it.
+
+## Adding new policies, datasets or environments
+
+Look at our implementations for [datasets](./lerobot/common/datasets/), [policies](./lerobot/common/policies/),
+environments ([aloha](https://github.com/huggingface/gym-aloha),
+[xarm](https://github.com/huggingface/gym-xarm),
+[pusht](https://github.com/huggingface/gym-pusht))
+and follow the same api design.
+
+When implementing a new dataset loadable with LeRobotDataset follow these steps:
+- Update `available_datasets_per_env` in `lerobot/__init__.py`
+
+When implementing a new environment (e.g. `gym_aloha`), follow these steps:
+- Update `available_tasks_per_env` and `available_datasets_per_env` in `lerobot/__init__.py`
+
+When implementing a new policy class (e.g. `DiffusionPolicy`) follow these steps:
+- Update `available_policies` and `available_policies_per_env`, in `lerobot/__init__.py`
+- Set the required `name` class attribute.
+- Update variables in `tests/test_available.py` by importing your new Policy class
+
+## Submitting a pull request (PR)
+
+Before writing code, we strongly advise you to search through the existing PRs or
+issues to make sure that nobody is already working on the same thing. If you are
+unsure, it is always a good idea to open an issue to get some feedback.
+
+You will need basic `git` proficiency to be able to contribute to
+🤗 LeRobot. `git` is not the easiest tool to use but it has the greatest
+manual. Type `git --help` in a shell and enjoy. If you prefer books, [Pro
+Git](https://git-scm.com/book/en/v2) is a very good reference.
+
+Follow these steps to start contributing:
+
+1. Fork the [repository](https://github.com/huggingface/lerobot) by
+   clicking on the 'Fork' button on the repository's page. This creates a copy of the code
+   under your GitHub user account.
+
+2. Clone your fork to your local disk, and add the base repository as a remote. The following command
+   assumes you have your public SSH key uploaded to GitHub. See the following guide for more
+   [information](https://docs.github.com/en/repositories/creating-and-managing-repositories/cloning-a-repository).
+
+   ```bash
+   git clone git@github.com:<your Github handle>/lerobot.git
+   cd lerobot
+   git remote add upstream https://github.com/huggingface/lerobot.git
+   ```
+
+3. Create a new branch to hold your development changes, and do this for every new PR you work on.
+
+   Start by synchronizing your `main` branch with the `upstream/main` branch (more details in the [GitHub Docs](https://docs.github.com/en/github/collaborating-with-issues-and-pull-requests/syncing-a-fork)):
+
+   ```bash
+   git checkout main
+   git fetch upstream
+   git rebase upstream/main
+   ```
+
+   Once your `main` branch is synchronized, create a new branch from it:
+
+   ```bash
+   git checkout -b a-descriptive-name-for-my-changes
+   ```
+
+   🚨 **Do not** work on the `main` branch.
+
+4. for development, we advise to use a tool like `poetry` or `uv` instead of just `pip` to easily track our dependencies.
+   Follow the instructions to [install poetry](https://python-poetry.org/docs/#installation) (use a version >=2.1.0) or to [install uv](https://docs.astral.sh/uv/getting-started/installation/#installation-methods) if you don't have one of them already.
+
+   Set up a development environment with conda or miniconda:
+   ```bash
+   conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev
+   ```
+
+   If you're using `uv`, it can manage python versions so you can instead do:
+   ```bash
+   uv venv --python 3.10 && source .venv/bin/activate
+   ```
+
+   To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
+
+   using `poetry`
+   ```bash
+   poetry sync --extras "dev test"
+   ```
+
+   using `uv`
+   ```bash
+   uv sync --extra dev --extra test
+   ```
+
+   You can also install the project with all its dependencies (including environments):
+
+   using `poetry`
+   ```bash
+   poetry sync --all-extras
+   ```
+
+   using `uv`
+   ```bash
+   uv sync --all-extras
+   ```
+
+   > **Note:** If you don't install simulation environments with `--all-extras`, the tests that require them will be skipped when running the pytest suite locally. However, they *will* be tested in the CI. In general, we advise you to install everything and test locally before pushing.
+
+   Whichever command you chose to install the project (e.g. `poetry sync --all-extras`), you should run it again when pulling code with an updated version of `pyproject.toml` and `poetry.lock` in order to synchronize your virtual environment with the new dependencies.
+
+   The equivalent of `pip install some-package`, would just be:
+
+   using `poetry`
+   ```bash
+   poetry add some-package
+   ```
+
+   using `uv`
+   ```bash
+   uv add some-package
+   ```
+
+   When making changes to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
+   using `poetry`
+   ```bash
+   poetry lock
+   ```
+
+   using `uv`
+   ```bash
+   uv lock
+   ```
+
+
+5. Develop the features on your branch.
+
+   As you work on the features, you should make sure that the test suite
+   passes. You should run the tests impacted by your changes like this (see
+   below an explanation regarding the environment variable):
+
+   ```bash
+   pytest tests/<TEST_TO_RUN>.py
+   ```
+
+6. Follow our style.
+
+   `lerobot` relies on `ruff` to format its source code
+   consistently. Set up [`pre-commit`](https://pre-commit.com/) to run these checks
+   automatically as Git commit hooks.
+
+   Install `pre-commit` hooks:
+   ```bash
+   pre-commit install
+   ```
+
+   You can run these hooks whenever you need on staged files with:
+   ```bash
+   pre-commit
+   ```
+
+   Once you're happy with your changes, add changed files using `git add` and
+   make a commit with `git commit` to record your changes locally:
+
+   ```bash
+   git add modified_file.py
+   git commit
+   ```
+
+   Note, if you already committed some changes that have a wrong formatting, you can use:
+   ```bash
+   pre-commit run --all-files
+   ```
+
+   Please write [good commit messages](https://chris.beams.io/posts/git-commit/).
+
+   It is a good idea to sync your copy of the code with the original
+   repository regularly. This way you can quickly account for changes:
+
+   ```bash
+   git fetch upstream
+   git rebase upstream/main
+   ```
+
+   Push the changes to your account using:
+
+   ```bash
+   git push -u origin a-descriptive-name-for-my-changes
+   ```
+
+6. Once you are satisfied (**and the checklist below is happy too**), go to the
+   webpage of your fork on GitHub. Click on 'Pull request' to send your changes
+   to the project maintainers for review.
+
+7. It's ok if maintainers ask you for changes. It happens to core contributors
+   too! So everyone can see the changes in the Pull request, work in your local
+   branch and push the changes to your fork. They will automatically appear in
+   the pull request.
+
+
+### Checklist
+
+1. The title of your pull request should be a summary of its contribution;
+2. If your pull request addresses an issue, please mention the issue number in
+   the pull request description to make sure they are linked (and people
+   consulting the issue know you are working on it);
+3. To indicate a work in progress please prefix the title with `[WIP]`, or preferably mark
+   the PR as a draft PR. These are useful to avoid duplicated work, and to differentiate
+   it from PRs ready to be merged;
+4. Make sure existing tests pass;
+<!-- 5. Add high-coverage tests. No quality testing = no merge.
+
+See an example of a good PR here: https://github.com/huggingface/lerobot/pull/ -->
+
+### Tests
+
+An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the [tests folder](https://github.com/huggingface/lerobot/tree/main/tests).
+
+Install [git lfs](https://git-lfs.com/) to retrieve test artifacts (if you don't have it already).
+
+On Mac:
+```bash
+brew install git-lfs
+git lfs install
+```
+
+On Ubuntu:
+```bash
+sudo apt-get install git-lfs
+git lfs install
+```
+
+Pull artifacts if they're not in [tests/artifacts](tests/artifacts)
+```bash
+git lfs pull
+```
+
+We use `pytest` in order to run the tests. From the root of the
+repository, here's how to run tests with `pytest` for the library:
+
+```bash
+python -m pytest -sv ./tests
+```
+
+
+You can specify a smaller set of tests in order to test only the feature
+you're working on.

LICENSE

@@ -0,0 +1,507 @@
+Copyright 2024 The Hugging Face team. All rights reserved.
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+
+## Some of lerobot's code is derived from Diffusion Policy, which is subject to the following copyright notice:
+
+MIT License
+
+Copyright (c) 2023 Columbia Artificial Intelligence and Robotics Lab
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+
+## Some of lerobot's code is derived from FOWM, which is subject to the following copyright notice:
+
+MIT License
+
+Copyright (c) 2023 Yunhai Feng
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+
+## Some of lerobot's code is derived from simxarm, which is subject to the following copyright notice:
+
+MIT License
+
+Copyright (c) 2023 Nicklas Hansen & Yanjie Ze
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+
+## Some of lerobot's code is derived from ALOHA, which is subject to the following copyright notice:
+
+MIT License
+
+Copyright (c) 2023 Tony Z. Zhao
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+## Some of lerobot's code is derived from DETR, which is subject to the following copyright notice:
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2020 - present, Facebook, Inc
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

Makefile

@@ -0,0 +1,142 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+.PHONY: tests
+
+PYTHON_PATH := $(shell which python)
+
+# If uv is installed and a virtual environment exists, use it
+UV_CHECK := $(shell command -v uv)
+ifneq ($(UV_CHECK),)
+	PYTHON_PATH := $(shell .venv/bin/python)
+endif
+
+export PATH := $(dir $(PYTHON_PATH)):$(PATH)
+
+DEVICE ?= cpu
+
+build-cpu:
+	docker build -t lerobot:latest -f docker/lerobot-cpu/Dockerfile .
+
+build-gpu:
+	docker build -t lerobot:latest -f docker/lerobot-gpu/Dockerfile .
+
+test-end-to-end:
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train-resume
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-eval
+
+test-act-ete-train:
+	python lerobot/scripts/train.py \
+		--policy.type=act \
+		--policy.dim_model=64 \
+		--policy.n_action_steps=20 \
+		--policy.chunk_size=20 \
+		--policy.device=$(DEVICE) \
+		--env.type=aloha \
+		--env.episode_length=5 \
+		--dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
+		--dataset.image_transforms.enable=true \
+		--dataset.episodes="[0]" \
+		--batch_size=2 \
+		--steps=4 \
+		--eval_freq=2 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1 \
+		--save_freq=2 \
+		--save_checkpoint=true \
+		--log_freq=1 \
+		--wandb.enable=false \
+		--output_dir=tests/outputs/act/
+
+test-act-ete-train-resume:
+	python lerobot/scripts/train.py \
+		--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
+		--resume=true
+
+test-act-ete-eval:
+	python lerobot/scripts/eval.py \
+		--policy.path=tests/outputs/act/checkpoints/000004/pretrained_model \
+		--policy.device=$(DEVICE) \
+		--env.type=aloha \
+		--env.episode_length=5 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1
+
+test-diffusion-ete-train:
+	python lerobot/scripts/train.py \
+		--policy.type=diffusion \
+		--policy.down_dims='[64,128,256]' \
+		--policy.diffusion_step_embed_dim=32 \
+		--policy.num_inference_steps=10 \
+		--policy.device=$(DEVICE) \
+		--env.type=pusht \
+		--env.episode_length=5 \
+		--dataset.repo_id=lerobot/pusht \
+		--dataset.image_transforms.enable=true \
+		--dataset.episodes="[0]" \
+		--batch_size=2 \
+		--steps=2 \
+		--eval_freq=2 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1 \
+		--save_checkpoint=true \
+		--save_freq=2 \
+		--log_freq=1 \
+		--wandb.enable=false \
+		--output_dir=tests/outputs/diffusion/
+
+test-diffusion-ete-eval:
+	python lerobot/scripts/eval.py \
+		--policy.path=tests/outputs/diffusion/checkpoints/000002/pretrained_model \
+		--policy.device=$(DEVICE) \
+		--env.type=pusht \
+		--env.episode_length=5 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1
+
+test-tdmpc-ete-train:
+	python lerobot/scripts/train.py \
+		--policy.type=tdmpc \
+		--policy.device=$(DEVICE) \
+		--env.type=xarm \
+		--env.task=XarmLift-v0 \
+		--env.episode_length=5 \
+		--dataset.repo_id=lerobot/xarm_lift_medium \
+		--dataset.image_transforms.enable=true \
+		--dataset.episodes="[0]" \
+		--batch_size=2 \
+		--steps=2 \
+		--eval_freq=2 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1 \
+		--save_checkpoint=true \
+		--save_freq=2 \
+		--log_freq=1 \
+		--wandb.enable=false \
+		--output_dir=tests/outputs/tdmpc/
+
+test-tdmpc-ete-eval:
+	python lerobot/scripts/eval.py \
+		--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
+		--policy.device=$(DEVICE) \
+		--env.type=xarm \
+		--env.episode_length=5 \
+		--env.task=XarmLift-v0 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1

README.md

@@ -1,83 +1,319 @@
-# LeRobot
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="media/lerobot-logo-thumbnail.png">
+    <source media="(prefers-color-scheme: light)" srcset="media/lerobot-logo-thumbnail.png">
+    <img alt="LeRobot, Hugging Face Robotics Library" src="media/lerobot-logo-thumbnail.png" style="max-width: 100%;">
+  </picture>
+  <br/>
+  <br/>
+</p>
+
+<div align="center">
+
+[![Tests](https://github.com/huggingface/lerobot/actions/workflows/nightly-tests.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/nightly-tests.yml?query=branch%3Amain)
+[![Coverage](https://codecov.io/gh/huggingface/lerobot/branch/main/graph/badge.svg?token=TODO)](https://codecov.io/gh/huggingface/lerobot)
+[![Python versions](https://img.shields.io/pypi/pyversions/lerobot)](https://www.python.org/downloads/)
+[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/huggingface/lerobot/blob/main/LICENSE)
+[![Status](https://img.shields.io/pypi/status/lerobot)](https://pypi.org/project/lerobot/)
+[![Version](https://img.shields.io/pypi/v/lerobot)](https://pypi.org/project/lerobot/)
+[![Examples](https://img.shields.io/badge/Examples-green.svg)](https://github.com/huggingface/lerobot/tree/main/examples)
+[![Contributor Covenant](https://img.shields.io/badge/Contributor%20Covenant-v2.1%20adopted-ff69b4.svg)](https://github.com/huggingface/lerobot/blob/main/CODE_OF_CONDUCT.md)
+[![Discord](https://dcbadge.vercel.app/api/server/C5P34WJ68S?style=flat)](https://discord.gg/s3KuuzsPFb)
+
+</div>
+
+<h2 align="center">
+    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
+        Build Your Own SO-100 Robot!</a></p>
+</h2>
+
+<div align="center">
+  <img src="media/so100/leader_follower.webp?raw=true" alt="SO-100 leader and follower arms" title="SO-100 leader and follower arms" width="50%">
+
+  <p><strong>Meet the SO-100 – Just $110 per arm!</strong></p>
+  <p>Train it in minutes with a few simple moves on your laptop.</p>
+  <p>Then sit back and watch your creation act autonomously! 🤯</p>
+
+  <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
+      Get the full SO-100 tutorial here.</a></p>
+
+  <p>Want to take it to the next level? Make your SO-100 mobile by building LeKiwi!</p>
+  <p>Check out the <a href="https://github.com/huggingface/lerobot/blob/main/examples/11_use_lekiwi.md">LeKiwi tutorial</a> and bring your robot to life on wheels.</p>
+
+  <img src="media/lekiwi/kiwi.webp?raw=true" alt="LeKiwi mobile robot" title="LeKiwi mobile robot" width="50%">
+</div>
+
+<br/>
+
+<h3 align="center">
+    <p>LeRobot: State-of-the-art AI for real-world robotics</p>
+</h3>
+
+---
+
+
+🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
+
+🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
+
+🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulation environments to get started without assembling a robot. In the coming weeks, the plan is to add more and more support for real-world robotics on the most affordable and capable robots out there.
+
+🤗 LeRobot hosts pretrained models and datasets on this Hugging Face community page: [huggingface.co/lerobot](https://huggingface.co/lerobot)
+
+#### Examples of pretrained models on simulation environments
+
+<table>
+  <tr>
+    <td><img src="media/gym/aloha_act.gif" width="100%" alt="ACT policy on ALOHA env"/></td>
+    <td><img src="media/gym/simxarm_tdmpc.gif" width="100%" alt="TDMPC policy on SimXArm env"/></td>
+    <td><img src="media/gym/pusht_diffusion.gif" width="100%" alt="Diffusion policy on PushT env"/></td>
+  </tr>
+  <tr>
+    <td align="center">ACT policy on ALOHA env</td>
+    <td align="center">TDMPC policy on SimXArm env</td>
+    <td align="center">Diffusion policy on PushT env</td>
+  </tr>
+</table>
+
+### Acknowledgment
+
+- Thanks to Tony Zhao, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
+- Thanks to Cheng Chi, Zhenjia Xu and colleagues for open sourcing Diffusion policy, Pusht environment and datasets, as well as UMI datasets. Ours are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu) and [UMI Gripper](https://umi-gripper.github.io).
+- Thanks to Nicklas Hansen, Yunhai Feng and colleagues for open sourcing TDMPC policy, Simxarm environments and datasets. Ours are adapted from [TDMPC](https://github.com/nicklashansen/tdmpc) and [FOWM](https://www.yunhaifeng.com/FOWM).
+- Thanks to Antonio Loquercio and Ashish Kumar for their early support.
+- Thanks to [Seungjae (Jay) Lee](https://sjlee.cc/), [Mahi Shafiullah](https://mahis.life/) and colleagues for open sourcing [VQ-BeT](https://sjlee.cc/vq-bet/) policy and helping us adapt the codebase to our repository. The policy is adapted from [VQ-BeT repo](https://github.com/jayLEE0301/vq_bet_official).
+
 
 ## Installation
 
-Create a virtual environment with python 3.10, e.g. using `conda`:
+Download our source code:
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
 ```
+
+Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/free/miniconda/index.html):
+```bash
 conda create -y -n lerobot python=3.10
 conda activate lerobot
 ```
 
-[Install `poetry`](https://python-poetry.org/docs/#installation) (if you don't have it already)
-```
-curl -sSL https://install.python-poetry.org | python -
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
 ```
 
-Install dependencies
-```
-poetry install
+Install 🤗 LeRobot:
+```bash
+pip install -e .
 ```
 
-If you encounter a disk space error, try to change your tmp dir to a location where you have enough disk space, e.g.
-```
-mkdir ~/tmp
-export TMPDIR='~/tmp'
+> **NOTE:** If you encounter build errors, you may need to install additional dependencies (`cmake`, `build-essential`, and `ffmpeg libs`). On Linux, run:
+`sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
+
+For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
+- [aloha](https://github.com/huggingface/gym-aloha)
+- [xarm](https://github.com/huggingface/gym-xarm)
+- [pusht](https://github.com/huggingface/gym-pusht)
+
+For instance, to install 🤗 LeRobot with aloha and pusht, use:
+```bash
+pip install -e ".[aloha, pusht]"
 ```
 
-Install `diffusion_policy` #HACK
-```
-# from this directory
-git clone https://github.com/real-stanford/diffusion_policy
-cp -r diffusion_policy/diffusion_policy $(poetry env info -p)/lib/python3.10/site-packages/
+To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
+```bash
+wandb login
 ```
 
-## Usage
+(note: you will also need to enable WandB in the configuration. See below.)
 
-
-### Train
+## Walkthrough
 
 ```
-python lerobot/scripts/train.py \
-hydra.job.name=pusht \
-env=pusht
+.
+├── examples             # contains demonstration examples, start here to learn about LeRobot
+|   └── advanced         # contains even more examples for those who have mastered the basics
+├── lerobot
+|   ├── configs          # contains config classes with all options that you can override in the command line
+|   ├── common           # contains classes and utilities
+|   |   ├── datasets       # various datasets of human demonstrations: aloha, pusht, xarm
+|   |   ├── envs           # various sim environments: aloha, pusht, xarm
+|   |   ├── policies       # various policies: act, diffusion, tdmpc
+|   |   ├── robot_devices  # various real devices: dynamixel motors, opencv cameras, koch robots
+|   |   └── utils          # various utilities
+|   └── scripts          # contains functions to execute via command line
+|       ├── eval.py                 # load policy and evaluate it on an environment
+|       ├── train.py                # train a policy via imitation learning and/or reinforcement learning
+|       ├── control_robot.py        # teleoperate a real robot, record data, run a policy
+|       ├── push_dataset_to_hub.py  # convert your dataset into LeRobot dataset format and upload it to the Hugging Face hub
+|       └── visualize_dataset.py    # load a dataset and render its demonstrations
+├── outputs               # contains results of scripts execution: logs, videos, model checkpoints
+└── tests                 # contains pytest utilities for continuous integration
 ```
 
-### Visualize offline buffer
+### Visualize datasets
 
-```
+Check out [example 1](./examples/1_load_lerobot_dataset.py) that illustrates how to use our dataset class which automatically downloads data from the Hugging Face hub.
+
+You can also locally visualize episodes from a dataset on the hub by executing our script from the command line:
+```bash
 python lerobot/scripts/visualize_dataset.py \
-hydra.run.dir=tmp/$(date +"%Y_%m_%d") \
-env=pusht
+    --repo-id lerobot/pusht \
+    --episode-index 0
 ```
 
-### Visualize online buffer / Eval
+or from a dataset in a local folder with the `root` option and the `--local-files-only` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
+```bash
+python lerobot/scripts/visualize_dataset.py \
+    --repo-id lerobot/pusht \
+    --root ./my_local_data_dir \
+    --local-files-only 1 \
+    --episode-index 0
+```
+
+
+It will open `rerun.io` and display the camera streams, robot states and actions, like this:
+
+https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-fd46b787-b532-47e2-bb6f-fd536a55a7ed.mov?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20240505%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240505T172924Z&X-Amz-Expires=300&X-Amz-Signature=d680b26c532eeaf80740f08af3320d22ad0b8a4e4da1bcc4f33142c15b509eda&X-Amz-SignedHeaders=host&actor_id=24889239&key_id=0&repo_id=748713144
+
+
+Our script can also visualize datasets stored on a distant server. See `python lerobot/scripts/visualize_dataset.py --help` for more instructions.
+
+### The `LeRobotDataset` format
+
+A dataset in `LeRobotDataset` format is very simple to use. It can be loaded from a repository on the Hugging Face hub or a local folder simply with e.g. `dataset = LeRobotDataset("lerobot/aloha_static_coffee")` and can be indexed into like any Hugging Face and PyTorch dataset. For instance `dataset[0]` will retrieve a single temporal frame from the dataset containing observation(s) and an action as PyTorch tensors ready to be fed to a model.
+
+A specificity of `LeRobotDataset` is that, rather than retrieving a single frame by its index, we can retrieve several frames based on their temporal relationship with the indexed frame, by setting `delta_timestamps` to a list of relative times with respect to the indexed frame. For example, with `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}`  one can retrieve, for a given index, 4 frames: 3 "previous" frames 1 second, 0.5 seconds, and 0.2 seconds before the indexed frame, and the indexed frame itself (corresponding to the 0 entry). See example [1_load_lerobot_dataset.py](examples/1_load_lerobot_dataset.py) for more details on `delta_timestamps`.
+
+Under the hood, the `LeRobotDataset` format makes use of several ways to serialize data which can be useful to understand if you plan to work more closely with this format. We tried to make a flexible yet simple dataset format that would cover most type of features and specificities present in reinforcement learning and robotics, in simulation and in real-world, with a focus on cameras and robot states but easily extended to other types of sensory inputs as long as they can be represented by a tensor.
+
+Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
 
 ```
+dataset attributes:
+  ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
+  │  ├ observation.images.cam_high (VideoFrame):
+  │  │   VideoFrame = {'path': path to a mp4 video, 'timestamp' (float32): timestamp in the video}
+  │  ├ observation.state (list of float32): position of an arm joints (for instance)
+  │  ... (more observations)
+  │  ├ action (list of float32): goal position of an arm joints (for instance)
+  │  ├ episode_index (int64): index of the episode for this sample
+  │  ├ frame_index (int64): index of the frame for this sample in the episode ; starts at 0 for each episode
+  │  ├ timestamp (float32): timestamp in the episode
+  │  ├ next.done (bool): indicates the end of en episode ; True for the last frame in each episode
+  │  └ index (int64): general index in the whole dataset
+  ├ episode_data_index: contains 2 tensors with the start and end indices of each episode
+  │  ├ from (1D int64 tensor): first frame index for each episode — shape (num episodes,) starts with 0
+  │  └ to: (1D int64 tensor): last frame index for each episode — shape (num episodes,)
+  ├ stats: a dictionary of statistics (max, mean, min, std) for each feature in the dataset, for instance
+  │  ├ observation.images.cam_high: {'max': tensor with same number of dimensions (e.g. `(c, 1, 1)` for images, `(c,)` for states), etc.}
+  │  ...
+  ├ info: a dictionary of metadata on the dataset
+  │  ├ codebase_version (str): this is to keep track of the codebase version the dataset was created with
+  │  ├ fps (float): frame per second the dataset is recorded/synchronized to
+  │  ├ video (bool): indicates if frames are encoded in mp4 video files to save space or stored as png files
+  │  └ encoding (dict): if video, this documents the main options that were used with ffmpeg to encode the videos
+  ├ videos_dir (Path): where the mp4 videos or png images are stored/accessed
+  └ camera_keys (list of string): the keys to access camera features in the item returned by the dataset (e.g. `["observation.images.cam_high", ...]`)
+```
+
+A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
+- hf_dataset stored using Hugging Face datasets library serialization to parquet
+- videos are stored in mp4 format to save space
+- metadata are stored in plain json/jsonl files
+
+Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
+
+### Evaluate a pretrained policy
+
+Check out [example 2](./examples/2_evaluate_pretrained_policy.py) that illustrates how to download a pretrained policy from Hugging Face hub, and run an evaluation on its corresponding environment.
+
+We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):
+```bash
 python lerobot/scripts/eval.py \
-hydra.run.dir=tmp/$(date +"%Y_%m_%d") \
-env=pusht
+    --policy.path=lerobot/diffusion_pusht \
+    --env.type=pusht \
+    --eval.batch_size=10 \
+    --eval.n_episodes=10 \
+    --policy.use_amp=false \
+    --policy.device=cuda
 ```
 
+Note: After training your own policy, you can re-evaluate the checkpoints with:
 
-## TODO
+```bash
+python lerobot/scripts/eval.py --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
+```
 
-- [x] priority update doesnt match FOWM or original paper
-- [x] self.step=100000 should be updated at every step to adjust to horizon of planner
-- [ ] prefetch replay buffer to speedup training
-- [ ] parallelize env to speedup eval
-- [ ] clean checkpointing / loading
-- [ ] clean logging
-- [ ] clean config
-- [ ] clean hyperparameter tuning
-- [ ] add pusht
-- [ ] add aloha
-- [ ] add act
-- [ ] add diffusion
-- [ ] add aloha 2
+See `python lerobot/scripts/eval.py --help` for more instructions.
 
-## Profile
+### Train your own policy
 
-**Example**
+Check out [example 3](./examples/3_train_policy.py) that illustrate how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
+
+To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding `--wandb.enable=true`.
+
+A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser. Please also check [here](./examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explanation of some commonly used metrics in logs.
+
+![](media/wandb.png)
+
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
+
+#### Reproduce state-of-the-art (SOTA)
+
+We provide some pretrained policies on our [hub page](https://huggingface.co/lerobot) that can achieve state-of-the-art performances.
+You can reproduce their training by loading the config from their run. Simply running:
+```bash
+python lerobot/scripts/train.py --config_path=lerobot/diffusion_pusht
+```
+reproduces SOTA results for Diffusion Policy on the PushT task.
+
+## Contribute
+
+If you would like to contribute to 🤗 LeRobot, please check out our [contribution guide](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md).
+
+<!-- ### Add a new dataset
+
+To add a dataset to the hub, you need to login using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Then point to your raw dataset folder (e.g. `data/aloha_static_pingpong_test_raw`), and push your dataset to the hub with:
+```bash
+python lerobot/scripts/push_dataset_to_hub.py \
+--raw-dir data/aloha_static_pingpong_test_raw \
+--out-dir data \
+--repo-id lerobot/aloha_static_pingpong_test \
+--raw-format aloha_hdf5
+```
+
+See `python lerobot/scripts/push_dataset_to_hub.py --help` for more instructions.
+
+If your dataset format is not supported, implement your own in `lerobot/common/datasets/push_dataset_to_hub/${raw_format}_format.py` by copying examples like [pusht_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py), [umi_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py), [aloha_hdf5](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py), or [xarm_pkl](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py). -->
+
+
+### Add a pretrained policy
+
+Once you have trained a policy you may upload it to the Hugging Face hub using a hub id that looks like `${hf_user}/${repo_name}` (e.g. [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)).
+
+You first need to find the checkpoint folder located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). Within that there is a `pretrained_model` directory which should contain:
+- `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
+- `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
+- `train_config.json`: A consolidated configuration containing all parameter userd for training. The policy configuration should match `config.json` exactly. Thisis useful for anyone who wants to evaluate your policy or for reproducibility.
+
+To upload these to the hub, run the following:
+```bash
+huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
+```
+
+See [eval.py](https://github.com/huggingface/lerobot/blob/main/lerobot/scripts/eval.py) for an example of how other people may use your policy.
+
+
+### Improve your code with profiling
+
+An example of a code snippet to profile the evaluation of a policy:
 ```python
 from torch.profiler import profile, record_function, ProfilerActivity
 
@@ -96,25 +332,62 @@ with profile(
     with record_function("eval_policy"):
         for i in range(num_episodes):
             prof.step()
+            # insert code to profile, potentially whole body of eval_policy function
 ```
 
-```bash
-python lerobot/scripts/eval.py \
-pretrained_model_path=/home/rcadene/code/fowm/logs/xarm_lift/all/default/2/models/final.pt \
-eval_episodes=7
+## Citation
+
+If you want, you can cite this work with:
+```bibtex
+@misc{cadene2024lerobot,
+    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Wolf, Thomas},
+    title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
+    howpublished = "\url{https://github.com/huggingface/lerobot}",
+    year = {2024}
+}
 ```
 
-## Contribute
+Additionally, if you are using any of the particular policy architecture, pretrained models, or datasets, it is recommended to cite the original authors of the work as they appear below:
 
-**Style**
+- [Diffusion Policy](https://diffusion-policy.cs.columbia.edu)
+```bibtex
+@article{chi2024diffusionpolicy,
+	author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
+	title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+	journal = {The International Journal of Robotics Research},
+	year = {2024},
+}
 ```
-# install if needed
-pre-commit install
-# apply style and linter checks before git commit
-pre-commit run -a
+- [ACT or ALOHA](https://tonyzhaozh.github.io/aloha)
+```bibtex
+@article{zhao2023learning,
+  title={Learning fine-grained bimanual manipulation with low-cost hardware},
+  author={Zhao, Tony Z and Kumar, Vikash and Levine, Sergey and Finn, Chelsea},
+  journal={arXiv preprint arXiv:2304.13705},
+  year={2023}
+}
 ```
 
-**Tests**
+- [TDMPC](https://www.nicklashansen.com/td-mpc/)
+
+```bibtex
+@inproceedings{Hansen2022tdmpc,
+	title={Temporal Difference Learning for Model Predictive Control},
+	author={Nicklas Hansen and Xiaolong Wang and Hao Su},
+	booktitle={ICML},
+	year={2022}
+}
 ```
-pytest -sx tests
+
+- [VQ-BeT](https://sjlee.cc/vq-bet/)
+```bibtex
+@article{lee2024behavior,
+  title={Behavior generation with latent actions},
+  author={Lee, Seungjae and Wang, Yibin and Etukuru, Haritheja and Kim, H Jin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},
+  journal={arXiv preprint arXiv:2403.03181},
+  year={2024}
+}
 ```
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)

benchmarks/video/README.md

@@ -0,0 +1,271 @@
+# Video benchmark
+
+
+## Questions
+What is the optimal trade-off between:
+- maximizing loading time with random access,
+- minimizing memory space on disk,
+- maximizing success rate of policies,
+- compatibility across devices/platforms for decoding videos (e.g. video players, web browsers).
+
+How to encode videos?
+- Which video codec (`-vcodec`) to use? h264, h265, AV1?
+- What pixel format to use (`-pix_fmt`)? `yuv444p` or `yuv420p`?
+- How much compression (`-crf`)? No compression with `0`, intermediate compression with `25` or extreme with `50+`?
+- Which frequency to chose for key frames (`-g`)? A key frame every `10` frames?
+
+How to decode videos?
+- Which `decoder`? `torchvision`, `torchaudio`, `ffmpegio`, `decord`, or `nvc`?
+- What scenarios to use for the requesting timestamps during benchmark? (`timestamps_mode`)
+
+
+## Variables
+**Image content & size**
+We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an apartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
+For these reasons, we run this benchmark on four representative datasets:
+- `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
+- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+
+Note: The datasets used for this benchmark need to be image datasets, not video datasets.
+
+**Data augmentations**
+We might revisit this benchmark and find better settings if we train our policies with various data augmentations to make them more robust (e.g. robust to color changes, compression, etc.).
+
+### Encoding parameters
+| parameter   | values                                                       |
+|-------------|--------------------------------------------------------------|
+| **vcodec**  | `libx264`, `libx265`, `libsvtav1`                            |
+| **pix_fmt** | `yuv444p`, `yuv420p`                                         |
+| **g**       | `1`, `2`, `3`, `4`, `5`, `6`, `10`, `15`, `20`, `40`, `None` |
+| **crf**     | `0`, `5`, `10`, `15`, `20`, `25`, `30`, `40`, `50`, `None`   |
+
+Note that `crf` value might be interpreted differently by various video codecs. In other words, the same value used with one codec doesn't necessarily translate into the same compression level with another codec. In fact, the default value (`None`) isn't the same amongst the different video codecs. Importantly, it is also the case for many other ffmpeg arguments like `g` which specifies the frequency of the key frames.
+
+For a comprehensive list and documentation of these parameters, see the ffmpeg documentation depending on the video codec used:
+- h264: https://trac.ffmpeg.org/wiki/Encode/H.264
+- h265: https://trac.ffmpeg.org/wiki/Encode/H.265
+- AV1: https://trac.ffmpeg.org/wiki/Encode/AV1
+
+### Decoding parameters
+**Decoder**
+We tested two video decoding backends from torchvision:
+- `pyav`
+- `video_reader` (requires to build torchvision from source)
+
+**Requested timestamps**
+Given the way video decoding works, once a keyframe has been loaded, the decoding of subsequent frames is fast.
+This of course is affected by the `-g` parameter during encoding, which specifies the frequency of the keyframes. Given our typical use cases in robotics policies which might request a few timestamps in different random places, we want to replicate these use cases with the following scenarios:
+- `1_frame`: 1 frame,
+- `2_frames`: 2 consecutive frames (e.g. `[t, t + 1 / fps]`),
+- `6_frames`: 6 consecutive frames (e.g. `[t + i / fps for i in range(6)]`)
+
+Note that this differs significantly from a typical use case like watching a movie, in which every frame is loaded sequentially from the beginning to the end and it's acceptable to have big values for `-g`.
+
+Additionally, because some policies might request single timestamps that are a few frames apart, we also have the following scenario:
+- `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
+
+However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
+
+
+## Metrics
+**Data compression ratio (lower is better)**
+`video_images_size_ratio` is the ratio of the memory space on disk taken by the encoded video over the memory space taken by the original images. For instance, `video_images_size_ratio=25%` means that the video takes 4 times less memory space on disk compared to the original images.
+
+**Loading time ratio (lower is better)**
+`video_images_load_time_ratio` is the ratio of the time it takes to decode frames from the video at a given timestamps over the time it takes to load the exact same original images. Lower is better. For instance, `video_images_load_time_ratio=200%` means that decoding from video is 2 times slower than loading the original images.
+
+**Average Mean Square Error (lower is better)**
+`avg_mse` is the average mean square error between each decoded frame and its corresponding original image over all requested timestamps, and also divided by the number of pixels in the image to be comparable when switching to different image sizes.
+
+**Average Peak Signal to Noise Ratio (higher is better)**
+`avg_psnr` measures the ratio between the maximum possible power of a signal and the power of corrupting noise that affects the fidelity of its representation. Higher PSNR indicates better quality.
+
+**Average Structural Similarity Index Measure (higher is better)**
+`avg_ssim` evaluates the perceived quality of images by comparing luminance, contrast, and structure. SSIM values range from -1 to 1, where 1 indicates perfect similarity.
+
+One aspect that can't be measured here with those metrics is the compatibility of the encoding across platforms, in particular on web browser, for visualization purposes.
+h264, h265 and AV1 are all commonly used codecs and should not pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
+- `yuv420p` is more widely supported across various platforms, including web browsers.
+- `yuv444p` offers higher color fidelity but might not be supported as broadly.
+
+
+<!-- **Loss of a pretrained policy (higher is better)** (not available)
+`loss_pretrained` is the result of evaluating with the selected encoding/decoding settings a policy pretrained on original images. It is easier to understand than `avg_l2_error`.
+
+**Success rate after retraining (higher is better)** (not available)
+`success_rate` is the result of training and evaluating a policy with the selected encoding/decoding settings. It is the most difficult metric to get but also the very best. -->
+
+
+## How the benchmark works
+The benchmark evaluates both encoding and decoding of video frames on the first episode of each dataset.
+
+**Encoding:** for each `vcodec` and `pix_fmt` pair, we use a default value for `g` and `crf` upon which we change a single value (either `g` or `crf`) to one of the specified values (we don't test every combination of those as this would be computationally too heavy).
+This gives a unique set of encoding parameters which is used to encode the episode.
+
+**Decoding:** Then, for each of those unique encodings, we iterate through every combination of the decoding parameters `backend` and `timestamps_mode`. For each of them, we record the metrics of a number of samples (given by `--num-samples`). This is parallelized for efficiency and the number of processes can be controlled with `--num-workers`. Ideally, it's best to have a `--num-samples` that is divisible by `--num-workers`.
+
+Intermediate results saved for each `vcodec` and `pix_fmt` combination in csv tables.
+These are then all concatenated to a single table ready for analysis.
+
+## Caveats
+We tried to measure the most impactful parameters for both encoding and decoding. However, for computational reasons we can't test out every combination.
+
+Additional encoding parameters exist that are not included in this benchmark. In particular:
+- `-preset` which allows for selecting encoding presets. This represents a collection of options that will provide a certain encoding speed to compression ratio. By leaving this parameter unspecified, it is considered to be `medium` for libx264 and libx265 and `8` for libsvtav1.
+- `-tune` which allows to optimize the encoding for certain aspects (e.g. film quality, fast decoding, etc.).
+
+See the documentation mentioned above for more detailed info on these settings and for a more comprehensive list of other parameters.
+
+Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
+- `torchaudio`
+- `ffmpegio`
+- `decord`
+- `nvc`
+
+Note as well that since we are mostly interested in the performance at decoding time (also because encoding is done only once before uploading a dataset), we did not measure encoding times nor have any metrics regarding encoding.
+However, besides the necessity to build ffmpeg from source, encoding did not pose any issue and it didn't take a significant amount of time during this benchmark.
+
+
+## Install
+Building ffmpeg from source is required to include libx265 and libaom/libsvtav1 (av1) video codecs ([compilation guide](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu)).
+
+**Note:** While you still need to build torchvision with a conda-installed `ffmpeg<4.3` to use the `video_reader` decoder (as described in [#220](https://github.com/huggingface/lerobot/pull/220)), you also need another version which is custom-built with all the video codecs for encoding. For the script to then use that version, you can prepend the command above with `PATH="$HOME/bin:$PATH"`, which is where ffmpeg should be built.
+
+
+## Adding a video decoder
+Right now, we're only benchmarking the two video decoder available with torchvision: `pyav` and `video_reader`.
+You can easily add a new decoder to benchmark by adding it to this function in the script:
+```diff
+def decode_video_frames(
+    video_path: str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str,
+) -> torch.Tensor:
+    if backend in ["pyav", "video_reader"]:
+        return decode_video_frames_torchvision(
+            video_path, timestamps, tolerance_s, backend
+        )
++    elif backend == ["your_decoder"]:
++        return your_decoder_function(
++            video_path, timestamps, tolerance_s, backend
++        )
+    else:
+        raise NotImplementedError(backend)
+```
+
+
+## Example
+For a quick run, you can try these parameters:
+```bash
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        aliberts/aloha_mobile_shrimp_image \
+    --vcodec libx264 libx265 \
+    --pix-fmt yuv444p yuv420p \
+    --g 2 20 None \
+    --crf 10 40 None \
+    --timestamps-modes 1_frame 2_frames \
+    --backends pyav video_reader \
+    --num-samples 5 \
+    --num-workers 5 \
+    --save-frames 0
+```
+
+
+## Results
+
+### Reproduce
+We ran the benchmark with the following parameters:
+```bash
+# h264 and h265 encodings
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/kitchen \
+    --vcodec libx264 libx265 \
+    --pix-fmt yuv444p yuv420p \
+    --g 1 2 3 4 5 6 10 15 20 40 None \
+    --crf 0 5 10 15 20 25 30 40 50 None \
+    --timestamps-modes 1_frame 2_frames 6_frames \
+    --backends pyav video_reader \
+    --num-samples 50 \
+    --num-workers 5 \
+    --save-frames 1
+
+# av1 encoding (only compatible with yuv420p and pyav decoder)
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/kitchen \
+    --vcodec libsvtav1 \
+    --pix-fmt yuv420p \
+    --g 1 2 3 4 5 6 10 15 20 40 None \
+    --crf 0 5 10 15 20 25 30 40 50 None \
+    --timestamps-modes 1_frame 2_frames 6_frames \
+    --backends pyav \
+    --num-samples 50 \
+    --num-workers 5 \
+    --save-frames 1
+```
+
+The full results are available [here](https://docs.google.com/spreadsheets/d/1OYJB43Qu8fC26k_OyoMFgGBBKfQRCi4BIuYitQnq3sw/edit?usp=sharing)
+
+
+### Parameters selected for LeRobotDataset
+Considering these results, we chose what we think is the best set of encoding parameter:
+- vcodec: `libsvtav1`
+- pix-fmt: `yuv420p`
+- g: `2`
+- crf: `30`
+
+Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_reader` does not support it (and `pyav` doesn't require a custom build of `torchvision`).
+
+### Summary
+
+These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
+
+| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
+|------------------------------------|------------|---------|-----------|-----------|-----------|
+|                                    | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+
+| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
+|------------------------------------|---------|---------|----------|---------|-----------|
+|                                    | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+
+|                                    |          | vcodec   | pix_fmt      |          |           |              |
+|------------------------------------|----------|----------|--------------|----------|-----------|--------------|
+|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

benchmarks/video/capture_camera_feed.py

@@ -0,0 +1,102 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Capture video feed from a camera as raw images."""
+
+import argparse
+import datetime as dt
+import os
+import time
+from pathlib import Path
+
+import cv2
+import rerun as rr
+
+# see https://rerun.io/docs/howto/visualization/limit-ram
+RERUN_MEMORY_LIMIT = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")
+
+
+def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int, duration: int):
+    rr.init("lerobot_capture_camera_feed")
+    rr.spawn(memory_limit=RERUN_MEMORY_LIMIT)
+
+    now = dt.datetime.now()
+    capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
+    if not capture_dir.exists():
+        capture_dir.mkdir(parents=True, exist_ok=True)
+
+    # Opens the default webcam
+    cap = cv2.VideoCapture(0)
+    if not cap.isOpened():
+        print("Error: Could not open video stream.")
+        return
+
+    cap.set(cv2.CAP_PROP_FPS, fps)
+    cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
+    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
+
+    frame_index = 0
+    start_time = time.time()
+    while time.time() - start_time < duration:
+        ret, frame = cap.read()
+
+        if not ret:
+            print("Error: Could not read frame.")
+            break
+        rr.log("video/stream", rr.Image(frame.numpy()), static=True)
+        cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
+        frame_index += 1
+
+    # Release the capture
+    cap.release()
+
+    # TODO(Steven): Add a graceful shutdown via a close() method for the Viewer context, though not currently supported in the Rerun API.
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path("outputs/cam_capture/"),
+        help="Directory where the capture images are written. A subfolder named with the current date & time will be created inside it for each capture.",
+    )
+    parser.add_argument(
+        "--fps",
+        type=int,
+        default=30,
+        help="Frames Per Second of the capture.",
+    )
+    parser.add_argument(
+        "--width",
+        type=int,
+        default=1280,
+        help="Width of the captured images.",
+    )
+    parser.add_argument(
+        "--height",
+        type=int,
+        default=720,
+        help="Height of the captured images.",
+    )
+    parser.add_argument(
+        "--duration",
+        type=int,
+        default=20,
+        help="Duration in seconds for which the video stream should be captured.",
+    )
+    args = parser.parse_args()
+    display_and_save_video_stream(**vars(args))

benchmarks/video/run_video_benchmark.py

@@ -0,0 +1,490 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Assess the performance of video decoding in various configurations.
+
+This script will benchmark different video encoding and decoding parameters.
+See the provided README.md or run `python benchmark/video/run_video_benchmark.py --help` for usage info.
+"""
+
+import argparse
+import datetime as dt
+import random
+import shutil
+from collections import OrderedDict
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+
+import einops
+import numpy as np
+import pandas as pd
+import PIL
+import torch
+from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
+from tqdm import tqdm
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.video_utils import (
+    decode_video_frames_torchvision,
+    encode_video_frames,
+)
+from lerobot.common.utils.benchmark import TimeBenchmark
+
+BASE_ENCODING = OrderedDict(
+    [
+        ("vcodec", "libx264"),
+        ("pix_fmt", "yuv444p"),
+        ("g", 2),
+        ("crf", None),
+        # TODO(aliberts): Add fastdecode
+        # ("fastdecode", 0),
+    ]
+)
+
+
+# TODO(rcadene, aliberts): move to `utils.py` folder when we want to refactor
+def parse_int_or_none(value) -> int | None:
+    if value.lower() == "none":
+        return None
+    try:
+        return int(value)
+    except ValueError as e:
+        raise argparse.ArgumentTypeError(f"Invalid int or None: {value}") from e
+
+
+def check_datasets_formats(repo_ids: list) -> None:
+    for repo_id in repo_ids:
+        dataset = LeRobotDataset(repo_id)
+        if len(dataset.meta.video_keys) > 0:
+            raise ValueError(
+                f"Use only image dataset for running this benchmark. Video dataset provided: {repo_id}"
+            )
+
+
+def get_directory_size(directory: Path) -> int:
+    total_size = 0
+    for item in directory.rglob("*"):
+        if item.is_file():
+            total_size += item.stat().st_size
+    return total_size
+
+
+def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> torch.Tensor:
+    frames = []
+    for ts in timestamps:
+        idx = int(ts * fps)
+        frame = PIL.Image.open(imgs_dir / f"frame_{idx:06d}.png")
+        frame = torch.from_numpy(np.array(frame))
+        frame = frame.type(torch.float32) / 255
+        frame = einops.rearrange(frame, "h w c -> c h w")
+        frames.append(frame)
+    return torch.stack(frames)
+
+
+def save_decoded_frames(
+    imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
+) -> None:
+    if save_dir.exists() and len(list(save_dir.glob("frame_*.png"))) == len(timestamps):
+        return
+
+    save_dir.mkdir(parents=True, exist_ok=True)
+    for i, ts in enumerate(timestamps):
+        idx = int(ts * fps)
+        frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
+        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame_{idx:06d}_decoded.png")
+        shutil.copyfile(imgs_dir / f"frame_{idx:06d}.png", save_dir / f"frame_{idx:06d}_original.png")
+
+
+def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
+    ep_num_images = dataset.episode_data_index["to"][0].item()
+    if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
+        return
+
+    imgs_dir.mkdir(parents=True, exist_ok=True)
+    hf_dataset = dataset.hf_dataset.with_format(None)
+
+    # We only save images from the first camera
+    img_keys = [key for key in hf_dataset.features if key.startswith("observation.image")]
+    imgs_dataset = hf_dataset.select_columns(img_keys[0])
+
+    for i, item in enumerate(
+        tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
+    ):
+        img = item[img_keys[0]]
+        img.save(str(imgs_dir / f"frame_{i:06d}.png"), quality=100)
+
+        if i >= ep_num_images - 1:
+            break
+
+
+def sample_timestamps(timestamps_mode: str, ep_num_images: int, fps: int) -> list[float]:
+    # Start at 5 to allow for 2_frames_4_space and 6_frames
+    idx = random.randint(5, ep_num_images - 1)
+    match timestamps_mode:
+        case "1_frame":
+            frame_indexes = [idx]
+        case "2_frames":
+            frame_indexes = [idx - 1, idx]
+        case "2_frames_4_space":
+            frame_indexes = [idx - 5, idx]
+        case "6_frames":
+            frame_indexes = [idx - i for i in range(6)][::-1]
+        case _:
+            raise ValueError(timestamps_mode)
+
+    return [idx / fps for idx in frame_indexes]
+
+
+def decode_video_frames(
+    video_path: str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str,
+) -> torch.Tensor:
+    if backend in ["pyav", "video_reader"]:
+        return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
+    else:
+        raise NotImplementedError(backend)
+
+
+def benchmark_decoding(
+    imgs_dir: Path,
+    video_path: Path,
+    timestamps_mode: str,
+    backend: str,
+    ep_num_images: int,
+    fps: int,
+    num_samples: int = 50,
+    num_workers: int = 4,
+    save_frames: bool = False,
+) -> dict:
+    def process_sample(sample: int):
+        time_benchmark = TimeBenchmark()
+        timestamps = sample_timestamps(timestamps_mode, ep_num_images, fps)
+        num_frames = len(timestamps)
+        result = {
+            "psnr_values": [],
+            "ssim_values": [],
+            "mse_values": [],
+        }
+
+        with time_benchmark:
+            frames = decode_video_frames(video_path, timestamps=timestamps, tolerance_s=5e-1, backend=backend)
+        result["load_time_video_ms"] = time_benchmark.result_ms / num_frames
+
+        with time_benchmark:
+            original_frames = load_original_frames(imgs_dir, timestamps, fps)
+        result["load_time_images_ms"] = time_benchmark.result_ms / num_frames
+
+        frames_np, original_frames_np = frames.numpy(), original_frames.numpy()
+        for i in range(num_frames):
+            result["mse_values"].append(mean_squared_error(original_frames_np[i], frames_np[i]))
+            result["psnr_values"].append(
+                peak_signal_noise_ratio(original_frames_np[i], frames_np[i], data_range=1.0)
+            )
+            result["ssim_values"].append(
+                structural_similarity(original_frames_np[i], frames_np[i], data_range=1.0, channel_axis=0)
+            )
+
+        if save_frames and sample == 0:
+            save_dir = video_path.with_suffix("") / f"{timestamps_mode}_{backend}"
+            save_decoded_frames(imgs_dir, save_dir, frames, timestamps, fps)
+
+        return result
+
+    load_times_video_ms = []
+    load_times_images_ms = []
+    mse_values = []
+    psnr_values = []
+    ssim_values = []
+
+    # A sample is a single set of decoded frames specified by timestamps_mode (e.g. a single frame, 2 frames, etc.).
+    # For each sample, we record metrics (loading time and quality metrics) which are then averaged over all samples.
+    # As these samples are independent, we run them in parallel threads to speed up the benchmark.
+    with ThreadPoolExecutor(max_workers=num_workers) as executor:
+        futures = [executor.submit(process_sample, i) for i in range(num_samples)]
+        for future in tqdm(as_completed(futures), total=num_samples, desc="samples", leave=False):
+            result = future.result()
+            load_times_video_ms.append(result["load_time_video_ms"])
+            load_times_images_ms.append(result["load_time_images_ms"])
+            psnr_values.extend(result["psnr_values"])
+            ssim_values.extend(result["ssim_values"])
+            mse_values.extend(result["mse_values"])
+
+    avg_load_time_video_ms = float(np.array(load_times_video_ms).mean())
+    avg_load_time_images_ms = float(np.array(load_times_images_ms).mean())
+    video_images_load_time_ratio = avg_load_time_video_ms / avg_load_time_images_ms
+
+    return {
+        "avg_load_time_video_ms": avg_load_time_video_ms,
+        "avg_load_time_images_ms": avg_load_time_images_ms,
+        "video_images_load_time_ratio": video_images_load_time_ratio,
+        "avg_mse": float(np.mean(mse_values)),
+        "avg_psnr": float(np.mean(psnr_values)),
+        "avg_ssim": float(np.mean(ssim_values)),
+    }
+
+
+def benchmark_encoding_decoding(
+    dataset: LeRobotDataset,
+    video_path: Path,
+    imgs_dir: Path,
+    encoding_cfg: dict,
+    decoding_cfg: dict,
+    num_samples: int,
+    num_workers: int,
+    save_frames: bool,
+    overwrite: bool = False,
+    seed: int = 1337,
+) -> list[dict]:
+    fps = dataset.fps
+
+    if overwrite or not video_path.is_file():
+        tqdm.write(f"encoding {video_path}")
+        encode_video_frames(
+            imgs_dir=imgs_dir,
+            video_path=video_path,
+            fps=fps,
+            vcodec=encoding_cfg["vcodec"],
+            pix_fmt=encoding_cfg["pix_fmt"],
+            g=encoding_cfg.get("g"),
+            crf=encoding_cfg.get("crf"),
+            # fast_decode=encoding_cfg.get("fastdecode"),
+            overwrite=True,
+        )
+
+    ep_num_images = dataset.episode_data_index["to"][0].item()
+    width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
+    num_pixels = width * height
+    video_size_bytes = video_path.stat().st_size
+    images_size_bytes = get_directory_size(imgs_dir)
+    video_images_size_ratio = video_size_bytes / images_size_bytes
+
+    random.seed(seed)
+    benchmark_table = []
+    for timestamps_mode in tqdm(
+        decoding_cfg["timestamps_modes"], desc="decodings (timestamps_modes)", leave=False
+    ):
+        for backend in tqdm(decoding_cfg["backends"], desc="decodings (backends)", leave=False):
+            benchmark_row = benchmark_decoding(
+                imgs_dir,
+                video_path,
+                timestamps_mode,
+                backend,
+                ep_num_images,
+                fps,
+                num_samples,
+                num_workers,
+                save_frames,
+            )
+            benchmark_row.update(
+                **{
+                    "repo_id": dataset.repo_id,
+                    "resolution": f"{width} x {height}",
+                    "num_pixels": num_pixels,
+                    "video_size_bytes": video_size_bytes,
+                    "images_size_bytes": images_size_bytes,
+                    "video_images_size_ratio": video_images_size_ratio,
+                    "timestamps_mode": timestamps_mode,
+                    "backend": backend,
+                },
+                **encoding_cfg,
+            )
+            benchmark_table.append(benchmark_row)
+
+    return benchmark_table
+
+
+def main(
+    output_dir: Path,
+    repo_ids: list[str],
+    vcodec: list[str],
+    pix_fmt: list[str],
+    g: list[int],
+    crf: list[int],
+    # fastdecode: list[int],
+    timestamps_modes: list[str],
+    backends: list[str],
+    num_samples: int,
+    num_workers: int,
+    save_frames: bool,
+):
+    check_datasets_formats(repo_ids)
+    encoding_benchmarks = {
+        "g": g,
+        "crf": crf,
+        # "fastdecode": fastdecode,
+    }
+    decoding_benchmarks = {
+        "timestamps_modes": timestamps_modes,
+        "backends": backends,
+    }
+    headers = ["repo_id", "resolution", "num_pixels"]
+    headers += list(BASE_ENCODING.keys())
+    headers += [
+        "timestamps_mode",
+        "backend",
+        "video_size_bytes",
+        "images_size_bytes",
+        "video_images_size_ratio",
+        "avg_load_time_video_ms",
+        "avg_load_time_images_ms",
+        "video_images_load_time_ratio",
+        "avg_mse",
+        "avg_psnr",
+        "avg_ssim",
+    ]
+    file_paths = []
+    for video_codec in tqdm(vcodec, desc="encodings (vcodec)"):
+        for pixel_format in tqdm(pix_fmt, desc="encodings (pix_fmt)", leave=False):
+            benchmark_table = []
+            for repo_id in tqdm(repo_ids, desc="encodings (datasets)", leave=False):
+                dataset = LeRobotDataset(repo_id)
+                imgs_dir = output_dir / "images" / dataset.repo_id.replace("/", "_")
+                # We only use the first episode
+                save_first_episode(imgs_dir, dataset)
+                for key, values in tqdm(encoding_benchmarks.items(), desc="encodings (g, crf)", leave=False):
+                    for value in tqdm(values, desc=f"encodings ({key})", leave=False):
+                        encoding_cfg = BASE_ENCODING.copy()
+                        encoding_cfg["vcodec"] = video_codec
+                        encoding_cfg["pix_fmt"] = pixel_format
+                        encoding_cfg[key] = value
+                        args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
+                        video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
+                        benchmark_table += benchmark_encoding_decoding(
+                            dataset,
+                            video_path,
+                            imgs_dir,
+                            encoding_cfg,
+                            decoding_benchmarks,
+                            num_samples,
+                            num_workers,
+                            save_frames,
+                        )
+
+            # Save intermediate results
+            benchmark_df = pd.DataFrame(benchmark_table, columns=headers)
+            now = dt.datetime.now()
+            csv_path = (
+                output_dir
+                / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_{video_codec}_{pixel_format}_{num_samples}-samples.csv"
+            )
+            benchmark_df.to_csv(csv_path, header=True, index=False)
+            file_paths.append(csv_path)
+            del benchmark_df
+
+    # Concatenate all results
+    df_list = [pd.read_csv(csv_path) for csv_path in file_paths]
+    concatenated_df = pd.concat(df_list, ignore_index=True)
+    concatenated_path = output_dir / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_all_{num_samples}-samples.csv"
+    concatenated_df.to_csv(concatenated_path, header=True, index=False)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path("outputs/video_benchmark"),
+        help="Directory where the video benchmark outputs are written.",
+    )
+    parser.add_argument(
+        "--repo-ids",
+        type=str,
+        nargs="*",
+        default=[
+            "lerobot/pusht_image",
+            "aliberts/aloha_mobile_shrimp_image",
+            "aliberts/paris_street",
+            "aliberts/kitchen",
+        ],
+        help="Datasets repo-ids to test against. First episodes only are used. Must be images.",
+    )
+    parser.add_argument(
+        "--vcodec",
+        type=str,
+        nargs="*",
+        default=["libx264", "libx265", "libsvtav1"],
+        help="Video codecs to be tested",
+    )
+    parser.add_argument(
+        "--pix-fmt",
+        type=str,
+        nargs="*",
+        default=["yuv444p", "yuv420p"],
+        help="Pixel formats (chroma subsampling) to be tested",
+    )
+    parser.add_argument(
+        "--g",
+        type=parse_int_or_none,
+        nargs="*",
+        default=[1, 2, 3, 4, 5, 6, 10, 15, 20, 40, 100, None],
+        help="Group of pictures sizes to be tested.",
+    )
+    parser.add_argument(
+        "--crf",
+        type=parse_int_or_none,
+        nargs="*",
+        default=[0, 5, 10, 15, 20, 25, 30, 40, 50, None],
+        help="Constant rate factors to be tested.",
+    )
+    # parser.add_argument(
+    #     "--fastdecode",
+    #     type=int,
+    #     nargs="*",
+    #     default=[0, 1],
+    #     help="Use the fastdecode tuning option. 0 disables it. "
+    #         "For libx264 and libx265, only 1 is possible. "
+    #         "For libsvtav1, 1, 2 or 3 are possible values with a higher number meaning a faster decoding optimization",
+    # )
+    parser.add_argument(
+        "--timestamps-modes",
+        type=str,
+        nargs="*",
+        default=[
+            "1_frame",
+            "2_frames",
+            "2_frames_4_space",
+            "6_frames",
+        ],
+        help="Timestamps scenarios to be tested.",
+    )
+    parser.add_argument(
+        "--backends",
+        type=str,
+        nargs="*",
+        default=["pyav", "video_reader"],
+        help="Torchvision decoding backend to be tested.",
+    )
+    parser.add_argument(
+        "--num-samples",
+        type=int,
+        default=50,
+        help="Number of samples for each encoding x decoding config.",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=10,
+        help="Number of processes for parallelized sample processing.",
+    )
+    parser.add_argument(
+        "--save-frames",
+        type=int,
+        default=0,
+        help="Whether to save decoded frames or not. Enter a non-zero number for true.",
+    )
+    args = parser.parse_args()
+    main(**vars(args))

docker/lerobot-cpu/Dockerfile

@@ -0,0 +1,29 @@
+# Configure image
+ARG PYTHON_VERSION=3.10
+FROM python:${PYTHON_VERSION}-slim
+
+# Configure environment variables
+ARG PYTHON_VERSION
+ENV DEBIAN_FRONTEND=noninteractive
+ENV MUJOCO_GL="egl"
+ENV PATH="/opt/venv/bin:$PATH"
+
+# Install dependencies and set up Python in a single layer
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake git \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    speech-dispatcher libgeos-dev \
+    && ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python \
+    && python -m venv /opt/venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/* \
+    && echo "source /opt/venv/bin/activate" >> /root/.bashrc
+
+# Clone repository and install LeRobot in a single layer
+COPY . /lerobot
+WORKDIR /lerobot
+RUN /opt/venv/bin/pip install --upgrade --no-cache-dir pip \
+    && /opt/venv/bin/pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]" \
+        --extra-index-url https://download.pytorch.org/whl/cpu
+
+# Execute in bash shell rather than python
+CMD ["/bin/bash"]

docker/lerobot-gpu-dev/Dockerfile

@@ -0,0 +1,68 @@
+FROM nvidia/cuda:12.2.2-devel-ubuntu22.04
+
+# Configure image
+ARG PYTHON_VERSION=3.10
+ARG DEBIAN_FRONTEND=noninteractive
+
+# Install apt dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake \
+    git git-lfs openssh-client \
+    nano vim less util-linux tree \
+    htop atop nvtop \
+    sed gawk grep curl wget zip unzip \
+    tcpdump sysstat screen tmux \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
+    speech-dispatcher portaudio19-dev libgeos-dev \
+    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Install ffmpeg build dependencies. See:
+# https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu
+# TODO(aliberts): create image to build dependencies from source instead
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    autoconf automake yasm \
+    libass-dev \
+    libfreetype6-dev \
+    libgnutls28-dev \
+    libunistring-dev \
+    libmp3lame-dev \
+    libtool \
+    libvorbis-dev \
+    meson \
+    ninja-build \
+    pkg-config \
+    texinfo \
+    yasm \
+    zlib1g-dev \
+    nasm \
+    libx264-dev \
+    libx265-dev libnuma-dev \
+    libvpx-dev \
+    libfdk-aac-dev \
+    libopus-dev \
+    libsvtav1-dev libsvtav1enc-dev libsvtav1dec-dev \
+    libdav1d-dev
+
+# Install gh cli tool
+RUN (type -p wget >/dev/null || (apt update && apt-get install wget -y)) \
+    && mkdir -p -m 755 /etc/apt/keyrings \
+    && wget -qO- https://cli.github.com/packages/githubcli-archive-keyring.gpg | tee /etc/apt/keyrings/githubcli-archive-keyring.gpg > /dev/null \
+    && chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg \
+    && echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" | tee /etc/apt/sources.list.d/github-cli.list > /dev/null \
+    && apt update \
+    && apt install gh -y \
+    && apt clean && rm -rf /var/lib/apt/lists/*
+
+# Setup `python`
+RUN ln -s /usr/bin/python3 /usr/bin/python
+
+# Install poetry
+RUN curl -sSL https://install.python-poetry.org | python -
+ENV PATH="/root/.local/bin:$PATH"
+RUN echo 'if [ "$HOME" != "/root" ]; then ln -sf /root/.local/bin/poetry $HOME/.local/bin/poetry; fi' >> /root/.bashrc
+RUN poetry config virtualenvs.create false
+RUN poetry config virtualenvs.in-project true
+
+# Set EGL as the rendering backend for MuJoCo
+ENV MUJOCO_GL="egl"

docker/lerobot-gpu/Dockerfile

@@ -0,0 +1,24 @@
+FROM nvidia/cuda:12.4.1-base-ubuntu22.04
+
+# Configure environment variables
+ARG PYTHON_VERSION=3.10
+ENV DEBIAN_FRONTEND=noninteractive
+ENV MUJOCO_GL="egl"
+ENV PATH="/opt/venv/bin:$PATH"
+
+# Install dependencies and set up Python in a single layer
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake git \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    speech-dispatcher libgeos-dev \
+    python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+    && ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python \
+    && python -m venv /opt/venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/* \
+    && echo "source /opt/venv/bin/activate" >> /root/.bashrc
+
+# Clone repository and install LeRobot in a single layer
+COPY . /lerobot
+WORKDIR /lerobot
+RUN /opt/venv/bin/pip install --upgrade --no-cache-dir pip \
+    && /opt/venv/bin/pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]"

examples/10_use_so100.md

@@ -0,0 +1,624 @@
+# Using the [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot
+
+## Table of Contents
+
+  - [A. Source the parts](#a-source-the-parts)
+  - [B. Install LeRobot](#b-install-lerobot)
+  - [C. Configure the Motors](#c-configure-the-motors)
+  - [D. Step-by-Step Assembly Instructions](#d-step-by-step-assembly-instructions)
+  - [E. Calibrate](#e-calibrate)
+  - [F. Teleoperate](#f-teleoperate)
+  - [G. Record a dataset](#g-record-a-dataset)
+  - [H. Visualize a dataset](#h-visualize-a-dataset)
+  - [I. Replay an episode](#i-replay-an-episode)
+  - [J. Train a policy](#j-train-a-policy)
+  - [K. Evaluate your policy](#k-evaluate-your-policy)
+  - [L. More Information](#l-more-information)
+
+## A. Source the parts
+
+Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts,
+and advice if it's your first time printing or if you don't own a 3D printer.
+
+Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
+
+## B. Install LeRobot
+
+> [!TIP]
+> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
+
+On your computer:
+
+#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
+
+#### 2. Restart shell
+Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
+
+#### 3. Create and activate a fresh conda environment for lerobot
+
+<details>
+<summary><strong>Video install instructions</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
+
+</details>
+
+```bash
+conda create -y -n lerobot python=3.10
+```
+
+Then activate your conda environment (do this each time you open a shell to use lerobot!):
+```bash
+conda activate lerobot
+```
+
+#### 4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+#### 5. Install ffmpeg in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+#### 6. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms :robot:.
+Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
+
+## C. Configure the motors
+
+> [!NOTE]
+> Throughout this tutorial you will find videos on how to do the steps, the full video tutorial can be found here: [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I).
+
+### 1. Find the USB ports associated to each arm
+
+Designate one bus servo adapter and 6 motors for your leader arm, and similarly the other bus servo adapter and 6 motors for the follower arm. It's convenient to label them and write on each motor if it's for the follower `F` or for the leader `L` and it's ID from 1 to 6 (F1...F6 and L1...L6).
+
+#### a. Run the script to find port
+
+<details>
+<summary><strong>Video finding port</strong></summary>
+  <video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
+  <video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
+</details>
+
+To find the port for each bus servo adapter, run the utility script:
+```bash
+python lerobot/scripts/find_motors_bus_port.py
+```
+
+#### b. Example outputs
+
+Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your MotorsBus and press Enter when done.
+
+[...Disconnect leader arm and press Enter...]
+
+The port of this MotorsBus is /dev/tty.usbmodem575E0031751
+Reconnect the usb cable.
+```
+Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your MotorsBus and press Enter when done.
+
+[...Disconnect follower arm and press Enter...]
+
+The port of this MotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the usb cable.
+```
+
+#### c. Troubleshooting
+On Linux, you might need to give access to the USB ports by running:
+```bash
+sudo chmod 666 /dev/ttyACM0
+sudo chmod 666 /dev/ttyACM1
+```
+
+#### d. Update config file
+
+IMPORTANTLY: Now that you have your ports, update the **port** default values of [`SO100RobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
+```python
+@RobotConfig.register_subclass("so100")
+@dataclass
+class So100RobotConfig(ManipulatorRobotConfig):
+    calibration_dir: str = ".cache/calibration/so100"
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
+
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+```
+
+### 2. Assembling the Base
+Let's begin with assembling the follower arm base
+
+#### a. Set IDs for all 12 motors
+
+<details>
+<summary><strong>Video configuring motor</strong></summary>
+  <video src="https://github.com/user-attachments/assets/ef9b3317-2e11-4858-b9d3-f0a02fb48ecf"></video>
+  <video src="https://github.com/user-attachments/assets/f36b5ed5-c803-4ebe-8947-b39278776a0d"></video>
+</details>
+
+Plug your first motor F1 and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate. Replace the text after --port to the corresponding follower control board port and run this command in cmd:
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 1
+```
+
+> [!NOTE]
+> These motors are currently limited. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
+
+Then unplug your motor and plug the second motor and set its ID to 2.
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 2
+```
+
+Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
+
+
+#### b. Remove the gears of the 6 leader motors
+
+<details>
+<summary><strong>Video removing gears</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/0c95b88c-5b85-413d-ba19-aee2f864f2a7"></video>
+
+</details>
+
+
+Follow the video for removing gears. You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
+
+## D. Step-by-Step Assembly Instructions
+
+**Step 1: Clean Parts**
+- Remove all support material from the 3D-printed parts.
+---
+
+### Additional Guidance
+
+<details>
+<summary><strong>Video assembling arms</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/488a39de-0189-4461-9de3-05b015f90cca"></video>
+
+</details>
+
+**Note:**
+This video provides visual guidance for assembling the arms, but it doesn't specify when or how to do the wiring. Inserting the cables beforehand is much easier than doing it afterward. The first arm may take a bit more than 1 hour to assemble, but once you get used to it, you can assemble the second arm in under 1 hour.
+
+---
+
+### First Motor
+
+**Step 2: Insert Wires**
+- Insert two wires into the first motor.
+
+  <img src="../media/tutorial/img1.jpg" style="height:300px;">
+
+**Step 3: Install in Base**
+- Place the first motor into the base.
+
+  <img src="../media/tutorial/img2.jpg" style="height:300px;">
+
+**Step 4: Secure Motor**
+- Fasten the motor with 4 screws. Two from the bottom and two from top.
+
+**Step 5: Attach Motor Holder**
+- Slide over the first motor holder and fasten it using two screws (one on each side).
+
+  <img src="../media/tutorial/img4.jpg" style="height:300px;">
+
+**Step 6: Attach Motor Horns**
+- Install both motor horns, securing the top horn with a screw. Try not to move the motor position when attaching the motor horn, especially for the leader arms, where we removed the gears.
+
+  <img src="../media/tutorial/img5.jpg" style="height:300px;">
+<details>
+  <summary><strong>Video adding motor horn</strong></summary>
+  <video src="https://github.com/user-attachments/assets/ef3391a4-ad05-4100-b2bd-1699bf86c969"></video>
+</details>
+
+**Step 7: Attach Shoulder Part**
+- Route one wire to the back of the robot and the other to the left or in photo towards you (see photo).
+- Attach the shoulder part.
+
+  <img src="../media/tutorial/img6.jpg" style="height:300px;">
+
+**Step 8: Secure Shoulder**
+- Tighten the shoulder part with 4 screws on top and 4 on the bottom
+*(access bottom holes by turning the shoulder).*
+
+---
+
+### Second Motor Assembly
+
+**Step 9: Install Motor 2**
+- Slide the second motor in from the top and link the wire from motor 1 to motor 2.
+
+  <img src="../media/tutorial/img8.jpg" style="height:300px;">
+
+**Step 10: Attach Shoulder Holder**
+- Add the shoulder motor holder.
+- Ensure the wire from motor 1 to motor 2 goes behind the holder while the other wire is routed upward (see photo).
+- This part can be tight to assemble, you can use a workbench like the image or a similar setup to push the part around the motor.
+
+  <div style="display: flex;">
+    <img src="../media/tutorial/img9.jpg" style="height:250px;">
+    <img src="../media/tutorial/img10.jpg" style="height:250px;">
+    <img src="../media/tutorial/img12.jpg" style="height:250px;">
+  </div>
+
+**Step 11: Secure Motor 2**
+- Fasten the second motor with 4 screws.
+
+**Step 12: Attach Motor Horn**
+- Attach both motor horns to motor 2, again use the horn screw.
+
+**Step 13: Attach Base**
+- Install the base attachment using 2 screws.
+
+  <img src="../media/tutorial/img11.jpg" style="height:300px;">
+
+**Step 14: Attach Upper Arm**
+- Attach the upper arm with 4 screws on each side.
+
+  <img src="../media/tutorial/img13.jpg" style="height:300px;">
+
+---
+
+### Third Motor Assembly
+
+**Step 15: Install Motor 3**
+- Route the motor cable from motor 2 through the cable holder to motor 3, then secure motor 3 with 4 screws.
+
+**Step 16: Attach Motor Horn**
+- Attach both motor horns to motor 3 and secure one again with a horn screw.
+
+  <img src="../media/tutorial/img14.jpg" style="height:300px;">
+
+**Step 17: Attach Forearm**
+- Connect the forearm to motor 3 using 4 screws on each side.
+
+  <img src="../media/tutorial/img15.jpg" style="height:300px;">
+
+---
+
+### Fourth Motor Assembly
+
+**Step 18: Install Motor 4**
+- Slide in motor 4, attach the cable from motor 3, and secure the cable in its holder with a screw.
+
+  <div style="display: flex;">
+    <img src="../media/tutorial/img16.jpg" style="height:300px;">
+    <img src="../media/tutorial/img19.jpg" style="height:300px;">
+  </div>
+
+**Step 19: Attach Motor Holder 4**
+- Install the fourth motor holder (a tight fit). Ensure one wire is routed upward and the wire from motor 3 is routed downward (see photo).
+
+  <img src="../media/tutorial/img17.jpg" style="height:300px;">
+
+**Step 20: Secure Motor 4 & Attach Horn**
+- Fasten motor 4 with 4 screws and attach its motor horns, use for one a horn screw.
+
+  <img src="../media/tutorial/img18.jpg" style="height:300px;">
+
+---
+
+### Wrist Assembly
+
+**Step 21: Install Motor 5**
+- Insert motor 5 into the wrist holder and secure it with 2 front screws.
+
+  <img src="../media/tutorial/img20.jpg" style="height:300px;">
+
+**Step 22: Attach Wrist**
+- Connect the wire from motor 4 to motor 5. And already insert the other wire for the gripper.
+- Secure the wrist to motor 4 using 4 screws on both sides.
+
+  <img src="../media/tutorial/img22.jpg" style="height:300px;">
+
+**Step 23: Attach Wrist Horn**
+- Install only one motor horn on the wrist motor and secure it with a horn screw.
+
+  <img src="../media/tutorial/img23.jpg" style="height:300px;">
+
+---
+
+### Follower Configuration
+
+**Step 24: Attach Gripper**
+- Attach the gripper to motor 5.
+
+  <img src="../media/tutorial/img24.jpg" style="height:300px;">
+
+**Step 25: Install Gripper Motor**
+- Insert the gripper motor, connect the motor wire from motor 5 to motor 6, and secure it with 3 screws on each side.
+
+  <img src="../media/tutorial/img25.jpg" style="height:300px;">
+
+**Step 26: Attach Gripper Horn & Claw**
+- Attach the motor horns and again use a horn screw.
+- Install the gripper claw and secure it with 4 screws on both sides.
+
+  <img src="../media/tutorial/img26.jpg" style="height:300px;">
+
+**Step 27: Mount Controller**
+- Attach the motor controller on the back.
+
+  <div style="display: flex;">
+    <img src="../media/tutorial/img27.jpg" style="height:300px;">
+    <img src="../media/tutorial/img28.jpg" style="height:300px;">
+  </div>
+
+*Assembly complete – proceed to Leader arm assembly.*
+
+---
+
+### Leader Configuration
+
+For the leader configuration, perform **Steps 1–23**. Make sure that you removed the motor gears from the motors.
+
+**Step 24: Attach Leader Holder**
+- Mount the leader holder onto the wrist and secure it with a screw.
+
+  <img src="../media/tutorial/img29.jpg" style="height:300px;">
+
+**Step 25: Attach Handle**
+- Attach the handle to motor 5 using 4 screws.
+
+  <img src="../media/tutorial/img30.jpg" style="height:300px;">
+
+**Step 26: Install Gripper Motor**
+- Insert the gripper motor, secure it with 3 screws on each side, attach a motor horn using a horn screw, and connect the motor wire.
+
+  <img src="../media/tutorial/img31.jpg" style="height:300px;">
+
+**Step 27: Attach Trigger**
+- Attach the follower trigger with 4 screws.
+
+  <img src="../media/tutorial/img32.jpg" style="height:300px;">
+
+**Step 28: Mount Controller**
+- Attach the motor controller on the back.
+
+  <div style="display: flex;">
+    <img src="../media/tutorial/img27.jpg" style="height:300px;">
+    <img src="../media/tutorial/img28.jpg" style="height:300px;">
+  </div>
+
+*Assembly complete – proceed to calibration.*
+
+
+## E. Calibrate
+
+Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one SO-100 robot to work on another.
+
+#### a. Manual calibration of follower arm
+
+> [!IMPORTANT]
+> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
+
+You will need to move the follower arm to these positions sequentially:
+
+| 1. Zero position                                                                                                                                             | 2. Rotated position                                                                                                                                                   | 3. Rest position                                                                                                                                             |
+| ------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| <img src="../media/so100/follower_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/so100/follower_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/so100/follower_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
+
+Make sure both arms are connected and run this script to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_follower"]'
+```
+
+#### b. Manual calibration of leader arm
+Follow step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
+
+| 1. Zero position                                                                                                                                       | 2. Rotated position                                                                                                                                             | 3. Rest position                                                                                                                                       |
+| ------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
+
+Run this script to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_leader"]'
+```
+
+## F. Teleoperate
+
+**Simple teleop**
+Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --robot.cameras='{}' \
+  --control.type=teleoperate
+```
+
+
+#### a. Teleop with displaying cameras
+Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
+
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --control.type=teleoperate
+```
+
+## G. Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset with SO-100.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/so100_test \
+  --control.tags='["so100","tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
+```
+
+Note: You can resume recording by adding `--control.resume=true`.
+
+## H. Visualize a dataset
+
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+```bash
+echo ${HF_USER}/so100_test
+```
+
+If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with (a window can be opened in the browser `http://127.0.0.1:9090` with the visualization tool):
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/so100_test \
+  --local-files-only 1
+```
+
+## I. Replay an episode
+
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/so100_test \
+  --control.episode=0
+```
+
+## J. Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+python lerobot/scripts/train.py \
+  --dataset.repo_id=${HF_USER}/so100_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_so100_test \
+  --job_name=act_so100_test \
+  --policy.device=cuda \
+  --wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so100_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+
+Training should take several hours. You will find checkpoints in `outputs/train/act_so100_test/checkpoints`.
+
+To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so100_test` policy:
+```bash
+python lerobot/scripts/train.py \
+  --config_path=outputs/train/act_so100_test/checkpoints/last/pretrained_model/train_config.json \
+  --resume=true
+```
+
+## K. Evaluate your policy
+
+You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/eval_act_so100_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=10 \
+  --control.push_to_hub=true \
+  --control.policy.path=outputs/train/act_so100_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so100_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so100_test`).
+
+## L. More Information
+
+Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
+
+> [!TIP]
+>  If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb) in the channel [`#so100-arm`](https://discord.com/channels/1216765309076115607/1237741463832363039).

examples/11_use_lekiwi.md

@@ -0,0 +1,597 @@
+# Using the [LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi) Robot with LeRobot
+
+## Table of Contents
+
+  - [A. Source the parts](#a-source-the-parts)
+  - [B. Install software Pi](#b-install-software-on-pi)
+  - [C. Setup LeRobot laptop/pc](#c-install-lerobot-on-laptop)
+  - [D. Assemble the arms](#d-assembly)
+  - [E. Calibrate](#e-calibration)
+  - [F. Teleoperate](#f-teleoperate)
+  - [G. Record a dataset](#g-record-a-dataset)
+  - [H. Visualize a dataset](#h-visualize-a-dataset)
+  - [I. Replay an episode](#i-replay-an-episode)
+  - [J. Train a policy](#j-train-a-policy)
+  - [K. Evaluate your policy](#k-evaluate-your-policy)
+
+> [!TIP]
+>  If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb) in the channel [`#mobile-so-100-arm`](https://discord.com/channels/1216765309076115607/1318390825528332371).
+
+## A. Source the parts
+
+Follow this [README](https://github.com/SIGRobotics-UIUC/LeKiwi). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts, and advice if it's your first time printing or if you don't own a 3D printer.
+
+Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
+
+### Wired version
+If you have the **wired** LeKiwi version you can skip the installation of the Raspberry Pi and setting up SSH. You can also run all commands directly on your PC for both the LeKiwi scripts and the leader arm scripts for teleoperating.
+
+## B. Install software on Pi
+Now we have to setup the remote PC that will run on the LeKiwi Robot. This is normally a Raspberry Pi, but can be any PC that can run on 5V and has enough usb ports (2 or more) for the cameras and motor control board.
+
+### Install OS
+For setting up the Raspberry Pi and its SD-card see: [Setup PI](https://www.raspberrypi.com/documentation/computers/getting-started.html). Here is explained how to download the [Imager](https://www.raspberrypi.com/software/) to install Raspberry Pi OS or Ubuntu.
+
+### Setup SSH
+After setting up your Pi, you should enable and setup [SSH](https://www.raspberrypi.com/news/coding-on-raspberry-pi-remotely-with-visual-studio-code/) (Secure Shell Protocol) so you can login into the Pi from your laptop without requiring a screen, keyboard and mouse in the Pi. A great tutorial on how to do this can be found [here](https://www.raspberrypi.com/documentation/computers/remote-access.html#ssh). Logging into your Pi can be done in your Command Prompt (cmd) or if you use VSCode you can use [this](https://marketplace.visualstudio.com/items?itemName=ms-vscode-remote.remote-ssh) extension.
+
+### Install LeRobot
+
+On your Raspberry Pi:
+
+#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
+
+#### 2. Restart shell
+Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
+
+#### 3. Create and activate a fresh conda environment for lerobot
+
+<details>
+<summary><strong>Video install instructions</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
+
+</details>
+
+```bash
+conda create -y -n lerobot python=3.10
+```
+
+Then activate your conda environment (do this each time you open a shell to use lerobot!):
+```bash
+conda activate lerobot
+```
+
+#### 4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+#### 5. Install ffmpeg in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+#### 6. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+## C. Install LeRobot on laptop
+If you already have install LeRobot on your laptop you can skip this step, otherwise please follow along as we do the same steps we did on the Pi.
+
+> [!TIP]
+> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
+
+On your computer:
+
+#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
+
+#### 2. Restart shell
+Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
+
+#### 3. Create and activate a fresh conda environment for lerobot
+
+<details>
+<summary><strong>Video install instructions</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
+
+</details>
+
+```bash
+conda create -y -n lerobot python=3.10
+```
+
+Then activate your conda environment (do this each time you open a shell to use lerobot!):
+```bash
+conda activate lerobot
+```
+
+#### 4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+#### 5. Install ffmpeg in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+#### 6. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms and Mobile base :robot:.
+Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
+
+# D. Assembly
+
+First we will assemble the two SO100 arms. One to attach to the mobile base and one for teleoperation. Then we will assemble the mobile base.
+
+## SO100 Arms
+### Configure motors
+The instructions for configuring the motors can be found [Here](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#c-configure-the-motors) in step C of the SO100 tutorial. Besides the ID's for the arm motors we also need to set the motor ID's for the mobile base. These needs to be in a specific order to work. Below an image of the motor ID's and motor mounting positions for the mobile base. Note that we only use one Motor Control board on LeKiwi. This means the motor ID's for the wheels are 7, 8 and 9.
+
+<img src="../media/lekiwi/motor_ids.webp?raw=true" alt="Motor ID's for mobile robot" title="Motor ID's for mobile robot" width="60%">
+
+### Assemble arms
+[Assemble arms instruction](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#d-assemble-the-arms)
+
+## Mobile base (LeKiwi)
+[Assemble LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi)
+
+### Update config
+Both config files on the LeKiwi LeRobot and on the laptop should be the same. First we should find the Ip address of the Raspberry Pi of the mobile manipulator. This is the same Ip address used in SSH. We also need the usb port of the control board of the leader arm on the laptop and the port of the control board on LeKiwi. We can find these ports with the following script.
+
+#### a. Run the script to find port
+
+<details>
+<summary><strong>Video finding port</strong></summary>
+  <video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
+  <video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
+</details>
+
+To find the port for each bus servo adapter, run the utility script:
+```bash
+python lerobot/scripts/find_motors_bus_port.py
+```
+
+#### b. Example outputs
+
+Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
+
+[...Disconnect leader arm and press Enter...]
+
+The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
+Reconnect the usb cable.
+```
+Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
+
+[...Disconnect follower arm and press Enter...]
+
+The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the usb cable.
+```
+
+#### c. Troubleshooting
+On Linux, you might need to give access to the USB ports by running:
+```bash
+sudo chmod 666 /dev/ttyACM0
+sudo chmod 666 /dev/ttyACM1
+```
+
+#### d. Update config file
+
+IMPORTANTLY: Now that you have your ports of leader and follower arm and ip address of the mobile-so100, update the **ip** in Network configuration, **port** in leader_arms and **port** in lekiwi. In the [`LeKiwiRobotConfig`](../lerobot/common/robot_devices/robots/configs.py) file. Where you will find something like:
+```python
+@RobotConfig.register_subclass("lekiwi")
+@dataclass
+class LeKiwiRobotConfig(RobotConfig):
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
+
+    # Network Configuration
+    ip: str = "172.17.133.91"
+    port: int = 5555
+    video_port: int = 5556
+
+    cameras: dict[str, CameraConfig] = field(
+        default_factory=lambda: {
+            "mobile": OpenCVCameraConfig(camera_index="/dev/video0", fps=30, width=640, height=480),
+            "mobile2": OpenCVCameraConfig(camera_index="/dev/video2", fps=30, width=640, height=480),
+        }
+    )
+
+    calibration_dir: str = ".cache/calibration/lekiwi"
+
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0077581",
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/ttyACM0",
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                    "left_wheel": (7, "sts3215"),
+                    "back_wheel": (8, "sts3215"),
+                    "right_wheel": (9, "sts3215"),
+                },
+            ),
+        }
+    )
+
+    teleop_keys: dict[str, str] = field(
+        default_factory=lambda: {
+            # Movement
+            "forward": "w",
+            "backward": "s",
+            "left": "a",
+            "right": "d",
+            "rotate_left": "z",
+            "rotate_right": "x",
+            # Speed control
+            "speed_up": "r",
+            "speed_down": "f",
+            # quit teleop
+            "quit": "q",
+        }
+    )
+
+    mock: bool = False
+```
+
+## Wired version
+
+For the wired LeKiwi version your configured IP address should refer to your own laptop (127.0.0.1), because leader arm and LeKiwi are in this case connected to own laptop. Below and example configuration for this wired setup:
+```python
+@RobotConfig.register_subclass("lekiwi")
+@dataclass
+class LeKiwiRobotConfig(RobotConfig):
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
+
+    # Network Configuration
+    ip: str = "127.0.0.1"
+    port: int = 5555
+    video_port: int = 5556
+
+    cameras: dict[str, CameraConfig] = field(
+        default_factory=lambda: {
+            "front": OpenCVCameraConfig(
+                camera_index=0, fps=30, width=640, height=480, rotation=90
+            ),
+            "wrist": OpenCVCameraConfig(
+                camera_index=1, fps=30, width=640, height=480, rotation=180
+            ),
+        }
+    )
+
+    calibration_dir: str = ".cache/calibration/lekiwi"
+
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0077581",
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem58760431061",
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                    "left_wheel": (7, "sts3215"),
+                    "back_wheel": (8, "sts3215"),
+                    "right_wheel": (9, "sts3215"),
+                },
+            ),
+        }
+    )
+
+    teleop_keys: dict[str, str] = field(
+        default_factory=lambda: {
+            # Movement
+            "forward": "w",
+            "backward": "s",
+            "left": "a",
+            "right": "d",
+            "rotate_left": "z",
+            "rotate_right": "x",
+            # Speed control
+            "speed_up": "r",
+            "speed_down": "f",
+            # quit teleop
+            "quit": "q",
+        }
+    )
+
+    mock: bool = False
+```
+
+# E. Calibration
+Now we have to calibrate the leader arm and the follower arm. The wheel motors don't have to be calibrated.
+
+
+### Calibrate follower arm (on mobile base)
+> [!IMPORTANT]
+> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
+
+You will need to move the follower arm to these positions sequentially:
+
+| 1. Zero position                                                                                                                                                  | 2. Rotated position                                                                                                                                                        | 3. Rest position                                                                                                                                                  |
+| ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <img src="../media/lekiwi/mobile_calib_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/lekiwi/mobile_calib_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/lekiwi/mobile_calib_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
+
+Make sure the arm is connected to the Raspberry Pi and run this script (on the Raspberry Pi) to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_follower"]'
+```
+
+### Wired version
+If you have the **wired** LeKiwi version please run all commands including this calibration command on your laptop.
+
+### Calibrate leader arm
+Then to calibrate the leader arm (which is attached to the laptop/pc). You will need to move the leader arm to these positions sequentially:
+
+| 1. Zero position                                                                                                                                       | 2. Rotated position                                                                                                                                             | 3. Rest position                                                                                                                                       |
+| ------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
+
+Run this script (on your laptop/pc) to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_leader"]'
+```
+
+# F. Teleoperate
+
+> [!TIP]
+> If you're using a Mac, you might need to give Terminal permission to access your keyboard. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
+
+To teleoperate SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=remote_robot
+```
+
+Then on your laptop, also run `conda activate lerobot` and this script:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=teleoperate \
+  --control.fps=30
+```
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`. For the `--control.type=remote_robot` you will also need to set `--control.viewer_ip` and `--control.viewer_port`
+
+You should see on your laptop something like this: ```[INFO] Connected to remote robot at tcp://172.17.133.91:5555 and video stream at tcp://172.17.133.91:5556.``` Now you can move the leader arm and use the keyboard (w,a,s,d) to drive forward, left, backwards, right. And use (z,x) to turn left or turn right. You can use (r,f) to increase and decrease the speed of the mobile robot. There are three speed modes, see the table below:
+| Speed Mode | Linear Speed (m/s) | Rotation Speed (deg/s) |
+| ---------- | ------------------ | ---------------------- |
+| Fast       | 0.4                | 90                     |
+| Medium     | 0.25               | 60                     |
+| Slow       | 0.1                | 30                     |
+
+
+| Key | Action         |
+| --- | -------------- |
+| W   | Move forward   |
+| A   | Move left      |
+| S   | Move backward  |
+| D   | Move right     |
+| Z   | Turn left      |
+| X   | Turn right     |
+| R   | Increase speed |
+| F   | Decrease speed |
+
+> [!TIP]
+>  If you use a different keyboard you can change the keys for each command in the [`LeKiwiRobotConfig`](../lerobot/common/robot_devices/robots/configs.py).
+
+### Wired version
+If you have the **wired** LeKiwi version please run all commands including both these teleoperation commands on your laptop.
+
+## Troubleshoot communication
+
+If you are having trouble connecting to the Mobile SO100, follow these steps to diagnose and resolve the issue.
+
+### 1. Verify IP Address Configuration
+Make sure that the correct ip for the Pi is set in the configuration file. To check the Raspberry Pi's IP address, run (on the Pi command line):
+```bash
+hostname -I
+```
+
+### 2. Check if Pi is reachable from laptop/pc
+Try pinging the Raspberry Pi from your laptop:
+```bach
+ping <your_pi_ip_address>
+```
+
+If the ping fails:
+- Ensure the Pi is powered on and connected to the same network.
+- Check if SSH is enabled on the Pi.
+
+### 3. Try SSH connection
+If you can't SSH into the Pi, it might not be properly connected. Use:
+```bash
+ssh <your_pi_user_name>@<your_pi_ip_address>
+```
+If you get a connection error:
+- Ensure SSH is enabled on the Pi by running:
+  ```bash
+  sudo raspi-config
+  ```
+  Then navigate to: **Interfacing Options -> SSH** and enable it.
+
+### 4. Same config file
+Make sure the configuration file on both your laptop/pc and the Raspberry Pi is the same.
+
+# G. Record a dataset
+Once you're familiar with teleoperation, you can record your first dataset with LeKiwi.
+
+To start the program on LeKiwi, SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=remote_robot
+```
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+On your laptop then run this command to record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/lekiwi_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
+```
+
+Note: You can resume recording by adding `--control.resume=true`.
+
+### Wired version
+If you have the **wired** LeKiwi version please run all commands including both these record dataset commands on your laptop.
+
+# H. Visualize a dataset
+
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+```bash
+echo ${HF_USER}/lekiwi_test
+```
+
+If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with (a window can be opened in the browser `http://127.0.0.1:9090` with the visualization tool):
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/lekiwi_test \
+  --local-files-only 1
+```
+
+# I. Replay an episode
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/lekiwi_test \
+  --control.episode=0
+```
+
+## J. Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+python lerobot/scripts/train.py \
+  --dataset.repo_id=${HF_USER}/lekiwi_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_lekiwi_test \
+  --job_name=act_lekiwi_test \
+  --policy.device=cuda \
+  --wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/lekiwi_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+
+Training should take several hours. You will find checkpoints in `outputs/train/act_lekiwi_test/checkpoints`.
+
+## K. Evaluate your policy
+
+You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Drive to the red block and pick it up" \
+  --control.repo_id=${HF_USER}/eval_act_lekiwi_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=10 \
+  --control.push_to_hub=true \
+  --control.policy.path=outputs/train/act_lekiwi_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_lekiwi_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_lekiwi_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_lekiwi_test`).

examples/11_use_moss.md

@@ -0,0 +1,337 @@
+This tutorial explains how to use [Moss v1](https://github.com/jess-moss/moss-robot-arms) with LeRobot.
+
+## Source the parts
+
+Follow this [README](https://github.com/jess-moss/moss-robot-arms). It contains the bill of materials with link to source the parts, as well as the instructions to 3D print the parts and advice if it's your first time printing or if you don't own a 3D printer already.
+
+**Important**: Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
+
+## Install LeRobot
+
+On your computer:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
+bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+rm ~/miniconda3/miniconda.sh
+~/miniconda3/bin/conda init bash
+```
+
+2. Restart shell or `source ~/.bashrc`
+
+3. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+5. Install ffmpeg in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+6. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+## Configure the motors
+
+Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the use of our scripts below.
+
+**Find USB ports associated to your arms**
+To find the correct ports for each arm, run the utility script twice:
+```bash
+python lerobot/scripts/find_motors_bus_port.py
+```
+
+Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
+
+[...Disconnect leader arm and press Enter...]
+
+The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
+Reconnect the usb cable.
+```
+
+Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
+
+[...Disconnect follower arm and press Enter...]
+
+The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the usb cable.
+```
+
+Troubleshooting: On Linux, you might need to give access to the USB ports by running:
+```bash
+sudo chmod 666 /dev/ttyACM0
+sudo chmod 666 /dev/ttyACM1
+```
+
+#### Update config file
+
+IMPORTANTLY: Now that you have your ports, update the **port** default values of [`MossRobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
+```python
+@RobotConfig.register_subclass("moss")
+@dataclass
+class MossRobotConfig(ManipulatorRobotConfig):
+    calibration_dir: str = ".cache/calibration/moss"
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
+
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+```
+
+**Configure your motors**
+Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 1
+```
+
+Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
+
+Then unplug your motor and plug the second motor and set its ID to 2.
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 2
+```
+
+Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
+
+**Remove the gears of the 6 leader motors**
+Follow step 2 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
+
+**Add motor horn to the motors**
+Follow step 3 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). For Moss v1, you need to align the holes on the motor horn to the motor spline to be approximately 3, 6, 9 and 12 o'clock.
+Try to avoid rotating the motor while doing so to keep position 2048 set during configuration. It is especially tricky for the leader motors as it is more sensible without the gears, but it's ok if it's a bit rotated.
+
+## Assemble the arms
+
+Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
+
+## Calibrate
+
+Next, you'll need to calibrate your Moss v1 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one Moss v1 robot to work on another.
+
+**Manual calibration of follower arm**
+/!\ Contrarily to step 6 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
+
+You will need to move the follower arm to these positions sequentially:
+
+| 1. Zero position                                                                                                                                              | 2. Rotated position                                                                                                                                                    | 3. Rest position                                                                                                                                              |
+| ------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <img src="../media/moss/follower_zero.webp?raw=true" alt="Moss v1 follower arm zero position" title="Moss v1 follower arm zero position" style="width:100%;"> | <img src="../media/moss/follower_rotated.webp?raw=true" alt="Moss v1 follower arm rotated position" title="Moss v1 follower arm rotated position" style="width:100%;"> | <img src="../media/moss/follower_rest.webp?raw=true" alt="Moss v1 follower arm rest position" title="Moss v1 follower arm rest position" style="width:100%;"> |
+
+Make sure both arms are connected and run this script to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_follower"]'
+```
+
+**Manual calibration of leader arm**
+Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
+
+| 1. Zero position                                                                                                                                        | 2. Rotated position                                                                                                                                              | 3. Rest position                                                                                                                                        |
+| ------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| <img src="../media/moss/leader_zero.webp?raw=true" alt="Moss v1 leader arm zero position" title="Moss v1 leader arm zero position" style="width:100%;"> | <img src="../media/moss/leader_rotated.webp?raw=true" alt="Moss v1 leader arm rotated position" title="Moss v1 leader arm rotated position" style="width:100%;"> | <img src="../media/moss/leader_rest.webp?raw=true" alt="Moss v1 leader arm rest position" title="Moss v1 leader arm rest position" style="width:100%;"> |
+
+Run this script to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_leader"]'
+```
+
+## Teleoperate
+
+**Simple teleop**
+Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --robot.cameras='{}' \
+  --control.type=teleoperate
+```
+
+
+**Teleop with displaying cameras**
+Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
+
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --control.type=teleoperate
+```
+
+## Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset with Moss v1.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/moss_test \
+  --control.tags='["moss","tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
+```
+
+Note: You can resume recording by adding `--control.resume=true`.
+
+## Visualize a dataset
+
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+```bash
+echo ${HF_USER}/moss_test
+```
+
+If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with:
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/moss_test \
+  --local-files-only 1
+```
+
+## Replay an episode
+
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/moss_test \
+  --control.episode=0
+```
+
+## Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+python lerobot/scripts/train.py \
+  --dataset.repo_id=${HF_USER}/moss_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_moss_test \
+  --job_name=act_moss_test \
+  --policy.device=cuda \
+  --wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/moss_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+
+Training should take several hours. You will find checkpoints in `outputs/train/act_moss_test/checkpoints`.
+
+## Evaluate your policy
+
+You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/eval_act_moss_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=10 \
+  --control.push_to_hub=true \
+  --control.policy.path=outputs/train/act_moss_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_moss_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_moss_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_moss_test`).
+
+## More
+
+Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
+
+If you have any question or need help, please reach out on Discord in the channel [`#moss-arm`](https://discord.com/channels/1216765309076115607/1275374638985252925).

examples/1_load_lerobot_dataset.py

@@ -0,0 +1,148 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script demonstrates the use of `LeRobotDataset` class for handling and processing robotic datasets from Hugging Face.
+It illustrates how to load datasets, manipulate them, and apply transformations suitable for machine learning tasks in PyTorch.
+
+Features included in this script:
+- Viewing a dataset's metadata and exploring its properties.
+- Loading an existing dataset from the hub or a subset of it.
+- Accessing frames by episode number.
+- Using advanced dataset features like timestamp-based frame selection.
+- Demonstrating compatibility with PyTorch DataLoader for batch processing.
+
+The script ends with examples of how to batch process data using PyTorch's DataLoader.
+"""
+
+from pprint import pprint
+
+import torch
+from huggingface_hub import HfApi
+
+import lerobot
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+
+# We ported a number of existing datasets ourselves, use this to see the list:
+print("List of available datasets:")
+pprint(lerobot.available_datasets)
+
+# You can also browse through the datasets created/ported by the community on the hub using the hub api:
+hub_api = HfApi()
+repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
+pprint(repo_ids)
+
+# Or simply explore them in your web browser directly at:
+# https://huggingface.co/datasets?other=LeRobot
+
+# Let's take this one for this example
+repo_id = "lerobot/aloha_mobile_cabinet"
+# We can have a look and fetch its metadata to know more about it:
+ds_meta = LeRobotDatasetMetadata(repo_id)
+
+# By instantiating just this class, you can quickly access useful information about the content and the
+# structure of the dataset without downloading the actual data yet (only metadata files — which are
+# lightweight).
+print(f"Total number of episodes: {ds_meta.total_episodes}")
+print(f"Average number of frames per episode: {ds_meta.total_frames / ds_meta.total_episodes:.3f}")
+print(f"Frames per second used during data collection: {ds_meta.fps}")
+print(f"Robot type: {ds_meta.robot_type}")
+print(f"keys to access images from cameras: {ds_meta.camera_keys=}\n")
+
+print("Tasks:")
+print(ds_meta.tasks)
+print("Features:")
+pprint(ds_meta.features)
+
+# You can also get a short summary by simply printing the object:
+print(ds_meta)
+
+# You can then load the actual dataset from the hub.
+# Either load any subset of episodes:
+dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
+
+# And see how many frames you have:
+print(f"Selected episodes: {dataset.episodes}")
+print(f"Number of episodes selected: {dataset.num_episodes}")
+print(f"Number of frames selected: {dataset.num_frames}")
+
+# Or simply load the entire dataset:
+dataset = LeRobotDataset(repo_id)
+print(f"Number of episodes selected: {dataset.num_episodes}")
+print(f"Number of frames selected: {dataset.num_frames}")
+
+# The previous metadata class is contained in the 'meta' attribute of the dataset:
+print(dataset.meta)
+
+# LeRobotDataset actually wraps an underlying Hugging Face dataset
+# (see https://huggingface.co/docs/datasets for more information).
+print(dataset.hf_dataset)
+
+# LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
+# with the latter, like iterating through the dataset.
+# The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
+# episodes, you can access the frame indices of any episode using the episode_data_index. Here, we access
+# frame indices associated to the first episode:
+episode_index = 0
+from_idx = dataset.episode_data_index["from"][episode_index].item()
+to_idx = dataset.episode_data_index["to"][episode_index].item()
+
+# Then we grab all the image frames from the first camera:
+camera_key = dataset.meta.camera_keys[0]
+frames = [dataset[idx][camera_key] for idx in range(from_idx, to_idx)]
+
+# The objects returned by the dataset are all torch.Tensors
+print(type(frames[0]))
+print(frames[0].shape)
+
+# Since we're using pytorch, the shape is in pytorch, channel-first convention (c, h, w).
+# We can compare this shape with the information available for that feature
+pprint(dataset.features[camera_key])
+# In particular:
+print(dataset.features[camera_key]["shape"])
+# The shape is in (h, w, c) which is a more universal format.
+
+# For many machine learning applications we need to load the history of past observations or trajectories of
+# future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
+# differences with the current loaded frame. For instance:
+delta_timestamps = {
+    # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
+    camera_key: [-1, -0.5, -0.20, 0],
+    # loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
+    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
+    # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
+    "action": [t / dataset.fps for t in range(64)],
+}
+# Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
+# timestamp, you still get a valid timestamp.
+
+dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
+print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
+print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
+print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
+
+# Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
+# PyTorch datasets.
+dataloader = torch.utils.data.DataLoader(
+    dataset,
+    num_workers=0,
+    batch_size=32,
+    shuffle=True,
+)
+
+for batch in dataloader:
+    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
+    print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
+    print(f"{batch['action'].shape=}")  # (32, 64, c)
+    break

examples/2_evaluate_pretrained_policy.py

@@ -0,0 +1,139 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This scripts demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
+training outputs directory. In the latter case, you might want to run examples/3_train_policy.py first.
+
+It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
+```bash
+pip install -e ".[pusht]"
+```
+"""
+
+from pathlib import Path
+
+import gym_pusht  # noqa: F401
+import gymnasium as gym
+import imageio
+import numpy
+import torch
+
+from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+
+# Create a directory to store the video of the evaluation
+output_directory = Path("outputs/eval/example_pusht_diffusion")
+output_directory.mkdir(parents=True, exist_ok=True)
+
+# Select your device
+device = "cuda"
+
+# Provide the [hugging face repo id](https://huggingface.co/lerobot/diffusion_pusht):
+pretrained_policy_path = "lerobot/diffusion_pusht"
+# OR a path to a local outputs/train folder.
+# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
+
+policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
+
+# Initialize evaluation environment to render two observation types:
+# an image of the scene and state/position of the agent. The environment
+# also automatically stops running after 300 interactions/steps.
+env = gym.make(
+    "gym_pusht/PushT-v0",
+    obs_type="pixels_agent_pos",
+    max_episode_steps=300,
+)
+
+# We can verify that the shapes of the features expected by the policy match the ones from the observations
+# produced by the environment
+print(policy.config.input_features)
+print(env.observation_space)
+
+# Similarly, we can check that the actions produced by the policy will match the actions expected by the
+# environment
+print(policy.config.output_features)
+print(env.action_space)
+
+# Reset the policy and environments to prepare for rollout
+policy.reset()
+numpy_observation, info = env.reset(seed=42)
+
+# Prepare to collect every rewards and all the frames of the episode,
+# from initial state to final state.
+rewards = []
+frames = []
+
+# Render frame of the initial state
+frames.append(env.render())
+
+step = 0
+done = False
+while not done:
+    # Prepare observation for the policy running in Pytorch
+    state = torch.from_numpy(numpy_observation["agent_pos"])
+    image = torch.from_numpy(numpy_observation["pixels"])
+
+    # Convert to float32 with image from channel first in [0,255]
+    # to channel last in [0,1]
+    state = state.to(torch.float32)
+    image = image.to(torch.float32) / 255
+    image = image.permute(2, 0, 1)
+
+    # Send data tensors from CPU to GPU
+    state = state.to(device, non_blocking=True)
+    image = image.to(device, non_blocking=True)
+
+    # Add extra (empty) batch dimension, required to forward the policy
+    state = state.unsqueeze(0)
+    image = image.unsqueeze(0)
+
+    # Create the policy input dictionary
+    observation = {
+        "observation.state": state,
+        "observation.image": image,
+    }
+
+    # Predict the next action with respect to the current observation
+    with torch.inference_mode():
+        action = policy.select_action(observation)
+
+    # Prepare the action for the environment
+    numpy_action = action.squeeze(0).to("cpu").numpy()
+
+    # Step through the environment and receive a new observation
+    numpy_observation, reward, terminated, truncated, info = env.step(numpy_action)
+    print(f"{step=} {reward=} {terminated=}")
+
+    # Keep track of all the rewards and frames
+    rewards.append(reward)
+    frames.append(env.render())
+
+    # The rollout is considered done when the success state is reach (i.e. terminated is True),
+    # or the maximum number of iterations is reached (i.e. truncated is True)
+    done = terminated | truncated | done
+    step += 1
+
+if terminated:
+    print("Success!")
+else:
+    print("Failure!")
+
+# Get the speed of environment (i.e. its number of frames per second).
+fps = env.metadata["render_fps"]
+
+# Encode all frames into a mp4 video.
+video_path = output_directory / "rollout.mp4"
+imageio.mimsave(str(video_path), numpy.stack(frames), fps=fps)
+
+print(f"Video of the evaluation is available in '{video_path}'.")

examples/3_train_policy.py

@@ -0,0 +1,120 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""This scripts demonstrates how to train Diffusion Policy on the PushT environment.
+
+Once you have trained a model with this script, you can try to evaluate it on
+examples/2_evaluate_pretrained_policy.py
+"""
+
+from pathlib import Path
+
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.datasets.utils import dataset_to_policy_features
+from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+from lerobot.configs.types import FeatureType
+
+
+def main():
+    # Create a directory to store the training checkpoint.
+    output_directory = Path("outputs/train/example_pusht_diffusion")
+    output_directory.mkdir(parents=True, exist_ok=True)
+
+    # # Select your device
+    device = torch.device("cuda")
+
+    # Number of offline training steps (we'll only do offline training for this example.)
+    # Adjust as you prefer. 5000 steps are needed to get something worth evaluating.
+    training_steps = 5000
+    log_freq = 1
+
+    # When starting from scratch (i.e. not from a pretrained policy), we need to specify 2 things before
+    # creating the policy:
+    #   - input/output shapes: to properly size the policy
+    #   - dataset stats: for normalization and denormalization of input/outputs
+    dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
+    features = dataset_to_policy_features(dataset_metadata.features)
+    output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    input_features = {key: ft for key, ft in features.items() if key not in output_features}
+
+    # Policies are initialized with a configuration class, in this case `DiffusionConfig`. For this example,
+    # we'll just use the defaults and so no arguments other than input/output features need to be passed.
+    cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
+
+    # We can now instantiate our policy with this config and the dataset stats.
+    policy = DiffusionPolicy(cfg, dataset_stats=dataset_metadata.stats)
+    policy.train()
+    policy.to(device)
+
+    # Another policy-dataset interaction is with the delta_timestamps. Each policy expects a given number frames
+    # which can differ for inputs, outputs and rewards (if there are some).
+    delta_timestamps = {
+        "observation.image": [i / dataset_metadata.fps for i in cfg.observation_delta_indices],
+        "observation.state": [i / dataset_metadata.fps for i in cfg.observation_delta_indices],
+        "action": [i / dataset_metadata.fps for i in cfg.action_delta_indices],
+    }
+
+    # In this case with the standard configuration for Diffusion Policy, it is equivalent to this:
+    delta_timestamps = {
+        # Load the previous image and state at -0.1 seconds before current frame,
+        # then load current image and state corresponding to 0.0 second.
+        "observation.image": [-0.1, 0.0],
+        "observation.state": [-0.1, 0.0],
+        # Load the previous action (-0.1), the next action to be executed (0.0),
+        # and 14 future actions with a 0.1 seconds spacing. All these actions will be
+        # used to supervise the policy.
+        "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
+    }
+
+    # We can then instantiate the dataset with these delta_timestamps configuration.
+    dataset = LeRobotDataset("lerobot/pusht", delta_timestamps=delta_timestamps)
+
+    # Then we create our optimizer and dataloader for offline training.
+    optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=4,
+        batch_size=64,
+        shuffle=True,
+        pin_memory=device.type != "cpu",
+        drop_last=True,
+    )
+
+    # Run training loop.
+    step = 0
+    done = False
+    while not done:
+        for batch in dataloader:
+            batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
+            loss, _ = policy.forward(batch)
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+
+            if step % log_freq == 0:
+                print(f"step: {step} loss: {loss.item():.3f}")
+            step += 1
+            if step >= training_steps:
+                done = True
+                break
+
+    # Save a policy checkpoint.
+    policy.save_pretrained(output_directory)
+
+
+if __name__ == "__main__":
+    main()

examples/4_train_policy_with_script.md

@@ -0,0 +1,274 @@
+This tutorial will explain the training script, how to use it, and particularly how to configure everything needed for the training run.
+> **Note:** The following assume you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
+
+
+## The training script
+
+LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
+
+- Initialize/load a configuration for the following steps using.
+- Instantiates a dataset.
+- (Optional) Instantiates a simulation environment corresponding to that dataset.
+- Instantiates a policy.
+- Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
+
+## Overview of the configuration system
+
+In the training script, the main function `train` expects a `TrainPipelineConfig` object:
+```python
+# train.py
+@parser.wrap()
+def train(cfg: TrainPipelineConfig):
+```
+
+You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
+
+When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated for this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
+
+Let's have a look at a simplified example. Amongst other attributes, the training config has the following attributes:
+```python
+@dataclass
+class TrainPipelineConfig:
+    dataset: DatasetConfig
+    env: envs.EnvConfig | None = None
+    policy: PreTrainedConfig | None = None
+```
+in which `DatasetConfig` for example is defined as such:
+```python
+@dataclass
+class DatasetConfig:
+    repo_id: str
+    episodes: list[int] | None = None
+    video_backend: str = "pyav"
+```
+
+This creates a hierarchical relationship where, for example assuming we have a `cfg` instance of `TrainPipelineConfig`, we can access the `repo_id` value with `cfg.dataset.repo_id`.
+From the command line, we can specify this value with using a very similar syntax `--dataset.repo_id=repo/id`.
+
+By default, every field takes its default value specified in the dataclass. If a field doesn't have a default value, it needs to be specified either from the command line or from a config file – which path is also given in the command line (more in this below). In the example above, the `dataset` field doesn't have a default value which means it must be specified.
+
+
+## Specifying values from the CLI
+
+Let's say that we want to train [Diffusion Policy](../../lerobot/common/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
+```bash
+python lerobot/scripts/train.py \
+    --dataset.repo_id=lerobot/pusht \
+    --policy.type=diffusion \
+    --env.type=pusht
+```
+
+Let's break this down:
+- To specify the dataset, we just need to specify its `repo_id` on the hub which is the only required argument in the `DatasetConfig`. The rest of the fields have default values and in this case we are fine with those so we can just add the option `--dataset.repo_id=lerobot/pusht`.
+- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../../lerobot/common/policies)
+- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../../lerobot/common/envs/configs.py)
+
+Let's see another example. Let's say you've been training [ACT](../../lerobot/common/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
+```bash
+python lerobot/scripts/train.py \
+    --policy.type=act \
+    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
+    --env.type=aloha \
+    --output_dir=outputs/train/act_aloha_insertion
+```
+> Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.
+
+We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
+Looking at the [`AlohaEnv`](../../lerobot/common/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
+```bash
+python lerobot/scripts/train.py \
+    --policy.type=act \
+    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
+    --env.type=aloha \
+    --env.task=AlohaTransferCube-v0 \
+    --output_dir=outputs/train/act_aloha_transfer
+```
+
+## Loading from a config file
+
+Now, let's assume that we want to reproduce the run just above. That run has produced a `train_config.json` file in its checkpoints, which serializes the `TrainPipelineConfig` instance it used:
+```json
+{
+    "dataset": {
+        "repo_id": "lerobot/aloha_sim_transfer_cube_human",
+        "episodes": null,
+        ...
+    },
+    "env": {
+        "type": "aloha",
+        "task": "AlohaTransferCube-v0",
+        "fps": 50,
+        ...
+    },
+    "policy": {
+        "type": "act",
+        "n_obs_steps": 1,
+        ...
+    },
+    ...
+}
+```
+
+We can then simply load the config values from this file using:
+```bash
+python lerobot/scripts/train.py \
+    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
+    --output_dir=outputs/train/act_aloha_transfer_2
+```
+`--config_path` is also a special argument which allows to initialize the config from a local config file. It can point to a directory that contains `train_config.json` or to the config file itself directly.
+
+Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
+```bash
+python lerobot/scripts/train.py \
+    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
+    --output_dir=outputs/train/act_aloha_transfer_2
+    --policy.n_action_steps=80
+```
+> Note: While `--output_dir` is not required in general, in this case we need to specify it since it will otherwise take the value from the `train_config.json` (which is `outputs/train/act_aloha_transfer`). In order to prevent accidental deletion of previous run checkpoints, we raise an error if you're trying to write in an existing directory. This is not the case when resuming a run, which is what you'll learn next.
+
+`--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:
+```bash
+python lerobot/scripts/train.py --config_path=lerobot/diffusion_pusht
+```
+will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
+
+
+## Resume training
+
+Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to that here.
+
+Let's reuse the command from the previous run and add a few more options:
+```bash
+python lerobot/scripts/train.py \
+    --policy.type=act \
+    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
+    --env.type=aloha \
+    --env.task=AlohaTransferCube-v0 \
+    --log_freq=25 \
+    --save_freq=100 \
+    --output_dir=outputs/train/run_resumption
+```
+
+Here we've taken care to set up the log frequency and checkpointing frequency to low numbers so we can showcase resumption. You should be able to see some logging and have a first checkpoint within 1 minute (depending on hardware). Wait for the first checkpoint to happen, you should see a line that looks like this in your terminal:
+```
+INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
+```
+Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:
+```bash
+python lerobot/scripts/train.py \
+    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
+    --resume=true
+```
+You should see from the logging that your training picks up from where it left off.
+
+Another reason for which you might want to resume a run is simply to extend training and add more training steps. The number of training steps is set by the option `--steps`, which is 100 000 by default.
+You could double the number of steps of the previous run with:
+```bash
+python lerobot/scripts/train.py \
+    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
+    --resume=true \
+    --steps=200000
+```
+
+## Outputs of a run
+In the output directory, there will be a folder called `checkpoints` with the following structure:
+```bash
+outputs/train/run_resumption/checkpoints
+├── 000100  # checkpoint_dir for training step 100
+│   ├── pretrained_model/
+│   │   ├── config.json  # policy config
+│   │   ├── model.safetensors  # policy weights
+│   │   └── train_config.json  # train config
+│   └── training_state/
+│       ├── optimizer_param_groups.json  #  optimizer param groups
+│       ├── optimizer_state.safetensors  # optimizer state
+│       ├── rng_state.safetensors  # rng states
+│       ├── scheduler_state.json  # scheduler state
+│       └── training_step.json  # training step
+├── 000200
+└── last -> 000200  # symlink to the last available checkpoint
+```
+
+## Fine-tuning a pre-trained policy
+
+In addition to the features currently in Draccus, we've added a special `.path` argument for the policy, which allows to load a policy as you would with `PreTrainedPolicy.from_pretrained()`. In that case, `path` can be a local directory that contains a checkpoint or a repo_id pointing to a pretrained policy on the hub.
+
+For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:
+```bash
+python lerobot/scripts/train.py \
+    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \
+    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
+    --env.type=aloha \
+    --env.task=AlohaInsertion-v0
+```
+
+When doing so, keep in mind that the features of the fine-tuning dataset would have to match the input/output features of the pretrained policy.
+
+## Typical logs and metrics
+
+When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you configured your run correctly. The final configuration will also be saved with the checkpoint.
+
+After that, you will see training log like this one:
+```
+INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
+```
+or evaluation log:
+```
+INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
+```
+
+These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
+- `smpl`: number of samples seen during training.
+- `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
+- `epch`: number of time all unique samples are seen (epoch).
+- `grdn`: gradient norm.
+- `∑rwrd`: compute the sum of rewards in every evaluation episode and then take an average of them.
+- `success`: average success rate of eval episodes. Reward and success are usually different except for the sparsing reward setting, where reward=1 only when the task is completed successfully.
+- `eval_s`: time to evaluate the policy in the environment, in second.
+- `updt_s`: time to update the network parameters, in second.
+- `data_s`: time to load a batch of data, in second.
+
+Some metrics are useful for initial performance profiling. For example, if you find the current GPU utilization is low via the `nvidia-smi` command and `data_s` sometimes is too high, you may need to modify batch size or number of dataloading workers to accelerate dataloading. We also recommend [pytorch profiler](https://github.com/huggingface/lerobot?tab=readme-ov-file#improve-your-code-with-profiling) for detailed performance probing.
+
+## In short
+
+We'll summarize here the main use cases to remember from this tutorial.
+
+#### Train a policy from scratch – CLI
+```bash
+python lerobot/scripts/train.py \
+    --policy.type=act \  # <- select 'act' policy
+    --env.type=pusht \  # <- select 'pusht' environment
+    --dataset.repo_id=lerobot/pusht  # <- train on this dataset
+```
+
+#### Train a policy from scratch - config file + CLI
+```bash
+python lerobot/scripts/train.py \
+    --config_path=path/to/pretrained_model \  # <- can also be a repo_id
+    --policy.n_action_steps=80  # <- you may still override values
+```
+
+#### Resume/continue a training run
+```bash
+python lerobot/scripts/train.py \
+    --config_path=checkpoint/pretrained_model/ \
+    --resume=true \
+    --steps=200000  # <- you can change some training parameters
+```
+
+#### Fine-tuning
+```bash
+python lerobot/scripts/train.py \
+    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \  # <- can also be a local path to a checkpoint
+    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
+    --env.type=aloha \
+    --env.task=AlohaInsertion-v0
+```
+
+---
+
+Now that you know the basics of how to train a policy, you might want to know how to apply this knowledge to actual robots, or how to record your own datasets and train policies on your specific task?
+If that's the case, head over to the next tutorial [`7_get_started_with_real_robot.md`](./7_get_started_with_real_robot.md).
+
+Or in the meantime, happy training! 🤗

examples/8_use_stretch.md

@@ -0,0 +1,161 @@
+This tutorial explains how to use [Stretch 3](https://hello-robot.com/stretch-3-product) with LeRobot.
+
+## Setup
+
+Familiarize yourself with Stretch by following its [tutorials](https://docs.hello-robot.com/0.3/getting_started/hello_robot/) (recommended).
+
+To use LeRobot on Stretch, 3 options are available:
+- [tethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#tethered-setup)
+- [untethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#untethered-setup)
+- ssh directly into Stretch (you will first need to install and configure openssh-server on stretch using one of the two above setups)
+
+
+## Install LeRobot
+
+On Stretch's CLI, follow these steps:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
+bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+rm ~/miniconda3/miniconda.sh
+~/miniconda3/bin/conda init bash
+```
+
+2. Comment out these lines in `~/.profile` (this can mess up paths used by conda and ~/.local/bin should already be in your PATH)
+```
+# set PATH so it includes user's private bin if it exists
+if [ -d "$HOME/.local/bin" ] ; then
+    PATH="$HOME/.local/bin:$PATH"
+fi
+```
+
+3. Restart shell or `source ~/.bashrc`
+
+4. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+5. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+6. When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+7. Install LeRobot with stretch dependencies:
+```bash
+cd ~/lerobot && pip install -e ".[stretch]"
+```
+
+> **Note:** If you get this message, you can ignore it: `ERROR: pip's dependency resolver does not currently take into account all the packages that are installed.`
+
+8. Run a [system check](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#system-check) to make sure your robot is ready:
+```bash
+stretch_system_check.py
+```
+
+> **Note:** You may need to free the "robot process" after booting Stretch by running `stretch_free_robot_process.py`. For more info this Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#turning-off-gamepad-teleoperation).
+
+You should get something like this:
+```bash
+For use with S T R E T C H (R) from Hello Robot Inc.
+---------------------------------------------------------------------
+
+Model = Stretch 3
+Tool = DexWrist 3 w/ Gripper
+Serial Number = stretch-se3-3054
+
+---- Checking Hardware ----
+[Pass] Comms are ready
+[Pass] Actuators are ready
+[Warn] Sensors not ready (IMU AZ = -10.19 out of range -10.1 to -9.5)
+[Pass] Battery voltage is 13.6 V
+
+---- Checking Software ----
+[Pass] Ubuntu 22.04 is ready
+[Pass] All APT pkgs are setup correctly
+[Pass] Firmware is up-to-date
+[Pass] Python pkgs are up-to-date
+[Pass] ROS2 Humble is ready
+```
+
+## Teleoperate, record a dataset and run a policy
+
+**Calibrate (Optional)**
+Before operating Stretch, you need to [home](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#homing) it first. Be mindful about giving Stretch some space as this procedure will move the robot's arm and gripper. Now run this command:
+```bash
+python lerobot/scripts/control_robot.py \
+    --robot.type=stretch \
+    --control.type=calibrate
+```
+This is equivalent to running `stretch_robot_home.py`
+
+> **Note:** If you run any of the LeRobot scripts below and Stretch is not properly homed, it will automatically home/calibrate first.
+
+**Teleoperate**
+Before trying teleoperation, you need activate the gamepad controller by pressing the middle button. For more info, see Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/hello_robot/#gamepad-teleoperation).
+
+Now try out teleoperation (see above documentation to learn about the gamepad controls):
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
+```bash
+python lerobot/scripts/control_robot.py \
+    --robot.type=stretch \
+    --control.type=teleoperate
+```
+This is essentially the same as running `stretch_gamepad_teleop.py`
+
+**Record a dataset**
+Once you're familiar with the gamepad controls and after a bit of practice, you can try to record your first dataset with Stretch.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record one episode:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=stretch \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/stretch_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
+```
+
+> **Note:** If you're using ssh to connect to Stretch and run this script, you won't be able to visualize its cameras feed (though they will still be recording). To see the cameras stream, use [tethered](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#tethered-setup) or [untethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#untethered-setup).
+
+**Replay an episode**
+Now try to replay this episode (make sure the robot's initial position is the same):
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=stretch \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/stretch_test \
+  --control.episode=0
+```
+
+Follow [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) to train a policy on your data and run inference on your robot. You will need to adapt the code for Stretch.
+
+> TODO(rcadene, aliberts): Add already setup environment and policy yaml configuration files
+
+If you need help, please reach out on Discord in the channel `#stretch3-mobile-arm`.

examples/9_use_aloha.md

@@ -0,0 +1,182 @@
+This tutorial explains how to use [Aloha and Aloha 2 stationary](https://www.trossenrobotics.com/aloha-stationary) with LeRobot.
+
+## Setup
+
+Follow the [documentation from Trossen Robotics](https://docs.trossenrobotics.com/aloha_docs/2.0/getting_started/stationary/hardware_setup.html) for setting up the hardware and plugging the 4 arms and 4 cameras to your computer.
+
+
+## Install LeRobot
+
+On your computer:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
+bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+rm ~/miniconda3/miniconda.sh
+~/miniconda3/bin/conda init bash
+```
+
+2. Restart shell or `source ~/.bashrc`
+
+3. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+5. When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+6. Install LeRobot with dependencies for the Aloha motors (dynamixel) and cameras (intelrealsense):
+```bash
+cd ~/lerobot && pip install -e ".[dynamixel, intelrealsense]"
+```
+
+## Teleoperate
+
+**/!\ FOR SAFETY, READ THIS /!\**
+Teleoperation consists in manually operating the leader arms to move the follower arms. Importantly:
+1. Make sure your leader arms are in the same position as the follower arms, so that the follower arms don't move too fast to match the leader arms,
+2. Our code assumes that your robot has been assembled following Trossen Robotics instructions. This allows us to skip calibration, as we use the pre-defined calibration files in `.cache/calibration/aloha_default`. If you replace a motor, make sure you follow the exact instructions from Trossen Robotics.
+
+By running the following code, you can start your first **SAFE** teleoperation:
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
+
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=aloha \
+  --robot.max_relative_target=5 \
+  --control.type=teleoperate
+```
+
+By adding `--robot.max_relative_target=5`, we override the default value for `max_relative_target` defined in [`AlohaRobotConfig`](lerobot/common/robot_devices/robots/configs.py). It is expected to be `5` to limit the magnitude of the movement for more safety, but the teleoperation won't be smooth. When you feel confident, you can disable this limit by adding `--robot.max_relative_target=null` to the command line:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=aloha \
+  --robot.max_relative_target=null \
+  --control.type=teleoperate
+```
+
+## Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset with Aloha.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=aloha \
+  --robot.max_relative_target=null \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/aloha_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
+```
+
+## Visualize a dataset
+
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+```bash
+echo ${HF_USER}/aloha_test
+```
+
+If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with:
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/aloha_test
+```
+
+## Replay an episode
+
+**/!\ FOR SAFETY, READ THIS /!\**
+Replay consists in automatically replaying the sequence of actions (i.e. goal positions for your motors) recorded in a given dataset episode. Make sure the current initial position of your robot is similar to the one in your episode, so that your follower arms don't move too fast to go to the first goal positions. For safety, you might want to add `--robot.max_relative_target=5` to your command line as explained above.
+
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=aloha \
+  --robot.max_relative_target=null \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/aloha_test \
+  --control.episode=0
+```
+
+## Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+python lerobot/scripts/train.py \
+  --dataset.repo_id=${HF_USER}/aloha_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_aloha_test \
+  --job_name=act_aloha_test \
+  --policy.device=cuda \
+  --wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/aloha_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+
+For more information on the `train` script see the previous tutorial: [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md)
+
+Training should take several hours. You will find checkpoints in `outputs/train/act_aloha_test/checkpoints`.
+
+## Evaluate your policy
+
+You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=aloha \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/eval_act_aloha_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=10 \
+  --control.push_to_hub=true \
+  --control.policy.path=outputs/train/act_aloha_test/checkpoints/last/pretrained_model \
+  --control.num_image_writer_processes=1
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_aloha_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_aloha_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_aloha_test`).
+3. We use `--control.num_image_writer_processes=1` instead of the default value (`0`). On our computer, using a dedicated process to write images from the 4 cameras on disk allows to reach constant 30 fps during inference. Feel free to explore different values for `--control.num_image_writer_processes`.
+
+## More
+
+Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth explanation.
+
+If you have any question or need help, please reach out on Discord in the channel `#aloha-arm`.

examples/advanced/1_add_image_transforms.py

@@ -0,0 +1,67 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script demonstrates how to use torchvision's image transformation with LeRobotDataset for data
+augmentation purposes. The transformations are passed to the dataset as an argument upon creation, and
+transforms are applied to the observation images before they are returned in the dataset's __getitem__.
+"""
+
+from pathlib import Path
+
+from torchvision.transforms import ToPILImage, v2
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+dataset_repo_id = "lerobot/aloha_static_screw_driver"
+
+# Create a LeRobotDataset with no transformations
+dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
+# This is equivalent to `dataset = LeRobotDataset(dataset_repo_id, image_transforms=None)`
+
+# Get the index of the first observation in the first episode
+first_idx = dataset.episode_data_index["from"][0].item()
+
+# Get the frame corresponding to the first camera
+frame = dataset[first_idx][dataset.meta.camera_keys[0]]
+
+
+# Define the transformations
+transforms = v2.Compose(
+    [
+        v2.ColorJitter(brightness=(0.5, 1.5)),
+        v2.ColorJitter(contrast=(0.5, 1.5)),
+        v2.ColorJitter(hue=(-0.1, 0.1)),
+        v2.RandomAdjustSharpness(sharpness_factor=2, p=1),
+    ]
+)
+
+# Create another LeRobotDataset with the defined transformations
+transformed_dataset = LeRobotDataset(dataset_repo_id, episodes=[0], image_transforms=transforms)
+
+# Get a frame from the transformed dataset
+transformed_frame = transformed_dataset[first_idx][transformed_dataset.meta.camera_keys[0]]
+
+# Create a directory to store output images
+output_dir = Path("outputs/image_transforms")
+output_dir.mkdir(parents=True, exist_ok=True)
+
+# Save the original frame
+to_pil = ToPILImage()
+to_pil(frame).save(output_dir / "original_frame.png", quality=100)
+print(f"Original frame saved to {output_dir / 'original_frame.png'}.")
+
+# Save the transformed frame
+to_pil(transformed_frame).save(output_dir / "transformed_frame.png", quality=100)
+print(f"Transformed frame saved to {output_dir / 'transformed_frame.png'}.")

examples/advanced/2_calculate_validation_loss.py

@@ -0,0 +1,104 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""This script demonstrates how to slice a dataset and calculate the loss on a subset of the data.
+
+This technique can be useful for debugging and testing purposes, as well as identifying whether a policy
+is learning effectively.
+
+Furthermore, relying on validation loss to evaluate performance is generally not considered a good practice,
+especially in the context of imitation learning. The most reliable approach is to evaluate the policy directly
+on the target environment, whether that be in simulation or the real world.
+"""
+
+import math
+
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+
+
+def main():
+    device = torch.device("cuda")
+
+    # Download the diffusion policy for pusht environment
+    pretrained_policy_path = "lerobot/diffusion_pusht"
+    # OR uncomment the following to evaluate a policy from the local outputs/train folder.
+    # pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
+
+    policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
+    policy.eval()
+    policy.to(device)
+
+    # Set up the dataset.
+    delta_timestamps = {
+        # Load the previous image and state at -0.1 seconds before current frame,
+        # then load current image and state corresponding to 0.0 second.
+        "observation.image": [-0.1, 0.0],
+        "observation.state": [-0.1, 0.0],
+        # Load the previous action (-0.1), the next action to be executed (0.0),
+        # and 14 future actions with a 0.1 seconds spacing. All these actions will be
+        # used to calculate the loss.
+        "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
+    }
+
+    # Load the last 10% of episodes of the dataset as a validation set.
+    # - Load dataset metadata
+    dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
+    # - Calculate train and val episodes
+    total_episodes = dataset_metadata.total_episodes
+    episodes = list(range(dataset_metadata.total_episodes))
+    num_train_episodes = math.floor(total_episodes * 90 / 100)
+    train_episodes = episodes[:num_train_episodes]
+    val_episodes = episodes[num_train_episodes:]
+    print(f"Number of episodes in full dataset: {total_episodes}")
+    print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
+    print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
+    # - Load train an val datasets
+    train_dataset = LeRobotDataset(
+        "lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps
+    )
+    val_dataset = LeRobotDataset("lerobot/pusht", episodes=val_episodes, delta_timestamps=delta_timestamps)
+    print(f"Number of frames in training dataset (90% subset): {len(train_dataset)}")
+    print(f"Number of frames in validation dataset (10% subset): {len(val_dataset)}")
+
+    # Create dataloader for evaluation.
+    val_dataloader = torch.utils.data.DataLoader(
+        val_dataset,
+        num_workers=4,
+        batch_size=64,
+        shuffle=False,
+        pin_memory=device != torch.device("cpu"),
+        drop_last=False,
+    )
+
+    # Run validation loop.
+    loss_cumsum = 0
+    n_examples_evaluated = 0
+    for batch in val_dataloader:
+        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+        loss, _ = policy.forward(batch)
+
+        loss_cumsum += loss.item()
+        n_examples_evaluated += batch["index"].shape[0]
+
+    # Calculate the average loss over the validation set.
+    average_loss = loss_cumsum / n_examples_evaluated
+
+    print(f"Average loss on validation set: {average_loss:.4f}")
+
+
+if __name__ == "__main__":
+    main()

lerobot/__init__.py

@@ -0,0 +1,217 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This file contains lists of available environments, dataset and policies to reflect the current state of LeRobot library.
+We do not want to import all the dependencies, but instead we keep it lightweight to ensure fast access to these variables.
+
+Example:
+    ```python
+        import lerobot
+        print(lerobot.available_envs)
+        print(lerobot.available_tasks_per_env)
+        print(lerobot.available_datasets)
+        print(lerobot.available_datasets_per_env)
+        print(lerobot.available_real_world_datasets)
+        print(lerobot.available_policies)
+        print(lerobot.available_policies_per_env)
+        print(lerobot.available_robots)
+        print(lerobot.available_cameras)
+        print(lerobot.available_motors)
+    ```
+
+When implementing a new dataset loadable with LeRobotDataset follow these steps:
+- Update `available_datasets_per_env` in `lerobot/__init__.py`
+
+When implementing a new environment (e.g. `gym_aloha`), follow these steps:
+- Update `available_tasks_per_env` and `available_datasets_per_env` in `lerobot/__init__.py`
+
+When implementing a new policy class (e.g. `DiffusionPolicy`) follow these steps:
+- Update `available_policies` and `available_policies_per_env`, in `lerobot/__init__.py`
+- Set the required `name` class attribute.
+- Update variables in `tests/test_available.py` by importing your new Policy class
+"""
+
+import itertools
+
+from lerobot.__version__ import __version__  # noqa: F401
+
+# TODO(rcadene): Improve policies and envs. As of now, an item in `available_policies`
+# refers to a yaml file AND a modeling name. Same for `available_envs` which refers to
+# a yaml file AND a environment name. The difference should be more obvious.
+available_tasks_per_env = {
+    "aloha": [
+        "AlohaInsertion-v0",
+        "AlohaTransferCube-v0",
+    ],
+    "pusht": ["PushT-v0"],
+    "xarm": ["XarmLift-v0"],
+}
+available_envs = list(available_tasks_per_env.keys())
+
+available_datasets_per_env = {
+    "aloha": [
+        "lerobot/aloha_sim_insertion_human",
+        "lerobot/aloha_sim_insertion_scripted",
+        "lerobot/aloha_sim_transfer_cube_human",
+        "lerobot/aloha_sim_transfer_cube_scripted",
+        "lerobot/aloha_sim_insertion_human_image",
+        "lerobot/aloha_sim_insertion_scripted_image",
+        "lerobot/aloha_sim_transfer_cube_human_image",
+        "lerobot/aloha_sim_transfer_cube_scripted_image",
+    ],
+    # TODO(alexander-soare): Add "lerobot/pusht_keypoints". Right now we can't because this is too tightly
+    # coupled with tests.
+    "pusht": ["lerobot/pusht", "lerobot/pusht_image"],
+    "xarm": [
+        "lerobot/xarm_lift_medium",
+        "lerobot/xarm_lift_medium_replay",
+        "lerobot/xarm_push_medium",
+        "lerobot/xarm_push_medium_replay",
+        "lerobot/xarm_lift_medium_image",
+        "lerobot/xarm_lift_medium_replay_image",
+        "lerobot/xarm_push_medium_image",
+        "lerobot/xarm_push_medium_replay_image",
+    ],
+}
+
+available_real_world_datasets = [
+    "lerobot/aloha_mobile_cabinet",
+    "lerobot/aloha_mobile_chair",
+    "lerobot/aloha_mobile_elevator",
+    "lerobot/aloha_mobile_shrimp",
+    "lerobot/aloha_mobile_wash_pan",
+    "lerobot/aloha_mobile_wipe_wine",
+    "lerobot/aloha_static_battery",
+    "lerobot/aloha_static_candy",
+    "lerobot/aloha_static_coffee",
+    "lerobot/aloha_static_coffee_new",
+    "lerobot/aloha_static_cups_open",
+    "lerobot/aloha_static_fork_pick_up",
+    "lerobot/aloha_static_pingpong_test",
+    "lerobot/aloha_static_pro_pencil",
+    "lerobot/aloha_static_screw_driver",
+    "lerobot/aloha_static_tape",
+    "lerobot/aloha_static_thread_velcro",
+    "lerobot/aloha_static_towel",
+    "lerobot/aloha_static_vinh_cup",
+    "lerobot/aloha_static_vinh_cup_left",
+    "lerobot/aloha_static_ziploc_slide",
+    "lerobot/umi_cup_in_the_wild",
+    "lerobot/unitreeh1_fold_clothes",
+    "lerobot/unitreeh1_rearrange_objects",
+    "lerobot/unitreeh1_two_robot_greeting",
+    "lerobot/unitreeh1_warehouse",
+    "lerobot/nyu_rot_dataset",
+    "lerobot/utokyo_saytap",
+    "lerobot/imperialcollege_sawyer_wrist_cam",
+    "lerobot/utokyo_xarm_bimanual",
+    "lerobot/tokyo_u_lsmo",
+    "lerobot/utokyo_pr2_opening_fridge",
+    "lerobot/cmu_franka_exploration_dataset",
+    "lerobot/cmu_stretch",
+    "lerobot/asu_table_top",
+    "lerobot/utokyo_pr2_tabletop_manipulation",
+    "lerobot/utokyo_xarm_pick_and_place",
+    "lerobot/ucsd_kitchen_dataset",
+    "lerobot/austin_buds_dataset",
+    "lerobot/dlr_sara_grid_clamp",
+    "lerobot/conq_hose_manipulation",
+    "lerobot/columbia_cairlab_pusht_real",
+    "lerobot/dlr_sara_pour",
+    "lerobot/dlr_edan_shared_control",
+    "lerobot/ucsd_pick_and_place_dataset",
+    "lerobot/berkeley_cable_routing",
+    "lerobot/nyu_franka_play_dataset",
+    "lerobot/austin_sirius_dataset",
+    "lerobot/cmu_play_fusion",
+    "lerobot/berkeley_gnm_sac_son",
+    "lerobot/nyu_door_opening_surprising_effectiveness",
+    "lerobot/berkeley_fanuc_manipulation",
+    "lerobot/jaco_play",
+    "lerobot/viola",
+    "lerobot/kaist_nonprehensile",
+    "lerobot/berkeley_mvp",
+    "lerobot/uiuc_d3field",
+    "lerobot/berkeley_gnm_recon",
+    "lerobot/austin_sailor_dataset",
+    "lerobot/utaustin_mutex",
+    "lerobot/roboturk",
+    "lerobot/stanford_hydra_dataset",
+    "lerobot/berkeley_autolab_ur5",
+    "lerobot/stanford_robocook",
+    "lerobot/toto",
+    "lerobot/fmb",
+    "lerobot/droid_100",
+    "lerobot/berkeley_rpt",
+    "lerobot/stanford_kuka_multimodal_dataset",
+    "lerobot/iamlab_cmu_pickup_insert",
+    "lerobot/taco_play",
+    "lerobot/berkeley_gnm_cory_hall",
+    "lerobot/usc_cloth_sim",
+]
+
+available_datasets = sorted(
+    set(itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets))
+)
+
+# lists all available policies from `lerobot/common/policies`
+available_policies = [
+    "act",
+    "diffusion",
+    "tdmpc",
+    "vqbet",
+]
+
+# lists all available robots from `lerobot/common/robot_devices/robots`
+available_robots = [
+    "koch",
+    "koch_bimanual",
+    "aloha",
+    "so100",
+    "moss",
+]
+
+# lists all available cameras from `lerobot/common/robot_devices/cameras`
+available_cameras = [
+    "opencv",
+    "intelrealsense",
+]
+
+# lists all available motors from `lerobot/common/robot_devices/motors`
+available_motors = [
+    "dynamixel",
+    "feetech",
+]
+
+# keys and values refer to yaml files
+available_policies_per_env = {
+    "aloha": ["act"],
+    "pusht": ["diffusion", "vqbet"],
+    "xarm": ["tdmpc"],
+    "koch_real": ["act_koch_real"],
+    "aloha_real": ["act_aloha_real"],
+}
+
+env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]
+env_dataset_pairs = [
+    (env, dataset) for env, datasets in available_datasets_per_env.items() for dataset in datasets
+]
+env_dataset_policy_triplets = [
+    (env, dataset, policy)
+    for env, datasets in available_datasets_per_env.items()
+    for dataset in datasets
+    for policy in available_policies_per_env[env]
+]

lerobot/__version__.py

@@ -1 +1,23 @@
-__version__ = "0.0.0"
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""To enable `lerobot.__version__`"""
+
+from importlib.metadata import PackageNotFoundError, version
+
+try:
+    __version__ = version("lerobot")
+except PackageNotFoundError:
+    __version__ = "unknown"

lerobot/common/constants.py

@@ -0,0 +1,45 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# keys
+import os
+from pathlib import Path
+
+from huggingface_hub.constants import HF_HOME
+
+OBS_ENV = "observation.environment_state"
+OBS_ROBOT = "observation.state"
+OBS_IMAGE = "observation.image"
+OBS_IMAGES = "observation.images"
+ACTION = "action"
+
+# files & directories
+CHECKPOINTS_DIR = "checkpoints"
+LAST_CHECKPOINT_LINK = "last"
+PRETRAINED_MODEL_DIR = "pretrained_model"
+TRAINING_STATE_DIR = "training_state"
+RNG_STATE = "rng_state.safetensors"
+TRAINING_STEP = "training_step.json"
+OPTIMIZER_STATE = "optimizer_state.safetensors"
+OPTIMIZER_PARAM_GROUPS = "optimizer_param_groups.json"
+SCHEDULER_STATE = "scheduler_state.json"
+
+# cache dir
+default_cache_path = Path(HF_HOME) / "lerobot"
+HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
+
+if "LEROBOT_HOME" in os.environ:
+    raise ValueError(
+        f"You have a 'LEROBOT_HOME' environment variable set to '{os.getenv('LEROBOT_HOME')}'.\n"
+        "'LEROBOT_HOME' is deprecated, please use 'HF_LEROBOT_HOME' instead."
+    )

lerobot/common/datasets/abstract.py

@@ -1,158 +0,0 @@
-import abc
-import logging
-from pathlib import Path
-from typing import Callable
-
-import einops
-import torch
-import torchrl
-import tqdm
-from tensordict import TensorDict
-from torchrl.data.datasets.utils import _get_root_dir
-from torchrl.data.replay_buffers.replay_buffers import TensorDictReplayBuffer
-from torchrl.data.replay_buffers.samplers import SliceSampler
-from torchrl.data.replay_buffers.storages import TensorStorage, _collate_id
-from torchrl.data.replay_buffers.writers import ImmutableDatasetWriter, Writer
-
-
-class AbstractExperienceReplay(TensorDictReplayBuffer):
-    def __init__(
-        self,
-        dataset_id: str,
-        batch_size: int = None,
-        *,
-        shuffle: bool = True,
-        root: Path = None,
-        pin_memory: bool = False,
-        prefetch: int = None,
-        sampler: SliceSampler = None,
-        collate_fn: Callable = None,
-        writer: Writer = None,
-        transform: "torchrl.envs.Transform" = None,
-    ):
-        self.dataset_id = dataset_id
-        self.shuffle = shuffle
-        self.root = _get_root_dir(self.dataset_id) if root is None else root
-        self.root = Path(self.root)
-        self.data_dir = self.root / self.dataset_id
-
-        storage = self._download_or_load_storage()
-
-        super().__init__(
-            storage=storage,
-            sampler=sampler,
-            writer=ImmutableDatasetWriter() if writer is None else writer,
-            collate_fn=_collate_id if collate_fn is None else collate_fn,
-            pin_memory=pin_memory,
-            prefetch=prefetch,
-            batch_size=batch_size,
-            transform=transform,
-        )
-
-    @property
-    def stats_patterns(self) -> dict:
-        return {
-            ("observation", "state"): "b c -> 1 c",
-            ("observation", "image"): "b c h w -> 1 c 1 1",
-            ("action"): "b c -> 1 c",
-        }
-
-    @property
-    def image_keys(self) -> list:
-        return [("observation", "image")]
-
-    @property
-    def num_cameras(self) -> int:
-        return len(self.image_keys)
-
-    @property
-    def num_samples(self) -> int:
-        return len(self)
-
-    @property
-    def num_episodes(self) -> int:
-        return len(self._storage._storage["episode"].unique())
-
-    def set_transform(self, transform):
-        self.transform = transform
-
-    def compute_or_load_stats(self, num_batch=100, batch_size=32) -> TensorDict:
-        stats_path = self.data_dir / "stats.pth"
-        if stats_path.exists():
-            stats = torch.load(stats_path)
-        else:
-            logging.info(f"compute_stats and save to {stats_path}")
-            stats = self._compute_stats(num_batch, batch_size)
-            torch.save(stats, stats_path)
-        return stats
-
-    @abc.abstractmethod
-    def _download_and_preproc(self) -> torch.StorageBase:
-        raise NotImplementedError()
-
-    def _download_or_load_storage(self):
-        if not self._is_downloaded():
-            storage = self._download_and_preproc()
-        else:
-            storage = TensorStorage(TensorDict.load_memmap(self.data_dir))
-        return storage
-
-    def _is_downloaded(self) -> bool:
-        return self.data_dir.is_dir()
-
-    def _compute_stats(self, num_batch=100, batch_size=32):
-        rb = TensorDictReplayBuffer(
-            storage=self._storage,
-            batch_size=batch_size,
-            prefetch=True,
-        )
-
-        mean, std, max, min = {}, {}, {}, {}
-
-        # compute mean, min, max
-        for _ in tqdm.tqdm(range(num_batch)):
-            batch = rb.sample()
-            for key, pattern in self.stats_patterns.items():
-                batch[key] = batch[key].float()
-                if key not in mean:
-                    # first batch initialize mean, min, max
-                    mean[key] = einops.reduce(batch[key], pattern, "mean")
-                    max[key] = einops.reduce(batch[key], pattern, "max")
-                    min[key] = einops.reduce(batch[key], pattern, "min")
-                else:
-                    mean[key] += einops.reduce(batch[key], pattern, "mean")
-                    max[key] = torch.maximum(max[key], einops.reduce(batch[key], pattern, "max"))
-                    min[key] = torch.minimum(min[key], einops.reduce(batch[key], pattern, "min"))
-                batch = rb.sample()
-
-        for key in self.stats_patterns:
-            mean[key] /= num_batch
-
-        # compute std, min, max
-        for _ in tqdm.tqdm(range(num_batch)):
-            batch = rb.sample()
-            for key, pattern in self.stats_patterns.items():
-                batch[key] = batch[key].float()
-                batch_mean = einops.reduce(batch[key], pattern, "mean")
-                if key not in std:
-                    # first batch initialize std
-                    std[key] = (batch_mean - mean[key]) ** 2
-                else:
-                    std[key] += (batch_mean - mean[key]) ** 2
-                max[key] = torch.maximum(max[key], einops.reduce(batch[key], pattern, "max"))
-                min[key] = torch.minimum(min[key], einops.reduce(batch[key], pattern, "min"))
-
-        for key in self.stats_patterns:
-            std[key] = torch.sqrt(std[key] / num_batch)
-
-        stats = TensorDict({}, batch_size=[])
-        for key in self.stats_patterns:
-            stats[(*key, "mean")] = mean[key]
-            stats[(*key, "std")] = std[key]
-            stats[(*key, "max")] = max[key]
-            stats[(*key, "min")] = min[key]
-
-            if key[0] == "observation":
-                # use same stats for the next observations
-                stats[("next", *key)] = stats[key]
-        return stats

lerobot/common/datasets/aloha.py

@@ -1,185 +0,0 @@
-import logging
-from pathlib import Path
-from typing import Callable
-
-import einops
-import gdown
-import h5py
-import torch
-import torchrl
-import tqdm
-from tensordict import TensorDict
-from torchrl.data.replay_buffers.samplers import SliceSampler
-from torchrl.data.replay_buffers.storages import TensorStorage
-from torchrl.data.replay_buffers.writers import Writer
-
-from lerobot.common.datasets.abstract import AbstractExperienceReplay
-
-DATASET_IDS = [
-    "aloha_sim_insertion_human",
-    "aloha_sim_insertion_scripted",
-    "aloha_sim_transfer_cube_human",
-    "aloha_sim_transfer_cube_scripted",
-]
-
-FOLDER_URLS = {
-    "aloha_sim_insertion_human": "https://drive.google.com/drive/folders/1RgyD0JgTX30H4IM5XZn8I3zSV_mr8pyF",
-    "aloha_sim_insertion_scripted": "https://drive.google.com/drive/folders/1TsojQQSXtHEoGnqgJ3gmpPQR2DPLtS2N",
-    "aloha_sim_transfer_cube_human": "https://drive.google.com/drive/folders/1sc-E4QYW7A0o23m1u2VWNGVq5smAsfCo",
-    "aloha_sim_transfer_cube_scripted": "https://drive.google.com/drive/folders/1aRyoOhQwxhyt1J8XgEig4s6kzaw__LXj",
-}
-
-EP48_URLS = {
-    "aloha_sim_insertion_human": "https://drive.google.com/file/d/18Cudl6nikDtgRolea7je8iF_gGKzynOP/view?usp=drive_link",
-    "aloha_sim_insertion_scripted": "https://drive.google.com/file/d/1wfMSZ24oOh5KR_0aaP3Cnu_c4ZCveduB/view?usp=drive_link",
-    "aloha_sim_transfer_cube_human": "https://drive.google.com/file/d/18smMymtr8tIxaNUQ61gW6dG50pt3MvGq/view?usp=drive_link",
-    "aloha_sim_transfer_cube_scripted": "https://drive.google.com/file/d/1pnGIOd-E4-rhz2P3VxpknMKRZCoKt6eI/view?usp=drive_link",
-}
-
-EP49_URLS = {
-    "aloha_sim_insertion_human": "https://drive.google.com/file/d/1C1kZYyROzs-PrLc0SkDgUgMi4-L3lauE/view?usp=drive_link",
-    "aloha_sim_insertion_scripted": "https://drive.google.com/file/d/17EuCUWS6uCCr6yyNzpXdcdE-_TTNCKtf/view?usp=drive_link",
-    "aloha_sim_transfer_cube_human": "https://drive.google.com/file/d/1Nk7l53d9sJoGDBKAOnNrExX5nLacATc6/view?usp=drive_link",
-    "aloha_sim_transfer_cube_scripted": "https://drive.google.com/file/d/1GKReZHrXU73NMiC5zKCq_UtqPVtYq8eo/view?usp=drive_link",
-}
-
-NUM_EPISODES = {
-    "aloha_sim_insertion_human": 50,
-    "aloha_sim_insertion_scripted": 50,
-    "aloha_sim_transfer_cube_human": 50,
-    "aloha_sim_transfer_cube_scripted": 50,
-}
-
-EPISODE_LEN = {
-    "aloha_sim_insertion_human": 500,
-    "aloha_sim_insertion_scripted": 400,
-    "aloha_sim_transfer_cube_human": 400,
-    "aloha_sim_transfer_cube_scripted": 400,
-}
-
-CAMERAS = {
-    "aloha_sim_insertion_human": ["top"],
-    "aloha_sim_insertion_scripted": ["top"],
-    "aloha_sim_transfer_cube_human": ["top"],
-    "aloha_sim_transfer_cube_scripted": ["top"],
-}
-
-
-def download(data_dir, dataset_id):
-    assert dataset_id in DATASET_IDS
-    assert dataset_id in FOLDER_URLS
-    assert dataset_id in EP48_URLS
-    assert dataset_id in EP49_URLS
-
-    data_dir.mkdir(parents=True, exist_ok=True)
-
-    gdown.download_folder(FOLDER_URLS[dataset_id], output=data_dir)
-
-    # because of the 50 files limit per directory, two files episode 48 and 49 were missing
-    gdown.download(EP48_URLS[dataset_id], output=data_dir / "episode_48.hdf5", fuzzy=True)
-    gdown.download(EP49_URLS[dataset_id], output=data_dir / "episode_49.hdf5", fuzzy=True)
-
-
-class AlohaExperienceReplay(AbstractExperienceReplay):
-    def __init__(
-        self,
-        dataset_id: str,
-        batch_size: int = None,
-        *,
-        shuffle: bool = True,
-        root: Path = None,
-        pin_memory: bool = False,
-        prefetch: int = None,
-        sampler: SliceSampler = None,
-        collate_fn: Callable = None,
-        writer: Writer = None,
-        transform: "torchrl.envs.Transform" = None,
-    ):
-        assert dataset_id in DATASET_IDS
-
-        super().__init__(
-            dataset_id,
-            batch_size,
-            shuffle=shuffle,
-            root=root,
-            pin_memory=pin_memory,
-            prefetch=prefetch,
-            sampler=sampler,
-            collate_fn=collate_fn,
-            writer=writer,
-            transform=transform,
-        )
-
-    @property
-    def stats_patterns(self) -> dict:
-        d = {
-            ("observation", "state"): "b c -> 1 c",
-            ("action"): "b c -> 1 c",
-        }
-        for cam in CAMERAS[self.dataset_id]:
-            d[("observation", "image", cam)] = "b c h w -> 1 c 1 1"
-        return d
-
-    @property
-    def image_keys(self) -> list:
-        return [("observation", "image", cam) for cam in CAMERAS[self.dataset_id]]
-
-    # def _is_downloaded(self) -> bool:
-    #     return False
-
-    def _download_and_preproc(self):
-        raw_dir = self.data_dir.parent / f"{self.data_dir.name}_raw"
-        if not raw_dir.is_dir():
-            download(raw_dir, self.dataset_id)
-
-        total_num_frames = 0
-        logging.info("Compute total number of frames to initialize offline buffer")
-        for ep_id in range(NUM_EPISODES[self.dataset_id]):
-            ep_path = raw_dir / f"episode_{ep_id}.hdf5"
-            with h5py.File(ep_path, "r") as ep:
-                total_num_frames += ep["/action"].shape[0] - 1
-        logging.info(f"{total_num_frames=}")
-
-        logging.info("Initialize and feed offline buffer")
-        idxtd = 0
-        for ep_id in tqdm.tqdm(range(NUM_EPISODES[self.dataset_id])):
-            ep_path = raw_dir / f"episode_{ep_id}.hdf5"
-            with h5py.File(ep_path, "r") as ep:
-                ep_num_frames = ep["/action"].shape[0]
-
-                # last step of demonstration is considered done
-                done = torch.zeros(ep_num_frames, 1, dtype=torch.bool)
-                done[-1] = True
-
-                state = torch.from_numpy(ep["/observations/qpos"][:])
-                action = torch.from_numpy(ep["/action"][:])
-
-                ep_td = TensorDict(
-                    {
-                        ("observation", "state"): state[:-1],
-                        "action": action[:-1],
-                        "episode": torch.tensor([ep_id] * (ep_num_frames - 1)),
-                        "frame_id": torch.arange(0, ep_num_frames - 1, 1),
-                        ("next", "observation", "state"): state[1:],
-                        # TODO: compute reward and success
-                        # ("next", "reward"): reward[1:],
-                        ("next", "done"): done[1:],
-                        # ("next", "success"): success[1:],
-                    },
-                    batch_size=ep_num_frames - 1,
-                )
-
-                for cam in CAMERAS[self.dataset_id]:
-                    image = torch.from_numpy(ep[f"/observations/images/{cam}"][:])
-                    image = einops.rearrange(image, "b h w c -> b c h w").contiguous()
-                    ep_td["observation", "image", cam] = image[:-1]
-                    ep_td["next", "observation", "image", cam] = image[1:]
-
-                if ep_id == 0:
-                    # hack to initialize tensordict data structure to store episodes
-                    td_data = ep_td[0].expand(total_num_frames).memmap_like(self.data_dir)
-
-                td_data[idxtd : idxtd + len(ep_td)] = ep_td
-                idxtd = idxtd + len(ep_td)
-
-        return TensorStorage(td_data.lock_())

lerobot/common/datasets/backward_compatibility.py

@@ -0,0 +1,68 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import packaging.version
+
+V2_MESSAGE = """
+The dataset you requested ({repo_id}) is in {version} format.
+
+We introduced a new format since v2.0 which is not backward compatible with v1.x.
+Please, use our conversion script. Modify the following command with your own task description:
+```
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \\
+    --repo-id {repo_id} \\
+    --single-task "TASK DESCRIPTION."  # <---- /!\\ Replace TASK DESCRIPTION /!\\
+```
+
+A few examples to replace TASK DESCRIPTION: "Pick up the blue cube and place it into the bin.", "Insert the
+peg into the socket.", "Slide open the ziploc bag.", "Take the elevator to the 1st floor.", "Open the top
+cabinet, store the pot inside it then close the cabinet.", "Push the T-shaped block onto the T-shaped
+target.", "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.", "Fold the
+sweatshirt.", ...
+
+If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
+or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
+"""
+
+V21_MESSAGE = """
+The dataset you requested ({repo_id}) is in {version} format.
+While current version of LeRobot is backward-compatible with it, the version of your dataset still uses global
+stats instead of per-episode stats. Update your dataset stats to the new format using this command:
+```
+python lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py --repo-id={repo_id}
+```
+
+If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
+or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
+"""
+
+FUTURE_MESSAGE = """
+The dataset you requested ({repo_id}) is only available in {version} format.
+As we cannot ensure forward compatibility with it, please update your current version of lerobot.
+"""
+
+
+class CompatibilityError(Exception): ...
+
+
+class BackwardCompatibilityError(CompatibilityError):
+    def __init__(self, repo_id: str, version: packaging.version.Version):
+        message = V2_MESSAGE.format(repo_id=repo_id, version=version)
+        super().__init__(message)
+
+
+class ForwardCompatibilityError(CompatibilityError):
+    def __init__(self, repo_id: str, version: packaging.version.Version):
+        message = FUTURE_MESSAGE.format(repo_id=repo_id, version=version)
+        super().__init__(message)

lerobot/common/datasets/card_template.md

@@ -0,0 +1,27 @@
+---
+# For reference on dataset card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/datasetcard.md?plain=1
+# Doc / guide: https://huggingface.co/docs/hub/datasets-cards
+{{ card_data }}
+---
+
+This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
+
+## Dataset Description
+
+{{ dataset_description | default("", true) }}
+
+- **Homepage:** {{ url | default("[More Information Needed]", true)}}
+- **Paper:** {{ paper | default("[More Information Needed]", true)}}
+- **License:** {{ license | default("[More Information Needed]", true)}}
+
+## Dataset Structure
+
+{{ dataset_structure | default("[More Information Needed]", true)}}
+
+## Citation
+
+**BibTeX:**
+
+```bibtex
+{{ citation_bibtex | default("[More Information Needed]", true)}}
+```

lerobot/common/datasets/compute_stats.py

@@ -0,0 +1,176 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import numpy as np
+
+from lerobot.common.datasets.utils import load_image_as_numpy
+
+
+def estimate_num_samples(
+    dataset_len: int, min_num_samples: int = 100, max_num_samples: int = 10_000, power: float = 0.75
+) -> int:
+    """Heuristic to estimate the number of samples based on dataset size.
+    The power controls the sample growth relative to dataset size.
+    Lower the power for less number of samples.
+
+    For default arguments, we have:
+    - from 1 to ~500, num_samples=100
+    - at 1000, num_samples=177
+    - at 2000, num_samples=299
+    - at 5000, num_samples=594
+    - at 10000, num_samples=1000
+    - at 20000, num_samples=1681
+    """
+    if dataset_len < min_num_samples:
+        min_num_samples = dataset_len
+    return max(min_num_samples, min(int(dataset_len**power), max_num_samples))
+
+
+def sample_indices(data_len: int) -> list[int]:
+    num_samples = estimate_num_samples(data_len)
+    return np.round(np.linspace(0, data_len - 1, num_samples)).astype(int).tolist()
+
+
+def auto_downsample_height_width(img: np.ndarray, target_size: int = 150, max_size_threshold: int = 300):
+    _, height, width = img.shape
+
+    if max(width, height) < max_size_threshold:
+        # no downsampling needed
+        return img
+
+    downsample_factor = int(width / target_size) if width > height else int(height / target_size)
+    return img[:, ::downsample_factor, ::downsample_factor]
+
+
+def sample_images(image_paths: list[str]) -> np.ndarray:
+    sampled_indices = sample_indices(len(image_paths))
+
+    images = None
+    for i, idx in enumerate(sampled_indices):
+        path = image_paths[idx]
+        # we load as uint8 to reduce memory usage
+        img = load_image_as_numpy(path, dtype=np.uint8, channel_first=True)
+        img = auto_downsample_height_width(img)
+
+        if images is None:
+            images = np.empty((len(sampled_indices), *img.shape), dtype=np.uint8)
+
+        images[i] = img
+
+    return images
+
+
+def get_feature_stats(array: np.ndarray, axis: tuple, keepdims: bool) -> dict[str, np.ndarray]:
+    return {
+        "min": np.min(array, axis=axis, keepdims=keepdims),
+        "max": np.max(array, axis=axis, keepdims=keepdims),
+        "mean": np.mean(array, axis=axis, keepdims=keepdims),
+        "std": np.std(array, axis=axis, keepdims=keepdims),
+        "count": np.array([len(array)]),
+    }
+
+
+def compute_episode_stats(episode_data: dict[str, list[str] | np.ndarray], features: dict) -> dict:
+    ep_stats = {}
+    for key, data in episode_data.items():
+        if features[key]["dtype"] == "string":
+            continue  # HACK: we should receive np.arrays of strings
+        elif features[key]["dtype"] in ["image", "video"]:
+            ep_ft_array = sample_images(data)  # data is a list of image paths
+            axes_to_reduce = (0, 2, 3)  # keep channel dim
+            keepdims = True
+        else:
+            ep_ft_array = data  # data is already a np.ndarray
+            axes_to_reduce = 0  # compute stats over the first axis
+            keepdims = data.ndim == 1  # keep as np.array
+
+        ep_stats[key] = get_feature_stats(ep_ft_array, axis=axes_to_reduce, keepdims=keepdims)
+
+        # finally, we normalize and remove batch dim for images
+        if features[key]["dtype"] in ["image", "video"]:
+            ep_stats[key] = {
+                k: v if k == "count" else np.squeeze(v / 255.0, axis=0) for k, v in ep_stats[key].items()
+            }
+
+    return ep_stats
+
+
+def _assert_type_and_shape(stats_list: list[dict[str, dict]]):
+    for i in range(len(stats_list)):
+        for fkey in stats_list[i]:
+            for k, v in stats_list[i][fkey].items():
+                if not isinstance(v, np.ndarray):
+                    raise ValueError(
+                        f"Stats must be composed of numpy array, but key '{k}' of feature '{fkey}' is of type '{type(v)}' instead."
+                    )
+                if v.ndim == 0:
+                    raise ValueError("Number of dimensions must be at least 1, and is 0 instead.")
+                if k == "count" and v.shape != (1,):
+                    raise ValueError(f"Shape of 'count' must be (1), but is {v.shape} instead.")
+                if "image" in fkey and k != "count" and v.shape != (3, 1, 1):
+                    raise ValueError(f"Shape of '{k}' must be (3,1,1), but is {v.shape} instead.")
+
+
+def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
+    """Aggregates stats for a single feature."""
+    means = np.stack([s["mean"] for s in stats_ft_list])
+    variances = np.stack([s["std"] ** 2 for s in stats_ft_list])
+    counts = np.stack([s["count"] for s in stats_ft_list])
+    total_count = counts.sum(axis=0)
+
+    # Prepare weighted mean by matching number of dimensions
+    while counts.ndim < means.ndim:
+        counts = np.expand_dims(counts, axis=-1)
+
+    # Compute the weighted mean
+    weighted_means = means * counts
+    total_mean = weighted_means.sum(axis=0) / total_count
+
+    # Compute the variance using the parallel algorithm
+    delta_means = means - total_mean
+    weighted_variances = (variances + delta_means**2) * counts
+    total_variance = weighted_variances.sum(axis=0) / total_count
+
+    return {
+        "min": np.min(np.stack([s["min"] for s in stats_ft_list]), axis=0),
+        "max": np.max(np.stack([s["max"] for s in stats_ft_list]), axis=0),
+        "mean": total_mean,
+        "std": np.sqrt(total_variance),
+        "count": total_count,
+    }
+
+
+def aggregate_stats(stats_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
+    """Aggregate stats from multiple compute_stats outputs into a single set of stats.
+
+    The final stats will have the union of all data keys from each of the stats dicts.
+
+    For instance:
+    - new_min = min(min_dataset_0, min_dataset_1, ...)
+    - new_max = max(max_dataset_0, max_dataset_1, ...)
+    - new_mean = (mean of all data, weighted by counts)
+    - new_std = (std of all data)
+    """
+
+    _assert_type_and_shape(stats_list)
+
+    data_keys = {key for stats in stats_list for key in stats}
+    aggregated_stats = {key: {} for key in data_keys}
+
+    for key in data_keys:
+        stats_with_key = [stats[key] for stats in stats_list if key in stats]
+        aggregated_stats[key] = aggregate_feature_stats(stats_with_key)
+
+    return aggregated_stats

lerobot/common/datasets/factory.py

@@ -1,114 +1,118 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 import logging
-import os
-from pathlib import Path
+from pprint import pformat
 
 import torch
-from torchrl.data.replay_buffers import PrioritizedSliceSampler, SliceSampler
 
-from lerobot.common.envs.transforms import NormalizeTransform
+from lerobot.common.datasets.lerobot_dataset import (
+    LeRobotDataset,
+    LeRobotDatasetMetadata,
+    MultiLeRobotDataset,
+)
+from lerobot.common.datasets.transforms import ImageTransforms
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.train import TrainPipelineConfig
 
-DATA_DIR = Path(os.environ.get("DATA_DIR", "data"))
+IMAGENET_STATS = {
+    "mean": [[[0.485]], [[0.456]], [[0.406]]],  # (c,1,1)
+    "std": [[[0.229]], [[0.224]], [[0.225]]],  # (c,1,1)
+}
 
 
-def make_offline_buffer(
-    cfg, overwrite_sampler=None, normalize=True, overwrite_batch_size=None, overwrite_prefetch=None
-):
-    if cfg.policy.balanced_sampling:
-        assert cfg.online_steps > 0
-        batch_size = None
-        pin_memory = False
-        prefetch = None
-    else:
-        assert cfg.online_steps == 0
-        num_slices = cfg.policy.batch_size
-        batch_size = cfg.policy.horizon * num_slices
-        pin_memory = cfg.device == "cuda"
-        prefetch = cfg.prefetch
+def resolve_delta_timestamps(
+    cfg: PreTrainedConfig, ds_meta: LeRobotDatasetMetadata
+) -> dict[str, list] | None:
+    """Resolves delta_timestamps by reading from the 'delta_indices' properties of the PreTrainedConfig.
 
-    if overwrite_batch_size is not None:
-        batch_size = overwrite_batch_size
+    Args:
+        cfg (PreTrainedConfig): The PreTrainedConfig to read delta_indices from.
+        ds_meta (LeRobotDatasetMetadata): The dataset from which features and fps are used to build
+            delta_timestamps against.
 
-    if overwrite_prefetch is not None:
-        prefetch = overwrite_prefetch
+    Returns:
+        dict[str, list] | None: A dictionary of delta_timestamps, e.g.:
+            {
+                "observation.state": [-0.04, -0.02, 0]
+                "observation.action": [-0.02, 0, 0.02]
+            }
+            returns `None` if the the resulting dict is empty.
+    """
+    delta_timestamps = {}
+    for key in ds_meta.features:
+        if key == "next.reward" and cfg.reward_delta_indices is not None:
+            delta_timestamps[key] = [i / ds_meta.fps for i in cfg.reward_delta_indices]
+        if key == "action" and cfg.action_delta_indices is not None:
+            delta_timestamps[key] = [i / ds_meta.fps for i in cfg.action_delta_indices]
+        if key.startswith("observation.") and cfg.observation_delta_indices is not None:
+            delta_timestamps[key] = [i / ds_meta.fps for i in cfg.observation_delta_indices]
 
-    if overwrite_sampler is None:
-        # TODO(rcadene): move batch_size outside
-        num_traj_per_batch = cfg.policy.batch_size  # // cfg.horizon
-        # TODO(rcadene): Sampler outputs a batch_size <= cfg.batch_size.
-        # We would need to add a transform to pad the tensordict to ensure batch_size == cfg.batch_size.
+    if len(delta_timestamps) == 0:
+        delta_timestamps = None
 
-        if cfg.offline_prioritized_sampler:
-            logging.info("use prioritized sampler for offline dataset")
-            sampler = PrioritizedSliceSampler(
-                max_capacity=100_000,
-                alpha=cfg.policy.per_alpha,
-                beta=cfg.policy.per_beta,
-                num_slices=num_traj_per_batch,
-                strict_length=False,
-            )
-        else:
-            logging.info("use simple sampler for offline dataset")
-            sampler = SliceSampler(
-                num_slices=num_traj_per_batch,
-                strict_length=False,
-            )
-    else:
-        sampler = overwrite_sampler
+    return delta_timestamps
 
-    if cfg.env.name == "simxarm":
-        from lerobot.common.datasets.simxarm import SimxarmExperienceReplay
 
-        clsfunc = SimxarmExperienceReplay
-        dataset_id = f"xarm_{cfg.env.task}_medium"
+def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDataset:
+    """Handles the logic of setting up delta timestamps and image transforms before creating a dataset.
 
-    elif cfg.env.name == "pusht":
-        from lerobot.common.datasets.pusht import PushtExperienceReplay
+    Args:
+        cfg (TrainPipelineConfig): A TrainPipelineConfig config which contains a DatasetConfig and a PreTrainedConfig.
 
-        clsfunc = PushtExperienceReplay
-        dataset_id = "pusht"
+    Raises:
+        NotImplementedError: The MultiLeRobotDataset is currently deactivated.
 
-    elif cfg.env.name == "aloha":
-        from lerobot.common.datasets.aloha import AlohaExperienceReplay
-
-        clsfunc = AlohaExperienceReplay
-        dataset_id = f"aloha_{cfg.env.task}"
-    else:
-        raise ValueError(cfg.env.name)
-
-    offline_buffer = clsfunc(
-        dataset_id=dataset_id,
-        root=DATA_DIR,
-        sampler=sampler,
-        batch_size=batch_size,
-        pin_memory=pin_memory,
-        prefetch=prefetch if isinstance(prefetch, int) else None,
+    Returns:
+        LeRobotDataset | MultiLeRobotDataset
+    """
+    image_transforms = (
+        ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None
     )
 
-    if normalize:
-        # TODO(rcadene): make normalization strategy configurable between mean_std, min_max, manual_min_max, min_max_from_spec
-        stats = offline_buffer.compute_or_load_stats()
-        in_keys = [("observation", "state"), ("action")]
+    if isinstance(cfg.dataset.repo_id, str):
+        ds_meta = LeRobotDatasetMetadata(
+            cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
+        )
+        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
+        dataset = LeRobotDataset(
+            cfg.dataset.repo_id,
+            root=cfg.dataset.root,
+            episodes=cfg.dataset.episodes,
+            delta_timestamps=delta_timestamps,
+            image_transforms=image_transforms,
+            revision=cfg.dataset.revision,
+            video_backend=cfg.dataset.video_backend,
+        )
+    else:
+        raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")
+        dataset = MultiLeRobotDataset(
+            cfg.dataset.repo_id,
+            # TODO(aliberts): add proper support for multi dataset
+            # delta_timestamps=delta_timestamps,
+            image_transforms=image_transforms,
+            video_backend=cfg.dataset.video_backend,
+        )
+        logging.info(
+            "Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
+            f"{pformat(dataset.repo_id_to_index, indent=2)}"
+        )
 
-        if cfg.policy == "tdmpc":
-            for key in offline_buffer.image_keys:
-                # TODO(rcadene): imagenet normalization is applied inside diffusion policy, but no normalization inside tdmpc
-                in_keys.append(key)
-                # since we use next observations in tdmpc
-                in_keys.append(("next", *key))
-            in_keys.append(("next", "observation", "state"))
+    if cfg.dataset.use_imagenet_stats:
+        for key in dataset.meta.camera_keys:
+            for stats_type, stats in IMAGENET_STATS.items():
+                dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
-        if cfg.policy == "diffusion" and cfg.env.name == "pusht":
-            # TODO(rcadene): we overwrite stats to have the same as pretrained model, but we should remove this
-            stats["observation", "state", "min"] = torch.tensor([13.456424, 32.938293], dtype=torch.float32)
-            stats["observation", "state", "max"] = torch.tensor([496.14618, 510.9579], dtype=torch.float32)
-            stats["action", "min"] = torch.tensor([12.0, 25.0], dtype=torch.float32)
-            stats["action", "max"] = torch.tensor([511.0, 511.0], dtype=torch.float32)
-
-        transform = NormalizeTransform(stats, in_keys, mode="min_max")
-        offline_buffer.set_transform(transform)
-
-    if not overwrite_sampler:
-        index = torch.arange(0, offline_buffer.num_samples, 1)
-        sampler.extend(index)
-
-    return offline_buffer
+    return dataset

lerobot/common/datasets/image_writer.py

@@ -0,0 +1,178 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import multiprocessing
+import queue
+import threading
+from pathlib import Path
+
+import numpy as np
+import PIL.Image
+import torch
+
+
+def safe_stop_image_writer(func):
+    def wrapper(*args, **kwargs):
+        try:
+            return func(*args, **kwargs)
+        except Exception as e:
+            dataset = kwargs.get("dataset")
+            image_writer = getattr(dataset, "image_writer", None) if dataset else None
+            if image_writer is not None:
+                print("Waiting for image writer to terminate...")
+                image_writer.stop()
+            raise e
+
+    return wrapper
+
+
+def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True) -> PIL.Image.Image:
+    # TODO(aliberts): handle 1 channel and 4 for depth images
+    if image_array.ndim != 3:
+        raise ValueError(f"The array has {image_array.ndim} dimensions, but 3 is expected for an image.")
+
+    if image_array.shape[0] == 3:
+        # Transpose from pytorch convention (C, H, W) to (H, W, C)
+        image_array = image_array.transpose(1, 2, 0)
+
+    elif image_array.shape[-1] != 3:
+        raise NotImplementedError(
+            f"The image has {image_array.shape[-1]} channels, but 3 is required for now."
+        )
+
+    if image_array.dtype != np.uint8:
+        if range_check:
+            max_ = image_array.max().item()
+            min_ = image_array.min().item()
+            if max_ > 1.0 or min_ < 0.0:
+                raise ValueError(
+                    "The image data type is float, which requires values in the range [0.0, 1.0]. "
+                    f"However, the provided range is [{min_}, {max_}]. Please adjust the range or "
+                    "provide a uint8 image with values in the range [0, 255]."
+                )
+
+        image_array = (image_array * 255).astype(np.uint8)
+
+    return PIL.Image.fromarray(image_array)
+
+
+def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path):
+    try:
+        if isinstance(image, np.ndarray):
+            img = image_array_to_pil_image(image)
+        elif isinstance(image, PIL.Image.Image):
+            img = image
+        else:
+            raise TypeError(f"Unsupported image type: {type(image)}")
+        img.save(fpath)
+    except Exception as e:
+        print(f"Error writing image {fpath}: {e}")
+
+
+def worker_thread_loop(queue: queue.Queue):
+    while True:
+        item = queue.get()
+        if item is None:
+            queue.task_done()
+            break
+        image_array, fpath = item
+        write_image(image_array, fpath)
+        queue.task_done()
+
+
+def worker_process(queue: queue.Queue, num_threads: int):
+    threads = []
+    for _ in range(num_threads):
+        t = threading.Thread(target=worker_thread_loop, args=(queue,))
+        t.daemon = True
+        t.start()
+        threads.append(t)
+    for t in threads:
+        t.join()
+
+
+class AsyncImageWriter:
+    """
+    This class abstract away the initialisation of processes or/and threads to
+    save images on disk asynchrounously, which is critical to control a robot and record data
+    at a high frame rate.
+
+    When `num_processes=0`, it creates a threads pool of size `num_threads`.
+    When `num_processes>0`, it creates processes pool of size `num_processes`, where each subprocess starts
+    their own threads pool of size `num_threads`.
+
+    The optimal number of processes and threads depends on your computer capabilities.
+    We advise to use 4 threads per camera with 0 processes. If the fps is not stable, try to increase or lower
+    the number of threads. If it is still not stable, try to use 1 subprocess, or more.
+    """
+
+    def __init__(self, num_processes: int = 0, num_threads: int = 1):
+        self.num_processes = num_processes
+        self.num_threads = num_threads
+        self.queue = None
+        self.threads = []
+        self.processes = []
+        self._stopped = False
+
+        if num_threads <= 0 and num_processes <= 0:
+            raise ValueError("Number of threads and processes must be greater than zero.")
+
+        if self.num_processes == 0:
+            # Use threading
+            self.queue = queue.Queue()
+            for _ in range(self.num_threads):
+                t = threading.Thread(target=worker_thread_loop, args=(self.queue,))
+                t.daemon = True
+                t.start()
+                self.threads.append(t)
+        else:
+            # Use multiprocessing
+            self.queue = multiprocessing.JoinableQueue()
+            for _ in range(self.num_processes):
+                p = multiprocessing.Process(target=worker_process, args=(self.queue, self.num_threads))
+                p.daemon = True
+                p.start()
+                self.processes.append(p)
+
+    def save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path):
+        if isinstance(image, torch.Tensor):
+            # Convert tensor to numpy array to minimize main process time
+            image = image.cpu().numpy()
+        self.queue.put((image, fpath))
+
+    def wait_until_done(self):
+        self.queue.join()
+
+    def stop(self):
+        if self._stopped:
+            return
+
+        if self.num_processes == 0:
+            for _ in self.threads:
+                self.queue.put(None)
+            for t in self.threads:
+                t.join()
+        else:
+            num_nones = self.num_processes * self.num_threads
+            for _ in range(num_nones):
+                self.queue.put(None)
+            for p in self.processes:
+                p.join()
+                if p.is_alive():
+                    p.terminate()
+            self.queue.close()
+            self.queue.join_thread()
+
+        self._stopped = True

lerobot/common/datasets/online_buffer.py

@@ -0,0 +1,384 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""An online buffer for the online training loop in train.py
+
+Note to maintainers: This duplicates some logic from LeRobotDataset and EpisodeAwareSampler. We should
+consider converging to one approach. Here we have opted to use numpy.memmap to back the data buffer. It's much
+faster than using HuggingFace Datasets as there's no conversion to an intermediate non-python object. Also it
+supports in-place slicing and mutation which is very handy for a dynamic buffer.
+"""
+
+import os
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+
+def _make_memmap_safe(**kwargs) -> np.memmap:
+    """Make a numpy memmap with checks on available disk space first.
+
+    Expected kwargs are: "filename", "dtype" (must by np.dtype), "mode" and "shape"
+
+    For information on dtypes:
+    https://numpy.org/doc/stable/reference/arrays.dtypes.html#arrays-dtypes-constructing
+    """
+    if kwargs["mode"].startswith("w"):
+        required_space = kwargs["dtype"].itemsize * np.prod(kwargs["shape"])  # bytes
+        stats = os.statvfs(Path(kwargs["filename"]).parent)
+        available_space = stats.f_bavail * stats.f_frsize  # bytes
+        if required_space >= available_space * 0.8:
+            raise RuntimeError(
+                f"You're about to take up {required_space} of {available_space} bytes available."
+            )
+    return np.memmap(**kwargs)
+
+
+class OnlineBuffer(torch.utils.data.Dataset):
+    """FIFO data buffer for the online training loop in train.py.
+
+    Follows the protocol of LeRobotDataset as much as is required to have it be used by the online training
+    loop in the same way that a LeRobotDataset would be used.
+
+    The underlying data structure will have data inserted in a circular fashion. Always insert after the
+    last index, and when you reach the end, wrap around to the start.
+
+    The data is stored in a numpy memmap.
+    """
+
+    NEXT_INDEX_KEY = "_next_index"
+    OCCUPANCY_MASK_KEY = "_occupancy_mask"
+    INDEX_KEY = "index"
+    FRAME_INDEX_KEY = "frame_index"
+    EPISODE_INDEX_KEY = "episode_index"
+    TIMESTAMP_KEY = "timestamp"
+    IS_PAD_POSTFIX = "_is_pad"
+
+    def __init__(
+        self,
+        write_dir: str | Path,
+        data_spec: dict[str, Any] | None,
+        buffer_capacity: int | None,
+        fps: float | None = None,
+        delta_timestamps: dict[str, list[float]] | dict[str, np.ndarray] | None = None,
+    ):
+        """
+        The online buffer can be provided from scratch or you can load an existing online buffer by passing
+        a `write_dir` associated with an existing buffer.
+
+        Args:
+            write_dir: Where to keep the numpy memmap files. One memmap file will be stored for each data key.
+                Note that if the files already exist, they are opened in read-write mode (used for training
+                resumption.)
+            data_spec: A mapping from data key to data specification, like {data_key: {"shape": tuple[int],
+                "dtype": np.dtype}}. This should include all the data that you wish to record into the buffer,
+                but note that "index", "frame_index" and "episode_index" are already accounted for by this
+                class, so you don't need to include them.
+            buffer_capacity: How many frames should be stored in the buffer as a maximum. Be aware of your
+                system's available disk space when choosing this.
+            fps: Same as the fps concept in LeRobot dataset. Here it needs to be provided for the
+                 delta_timestamps logic. You can pass None if you are not using delta_timestamps.
+            delta_timestamps: Same as the delta_timestamps concept in LeRobotDataset. This is internally
+                converted to dict[str, np.ndarray] for optimization purposes.
+
+        """
+        self.set_delta_timestamps(delta_timestamps)
+        self._fps = fps
+        # Tolerance in seconds used to discard loaded frames when their timestamps are not close enough from
+        # the requested frames. It is only used when `delta_timestamps` is provided.
+        # minus 1e-4 to account for possible numerical error
+        self.tolerance_s = 1 / self.fps - 1e-4 if fps is not None else None
+        self._buffer_capacity = buffer_capacity
+        data_spec = self._make_data_spec(data_spec, buffer_capacity)
+        Path(write_dir).mkdir(parents=True, exist_ok=True)
+        self._data = {}
+        for k, v in data_spec.items():
+            self._data[k] = _make_memmap_safe(
+                filename=Path(write_dir) / k,
+                dtype=v["dtype"] if v is not None else None,
+                mode="r+" if (Path(write_dir) / k).exists() else "w+",
+                shape=tuple(v["shape"]) if v is not None else None,
+            )
+
+    @property
+    def delta_timestamps(self) -> dict[str, np.ndarray] | None:
+        return self._delta_timestamps
+
+    def set_delta_timestamps(self, value: dict[str, list[float]] | None):
+        """Set delta_timestamps converting the values to numpy arrays.
+
+        The conversion is for an optimization in the __getitem__. The loop is much slower if the arrays
+        need to be converted into numpy arrays.
+        """
+        if value is not None:
+            self._delta_timestamps = {k: np.array(v) for k, v in value.items()}
+        else:
+            self._delta_timestamps = None
+
+    def _make_data_spec(self, data_spec: dict[str, Any], buffer_capacity: int) -> dict[str, dict[str, Any]]:
+        """Makes the data spec for np.memmap."""
+        if any(k.startswith("_") for k in data_spec):
+            raise ValueError(
+                "data_spec keys should not start with '_'. This prefix is reserved for internal logic."
+            )
+        preset_keys = {
+            OnlineBuffer.INDEX_KEY,
+            OnlineBuffer.FRAME_INDEX_KEY,
+            OnlineBuffer.EPISODE_INDEX_KEY,
+            OnlineBuffer.TIMESTAMP_KEY,
+        }
+        if len(intersection := set(data_spec).intersection(preset_keys)) > 0:
+            raise ValueError(
+                f"data_spec should not contain any of {preset_keys} as these are handled internally. "
+                f"The provided data_spec has {intersection}."
+            )
+        complete_data_spec = {
+            # _next_index will be a pointer to the next index that we should start filling from when we add
+            # more data.
+            OnlineBuffer.NEXT_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": ()},
+            # Since the memmap is initialized with all-zeros, this keeps track of which indices are occupied
+            # with real data rather than the dummy initialization.
+            OnlineBuffer.OCCUPANCY_MASK_KEY: {"dtype": np.dtype("?"), "shape": (buffer_capacity,)},
+            OnlineBuffer.INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.FRAME_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.EPISODE_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.TIMESTAMP_KEY: {"dtype": np.dtype("float64"), "shape": (buffer_capacity,)},
+        }
+        for k, v in data_spec.items():
+            complete_data_spec[k] = {"dtype": v["dtype"], "shape": (buffer_capacity, *v["shape"])}
+        return complete_data_spec
+
+    def add_data(self, data: dict[str, np.ndarray]):
+        """Add new data to the buffer, which could potentially mean shifting old data out.
+
+        The new data should contain all the frames (in order) of any number of episodes. The indices should
+        start from 0 (note to the developer: this can easily be generalized). See the `rollout` and
+        `eval_policy` functions in `eval.py` for more information on how the data is constructed.
+
+        Shift the incoming data index and episode_index to continue on from the last frame. Note that this
+        will be done in place!
+        """
+        if len(missing_keys := (set(self.data_keys).difference(set(data)))) > 0:
+            raise ValueError(f"Missing data keys: {missing_keys}")
+        new_data_length = len(data[self.data_keys[0]])
+        if not all(len(data[k]) == new_data_length for k in self.data_keys):
+            raise ValueError("All data items should have the same length")
+
+        next_index = self._data[OnlineBuffer.NEXT_INDEX_KEY]
+
+        # Sanity check to make sure that the new data indices start from 0.
+        assert data[OnlineBuffer.EPISODE_INDEX_KEY][0].item() == 0
+        assert data[OnlineBuffer.INDEX_KEY][0].item() == 0
+
+        # Shift the incoming indices if necessary.
+        if self.num_frames > 0:
+            last_episode_index = self._data[OnlineBuffer.EPISODE_INDEX_KEY][next_index - 1]
+            last_data_index = self._data[OnlineBuffer.INDEX_KEY][next_index - 1]
+            data[OnlineBuffer.EPISODE_INDEX_KEY] += last_episode_index + 1
+            data[OnlineBuffer.INDEX_KEY] += last_data_index + 1
+
+        # Insert the new data starting from next_index. It may be necessary to wrap around to the start.
+        n_surplus = max(0, new_data_length - (self._buffer_capacity - next_index))
+        for k in self.data_keys:
+            if n_surplus == 0:
+                slc = slice(next_index, next_index + new_data_length)
+                self._data[k][slc] = data[k]
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][slc] = True
+            else:
+                self._data[k][next_index:] = data[k][:-n_surplus]
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][next_index:] = True
+                self._data[k][:n_surplus] = data[k][-n_surplus:]
+        if n_surplus == 0:
+            self._data[OnlineBuffer.NEXT_INDEX_KEY] = next_index + new_data_length
+        else:
+            self._data[OnlineBuffer.NEXT_INDEX_KEY] = n_surplus
+
+    @property
+    def data_keys(self) -> list[str]:
+        keys = set(self._data)
+        keys.remove(OnlineBuffer.OCCUPANCY_MASK_KEY)
+        keys.remove(OnlineBuffer.NEXT_INDEX_KEY)
+        return sorted(keys)
+
+    @property
+    def fps(self) -> float | None:
+        return self._fps
+
+    @property
+    def num_episodes(self) -> int:
+        return len(
+            np.unique(self._data[OnlineBuffer.EPISODE_INDEX_KEY][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
+        )
+
+    @property
+    def num_frames(self) -> int:
+        return np.count_nonzero(self._data[OnlineBuffer.OCCUPANCY_MASK_KEY])
+
+    def __len__(self):
+        return self.num_frames
+
+    def _item_to_tensors(self, item: dict) -> dict:
+        item_ = {}
+        for k, v in item.items():
+            if isinstance(v, torch.Tensor):
+                item_[k] = v
+            elif isinstance(v, np.ndarray):
+                item_[k] = torch.from_numpy(v)
+            else:
+                item_[k] = torch.tensor(v)
+        return item_
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        if idx >= len(self) or idx < -len(self):
+            raise IndexError
+
+        item = {k: v[idx] for k, v in self._data.items() if not k.startswith("_")}
+
+        if self.delta_timestamps is None:
+            return self._item_to_tensors(item)
+
+        episode_index = item[OnlineBuffer.EPISODE_INDEX_KEY]
+        current_ts = item[OnlineBuffer.TIMESTAMP_KEY]
+        episode_data_indices = np.where(
+            np.bitwise_and(
+                self._data[OnlineBuffer.EPISODE_INDEX_KEY] == episode_index,
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY],
+            )
+        )[0]
+        episode_timestamps = self._data[OnlineBuffer.TIMESTAMP_KEY][episode_data_indices]
+
+        for data_key in self.delta_timestamps:
+            # Note: The logic in this loop is copied from `load_previous_and_future_frames`.
+            # Get timestamps used as query to retrieve data of previous/future frames.
+            query_ts = current_ts + self.delta_timestamps[data_key]
+
+            # Compute distances between each query timestamp and all timestamps of all the frames belonging to
+            # the episode.
+            dist = np.abs(query_ts[:, None] - episode_timestamps[None, :])
+            argmin_ = np.argmin(dist, axis=1)
+            min_ = dist[np.arange(dist.shape[0]), argmin_]
+
+            is_pad = min_ > self.tolerance_s
+
+            # Check violated query timestamps are all outside the episode range.
+            assert (
+                (query_ts[is_pad] < episode_timestamps[0]) | (episode_timestamps[-1] < query_ts[is_pad])
+            ).all(), (
+                f"One or several timestamps unexpectedly violate the tolerance ({min_} > {self.tolerance_s=}"
+                ") inside the episode range."
+            )
+
+            # Load frames for this data key.
+            item[data_key] = self._data[data_key][episode_data_indices[argmin_]]
+
+            item[f"{data_key}{OnlineBuffer.IS_PAD_POSTFIX}"] = is_pad
+
+        return self._item_to_tensors(item)
+
+    def get_data_by_key(self, key: str) -> torch.Tensor:
+        """Returns all data for a given data key as a Tensor."""
+        return torch.from_numpy(self._data[key][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
+
+
+def compute_sampler_weights(
+    offline_dataset: LeRobotDataset,
+    offline_drop_n_last_frames: int = 0,
+    online_dataset: OnlineBuffer | None = None,
+    online_sampling_ratio: float | None = None,
+    online_drop_n_last_frames: int = 0,
+) -> torch.Tensor:
+    """Compute the sampling weights for the online training dataloader in train.py.
+
+    Args:
+        offline_dataset: The LeRobotDataset used for offline pre-training.
+        online_drop_n_last_frames: Number of frames to drop from the end of each offline dataset episode.
+        online_dataset: The OnlineBuffer used in online training.
+        online_sampling_ratio: The proportion of data that should be sampled from the online dataset. If an
+            online dataset is provided, this value must also be provided.
+        online_drop_n_first_frames: See `offline_drop_n_last_frames`. This is the same, but for the online
+            dataset.
+    Returns:
+        Tensor of weights for [offline_dataset; online_dataset], normalized to 1.
+
+    Notes to maintainers:
+        - This duplicates some logic from EpisodeAwareSampler. We should consider converging to one approach.
+        - When used with `torch.utils.data.WeightedRandomSampler`, it could completely replace
+          `EpisodeAwareSampler` as the online dataset related arguments are optional. The only missing feature
+          is the ability to turn shuffling off.
+        - Options `drop_first_n_frames` and `episode_indices_to_use` can be added easily. They were not
+          included here to avoid adding complexity.
+    """
+    if len(offline_dataset) == 0 and (online_dataset is None or len(online_dataset) == 0):
+        raise ValueError("At least one of `offline_dataset` or `online_dataset` should be contain data.")
+    if (online_dataset is None) ^ (online_sampling_ratio is None):
+        raise ValueError(
+            "`online_dataset` and `online_sampling_ratio` must be provided together or not at all."
+        )
+    offline_sampling_ratio = 0 if online_sampling_ratio is None else 1 - online_sampling_ratio
+
+    weights = []
+
+    if len(offline_dataset) > 0:
+        offline_data_mask_indices = []
+        for start_index, end_index in zip(
+            offline_dataset.episode_data_index["from"],
+            offline_dataset.episode_data_index["to"],
+            strict=True,
+        ):
+            offline_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - offline_drop_n_last_frames)
+            )
+        offline_data_mask = torch.zeros(len(offline_dataset), dtype=torch.bool)
+        offline_data_mask[torch.tensor(offline_data_mask_indices)] = True
+        weights.append(
+            torch.full(
+                size=(len(offline_dataset),),
+                fill_value=offline_sampling_ratio / offline_data_mask.sum(),
+            )
+            * offline_data_mask
+        )
+
+    if online_dataset is not None and len(online_dataset) > 0:
+        online_data_mask_indices = []
+        episode_indices = online_dataset.get_data_by_key("episode_index")
+        for episode_idx in torch.unique(episode_indices):
+            where_episode = torch.where(episode_indices == episode_idx)
+            start_index = where_episode[0][0]
+            end_index = where_episode[0][-1] + 1
+            online_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - online_drop_n_last_frames)
+            )
+        online_data_mask = torch.zeros(len(online_dataset), dtype=torch.bool)
+        online_data_mask[torch.tensor(online_data_mask_indices)] = True
+        weights.append(
+            torch.full(
+                size=(len(online_dataset),),
+                fill_value=online_sampling_ratio / online_data_mask.sum(),
+            )
+            * online_data_mask
+        )
+
+    weights = torch.cat(weights)
+
+    if weights.sum() == 0:
+        weights += 1 / len(weights)
+    else:
+        weights /= weights.sum()
+
+    return weights

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -0,0 +1,131 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import inspect
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from typing import Dict
+
+import datasets
+import numpy
+import PIL
+import torch
+
+from lerobot.common.datasets.video_utils import encode_video_frames
+
+
+def concatenate_episodes(ep_dicts):
+    data_dict = {}
+
+    keys = ep_dicts[0].keys()
+    for key in keys:
+        if torch.is_tensor(ep_dicts[0][key][0]):
+            data_dict[key] = torch.cat([ep_dict[key] for ep_dict in ep_dicts])
+        else:
+            if key not in data_dict:
+                data_dict[key] = []
+            for ep_dict in ep_dicts:
+                for x in ep_dict[key]:
+                    data_dict[key].append(x)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def save_images_concurrently(imgs_array: numpy.array, out_dir: Path, max_workers: int = 4):
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    def save_image(img_array, i, out_dir):
+        img = PIL.Image.fromarray(img_array)
+        img.save(str(out_dir / f"frame_{i:06d}.png"), quality=100)
+
+    num_images = len(imgs_array)
+    with ThreadPoolExecutor(max_workers=max_workers) as executor:
+        [executor.submit(save_image, imgs_array[i], i, out_dir) for i in range(num_images)]
+
+
+def get_default_encoding() -> dict:
+    """Returns the default ffmpeg encoding parameters used by `encode_video_frames`."""
+    signature = inspect.signature(encode_video_frames)
+    return {
+        k: v.default
+        for k, v in signature.parameters.items()
+        if v.default is not inspect.Parameter.empty and k in ["vcodec", "pix_fmt", "g", "crf"]
+    }
+
+
+def check_repo_id(repo_id: str) -> None:
+    if len(repo_id.split("/")) != 2:
+        raise ValueError(
+            f"""`repo_id` is expected to contain a community or user id `/` the name of the dataset
+            (e.g. 'lerobot/pusht'), but contains '{repo_id}'."""
+        )
+
+
+# TODO(aliberts): remove
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
+    """
+    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
+
+    Parameters:
+    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
+
+    Returns:
+    - episode_data_index: A dictionary containing the data index for each episode. The dictionary has two keys:
+        - "from": A tensor containing the starting index of each episode.
+        - "to": A tensor containing the ending index of each episode.
+    """
+    episode_data_index = {"from": [], "to": []}
+
+    current_episode = None
+    """
+    The episode_index is a list of integers, each representing the episode index of the corresponding example.
+    For instance, the following is a valid episode_index:
+      [0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2]
+
+    Below, we iterate through the episode_index and populate the episode_data_index dictionary with the starting and
+    ending index of each episode. For the episode_index above, the episode_data_index dictionary will look like this:
+        {
+            "from": [0, 3, 7],
+            "to": [3, 7, 12]
+        }
+    """
+    if len(hf_dataset) == 0:
+        episode_data_index = {
+            "from": torch.tensor([]),
+            "to": torch.tensor([]),
+        }
+        return episode_data_index
+    for idx, episode_idx in enumerate(hf_dataset["episode_index"]):
+        if episode_idx != current_episode:
+            # We encountered a new episode, so we append its starting location to the "from" list
+            episode_data_index["from"].append(idx)
+            # If this is not the first episode, we append the ending location of the previous episode to the "to" list
+            if current_episode is not None:
+                episode_data_index["to"].append(idx)
+            # Let's keep track of the current episode index
+            current_episode = episode_idx
+        else:
+            # We are still in the same episode, so there is nothing for us to do here
+            pass
+    # We have reached the end of the dataset, so we append the ending location of the last episode to the "to" list
+    episode_data_index["to"].append(idx + 1)
+
+    for k in ["from", "to"]:
+        episode_data_index[k] = torch.tensor(episode_data_index[k])
+
+    return episode_data_index

lerobot/common/datasets/pusht.py

@@ -1,212 +0,0 @@
-from pathlib import Path
-from typing import Callable
-
-import einops
-import numpy as np
-import pygame
-import pymunk
-import torch
-import torchrl
-import tqdm
-from diffusion_policy.common.replay_buffer import ReplayBuffer as DiffusionPolicyReplayBuffer
-from diffusion_policy.env.pusht.pusht_env import pymunk_to_shapely
-from tensordict import TensorDict
-from torchrl.data.replay_buffers.samplers import SliceSampler
-from torchrl.data.replay_buffers.storages import TensorStorage
-from torchrl.data.replay_buffers.writers import Writer
-
-from lerobot.common.datasets.abstract import AbstractExperienceReplay
-from lerobot.common.datasets.utils import download_and_extract_zip
-
-# as define in env
-SUCCESS_THRESHOLD = 0.95  # 95% coverage,
-
-DEFAULT_TEE_MASK = pymunk.ShapeFilter.ALL_MASKS()
-PUSHT_URL = "https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip"
-PUSHT_ZARR = Path("pusht/pusht_cchi_v7_replay.zarr")
-
-
-def get_goal_pose_body(pose):
-    mass = 1
-    inertia = pymunk.moment_for_box(mass, (50, 100))
-    body = pymunk.Body(mass, inertia)
-    # preserving the legacy assignment order for compatibility
-    # the order here doesn't matter somehow, maybe because CoM is aligned with body origin
-    body.position = pose[:2].tolist()
-    body.angle = pose[2]
-    return body
-
-
-def add_segment(space, a, b, radius):
-    shape = pymunk.Segment(space.static_body, a, b, radius)
-    shape.color = pygame.Color("LightGray")  # https://htmlcolorcodes.com/color-names
-    return shape
-
-
-def add_tee(
-    space,
-    position,
-    angle,
-    scale=30,
-    color="LightSlateGray",
-    mask=DEFAULT_TEE_MASK,
-):
-    mass = 1
-    length = 4
-    vertices1 = [
-        (-length * scale / 2, scale),
-        (length * scale / 2, scale),
-        (length * scale / 2, 0),
-        (-length * scale / 2, 0),
-    ]
-    inertia1 = pymunk.moment_for_poly(mass, vertices=vertices1)
-    vertices2 = [
-        (-scale / 2, scale),
-        (-scale / 2, length * scale),
-        (scale / 2, length * scale),
-        (scale / 2, scale),
-    ]
-    inertia2 = pymunk.moment_for_poly(mass, vertices=vertices1)
-    body = pymunk.Body(mass, inertia1 + inertia2)
-    shape1 = pymunk.Poly(body, vertices1)
-    shape2 = pymunk.Poly(body, vertices2)
-    shape1.color = pygame.Color(color)
-    shape2.color = pygame.Color(color)
-    shape1.filter = pymunk.ShapeFilter(mask=mask)
-    shape2.filter = pymunk.ShapeFilter(mask=mask)
-    body.center_of_gravity = (shape1.center_of_gravity + shape2.center_of_gravity) / 2
-    body.position = position
-    body.angle = angle
-    body.friction = 1
-    space.add(body, shape1, shape2)
-    return body
-
-
-class PushtExperienceReplay(AbstractExperienceReplay):
-    def __init__(
-        self,
-        dataset_id: str,
-        batch_size: int = None,
-        *,
-        shuffle: bool = True,
-        root: Path = None,
-        pin_memory: bool = False,
-        prefetch: int = None,
-        sampler: SliceSampler = None,
-        collate_fn: Callable = None,
-        writer: Writer = None,
-        transform: "torchrl.envs.Transform" = None,
-    ):
-        super().__init__(
-            dataset_id,
-            batch_size,
-            shuffle=shuffle,
-            root=root,
-            pin_memory=pin_memory,
-            prefetch=prefetch,
-            sampler=sampler,
-            collate_fn=collate_fn,
-            writer=writer,
-            transform=transform,
-        )
-
-    def _download_and_preproc(self):
-        raw_dir = self.data_dir.parent / f"{self.data_dir.name}_raw"
-        zarr_path = (raw_dir / PUSHT_ZARR).resolve()
-        if not zarr_path.is_dir():
-            raw_dir.mkdir(parents=True, exist_ok=True)
-            download_and_extract_zip(PUSHT_URL, raw_dir)
-
-        # load
-        dataset_dict = DiffusionPolicyReplayBuffer.copy_from_path(
-            zarr_path
-        )  # , keys=['img', 'state', 'action'])
-
-        episode_ids = torch.from_numpy(dataset_dict.get_episode_idxs())
-        num_episodes = dataset_dict.meta["episode_ends"].shape[0]
-        total_frames = dataset_dict["action"].shape[0]
-        assert len(
-            {dataset_dict[key].shape[0] for key in dataset_dict.keys()}  # noqa: SIM118
-        ), "Some data type dont have the same number of total frames."
-
-        # TODO: verify that goal pose is expected to be fixed
-        goal_pos_angle = np.array([256, 256, np.pi / 4])  # x, y, theta (in radians)
-        goal_body = get_goal_pose_body(goal_pos_angle)
-
-        imgs = torch.from_numpy(dataset_dict["img"])
-        imgs = einops.rearrange(imgs, "b h w c -> b c h w")
-        states = torch.from_numpy(dataset_dict["state"])
-        actions = torch.from_numpy(dataset_dict["action"])
-
-        idx0 = 0
-        idxtd = 0
-        for episode_id in tqdm.tqdm(range(num_episodes)):
-            idx1 = dataset_dict.meta["episode_ends"][episode_id]
-
-            num_frames = idx1 - idx0
-
-            assert (episode_ids[idx0:idx1] == episode_id).all()
-
-            image = imgs[idx0:idx1]
-
-            state = states[idx0:idx1]
-            agent_pos = state[:, :2]
-            block_pos = state[:, 2:4]
-            block_angle = state[:, 4]
-
-            reward = torch.zeros(num_frames, 1)
-            success = torch.zeros(num_frames, 1, dtype=torch.bool)
-            done = torch.zeros(num_frames, 1, dtype=torch.bool)
-            for i in range(num_frames):
-                space = pymunk.Space()
-                space.gravity = 0, 0
-                space.damping = 0
-
-                # Add walls.
-                walls = [
-                    add_segment(space, (5, 506), (5, 5), 2),
-                    add_segment(space, (5, 5), (506, 5), 2),
-                    add_segment(space, (506, 5), (506, 506), 2),
-                    add_segment(space, (5, 506), (506, 506), 2),
-                ]
-                space.add(*walls)
-
-                block_body = add_tee(space, block_pos[i].tolist(), block_angle[i].item())
-                goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
-                block_geom = pymunk_to_shapely(block_body, block_body.shapes)
-                intersection_area = goal_geom.intersection(block_geom).area
-                goal_area = goal_geom.area
-                coverage = intersection_area / goal_area
-                reward[i] = np.clip(coverage / SUCCESS_THRESHOLD, 0, 1)
-                success[i] = coverage > SUCCESS_THRESHOLD
-
-            # last step of demonstration is considered done
-            done[-1] = True
-
-            ep_td = TensorDict(
-                {
-                    ("observation", "image"): image[:-1],
-                    ("observation", "state"): agent_pos[:-1],
-                    "action": actions[idx0:idx1][:-1],
-                    "episode": episode_ids[idx0:idx1][:-1],
-                    "frame_id": torch.arange(0, num_frames - 1, 1),
-                    ("next", "observation", "image"): image[1:],
-                    ("next", "observation", "state"): agent_pos[1:],
-                    # TODO: verify that reward and done are aligned with image and agent_pos
-                    ("next", "reward"): reward[1:],
-                    ("next", "done"): done[1:],
-                    ("next", "success"): success[1:],
-                },
-                batch_size=num_frames - 1,
-            )
-
-            if episode_id == 0:
-                # hack to initialize tensordict data structure to store episodes
-                td_data = ep_td[0].expand(total_frames).memmap_like(self.data_dir)
-
-            td_data[idxtd : idxtd + len(ep_td)] = ep_td
-
-            idx0 = idx1
-            idxtd = idxtd + len(ep_td)
-
-        return TensorStorage(td_data.lock_())

lerobot/common/datasets/sampler.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Iterator, Union
+
+import torch
+
+
+class EpisodeAwareSampler:
+    def __init__(
+        self,
+        episode_data_index: dict,
+        episode_indices_to_use: Union[list, None] = None,
+        drop_n_first_frames: int = 0,
+        drop_n_last_frames: int = 0,
+        shuffle: bool = False,
+    ):
+        """Sampler that optionally incorporates episode boundary information.
+
+        Args:
+            episode_data_index: Dictionary with keys 'from' and 'to' containing the start and end indices of each episode.
+            episode_indices_to_use: List of episode indices to use. If None, all episodes are used.
+                                    Assumes that episodes are indexed from 0 to N-1.
+            drop_n_first_frames: Number of frames to drop from the start of each episode.
+            drop_n_last_frames: Number of frames to drop from the end of each episode.
+            shuffle: Whether to shuffle the indices.
+        """
+        indices = []
+        for episode_idx, (start_index, end_index) in enumerate(
+            zip(episode_data_index["from"], episode_data_index["to"], strict=True)
+        ):
+            if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
+                indices.extend(
+                    range(start_index.item() + drop_n_first_frames, end_index.item() - drop_n_last_frames)
+                )
+
+        self.indices = indices
+        self.shuffle = shuffle
+
+    def __iter__(self) -> Iterator[int]:
+        if self.shuffle:
+            for i in torch.randperm(len(self.indices)):
+                yield self.indices[i]
+        else:
+            for i in self.indices:
+                yield i
+
+    def __len__(self) -> int:
+        return len(self.indices)

lerobot/common/datasets/simxarm.py

@@ -1,103 +0,0 @@
-import pickle
-from pathlib import Path
-from typing import Callable
-
-import torch
-import torchrl
-import tqdm
-from tensordict import TensorDict
-from torchrl.data.replay_buffers.samplers import (
-    SliceSampler,
-)
-from torchrl.data.replay_buffers.storages import TensorStorage
-from torchrl.data.replay_buffers.writers import Writer
-
-from lerobot.common.datasets.abstract import AbstractExperienceReplay
-
-
-class SimxarmExperienceReplay(AbstractExperienceReplay):
-    available_datasets = [
-        "xarm_lift_medium",
-    ]
-
-    def __init__(
-        self,
-        dataset_id: str,
-        batch_size: int = None,
-        *,
-        shuffle: bool = True,
-        root: Path = None,
-        pin_memory: bool = False,
-        prefetch: int = None,
-        sampler: SliceSampler = None,
-        collate_fn: Callable = None,
-        writer: Writer = None,
-        transform: "torchrl.envs.Transform" = None,
-    ):
-        super().__init__(
-            dataset_id,
-            batch_size,
-            shuffle=shuffle,
-            root=root,
-            pin_memory=pin_memory,
-            prefetch=prefetch,
-            sampler=sampler,
-            collate_fn=collate_fn,
-            writer=writer,
-            transform=transform,
-        )
-
-    def _download_and_preproc(self):
-        # download
-        # TODO(rcadene)
-
-        dataset_path = self.data_dir / "buffer.pkl"
-        print(f"Using offline dataset '{dataset_path}'")
-        with open(dataset_path, "rb") as f:
-            dataset_dict = pickle.load(f)
-
-        total_frames = dataset_dict["actions"].shape[0]
-
-        idx0 = 0
-        idx1 = 0
-        episode_id = 0
-        for i in tqdm.tqdm(range(total_frames)):
-            idx1 += 1
-
-            if not dataset_dict["dones"][i]:
-                continue
-
-            num_frames = idx1 - idx0
-
-            image = torch.tensor(dataset_dict["observations"]["rgb"][idx0:idx1])
-            state = torch.tensor(dataset_dict["observations"]["state"][idx0:idx1])
-            next_image = torch.tensor(dataset_dict["next_observations"]["rgb"][idx0:idx1])
-            next_state = torch.tensor(dataset_dict["next_observations"]["state"][idx0:idx1])
-            next_reward = torch.tensor(dataset_dict["rewards"][idx0:idx1])
-            next_done = torch.tensor(dataset_dict["dones"][idx0:idx1])
-
-            episode = TensorDict(
-                {
-                    ("observation", "image"): image,
-                    ("observation", "state"): state,
-                    "action": torch.tensor(dataset_dict["actions"][idx0:idx1]),
-                    "episode": torch.tensor([episode_id] * num_frames, dtype=torch.int),
-                    "frame_id": torch.arange(0, num_frames, 1),
-                    ("next", "observation", "image"): next_image,
-                    ("next", "observation", "state"): next_state,
-                    ("next", "observation", "reward"): next_reward,
-                    ("next", "observation", "done"): next_done,
-                },
-                batch_size=num_frames,
-            )
-
-            if episode_id == 0:
-                # hack to initialize tensordict data structure to store episodes
-                td_data = episode[0].expand(total_frames).memmap_like(self.data_dir)
-
-            td_data[idx0:idx1] = episode
-
-            episode_id += 1
-            idx0 = idx1
-
-        return TensorStorage(td_data.lock_())

lerobot/common/datasets/transforms.py

@@ -0,0 +1,249 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import collections
+from dataclasses import dataclass, field
+from typing import Any, Callable, Sequence
+
+import torch
+from torchvision.transforms import v2
+from torchvision.transforms.v2 import Transform
+from torchvision.transforms.v2 import functional as F  # noqa: N812
+
+
+class RandomSubsetApply(Transform):
+    """Apply a random subset of N transformations from a list of transformations.
+
+    Args:
+        transforms: list of transformations.
+        p: represents the multinomial probabilities (with no replacement) used for sampling the transform.
+            If the sum of the weights is not 1, they will be normalized. If ``None`` (default), all transforms
+            have the same probability.
+        n_subset: number of transformations to apply. If ``None``, all transforms are applied.
+            Must be in [1, len(transforms)].
+        random_order: apply transformations in a random order.
+    """
+
+    def __init__(
+        self,
+        transforms: Sequence[Callable],
+        p: list[float] | None = None,
+        n_subset: int | None = None,
+        random_order: bool = False,
+    ) -> None:
+        super().__init__()
+        if not isinstance(transforms, Sequence):
+            raise TypeError("Argument transforms should be a sequence of callables")
+        if p is None:
+            p = [1] * len(transforms)
+        elif len(p) != len(transforms):
+            raise ValueError(
+                f"Length of p doesn't match the number of transforms: {len(p)} != {len(transforms)}"
+            )
+
+        if n_subset is None:
+            n_subset = len(transforms)
+        elif not isinstance(n_subset, int):
+            raise TypeError("n_subset should be an int or None")
+        elif not (1 <= n_subset <= len(transforms)):
+            raise ValueError(f"n_subset should be in the interval [1, {len(transforms)}]")
+
+        self.transforms = transforms
+        total = sum(p)
+        self.p = [prob / total for prob in p]
+        self.n_subset = n_subset
+        self.random_order = random_order
+
+        self.selected_transforms = None
+
+    def forward(self, *inputs: Any) -> Any:
+        needs_unpacking = len(inputs) > 1
+
+        selected_indices = torch.multinomial(torch.tensor(self.p), self.n_subset)
+        if not self.random_order:
+            selected_indices = selected_indices.sort().values
+
+        self.selected_transforms = [self.transforms[i] for i in selected_indices]
+
+        for transform in self.selected_transforms:
+            outputs = transform(*inputs)
+            inputs = outputs if needs_unpacking else (outputs,)
+
+        return outputs
+
+    def extra_repr(self) -> str:
+        return (
+            f"transforms={self.transforms}, "
+            f"p={self.p}, "
+            f"n_subset={self.n_subset}, "
+            f"random_order={self.random_order}"
+        )
+
+
+class SharpnessJitter(Transform):
+    """Randomly change the sharpness of an image or video.
+
+    Similar to a v2.RandomAdjustSharpness with p=1 and a sharpness_factor sampled randomly.
+    While v2.RandomAdjustSharpness applies — with a given probability — a fixed sharpness_factor to an image,
+    SharpnessJitter applies a random sharpness_factor each time. This is to have a more diverse set of
+    augmentations as a result.
+
+    A sharpness_factor of 0 gives a blurred image, 1 gives the original image while 2 increases the sharpness
+    by a factor of 2.
+
+    If the input is a :class:`torch.Tensor`,
+    it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions.
+
+    Args:
+        sharpness: How much to jitter sharpness. sharpness_factor is chosen uniformly from
+            [max(0, 1 - sharpness), 1 + sharpness] or the given
+            [min, max]. Should be non negative numbers.
+    """
+
+    def __init__(self, sharpness: float | Sequence[float]) -> None:
+        super().__init__()
+        self.sharpness = self._check_input(sharpness)
+
+    def _check_input(self, sharpness):
+        if isinstance(sharpness, (int, float)):
+            if sharpness < 0:
+                raise ValueError("If sharpness is a single number, it must be non negative.")
+            sharpness = [1.0 - sharpness, 1.0 + sharpness]
+            sharpness[0] = max(sharpness[0], 0.0)
+        elif isinstance(sharpness, collections.abc.Sequence) and len(sharpness) == 2:
+            sharpness = [float(v) for v in sharpness]
+        else:
+            raise TypeError(f"{sharpness=} should be a single number or a sequence with length 2.")
+
+        if not 0.0 <= sharpness[0] <= sharpness[1]:
+            raise ValueError(f"sharpnesss values should be between (0., inf), but got {sharpness}.")
+
+        return float(sharpness[0]), float(sharpness[1])
+
+    def make_params(self, flat_inputs: list[Any]) -> dict[str, Any]:
+        sharpness_factor = torch.empty(1).uniform_(self.sharpness[0], self.sharpness[1]).item()
+        return {"sharpness_factor": sharpness_factor}
+
+    def transform(self, inpt: Any, params: dict[str, Any]) -> Any:
+        sharpness_factor = params["sharpness_factor"]
+        return self._call_kernel(F.adjust_sharpness, inpt, sharpness_factor=sharpness_factor)
+
+
+@dataclass
+class ImageTransformConfig:
+    """
+    For each transform, the following parameters are available:
+      weight: This represents the multinomial probability (with no replacement)
+            used for sampling the transform. If the sum of the weights is not 1,
+            they will be normalized.
+      type: The name of the class used. This is either a class available under torchvision.transforms.v2 or a
+            custom transform defined here.
+      kwargs: Lower & upper bound respectively used for sampling the transform's parameter
+            (following uniform distribution) when it's applied.
+    """
+
+    weight: float = 1.0
+    type: str = "Identity"
+    kwargs: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ImageTransformsConfig:
+    """
+    These transforms are all using standard torchvision.transforms.v2
+    You can find out how these transformations affect images here:
+    https://pytorch.org/vision/0.18/auto_examples/transforms/plot_transforms_illustrations.html
+    We use a custom RandomSubsetApply container to sample them.
+    """
+
+    # Set this flag to `true` to enable transforms during training
+    enable: bool = False
+    # This is the maximum number of transforms (sampled from these below) that will be applied to each frame.
+    # It's an integer in the interval [1, number_of_available_transforms].
+    max_num_transforms: int = 3
+    # By default, transforms are applied in Torchvision's suggested order (shown below).
+    # Set this to True to apply them in a random order.
+    random_order: bool = False
+    tfs: dict[str, ImageTransformConfig] = field(
+        default_factory=lambda: {
+            "brightness": ImageTransformConfig(
+                weight=1.0,
+                type="ColorJitter",
+                kwargs={"brightness": (0.8, 1.2)},
+            ),
+            "contrast": ImageTransformConfig(
+                weight=1.0,
+                type="ColorJitter",
+                kwargs={"contrast": (0.8, 1.2)},
+            ),
+            "saturation": ImageTransformConfig(
+                weight=1.0,
+                type="ColorJitter",
+                kwargs={"saturation": (0.5, 1.5)},
+            ),
+            "hue": ImageTransformConfig(
+                weight=1.0,
+                type="ColorJitter",
+                kwargs={"hue": (-0.05, 0.05)},
+            ),
+            "sharpness": ImageTransformConfig(
+                weight=1.0,
+                type="SharpnessJitter",
+                kwargs={"sharpness": (0.5, 1.5)},
+            ),
+        }
+    )
+
+
+def make_transform_from_config(cfg: ImageTransformConfig):
+    if cfg.type == "Identity":
+        return v2.Identity(**cfg.kwargs)
+    elif cfg.type == "ColorJitter":
+        return v2.ColorJitter(**cfg.kwargs)
+    elif cfg.type == "SharpnessJitter":
+        return SharpnessJitter(**cfg.kwargs)
+    else:
+        raise ValueError(f"Transform '{cfg.type}' is not valid.")
+
+
+class ImageTransforms(Transform):
+    """A class to compose image transforms based on configuration."""
+
+    def __init__(self, cfg: ImageTransformsConfig) -> None:
+        super().__init__()
+        self._cfg = cfg
+
+        self.weights = []
+        self.transforms = {}
+        for tf_name, tf_cfg in cfg.tfs.items():
+            if tf_cfg.weight <= 0.0:
+                continue
+
+            self.transforms[tf_name] = make_transform_from_config(tf_cfg)
+            self.weights.append(tf_cfg.weight)
+
+        n_subset = min(len(self.transforms), cfg.max_num_transforms)
+        if n_subset == 0 or not cfg.enable:
+            self.tf = v2.Identity()
+        else:
+            self.tf = RandomSubsetApply(
+                transforms=list(self.transforms.values()),
+                p=self.weights,
+                n_subset=n_subset,
+                random_order=cfg.random_order,
+            )
+
+    def forward(self, *inputs: Any) -> Any:
+        return self.tf(*inputs)

lerobot/common/datasets/utils.py

@@ -1,30 +1,813 @@
-import io
-import zipfile
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import contextlib
+import importlib.resources
+import json
+import logging
+from collections.abc import Iterator
+from itertools import accumulate
 from pathlib import Path
+from pprint import pformat
+from types import SimpleNamespace
+from typing import Any
 
-import requests
-import tqdm
+import datasets
+import jsonlines
+import numpy as np
+import packaging.version
+import torch
+from datasets.table import embed_table_storage
+from huggingface_hub import DatasetCard, DatasetCardData, HfApi
+from huggingface_hub.errors import RevisionNotFoundError
+from PIL import Image as PILImage
+from torchvision import transforms
+
+from lerobot.common.datasets.backward_compatibility import (
+    V21_MESSAGE,
+    BackwardCompatibilityError,
+    ForwardCompatibilityError,
+)
+from lerobot.common.robot_devices.robots.utils import Robot
+from lerobot.common.utils.utils import is_valid_numpy_dtype_string
+from lerobot.configs.types import DictLike, FeatureType, PolicyFeature
+
+DEFAULT_CHUNK_SIZE = 1000  # Max number of episodes per chunk
+
+INFO_PATH = "meta/info.json"
+EPISODES_PATH = "meta/episodes.jsonl"
+STATS_PATH = "meta/stats.json"
+EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
+TASKS_PATH = "meta/tasks.jsonl"
+
+DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
+DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+DEFAULT_IMAGE_PATH = "images/{image_key}/episode_{episode_index:06d}/frame_{frame_index:06d}.png"
+
+DATASET_CARD_TEMPLATE = """
+---
+# Metadata will go there
+---
+This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
+
+## {}
+
+"""
+
+DEFAULT_FEATURES = {
+    "timestamp": {"dtype": "float32", "shape": (1,), "names": None},
+    "frame_index": {"dtype": "int64", "shape": (1,), "names": None},
+    "episode_index": {"dtype": "int64", "shape": (1,), "names": None},
+    "index": {"dtype": "int64", "shape": (1,), "names": None},
+    "task_index": {"dtype": "int64", "shape": (1,), "names": None},
+}
 
 
-def download_and_extract_zip(url: str, destination_folder: Path) -> bool:
-    print(f"downloading from {url}")
-    response = requests.get(url, stream=True)
-    if response.status_code == 200:
-        total_size = int(response.headers.get("content-length", 0))
-        progress_bar = tqdm.tqdm(total=total_size, unit="B", unit_scale=True)
+def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
+    """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
 
-        zip_file = io.BytesIO()
-        for chunk in response.iter_content(chunk_size=1024):
-            if chunk:
-                zip_file.write(chunk)
-                progress_bar.update(len(chunk))
+    For example:
+    ```
+    >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}`
+    >>> print(flatten_dict(dct))
+    {"a/b": 1, "a/c/d": 2, "e": 3}
+    """
+    items = []
+    for k, v in d.items():
+        new_key = f"{parent_key}{sep}{k}" if parent_key else k
+        if isinstance(v, dict):
+            items.extend(flatten_dict(v, new_key, sep=sep).items())
+        else:
+            items.append((new_key, v))
+    return dict(items)
 
-        progress_bar.close()
 
-        zip_file.seek(0)
+def unflatten_dict(d: dict, sep: str = "/") -> dict:
+    outdict = {}
+    for key, value in d.items():
+        parts = key.split(sep)
+        d = outdict
+        for part in parts[:-1]:
+            if part not in d:
+                d[part] = {}
+            d = d[part]
+        d[parts[-1]] = value
+    return outdict
 
-        with zipfile.ZipFile(zip_file, "r") as zip_ref:
-            zip_ref.extractall(destination_folder)
+
+def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
+    split_keys = flattened_key.split(sep)
+    getter = obj[split_keys[0]]
+    if len(split_keys) == 1:
+        return getter
+
+    for key in split_keys[1:]:
+        getter = getter[key]
+
+    return getter
+
+
+def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
+    serialized_dict = {}
+    for key, value in flatten_dict(stats).items():
+        if isinstance(value, (torch.Tensor, np.ndarray)):
+            serialized_dict[key] = value.tolist()
+        elif isinstance(value, np.generic):
+            serialized_dict[key] = value.item()
+        elif isinstance(value, (int, float)):
+            serialized_dict[key] = value
+        else:
+            raise NotImplementedError(f"The value '{value}' of type '{type(value)}' is not supported.")
+    return unflatten_dict(serialized_dict)
+
+
+def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
+    # Embed image bytes into the table before saving to parquet
+    format = dataset.format
+    dataset = dataset.with_format("arrow")
+    dataset = dataset.map(embed_table_storage, batched=False)
+    dataset = dataset.with_format(**format)
+    return dataset
+
+
+def load_json(fpath: Path) -> Any:
+    with open(fpath) as f:
+        return json.load(f)
+
+
+def write_json(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with open(fpath, "w") as f:
+        json.dump(data, f, indent=4, ensure_ascii=False)
+
+
+def load_jsonlines(fpath: Path) -> list[Any]:
+    with jsonlines.open(fpath, "r") as reader:
+        return list(reader)
+
+
+def write_jsonlines(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "w") as writer:
+        writer.write_all(data)
+
+
+def append_jsonlines(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "a") as writer:
+        writer.write(data)
+
+
+def write_info(info: dict, local_dir: Path):
+    write_json(info, local_dir / INFO_PATH)
+
+
+def load_info(local_dir: Path) -> dict:
+    info = load_json(local_dir / INFO_PATH)
+    for ft in info["features"].values():
+        ft["shape"] = tuple(ft["shape"])
+    return info
+
+
+def write_stats(stats: dict, local_dir: Path):
+    serialized_stats = serialize_dict(stats)
+    write_json(serialized_stats, local_dir / STATS_PATH)
+
+
+def cast_stats_to_numpy(stats) -> dict[str, dict[str, np.ndarray]]:
+    stats = {key: np.array(value) for key, value in flatten_dict(stats).items()}
+    return unflatten_dict(stats)
+
+
+def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]]:
+    if not (local_dir / STATS_PATH).exists():
+        return None
+    stats = load_json(local_dir / STATS_PATH)
+    return cast_stats_to_numpy(stats)
+
+
+def write_task(task_index: int, task: dict, local_dir: Path):
+    task_dict = {
+        "task_index": task_index,
+        "task": task,
+    }
+    append_jsonlines(task_dict, local_dir / TASKS_PATH)
+
+
+def load_tasks(local_dir: Path) -> tuple[dict, dict]:
+    tasks = load_jsonlines(local_dir / TASKS_PATH)
+    tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
+    task_to_task_index = {task: task_index for task_index, task in tasks.items()}
+    return tasks, task_to_task_index
+
+
+def write_episode(episode: dict, local_dir: Path):
+    append_jsonlines(episode, local_dir / EPISODES_PATH)
+
+
+def load_episodes(local_dir: Path) -> dict:
+    episodes = load_jsonlines(local_dir / EPISODES_PATH)
+    return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
+
+
+def write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path):
+    # We wrap episode_stats in a dictionary since `episode_stats["episode_index"]`
+    # is a dictionary of stats and not an integer.
+    episode_stats = {"episode_index": episode_index, "stats": serialize_dict(episode_stats)}
+    append_jsonlines(episode_stats, local_dir / EPISODES_STATS_PATH)
+
+
+def load_episodes_stats(local_dir: Path) -> dict:
+    episodes_stats = load_jsonlines(local_dir / EPISODES_STATS_PATH)
+    return {
+        item["episode_index"]: cast_stats_to_numpy(item["stats"])
+        for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
+    }
+
+
+def backward_compatible_episodes_stats(
+    stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
+) -> dict[str, dict[str, np.ndarray]]:
+    return dict.fromkeys(episodes, stats)
+
+
+def load_image_as_numpy(
+    fpath: str | Path, dtype: np.dtype = np.float32, channel_first: bool = True
+) -> np.ndarray:
+    img = PILImage.open(fpath).convert("RGB")
+    img_array = np.array(img, dtype=dtype)
+    if channel_first:  # (H, W, C) -> (C, H, W)
+        img_array = np.transpose(img_array, (2, 0, 1))
+    if np.issubdtype(dtype, np.floating):
+        img_array /= 255.0
+    return img_array
+
+
+def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
+    """Get a transform function that convert items from Hugging Face dataset (pyarrow)
+    to torch tensors. Importantly, images are converted from PIL, which corresponds to
+    a channel last representation (h w c) of uint8 type, to a torch image representation
+    with channel first (c h w) of float32 type in range [0,1].
+    """
+    for key in items_dict:
+        first_item = items_dict[key][0]
+        if isinstance(first_item, PILImage.Image):
+            to_tensor = transforms.ToTensor()
+            items_dict[key] = [to_tensor(img) for img in items_dict[key]]
+        elif first_item is None:
+            pass
+        else:
+            items_dict[key] = [x if isinstance(x, str) else torch.tensor(x) for x in items_dict[key]]
+    return items_dict
+
+
+def is_valid_version(version: str) -> bool:
+    try:
+        packaging.version.parse(version)
         return True
-    else:
+    except packaging.version.InvalidVersion:
         return False
+
+
+def check_version_compatibility(
+    repo_id: str,
+    version_to_check: str | packaging.version.Version,
+    current_version: str | packaging.version.Version,
+    enforce_breaking_major: bool = True,
+) -> None:
+    v_check = (
+        packaging.version.parse(version_to_check)
+        if not isinstance(version_to_check, packaging.version.Version)
+        else version_to_check
+    )
+    v_current = (
+        packaging.version.parse(current_version)
+        if not isinstance(current_version, packaging.version.Version)
+        else current_version
+    )
+    if v_check.major < v_current.major and enforce_breaking_major:
+        raise BackwardCompatibilityError(repo_id, v_check)
+    elif v_check.minor < v_current.minor:
+        logging.warning(V21_MESSAGE.format(repo_id=repo_id, version=v_check))
+
+
+def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
+    """Returns available valid versions (branches and tags) on given repo."""
+    api = HfApi()
+    repo_refs = api.list_repo_refs(repo_id, repo_type="dataset")
+    repo_refs = [b.name for b in repo_refs.branches + repo_refs.tags]
+    repo_versions = []
+    for ref in repo_refs:
+        with contextlib.suppress(packaging.version.InvalidVersion):
+            repo_versions.append(packaging.version.parse(ref))
+
+    return repo_versions
+
+
+def get_safe_version(repo_id: str, version: str | packaging.version.Version) -> str:
+    """
+    Returns the version if available on repo or the latest compatible one.
+    Otherwise, will throw a `CompatibilityError`.
+    """
+    target_version = (
+        packaging.version.parse(version) if not isinstance(version, packaging.version.Version) else version
+    )
+    hub_versions = get_repo_versions(repo_id)
+
+    if not hub_versions:
+        raise RevisionNotFoundError(
+            f"""Your dataset must be tagged with a codebase version.
+            Assuming _version_ is the codebase_version value in the info.json, you can run this:
+            ```python
+            from huggingface_hub import HfApi
+
+            hub_api = HfApi()
+            hub_api.create_tag("{repo_id}", tag="_version_", repo_type="dataset")
+            ```
+            """
+        )
+
+    if target_version in hub_versions:
+        return f"v{target_version}"
+
+    compatibles = [
+        v for v in hub_versions if v.major == target_version.major and v.minor <= target_version.minor
+    ]
+    if compatibles:
+        return_version = max(compatibles)
+        if return_version < target_version:
+            logging.warning(f"Revision {version} for {repo_id} not found, using version v{return_version}")
+        return f"v{return_version}"
+
+    lower_major = [v for v in hub_versions if v.major < target_version.major]
+    if lower_major:
+        raise BackwardCompatibilityError(repo_id, max(lower_major))
+
+    upper_versions = [v for v in hub_versions if v > target_version]
+    assert len(upper_versions) > 0
+    raise ForwardCompatibilityError(repo_id, min(upper_versions))
+
+
+def get_hf_features_from_features(features: dict) -> datasets.Features:
+    hf_features = {}
+    for key, ft in features.items():
+        if ft["dtype"] == "video":
+            continue
+        elif ft["dtype"] == "image":
+            hf_features[key] = datasets.Image()
+        elif ft["shape"] == (1,):
+            hf_features[key] = datasets.Value(dtype=ft["dtype"])
+        elif len(ft["shape"]) == 1:
+            hf_features[key] = datasets.Sequence(
+                length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
+            )
+        elif len(ft["shape"]) == 2:
+            hf_features[key] = datasets.Array2D(shape=ft["shape"], dtype=ft["dtype"])
+        elif len(ft["shape"]) == 3:
+            hf_features[key] = datasets.Array3D(shape=ft["shape"], dtype=ft["dtype"])
+        elif len(ft["shape"]) == 4:
+            hf_features[key] = datasets.Array4D(shape=ft["shape"], dtype=ft["dtype"])
+        elif len(ft["shape"]) == 5:
+            hf_features[key] = datasets.Array5D(shape=ft["shape"], dtype=ft["dtype"])
+        else:
+            raise ValueError(f"Corresponding feature is not valid: {ft}")
+
+    return datasets.Features(hf_features)
+
+
+def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
+    camera_ft = {}
+    if robot.cameras:
+        camera_ft = {
+            key: {"dtype": "video" if use_videos else "image", **ft}
+            for key, ft in robot.camera_features.items()
+        }
+    return {**robot.motor_features, **camera_ft, **DEFAULT_FEATURES}
+
+
+def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFeature]:
+    # TODO(aliberts): Implement "type" in dataset features and simplify this
+    policy_features = {}
+    for key, ft in features.items():
+        shape = ft["shape"]
+        if ft["dtype"] in ["image", "video"]:
+            type = FeatureType.VISUAL
+            if len(shape) != 3:
+                raise ValueError(f"Number of dimensions of {key} != 3 (shape={shape})")
+
+            names = ft["names"]
+            # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
+            if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
+                shape = (shape[2], shape[0], shape[1])
+        elif key == "observation.environment_state":
+            type = FeatureType.ENV
+        elif key.startswith("observation"):
+            type = FeatureType.STATE
+        elif key == "action":
+            type = FeatureType.ACTION
+        else:
+            continue
+
+        policy_features[key] = PolicyFeature(
+            type=type,
+            shape=shape,
+        )
+
+    return policy_features
+
+
+def create_empty_dataset_info(
+    codebase_version: str,
+    fps: int,
+    robot_type: str,
+    features: dict,
+    use_videos: bool,
+) -> dict:
+    return {
+        "codebase_version": codebase_version,
+        "robot_type": robot_type,
+        "total_episodes": 0,
+        "total_frames": 0,
+        "total_tasks": 0,
+        "total_videos": 0,
+        "total_chunks": 0,
+        "chunks_size": DEFAULT_CHUNK_SIZE,
+        "fps": fps,
+        "splits": {},
+        "data_path": DEFAULT_PARQUET_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
+        "features": features,
+    }
+
+
+def get_episode_data_index(
+    episode_dicts: dict[dict], episodes: list[int] | None = None
+) -> dict[str, torch.Tensor]:
+    episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in episode_dicts.items()}
+    if episodes is not None:
+        episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
+
+    cumulative_lengths = list(accumulate(episode_lengths.values()))
+    return {
+        "from": torch.LongTensor([0] + cumulative_lengths[:-1]),
+        "to": torch.LongTensor(cumulative_lengths),
+    }
+
+
+def check_timestamps_sync(
+    timestamps: np.ndarray,
+    episode_indices: np.ndarray,
+    episode_data_index: dict[str, np.ndarray],
+    fps: int,
+    tolerance_s: float,
+    raise_value_error: bool = True,
+) -> bool:
+    """
+    This check is to make sure that each timestamp is separated from the next by (1/fps) +/- tolerance
+    to account for possible numerical error.
+
+    Args:
+        timestamps (np.ndarray): Array of timestamps in seconds.
+        episode_indices (np.ndarray): Array indicating the episode index for each timestamp.
+        episode_data_index (dict[str, np.ndarray]): A dictionary that includes 'to',
+            which identifies indices for the end of each episode.
+        fps (int): Frames per second. Used to check the expected difference between consecutive timestamps.
+        tolerance_s (float): Allowed deviation from the expected (1/fps) difference.
+        raise_value_error (bool): Whether to raise a ValueError if the check fails.
+
+    Returns:
+        bool: True if all checked timestamp differences lie within tolerance, False otherwise.
+
+    Raises:
+        ValueError: If the check fails and `raise_value_error` is True.
+    """
+    if timestamps.shape != episode_indices.shape:
+        raise ValueError(
+            "timestamps and episode_indices should have the same shape. "
+            f"Found {timestamps.shape=} and {episode_indices.shape=}."
+        )
+
+    # Consecutive differences
+    diffs = np.diff(timestamps)
+    within_tolerance = np.abs(diffs - (1.0 / fps)) <= tolerance_s
+
+    # Mask to ignore differences at the boundaries between episodes
+    mask = np.ones(len(diffs), dtype=bool)
+    ignored_diffs = episode_data_index["to"][:-1] - 1  # indices at the end of each episode
+    mask[ignored_diffs] = False
+    filtered_within_tolerance = within_tolerance[mask]
+
+    # Check if all remaining diffs are within tolerance
+    if not np.all(filtered_within_tolerance):
+        # Track original indices before masking
+        original_indices = np.arange(len(diffs))
+        filtered_indices = original_indices[mask]
+        outside_tolerance_filtered_indices = np.nonzero(~filtered_within_tolerance)[0]
+        outside_tolerance_indices = filtered_indices[outside_tolerance_filtered_indices]
+
+        outside_tolerances = []
+        for idx in outside_tolerance_indices:
+            entry = {
+                "timestamps": [timestamps[idx], timestamps[idx + 1]],
+                "diff": diffs[idx],
+                "episode_index": episode_indices[idx].item()
+                if hasattr(episode_indices[idx], "item")
+                else episode_indices[idx],
+            }
+            outside_tolerances.append(entry)
+
+        if raise_value_error:
+            raise ValueError(
+                f"""One or several timestamps unexpectedly violate the tolerance inside episode range.
+                This might be due to synchronization issues during data collection.
+                \n{pformat(outside_tolerances)}"""
+            )
+        return False
+
+    return True
+
+
+def check_delta_timestamps(
+    delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
+) -> bool:
+    """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
+    This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be
+    actual timestamps from the dataset.
+    """
+    outside_tolerance = {}
+    for key, delta_ts in delta_timestamps.items():
+        within_tolerance = [abs(ts * fps - round(ts * fps)) / fps <= tolerance_s for ts in delta_ts]
+        if not all(within_tolerance):
+            outside_tolerance[key] = [
+                ts for ts, is_within in zip(delta_ts, within_tolerance, strict=True) if not is_within
+            ]
+
+    if len(outside_tolerance) > 0:
+        if raise_value_error:
+            raise ValueError(
+                f"""
+                The following delta_timestamps are found outside of tolerance range.
+                Please make sure they are multiples of 1/{fps} +/- tolerance and adjust
+                their values accordingly.
+                \n{pformat(outside_tolerance)}
+                """
+            )
+        return False
+
+    return True
+
+
+def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
+    delta_indices = {}
+    for key, delta_ts in delta_timestamps.items():
+        delta_indices[key] = [round(d * fps) for d in delta_ts]
+
+    return delta_indices
+
+
+def cycle(iterable):
+    """The equivalent of itertools.cycle, but safe for Pytorch dataloaders.
+
+    See https://github.com/pytorch/pytorch/issues/23900 for information on why itertools.cycle is not safe.
+    """
+    iterator = iter(iterable)
+    while True:
+        try:
+            yield next(iterator)
+        except StopIteration:
+            iterator = iter(iterable)
+
+
+def create_branch(repo_id, *, branch: str, repo_type: str | None = None) -> None:
+    """Create a branch on a existing Hugging Face repo. Delete the branch if it already
+    exists before creating it.
+    """
+    api = HfApi()
+
+    branches = api.list_repo_refs(repo_id, repo_type=repo_type).branches
+    refs = [branch.ref for branch in branches]
+    ref = f"refs/heads/{branch}"
+    if ref in refs:
+        api.delete_branch(repo_id, repo_type=repo_type, branch=branch)
+
+    api.create_branch(repo_id, repo_type=repo_type, branch=branch)
+
+
+def create_lerobot_dataset_card(
+    tags: list | None = None,
+    dataset_info: dict | None = None,
+    **kwargs,
+) -> DatasetCard:
+    """
+    Keyword arguments will be used to replace values in ./lerobot/common/datasets/card_template.md.
+    Note: If specified, license must be one of https://huggingface.co/docs/hub/repositories-licenses.
+    """
+    card_tags = ["LeRobot"]
+
+    if tags:
+        card_tags += tags
+    if dataset_info:
+        dataset_structure = "[meta/info.json](meta/info.json):\n"
+        dataset_structure += f"```json\n{json.dumps(dataset_info, indent=4)}\n```\n"
+        kwargs = {**kwargs, "dataset_structure": dataset_structure}
+    card_data = DatasetCardData(
+        license=kwargs.get("license"),
+        tags=card_tags,
+        task_categories=["robotics"],
+        configs=[
+            {
+                "config_name": "default",
+                "data_files": "data/*/*.parquet",
+            }
+        ],
+    )
+
+    card_template = (importlib.resources.files("lerobot.common.datasets") / "card_template.md").read_text()
+
+    return DatasetCard.from_template(
+        card_data=card_data,
+        template_str=card_template,
+        **kwargs,
+    )
+
+
+class IterableNamespace(SimpleNamespace):
+    """
+    A namespace object that supports both dictionary-like iteration and dot notation access.
+    Automatically converts nested dictionaries into IterableNamespaces.
+
+    This class extends SimpleNamespace to provide:
+    - Dictionary-style iteration over keys
+    - Access to items via both dot notation (obj.key) and brackets (obj["key"])
+    - Dictionary-like methods: items(), keys(), values()
+    - Recursive conversion of nested dictionaries
+
+    Args:
+        dictionary: Optional dictionary to initialize the namespace
+        **kwargs: Additional keyword arguments passed to SimpleNamespace
+
+    Examples:
+        >>> data = {"name": "Alice", "details": {"age": 25}}
+        >>> ns = IterableNamespace(data)
+        >>> ns.name
+        'Alice'
+        >>> ns.details.age
+        25
+        >>> list(ns.keys())
+        ['name', 'details']
+        >>> for key, value in ns.items():
+        ...     print(f"{key}: {value}")
+        name: Alice
+        details: IterableNamespace(age=25)
+    """
+
+    def __init__(self, dictionary: dict[str, Any] = None, **kwargs):
+        super().__init__(**kwargs)
+        if dictionary is not None:
+            for key, value in dictionary.items():
+                if isinstance(value, dict):
+                    setattr(self, key, IterableNamespace(value))
+                else:
+                    setattr(self, key, value)
+
+    def __iter__(self) -> Iterator[str]:
+        return iter(vars(self))
+
+    def __getitem__(self, key: str) -> Any:
+        return vars(self)[key]
+
+    def items(self):
+        return vars(self).items()
+
+    def values(self):
+        return vars(self).values()
+
+    def keys(self):
+        return vars(self).keys()
+
+
+def validate_frame(frame: dict, features: dict):
+    optional_features = {"timestamp"}
+    expected_features = (set(features) - set(DEFAULT_FEATURES.keys())) | {"task"}
+    actual_features = set(frame.keys())
+
+    error_message = validate_features_presence(actual_features, expected_features, optional_features)
+
+    if "task" in frame:
+        error_message += validate_feature_string("task", frame["task"])
+
+    common_features = actual_features & (expected_features | optional_features)
+    for name in common_features - {"task"}:
+        error_message += validate_feature_dtype_and_shape(name, features[name], frame[name])
+
+    if error_message:
+        raise ValueError(error_message)
+
+
+def validate_features_presence(
+    actual_features: set[str], expected_features: set[str], optional_features: set[str]
+):
+    error_message = ""
+    missing_features = expected_features - actual_features
+    extra_features = actual_features - (expected_features | optional_features)
+
+    if missing_features or extra_features:
+        error_message += "Feature mismatch in `frame` dictionary:\n"
+        if missing_features:
+            error_message += f"Missing features: {missing_features}\n"
+        if extra_features:
+            error_message += f"Extra features: {extra_features}\n"
+
+    return error_message
+
+
+def validate_feature_dtype_and_shape(name: str, feature: dict, value: np.ndarray | PILImage.Image | str):
+    expected_dtype = feature["dtype"]
+    expected_shape = feature["shape"]
+    if is_valid_numpy_dtype_string(expected_dtype):
+        return validate_feature_numpy_array(name, expected_dtype, expected_shape, value)
+    elif expected_dtype in ["image", "video"]:
+        return validate_feature_image_or_video(name, expected_shape, value)
+    elif expected_dtype == "string":
+        return validate_feature_string(name, value)
+    else:
+        raise NotImplementedError(f"The feature dtype '{expected_dtype}' is not implemented yet.")
+
+
+def validate_feature_numpy_array(
+    name: str, expected_dtype: str, expected_shape: list[int], value: np.ndarray
+):
+    error_message = ""
+    if isinstance(value, np.ndarray):
+        actual_dtype = value.dtype
+        actual_shape = value.shape
+
+        if actual_dtype != np.dtype(expected_dtype):
+            error_message += f"The feature '{name}' of dtype '{actual_dtype}' is not of the expected dtype '{expected_dtype}'.\n"
+
+        if actual_shape != expected_shape:
+            error_message += f"The feature '{name}' of shape '{actual_shape}' does not have the expected shape '{expected_shape}'.\n"
+    else:
+        error_message += f"The feature '{name}' is not a 'np.ndarray'. Expected type is '{expected_dtype}', but type '{type(value)}' provided instead.\n"
+
+    return error_message
+
+
+def validate_feature_image_or_video(name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image):
+    # Note: The check of pixels range ([0,1] for float and [0,255] for uint8) is done by the image writer threads.
+    error_message = ""
+    if isinstance(value, np.ndarray):
+        actual_shape = value.shape
+        c, h, w = expected_shape
+        if len(actual_shape) != 3 or (actual_shape != (c, h, w) and actual_shape != (h, w, c)):
+            error_message += f"The feature '{name}' of shape '{actual_shape}' does not have the expected shape '{(c, h, w)}' or '{(h, w, c)}'.\n"
+    elif isinstance(value, PILImage.Image):
+        pass
+    else:
+        error_message += f"The feature '{name}' is expected to be of type 'PIL.Image' or 'np.ndarray' channel first or channel last, but type '{type(value)}' provided instead.\n"
+
+    return error_message
+
+
+def validate_feature_string(name: str, value: str):
+    if not isinstance(value, str):
+        return f"The feature '{name}' is expected to be of type 'str', but type '{type(value)}' provided instead.\n"
+    return ""
+
+
+def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict):
+    if "size" not in episode_buffer:
+        raise ValueError("size key not found in episode_buffer")
+
+    if "task" not in episode_buffer:
+        raise ValueError("task key not found in episode_buffer")
+
+    if episode_buffer["episode_index"] != total_episodes:
+        # TODO(aliberts): Add option to use existing episode_index
+        raise NotImplementedError(
+            "You might have manually provided the episode_buffer with an episode_index that doesn't "
+            "match the total number of episodes already in the dataset. This is not supported for now."
+        )
+
+    if episode_buffer["size"] == 0:
+        raise ValueError("You must add one or several frames with `add_frame` before calling `add_episode`.")
+
+    buffer_keys = set(episode_buffer.keys()) - {"task", "size"}
+    if not buffer_keys == set(features):
+        raise ValueError(
+            f"Features from `episode_buffer` don't match the ones in `features`."
+            f"In episode_buffer not in features: {buffer_keys - set(features)}"
+            f"In features not in episode_buffer: {set(features) - buffer_keys}"
+        )

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -0,0 +1,884 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script is for internal use to convert all datasets under the 'lerobot' hub user account to v2.
+
+Note: Since the original Aloha datasets don't use shadow motors, you need to comment those out in
+lerobot/configs/robot/aloha.yaml before running this script.
+"""
+
+import traceback
+from pathlib import Path
+from textwrap import dedent
+
+from lerobot import available_datasets
+from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset
+from lerobot.common.robot_devices.robots.configs import AlohaRobotConfig
+
+LOCAL_DIR = Path("data/")
+
+# spellchecker:off
+ALOHA_MOBILE_INFO = {
+    "robot_config": AlohaRobotConfig(),
+    "license": "mit",
+    "url": "https://mobile-aloha.github.io/",
+    "paper": "https://arxiv.org/abs/2401.02117",
+    "citation_bibtex": dedent(r"""
+        @inproceedings{fu2024mobile,
+            author    = {Fu, Zipeng and Zhao, Tony Z. and Finn, Chelsea},
+            title     = {Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation},
+            booktitle = {arXiv},
+            year      = {2024},
+        }""").lstrip(),
+}
+ALOHA_STATIC_INFO = {
+    "robot_config": AlohaRobotConfig(),
+    "license": "mit",
+    "url": "https://tonyzhaozh.github.io/aloha/",
+    "paper": "https://arxiv.org/abs/2304.13705",
+    "citation_bibtex": dedent(r"""
+        @article{Zhao2023LearningFB,
+            title={Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware},
+            author={Tony Zhao and Vikash Kumar and Sergey Levine and Chelsea Finn},
+            journal={RSS},
+            year={2023},
+            volume={abs/2304.13705},
+            url={https://arxiv.org/abs/2304.13705}
+        }""").lstrip(),
+}
+PUSHT_INFO = {
+    "license": "mit",
+    "url": "https://diffusion-policy.cs.columbia.edu/",
+    "paper": "https://arxiv.org/abs/2303.04137v5",
+    "citation_bibtex": dedent(r"""
+        @article{chi2024diffusionpolicy,
+            author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
+            title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+            journal = {The International Journal of Robotics Research},
+            year = {2024},
+        }""").lstrip(),
+}
+XARM_INFO = {
+    "license": "mit",
+    "url": "https://www.nicklashansen.com/td-mpc/",
+    "paper": "https://arxiv.org/abs/2203.04955",
+    "citation_bibtex": dedent(r"""
+        @inproceedings{Hansen2022tdmpc,
+            title={Temporal Difference Learning for Model Predictive Control},
+            author={Nicklas Hansen and Xiaolong Wang and Hao Su},
+            booktitle={ICML},
+            year={2022}
+        }
+    """),
+}
+UNITREEH_INFO = {
+    "license": "apache-2.0",
+}
+
+DATASETS = {
+    "aloha_mobile_cabinet": {
+        "single_task": "Open the top cabinet, store the pot inside it then close the cabinet.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_chair": {
+        "single_task": "Push the chairs in front of the desk to place them against it.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_elevator": {
+        "single_task": "Take the elevator to the 1st floor.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_shrimp": {
+        "single_task": "Sauté the raw shrimp on both sides, then serve it in the bowl.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_wash_pan": {
+        "single_task": "Pick up the pan, rinse it in the sink and then place it in the drying rack.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_wipe_wine": {
+        "single_task": "Pick up the wet cloth on the faucet and use it to clean the spilled wine on the table and underneath the glass.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_static_battery": {
+        "single_task": "Place the battery into the slot of the remote controller.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_candy": {"single_task": "Pick up the candy and unwrap it.", **ALOHA_STATIC_INFO},
+    "aloha_static_coffee": {
+        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray, then push the 'Hot Water' and 'Travel Mug' buttons.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_coffee_new": {
+        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_cups_open": {
+        "single_task": "Pick up the plastic cup and open its lid.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_fork_pick_up": {
+        "single_task": "Pick up the fork and place it on the plate.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_pingpong_test": {
+        "single_task": "Transfer one of the two balls in the right glass into the left glass, then transfer it back to the right glass.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_pro_pencil": {
+        "single_task": "Pick up the pencil with the right arm, hand it over to the left arm then place it back onto the table.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_screw_driver": {
+        "single_task": "Pick up the screwdriver with the right arm, hand it over to the left arm then place it into the cup.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_tape": {
+        "single_task": "Cut a small piece of tape from the tape dispenser then place it on the cardboard box's edge.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_thread_velcro": {
+        "single_task": "Pick up the velcro cable tie with the left arm, then insert the end of the velcro tie into the other end's loop with the right arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_towel": {
+        "single_task": "Pick up a piece of paper towel and place it on the spilled liquid.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_vinh_cup": {
+        "single_task": "Pick up the plastic cup with the right arm, then pop its lid open with the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_vinh_cup_left": {
+        "single_task": "Pick up the plastic cup with the left arm, then pop its lid open with the right arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_ziploc_slide": {"single_task": "Slide open the ziploc bag.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted_image": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_insertion_human": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_human_image": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_scripted": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_scripted_image": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_human": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_human_image": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "pusht": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "pusht_image": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "unitreeh1_fold_clothes": {"single_task": "Fold the sweatshirt.", **UNITREEH_INFO},
+    "unitreeh1_rearrange_objects": {"single_task": "Put the object into the bin.", **UNITREEH_INFO},
+    "unitreeh1_two_robot_greeting": {
+        "single_task": "Greet the other robot with a high five.",
+        **UNITREEH_INFO,
+    },
+    "unitreeh1_warehouse": {
+        "single_task": "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.",
+        **UNITREEH_INFO,
+    },
+    "xarm_lift_medium": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_push_medium": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "umi_cup_in_the_wild": {
+        "single_task": "Put the cup on the plate.",
+        "license": "apache-2.0",
+    },
+    "asu_table_top": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://link.springer.com/article/10.1007/s10514-023-10129-1",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{zhou2023modularity,
+                title={Modularity through Attention: Efficient Training and Transfer of Language-Conditioned Policies for Robot Manipulation},
+                author={Zhou, Yifan and Sonawani, Shubham and Phielipp, Mariano and Stepputtis, Simon and Amor, Heni},
+                booktitle={Conference on Robot Learning},
+                pages={1684--1695},
+                year={2023},
+                organization={PMLR}
+            }
+            @article{zhou2023learning,
+                title={Learning modular language-conditioned robot policies through attention},
+                author={Zhou, Yifan and Sonawani, Shubham and Phielipp, Mariano and Ben Amor, Heni and Stepputtis, Simon},
+                journal={Autonomous Robots},
+                pages={1--21},
+                year={2023},
+                publisher={Springer}
+            }""").lstrip(),
+    },
+    "austin_buds_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/BUDS-website/",
+        "paper": "https://arxiv.org/abs/2109.13841",
+        "citation_bibtex": dedent(r"""
+            @article{zhu2022bottom,
+                title={Bottom-Up Skill Discovery From Unsegmented Demonstrations for Long-Horizon Robot Manipulation},
+                author={Zhu, Yifeng and Stone, Peter and Zhu, Yuke},
+                journal={IEEE Robotics and Automation Letters},
+                volume={7},
+                number={2},
+                pages={4126--4133},
+                year={2022},
+                publisher={IEEE}
+            }""").lstrip(),
+    },
+    "austin_sailor_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/sailor/",
+        "paper": "https://arxiv.org/abs/2210.11435",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{nasiriany2022sailor,
+                title={Learning and Retrieval from Prior Data for Skill-based Imitation Learning},
+                author={Soroush Nasiriany and Tian Gao and Ajay Mandlekar and Yuke Zhu},
+                booktitle={Conference on Robot Learning (CoRL)},
+                year={2022}
+            }""").lstrip(),
+    },
+    "austin_sirius_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/sirius/",
+        "paper": "https://arxiv.org/abs/2211.08416",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{liu2022robot,
+                title = {Robot Learning on the Job: Human-in-the-Loop Autonomy and Learning During Deployment},
+                author = {Huihan Liu and Soroush Nasiriany and Lance Zhang and Zhiyao Bao and Yuke Zhu},
+                booktitle = {Robotics: Science and Systems (RSS)},
+                year = {2023}
+            }""").lstrip(),
+    },
+    "berkeley_autolab_ur5": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://sites.google.com/view/berkeley-ur5/home",
+        "citation_bibtex": dedent(r"""
+            @misc{BerkeleyUR5Website,
+                title = {Berkeley {UR5} Demonstration Dataset},
+                author = {Lawrence Yunliang Chen and Simeon Adebola and Ken Goldberg},
+                howpublished = {https://sites.google.com/view/berkeley-ur5/home},
+            }""").lstrip(),
+    },
+    "berkeley_cable_routing": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://sites.google.com/view/cablerouting/home",
+        "paper": "https://arxiv.org/abs/2307.08927",
+        "citation_bibtex": dedent(r"""
+            @article{luo2023multistage,
+                author    = {Jianlan Luo and Charles Xu and Xinyang Geng and Gilbert Feng and Kuan Fang and Liam Tan and Stefan Schaal and Sergey Levine},
+                title     = {Multi-Stage Cable Routing through Hierarchical Imitation Learning},
+                journal   = {arXiv pre-print},
+                year      = {2023},
+                url       = {https://arxiv.org/abs/2307.08927},
+            }""").lstrip(),
+    },
+    "berkeley_fanuc_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/berkeley.edu/fanuc-manipulation",
+        "citation_bibtex": dedent(r"""
+            @article{fanuc_manipulation2023,
+                title={Fanuc Manipulation: A Dataset for Learning-based Manipulation with FANUC Mate 200iD Robot},
+                author={Zhu, Xinghao and Tian, Ran and Xu, Chenfeng and Ding, Mingyu and Zhan, Wei and Tomizuka, Masayoshi},
+                year={2023},
+            }""").lstrip(),
+    },
+    "berkeley_gnm_cory_hall": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/1709.10489",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{kahn2018self,
+                title={Self-supervised deep reinforcement learning with generalized computation graphs for robot navigation},
+                author={Kahn, Gregory and Villaflor, Adam and Ding, Bosen and Abbeel, Pieter and Levine, Sergey},
+                booktitle={2018 IEEE international conference on robotics and automation (ICRA)},
+                pages={5129--5136},
+                year={2018},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "berkeley_gnm_recon": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/recon-robot",
+        "paper": "https://arxiv.org/abs/2104.05859",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{shah2021rapid,
+                title={Rapid Exploration for Open-World Navigation with Latent Goal Models},
+                author={Dhruv Shah and Benjamin Eysenbach and Nicholas Rhinehart and Sergey Levine},
+                booktitle={5th Annual Conference on Robot Learning },
+                year={2021},
+                url={https://openreview.net/forum?id=d_SWJhyKfVw}
+            }""").lstrip(),
+    },
+    "berkeley_gnm_sac_son": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/SACSoN-review",
+        "paper": "https://arxiv.org/abs/2306.01874",
+        "citation_bibtex": dedent(r"""
+            @article{hirose2023sacson,
+                title={SACSoN: Scalable Autonomous Data Collection for Social Navigation},
+                author={Hirose, Noriaki and Shah, Dhruv and Sridhar, Ajay and Levine, Sergey},
+                journal={arXiv preprint arXiv:2306.01874},
+                year={2023}
+            }""").lstrip(),
+    },
+    "berkeley_mvp": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/2203.06173",
+        "citation_bibtex": dedent(r"""
+            @InProceedings{Radosavovic2022,
+                title = {Real-World Robot Learning with Masked Visual Pre-training},
+                author = {Ilija Radosavovic and Tete Xiao and Stephen James and Pieter Abbeel and Jitendra Malik and Trevor Darrell},
+                booktitle = {CoRL},
+                year = {2022}
+            }""").lstrip(),
+    },
+    "berkeley_rpt": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/2306.10007",
+        "citation_bibtex": dedent(r"""
+            @article{Radosavovic2023,
+                title={Robot Learning with Sensorimotor Pre-training},
+                author={Ilija Radosavovic and Baifeng Shi and Letian Fu and Ken Goldberg and Trevor Darrell and Jitendra Malik},
+                year={2023},
+                journal={arXiv:2306.10007}
+            }""").lstrip(),
+    },
+    "cmu_franka_exploration_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://human-world-model.github.io/",
+        "paper": "https://arxiv.org/abs/2308.10901",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{mendonca2023structured,
+                title={Structured World Models from Human Videos},
+                author={Mendonca, Russell  and Bahl, Shikhar and Pathak, Deepak},
+                journal={RSS},
+                year={2023}
+            }""").lstrip(),
+    },
+    "cmu_play_fusion": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://play-fusion.github.io/",
+        "paper": "https://arxiv.org/abs/2312.04549",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{chen2023playfusion,
+                title={PlayFusion: Skill Acquisition via Diffusion from Language-Annotated Play},
+                author={Chen, Lili and Bahl, Shikhar and Pathak, Deepak},
+                booktitle={CoRL},
+                year={2023}
+            }""").lstrip(),
+    },
+    "cmu_stretch": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://robo-affordances.github.io/",
+        "paper": "https://arxiv.org/abs/2304.08488",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{bahl2023affordances,
+                title={Affordances from Human Videos as a Versatile Representation for Robotics},
+                author={Bahl, Shikhar and Mendonca, Russell and Chen, Lili and Jain, Unnat and Pathak, Deepak},
+                booktitle={CVPR},
+                year={2023}
+            }
+                @article{mendonca2023structured,
+                title={Structured World Models from Human Videos},
+                author={Mendonca, Russell and Bahl, Shikhar and Pathak, Deepak},
+                journal={CoRL},
+                year={2023}
+            }""").lstrip(),
+    },
+    "columbia_cairlab_pusht_real": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://diffusion-policy.cs.columbia.edu/",
+        "paper": "https://arxiv.org/abs/2303.04137v5",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{chi2023diffusionpolicy,
+                title={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+                author={Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin and Song, Shuran},
+                booktitle={Proceedings of Robotics: Science and Systems (RSS)},
+                year={2023}
+            }""").lstrip(),
+    },
+    "conq_hose_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/conq-hose-manipulation-dataset/home",
+        "citation_bibtex": dedent(r"""
+            @misc{ConqHoseManipData,
+                author={Peter Mitrano and Dmitry Berenson},
+                title={Conq Hose Manipulation Dataset, v1.15.0},
+                year={2024},
+                howpublished={https://sites.google.com/view/conq-hose-manipulation-dataset}
+            }""").lstrip(),
+    },
+    "dlr_edan_shared_control": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://ieeexplore.ieee.org/document/9341156",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{vogel_edan_2020,
+                title = {EDAN - an EMG-Controlled Daily Assistant to Help People with Physical Disabilities},
+                language = {en},
+                booktitle = {2020 {IEEE}/{RSJ} {International} {Conference} on {Intelligent} {Robots} and {Systems} ({IROS})},
+                author = {Vogel, Jörn and Hagengruber, Annette and Iskandar, Maged and Quere, Gabriel and Leipscher, Ulrike and Bustamante, Samuel and Dietrich, Alexander and Hoeppner, Hannes and Leidner, Daniel and Albu-Schäffer, Alin},
+                year = {2020}
+            }
+            @inproceedings{quere_shared_2020,
+                address = {Paris, France},
+                title = {Shared {Control} {Templates} for {Assistive} {Robotics}},
+                language = {en},
+                booktitle = {2020 {IEEE} {International} {Conference} on {Robotics} and {Automation} ({ICRA})},
+                author = {Quere, Gabriel and Hagengruber, Annette and Iskandar, Maged and Bustamante, Samuel and Leidner, Daniel and Stulp, Freek and Vogel, Joern},
+                year = {2020},
+                pages = {7},
+            }""").lstrip(),
+    },
+    "dlr_sara_grid_clamp": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://www.researchsquare.com/article/rs-3289569/v1",
+        "citation_bibtex": dedent(r"""
+            @article{padalkar2023guided,
+                title={A guided reinforcement learning approach using shared control templates for learning manipulation skills in the real world},
+                author={Padalkar, Abhishek and Quere, Gabriel and Raffin, Antonin and Silv{\'e}rio, Jo{\~a}o and Stulp, Freek},
+                journal={Research square preprint rs-3289569/v1},
+                year={2023}
+            }""").lstrip(),
+    },
+    "dlr_sara_pour": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://elib.dlr.de/193739/1/padalkar2023rlsct.pdf",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{padalkar2023guiding,
+                title={Guiding Reinforcement Learning with Shared Control Templates},
+                author={Padalkar, Abhishek and Quere, Gabriel and Steinmetz, Franz and Raffin, Antonin and Nieuwenhuisen, Matthias and Silv{\'e}rio, Jo{\~a}o and Stulp, Freek},
+                booktitle={40th IEEE International Conference on Robotics and Automation, ICRA 2023},
+                year={2023},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "droid_100": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://droid-dataset.github.io/",
+        "paper": "https://arxiv.org/abs/2403.12945",
+        "citation_bibtex": dedent(r"""
+            @article{khazatsky2024droid,
+                title   = {DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset},
+                author  = {Alexander Khazatsky and Karl Pertsch and Suraj Nair and Ashwin Balakrishna and Sudeep Dasari and Siddharth Karamcheti and Soroush Nasiriany and Mohan Kumar Srirama and Lawrence Yunliang Chen and Kirsty Ellis and Peter David Fagan and Joey Hejna and Masha Itkina and Marion Lepert and Yecheng Jason Ma and Patrick Tree Miller and Jimmy Wu and Suneel Belkhale and Shivin Dass and Huy Ha and Arhan Jain and Abraham Lee and Youngwoon Lee and Marius Memmel and Sungjae Park and Ilija Radosavovic and Kaiyuan Wang and Albert Zhan and Kevin Black and Cheng Chi and Kyle Beltran Hatch and Shan Lin and Jingpei Lu and Jean Mercat and Abdul Rehman and Pannag R Sanketi and Archit Sharma and Cody Simpson and Quan Vuong and Homer Rich Walke and Blake Wulfe and Ted Xiao and Jonathan Heewon Yang and Arefeh Yavary and Tony Z. Zhao and Christopher Agia and Rohan Baijal and Mateo Guaman Castro and Daphne Chen and Qiuyu Chen and Trinity Chung and Jaimyn Drake and Ethan Paul Foster and Jensen Gao and David Antonio Herrera and Minho Heo and Kyle Hsu and Jiaheng Hu and Donovon Jackson and Charlotte Le and Yunshuang Li and Kevin Lin and Roy Lin and Zehan Ma and Abhiram Maddukuri and Suvir Mirchandani and Daniel Morton and Tony Nguyen and Abigail O'Neill and Rosario Scalise and Derick Seale and Victor Son and Stephen Tian and Emi Tran and Andrew E. Wang and Yilin Wu and Annie Xie and Jingyun Yang and Patrick Yin and Yunchu Zhang and Osbert Bastani and Glen Berseth and Jeannette Bohg and Ken Goldberg and Abhinav Gupta and Abhishek Gupta and Dinesh Jayaraman and Joseph J Lim and Jitendra Malik and Roberto Martín-Martín and Subramanian Ramamoorthy and Dorsa Sadigh and Shuran Song and Jiajun Wu and Michael C. Yip and Yuke Zhu and Thomas Kollar and Sergey Levine and Chelsea Finn},
+                year    = {2024},
+            }""").lstrip(),
+    },
+    "fmb": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://functional-manipulation-benchmark.github.io/",
+        "paper": "https://arxiv.org/abs/2401.08553",
+        "citation_bibtex": dedent(r"""
+            @article{luo2024fmb,
+                title={FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning},
+                author={Luo, Jianlan and Xu, Charles and Liu, Fangchen and Tan, Liam and Lin, Zipeng and Wu, Jeffrey and Abbeel, Pieter and Levine, Sergey},
+                journal={arXiv preprint arXiv:2401.08553},
+                year={2024}
+            }""").lstrip(),
+    },
+    "iamlab_cmu_pickup_insert": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://openreview.net/forum?id=WuBv9-IGDUA",
+        "paper": "https://arxiv.org/abs/2401.14502",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{saxena2023multiresolution,
+                title={Multi-Resolution Sensing for Real-Time Control with Vision-Language Models},
+                author={Saumya Saxena and Mohit Sharma and Oliver Kroemer},
+                booktitle={7th Annual Conference on Robot Learning},
+                year={2023},
+                url={https://openreview.net/forum?id=WuBv9-IGDUA}
+            }""").lstrip(),
+    },
+    "imperialcollege_sawyer_wrist_cam": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+    },
+    "jaco_play": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://github.com/clvrai/clvr_jaco_play_dataset",
+        "citation_bibtex": dedent(r"""
+            @software{dass2023jacoplay,
+                author = {Dass, Shivin and Yapeter, Jullian and Zhang, Jesse and Zhang, Jiahui
+                            and Pertsch, Karl and Nikolaidis, Stefanos and Lim, Joseph J.},
+                title = {CLVR Jaco Play Dataset},
+                url = {https://github.com/clvrai/clvr_jaco_play_dataset},
+                version = {1.0.0},
+                year = {2023}
+            }""").lstrip(),
+    },
+    "kaist_nonprehensile": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://github.com/JaeHyung-Kim/rlds_dataset_builder",
+        "citation_bibtex": dedent(r"""
+            @article{kimpre,
+                title={Pre-and post-contact policy decomposition for non-prehensile manipulation with zero-shot sim-to-real transfer},
+                author={Kim, Minchan and Han, Junhyek and Kim, Jaehyung and Kim, Beomjoon},
+                booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+                year={2023},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "nyu_door_opening_surprising_effectiveness": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://jyopari.github.io/VINN/",
+        "paper": "https://arxiv.org/abs/2112.01511",
+        "citation_bibtex": dedent(r"""
+            @misc{pari2021surprising,
+                title={The Surprising Effectiveness of Representation Learning for Visual Imitation},
+                author={Jyothish Pari and Nur Muhammad Shafiullah and Sridhar Pandian Arunachalam and Lerrel Pinto},
+                year={2021},
+                eprint={2112.01511},
+                archivePrefix={arXiv},
+                primaryClass={cs.RO}
+            }""").lstrip(),
+    },
+    "nyu_franka_play_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://play-to-policy.github.io/",
+        "paper": "https://arxiv.org/abs/2210.10047",
+        "citation_bibtex": dedent(r"""
+            @article{cui2022play,
+                title   = {From Play to Policy: Conditional Behavior Generation from Uncurated Robot Data},
+                author  = {Cui, Zichen Jeff and Wang, Yibin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},
+                journal = {arXiv preprint arXiv:2210.10047},
+                year    = {2022}
+            }""").lstrip(),
+    },
+    "nyu_rot_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://rot-robot.github.io/",
+        "paper": "https://arxiv.org/abs/2206.15469",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{haldar2023watch,
+                title={Watch and match: Supercharging imitation with regularized optimal transport},
+                author={Haldar, Siddhant and Mathur, Vaibhav and Yarats, Denis and Pinto, Lerrel},
+                booktitle={Conference on Robot Learning},
+                pages={32--43},
+                year={2023},
+                organization={PMLR}
+            }""").lstrip(),
+    },
+    "roboturk": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://roboturk.stanford.edu/dataset_real.html",
+        "paper": "PAPER",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{mandlekar2019scaling,
+                title={Scaling robot supervision to hundreds of hours with roboturk: Robotic manipulation dataset through human reasoning and dexterity},
+                author={Mandlekar, Ajay and Booher, Jonathan and Spero, Max and Tung, Albert and Gupta, Anchit and Zhu, Yuke and Garg, Animesh and Savarese, Silvio and Fei-Fei, Li},
+                booktitle={2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+                pages={1048--1055},
+                year={2019},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "stanford_hydra_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/hydra-il-2023",
+        "paper": "https://arxiv.org/abs/2306.17237",
+        "citation_bibtex": dedent(r"""
+            @article{belkhale2023hydra,
+                title={HYDRA: Hybrid Robot Actions for Imitation Learning},
+                author={Belkhale, Suneel and Cui, Yuchen and Sadigh, Dorsa},
+                journal={arxiv},
+                year={2023}
+            }""").lstrip(),
+    },
+    "stanford_kuka_multimodal_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/visionandtouch",
+        "paper": "https://arxiv.org/abs/1810.10191",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{lee2019icra,
+                title={Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks},
+                author={Lee, Michelle A and Zhu, Yuke and Srinivasan, Krishnan and Shah, Parth and Savarese, Silvio and Fei-Fei, Li and  Garg, Animesh and Bohg, Jeannette},
+                booktitle={2019 IEEE International Conference on Robotics and Automation (ICRA)},
+                year={2019},
+                url={https://arxiv.org/abs/1810.10191}
+            }""").lstrip(),
+    },
+    "stanford_robocook": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://hshi74.github.io/robocook/",
+        "paper": "https://arxiv.org/abs/2306.14447",
+        "citation_bibtex": dedent(r"""
+            @article{shi2023robocook,
+                title={RoboCook: Long-Horizon Elasto-Plastic Object Manipulation with Diverse Tools},
+                author={Shi, Haochen and Xu, Huazhe and Clarke, Samuel and Li, Yunzhu and Wu, Jiajun},
+                journal={arXiv preprint arXiv:2306.14447},
+                year={2023}
+            }""").lstrip(),
+    },
+    "taco_play": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://www.kaggle.com/datasets/oiermees/taco-robot",
+        "paper": "https://arxiv.org/abs/2209.08959, https://arxiv.org/abs/2210.01911",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{rosete2022tacorl,
+                author = {Erick Rosete-Beas and Oier Mees and Gabriel Kalweit and Joschka Boedecker and Wolfram Burgard},
+                title = {Latent Plans for Task Agnostic Offline Reinforcement Learning},
+                journal = {Proceedings of the 6th Conference on Robot Learning (CoRL)},
+                year = {2022}
+            }
+            @inproceedings{mees23hulc2,
+                title={Grounding  Language  with  Visual  Affordances  over  Unstructured  Data},
+                author={Oier Mees and Jessica Borja-Diaz and Wolfram Burgard},
+                booktitle = {Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
+                year={2023},
+                address = {London, UK}
+            }""").lstrip(),
+    },
+    "tokyo_u_lsmo": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "URL",
+        "paper": "https://arxiv.org/abs/2107.05842",
+        "citation_bibtex": dedent(r"""
+            @Article{Osa22,
+                author  = {Takayuki Osa},
+                journal = {The International Journal of Robotics Research},
+                title   = {Motion Planning by Learning the Solution Manifold in Trajectory Optimization},
+                year    = {2022},
+                number  = {3},
+                pages   = {291--311},
+                volume  = {41},
+            }""").lstrip(),
+    },
+    "toto": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://toto-benchmark.org/",
+        "paper": "https://arxiv.org/abs/2306.00942",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{zhou2023train,
+                author={Zhou, Gaoyue and Dean, Victoria and Srirama, Mohan Kumar and Rajeswaran, Aravind and Pari, Jyothish and Hatch, Kyle and Jain, Aryan and Yu, Tianhe and Abbeel, Pieter and Pinto, Lerrel and Finn, Chelsea and Gupta, Abhinav},
+                booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
+                title={Train Offline, Test Online: A Real Robot Learning Benchmark},
+                year={2023},
+            }""").lstrip(),
+    },
+    "ucsd_kitchen_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @ARTICLE{ucsd_kitchens,
+                author = {Ge Yan, Kris Wu, and Xiaolong Wang},
+                title = {{ucsd kitchens Dataset}},
+                year = {2023},
+                month = {August}
+            }""").lstrip(),
+    },
+    "ucsd_pick_and_place_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://owmcorl.github.io/#",
+        "paper": "https://arxiv.org/abs/2310.16029",
+        "citation_bibtex": dedent(r"""
+            @preprint{Feng2023Finetuning,
+                title={Finetuning Offline World Models in the Real World},
+                author={Yunhai Feng, Nicklas Hansen, Ziyan Xiong, Chandramouli Rajagopalan, Xiaolong Wang},
+                year={2023}
+            }""").lstrip(),
+    },
+    "uiuc_d3field": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://robopil.github.io/d3fields/",
+        "paper": "https://arxiv.org/abs/2309.16118",
+        "citation_bibtex": dedent(r"""
+            @article{wang2023d3field,
+                title={D^3Field: Dynamic 3D Descriptor Fields for Generalizable Robotic Manipulation},
+                author={Wang, Yixuan and Li, Zhuoran and Zhang, Mingtong and Driggs-Campbell, Katherine and Wu, Jiajun and Fei-Fei, Li and Li, Yunzhu},
+                journal={arXiv preprint arXiv:},
+                year={2023},
+            }""").lstrip(),
+    },
+    "usc_cloth_sim": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://uscresl.github.io/dmfd/",
+        "paper": "https://arxiv.org/abs/2207.10148",
+        "citation_bibtex": dedent(r"""
+            @article{salhotra2022dmfd,
+                author={Salhotra, Gautam and Liu, I-Chun Arthur and Dominguez-Kuhne, Marcus and Sukhatme, Gaurav S.},
+                journal={IEEE Robotics and Automation Letters},
+                title={Learning Deformable Object Manipulation From Expert Demonstrations},
+                year={2022},
+                volume={7},
+                number={4},
+                pages={8775-8782},
+                doi={10.1109/LRA.2022.3187843}
+            }""").lstrip(),
+    },
+    "utaustin_mutex": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/MUTEX/",
+        "paper": "https://arxiv.org/abs/2309.14320",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{shah2023mutex,
+                title={{MUTEX}: Learning Unified Policies from Multimodal Task Specifications},
+                author={Rutav Shah and Roberto Mart{\'\i}n-Mart{\'\i}n and Yuke Zhu},
+                booktitle={7th Annual Conference on Robot Learning},
+                year={2023},
+                url={https://openreview.net/forum?id=PwqiqaaEzJ}
+            }""").lstrip(),
+    },
+    "utokyo_pr2_opening_fridge": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @misc{oh2023pr2utokyodatasets,
+                author={Jihoon Oh and Naoaki Kanazawa and Kento Kawaharazuka},
+                title={X-Embodiment U-Tokyo PR2 Datasets},
+                year={2023},
+                url={https://github.com/ojh6404/rlds_dataset_builder},
+            }""").lstrip(),
+    },
+    "utokyo_pr2_tabletop_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @misc{oh2023pr2utokyodatasets,
+                author={Jihoon Oh and Naoaki Kanazawa and Kento Kawaharazuka},
+                title={X-Embodiment U-Tokyo PR2 Datasets},
+                year={2023},
+                url={https://github.com/ojh6404/rlds_dataset_builder},
+            }""").lstrip(),
+    },
+    "utokyo_saytap": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://saytap.github.io/",
+        "paper": "https://arxiv.org/abs/2306.07580",
+        "citation_bibtex": dedent(r"""
+            @article{saytap2023,
+                author = {Yujin Tang and Wenhao Yu and Jie Tan and Heiga Zen and Aleksandra Faust and
+                Tatsuya Harada},
+                title  = {SayTap: Language to Quadrupedal Locomotion},
+                eprint = {arXiv:2306.07580},
+                url    = {https://saytap.github.io},
+                note   = {https://saytap.github.io},
+                year   = {2023}
+            }""").lstrip(),
+    },
+    "utokyo_xarm_bimanual": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "citation_bibtex": dedent(r"""
+            @misc{matsushima2023weblab,
+                title={Weblab xArm Dataset},
+                author={Tatsuya Matsushima and Hiroki Furuta and Yusuke Iwasawa and Yutaka Matsuo},
+                year={2023},
+            }""").lstrip(),
+    },
+    "utokyo_xarm_pick_and_place": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "citation_bibtex": dedent(r"""
+            @misc{matsushima2023weblab,
+                title={Weblab xArm Dataset},
+                author={Tatsuya Matsushima and Hiroki Furuta and Yusuke Iwasawa and Yutaka Matsuo},
+                year={2023},
+            }""").lstrip(),
+    },
+    "viola": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/VIOLA/",
+        "paper": "https://arxiv.org/abs/2210.11339",
+        "citation_bibtex": dedent(r"""
+            @article{zhu2022viola,
+                title={VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Priors},
+                author={Zhu, Yifeng and Joshi, Abhishek and Stone, Peter and Zhu, Yuke},
+                journal={6th Annual Conference on Robot Learning (CoRL)},
+                year={2022}
+            }""").lstrip(),
+    },
+}
+# spellchecker:on
+
+
+def batch_convert():
+    status = {}
+    logfile = LOCAL_DIR / "conversion_log.txt"
+    assert set(DATASETS) == {id_.split("/")[1] for id_ in available_datasets}
+    for num, (name, kwargs) in enumerate(DATASETS.items()):
+        repo_id = f"lerobot/{name}"
+        print(f"\nConverting {repo_id} ({num}/{len(DATASETS)})")
+        print("---------------------------------------------------------")
+        try:
+            convert_dataset(repo_id, LOCAL_DIR, **kwargs)
+            status = f"{repo_id}: success."
+            with open(logfile, "a") as file:
+                file.write(status + "\n")
+        except Exception:
+            status = f"{repo_id}: failed\n    {traceback.format_exc()}"
+            with open(logfile, "a") as file:
+                file.write(status + "\n")
+            continue
+
+
+if __name__ == "__main__":
+    batch_convert()

lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py

@@ -0,0 +1,664 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 1.6 to
+2.0. You will be required to provide the 'tasks', which is a short but accurate description in plain English
+for each of the task performed in the dataset. This will allow to easily train models with task-conditioning.
+
+We support 3 different scenarios for these tasks (see instructions below):
+    1. Single task dataset: all episodes of your dataset have the same single task.
+    2. Single task episodes: the episodes of your dataset each contain a single task but they can differ from
+      one episode to the next.
+    3. Multi task episodes: episodes of your dataset may each contain several different tasks.
+
+
+Can you can also provide a robot config .yaml file (not mandatory) to this script via the option
+'--robot-config' so that it writes information about the robot (robot type, motors names) this dataset was
+recorded with. For now, only Aloha/Koch type robots are supported with this option.
+
+
+# 1. Single task dataset
+If your dataset contains a single task, you can simply provide it directly via the CLI with the
+'--single-task' option.
+
+Examples:
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id lerobot/aloha_sim_insertion_human_image \
+    --single-task "Insert the peg into the socket." \
+    --robot-config lerobot/configs/robot/aloha.yaml \
+    --local-dir data
+```
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id aliberts/koch_tutorial \
+    --single-task "Pick the Lego block and drop it in the box on the right." \
+    --robot-config lerobot/configs/robot/koch.yaml \
+    --local-dir data
+```
+
+
+# 2. Single task episodes
+If your dataset is a multi-task dataset, you have two options to provide the tasks to this script:
+
+- If your dataset already contains a language instruction column in its parquet file, you can simply provide
+  this column's name with the '--tasks-col' arg.
+
+    Example:
+
+    ```bash
+    python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+        --repo-id lerobot/stanford_kuka_multimodal_dataset \
+        --tasks-col "language_instruction" \
+        --local-dir data
+    ```
+
+- If your dataset doesn't contain a language instruction, you should provide the path to a .json file with the
+  '--tasks-path' arg. This file should have the following structure where keys correspond to each
+  episode_index in the dataset, and values are the language instruction for that episode.
+
+    Example:
+
+    ```json
+    {
+        "0": "Do something",
+        "1": "Do something else",
+        "2": "Do something",
+        "3": "Go there",
+        ...
+    }
+    ```
+
+# 3. Multi task episodes
+If you have multiple tasks per episodes, your dataset should contain a language instruction column in its
+parquet file, and you must provide this column's name with the '--tasks-col' arg.
+
+Example:
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id lerobot/stanford_kuka_multimodal_dataset \
+    --tasks-col "language_instruction" \
+    --local-dir data
+```
+"""
+
+import argparse
+import contextlib
+import filecmp
+import json
+import logging
+import math
+import shutil
+import subprocess
+import tempfile
+from pathlib import Path
+
+import datasets
+import pyarrow.compute as pc
+import pyarrow.parquet as pq
+import torch
+from datasets import Dataset
+from huggingface_hub import HfApi
+from huggingface_hub.errors import EntryNotFoundError, HfHubHTTPError
+from safetensors.torch import load_file
+
+from lerobot.common.datasets.utils import (
+    DEFAULT_CHUNK_SIZE,
+    DEFAULT_PARQUET_PATH,
+    DEFAULT_VIDEO_PATH,
+    EPISODES_PATH,
+    INFO_PATH,
+    STATS_PATH,
+    TASKS_PATH,
+    create_branch,
+    create_lerobot_dataset_card,
+    flatten_dict,
+    get_safe_version,
+    load_json,
+    unflatten_dict,
+    write_json,
+    write_jsonlines,
+)
+from lerobot.common.datasets.video_utils import (
+    VideoFrame,  # noqa: F401
+    get_image_pixel_channels,
+    get_video_info,
+)
+from lerobot.common.robot_devices.robots.configs import RobotConfig
+from lerobot.common.robot_devices.robots.utils import make_robot_config
+
+V16 = "v1.6"
+V20 = "v2.0"
+
+GITATTRIBUTES_REF = "aliberts/gitattributes_reference"
+V1_VIDEO_FILE = "{video_key}_episode_{episode_index:06d}.mp4"
+V1_INFO_PATH = "meta_data/info.json"
+V1_STATS_PATH = "meta_data/stats.safetensors"
+
+
+def parse_robot_config(robot_cfg: RobotConfig) -> tuple[str, dict]:
+    if robot_cfg.type in ["aloha", "koch"]:
+        state_names = [
+            f"{arm}_{motor}" if len(robot_cfg.follower_arms) > 1 else motor
+            for arm in robot_cfg.follower_arms
+            for motor in robot_cfg.follower_arms[arm].motors
+        ]
+        action_names = [
+            # f"{arm}_{motor}" for arm in ["left", "right"] for motor in robot_cfg["leader_arms"][arm]["motors"]
+            f"{arm}_{motor}" if len(robot_cfg.leader_arms) > 1 else motor
+            for arm in robot_cfg.leader_arms
+            for motor in robot_cfg.leader_arms[arm].motors
+        ]
+    # elif robot_cfg["robot_type"] == "stretch3": TODO
+    else:
+        raise NotImplementedError(
+            "Please provide robot_config={'robot_type': ..., 'names': ...} directly to convert_dataset()."
+        )
+
+    return {
+        "robot_type": robot_cfg.type,
+        "names": {
+            "observation.state": state_names,
+            "observation.effort": state_names,
+            "action": action_names,
+        },
+    }
+
+
+def convert_stats_to_json(v1_dir: Path, v2_dir: Path) -> None:
+    safetensor_path = v1_dir / V1_STATS_PATH
+    stats = load_file(safetensor_path)
+    serialized_stats = {key: value.tolist() for key, value in stats.items()}
+    serialized_stats = unflatten_dict(serialized_stats)
+
+    json_path = v2_dir / STATS_PATH
+    json_path.parent.mkdir(exist_ok=True, parents=True)
+    with open(json_path, "w") as f:
+        json.dump(serialized_stats, f, indent=4)
+
+    # Sanity check
+    with open(json_path) as f:
+        stats_json = json.load(f)
+
+    stats_json = flatten_dict(stats_json)
+    stats_json = {key: torch.tensor(value) for key, value in stats_json.items()}
+    for key in stats:
+        torch.testing.assert_close(stats_json[key], stats[key])
+
+
+def get_features_from_hf_dataset(
+    dataset: Dataset, robot_config: RobotConfig | None = None
+) -> dict[str, list]:
+    robot_config = parse_robot_config(robot_config)
+    features = {}
+    for key, ft in dataset.features.items():
+        if isinstance(ft, datasets.Value):
+            dtype = ft.dtype
+            shape = (1,)
+            names = None
+        if isinstance(ft, datasets.Sequence):
+            assert isinstance(ft.feature, datasets.Value)
+            dtype = ft.feature.dtype
+            shape = (ft.length,)
+            motor_names = (
+                robot_config["names"][key] if robot_config else [f"motor_{i}" for i in range(ft.length)]
+            )
+            assert len(motor_names) == shape[0]
+            names = {"motors": motor_names}
+        elif isinstance(ft, datasets.Image):
+            dtype = "image"
+            image = dataset[0][key]  # Assuming first row
+            channels = get_image_pixel_channels(image)
+            shape = (image.height, image.width, channels)
+            names = ["height", "width", "channels"]
+        elif ft._type == "VideoFrame":
+            dtype = "video"
+            shape = None  # Add shape later
+            names = ["height", "width", "channels"]
+
+        features[key] = {
+            "dtype": dtype,
+            "shape": shape,
+            "names": names,
+        }
+
+    return features
+
+
+def add_task_index_by_episodes(dataset: Dataset, tasks_by_episodes: dict) -> tuple[Dataset, list[str]]:
+    df = dataset.to_pandas()
+    tasks = list(set(tasks_by_episodes.values()))
+    tasks_to_task_index = {task: task_idx for task_idx, task in enumerate(tasks)}
+    episodes_to_task_index = {ep_idx: tasks_to_task_index[task] for ep_idx, task in tasks_by_episodes.items()}
+    df["task_index"] = df["episode_index"].map(episodes_to_task_index).astype(int)
+
+    features = dataset.features
+    features["task_index"] = datasets.Value(dtype="int64")
+    dataset = Dataset.from_pandas(df, features=features, split="train")
+    return dataset, tasks
+
+
+def add_task_index_from_tasks_col(
+    dataset: Dataset, tasks_col: str
+) -> tuple[Dataset, dict[str, list[str]], list[str]]:
+    df = dataset.to_pandas()
+
+    # HACK: This is to clean some of the instructions in our version of Open X datasets
+    prefix_to_clean = "tf.Tensor(b'"
+    suffix_to_clean = "', shape=(), dtype=string)"
+    df[tasks_col] = df[tasks_col].str.removeprefix(prefix_to_clean).str.removesuffix(suffix_to_clean)
+
+    # Create task_index col
+    tasks_by_episode = df.groupby("episode_index")[tasks_col].unique().apply(lambda x: x.tolist()).to_dict()
+    tasks = df[tasks_col].unique().tolist()
+    tasks_to_task_index = {task: idx for idx, task in enumerate(tasks)}
+    df["task_index"] = df[tasks_col].map(tasks_to_task_index).astype(int)
+
+    # Build the dataset back from df
+    features = dataset.features
+    features["task_index"] = datasets.Value(dtype="int64")
+    dataset = Dataset.from_pandas(df, features=features, split="train")
+    dataset = dataset.remove_columns(tasks_col)
+
+    return dataset, tasks, tasks_by_episode
+
+
+def split_parquet_by_episodes(
+    dataset: Dataset,
+    total_episodes: int,
+    total_chunks: int,
+    output_dir: Path,
+) -> list:
+    table = dataset.data.table
+    episode_lengths = []
+    for ep_chunk in range(total_chunks):
+        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
+        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
+        chunk_dir = "/".join(DEFAULT_PARQUET_PATH.split("/")[:-1]).format(episode_chunk=ep_chunk)
+        (output_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
+        for ep_idx in range(ep_chunk_start, ep_chunk_end):
+            ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
+            episode_lengths.insert(ep_idx, len(ep_table))
+            output_file = output_dir / DEFAULT_PARQUET_PATH.format(
+                episode_chunk=ep_chunk, episode_index=ep_idx
+            )
+            pq.write_table(ep_table, output_file)
+
+    return episode_lengths
+
+
+def move_videos(
+    repo_id: str,
+    video_keys: list[str],
+    total_episodes: int,
+    total_chunks: int,
+    work_dir: Path,
+    clean_gittatributes: Path,
+    branch: str = "main",
+) -> None:
+    """
+    HACK: Since HfApi() doesn't provide a way to move files directly in a repo, this function will run git
+    commands to fetch git lfs video files references to move them into subdirectories without having to
+    actually download them.
+    """
+    _lfs_clone(repo_id, work_dir, branch)
+
+    videos_moved = False
+    video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*.mp4")]
+    if len(video_files) == 0:
+        video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*/*/*.mp4")]
+        videos_moved = True  # Videos have already been moved
+
+    assert len(video_files) == total_episodes * len(video_keys)
+
+    lfs_untracked_videos = _get_lfs_untracked_videos(work_dir, video_files)
+
+    current_gittatributes = work_dir / ".gitattributes"
+    if not filecmp.cmp(current_gittatributes, clean_gittatributes, shallow=False):
+        fix_gitattributes(work_dir, current_gittatributes, clean_gittatributes)
+
+    if lfs_untracked_videos:
+        fix_lfs_video_files_tracking(work_dir, video_files)
+
+    if videos_moved:
+        return
+
+    video_dirs = sorted(work_dir.glob("videos*/"))
+    for ep_chunk in range(total_chunks):
+        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
+        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
+        for vid_key in video_keys:
+            chunk_dir = "/".join(DEFAULT_VIDEO_PATH.split("/")[:-1]).format(
+                episode_chunk=ep_chunk, video_key=vid_key
+            )
+            (work_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
+
+            for ep_idx in range(ep_chunk_start, ep_chunk_end):
+                target_path = DEFAULT_VIDEO_PATH.format(
+                    episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_idx
+                )
+                video_file = V1_VIDEO_FILE.format(video_key=vid_key, episode_index=ep_idx)
+                if len(video_dirs) == 1:
+                    video_path = video_dirs[0] / video_file
+                else:
+                    for dir in video_dirs:
+                        if (dir / video_file).is_file():
+                            video_path = dir / video_file
+                            break
+
+                video_path.rename(work_dir / target_path)
+
+    commit_message = "Move video files into chunk subdirectories"
+    subprocess.run(["git", "add", "."], cwd=work_dir, check=True)
+    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def fix_lfs_video_files_tracking(work_dir: Path, lfs_untracked_videos: list[str]) -> None:
+    """
+    HACK: This function fixes the tracking by git lfs which was not properly set on some repos. In that case,
+    there's no other option than to download the actual files and reupload them with lfs tracking.
+    """
+    for i in range(0, len(lfs_untracked_videos), 100):
+        files = lfs_untracked_videos[i : i + 100]
+        try:
+            subprocess.run(["git", "rm", "--cached", *files], cwd=work_dir, capture_output=True, check=True)
+        except subprocess.CalledProcessError as e:
+            print("git rm --cached ERROR:")
+            print(e.stderr)
+        subprocess.run(["git", "add", *files], cwd=work_dir, check=True)
+
+    commit_message = "Track video files with git lfs"
+    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def fix_gitattributes(work_dir: Path, current_gittatributes: Path, clean_gittatributes: Path) -> None:
+    shutil.copyfile(clean_gittatributes, current_gittatributes)
+    subprocess.run(["git", "add", ".gitattributes"], cwd=work_dir, check=True)
+    subprocess.run(["git", "commit", "-m", "Fix .gitattributes"], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:
+    subprocess.run(["git", "lfs", "install"], cwd=work_dir, check=True)
+    repo_url = f"https://huggingface.co/datasets/{repo_id}"
+    env = {"GIT_LFS_SKIP_SMUDGE": "1"}  # Prevent downloading LFS files
+    subprocess.run(
+        ["git", "clone", "--branch", branch, "--single-branch", "--depth", "1", repo_url, str(work_dir)],
+        check=True,
+        env=env,
+    )
+
+
+def _get_lfs_untracked_videos(work_dir: Path, video_files: list[str]) -> list[str]:
+    lfs_tracked_files = subprocess.run(
+        ["git", "lfs", "ls-files", "-n"], cwd=work_dir, capture_output=True, text=True, check=True
+    )
+    lfs_tracked_files = set(lfs_tracked_files.stdout.splitlines())
+    return [f for f in video_files if f not in lfs_tracked_files]
+
+
+def get_videos_info(repo_id: str, local_dir: Path, video_keys: list[str], branch: str) -> dict:
+    # Assumes first episode
+    video_files = [
+        DEFAULT_VIDEO_PATH.format(episode_chunk=0, video_key=vid_key, episode_index=0)
+        for vid_key in video_keys
+    ]
+    hub_api = HfApi()
+    hub_api.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", local_dir=local_dir, revision=branch, allow_patterns=video_files
+    )
+    videos_info_dict = {}
+    for vid_key, vid_path in zip(video_keys, video_files, strict=True):
+        videos_info_dict[vid_key] = get_video_info(local_dir / vid_path)
+
+    return videos_info_dict
+
+
+def convert_dataset(
+    repo_id: str,
+    local_dir: Path,
+    single_task: str | None = None,
+    tasks_path: Path | None = None,
+    tasks_col: Path | None = None,
+    robot_config: RobotConfig | None = None,
+    test_branch: str | None = None,
+    **card_kwargs,
+):
+    v1 = get_safe_version(repo_id, V16)
+    v1x_dir = local_dir / V16 / repo_id
+    v20_dir = local_dir / V20 / repo_id
+    v1x_dir.mkdir(parents=True, exist_ok=True)
+    v20_dir.mkdir(parents=True, exist_ok=True)
+
+    hub_api = HfApi()
+    hub_api.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", revision=v1, local_dir=v1x_dir, ignore_patterns="videos*/"
+    )
+    branch = "main"
+    if test_branch:
+        branch = test_branch
+        create_branch(repo_id=repo_id, branch=test_branch, repo_type="dataset")
+
+    metadata_v1 = load_json(v1x_dir / V1_INFO_PATH)
+    dataset = datasets.load_dataset("parquet", data_dir=v1x_dir / "data", split="train")
+    features = get_features_from_hf_dataset(dataset, robot_config)
+    video_keys = [key for key, ft in features.items() if ft["dtype"] == "video"]
+
+    if single_task and "language_instruction" in dataset.column_names:
+        logging.warning(
+            "'single_task' provided but 'language_instruction' tasks_col found. Using 'language_instruction'.",
+        )
+        single_task = None
+        tasks_col = "language_instruction"
+
+    # Episodes & chunks
+    episode_indices = sorted(dataset.unique("episode_index"))
+    total_episodes = len(episode_indices)
+    assert episode_indices == list(range(total_episodes))
+    total_videos = total_episodes * len(video_keys)
+    total_chunks = total_episodes // DEFAULT_CHUNK_SIZE
+    if total_episodes % DEFAULT_CHUNK_SIZE != 0:
+        total_chunks += 1
+
+    # Tasks
+    if single_task:
+        tasks_by_episodes = dict.fromkeys(episode_indices, single_task)
+        dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
+        tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
+    elif tasks_path:
+        tasks_by_episodes = load_json(tasks_path)
+        tasks_by_episodes = {int(ep_idx): task for ep_idx, task in tasks_by_episodes.items()}
+        dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
+        tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
+    elif tasks_col:
+        dataset, tasks, tasks_by_episodes = add_task_index_from_tasks_col(dataset, tasks_col)
+    else:
+        raise ValueError
+
+    assert set(tasks) == {task for ep_tasks in tasks_by_episodes.values() for task in ep_tasks}
+    tasks = [{"task_index": task_idx, "task": task} for task_idx, task in enumerate(tasks)]
+    write_jsonlines(tasks, v20_dir / TASKS_PATH)
+    features["task_index"] = {
+        "dtype": "int64",
+        "shape": (1,),
+        "names": None,
+    }
+
+    # Videos
+    if video_keys:
+        assert metadata_v1.get("video", False)
+        dataset = dataset.remove_columns(video_keys)
+        clean_gitattr = Path(
+            hub_api.hf_hub_download(
+                repo_id=GITATTRIBUTES_REF, repo_type="dataset", local_dir=local_dir, filename=".gitattributes"
+            )
+        ).absolute()
+        with tempfile.TemporaryDirectory() as tmp_video_dir:
+            move_videos(
+                repo_id, video_keys, total_episodes, total_chunks, Path(tmp_video_dir), clean_gitattr, branch
+            )
+        videos_info = get_videos_info(repo_id, v1x_dir, video_keys=video_keys, branch=branch)
+        for key in video_keys:
+            features[key]["shape"] = (
+                videos_info[key].pop("video.height"),
+                videos_info[key].pop("video.width"),
+                videos_info[key].pop("video.channels"),
+            )
+            features[key]["video_info"] = videos_info[key]
+            assert math.isclose(videos_info[key]["video.fps"], metadata_v1["fps"], rel_tol=1e-3)
+            if "encoding" in metadata_v1:
+                assert videos_info[key]["video.pix_fmt"] == metadata_v1["encoding"]["pix_fmt"]
+    else:
+        assert metadata_v1.get("video", 0) == 0
+        videos_info = None
+
+    # Split data into 1 parquet file by episode
+    episode_lengths = split_parquet_by_episodes(dataset, total_episodes, total_chunks, v20_dir)
+
+    if robot_config is not None:
+        robot_type = robot_config.type
+        repo_tags = [robot_type]
+    else:
+        robot_type = "unknown"
+        repo_tags = None
+
+    # Episodes
+    episodes = [
+        {"episode_index": ep_idx, "tasks": tasks_by_episodes[ep_idx], "length": episode_lengths[ep_idx]}
+        for ep_idx in episode_indices
+    ]
+    write_jsonlines(episodes, v20_dir / EPISODES_PATH)
+
+    # Assemble metadata v2.0
+    metadata_v2_0 = {
+        "codebase_version": V20,
+        "robot_type": robot_type,
+        "total_episodes": total_episodes,
+        "total_frames": len(dataset),
+        "total_tasks": len(tasks),
+        "total_videos": total_videos,
+        "total_chunks": total_chunks,
+        "chunks_size": DEFAULT_CHUNK_SIZE,
+        "fps": metadata_v1["fps"],
+        "splits": {"train": f"0:{total_episodes}"},
+        "data_path": DEFAULT_PARQUET_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if video_keys else None,
+        "features": features,
+    }
+    write_json(metadata_v2_0, v20_dir / INFO_PATH)
+    convert_stats_to_json(v1x_dir, v20_dir)
+    card = create_lerobot_dataset_card(tags=repo_tags, dataset_info=metadata_v2_0, **card_kwargs)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta_data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta", repo_type="dataset", revision=branch)
+
+    hub_api.upload_folder(
+        repo_id=repo_id,
+        path_in_repo="data",
+        folder_path=v20_dir / "data",
+        repo_type="dataset",
+        revision=branch,
+    )
+    hub_api.upload_folder(
+        repo_id=repo_id,
+        path_in_repo="meta",
+        folder_path=v20_dir / "meta",
+        repo_type="dataset",
+        revision=branch,
+    )
+
+    card.push_to_hub(repo_id=repo_id, repo_type="dataset", revision=branch)
+
+    if not test_branch:
+        create_branch(repo_id=repo_id, branch=V20, repo_type="dataset")
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    task_args = parser.add_mutually_exclusive_group(required=True)
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset (e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
+    )
+    task_args.add_argument(
+        "--single-task",
+        type=str,
+        help="A short but accurate description of the single task performed in the dataset.",
+    )
+    task_args.add_argument(
+        "--tasks-col",
+        type=str,
+        help="The name of the column containing language instructions",
+    )
+    task_args.add_argument(
+        "--tasks-path",
+        type=Path,
+        help="The path to a .json file containing one language instruction for each episode_index",
+    )
+    parser.add_argument(
+        "--robot",
+        type=str,
+        default=None,
+        help="Robot config used for the dataset during conversion (e.g. 'koch', 'aloha', 'so100', etc.)",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="Local directory to store the dataset during conversion. Defaults to /tmp/lerobot_dataset_v2",
+    )
+    parser.add_argument(
+        "--license",
+        type=str,
+        default="apache-2.0",
+        help="Repo license. Must be one of https://huggingface.co/docs/hub/repositories-licenses. Defaults to mit.",
+    )
+    parser.add_argument(
+        "--test-branch",
+        type=str,
+        default=None,
+        help="Repo branch to test your conversion first (e.g. 'v2.0.test')",
+    )
+
+    args = parser.parse_args()
+    if not args.local_dir:
+        args.local_dir = Path("/tmp/lerobot_dataset_v2")
+
+    if args.robot is not None:
+        robot_config = make_robot_config(args.robot)
+
+    del args.robot
+
+    convert_dataset(**vars(args), robot_config=robot_config)
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/v21/_remove_language_instruction.py

@@ -0,0 +1,87 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import traceback
+from pathlib import Path
+
+from datasets import get_dataset_config_info
+from huggingface_hub import HfApi
+
+from lerobot import available_datasets
+from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.common.datasets.utils import INFO_PATH, write_info
+from lerobot.common.datasets.v21.convert_dataset_v20_to_v21 import V20, SuppressWarnings
+
+LOCAL_DIR = Path("data/")
+
+hub_api = HfApi()
+
+
+def fix_dataset(repo_id: str) -> str:
+    if not hub_api.revision_exists(repo_id, V20, repo_type="dataset"):
+        return f"{repo_id}: skipped (not in {V20})."
+
+    dataset_info = get_dataset_config_info(repo_id, "default")
+    with SuppressWarnings():
+        lerobot_metadata = LeRobotDatasetMetadata(repo_id, revision=V20, force_cache_sync=True)
+
+    meta_features = {key for key, ft in lerobot_metadata.features.items() if ft["dtype"] != "video"}
+    parquet_features = set(dataset_info.features)
+
+    diff_parquet_meta = parquet_features - meta_features
+    diff_meta_parquet = meta_features - parquet_features
+
+    if diff_parquet_meta:
+        raise ValueError(f"In parquet not in info.json: {parquet_features - meta_features}")
+
+    if not diff_meta_parquet:
+        return f"{repo_id}: skipped (no diff)"
+
+    if diff_meta_parquet:
+        logging.warning(f"In info.json not in parquet: {meta_features - parquet_features}")
+        assert diff_meta_parquet == {"language_instruction"}
+        lerobot_metadata.features.pop("language_instruction")
+        write_info(lerobot_metadata.info, lerobot_metadata.root)
+        commit_info = hub_api.upload_file(
+            path_or_fileobj=lerobot_metadata.root / INFO_PATH,
+            path_in_repo=INFO_PATH,
+            repo_id=repo_id,
+            repo_type="dataset",
+            revision=V20,
+            commit_message="Remove 'language_instruction'",
+            create_pr=True,
+        )
+        return f"{repo_id}: success - PR: {commit_info.pr_url}"
+
+
+def batch_fix():
+    status = {}
+    LOCAL_DIR.mkdir(parents=True, exist_ok=True)
+    logfile = LOCAL_DIR / "fix_features_v20.txt"
+    for num, repo_id in enumerate(available_datasets):
+        print(f"\nConverting {repo_id} ({num}/{len(available_datasets)})")
+        print("---------------------------------------------------------")
+        try:
+            status = fix_dataset(repo_id)
+        except Exception:
+            status = f"{repo_id}: failed\n    {traceback.format_exc()}"
+
+        logging.info(status)
+        with open(logfile, "a") as file:
+            file.write(status + "\n")
+
+
+if __name__ == "__main__":
+    batch_fix()

lerobot/common/datasets/v21/batch_convert_dataset_v20_to_v21.py

@@ -0,0 +1,54 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script is for internal use to convert all datasets under the 'lerobot' hub user account to v2.1.
+"""
+
+import traceback
+from pathlib import Path
+
+from huggingface_hub import HfApi
+
+from lerobot import available_datasets
+from lerobot.common.datasets.v21.convert_dataset_v20_to_v21 import V21, convert_dataset
+
+LOCAL_DIR = Path("data/")
+
+
+def batch_convert():
+    status = {}
+    LOCAL_DIR.mkdir(parents=True, exist_ok=True)
+    logfile = LOCAL_DIR / "conversion_log_v21.txt"
+    hub_api = HfApi()
+    for num, repo_id in enumerate(available_datasets):
+        print(f"\nConverting {repo_id} ({num}/{len(available_datasets)})")
+        print("---------------------------------------------------------")
+        try:
+            if hub_api.revision_exists(repo_id, V21, repo_type="dataset"):
+                status = f"{repo_id}: success (already in {V21})."
+            else:
+                convert_dataset(repo_id)
+                status = f"{repo_id}: success."
+        except Exception:
+            status = f"{repo_id}: failed\n    {traceback.format_exc()}"
+
+        with open(logfile, "a") as file:
+            file.write(status + "\n")
+
+
+if __name__ == "__main__":
+    batch_convert()

lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py

@@ -0,0 +1,114 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 2.0 to
+2.1. It will:
+
+- Generate per-episodes stats and writes them in `episodes_stats.jsonl`
+- Check consistency between these new stats and the old ones.
+- Remove the deprecated `stats.json`.
+- Update codebase_version in `info.json`.
+- Push this new version to the hub on the 'main' branch and tags it with "v2.1".
+
+Usage:
+
+```bash
+python lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py \
+    --repo-id=aliberts/koch_tutorial
+```
+
+"""
+
+import argparse
+import logging
+
+from huggingface_hub import HfApi
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
+from lerobot.common.datasets.utils import EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
+from lerobot.common.datasets.v21.convert_stats import check_aggregate_stats, convert_stats
+
+V20 = "v2.0"
+V21 = "v2.1"
+
+
+class SuppressWarnings:
+    def __enter__(self):
+        self.previous_level = logging.getLogger().getEffectiveLevel()
+        logging.getLogger().setLevel(logging.ERROR)
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        logging.getLogger().setLevel(self.previous_level)
+
+
+def convert_dataset(
+    repo_id: str,
+    branch: str | None = None,
+    num_workers: int = 4,
+):
+    with SuppressWarnings():
+        dataset = LeRobotDataset(repo_id, revision=V20, force_cache_sync=True)
+
+    if (dataset.root / EPISODES_STATS_PATH).is_file():
+        (dataset.root / EPISODES_STATS_PATH).unlink()
+
+    convert_stats(dataset, num_workers=num_workers)
+    ref_stats = load_stats(dataset.root)
+    check_aggregate_stats(dataset, ref_stats)
+
+    dataset.meta.info["codebase_version"] = CODEBASE_VERSION
+    write_info(dataset.meta.info, dataset.root)
+
+    dataset.push_to_hub(branch=branch, tag_version=False, allow_patterns="meta/")
+
+    # delete old stats.json file
+    if (dataset.root / STATS_PATH).is_file:
+        (dataset.root / STATS_PATH).unlink()
+
+    hub_api = HfApi()
+    if hub_api.file_exists(
+        repo_id=dataset.repo_id, filename=STATS_PATH, revision=branch, repo_type="dataset"
+    ):
+        hub_api.delete_file(
+            path_in_repo=STATS_PATH, repo_id=dataset.repo_id, revision=branch, repo_type="dataset"
+        )
+
+    hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
+        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
+    )
+    parser.add_argument(
+        "--branch",
+        type=str,
+        default=None,
+        help="Repo branch to push your dataset. Defaults to the main branch.",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=4,
+        help="Number of workers for parallelizing stats compute. Defaults to 4.",
+    )
+
+    args = parser.parse_args()
+    convert_dataset(**vars(args))

lerobot/common/datasets/v21/convert_stats.py

@@ -0,0 +1,99 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+import numpy as np
+from tqdm import tqdm
+
+from lerobot.common.datasets.compute_stats import aggregate_stats, get_feature_stats, sample_indices
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.utils import write_episode_stats
+
+
+def sample_episode_video_frames(dataset: LeRobotDataset, episode_index: int, ft_key: str) -> np.ndarray:
+    ep_len = dataset.meta.episodes[episode_index]["length"]
+    sampled_indices = sample_indices(ep_len)
+    query_timestamps = dataset._get_query_timestamps(0.0, {ft_key: sampled_indices})
+    video_frames = dataset._query_videos(query_timestamps, episode_index)
+    return video_frames[ft_key].numpy()
+
+
+def convert_episode_stats(dataset: LeRobotDataset, ep_idx: int):
+    ep_start_idx = dataset.episode_data_index["from"][ep_idx]
+    ep_end_idx = dataset.episode_data_index["to"][ep_idx]
+    ep_data = dataset.hf_dataset.select(range(ep_start_idx, ep_end_idx))
+
+    ep_stats = {}
+    for key, ft in dataset.features.items():
+        if ft["dtype"] == "video":
+            # We sample only for videos
+            ep_ft_data = sample_episode_video_frames(dataset, ep_idx, key)
+        else:
+            ep_ft_data = np.array(ep_data[key])
+
+        axes_to_reduce = (0, 2, 3) if ft["dtype"] in ["image", "video"] else 0
+        keepdims = True if ft["dtype"] in ["image", "video"] else ep_ft_data.ndim == 1
+        ep_stats[key] = get_feature_stats(ep_ft_data, axis=axes_to_reduce, keepdims=keepdims)
+
+        if ft["dtype"] in ["image", "video"]:  # remove batch dim
+            ep_stats[key] = {
+                k: v if k == "count" else np.squeeze(v, axis=0) for k, v in ep_stats[key].items()
+            }
+
+    dataset.meta.episodes_stats[ep_idx] = ep_stats
+
+
+def convert_stats(dataset: LeRobotDataset, num_workers: int = 0):
+    assert dataset.episodes is None
+    print("Computing episodes stats")
+    total_episodes = dataset.meta.total_episodes
+    if num_workers > 0:
+        with ThreadPoolExecutor(max_workers=num_workers) as executor:
+            futures = {
+                executor.submit(convert_episode_stats, dataset, ep_idx): ep_idx
+                for ep_idx in range(total_episodes)
+            }
+            for future in tqdm(as_completed(futures), total=total_episodes):
+                future.result()
+    else:
+        for ep_idx in tqdm(range(total_episodes)):
+            convert_episode_stats(dataset, ep_idx)
+
+    for ep_idx in tqdm(range(total_episodes)):
+        write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
+
+
+def check_aggregate_stats(
+    dataset: LeRobotDataset,
+    reference_stats: dict[str, dict[str, np.ndarray]],
+    video_rtol_atol: tuple[float] = (1e-2, 1e-2),
+    default_rtol_atol: tuple[float] = (5e-6, 6e-5),
+):
+    """Verifies that the aggregated stats from episodes_stats are close to reference stats."""
+    agg_stats = aggregate_stats(list(dataset.meta.episodes_stats.values()))
+    for key, ft in dataset.features.items():
+        # These values might need some fine-tuning
+        if ft["dtype"] == "video":
+            # to account for image sub-sampling
+            rtol, atol = video_rtol_atol
+        else:
+            rtol, atol = default_rtol_atol
+
+        for stat, val in agg_stats[key].items():
+            if key in reference_stats and stat in reference_stats[key]:
+                err_msg = f"feature='{key}' stats='{stat}'"
+                np.testing.assert_allclose(
+                    val, reference_stats[key][stat], rtol=rtol, atol=atol, err_msg=err_msg
+                )

lerobot/common/datasets/video_utils.py

@@ -0,0 +1,432 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib
+import json
+import logging
+import subprocess
+import warnings
+from collections import OrderedDict
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, ClassVar
+
+import pyarrow as pa
+import torch
+import torchvision
+from datasets.features.features import register_feature
+from PIL import Image
+
+
+def get_safe_default_codec():
+    if importlib.util.find_spec("torchcodec"):
+        return "torchcodec"
+    else:
+        logging.warning(
+            "'torchcodec' is not available in your platform, falling back to 'pyav' as a default decoder"
+        )
+        return "pyav"
+
+
+def decode_video_frames(
+    video_path: Path | str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str | None = None,
+) -> torch.Tensor:
+    """
+    Decodes video frames using the specified backend.
+
+    Args:
+        video_path (Path): Path to the video file.
+        timestamps (list[float]): List of timestamps to extract frames.
+        tolerance_s (float): Allowed deviation in seconds for frame retrieval.
+        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav"..
+
+    Returns:
+        torch.Tensor: Decoded frames.
+
+    Currently supports torchcodec on cpu and pyav.
+    """
+    if backend is None:
+        backend = get_safe_default_codec()
+    if backend == "torchcodec":
+        return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s)
+    elif backend in ["pyav", "video_reader"]:
+        return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
+    else:
+        raise ValueError(f"Unsupported video backend: {backend}")
+
+
+def decode_video_frames_torchvision(
+    video_path: Path | str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str = "pyav",
+    log_loaded_timestamps: bool = False,
+) -> torch.Tensor:
+    """Loads frames associated to the requested timestamps of a video
+
+    The backend can be either "pyav" (default) or "video_reader".
+    "video_reader" requires installing torchvision from source, see:
+    https://github.com/pytorch/vision/blob/main/torchvision/csrc/io/decoder/gpu/README.rst
+    (note that you need to compile against ffmpeg<4.3)
+
+    While both use cpu, "video_reader" is supposedly faster than "pyav" but requires additional setup.
+    For more info on video decoding, see `benchmark/video/README.md`
+
+    See torchvision doc for more info on these two backends:
+    https://pytorch.org/vision/0.18/index.html?highlight=backend#torchvision.set_video_backend
+
+    Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
+    the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
+    that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
+    and all subsequent frames until reaching the requested frame. The number of key frames in a video
+    can be adjusted during encoding to take into account decoding time and video size in bytes.
+    """
+    video_path = str(video_path)
+
+    # set backend
+    keyframes_only = False
+    torchvision.set_video_backend(backend)
+    if backend == "pyav":
+        keyframes_only = True  # pyav doesnt support accuracte seek
+
+    # set a video stream reader
+    # TODO(rcadene): also load audio stream at the same time
+    reader = torchvision.io.VideoReader(video_path, "video")
+
+    # set the first and last requested timestamps
+    # Note: previous timestamps are usually loaded, since we need to access the previous key frame
+    first_ts = min(timestamps)
+    last_ts = max(timestamps)
+
+    # access closest key frame of the first requested frame
+    # Note: closest key frame timestamp is usually smaller than `first_ts` (e.g. key frame can be the first frame of the video)
+    # for details on what `seek` is doing see: https://pyav.basswood-io.com/docs/stable/api/container.html?highlight=inputcontainer#av.container.InputContainer.seek
+    reader.seek(first_ts, keyframes_only=keyframes_only)
+
+    # load all frames until last requested frame
+    loaded_frames = []
+    loaded_ts = []
+    for frame in reader:
+        current_ts = frame["pts"]
+        if log_loaded_timestamps:
+            logging.info(f"frame loaded at timestamp={current_ts:.4f}")
+        loaded_frames.append(frame["data"])
+        loaded_ts.append(current_ts)
+        if current_ts >= last_ts:
+            break
+
+    if backend == "pyav":
+        reader.container.close()
+
+    reader = None
+
+    query_ts = torch.tensor(timestamps)
+    loaded_ts = torch.tensor(loaded_ts)
+
+    # compute distances between each query timestamp and timestamps of all loaded frames
+    dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
+    min_, argmin_ = dist.min(1)
+
+    is_within_tol = min_ < tolerance_s
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
+        f"\nbackend: {backend}"
+    )
+
+    # get closest frames to the query timestamps
+    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
+    closest_ts = loaded_ts[argmin_]
+
+    if log_loaded_timestamps:
+        logging.info(f"{closest_ts=}")
+
+    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
+    closest_frames = closest_frames.type(torch.float32) / 255
+
+    assert len(timestamps) == len(closest_frames)
+    return closest_frames
+
+
+def decode_video_frames_torchcodec(
+    video_path: Path | str,
+    timestamps: list[float],
+    tolerance_s: float,
+    device: str = "cpu",
+    log_loaded_timestamps: bool = False,
+) -> torch.Tensor:
+    """Loads frames associated with the requested timestamps of a video using torchcodec.
+
+    Note: Setting device="cuda" outside the main process, e.g. in data loader workers, will lead to CUDA initialization errors.
+
+    Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
+    the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
+    that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
+    and all subsequent frames until reaching the requested frame. The number of key frames in a video
+    can be adjusted during encoding to take into account decoding time and video size in bytes.
+    """
+
+    if importlib.util.find_spec("torchcodec"):
+        from torchcodec.decoders import VideoDecoder
+    else:
+        raise ImportError("torchcodec is required but not available.")
+
+    # initialize video decoder
+    decoder = VideoDecoder(video_path, device=device, seek_mode="approximate")
+    loaded_frames = []
+    loaded_ts = []
+    # get metadata for frame information
+    metadata = decoder.metadata
+    average_fps = metadata.average_fps
+
+    # convert timestamps to frame indices
+    frame_indices = [round(ts * average_fps) for ts in timestamps]
+
+    # retrieve frames based on indices
+    frames_batch = decoder.get_frames_at(indices=frame_indices)
+
+    for frame, pts in zip(frames_batch.data, frames_batch.pts_seconds, strict=False):
+        loaded_frames.append(frame)
+        loaded_ts.append(pts.item())
+        if log_loaded_timestamps:
+            logging.info(f"Frame loaded at timestamp={pts:.4f}")
+
+    query_ts = torch.tensor(timestamps)
+    loaded_ts = torch.tensor(loaded_ts)
+
+    # compute distances between each query timestamp and loaded timestamps
+    dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
+    min_, argmin_ = dist.min(1)
+
+    is_within_tol = min_ < tolerance_s
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
+    )
+
+    # get closest frames to the query timestamps
+    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
+    closest_ts = loaded_ts[argmin_]
+
+    if log_loaded_timestamps:
+        logging.info(f"{closest_ts=}")
+
+    # convert to float32 in [0,1] range (channel first)
+    closest_frames = closest_frames.type(torch.float32) / 255
+
+    assert len(timestamps) == len(closest_frames)
+    return closest_frames
+
+
+def encode_video_frames(
+    imgs_dir: Path | str,
+    video_path: Path | str,
+    fps: int,
+    vcodec: str = "libsvtav1",
+    pix_fmt: str = "yuv420p",
+    g: int | None = 2,
+    crf: int | None = 30,
+    fast_decode: int = 0,
+    log_level: str | None = "error",
+    overwrite: bool = False,
+) -> None:
+    """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
+    video_path = Path(video_path)
+    imgs_dir = Path(imgs_dir)
+    video_path.parent.mkdir(parents=True, exist_ok=True)
+
+    ffmpeg_args = OrderedDict(
+        [
+            ("-f", "image2"),
+            ("-r", str(fps)),
+            ("-i", str(imgs_dir / "frame_%06d.png")),
+            ("-vcodec", vcodec),
+            ("-pix_fmt", pix_fmt),
+        ]
+    )
+
+    if g is not None:
+        ffmpeg_args["-g"] = str(g)
+
+    if crf is not None:
+        ffmpeg_args["-crf"] = str(crf)
+
+    if fast_decode:
+        key = "-svtav1-params" if vcodec == "libsvtav1" else "-tune"
+        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
+        ffmpeg_args[key] = value
+
+    if log_level is not None:
+        ffmpeg_args["-loglevel"] = str(log_level)
+
+    ffmpeg_args = [item for pair in ffmpeg_args.items() for item in pair]
+    if overwrite:
+        ffmpeg_args.append("-y")
+
+    ffmpeg_cmd = ["ffmpeg"] + ffmpeg_args + [str(video_path)]
+    # redirect stdin to subprocess.DEVNULL to prevent reading random keyboard inputs from terminal
+    subprocess.run(ffmpeg_cmd, check=True, stdin=subprocess.DEVNULL)
+
+    if not video_path.exists():
+        raise OSError(
+            f"Video encoding did not work. File not found: {video_path}. "
+            f"Try running the command manually to debug: `{''.join(ffmpeg_cmd)}`"
+        )
+
+
+@dataclass
+class VideoFrame:
+    # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
+    """
+    Provides a type for a dataset containing video frames.
+
+    Example:
+
+    ```python
+    data_dict = [{"image": {"path": "videos/episode_0.mp4", "timestamp": 0.3}}]
+    features = {"image": VideoFrame()}
+    Dataset.from_dict(data_dict, features=Features(features))
+    ```
+    """
+
+    pa_type: ClassVar[Any] = pa.struct({"path": pa.string(), "timestamp": pa.float32()})
+    _type: str = field(default="VideoFrame", init=False, repr=False)
+
+    def __call__(self):
+        return self.pa_type
+
+
+with warnings.catch_warnings():
+    warnings.filterwarnings(
+        "ignore",
+        "'register_feature' is experimental and might be subject to breaking changes in the future.",
+        category=UserWarning,
+    )
+    # to make VideoFrame available in HuggingFace `datasets`
+    register_feature(VideoFrame, "VideoFrame")
+
+
+def get_audio_info(video_path: Path | str) -> dict:
+    ffprobe_audio_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "a:0",
+        "-show_entries",
+        "stream=channels,codec_name,bit_rate,sample_rate,bit_depth,channel_layout,duration",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_audio_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+
+    info = json.loads(result.stdout)
+    audio_stream_info = info["streams"][0] if info.get("streams") else None
+    if audio_stream_info is None:
+        return {"has_audio": False}
+
+    # Return the information, defaulting to None if no audio stream is present
+    return {
+        "has_audio": True,
+        "audio.channels": audio_stream_info.get("channels", None),
+        "audio.codec": audio_stream_info.get("codec_name", None),
+        "audio.bit_rate": int(audio_stream_info["bit_rate"]) if audio_stream_info.get("bit_rate") else None,
+        "audio.sample_rate": int(audio_stream_info["sample_rate"])
+        if audio_stream_info.get("sample_rate")
+        else None,
+        "audio.bit_depth": audio_stream_info.get("bit_depth", None),
+        "audio.channel_layout": audio_stream_info.get("channel_layout", None),
+    }
+
+
+def get_video_info(video_path: Path | str) -> dict:
+    ffprobe_video_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "v:0",
+        "-show_entries",
+        "stream=r_frame_rate,width,height,codec_name,nb_frames,duration,pix_fmt",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_video_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+
+    info = json.loads(result.stdout)
+    video_stream_info = info["streams"][0]
+
+    # Calculate fps from r_frame_rate
+    r_frame_rate = video_stream_info["r_frame_rate"]
+    num, denom = map(int, r_frame_rate.split("/"))
+    fps = num / denom
+
+    pixel_channels = get_video_pixel_channels(video_stream_info["pix_fmt"])
+
+    video_info = {
+        "video.fps": fps,
+        "video.height": video_stream_info["height"],
+        "video.width": video_stream_info["width"],
+        "video.channels": pixel_channels,
+        "video.codec": video_stream_info["codec_name"],
+        "video.pix_fmt": video_stream_info["pix_fmt"],
+        "video.is_depth_map": False,
+        **get_audio_info(video_path),
+    }
+
+    return video_info
+
+
+def get_video_pixel_channels(pix_fmt: str) -> int:
+    if "gray" in pix_fmt or "depth" in pix_fmt or "monochrome" in pix_fmt:
+        return 1
+    elif "rgba" in pix_fmt or "yuva" in pix_fmt:
+        return 4
+    elif "rgb" in pix_fmt or "yuv" in pix_fmt:
+        return 3
+    else:
+        raise ValueError("Unknown format")
+
+
+def get_image_pixel_channels(image: Image):
+    if image.mode == "L":
+        return 1  # Grayscale
+    elif image.mode == "LA":
+        return 2  # Grayscale + Alpha
+    elif image.mode == "RGB":
+        return 3  # RGB
+    elif image.mode == "RGBA":
+        return 4  # RGBA
+    else:
+        raise ValueError("Unknown format")

lerobot/common/envs/__init__.py

@@ -0,0 +1,15 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv  # noqa: F401

lerobot/common/envs/configs.py

@@ -0,0 +1,156 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import abc
+from dataclasses import dataclass, field
+
+import draccus
+
+from lerobot.common.constants import ACTION, OBS_ENV, OBS_IMAGE, OBS_IMAGES, OBS_ROBOT
+from lerobot.configs.types import FeatureType, PolicyFeature
+
+
+@dataclass
+class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
+    task: str | None = None
+    fps: int = 30
+    features: dict[str, PolicyFeature] = field(default_factory=dict)
+    features_map: dict[str, str] = field(default_factory=dict)
+
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)
+
+    @abc.abstractproperty
+    def gym_kwargs(self) -> dict:
+        raise NotImplementedError()
+
+
+@EnvConfig.register_subclass("aloha")
+@dataclass
+class AlohaEnv(EnvConfig):
+    task: str = "AlohaInsertion-v0"
+    fps: int = 50
+    episode_length: int = 400
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            "action": ACTION,
+            "agent_pos": OBS_ROBOT,
+            "top": f"{OBS_IMAGE}.top",
+            "pixels/top": f"{OBS_IMAGES}.top",
+        }
+    )
+
+    def __post_init__(self):
+        if self.obs_type == "pixels":
+            self.features["top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
+        elif self.obs_type == "pixels_agent_pos":
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(14,))
+            self.features["pixels/top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "max_episode_steps": self.episode_length,
+        }
+
+
+@EnvConfig.register_subclass("pusht")
+@dataclass
+class PushtEnv(EnvConfig):
+    task: str = "PushT-v0"
+    fps: int = 10
+    episode_length: int = 300
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    visualization_width: int = 384
+    visualization_height: int = 384
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(2,)),
+            "agent_pos": PolicyFeature(type=FeatureType.STATE, shape=(2,)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            "action": ACTION,
+            "agent_pos": OBS_ROBOT,
+            "environment_state": OBS_ENV,
+            "pixels": OBS_IMAGE,
+        }
+    )
+
+    def __post_init__(self):
+        if self.obs_type == "pixels_agent_pos":
+            self.features["pixels"] = PolicyFeature(type=FeatureType.VISUAL, shape=(384, 384, 3))
+        elif self.obs_type == "environment_state_agent_pos":
+            self.features["environment_state"] = PolicyFeature(type=FeatureType.ENV, shape=(16,))
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "visualization_width": self.visualization_width,
+            "visualization_height": self.visualization_height,
+            "max_episode_steps": self.episode_length,
+        }
+
+
+@EnvConfig.register_subclass("xarm")
+@dataclass
+class XarmEnv(EnvConfig):
+    task: str = "XarmLift-v0"
+    fps: int = 15
+    episode_length: int = 200
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    visualization_width: int = 384
+    visualization_height: int = 384
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
+            "pixels": PolicyFeature(type=FeatureType.VISUAL, shape=(84, 84, 3)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            "action": ACTION,
+            "agent_pos": OBS_ROBOT,
+            "pixels": OBS_IMAGE,
+        }
+    )
+
+    def __post_init__(self):
+        if self.obs_type == "pixels_agent_pos":
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(4,))
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "visualization_width": self.visualization_width,
+            "visualization_height": self.visualization_height,
+            "max_episode_steps": self.episode_length,
+        }

lerobot/common/envs/factory.py

@@ -1,62 +1,69 @@
-from torchrl.envs.transforms import StepCounter, TransformedEnv
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib
+
+import gymnasium as gym
+
+from lerobot.common.envs.configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv
 
 
-def make_env(cfg, transform=None):
-    kwargs = {
-        "frame_skip": cfg.env.action_repeat,
-        "from_pixels": cfg.env.from_pixels,
-        "pixels_only": cfg.env.pixels_only,
-        "image_size": cfg.env.image_size,
-        # TODO(rcadene): do we want a specific eval_env_seed?
-        "seed": cfg.seed,
-        "num_prev_obs": cfg.n_obs_steps - 1,
-    }
-
-    if cfg.env.name == "simxarm":
-        from lerobot.common.envs.simxarm import SimxarmEnv
-
-        kwargs["task"] = cfg.env.task
-        clsfunc = SimxarmEnv
-    elif cfg.env.name == "pusht":
-        from lerobot.common.envs.pusht import PushtEnv
-
-        # assert kwargs["seed"] > 200, "Seed 0-200 are used for the demonstration dataset, so we don't want to seed the eval env with this range."
-
-        clsfunc = PushtEnv
+def make_env_config(env_type: str, **kwargs) -> EnvConfig:
+    if env_type == "aloha":
+        return AlohaEnv(**kwargs)
+    elif env_type == "pusht":
+        return PushtEnv(**kwargs)
+    elif env_type == "xarm":
+        return XarmEnv(**kwargs)
     else:
-        raise ValueError(cfg.env.name)
+        raise ValueError(f"Policy type '{env_type}' is not available.")
 
-    env = clsfunc(**kwargs)
 
-    # limit rollout to max_steps
-    env = TransformedEnv(env, StepCounter(max_steps=cfg.env.episode_length))
+def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> gym.vector.VectorEnv | None:
+    """Makes a gym vector environment according to the config.
 
-    if transform is not None:
-        # useful to add normalization
-        env.append_transform(transform)
+    Args:
+        cfg (EnvConfig): the config of the environment to instantiate.
+        n_envs (int, optional): The number of parallelized env to return. Defaults to 1.
+        use_async_envs (bool, optional): Whether to return an AsyncVectorEnv or a SyncVectorEnv. Defaults to
+            False.
+
+    Raises:
+        ValueError: if n_envs < 1
+        ModuleNotFoundError: If the requested env package is not installed
+
+    Returns:
+        gym.vector.VectorEnv: The parallelized gym.env instance.
+    """
+    if n_envs < 1:
+        raise ValueError("`n_envs must be at least 1")
+
+    package_name = f"gym_{cfg.type}"
+
+    try:
+        importlib.import_module(package_name)
+    except ModuleNotFoundError as e:
+        print(f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.type}]'`")
+        raise e
+
+    gym_handle = f"{package_name}/{cfg.task}"
+
+    # batched version of the env that returns an observation of shape (b, c)
+    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+    env = env_cls(
+        [lambda: gym.make(gym_handle, disable_env_checker=True, **cfg.gym_kwargs) for _ in range(n_envs)]
+    )
 
     return env
-
-
-# def make_env(env_name, frame_skip, device, is_test=False):
-#     env = GymEnv(
-#         env_name,
-#         frame_skip=frame_skip,
-#         from_pixels=True,
-#         pixels_only=False,
-#         device=device,
-#     )
-#     env = TransformedEnv(env)
-#     env.append_transform(NoopResetEnv(noops=30, random=True))
-#     if not is_test:
-#         env.append_transform(EndOfLifeTransform())
-#         env.append_transform(RewardClipping(-1, 1))
-#     env.append_transform(ToTensorImage())
-#     env.append_transform(GrayScale())
-#     env.append_transform(Resize(84, 84))
-#     env.append_transform(CatFrames(N=4, dim=-3))
-#     env.append_transform(RewardSum())
-#     env.append_transform(StepCounter(max_steps=4500))
-#     env.append_transform(DoubleToFloat())
-#     env.append_transform(VecNorm(in_keys=["pixels"]))
-#     return env

lerobot/common/envs/pusht.py

@@ -1,242 +0,0 @@
-import importlib
-from collections import deque
-from typing import Optional
-
-import einops
-import torch
-from tensordict import TensorDict
-from torchrl.data.tensor_specs import (
-    BoundedTensorSpec,
-    CompositeSpec,
-    DiscreteTensorSpec,
-    UnboundedContinuousTensorSpec,
-)
-from torchrl.envs import EnvBase
-from torchrl.envs.libs.gym import _gym_to_torchrl_spec_transform
-
-from lerobot.common.utils import set_seed
-
-_has_gym = importlib.util.find_spec("gym") is not None
-_has_diffpolicy = importlib.util.find_spec("diffusion_policy") is not None and _has_gym
-
-
-class PushtEnv(EnvBase):
-    def __init__(
-        self,
-        frame_skip: int = 1,
-        from_pixels: bool = False,
-        pixels_only: bool = False,
-        image_size=None,
-        seed=1337,
-        device="cpu",
-        num_prev_obs=0,
-        num_prev_action=0,
-    ):
-        super().__init__(device=device, batch_size=[])
-        self.frame_skip = frame_skip
-        self.from_pixels = from_pixels
-        self.pixels_only = pixels_only
-        self.image_size = image_size
-        self.num_prev_obs = num_prev_obs
-        self.num_prev_action = num_prev_action
-
-        if pixels_only:
-            assert from_pixels
-        if from_pixels:
-            assert image_size
-
-        if not _has_diffpolicy:
-            raise ImportError("Cannot import diffusion_policy.")
-        if not _has_gym:
-            raise ImportError("Cannot import gym.")
-
-        # TODO(rcadene) (PushTEnv is similar to PushTImageEnv, but without the image rendering, it's faster to iterate on)
-        # from diffusion_policy.env.pusht.pusht_env import PushTEnv
-
-        if not from_pixels:
-            raise NotImplementedError("Use PushTEnv, instead of PushTImageEnv")
-        from diffusion_policy.env.pusht.pusht_image_env import PushTImageEnv
-
-        self._env = PushTImageEnv(render_size=self.image_size)
-
-        self._make_spec()
-        self._current_seed = self.set_seed(seed)
-
-        if self.num_prev_obs > 0:
-            self._prev_obs_image_queue = deque(maxlen=self.num_prev_obs)
-            self._prev_obs_state_queue = deque(maxlen=self.num_prev_obs)
-        if self.num_prev_action > 0:
-            raise NotImplementedError()
-            # self._prev_action_queue = deque(maxlen=self.num_prev_action)
-
-    def render(self, mode="rgb_array", width=384, height=384):
-        if width != height:
-            raise NotImplementedError()
-        tmp = self._env.render_size
-        self._env.render_size = width
-        out = self._env.render(mode)
-        self._env.render_size = tmp
-        return out
-
-    def _format_raw_obs(self, raw_obs):
-        if self.from_pixels:
-            image = torch.from_numpy(raw_obs["image"])
-            obs = {"image": image}
-
-            if not self.pixels_only:
-                obs["state"] = torch.from_numpy(raw_obs["agent_pos"]).type(torch.float32)
-        else:
-            # TODO:
-            obs = {"state": torch.from_numpy(raw_obs["observation"]).type(torch.float32)}
-
-        return obs
-
-    def _reset(self, tensordict: Optional[TensorDict] = None):
-        td = tensordict
-        if td is None or td.is_empty():
-            # we need to handle seed iteration, since self._env.reset() rely an internal _seed.
-            self._current_seed += 1
-            self.set_seed(self._current_seed)
-            raw_obs = self._env.reset()
-            assert self._current_seed == self._env._seed
-
-            obs = self._format_raw_obs(raw_obs)
-
-            if self.num_prev_obs > 0:
-                stacked_obs = {}
-                if "image" in obs:
-                    self._prev_obs_image_queue = deque(
-                        [obs["image"]] * (self.num_prev_obs + 1), maxlen=(self.num_prev_obs + 1)
-                    )
-                    stacked_obs["image"] = torch.stack(list(self._prev_obs_image_queue))
-                if "state" in obs:
-                    self._prev_obs_state_queue = deque(
-                        [obs["state"]] * (self.num_prev_obs + 1), maxlen=(self.num_prev_obs + 1)
-                    )
-                    stacked_obs["state"] = torch.stack(list(self._prev_obs_state_queue))
-                obs = stacked_obs
-
-            td = TensorDict(
-                {
-                    "observation": TensorDict(obs, batch_size=[]),
-                    "done": torch.tensor([False], dtype=torch.bool),
-                },
-                batch_size=[],
-            )
-        else:
-            raise NotImplementedError()
-        return td
-
-    def _step(self, tensordict: TensorDict):
-        td = tensordict
-        action = td["action"].numpy()
-        # step expects shape=(4,) so we pad if necessary
-        # TODO(rcadene): add info["is_success"] and info["success"] ?
-        sum_reward = 0
-
-        if action.ndim == 1:
-            action = einops.repeat(action, "c -> t c", t=self.frame_skip)
-        else:
-            if self.frame_skip > 1:
-                raise NotImplementedError()
-
-        num_action_steps = action.shape[0]
-        for i in range(num_action_steps):
-            raw_obs, reward, done, info = self._env.step(action[i])
-            sum_reward += reward
-
-            obs = self._format_raw_obs(raw_obs)
-
-            if self.num_prev_obs > 0:
-                stacked_obs = {}
-                if "image" in obs:
-                    self._prev_obs_image_queue.append(obs["image"])
-                    stacked_obs["image"] = torch.stack(list(self._prev_obs_image_queue))
-                if "state" in obs:
-                    self._prev_obs_state_queue.append(obs["state"])
-                    stacked_obs["state"] = torch.stack(list(self._prev_obs_state_queue))
-                obs = stacked_obs
-
-        td = TensorDict(
-            {
-                "observation": TensorDict(obs, batch_size=[]),
-                "reward": torch.tensor([sum_reward], dtype=torch.float32),
-                # succes and done are true when coverage > self.success_threshold in env
-                "done": torch.tensor([done], dtype=torch.bool),
-                "success": torch.tensor([done], dtype=torch.bool),
-            },
-            batch_size=[],
-        )
-        return td
-
-    def _make_spec(self):
-        obs = {}
-        if self.from_pixels:
-            image_shape = (3, self.image_size, self.image_size)
-            if self.num_prev_obs > 0:
-                image_shape = (self.num_prev_obs + 1, *image_shape)
-
-            obs["image"] = BoundedTensorSpec(
-                low=0,
-                high=1,
-                shape=image_shape,
-                dtype=torch.float32,
-                device=self.device,
-            )
-            if not self.pixels_only:
-                state_shape = self._env.observation_space["agent_pos"].shape
-                if self.num_prev_obs > 0:
-                    state_shape = (self.num_prev_obs + 1, *state_shape)
-
-                obs["state"] = BoundedTensorSpec(
-                    low=0,
-                    high=512,
-                    shape=state_shape,
-                    dtype=torch.float32,
-                    device=self.device,
-                )
-        else:
-            # TODO(rcadene): add observation_space achieved_goal and desired_goal?
-            state_shape = self._env.observation_space["observation"].shape
-            if self.num_prev_obs > 0:
-                state_shape = (self.num_prev_obs + 1, *state_shape)
-
-            obs["state"] = UnboundedContinuousTensorSpec(
-                # TODO:
-                shape=state_shape,
-                dtype=torch.float32,
-                device=self.device,
-            )
-        self.observation_spec = CompositeSpec({"observation": obs})
-
-        self.action_spec = _gym_to_torchrl_spec_transform(
-            self._env.action_space,
-            device=self.device,
-        )
-
-        self.reward_spec = UnboundedContinuousTensorSpec(
-            shape=(1,),
-            dtype=torch.float32,
-            device=self.device,
-        )
-
-        self.done_spec = CompositeSpec(
-            {
-                "done": DiscreteTensorSpec(
-                    2,
-                    shape=(1,),
-                    dtype=torch.bool,
-                    device=self.device,
-                ),
-                "success": DiscreteTensorSpec(
-                    2,
-                    shape=(1,),
-                    dtype=torch.bool,
-                    device=self.device,
-                ),
-            }
-        )
-
-    def _set_seed(self, seed: Optional[int]):
-        set_seed(seed)
-        self._env.seed(seed)

lerobot/common/envs/simxarm.py

@@ -1,181 +0,0 @@
-import importlib
-from typing import Optional
-
-import numpy as np
-import torch
-from tensordict import TensorDict
-from torchrl.data.tensor_specs import (
-    BoundedTensorSpec,
-    CompositeSpec,
-    DiscreteTensorSpec,
-    UnboundedContinuousTensorSpec,
-)
-from torchrl.envs import EnvBase
-from torchrl.envs.libs.gym import _gym_to_torchrl_spec_transform
-
-from lerobot.common.utils import set_seed
-
-MAX_NUM_ACTIONS = 4
-
-_has_gym = importlib.util.find_spec("gym") is not None
-_has_simxarm = importlib.util.find_spec("simxarm") is not None and _has_gym
-
-
-class SimxarmEnv(EnvBase):
-    def __init__(
-        self,
-        task,
-        frame_skip: int = 1,
-        from_pixels: bool = False,
-        pixels_only: bool = False,
-        image_size=None,
-        seed=1337,
-        device="cpu",
-    ):
-        super().__init__(device=device, batch_size=[])
-        self.task = task
-        self.frame_skip = frame_skip
-        self.from_pixels = from_pixels
-        self.pixels_only = pixels_only
-        self.image_size = image_size
-
-        if pixels_only:
-            assert from_pixels
-        if from_pixels:
-            assert image_size
-
-        if not _has_simxarm:
-            raise ImportError("Cannot import simxarm.")
-        if not _has_gym:
-            raise ImportError("Cannot import gym.")
-
-        import gym
-        from simxarm import TASKS
-
-        if self.task not in TASKS:
-            raise ValueError(f"Unknown task {self.task}. Must be one of {list(TASKS.keys())}")
-
-        self._env = TASKS[self.task]["env"]()
-
-        num_actions = len(TASKS[self.task]["action_space"])
-        self._action_space = gym.spaces.Box(low=-1.0, high=1.0, shape=(num_actions,))
-        self._action_padding = np.zeros((MAX_NUM_ACTIONS - num_actions), dtype=np.float32)
-        if "w" not in TASKS[self.task]["action_space"]:
-            self._action_padding[-1] = 1.0
-
-        self._make_spec()
-        self.set_seed(seed)
-
-    def render(self, mode="rgb_array", width=384, height=384):
-        return self._env.render(mode, width=width, height=height)
-
-    def _format_raw_obs(self, raw_obs):
-        if self.from_pixels:
-            image = self.render(mode="rgb_array", width=self.image_size, height=self.image_size)
-            image = image.transpose(2, 0, 1)  # (H, W, C) -> (C, H, W)
-            image = torch.tensor(image.copy(), dtype=torch.uint8)
-
-            obs = {"image": image}
-
-            if not self.pixels_only:
-                obs["state"] = torch.tensor(self._env.robot_state, dtype=torch.float32)
-        else:
-            obs = {"state": torch.tensor(raw_obs["observation"], dtype=torch.float32)}
-
-        obs = TensorDict(obs, batch_size=[])
-        return obs
-
-    def _reset(self, tensordict: Optional[TensorDict] = None):
-        td = tensordict
-        if td is None or td.is_empty():
-            raw_obs = self._env.reset()
-
-            td = TensorDict(
-                {
-                    "observation": self._format_raw_obs(raw_obs),
-                    "done": torch.tensor([False], dtype=torch.bool),
-                },
-                batch_size=[],
-            )
-        else:
-            raise NotImplementedError()
-        return td
-
-    def _step(self, tensordict: TensorDict):
-        td = tensordict
-        action = td["action"].numpy()
-        # step expects shape=(4,) so we pad if necessary
-        action = np.concatenate([action, self._action_padding])
-        # TODO(rcadene): add info["is_success"] and info["success"] ?
-        sum_reward = 0
-        for _ in range(self.frame_skip):
-            raw_obs, reward, done, info = self._env.step(action)
-            sum_reward += reward
-
-        td = TensorDict(
-            {
-                "observation": self._format_raw_obs(raw_obs),
-                "reward": torch.tensor([sum_reward], dtype=torch.float32),
-                "done": torch.tensor([done], dtype=torch.bool),
-                "success": torch.tensor([info["success"]], dtype=torch.bool),
-            },
-            batch_size=[],
-        )
-        return td
-
-    def _make_spec(self):
-        obs = {}
-        if self.from_pixels:
-            obs["image"] = BoundedTensorSpec(
-                low=0,
-                high=255,
-                shape=(3, self.image_size, self.image_size),
-                dtype=torch.uint8,
-                device=self.device,
-            )
-            if not self.pixels_only:
-                obs["state"] = UnboundedContinuousTensorSpec(
-                    shape=(len(self._env.robot_state),),
-                    dtype=torch.float32,
-                    device=self.device,
-                )
-        else:
-            # TODO(rcadene): add observation_space achieved_goal and desired_goal?
-            obs["state"] = UnboundedContinuousTensorSpec(
-                shape=self._env.observation_space["observation"].shape,
-                dtype=torch.float32,
-                device=self.device,
-            )
-        self.observation_spec = CompositeSpec({"observation": obs})
-
-        self.action_spec = _gym_to_torchrl_spec_transform(
-            self._action_space,
-            device=self.device,
-        )
-
-        self.reward_spec = UnboundedContinuousTensorSpec(
-            shape=(1,),
-            dtype=torch.float32,
-            device=self.device,
-        )
-
-        self.done_spec = CompositeSpec(
-            {
-                "done": DiscreteTensorSpec(
-                    2,
-                    shape=(1,),
-                    dtype=torch.bool,
-                    device=self.device,
-                ),
-                "success": DiscreteTensorSpec(
-                    2,
-                    shape=(1,),
-                    dtype=torch.bool,
-                    device=self.device,
-                ),
-            }
-        )
-
-    def _set_seed(self, seed: Optional[int]):
-        set_seed(seed)
-        self._env.seed(seed)

lerobot/common/envs/transforms.py

@@ -1,92 +0,0 @@
-from typing import Sequence
-
-from tensordict import TensorDictBase
-from tensordict.nn import dispatch
-from tensordict.utils import NestedKey
-from torchrl.envs.transforms import ObservationTransform, Transform
-
-
-class Prod(ObservationTransform):
-    def __init__(self, in_keys: Sequence[NestedKey], prod: float):
-        super().__init__()
-        self.in_keys = in_keys
-        self.prod = prod
-
-    def _call(self, td):
-        for key in self.in_keys:
-            td[key] *= self.prod
-        return td
-
-    def transform_observation_spec(self, obs_spec):
-        for key in self.in_keys:
-            obs_spec[key].space.high *= self.prod
-        return obs_spec
-
-
-class NormalizeTransform(Transform):
-    invertible = True
-
-    def __init__(
-        self,
-        stats: TensorDictBase,
-        in_keys: Sequence[NestedKey] = None,
-        out_keys: Sequence[NestedKey] | None = None,
-        in_keys_inv: Sequence[NestedKey] | None = None,
-        out_keys_inv: Sequence[NestedKey] | None = None,
-        mode="mean_std",
-    ):
-        if out_keys is None:
-            out_keys = in_keys
-        if in_keys_inv is None:
-            in_keys_inv = out_keys
-        if out_keys_inv is None:
-            out_keys_inv = in_keys
-        super().__init__(
-            in_keys=in_keys, out_keys=out_keys, in_keys_inv=in_keys_inv, out_keys_inv=out_keys_inv
-        )
-        self.stats = stats
-        assert mode in ["mean_std", "min_max"]
-        self.mode = mode
-
-    def _reset(self, tensordict: TensorDictBase, tensordict_reset: TensorDictBase) -> TensorDictBase:
-        # _reset is called once when the environment reset to normalize the first observation
-        tensordict_reset = self._call(tensordict_reset)
-        return tensordict_reset
-
-    @dispatch(source="in_keys", dest="out_keys")
-    def forward(self, tensordict: TensorDictBase) -> TensorDictBase:
-        return self._call(tensordict)
-
-    def _call(self, td: TensorDictBase) -> TensorDictBase:
-        for inkey, outkey in zip(self.in_keys, self.out_keys, strict=False):
-            # TODO(rcadene): don't know how to do `inkey not in td`
-            if td.get(inkey, None) is None:
-                continue
-            if self.mode == "mean_std":
-                mean = self.stats[inkey]["mean"]
-                std = self.stats[inkey]["std"]
-                td[outkey] = (td[inkey] - mean) / (std + 1e-8)
-            else:
-                min = self.stats[inkey]["min"]
-                max = self.stats[inkey]["max"]
-                # normalize to [0,1]
-                td[outkey] = (td[inkey] - min) / (max - min)
-                # normalize to [-1, 1]
-                td[outkey] = td[outkey] * 2 - 1
-        return td
-
-    def _inv_call(self, td: TensorDictBase) -> TensorDictBase:
-        for inkey, outkey in zip(self.in_keys_inv, self.out_keys_inv, strict=False):
-            # TODO(rcadene): don't know how to do `inkey not in td`
-            if td.get(inkey, None) is None:
-                continue
-            if self.mode == "mean_std":
-                mean = self.stats[inkey]["mean"]
-                std = self.stats[inkey]["std"]
-                td[outkey] = td[inkey] * std + mean
-            else:
-                min = self.stats[inkey]["min"]
-                max = self.stats[inkey]["max"]
-                td[outkey] = (td[inkey] + 1) / 2
-                td[outkey] = td[outkey] * (max - min) + min
-        return td

lerobot/common/envs/utils.py

@@ -0,0 +1,127 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import warnings
+from typing import Any
+
+import einops
+import gymnasium as gym
+import numpy as np
+import torch
+from torch import Tensor
+
+from lerobot.common.envs.configs import EnvConfig
+from lerobot.common.utils.utils import get_channel_first_image_shape
+from lerobot.configs.types import FeatureType, PolicyFeature
+
+
+def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
+    # TODO(aliberts, rcadene): refactor this to use features from the environment (no hardcoding)
+    """Convert environment observation to LeRobot format observation.
+    Args:
+        observation: Dictionary of observation batches from a Gym vector environment.
+    Returns:
+        Dictionary of observation batches with keys renamed to LeRobot format and values as tensors.
+    """
+    # map to expected inputs for the policy
+    return_observations = {}
+    if "pixels" in observations:
+        if isinstance(observations["pixels"], dict):
+            imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
+        else:
+            imgs = {"observation.image": observations["pixels"]}
+
+        for imgkey, img in imgs.items():
+            # TODO(aliberts, rcadene): use transforms.ToTensor()?
+            img = torch.from_numpy(img)
+
+            # sanity check that images are channel last
+            _, h, w, c = img.shape
+            assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
+
+            # sanity check that images are uint8
+            assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
+
+            # convert to channel first of type float32 in range [0,1]
+            img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
+            img = img.type(torch.float32)
+            img /= 255
+
+            return_observations[imgkey] = img
+
+    if "environment_state" in observations:
+        return_observations["observation.environment_state"] = torch.from_numpy(
+            observations["environment_state"]
+        ).float()
+
+    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
+    # requirement for "agent_pos"
+    return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
+    return return_observations
+
+
+def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
+    # TODO(aliberts, rcadene): remove this hardcoding of keys and just use the nested keys as is
+    # (need to also refactor preprocess_observation and externalize normalization from policies)
+    policy_features = {}
+    for key, ft in env_cfg.features.items():
+        if ft.type is FeatureType.VISUAL:
+            if len(ft.shape) != 3:
+                raise ValueError(f"Number of dimensions of {key} != 3 (shape={ft.shape})")
+
+            shape = get_channel_first_image_shape(ft.shape)
+            feature = PolicyFeature(type=ft.type, shape=shape)
+        else:
+            feature = ft
+
+        policy_key = env_cfg.features_map[key]
+        policy_features[policy_key] = feature
+
+    return policy_features
+
+
+def are_all_envs_same_type(env: gym.vector.VectorEnv) -> bool:
+    first_type = type(env.envs[0])  # Get type of first env
+    return all(type(e) is first_type for e in env.envs)  # Fast type check
+
+
+def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
+    with warnings.catch_warnings():
+        warnings.simplefilter("once", UserWarning)  # Apply filter only in this function
+
+        if not (hasattr(env.envs[0], "task_description") and hasattr(env.envs[0], "task")):
+            warnings.warn(
+                "The environment does not have 'task_description' and 'task'. Some policies require these features.",
+                UserWarning,
+                stacklevel=2,
+            )
+        if not are_all_envs_same_type(env):
+            warnings.warn(
+                "The environments have different types. Make sure you infer the right task from each environment. Empty task will be passed instead.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+
+def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
+    """Adds task feature to the observation dict with respect to the first environment attribute."""
+    if hasattr(env.envs[0], "task_description"):
+        observation["task"] = env.call("task_description")
+    elif hasattr(env.envs[0], "task"):
+        observation["task"] = env.call("task")
+    else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
+        num_envs = observation[list(observation.keys())[0]].shape[0]
+        observation["task"] = ["" for _ in range(num_envs)]
+    return observation

lerobot/common/logger.py

@@ -1,110 +0,0 @@
-import logging
-import os
-from pathlib import Path
-
-from omegaconf import OmegaConf
-from termcolor import colored
-
-
-def log_output_dir(out_dir):
-    logging.info(colored("Output dir:", "yellow", attrs=["bold"]) + f" {out_dir}")
-
-
-def cfg_to_group(cfg, return_list=False):
-    """Return a wandb-safe group name for logging. Optionally returns group name as list."""
-    # lst = [cfg.task, cfg.modality, re.sub("[^0-9a-zA-Z]+", "-", cfg.exp_name)]
-    lst = [
-        f"env:{cfg.env.name}",
-        f"seed:{cfg.seed}",
-    ]
-    return lst if return_list else "-".join(lst)
-
-
-class Logger:
-    """Primary logger object. Logs either locally or using wandb."""
-
-    def __init__(self, log_dir, job_name, cfg):
-        self._log_dir = Path(log_dir)
-        self._log_dir.mkdir(parents=True, exist_ok=True)
-        self._job_name = job_name
-        self._model_dir = self._log_dir / "models"
-        self._buffer_dir = self._log_dir / "buffers"
-        self._save_model = cfg.save_model
-        self._save_buffer = cfg.save_buffer
-        self._group = cfg_to_group(cfg)
-        self._seed = cfg.seed
-        self._cfg = cfg
-        self._eval = []
-        project = cfg.get("wandb", {}).get("project")
-        entity = cfg.get("wandb", {}).get("entity")
-        enable_wandb = cfg.get("wandb", {}).get("enable", False)
-        run_offline = not enable_wandb or not project or not entity
-        if run_offline:
-            logging.info(colored("Logs will be saved locally.", "yellow", attrs=["bold"]))
-            self._wandb = None
-        else:
-            os.environ["WANDB_SILENT"] = "true"
-            import wandb
-
-            wandb.init(
-                project=project,
-                entity=entity,
-                name=job_name,
-                notes=cfg.get("wandb", {}).get("notes"),
-                # group=self._group,
-                tags=cfg_to_group(cfg, return_list=True),
-                dir=self._log_dir,
-                config=OmegaConf.to_container(cfg, resolve=True),
-                # TODO(rcadene): try set to True
-                save_code=False,
-                # TODO(rcadene): split train and eval, and run async eval with job_type="eval"
-                job_type="train_eval",
-                # TODO(rcadene): add resume option
-                resume=None,
-            )
-            print(colored("Logs will be synced with wandb.", "blue", attrs=["bold"]))
-            self._wandb = wandb
-
-    def save_model(self, policy, identifier):
-        if self._save_model:
-            self._model_dir.mkdir(parents=True, exist_ok=True)
-            fp = self._model_dir / f"{str(identifier)}.pt"
-            policy.save(fp)
-            if self._wandb:
-                artifact = self._wandb.Artifact(
-                    self._group + "-" + str(self._seed) + "-" + str(identifier),
-                    type="model",
-                )
-                artifact.add_file(fp)
-                self._wandb.log_artifact(artifact)
-
-    def save_buffer(self, buffer, identifier):
-        self._buffer_dir.mkdir(parents=True, exist_ok=True)
-        fp = self._buffer_dir / f"{str(identifier)}.pkl"
-        buffer.save(fp)
-        if self._wandb:
-            artifact = self._wandb.Artifact(
-                self._group + "-" + str(self._seed) + "-" + str(identifier),
-                type="buffer",
-            )
-            artifact.add_file(fp)
-            self._wandb.log_artifact(artifact)
-
-    def finish(self, agent, buffer):
-        if self._save_model:
-            self.save_model(agent, identifier="final")
-        if self._save_buffer:
-            self.save_buffer(buffer, identifier="buffer")
-        if self._wandb:
-            self._wandb.finish()
-
-    def log_dict(self, d, step, mode="train"):
-        assert mode in {"train", "eval"}
-        if self._wandb is not None:
-            for k, v in d.items():
-                self._wandb.log({f"{mode}/{k}": v}, step=step)
-
-    def log_video(self, video, step, mode="train"):
-        assert mode in {"train", "eval"}
-        wandb_video = self._wandb.Video(video, fps=self._cfg.fps, format="mp4")
-        self._wandb.log({f"{mode}/video": wandb_video}, step=step)

lerobot/common/optim/__init__.py

@@ -0,0 +1,15 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .optimizers import OptimizerConfig as OptimizerConfig

lerobot/common/optim/factory.py

@@ -0,0 +1,40 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from torch.optim import Optimizer
+from torch.optim.lr_scheduler import LRScheduler
+
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+from lerobot.configs.train import TrainPipelineConfig
+
+
+def make_optimizer_and_scheduler(
+    cfg: TrainPipelineConfig, policy: PreTrainedPolicy
+) -> tuple[Optimizer, LRScheduler | None]:
+    """Generates the optimizer and scheduler based on configs.
+
+    Args:
+        cfg (TrainPipelineConfig): The training config that contains optimizer and scheduler configs
+        policy (PreTrainedPolicy): The policy config from which parameters and presets must be taken from.
+
+    Returns:
+        tuple[Optimizer, LRScheduler | None]: The couple (Optimizer, Scheduler). Scheduler can be `None`.
+    """
+    params = policy.get_optim_params() if cfg.use_policy_training_preset else policy.parameters()
+    optimizer = cfg.optimizer.build(params)
+    lr_scheduler = cfg.scheduler.build(optimizer, cfg.steps) if cfg.scheduler is not None else None
+    return optimizer, lr_scheduler

lerobot/common/optim/optimizers.py

@@ -0,0 +1,118 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import abc
+from dataclasses import asdict, dataclass
+from pathlib import Path
+
+import draccus
+import torch
+from safetensors.torch import load_file, save_file
+
+from lerobot.common.constants import (
+    OPTIMIZER_PARAM_GROUPS,
+    OPTIMIZER_STATE,
+)
+from lerobot.common.datasets.utils import flatten_dict, unflatten_dict, write_json
+from lerobot.common.utils.io_utils import deserialize_json_into_object
+
+
+@dataclass
+class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
+    lr: float
+    weight_decay: float
+    grad_clip_norm: float
+
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)
+
+    @classmethod
+    def default_choice_name(cls) -> str | None:
+        return "adam"
+
+    @abc.abstractmethod
+    def build(self) -> torch.optim.Optimizer:
+        raise NotImplementedError
+
+
+@OptimizerConfig.register_subclass("adam")
+@dataclass
+class AdamConfig(OptimizerConfig):
+    lr: float = 1e-3
+    betas: tuple[float, float] = (0.9, 0.999)
+    eps: float = 1e-8
+    weight_decay: float = 0.0
+    grad_clip_norm: float = 10.0
+
+    def build(self, params: dict) -> torch.optim.Optimizer:
+        kwargs = asdict(self)
+        kwargs.pop("grad_clip_norm")
+        return torch.optim.Adam(params, **kwargs)
+
+
+@OptimizerConfig.register_subclass("adamw")
+@dataclass
+class AdamWConfig(OptimizerConfig):
+    lr: float = 1e-3
+    betas: tuple[float, float] = (0.9, 0.999)
+    eps: float = 1e-8
+    weight_decay: float = 1e-2
+    grad_clip_norm: float = 10.0
+
+    def build(self, params: dict) -> torch.optim.Optimizer:
+        kwargs = asdict(self)
+        kwargs.pop("grad_clip_norm")
+        return torch.optim.AdamW(params, **kwargs)
+
+
+@OptimizerConfig.register_subclass("sgd")
+@dataclass
+class SGDConfig(OptimizerConfig):
+    lr: float = 1e-3
+    momentum: float = 0.0
+    dampening: float = 0.0
+    nesterov: bool = False
+    weight_decay: float = 0.0
+    grad_clip_norm: float = 10.0
+
+    def build(self, params: dict) -> torch.optim.Optimizer:
+        kwargs = asdict(self)
+        kwargs.pop("grad_clip_norm")
+        return torch.optim.SGD(params, **kwargs)
+
+
+def save_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
+    state = optimizer.state_dict()
+    param_groups = state.pop("param_groups")
+    flat_state = flatten_dict(state)
+    save_file(flat_state, save_dir / OPTIMIZER_STATE)
+    write_json(param_groups, save_dir / OPTIMIZER_PARAM_GROUPS)
+
+
+def load_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> torch.optim.Optimizer:
+    current_state_dict = optimizer.state_dict()
+    flat_state = load_file(save_dir / OPTIMIZER_STATE)
+    state = unflatten_dict(flat_state)
+    loaded_state_dict = {"state": {int(k): v for k, v in state["state"].items()}}
+
+    if "param_groups" in current_state_dict:
+        param_groups = deserialize_json_into_object(
+            save_dir / OPTIMIZER_PARAM_GROUPS, current_state_dict["param_groups"]
+        )
+        loaded_state_dict["param_groups"] = param_groups
+
+    optimizer.load_state_dict(loaded_state_dict)
+    return optimizer

lerobot/common/optim/schedulers.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import abc
+import math
+from dataclasses import asdict, dataclass
+from pathlib import Path
+
+import draccus
+from torch.optim import Optimizer
+from torch.optim.lr_scheduler import LambdaLR, LRScheduler
+
+from lerobot.common.constants import SCHEDULER_STATE
+from lerobot.common.datasets.utils import write_json
+from lerobot.common.utils.io_utils import deserialize_json_into_object
+
+
+@dataclass
+class LRSchedulerConfig(draccus.ChoiceRegistry, abc.ABC):
+    num_warmup_steps: int
+
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)
+
+    @abc.abstractmethod
+    def build(self, optimizer: Optimizer, num_training_steps: int) -> LRScheduler | None:
+        raise NotImplementedError
+
+
+@LRSchedulerConfig.register_subclass("diffuser")
+@dataclass
+class DiffuserSchedulerConfig(LRSchedulerConfig):
+    name: str = "cosine"
+    num_warmup_steps: int | None = None
+
+    def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
+        from diffusers.optimization import get_scheduler
+
+        kwargs = {**asdict(self), "num_training_steps": num_training_steps, "optimizer": optimizer}
+        return get_scheduler(**kwargs)
+
+
+@LRSchedulerConfig.register_subclass("vqbet")
+@dataclass
+class VQBeTSchedulerConfig(LRSchedulerConfig):
+    num_warmup_steps: int
+    num_vqvae_training_steps: int
+    num_cycles: float = 0.5
+
+    def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
+        def lr_lambda(current_step):
+            if current_step < self.num_vqvae_training_steps:
+                return float(1)
+            else:
+                adjusted_step = current_step - self.num_vqvae_training_steps
+                if adjusted_step < self.num_warmup_steps:
+                    return float(adjusted_step) / float(max(1, self.num_warmup_steps))
+                progress = float(adjusted_step - self.num_warmup_steps) / float(
+                    max(1, num_training_steps - self.num_warmup_steps)
+                )
+                return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(self.num_cycles) * 2.0 * progress)))
+
+        return LambdaLR(optimizer, lr_lambda, -1)
+
+
+@LRSchedulerConfig.register_subclass("cosine_decay_with_warmup")
+@dataclass
+class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
+    """Used by Physical Intelligence to train Pi0"""
+
+    num_warmup_steps: int
+    num_decay_steps: int
+    peak_lr: float
+    decay_lr: float
+
+    def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
+        del num_training_steps
+
+        def lr_lambda(current_step):
+            def linear_warmup_schedule(current_step):
+                if current_step <= 0:
+                    return 1 / (self.num_warmup_steps + 1)
+                frac = 1 - current_step / self.num_warmup_steps
+                return (1 / (self.num_warmup_steps + 1) - 1) * frac + 1
+
+            def cosine_decay_schedule(current_step):
+                step = min(current_step, self.num_decay_steps)
+                cosine_decay = 0.5 * (1 + math.cos(math.pi * step / self.num_decay_steps))
+                alpha = self.decay_lr / self.peak_lr
+                decayed = (1 - alpha) * cosine_decay + alpha
+                return decayed
+
+            if current_step < self.num_warmup_steps:
+                return linear_warmup_schedule(current_step)
+
+            return cosine_decay_schedule(current_step)
+
+        return LambdaLR(optimizer, lr_lambda, -1)
+
+
+def save_scheduler_state(scheduler: LRScheduler, save_dir: Path) -> None:
+    state_dict = scheduler.state_dict()
+    write_json(state_dict, save_dir / SCHEDULER_STATE)
+
+
+def load_scheduler_state(scheduler: LRScheduler, save_dir: Path) -> LRScheduler:
+    state_dict = deserialize_json_into_object(save_dir / SCHEDULER_STATE, scheduler.state_dict())
+    scheduler.load_state_dict(state_dict)
+    return scheduler

lerobot/common/policies/__init__.py

@@ -0,0 +1,19 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .act.configuration_act import ACTConfig as ACTConfig
+from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
+from .pi0.configuration_pi0 import PI0Config as PI0Config
+from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
+from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig

lerobot/common/policies/act/configuration_act.py

@@ -0,0 +1,186 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Tony Z. Zhao and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass, field
+
+from lerobot.common.optim.optimizers import AdamWConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+
+
+@PreTrainedConfig.register_subclass("act")
+@dataclass
+class ACTConfig(PreTrainedConfig):
+    """Configuration class for the Action Chunking Transformers policy.
+
+    Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
+
+    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
+    Those are: `input_shapes` and 'output_shapes`.
+
+    Notes on the inputs and outputs:
+        - Either:
+            - At least one key starting with "observation.image is required as an input.
+              AND/OR
+            - The key "observation.environment_state" is required as input.
+        - If there are multiple keys beginning with "observation.images." they are treated as multiple camera
+          views. Right now we only support all images having the same shape.
+        - May optionally work without an "observation.state" key for the proprioceptive robot state.
+        - "action" is required as an output key.
+
+    Args:
+        n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
+            current step and additional steps going back).
+        chunk_size: The size of the action prediction "chunks" in units of environment steps.
+        n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
+            This should be no greater than the chunk size. For example, if the chunk size size 100, you may
+            set this to 50. This would mean that the model predicts 100 steps worth of actions, runs 50 in the
+            environment, and throws the other 50 out.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
+        vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
+        pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
+            `None` means no pretrained weights.
+        replace_final_stride_with_dilation: Whether to replace the ResNet's final 2x2 stride with a dilated
+            convolution.
+        pre_norm: Whether to use "pre-norm" in the transformer blocks.
+        dim_model: The transformer blocks' main hidden dimension.
+        n_heads: The number of heads to use in the transformer blocks' multi-head attention.
+        dim_feedforward: The dimension to expand the transformer's hidden dimension to in the feed-forward
+            layers.
+        feedforward_activation: The activation to use in the transformer block's feed-forward layers.
+        n_encoder_layers: The number of transformer layers to use for the transformer encoder.
+        n_decoder_layers: The number of transformer layers to use for the transformer decoder.
+        use_vae: Whether to use a variational objective during training. This introduces another transformer
+            which is used as the VAE's encoder (not to be confused with the transformer encoder - see
+            documentation in the policy class).
+        latent_dim: The VAE's latent dimension.
+        n_vae_encoder_layers: The number of transformer layers to use for the VAE's encoder.
+        temporal_ensemble_coeff: Coefficient for the exponential weighting scheme to apply for temporal
+            ensembling. Defaults to None which means temporal ensembling is not used. `n_action_steps` must be
+            1 when using this feature, as inference needs to happen at every step to form an ensemble. For
+            more information on how ensembling works, please see `ACTTemporalEnsembler`.
+        dropout: Dropout to use in the transformer layers (see code for details).
+        kl_weight: The weight to use for the KL-divergence component of the loss if the variational objective
+            is enabled. Loss is then calculated as: `reconstruction_loss + kl_weight * kld_loss`.
+    """
+
+    # Input / output structure.
+    n_obs_steps: int = 1
+    chunk_size: int = 100
+    n_action_steps: int = 100
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.MEAN_STD,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MEAN_STD,
+        }
+    )
+
+    # Architecture.
+    # Vision backbone.
+    vision_backbone: str = "resnet18"
+    pretrained_backbone_weights: str | None = "ResNet18_Weights.IMAGENET1K_V1"
+    replace_final_stride_with_dilation: int = False
+    # Transformer layers.
+    pre_norm: bool = False
+    dim_model: int = 512
+    n_heads: int = 8
+    dim_feedforward: int = 3200
+    feedforward_activation: str = "relu"
+    n_encoder_layers: int = 4
+    # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
+    # that means only the first layer is used. Here we match the original implementation by setting this to 1.
+    # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
+    n_decoder_layers: int = 1
+    # VAE.
+    use_vae: bool = True
+    latent_dim: int = 32
+    n_vae_encoder_layers: int = 4
+
+    # Inference.
+    # Note: the value used in ACT when temporal ensembling is enabled is 0.01.
+    temporal_ensemble_coeff: float | None = None
+
+    # Training and loss computation.
+    dropout: float = 0.1
+    kl_weight: float = 10.0
+
+    # Training preset
+    optimizer_lr: float = 1e-5
+    optimizer_weight_decay: float = 1e-4
+    optimizer_lr_backbone: float = 1e-5
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        """Input validation (not exhaustive)."""
+        if not self.vision_backbone.startswith("resnet"):
+            raise ValueError(
+                f"`vision_backbone` must be one of the ResNet variants. Got {self.vision_backbone}."
+            )
+        if self.temporal_ensemble_coeff is not None and self.n_action_steps > 1:
+            raise NotImplementedError(
+                "`n_action_steps` must be 1 when using temporal ensembling. This is "
+                "because the policy needs to be queried every step to compute the ensembled action."
+            )
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
+                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
+            )
+        if self.n_obs_steps != 1:
+            raise ValueError(
+                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
+            )
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            weight_decay=self.optimizer_weight_decay,
+        )
+
+    def get_scheduler_preset(self) -> None:
+        return None
+
+    def validate_features(self) -> None:
+        if not self.image_features and not self.env_state_feature:
+            raise ValueError("You must provide at least one image or the environment state among the inputs.")
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

lerobot/common/policies/act/modeling_act.py

@@ -0,0 +1,765 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Tony Z. Zhao and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Action Chunking Transformer Policy
+
+As per Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (https://arxiv.org/abs/2304.13705).
+The majority of changes here involve removing unused code, unifying naming, and adding helpful comments.
+"""
+
+import math
+from collections import deque
+from itertools import chain
+from typing import Callable
+
+import einops
+import numpy as np
+import torch
+import torch.nn.functional as F  # noqa: N812
+import torchvision
+from torch import Tensor, nn
+from torchvision.models._utils import IntermediateLayerGetter
+from torchvision.ops.misc import FrozenBatchNorm2d
+
+from lerobot.common.policies.act.configuration_act import ACTConfig
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+
+
+class ACTPolicy(PreTrainedPolicy):
+    """
+    Action Chunking Transformer Policy as per Learning Fine-Grained Bimanual Manipulation with Low-Cost
+    Hardware (paper: https://arxiv.org/abs/2304.13705, code: https://github.com/tonyzhaozh/act)
+    """
+
+    config_class = ACTConfig
+    name = "act"
+
+    def __init__(
+        self,
+        config: ACTConfig,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                    the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_targets = Normalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+
+        self.model = ACT(config)
+
+        if config.temporal_ensemble_coeff is not None:
+            self.temporal_ensembler = ACTTemporalEnsembler(config.temporal_ensemble_coeff, config.chunk_size)
+
+        self.reset()
+
+    def get_optim_params(self) -> dict:
+        # TODO(aliberts, rcadene): As of now, lr_backbone == lr
+        # Should we remove this and just `return self.parameters()`?
+        return [
+            {
+                "params": [
+                    p
+                    for n, p in self.named_parameters()
+                    if not n.startswith("model.backbone") and p.requires_grad
+                ]
+            },
+            {
+                "params": [
+                    p
+                    for n, p in self.named_parameters()
+                    if n.startswith("model.backbone") and p.requires_grad
+                ],
+                "lr": self.config.optimizer_lr_backbone,
+            },
+        ]
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        if self.config.temporal_ensemble_coeff is not None:
+            self.temporal_ensembler.reset()
+        else:
+            self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    @torch.no_grad
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method wraps `select_actions` in order to return one action at a time for execution in the
+        environment. It works by managing the actions in a queue and only calling `select_actions` when the
+        queue is empty.
+        """
+        self.eval()
+
+        batch = self.normalize_inputs(batch)
+        if self.config.image_features:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.images"] = [batch[key] for key in self.config.image_features]
+
+        # If we are doing temporal ensembling, do online updates where we keep track of the number of actions
+        # we are ensembling over.
+        if self.config.temporal_ensemble_coeff is not None:
+            actions = self.model(batch)[0]  # (batch_size, chunk_size, action_dim)
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+            action = self.temporal_ensembler.update(actions)
+            return action
+
+        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
+        # querying the policy.
+        if len(self._action_queue) == 0:
+            actions = self.model(batch)[0][:, : self.config.n_action_steps]
+
+            # TODO(rcadene): make _forward return output dictionary?
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
+        """Run the batch through the model and compute the loss for training or validation."""
+        batch = self.normalize_inputs(batch)
+        if self.config.image_features:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.images"] = [batch[key] for key in self.config.image_features]
+
+        batch = self.normalize_targets(batch)
+        actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
+
+        l1_loss = (
+            F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
+        ).mean()
+
+        loss_dict = {"l1_loss": l1_loss.item()}
+        if self.config.use_vae:
+            # Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for
+            # each dimension independently, we sum over the latent dimension to get the total
+            # KL-divergence per batch element, then take the mean over the batch.
+            # (See App. B of https://arxiv.org/abs/1312.6114 for more details).
+            mean_kld = (
+                (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
+            )
+            loss_dict["kld_loss"] = mean_kld.item()
+            loss = l1_loss + mean_kld * self.config.kl_weight
+        else:
+            loss = l1_loss
+
+        return loss, loss_dict
+
+
+class ACTTemporalEnsembler:
+    def __init__(self, temporal_ensemble_coeff: float, chunk_size: int) -> None:
+        """Temporal ensembling as described in Algorithm 2 of https://arxiv.org/abs/2304.13705.
+
+        The weights are calculated as wᵢ = exp(-temporal_ensemble_coeff * i) where w₀ is the oldest action.
+        They are then normalized to sum to 1 by dividing by Σwᵢ. Here's some intuition around how the
+        coefficient works:
+            - Setting it to 0 uniformly weighs all actions.
+            - Setting it positive gives more weight to older actions.
+            - Setting it negative gives more weight to newer actions.
+        NOTE: The default value for `temporal_ensemble_coeff` used by the original ACT work is 0.01. This
+        results in older actions being weighed more highly than newer actions (the experiments documented in
+        https://github.com/huggingface/lerobot/pull/319 hint at why highly weighing new actions might be
+        detrimental: doing so aggressively may diminish the benefits of action chunking).
+
+        Here we use an online method for computing the average rather than caching a history of actions in
+        order to compute the average offline. For a simple 1D sequence it looks something like:
+
+        ```
+        import torch
+
+        seq = torch.linspace(8, 8.5, 100)
+        print(seq)
+
+        m = 0.01
+        exp_weights = torch.exp(-m * torch.arange(len(seq)))
+        print(exp_weights)
+
+        # Calculate offline
+        avg = (exp_weights * seq).sum() / exp_weights.sum()
+        print("offline", avg)
+
+        # Calculate online
+        for i, item in enumerate(seq):
+            if i == 0:
+                avg = item
+                continue
+            avg *= exp_weights[:i].sum()
+            avg += item * exp_weights[i]
+            avg /= exp_weights[:i+1].sum()
+        print("online", avg)
+        ```
+        """
+        self.chunk_size = chunk_size
+        self.ensemble_weights = torch.exp(-temporal_ensemble_coeff * torch.arange(chunk_size))
+        self.ensemble_weights_cumsum = torch.cumsum(self.ensemble_weights, dim=0)
+        self.reset()
+
+    def reset(self):
+        """Resets the online computation variables."""
+        self.ensembled_actions = None
+        # (chunk_size,) count of how many actions are in the ensemble for each time step in the sequence.
+        self.ensembled_actions_count = None
+
+    def update(self, actions: Tensor) -> Tensor:
+        """
+        Takes a (batch, chunk_size, action_dim) sequence of actions, update the temporal ensemble for all
+        time steps, and pop/return the next batch of actions in the sequence.
+        """
+        self.ensemble_weights = self.ensemble_weights.to(device=actions.device)
+        self.ensemble_weights_cumsum = self.ensemble_weights_cumsum.to(device=actions.device)
+        if self.ensembled_actions is None:
+            # Initializes `self._ensembled_action` to the sequence of actions predicted during the first
+            # time step of the episode.
+            self.ensembled_actions = actions.clone()
+            # Note: The last dimension is unsqueeze to make sure we can broadcast properly for tensor
+            # operations later.
+            self.ensembled_actions_count = torch.ones(
+                (self.chunk_size, 1), dtype=torch.long, device=self.ensembled_actions.device
+            )
+        else:
+            # self.ensembled_actions will have shape (batch_size, chunk_size - 1, action_dim). Compute
+            # the online update for those entries.
+            self.ensembled_actions *= self.ensemble_weights_cumsum[self.ensembled_actions_count - 1]
+            self.ensembled_actions += actions[:, :-1] * self.ensemble_weights[self.ensembled_actions_count]
+            self.ensembled_actions /= self.ensemble_weights_cumsum[self.ensembled_actions_count]
+            self.ensembled_actions_count = torch.clamp(self.ensembled_actions_count + 1, max=self.chunk_size)
+            # The last action, which has no prior online average, needs to get concatenated onto the end.
+            self.ensembled_actions = torch.cat([self.ensembled_actions, actions[:, -1:]], dim=1)
+            self.ensembled_actions_count = torch.cat(
+                [self.ensembled_actions_count, torch.ones_like(self.ensembled_actions_count[-1:])]
+            )
+        # "Consume" the first action.
+        action, self.ensembled_actions, self.ensembled_actions_count = (
+            self.ensembled_actions[:, 0],
+            self.ensembled_actions[:, 1:],
+            self.ensembled_actions_count[1:],
+        )
+        return action
+
+
+class ACT(nn.Module):
+    """Action Chunking Transformer: The underlying neural network for ACTPolicy.
+
+    Note: In this code we use the terms `vae_encoder`, 'encoder', `decoder`. The meanings are as follows.
+        - The `vae_encoder` is, as per the literature around variational auto-encoders (VAE), the part of the
+          model that encodes the target data (a sequence of actions), and the condition (the robot
+          joint-space).
+        - A transformer with an `encoder` (not the VAE encoder) and `decoder` (not the VAE decoder) with
+          cross-attention is used as the VAE decoder. For these terms, we drop the `vae_` prefix because we
+          have an option to train this model without the variational objective (in which case we drop the
+          `vae_encoder` altogether, and nothing about this model has anything to do with a VAE).
+
+                                 Transformer
+                                 Used alone for inference
+                                 (acts as VAE decoder
+                                  during training)
+                                ┌───────────────────────┐
+                                │             Outputs   │
+                                │                ▲      │
+                                │     ┌─────►┌───────┐  │
+                   ┌──────┐     │     │      │Transf.│  │
+                   │      │     │     ├─────►│decoder│  │
+              ┌────┴────┐ │     │     │      │       │  │
+              │         │ │     │ ┌───┴───┬─►│       │  │
+              │ VAE     │ │     │ │       │  └───────┘  │
+              │ encoder │ │     │ │Transf.│             │
+              │         │ │     │ │encoder│             │
+              └───▲─────┘ │     │ │       │             │
+                  │       │     │ └▲──▲─▲─┘             │
+                  │       │     │  │  │ │               │
+                inputs    └─────┼──┘  │ image emb.      │
+                                │    state emb.         │
+                                └───────────────────────┘
+    """
+
+    def __init__(self, config: ACTConfig):
+        # BERT style VAE encoder with input tokens [cls, robot_state, *action_sequence].
+        # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
+        super().__init__()
+        self.config = config
+
+        if self.config.use_vae:
+            self.vae_encoder = ACTEncoder(config, is_vae_encoder=True)
+            self.vae_encoder_cls_embed = nn.Embedding(1, config.dim_model)
+            # Projection layer for joint-space configuration to hidden dimension.
+            if self.config.robot_state_feature:
+                self.vae_encoder_robot_state_input_proj = nn.Linear(
+                    self.config.robot_state_feature.shape[0], config.dim_model
+                )
+            # Projection layer for action (joint-space target) to hidden dimension.
+            self.vae_encoder_action_input_proj = nn.Linear(
+                self.config.action_feature.shape[0],
+                config.dim_model,
+            )
+            # Projection layer from the VAE encoder's output to the latent distribution's parameter space.
+            self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, config.latent_dim * 2)
+            # Fixed sinusoidal positional embedding for the input to the VAE encoder. Unsqueeze for batch
+            # dimension.
+            num_input_token_encoder = 1 + config.chunk_size
+            if self.config.robot_state_feature:
+                num_input_token_encoder += 1
+            self.register_buffer(
+                "vae_encoder_pos_enc",
+                create_sinusoidal_pos_embedding(num_input_token_encoder, config.dim_model).unsqueeze(0),
+            )
+
+        # Backbone for image feature extraction.
+        if self.config.image_features:
+            backbone_model = getattr(torchvision.models, config.vision_backbone)(
+                replace_stride_with_dilation=[False, False, config.replace_final_stride_with_dilation],
+                weights=config.pretrained_backbone_weights,
+                norm_layer=FrozenBatchNorm2d,
+            )
+            # Note: The assumption here is that we are using a ResNet model (and hence layer4 is the final
+            # feature map).
+            # Note: The forward method of this returns a dict: {"feature_map": output}.
+            self.backbone = IntermediateLayerGetter(backbone_model, return_layers={"layer4": "feature_map"})
+
+        # Transformer (acts as VAE decoder when training with the variational objective).
+        self.encoder = ACTEncoder(config)
+        self.decoder = ACTDecoder(config)
+
+        # Transformer encoder input projections. The tokens will be structured like
+        # [latent, (robot_state), (env_state), (image_feature_map_pixels)].
+        if self.config.robot_state_feature:
+            self.encoder_robot_state_input_proj = nn.Linear(
+                self.config.robot_state_feature.shape[0], config.dim_model
+            )
+        if self.config.env_state_feature:
+            self.encoder_env_state_input_proj = nn.Linear(
+                self.config.env_state_feature.shape[0], config.dim_model
+            )
+        self.encoder_latent_input_proj = nn.Linear(config.latent_dim, config.dim_model)
+        if self.config.image_features:
+            self.encoder_img_feat_input_proj = nn.Conv2d(
+                backbone_model.fc.in_features, config.dim_model, kernel_size=1
+            )
+        # Transformer encoder positional embeddings.
+        n_1d_tokens = 1  # for the latent
+        if self.config.robot_state_feature:
+            n_1d_tokens += 1
+        if self.config.env_state_feature:
+            n_1d_tokens += 1
+        self.encoder_1d_feature_pos_embed = nn.Embedding(n_1d_tokens, config.dim_model)
+        if self.config.image_features:
+            self.encoder_cam_feat_pos_embed = ACTSinusoidalPositionEmbedding2d(config.dim_model // 2)
+
+        # Transformer decoder.
+        # Learnable positional embedding for the transformer's decoder (in the style of DETR object queries).
+        self.decoder_pos_embed = nn.Embedding(config.chunk_size, config.dim_model)
+
+        # Final action regression head on the output of the transformer's decoder.
+        self.action_head = nn.Linear(config.dim_model, self.config.action_feature.shape[0])
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        """Xavier-uniform initialization of the transformer parameters as in the original code."""
+        for p in chain(self.encoder.parameters(), self.decoder.parameters()):
+            if p.dim() > 1:
+                nn.init.xavier_uniform_(p)
+
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, tuple[Tensor, Tensor] | tuple[None, None]]:
+        """A forward pass through the Action Chunking Transformer (with optional VAE encoder).
+
+        `batch` should have the following structure:
+        {
+            [robot_state_feature] (optional): (B, state_dim) batch of robot states.
+
+            [image_features]: (B, n_cameras, C, H, W) batch of images.
+                AND/OR
+            [env_state_feature]: (B, env_dim) batch of environment states.
+
+            [action_feature] (optional, only if training with VAE): (B, chunk_size, action dim) batch of actions.
+        }
+
+        Returns:
+            (B, chunk_size, action_dim) batch of action sequences
+            Tuple containing the latent PDF's parameters (mean, log(σ²)) both as (B, L) tensors where L is the
+            latent dimension.
+        """
+        if self.config.use_vae and self.training:
+            assert "action" in batch, (
+                "actions must be provided when using the variational objective in training mode."
+            )
+
+        if "observation.images" in batch:
+            batch_size = batch["observation.images"][0].shape[0]
+        else:
+            batch_size = batch["observation.environment_state"].shape[0]
+
+        # Prepare the latent for input to the transformer encoder.
+        if self.config.use_vae and "action" in batch:
+            # Prepare the input to the VAE encoder: [cls, *joint_space_configuration, *action_sequence].
+            cls_embed = einops.repeat(
+                self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
+            )  # (B, 1, D)
+            if self.config.robot_state_feature:
+                robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
+                robot_state_embed = robot_state_embed.unsqueeze(1)  # (B, 1, D)
+            action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
+
+            if self.config.robot_state_feature:
+                vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
+            else:
+                vae_encoder_input = [cls_embed, action_embed]
+            vae_encoder_input = torch.cat(vae_encoder_input, axis=1)
+
+            # Prepare fixed positional embedding.
+            # Note: detach() shouldn't be necessary but leaving it the same as the original code just in case.
+            pos_embed = self.vae_encoder_pos_enc.clone().detach()  # (1, S+2, D)
+
+            # Prepare key padding mask for the transformer encoder. We have 1 or 2 extra tokens at the start of the
+            # sequence depending whether we use the input states or not (cls and robot state)
+            # False means not a padding token.
+            cls_joint_is_pad = torch.full(
+                (batch_size, 2 if self.config.robot_state_feature else 1),
+                False,
+                device=batch["observation.state"].device,
+            )
+            key_padding_mask = torch.cat(
+                [cls_joint_is_pad, batch["action_is_pad"]], axis=1
+            )  # (bs, seq+1 or 2)
+
+            # Forward pass through VAE encoder to get the latent PDF parameters.
+            cls_token_out = self.vae_encoder(
+                vae_encoder_input.permute(1, 0, 2),
+                pos_embed=pos_embed.permute(1, 0, 2),
+                key_padding_mask=key_padding_mask,
+            )[0]  # select the class token, with shape (B, D)
+            latent_pdf_params = self.vae_encoder_latent_output_proj(cls_token_out)
+            mu = latent_pdf_params[:, : self.config.latent_dim]
+            # This is 2log(sigma). Done this way to match the original implementation.
+            log_sigma_x2 = latent_pdf_params[:, self.config.latent_dim :]
+
+            # Sample the latent with the reparameterization trick.
+            latent_sample = mu + log_sigma_x2.div(2).exp() * torch.randn_like(mu)
+        else:
+            # When not using the VAE encoder, we set the latent to be all zeros.
+            mu = log_sigma_x2 = None
+            # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
+            latent_sample = torch.zeros([batch_size, self.config.latent_dim], dtype=torch.float32).to(
+                batch["observation.state"].device
+            )
+
+        # Prepare transformer encoder inputs.
+        encoder_in_tokens = [self.encoder_latent_input_proj(latent_sample)]
+        encoder_in_pos_embed = list(self.encoder_1d_feature_pos_embed.weight.unsqueeze(1))
+        # Robot state token.
+        if self.config.robot_state_feature:
+            encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch["observation.state"]))
+        # Environment state token.
+        if self.config.env_state_feature:
+            encoder_in_tokens.append(
+                self.encoder_env_state_input_proj(batch["observation.environment_state"])
+            )
+
+        # Camera observation features and positional embeddings.
+        if self.config.image_features:
+            all_cam_features = []
+            all_cam_pos_embeds = []
+
+            # For a list of images, the H and W may vary but H*W is constant.
+            for img in batch["observation.images"]:
+                cam_features = self.backbone(img)["feature_map"]
+                cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
+                cam_features = self.encoder_img_feat_input_proj(cam_features)
+
+                # Rearrange features to (sequence, batch, dim).
+                cam_features = einops.rearrange(cam_features, "b c h w -> (h w) b c")
+                cam_pos_embed = einops.rearrange(cam_pos_embed, "b c h w -> (h w) b c")
+
+                all_cam_features.append(cam_features)
+                all_cam_pos_embeds.append(cam_pos_embed)
+
+            encoder_in_tokens.extend(torch.cat(all_cam_features, axis=0))
+            encoder_in_pos_embed.extend(torch.cat(all_cam_pos_embeds, axis=0))
+
+        # Stack all tokens along the sequence dimension.
+        encoder_in_tokens = torch.stack(encoder_in_tokens, axis=0)
+        encoder_in_pos_embed = torch.stack(encoder_in_pos_embed, axis=0)
+
+        # Forward pass through the transformer modules.
+        encoder_out = self.encoder(encoder_in_tokens, pos_embed=encoder_in_pos_embed)
+        # TODO(rcadene, alexander-soare): remove call to `device` ; precompute and use buffer
+        decoder_in = torch.zeros(
+            (self.config.chunk_size, batch_size, self.config.dim_model),
+            dtype=encoder_in_pos_embed.dtype,
+            device=encoder_in_pos_embed.device,
+        )
+        decoder_out = self.decoder(
+            decoder_in,
+            encoder_out,
+            encoder_pos_embed=encoder_in_pos_embed,
+            decoder_pos_embed=self.decoder_pos_embed.weight.unsqueeze(1),
+        )
+
+        # Move back to (B, S, C).
+        decoder_out = decoder_out.transpose(0, 1)
+
+        actions = self.action_head(decoder_out)
+
+        return actions, (mu, log_sigma_x2)
+
+
+class ACTEncoder(nn.Module):
+    """Convenience module for running multiple encoder layers, maybe followed by normalization."""
+
+    def __init__(self, config: ACTConfig, is_vae_encoder: bool = False):
+        super().__init__()
+        self.is_vae_encoder = is_vae_encoder
+        num_layers = config.n_vae_encoder_layers if self.is_vae_encoder else config.n_encoder_layers
+        self.layers = nn.ModuleList([ACTEncoderLayer(config) for _ in range(num_layers)])
+        self.norm = nn.LayerNorm(config.dim_model) if config.pre_norm else nn.Identity()
+
+    def forward(
+        self, x: Tensor, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None
+    ) -> Tensor:
+        for layer in self.layers:
+            x = layer(x, pos_embed=pos_embed, key_padding_mask=key_padding_mask)
+        x = self.norm(x)
+        return x
+
+
+class ACTEncoderLayer(nn.Module):
+    def __init__(self, config: ACTConfig):
+        super().__init__()
+        self.self_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
+
+        # Feed forward layers.
+        self.linear1 = nn.Linear(config.dim_model, config.dim_feedforward)
+        self.dropout = nn.Dropout(config.dropout)
+        self.linear2 = nn.Linear(config.dim_feedforward, config.dim_model)
+
+        self.norm1 = nn.LayerNorm(config.dim_model)
+        self.norm2 = nn.LayerNorm(config.dim_model)
+        self.dropout1 = nn.Dropout(config.dropout)
+        self.dropout2 = nn.Dropout(config.dropout)
+
+        self.activation = get_activation_fn(config.feedforward_activation)
+        self.pre_norm = config.pre_norm
+
+    def forward(self, x, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None) -> Tensor:
+        skip = x
+        if self.pre_norm:
+            x = self.norm1(x)
+        q = k = x if pos_embed is None else x + pos_embed
+        x = self.self_attn(q, k, value=x, key_padding_mask=key_padding_mask)
+        x = x[0]  # note: [0] to select just the output, not the attention weights
+        x = skip + self.dropout1(x)
+        if self.pre_norm:
+            skip = x
+            x = self.norm2(x)
+        else:
+            x = self.norm1(x)
+            skip = x
+        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
+        x = skip + self.dropout2(x)
+        if not self.pre_norm:
+            x = self.norm2(x)
+        return x
+
+
+class ACTDecoder(nn.Module):
+    def __init__(self, config: ACTConfig):
+        """Convenience module for running multiple decoder layers followed by normalization."""
+        super().__init__()
+        self.layers = nn.ModuleList([ACTDecoderLayer(config) for _ in range(config.n_decoder_layers)])
+        self.norm = nn.LayerNorm(config.dim_model)
+
+    def forward(
+        self,
+        x: Tensor,
+        encoder_out: Tensor,
+        decoder_pos_embed: Tensor | None = None,
+        encoder_pos_embed: Tensor | None = None,
+    ) -> Tensor:
+        for layer in self.layers:
+            x = layer(
+                x, encoder_out, decoder_pos_embed=decoder_pos_embed, encoder_pos_embed=encoder_pos_embed
+            )
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+
+class ACTDecoderLayer(nn.Module):
+    def __init__(self, config: ACTConfig):
+        super().__init__()
+        self.self_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
+        self.multihead_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
+
+        # Feed forward layers.
+        self.linear1 = nn.Linear(config.dim_model, config.dim_feedforward)
+        self.dropout = nn.Dropout(config.dropout)
+        self.linear2 = nn.Linear(config.dim_feedforward, config.dim_model)
+
+        self.norm1 = nn.LayerNorm(config.dim_model)
+        self.norm2 = nn.LayerNorm(config.dim_model)
+        self.norm3 = nn.LayerNorm(config.dim_model)
+        self.dropout1 = nn.Dropout(config.dropout)
+        self.dropout2 = nn.Dropout(config.dropout)
+        self.dropout3 = nn.Dropout(config.dropout)
+
+        self.activation = get_activation_fn(config.feedforward_activation)
+        self.pre_norm = config.pre_norm
+
+    def maybe_add_pos_embed(self, tensor: Tensor, pos_embed: Tensor | None) -> Tensor:
+        return tensor if pos_embed is None else tensor + pos_embed
+
+    def forward(
+        self,
+        x: Tensor,
+        encoder_out: Tensor,
+        decoder_pos_embed: Tensor | None = None,
+        encoder_pos_embed: Tensor | None = None,
+    ) -> Tensor:
+        """
+        Args:
+            x: (Decoder Sequence, Batch, Channel) tensor of input tokens.
+            encoder_out: (Encoder Sequence, B, C) output features from the last layer of the encoder we are
+                cross-attending with.
+            decoder_pos_embed: (ES, 1, C) positional embedding for keys (from the encoder).
+            encoder_pos_embed: (DS, 1, C) Positional_embedding for the queries (from the decoder).
+        Returns:
+            (DS, B, C) tensor of decoder output features.
+        """
+        skip = x
+        if self.pre_norm:
+            x = self.norm1(x)
+        q = k = self.maybe_add_pos_embed(x, decoder_pos_embed)
+        x = self.self_attn(q, k, value=x)[0]  # select just the output, not the attention weights
+        x = skip + self.dropout1(x)
+        if self.pre_norm:
+            skip = x
+            x = self.norm2(x)
+        else:
+            x = self.norm1(x)
+            skip = x
+        x = self.multihead_attn(
+            query=self.maybe_add_pos_embed(x, decoder_pos_embed),
+            key=self.maybe_add_pos_embed(encoder_out, encoder_pos_embed),
+            value=encoder_out,
+        )[0]  # select just the output, not the attention weights
+        x = skip + self.dropout2(x)
+        if self.pre_norm:
+            skip = x
+            x = self.norm3(x)
+        else:
+            x = self.norm2(x)
+            skip = x
+        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
+        x = skip + self.dropout3(x)
+        if not self.pre_norm:
+            x = self.norm3(x)
+        return x
+
+
+def create_sinusoidal_pos_embedding(num_positions: int, dimension: int) -> Tensor:
+    """1D sinusoidal positional embeddings as in Attention is All You Need.
+
+    Args:
+        num_positions: Number of token positions required.
+    Returns: (num_positions, dimension) position embeddings (the first dimension is the batch dimension).
+
+    """
+
+    def get_position_angle_vec(position):
+        return [position / np.power(10000, 2 * (hid_j // 2) / dimension) for hid_j in range(dimension)]
+
+    sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(num_positions)])
+    sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])  # dim 2i
+    sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])  # dim 2i+1
+    return torch.from_numpy(sinusoid_table).float()
+
+
+class ACTSinusoidalPositionEmbedding2d(nn.Module):
+    """2D sinusoidal positional embeddings similar to what's presented in Attention Is All You Need.
+
+    The variation is that the position indices are normalized in [0, 2π] (not quite: the lower bound is 1/H
+    for the vertical direction, and 1/W for the horizontal direction.
+    """
+
+    def __init__(self, dimension: int):
+        """
+        Args:
+            dimension: The desired dimension of the embeddings.
+        """
+        super().__init__()
+        self.dimension = dimension
+        self._two_pi = 2 * math.pi
+        self._eps = 1e-6
+        # Inverse "common ratio" for the geometric progression in sinusoid frequencies.
+        self._temperature = 10000
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        Args:
+            x: A (B, C, H, W) batch of 2D feature map to generate the embeddings for.
+        Returns:
+            A (1, C, H, W) batch of corresponding sinusoidal positional embeddings.
+        """
+        not_mask = torch.ones_like(x[0, :1])  # (1, H, W)
+        # Note: These are like range(1, H+1) and range(1, W+1) respectively, but in most implementations
+        # they would be range(0, H) and range(0, W). Keeping it at as is to match the original code.
+        y_range = not_mask.cumsum(1, dtype=torch.float32)
+        x_range = not_mask.cumsum(2, dtype=torch.float32)
+
+        # "Normalize" the position index such that it ranges in [0, 2π].
+        # Note: Adding epsilon on the denominator should not be needed as all values of y_embed and x_range
+        # are non-zero by construction. This is an artifact of the original code.
+        y_range = y_range / (y_range[:, -1:, :] + self._eps) * self._two_pi
+        x_range = x_range / (x_range[:, :, -1:] + self._eps) * self._two_pi
+
+        inverse_frequency = self._temperature ** (
+            2 * (torch.arange(self.dimension, dtype=torch.float32, device=x.device) // 2) / self.dimension
+        )
+
+        x_range = x_range.unsqueeze(-1) / inverse_frequency  # (1, H, W, 1)
+        y_range = y_range.unsqueeze(-1) / inverse_frequency  # (1, H, W, 1)
+
+        # Note: this stack then flatten operation results in interleaved sine and cosine terms.
+        # pos_embed_x and pos_embed_y are (1, H, W, C // 2).
+        pos_embed_x = torch.stack((x_range[..., 0::2].sin(), x_range[..., 1::2].cos()), dim=-1).flatten(3)
+        pos_embed_y = torch.stack((y_range[..., 0::2].sin(), y_range[..., 1::2].cos()), dim=-1).flatten(3)
+        pos_embed = torch.cat((pos_embed_y, pos_embed_x), dim=3).permute(0, 3, 1, 2)  # (1, C, H, W)
+
+        return pos_embed
+
+
+def get_activation_fn(activation: str) -> Callable:
+    """Return an activation function given a string."""
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    raise RuntimeError(f"activation should be relu/gelu/glu, not {activation}.")

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -0,0 +1,237 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Columbia Artificial Intelligence, Robotics Lab,
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass, field
+
+from lerobot.common.optim.optimizers import AdamConfig
+from lerobot.common.optim.schedulers import DiffuserSchedulerConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+
+
+@PreTrainedConfig.register_subclass("diffusion")
+@dataclass
+class DiffusionConfig(PreTrainedConfig):
+    """Configuration class for DiffusionPolicy.
+
+    Defaults are configured for training with PushT providing proprioceptive and single camera observations.
+
+    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
+    Those are: `input_shapes` and `output_shapes`.
+
+    Notes on the inputs and outputs:
+        - "observation.state" is required as an input key.
+        - Either:
+            - At least one key starting with "observation.image is required as an input.
+              AND/OR
+            - The key "observation.environment_state" is required as input.
+        - If there are multiple keys beginning with "observation.image" they are treated as multiple camera
+          views. Right now we only support all images having the same shape.
+        - "action" is required as an output key.
+
+    Args:
+        n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
+            current step and additional steps going back).
+        horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
+        n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
+            See `DiffusionPolicy.select_action` for more details.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
+        vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
+        crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
+            within the image size. If None, no cropping is done.
+        crop_is_random: Whether the crop should be random at training time (it's always a center crop in eval
+            mode).
+        pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
+            `None` means no pretrained weights.
+        use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
+            The group sizes are set to be about 16 (to be precise, feature_dim // 16).
+        spatial_softmax_num_keypoints: Number of keypoints for SpatialSoftmax.
+        use_separate_rgb_encoders_per_camera: Whether to use a separate RGB encoder for each camera view.
+        down_dims: Feature dimension for each stage of temporal downsampling in the diffusion modeling Unet.
+            You may provide a variable number of dimensions, therefore also controlling the degree of
+            downsampling.
+        kernel_size: The convolutional kernel size of the diffusion modeling Unet.
+        n_groups: Number of groups used in the group norm of the Unet's convolutional blocks.
+        diffusion_step_embed_dim: The Unet is conditioned on the diffusion timestep via a small non-linear
+            network. This is the output dimension of that network, i.e., the embedding dimension.
+        use_film_scale_modulation: FiLM (https://arxiv.org/abs/1709.07871) is used for the Unet conditioning.
+            Bias modulation is used be default, while this parameter indicates whether to also use scale
+            modulation.
+        noise_scheduler_type: Name of the noise scheduler to use. Supported options: ["DDPM", "DDIM"].
+        num_train_timesteps: Number of diffusion steps for the forward diffusion schedule.
+        beta_schedule: Name of the diffusion beta schedule as per DDPMScheduler from Hugging Face diffusers.
+        beta_start: Beta value for the first forward-diffusion step.
+        beta_end: Beta value for the last forward-diffusion step.
+        prediction_type: The type of prediction that the diffusion modeling Unet makes. Choose from "epsilon"
+            or "sample". These have equivalent outcomes from a latent variable modeling perspective, but
+            "epsilon" has been shown to work better in many deep neural network settings.
+        clip_sample: Whether to clip the sample to [-`clip_sample_range`, +`clip_sample_range`] for each
+            denoising step at inference time. WARNING: you will need to make sure your action-space is
+            normalized to fit within this range.
+        clip_sample_range: The magnitude of the clipping range as described above.
+        num_inference_steps: Number of reverse diffusion steps to use at inference time (steps are evenly
+            spaced). If not provided, this defaults to be the same as `num_train_timesteps`.
+        do_mask_loss_for_padding: Whether to mask the loss when there are copy-padded actions. See
+            `LeRobotDataset` and `load_previous_and_future_frames` for more information. Note, this defaults
+            to False as the original Diffusion Policy implementation does the same.
+    """
+
+    # Inputs / output structure.
+    n_obs_steps: int = 2
+    horizon: int = 16
+    n_action_steps: int = 8
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.MEAN_STD,
+            "STATE": NormalizationMode.MIN_MAX,
+            "ACTION": NormalizationMode.MIN_MAX,
+        }
+    )
+
+    # The original implementation doesn't sample frames for the last 7 steps,
+    # which avoids excessive padding and leads to improved training results.
+    drop_n_last_frames: int = 7  # horizon - n_action_steps - n_obs_steps + 1
+
+    # Architecture / modeling.
+    # Vision backbone.
+    vision_backbone: str = "resnet18"
+    crop_shape: tuple[int, int] | None = (84, 84)
+    crop_is_random: bool = True
+    pretrained_backbone_weights: str | None = None
+    use_group_norm: bool = True
+    spatial_softmax_num_keypoints: int = 32
+    use_separate_rgb_encoder_per_camera: bool = False
+    # Unet.
+    down_dims: tuple[int, ...] = (512, 1024, 2048)
+    kernel_size: int = 5
+    n_groups: int = 8
+    diffusion_step_embed_dim: int = 128
+    use_film_scale_modulation: bool = True
+    # Noise scheduler.
+    noise_scheduler_type: str = "DDPM"
+    num_train_timesteps: int = 100
+    beta_schedule: str = "squaredcos_cap_v2"
+    beta_start: float = 0.0001
+    beta_end: float = 0.02
+    prediction_type: str = "epsilon"
+    clip_sample: bool = True
+    clip_sample_range: float = 1.0
+
+    # Inference
+    num_inference_steps: int | None = None
+
+    # Loss computation
+    do_mask_loss_for_padding: bool = False
+
+    # Training presets
+    optimizer_lr: float = 1e-4
+    optimizer_betas: tuple = (0.95, 0.999)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-6
+    scheduler_name: str = "cosine"
+    scheduler_warmup_steps: int = 500
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        """Input validation (not exhaustive)."""
+        if not self.vision_backbone.startswith("resnet"):
+            raise ValueError(
+                f"`vision_backbone` must be one of the ResNet variants. Got {self.vision_backbone}."
+            )
+
+        supported_prediction_types = ["epsilon", "sample"]
+        if self.prediction_type not in supported_prediction_types:
+            raise ValueError(
+                f"`prediction_type` must be one of {supported_prediction_types}. Got {self.prediction_type}."
+            )
+        supported_noise_schedulers = ["DDPM", "DDIM"]
+        if self.noise_scheduler_type not in supported_noise_schedulers:
+            raise ValueError(
+                f"`noise_scheduler_type` must be one of {supported_noise_schedulers}. "
+                f"Got {self.noise_scheduler_type}."
+            )
+
+        # Check that the horizon size and U-Net downsampling is compatible.
+        # U-Net downsamples by 2 with each stage.
+        downsampling_factor = 2 ** len(self.down_dims)
+        if self.horizon % downsampling_factor != 0:
+            raise ValueError(
+                "The horizon should be an integer multiple of the downsampling factor (which is determined "
+                f"by `len(down_dims)`). Got {self.horizon=} and {self.down_dims=}"
+            )
+
+    def get_optimizer_preset(self) -> AdamConfig:
+        return AdamConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+        )
+
+    def get_scheduler_preset(self) -> DiffuserSchedulerConfig:
+        return DiffuserSchedulerConfig(
+            name=self.scheduler_name,
+            num_warmup_steps=self.scheduler_warmup_steps,
+        )
+
+    def validate_features(self) -> None:
+        if len(self.image_features) == 0 and self.env_state_feature is None:
+            raise ValueError("You must provide at least one image or the environment state among the inputs.")
+
+        if self.crop_shape is not None:
+            for key, image_ft in self.image_features.items():
+                if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
+                    raise ValueError(
+                        f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
+                        f"for `crop_shape` and {image_ft.shape} for "
+                        f"`{key}`."
+                    )
+
+        # Check that all input images have the same shape.
+        first_image_key, first_image_ft = next(iter(self.image_features.items()))
+        for key, image_ft in self.image_features.items():
+            if image_ft.shape != first_image_ft.shape:
+                raise ValueError(
+                    f"`{key}` does not match `{first_image_key}`, but we expect all image shapes to match."
+                )
+
+    @property
+    def observation_delta_indices(self) -> list:
+        return list(range(1 - self.n_obs_steps, 1))
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(1 - self.n_obs_steps, 1 - self.n_obs_steps + self.horizon))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

lerobot/common/policies/diffusion/diffusion_unet_image_policy.py

@@ -1,246 +0,0 @@
-from typing import Dict
-
-import torch
-import torch.nn.functional as F  # noqa: N812
-from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
-from einops import reduce
-
-from diffusion_policy.common.pytorch_util import dict_apply
-from diffusion_policy.model.diffusion.conditional_unet1d import ConditionalUnet1D
-from diffusion_policy.model.diffusion.mask_generator import LowdimMaskGenerator
-from diffusion_policy.model.vision.multi_image_obs_encoder import MultiImageObsEncoder
-from diffusion_policy.policy.base_image_policy import BaseImagePolicy
-
-
-class DiffusionUnetImagePolicy(BaseImagePolicy):
-    def __init__(
-        self,
-        shape_meta: dict,
-        noise_scheduler: DDPMScheduler,
-        obs_encoder: MultiImageObsEncoder,
-        horizon,
-        n_action_steps,
-        n_obs_steps,
-        num_inference_steps=None,
-        obs_as_global_cond=True,
-        diffusion_step_embed_dim=256,
-        down_dims=(256, 512, 1024),
-        kernel_size=5,
-        n_groups=8,
-        cond_predict_scale=True,
-        # parameters passed to step
-        **kwargs,
-    ):
-        super().__init__()
-
-        # parse shapes
-        action_shape = shape_meta["action"]["shape"]
-        assert len(action_shape) == 1
-        action_dim = action_shape[0]
-        # get feature dim
-        obs_feature_dim = obs_encoder.output_shape()[0]
-
-        # create diffusion model
-        input_dim = action_dim + obs_feature_dim
-        global_cond_dim = None
-        if obs_as_global_cond:
-            input_dim = action_dim
-            global_cond_dim = obs_feature_dim * n_obs_steps
-
-        model = ConditionalUnet1D(
-            input_dim=input_dim,
-            local_cond_dim=None,
-            global_cond_dim=global_cond_dim,
-            diffusion_step_embed_dim=diffusion_step_embed_dim,
-            down_dims=down_dims,
-            kernel_size=kernel_size,
-            n_groups=n_groups,
-            cond_predict_scale=cond_predict_scale,
-        )
-
-        self.obs_encoder = obs_encoder
-        self.model = model
-        self.noise_scheduler = noise_scheduler
-        self.mask_generator = LowdimMaskGenerator(
-            action_dim=action_dim,
-            obs_dim=0 if obs_as_global_cond else obs_feature_dim,
-            max_n_obs_steps=n_obs_steps,
-            fix_obs_steps=True,
-            action_visible=False,
-        )
-        self.horizon = horizon
-        self.obs_feature_dim = obs_feature_dim
-        self.action_dim = action_dim
-        self.n_action_steps = n_action_steps
-        self.n_obs_steps = n_obs_steps
-        self.obs_as_global_cond = obs_as_global_cond
-        self.kwargs = kwargs
-
-        if num_inference_steps is None:
-            num_inference_steps = noise_scheduler.config.num_train_timesteps
-        self.num_inference_steps = num_inference_steps
-
-    # ========= inference  ============
-    def conditional_sample(
-        self,
-        condition_data,
-        condition_mask,
-        local_cond=None,
-        global_cond=None,
-        generator=None,
-        # keyword arguments to scheduler.step
-        **kwargs,
-    ):
-        model = self.model
-        scheduler = self.noise_scheduler
-
-        trajectory = torch.randn(
-            size=condition_data.shape,
-            dtype=condition_data.dtype,
-            device=condition_data.device,
-            generator=generator,
-        )
-
-        # set step values
-        scheduler.set_timesteps(self.num_inference_steps)
-
-        for t in scheduler.timesteps:
-            # 1. apply conditioning
-            trajectory[condition_mask] = condition_data[condition_mask]
-
-            # 2. predict model output
-            model_output = model(trajectory, t, local_cond=local_cond, global_cond=global_cond)
-
-            # 3. compute previous image: x_t -> x_t-1
-            trajectory = scheduler.step(
-                model_output,
-                t,
-                trajectory,
-                generator=generator,
-                # **kwargs  # TODO(rcadene): in diffusion_policy, expected to be {}
-            ).prev_sample
-
-        # finally make sure conditioning is enforced
-        trajectory[condition_mask] = condition_data[condition_mask]
-
-        return trajectory
-
-    def predict_action(self, obs_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
-        """
-        obs_dict: must include "obs" key
-        result: must include "action" key
-        """
-        assert "past_action" not in obs_dict  # not implemented yet
-        nobs = obs_dict
-        value = next(iter(nobs.values()))
-        bsize, n_obs_steps = value.shape[:2]
-        horizon = self.horizon
-        action_dim = self.action_dim
-        obs_dim = self.obs_feature_dim
-        assert self.n_obs_steps == n_obs_steps
-
-        # build input
-        device = self.device
-        dtype = self.dtype
-
-        # handle different ways of passing observation
-        local_cond = None
-        global_cond = None
-        if self.obs_as_global_cond:
-            # condition through global feature
-            this_nobs = dict_apply(nobs, lambda x: x[:, :n_obs_steps, ...].reshape(-1, *x.shape[2:]))
-            nobs_features = self.obs_encoder(this_nobs)
-            # reshape back to B, Do
-            global_cond = nobs_features.reshape(bsize, -1)
-            # empty data for action
-            cond_data = torch.zeros(size=(bsize, horizon, action_dim), device=device, dtype=dtype)
-            cond_mask = torch.zeros_like(cond_data, dtype=torch.bool)
-        else:
-            # condition through impainting
-            this_nobs = dict_apply(nobs, lambda x: x[:, :n_obs_steps, ...].reshape(-1, *x.shape[2:]))
-            nobs_features = self.obs_encoder(this_nobs)
-            # reshape back to B, T, Do
-            nobs_features = nobs_features.reshape(bsize, n_obs_steps, -1)
-            cond_data = torch.zeros(size=(bsize, horizon, action_dim + obs_dim), device=device, dtype=dtype)
-            cond_mask = torch.zeros_like(cond_data, dtype=torch.bool)
-            cond_data[:, :n_obs_steps, action_dim:] = nobs_features
-            cond_mask[:, :n_obs_steps, action_dim:] = True
-
-        # run sampling
-        nsample = self.conditional_sample(
-            cond_data, cond_mask, local_cond=local_cond, global_cond=global_cond, **self.kwargs
-        )
-
-        action_pred = nsample[..., :action_dim]
-
-        # get action
-        start = n_obs_steps - 1
-        end = start + self.n_action_steps
-        action = action_pred[:, start:end]
-
-        result = {"action": action, "action_pred": action_pred}
-        return result
-
-    def compute_loss(self, batch):
-        assert "valid_mask" not in batch
-        nobs = batch["obs"]
-        nactions = batch["action"]
-        batch_size = nactions.shape[0]
-        horizon = nactions.shape[1]
-
-        # handle different ways of passing observation
-        local_cond = None
-        global_cond = None
-        trajectory = nactions
-        cond_data = trajectory
-        if self.obs_as_global_cond:
-            # reshape B, T, ... to B*T
-            this_nobs = dict_apply(nobs, lambda x: x[:, : self.n_obs_steps, ...].reshape(-1, *x.shape[2:]))
-            nobs_features = self.obs_encoder(this_nobs)
-            # reshape back to B, Do
-            global_cond = nobs_features.reshape(batch_size, -1)
-        else:
-            # reshape B, T, ... to B*T
-            this_nobs = dict_apply(nobs, lambda x: x.reshape(-1, *x.shape[2:]))
-            nobs_features = self.obs_encoder(this_nobs)
-            # reshape back to B, T, Do
-            nobs_features = nobs_features.reshape(batch_size, horizon, -1)
-            cond_data = torch.cat([nactions, nobs_features], dim=-1)
-            trajectory = cond_data.detach()
-
-        # generate impainting mask
-        condition_mask = self.mask_generator(trajectory.shape)
-
-        # Sample noise that we'll add to the images
-        noise = torch.randn(trajectory.shape, device=trajectory.device)
-        bsz = trajectory.shape[0]
-        # Sample a random timestep for each image
-        timesteps = torch.randint(
-            0, self.noise_scheduler.config.num_train_timesteps, (bsz,), device=trajectory.device
-        ).long()
-        # Add noise to the clean images according to the noise magnitude at each timestep
-        # (this is the forward diffusion process)
-        noisy_trajectory = self.noise_scheduler.add_noise(trajectory, noise, timesteps)
-
-        # compute loss mask
-        loss_mask = ~condition_mask
-
-        # apply conditioning
-        noisy_trajectory[condition_mask] = cond_data[condition_mask]
-
-        # Predict the noise residual
-        pred = self.model(noisy_trajectory, timesteps, local_cond=local_cond, global_cond=global_cond)
-
-        pred_type = self.noise_scheduler.config.prediction_type
-        if pred_type == "epsilon":
-            target = noise
-        elif pred_type == "sample":
-            target = trajectory
-        else:
-            raise ValueError(f"Unsupported prediction type {pred_type}")
-
-        loss = F.mse_loss(pred, target, reduction="none")
-        loss = loss * loss_mask.type(loss.dtype)
-        loss = reduce(loss, "b ... -> b (...)", "mean")
-        loss = loss.mean()
-        return loss

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -0,0 +1,765 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Columbia Artificial Intelligence, Robotics Lab,
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
+
+TODO(alexander-soare):
+  - Remove reliance on diffusers for DDPMScheduler and LR scheduler.
+"""
+
+import math
+from collections import deque
+from typing import Callable
+
+import einops
+import numpy as np
+import torch
+import torch.nn.functional as F  # noqa: N812
+import torchvision
+from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
+from torch import Tensor, nn
+
+from lerobot.common.constants import OBS_ENV, OBS_ROBOT
+from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+from lerobot.common.policies.utils import (
+    get_device_from_parameters,
+    get_dtype_from_parameters,
+    get_output_shape,
+    populate_queues,
+)
+
+
+class DiffusionPolicy(PreTrainedPolicy):
+    """
+    Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
+    (paper: https://arxiv.org/abs/2303.04137, code: https://github.com/real-stanford/diffusion_policy).
+    """
+
+    config_class = DiffusionConfig
+    name = "diffusion"
+
+    def __init__(
+        self,
+        config: DiffusionConfig,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_targets = Normalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+
+        # queues are populated during rollout of the policy, they contain the n latest observations and actions
+        self._queues = None
+
+        self.diffusion = DiffusionModel(config)
+
+        self.reset()
+
+    def get_optim_params(self) -> dict:
+        return self.diffusion.parameters()
+
+    def reset(self):
+        """Clear observation and action queues. Should be called on `env.reset()`"""
+        self._queues = {
+            "observation.state": deque(maxlen=self.config.n_obs_steps),
+            "action": deque(maxlen=self.config.n_action_steps),
+        }
+        if self.config.image_features:
+            self._queues["observation.images"] = deque(maxlen=self.config.n_obs_steps)
+        if self.config.env_state_feature:
+            self._queues["observation.environment_state"] = deque(maxlen=self.config.n_obs_steps)
+
+    @torch.no_grad
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method handles caching a history of observations and an action trajectory generated by the
+        underlying diffusion model. Here's how it works:
+          - `n_obs_steps` steps worth of observations are cached (for the first steps, the observation is
+            copied `n_obs_steps` times to fill the cache).
+          - The diffusion model generates `horizon` steps worth of actions.
+          - `n_action_steps` worth of actions are actually kept for execution, starting from the current step.
+        Schematically this looks like:
+            ----------------------------------------------------------------------------------------------
+            (legend: o = n_obs_steps, h = horizon, a = n_action_steps)
+            |timestep            | n-o+1 | n-o+2 | ..... | n     | ..... | n+a-1 | n+a   | ..... | n-o+h |
+            |observation is used | YES   | YES   | YES   | YES   | NO    | NO    | NO    | NO    | NO    |
+            |action is generated | YES   | YES   | YES   | YES   | YES   | YES   | YES   | YES   | YES   |
+            |action is used      | NO    | NO    | NO    | YES   | YES   | YES   | NO    | NO    | NO    |
+            ----------------------------------------------------------------------------------------------
+        Note that this means we require: `n_action_steps <= horizon - n_obs_steps + 1`. Also, note that
+        "horizon" may not the best name to describe what the variable actually means, because this period is
+        actually measured from the first observation which (if `n_obs_steps` > 1) happened in the past.
+        """
+        batch = self.normalize_inputs(batch)
+        if self.config.image_features:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.images"] = torch.stack(
+                [batch[key] for key in self.config.image_features], dim=-4
+            )
+        # Note: It's important that this happens after stacking the images into a single key.
+        self._queues = populate_queues(self._queues, batch)
+
+        if len(self._queues["action"]) == 0:
+            # stack n latest observations from the queue
+            batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
+            actions = self.diffusion.generate_actions(batch)
+
+            # TODO(rcadene): make above methods return output dictionary?
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            self._queues["action"].extend(actions.transpose(0, 1))
+
+        action = self._queues["action"].popleft()
+        return action
+
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, None]:
+        """Run the batch through the model and compute the loss for training or validation."""
+        batch = self.normalize_inputs(batch)
+        if self.config.image_features:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.images"] = torch.stack(
+                [batch[key] for key in self.config.image_features], dim=-4
+            )
+        batch = self.normalize_targets(batch)
+        loss = self.diffusion.compute_loss(batch)
+        # no output_dict so returning None
+        return loss, None
+
+
+def _make_noise_scheduler(name: str, **kwargs: dict) -> DDPMScheduler | DDIMScheduler:
+    """
+    Factory for noise scheduler instances of the requested type. All kwargs are passed
+    to the scheduler.
+    """
+    if name == "DDPM":
+        return DDPMScheduler(**kwargs)
+    elif name == "DDIM":
+        return DDIMScheduler(**kwargs)
+    else:
+        raise ValueError(f"Unsupported noise scheduler type {name}")
+
+
+class DiffusionModel(nn.Module):
+    def __init__(self, config: DiffusionConfig):
+        super().__init__()
+        self.config = config
+
+        # Build observation encoders (depending on which observations are provided).
+        global_cond_dim = self.config.robot_state_feature.shape[0]
+        if self.config.image_features:
+            num_images = len(self.config.image_features)
+            if self.config.use_separate_rgb_encoder_per_camera:
+                encoders = [DiffusionRgbEncoder(config) for _ in range(num_images)]
+                self.rgb_encoder = nn.ModuleList(encoders)
+                global_cond_dim += encoders[0].feature_dim * num_images
+            else:
+                self.rgb_encoder = DiffusionRgbEncoder(config)
+                global_cond_dim += self.rgb_encoder.feature_dim * num_images
+        if self.config.env_state_feature:
+            global_cond_dim += self.config.env_state_feature.shape[0]
+
+        self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
+
+        self.noise_scheduler = _make_noise_scheduler(
+            config.noise_scheduler_type,
+            num_train_timesteps=config.num_train_timesteps,
+            beta_start=config.beta_start,
+            beta_end=config.beta_end,
+            beta_schedule=config.beta_schedule,
+            clip_sample=config.clip_sample,
+            clip_sample_range=config.clip_sample_range,
+            prediction_type=config.prediction_type,
+        )
+
+        if config.num_inference_steps is None:
+            self.num_inference_steps = self.noise_scheduler.config.num_train_timesteps
+        else:
+            self.num_inference_steps = config.num_inference_steps
+
+    # ========= inference  ============
+    def conditional_sample(
+        self, batch_size: int, global_cond: Tensor | None = None, generator: torch.Generator | None = None
+    ) -> Tensor:
+        device = get_device_from_parameters(self)
+        dtype = get_dtype_from_parameters(self)
+
+        # Sample prior.
+        sample = torch.randn(
+            size=(batch_size, self.config.horizon, self.config.action_feature.shape[0]),
+            dtype=dtype,
+            device=device,
+            generator=generator,
+        )
+
+        self.noise_scheduler.set_timesteps(self.num_inference_steps)
+
+        for t in self.noise_scheduler.timesteps:
+            # Predict model output.
+            model_output = self.unet(
+                sample,
+                torch.full(sample.shape[:1], t, dtype=torch.long, device=sample.device),
+                global_cond=global_cond,
+            )
+            # Compute previous image: x_t -> x_t-1
+            sample = self.noise_scheduler.step(model_output, t, sample, generator=generator).prev_sample
+
+        return sample
+
+    def _prepare_global_conditioning(self, batch: dict[str, Tensor]) -> Tensor:
+        """Encode image features and concatenate them all together along with the state vector."""
+        batch_size, n_obs_steps = batch[OBS_ROBOT].shape[:2]
+        global_cond_feats = [batch[OBS_ROBOT]]
+        # Extract image features.
+        if self.config.image_features:
+            if self.config.use_separate_rgb_encoder_per_camera:
+                # Combine batch and sequence dims while rearranging to make the camera index dimension first.
+                images_per_camera = einops.rearrange(batch["observation.images"], "b s n ... -> n (b s) ...")
+                img_features_list = torch.cat(
+                    [
+                        encoder(images)
+                        for encoder, images in zip(self.rgb_encoder, images_per_camera, strict=True)
+                    ]
+                )
+                # Separate batch and sequence dims back out. The camera index dim gets absorbed into the
+                # feature dim (effectively concatenating the camera features).
+                img_features = einops.rearrange(
+                    img_features_list, "(n b s) ... -> b s (n ...)", b=batch_size, s=n_obs_steps
+                )
+            else:
+                # Combine batch, sequence, and "which camera" dims before passing to shared encoder.
+                img_features = self.rgb_encoder(
+                    einops.rearrange(batch["observation.images"], "b s n ... -> (b s n) ...")
+                )
+                # Separate batch dim and sequence dim back out. The camera index dim gets absorbed into the
+                # feature dim (effectively concatenating the camera features).
+                img_features = einops.rearrange(
+                    img_features, "(b s n) ... -> b s (n ...)", b=batch_size, s=n_obs_steps
+                )
+            global_cond_feats.append(img_features)
+
+        if self.config.env_state_feature:
+            global_cond_feats.append(batch[OBS_ENV])
+
+        # Concatenate features then flatten to (B, global_cond_dim).
+        return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)
+
+    def generate_actions(self, batch: dict[str, Tensor]) -> Tensor:
+        """
+        This function expects `batch` to have:
+        {
+            "observation.state": (B, n_obs_steps, state_dim)
+
+            "observation.images": (B, n_obs_steps, num_cameras, C, H, W)
+                AND/OR
+            "observation.environment_state": (B, environment_dim)
+        }
+        """
+        batch_size, n_obs_steps = batch["observation.state"].shape[:2]
+        assert n_obs_steps == self.config.n_obs_steps
+
+        # Encode image features and concatenate them all together along with the state vector.
+        global_cond = self._prepare_global_conditioning(batch)  # (B, global_cond_dim)
+
+        # run sampling
+        actions = self.conditional_sample(batch_size, global_cond=global_cond)
+
+        # Extract `n_action_steps` steps worth of actions (from the current observation).
+        start = n_obs_steps - 1
+        end = start + self.config.n_action_steps
+        actions = actions[:, start:end]
+
+        return actions
+
+    def compute_loss(self, batch: dict[str, Tensor]) -> Tensor:
+        """
+        This function expects `batch` to have (at least):
+        {
+            "observation.state": (B, n_obs_steps, state_dim)
+
+            "observation.images": (B, n_obs_steps, num_cameras, C, H, W)
+                AND/OR
+            "observation.environment_state": (B, environment_dim)
+
+            "action": (B, horizon, action_dim)
+            "action_is_pad": (B, horizon)
+        }
+        """
+        # Input validation.
+        assert set(batch).issuperset({"observation.state", "action", "action_is_pad"})
+        assert "observation.images" in batch or "observation.environment_state" in batch
+        n_obs_steps = batch["observation.state"].shape[1]
+        horizon = batch["action"].shape[1]
+        assert horizon == self.config.horizon
+        assert n_obs_steps == self.config.n_obs_steps
+
+        # Encode image features and concatenate them all together along with the state vector.
+        global_cond = self._prepare_global_conditioning(batch)  # (B, global_cond_dim)
+
+        # Forward diffusion.
+        trajectory = batch["action"]
+        # Sample noise to add to the trajectory.
+        eps = torch.randn(trajectory.shape, device=trajectory.device)
+        # Sample a random noising timestep for each item in the batch.
+        timesteps = torch.randint(
+            low=0,
+            high=self.noise_scheduler.config.num_train_timesteps,
+            size=(trajectory.shape[0],),
+            device=trajectory.device,
+        ).long()
+        # Add noise to the clean trajectories according to the noise magnitude at each timestep.
+        noisy_trajectory = self.noise_scheduler.add_noise(trajectory, eps, timesteps)
+
+        # Run the denoising network (that might denoise the trajectory, or attempt to predict the noise).
+        pred = self.unet(noisy_trajectory, timesteps, global_cond=global_cond)
+
+        # Compute the loss.
+        # The target is either the original trajectory, or the noise.
+        if self.config.prediction_type == "epsilon":
+            target = eps
+        elif self.config.prediction_type == "sample":
+            target = batch["action"]
+        else:
+            raise ValueError(f"Unsupported prediction type {self.config.prediction_type}")
+
+        loss = F.mse_loss(pred, target, reduction="none")
+
+        # Mask loss wherever the action is padded with copies (edges of the dataset trajectory).
+        if self.config.do_mask_loss_for_padding:
+            if "action_is_pad" not in batch:
+                raise ValueError(
+                    "You need to provide 'action_is_pad' in the batch when "
+                    f"{self.config.do_mask_loss_for_padding=}."
+                )
+            in_episode_bound = ~batch["action_is_pad"]
+            loss = loss * in_episode_bound.unsqueeze(-1)
+
+        return loss.mean()
+
+
+class SpatialSoftmax(nn.Module):
+    """
+    Spatial Soft Argmax operation described in "Deep Spatial Autoencoders for Visuomotor Learning" by Finn et al.
+    (https://arxiv.org/pdf/1509.06113). A minimal port of the robomimic implementation.
+
+    At a high level, this takes 2D feature maps (from a convnet/ViT) and returns the "center of mass"
+    of activations of each channel, i.e., keypoints in the image space for the policy to focus on.
+
+    Example: take feature maps of size (512x10x12). We generate a grid of normalized coordinates (10x12x2):
+    -----------------------------------------------------
+    | (-1., -1.)   | (-0.82, -1.)   | ... | (1., -1.)   |
+    | (-1., -0.78) | (-0.82, -0.78) | ... | (1., -0.78) |
+    | ...          | ...            | ... | ...         |
+    | (-1., 1.)    | (-0.82, 1.)    | ... | (1., 1.)    |
+    -----------------------------------------------------
+    This is achieved by applying channel-wise softmax over the activations (512x120) and computing the dot
+    product with the coordinates (120x2) to get expected points of maximal activation (512x2).
+
+    The example above results in 512 keypoints (corresponding to the 512 input channels). We can optionally
+    provide num_kp != None to control the number of keypoints. This is achieved by a first applying a learnable
+    linear mapping (in_channels, H, W) -> (num_kp, H, W).
+    """
+
+    def __init__(self, input_shape, num_kp=None):
+        """
+        Args:
+            input_shape (list): (C, H, W) input feature map shape.
+            num_kp (int): number of keypoints in output. If None, output will have the same number of channels as input.
+        """
+        super().__init__()
+
+        assert len(input_shape) == 3
+        self._in_c, self._in_h, self._in_w = input_shape
+
+        if num_kp is not None:
+            self.nets = torch.nn.Conv2d(self._in_c, num_kp, kernel_size=1)
+            self._out_c = num_kp
+        else:
+            self.nets = None
+            self._out_c = self._in_c
+
+        # we could use torch.linspace directly but that seems to behave slightly differently than numpy
+        # and causes a small degradation in pc_success of pre-trained models.
+        pos_x, pos_y = np.meshgrid(np.linspace(-1.0, 1.0, self._in_w), np.linspace(-1.0, 1.0, self._in_h))
+        pos_x = torch.from_numpy(pos_x.reshape(self._in_h * self._in_w, 1)).float()
+        pos_y = torch.from_numpy(pos_y.reshape(self._in_h * self._in_w, 1)).float()
+        # register as buffer so it's moved to the correct device.
+        self.register_buffer("pos_grid", torch.cat([pos_x, pos_y], dim=1))
+
+    def forward(self, features: Tensor) -> Tensor:
+        """
+        Args:
+            features: (B, C, H, W) input feature maps.
+        Returns:
+            (B, K, 2) image-space coordinates of keypoints.
+        """
+        if self.nets is not None:
+            features = self.nets(features)
+
+        # [B, K, H, W] -> [B * K, H * W] where K is number of keypoints
+        features = features.reshape(-1, self._in_h * self._in_w)
+        # 2d softmax normalization
+        attention = F.softmax(features, dim=-1)
+        # [B * K, H * W] x [H * W, 2] -> [B * K, 2] for spatial coordinate mean in x and y dimensions
+        expected_xy = attention @ self.pos_grid
+        # reshape to [B, K, 2]
+        feature_keypoints = expected_xy.view(-1, self._out_c, 2)
+
+        return feature_keypoints
+
+
+class DiffusionRgbEncoder(nn.Module):
+    """Encodes an RGB image into a 1D feature vector.
+
+    Includes the ability to normalize and crop the image first.
+    """
+
+    def __init__(self, config: DiffusionConfig):
+        super().__init__()
+        # Set up optional preprocessing.
+        if config.crop_shape is not None:
+            self.do_crop = True
+            # Always use center crop for eval
+            self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
+            if config.crop_is_random:
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
+            else:
+                self.maybe_random_crop = self.center_crop
+        else:
+            self.do_crop = False
+
+        # Set up backbone.
+        backbone_model = getattr(torchvision.models, config.vision_backbone)(
+            weights=config.pretrained_backbone_weights
+        )
+        # Note: This assumes that the layer4 feature map is children()[-3]
+        # TODO(alexander-soare): Use a safer alternative.
+        self.backbone = nn.Sequential(*(list(backbone_model.children())[:-2]))
+        if config.use_group_norm:
+            if config.pretrained_backbone_weights:
+                raise ValueError(
+                    "You can't replace BatchNorm in a pretrained model without ruining the weights!"
+                )
+            self.backbone = _replace_submodules(
+                root_module=self.backbone,
+                predicate=lambda x: isinstance(x, nn.BatchNorm2d),
+                func=lambda x: nn.GroupNorm(num_groups=x.num_features // 16, num_channels=x.num_features),
+            )
+
+        # Set up pooling and final layers.
+        # Use a dry run to get the feature map shape.
+        # The dummy input should take the number of image channels from `config.image_features` and it should
+        # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
+        # height and width from `config.image_features`.
+
+        # Note: we have a check in the config class to make sure all images have the same shape.
+        images_shape = next(iter(config.image_features.values())).shape
+        dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
+        dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
+        feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
+
+        self.pool = SpatialSoftmax(feature_map_shape, num_kp=config.spatial_softmax_num_keypoints)
+        self.feature_dim = config.spatial_softmax_num_keypoints * 2
+        self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
+        self.relu = nn.ReLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        Args:
+            x: (B, C, H, W) image tensor with pixel values in [0, 1].
+        Returns:
+            (B, D) image feature.
+        """
+        # Preprocess: maybe crop (if it was set up in the __init__).
+        if self.do_crop:
+            if self.training:  # noqa: SIM108
+                x = self.maybe_random_crop(x)
+            else:
+                # Always use center crop for eval.
+                x = self.center_crop(x)
+        # Extract backbone feature.
+        x = torch.flatten(self.pool(self.backbone(x)), start_dim=1)
+        # Final linear layer with non-linearity.
+        x = self.relu(self.out(x))
+        return x
+
+
+def _replace_submodules(
+    root_module: nn.Module, predicate: Callable[[nn.Module], bool], func: Callable[[nn.Module], nn.Module]
+) -> nn.Module:
+    """
+    Args:
+        root_module: The module for which the submodules need to be replaced
+        predicate: Takes a module as an argument and must return True if the that module is to be replaced.
+        func: Takes a module as an argument and returns a new module to replace it with.
+    Returns:
+        The root module with its submodules replaced.
+    """
+    if predicate(root_module):
+        return func(root_module)
+
+    replace_list = [k.split(".") for k, m in root_module.named_modules(remove_duplicate=True) if predicate(m)]
+    for *parents, k in replace_list:
+        parent_module = root_module
+        if len(parents) > 0:
+            parent_module = root_module.get_submodule(".".join(parents))
+        if isinstance(parent_module, nn.Sequential):
+            src_module = parent_module[int(k)]
+        else:
+            src_module = getattr(parent_module, k)
+        tgt_module = func(src_module)
+        if isinstance(parent_module, nn.Sequential):
+            parent_module[int(k)] = tgt_module
+        else:
+            setattr(parent_module, k, tgt_module)
+    # verify that all BN are replaced
+    assert not any(predicate(m) for _, m in root_module.named_modules(remove_duplicate=True))
+    return root_module
+
+
+class DiffusionSinusoidalPosEmb(nn.Module):
+    """1D sinusoidal positional embeddings as in Attention is All You Need."""
+
+    def __init__(self, dim: int):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x: Tensor) -> Tensor:
+        device = x.device
+        half_dim = self.dim // 2
+        emb = math.log(10000) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
+        emb = x.unsqueeze(-1) * emb.unsqueeze(0)
+        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
+        return emb
+
+
+class DiffusionConv1dBlock(nn.Module):
+    """Conv1d --> GroupNorm --> Mish"""
+
+    def __init__(self, inp_channels, out_channels, kernel_size, n_groups=8):
+        super().__init__()
+
+        self.block = nn.Sequential(
+            nn.Conv1d(inp_channels, out_channels, kernel_size, padding=kernel_size // 2),
+            nn.GroupNorm(n_groups, out_channels),
+            nn.Mish(),
+        )
+
+    def forward(self, x):
+        return self.block(x)
+
+
+class DiffusionConditionalUnet1d(nn.Module):
+    """A 1D convolutional UNet with FiLM modulation for conditioning.
+
+    Note: this removes local conditioning as compared to the original diffusion policy code.
+    """
+
+    def __init__(self, config: DiffusionConfig, global_cond_dim: int):
+        super().__init__()
+
+        self.config = config
+
+        # Encoder for the diffusion timestep.
+        self.diffusion_step_encoder = nn.Sequential(
+            DiffusionSinusoidalPosEmb(config.diffusion_step_embed_dim),
+            nn.Linear(config.diffusion_step_embed_dim, config.diffusion_step_embed_dim * 4),
+            nn.Mish(),
+            nn.Linear(config.diffusion_step_embed_dim * 4, config.diffusion_step_embed_dim),
+        )
+
+        # The FiLM conditioning dimension.
+        cond_dim = config.diffusion_step_embed_dim + global_cond_dim
+
+        # In channels / out channels for each downsampling block in the Unet's encoder. For the decoder, we
+        # just reverse these.
+        in_out = [(config.action_feature.shape[0], config.down_dims[0])] + list(
+            zip(config.down_dims[:-1], config.down_dims[1:], strict=True)
+        )
+
+        # Unet encoder.
+        common_res_block_kwargs = {
+            "cond_dim": cond_dim,
+            "kernel_size": config.kernel_size,
+            "n_groups": config.n_groups,
+            "use_film_scale_modulation": config.use_film_scale_modulation,
+        }
+        self.down_modules = nn.ModuleList([])
+        for ind, (dim_in, dim_out) in enumerate(in_out):
+            is_last = ind >= (len(in_out) - 1)
+            self.down_modules.append(
+                nn.ModuleList(
+                    [
+                        DiffusionConditionalResidualBlock1d(dim_in, dim_out, **common_res_block_kwargs),
+                        DiffusionConditionalResidualBlock1d(dim_out, dim_out, **common_res_block_kwargs),
+                        # Downsample as long as it is not the last block.
+                        nn.Conv1d(dim_out, dim_out, 3, 2, 1) if not is_last else nn.Identity(),
+                    ]
+                )
+            )
+
+        # Processing in the middle of the auto-encoder.
+        self.mid_modules = nn.ModuleList(
+            [
+                DiffusionConditionalResidualBlock1d(
+                    config.down_dims[-1], config.down_dims[-1], **common_res_block_kwargs
+                ),
+                DiffusionConditionalResidualBlock1d(
+                    config.down_dims[-1], config.down_dims[-1], **common_res_block_kwargs
+                ),
+            ]
+        )
+
+        # Unet decoder.
+        self.up_modules = nn.ModuleList([])
+        for ind, (dim_out, dim_in) in enumerate(reversed(in_out[1:])):
+            is_last = ind >= (len(in_out) - 1)
+            self.up_modules.append(
+                nn.ModuleList(
+                    [
+                        # dim_in * 2, because it takes the encoder's skip connection as well
+                        DiffusionConditionalResidualBlock1d(dim_in * 2, dim_out, **common_res_block_kwargs),
+                        DiffusionConditionalResidualBlock1d(dim_out, dim_out, **common_res_block_kwargs),
+                        # Upsample as long as it is not the last block.
+                        nn.ConvTranspose1d(dim_out, dim_out, 4, 2, 1) if not is_last else nn.Identity(),
+                    ]
+                )
+            )
+
+        self.final_conv = nn.Sequential(
+            DiffusionConv1dBlock(config.down_dims[0], config.down_dims[0], kernel_size=config.kernel_size),
+            nn.Conv1d(config.down_dims[0], config.action_feature.shape[0], 1),
+        )
+
+    def forward(self, x: Tensor, timestep: Tensor | int, global_cond=None) -> Tensor:
+        """
+        Args:
+            x: (B, T, input_dim) tensor for input to the Unet.
+            timestep: (B,) tensor of (timestep_we_are_denoising_from - 1).
+            global_cond: (B, global_cond_dim)
+            output: (B, T, input_dim)
+        Returns:
+            (B, T, input_dim) diffusion model prediction.
+        """
+        # For 1D convolutions we'll need feature dimension first.
+        x = einops.rearrange(x, "b t d -> b d t")
+
+        timesteps_embed = self.diffusion_step_encoder(timestep)
+
+        # If there is a global conditioning feature, concatenate it to the timestep embedding.
+        if global_cond is not None:
+            global_feature = torch.cat([timesteps_embed, global_cond], axis=-1)
+        else:
+            global_feature = timesteps_embed
+
+        # Run encoder, keeping track of skip features to pass to the decoder.
+        encoder_skip_features: list[Tensor] = []
+        for resnet, resnet2, downsample in self.down_modules:
+            x = resnet(x, global_feature)
+            x = resnet2(x, global_feature)
+            encoder_skip_features.append(x)
+            x = downsample(x)
+
+        for mid_module in self.mid_modules:
+            x = mid_module(x, global_feature)
+
+        # Run decoder, using the skip features from the encoder.
+        for resnet, resnet2, upsample in self.up_modules:
+            x = torch.cat((x, encoder_skip_features.pop()), dim=1)
+            x = resnet(x, global_feature)
+            x = resnet2(x, global_feature)
+            x = upsample(x)
+
+        x = self.final_conv(x)
+
+        x = einops.rearrange(x, "b d t -> b t d")
+        return x
+
+
+class DiffusionConditionalResidualBlock1d(nn.Module):
+    """ResNet style 1D convolutional block with FiLM modulation for conditioning."""
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        cond_dim: int,
+        kernel_size: int = 3,
+        n_groups: int = 8,
+        # Set to True to do scale modulation with FiLM as well as bias modulation (defaults to False meaning
+        # FiLM just modulates bias).
+        use_film_scale_modulation: bool = False,
+    ):
+        super().__init__()
+
+        self.use_film_scale_modulation = use_film_scale_modulation
+        self.out_channels = out_channels
+
+        self.conv1 = DiffusionConv1dBlock(in_channels, out_channels, kernel_size, n_groups=n_groups)
+
+        # FiLM modulation (https://arxiv.org/abs/1709.07871) outputs per-channel bias and (maybe) scale.
+        cond_channels = out_channels * 2 if use_film_scale_modulation else out_channels
+        self.cond_encoder = nn.Sequential(nn.Mish(), nn.Linear(cond_dim, cond_channels))
+
+        self.conv2 = DiffusionConv1dBlock(out_channels, out_channels, kernel_size, n_groups=n_groups)
+
+        # A final convolution for dimension matching the residual (if needed).
+        self.residual_conv = (
+            nn.Conv1d(in_channels, out_channels, 1) if in_channels != out_channels else nn.Identity()
+        )
+
+    def forward(self, x: Tensor, cond: Tensor) -> Tensor:
+        """
+        Args:
+            x: (B, in_channels, T)
+            cond: (B, cond_dim)
+        Returns:
+            (B, out_channels, T)
+        """
+        out = self.conv1(x)
+
+        # Get condition embedding. Unsqueeze for broadcasting to `out`, resulting in (B, out_channels, 1).
+        cond_embed = self.cond_encoder(cond).unsqueeze(-1)
+        if self.use_film_scale_modulation:
+            # Treat the embedding as a list of scales and biases.
+            scale = cond_embed[:, : self.out_channels]
+            bias = cond_embed[:, self.out_channels :]
+            out = scale * out + bias
+        else:
+            # Treat the embedding as biases.
+            out = out + cond_embed
+
+        out = self.conv2(out)
+        out = out + self.residual_conv(x)
+        return out

lerobot/common/policies/diffusion/multi_image_obs_encoder.py

@@ -1,189 +0,0 @@
-import copy
-from typing import Dict, Tuple, Union
-
-import torch
-import torch.nn as nn
-import torchvision
-
-from diffusion_policy.common.pytorch_util import replace_submodules
-from diffusion_policy.model.common.module_attr_mixin import ModuleAttrMixin
-from diffusion_policy.model.vision.crop_randomizer import CropRandomizer
-
-
-class MultiImageObsEncoder(ModuleAttrMixin):
-    def __init__(
-        self,
-        shape_meta: dict,
-        rgb_model: Union[nn.Module, Dict[str, nn.Module]],
-        resize_shape: Union[Tuple[int, int], Dict[str, tuple], None] = None,
-        crop_shape: Union[Tuple[int, int], Dict[str, tuple], None] = None,
-        random_crop: bool = True,
-        # replace BatchNorm with GroupNorm
-        use_group_norm: bool = False,
-        # use single rgb model for all rgb inputs
-        share_rgb_model: bool = False,
-        # renormalize rgb input with imagenet normalization
-        # assuming input in [0,1]
-        imagenet_norm: bool = False,
-    ):
-        """
-        Assumes rgb input: B,C,H,W
-        Assumes low_dim input: B,D
-        """
-        super().__init__()
-
-        rgb_keys = []
-        low_dim_keys = []
-        key_model_map = nn.ModuleDict()
-        key_transform_map = nn.ModuleDict()
-        key_shape_map = {}
-
-        # handle sharing vision backbone
-        if share_rgb_model:
-            assert isinstance(rgb_model, nn.Module)
-            key_model_map["rgb"] = rgb_model
-
-        obs_shape_meta = shape_meta["obs"]
-        for key, attr in obs_shape_meta.items():
-            shape = tuple(attr["shape"])
-            type = attr.get("type", "low_dim")
-            key_shape_map[key] = shape
-            if type == "rgb":
-                rgb_keys.append(key)
-                # configure model for this key
-                this_model = None
-                if not share_rgb_model:
-                    if isinstance(rgb_model, dict):
-                        # have provided model for each key
-                        this_model = rgb_model[key]
-                    else:
-                        assert isinstance(rgb_model, nn.Module)
-                        # have a copy of the rgb model
-                        this_model = copy.deepcopy(rgb_model)
-
-                if this_model is not None:
-                    if use_group_norm:
-                        this_model = replace_submodules(
-                            root_module=this_model,
-                            predicate=lambda x: isinstance(x, nn.BatchNorm2d),
-                            func=lambda x: nn.GroupNorm(
-                                num_groups=x.num_features // 16, num_channels=x.num_features
-                            ),
-                        )
-                    key_model_map[key] = this_model
-
-                # configure resize
-                input_shape = shape
-                this_resizer = nn.Identity()
-                if resize_shape is not None:
-                    if isinstance(resize_shape, dict):
-                        h, w = resize_shape[key]
-                    else:
-                        h, w = resize_shape
-                    this_resizer = torchvision.transforms.Resize(size=(h, w))
-                    input_shape = (shape[0], h, w)
-
-                # configure randomizer
-                this_randomizer = nn.Identity()
-                if crop_shape is not None:
-                    if isinstance(crop_shape, dict):
-                        h, w = crop_shape[key]
-                    else:
-                        h, w = crop_shape
-                    if random_crop:
-                        this_randomizer = CropRandomizer(
-                            input_shape=input_shape, crop_height=h, crop_width=w, num_crops=1, pos_enc=False
-                        )
-                    else:
-                        this_normalizer = torchvision.transforms.CenterCrop(size=(h, w))
-                # configure normalizer
-                this_normalizer = nn.Identity()
-                if imagenet_norm:
-                    # TODO(rcadene): move normalizer to dataset and env
-                    this_normalizer = torchvision.transforms.Normalize(
-                        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
-                    )
-
-                this_transform = nn.Sequential(this_resizer, this_randomizer, this_normalizer)
-                key_transform_map[key] = this_transform
-            elif type == "low_dim":
-                low_dim_keys.append(key)
-            else:
-                raise RuntimeError(f"Unsupported obs type: {type}")
-        rgb_keys = sorted(rgb_keys)
-        low_dim_keys = sorted(low_dim_keys)
-
-        self.shape_meta = shape_meta
-        self.key_model_map = key_model_map
-        self.key_transform_map = key_transform_map
-        self.share_rgb_model = share_rgb_model
-        self.rgb_keys = rgb_keys
-        self.low_dim_keys = low_dim_keys
-        self.key_shape_map = key_shape_map
-
-    def forward(self, obs_dict):
-        batch_size = None
-        features = []
-        # process rgb input
-        if self.share_rgb_model:
-            # pass all rgb obs to rgb model
-            imgs = []
-            for key in self.rgb_keys:
-                img = obs_dict[key]
-                if batch_size is None:
-                    batch_size = img.shape[0]
-                else:
-                    assert batch_size == img.shape[0]
-                assert img.shape[1:] == self.key_shape_map[key]
-                img = self.key_transform_map[key](img)
-                imgs.append(img)
-            # (N*B,C,H,W)
-            imgs = torch.cat(imgs, dim=0)
-            # (N*B,D)
-            feature = self.key_model_map["rgb"](imgs)
-            # (N,B,D)
-            feature = feature.reshape(-1, batch_size, *feature.shape[1:])
-            # (B,N,D)
-            feature = torch.moveaxis(feature, 0, 1)
-            # (B,N*D)
-            feature = feature.reshape(batch_size, -1)
-            features.append(feature)
-        else:
-            # run each rgb obs to independent models
-            for key in self.rgb_keys:
-                img = obs_dict[key]
-                if batch_size is None:
-                    batch_size = img.shape[0]
-                else:
-                    assert batch_size == img.shape[0]
-                assert img.shape[1:] == self.key_shape_map[key]
-                img = self.key_transform_map[key](img)
-                feature = self.key_model_map[key](img)
-                features.append(feature)
-
-        # process lowdim input
-        for key in self.low_dim_keys:
-            data = obs_dict[key]
-            if batch_size is None:
-                batch_size = data.shape[0]
-            else:
-                assert batch_size == data.shape[0]
-            assert data.shape[1:] == self.key_shape_map[key]
-            features.append(data)
-
-        # concatenate all features
-        result = torch.cat(features, dim=-1)
-        return result
-
-    @torch.no_grad()
-    def output_shape(self):
-        example_obs_dict = {}
-        obs_shape_meta = self.shape_meta["obs"]
-        batch_size = 1
-        for key, attr in obs_shape_meta.items():
-            shape = tuple(attr["shape"])
-            this_obs = torch.zeros((batch_size,) + shape, dtype=self.dtype, device=self.device)
-            example_obs_dict[key] = this_obs
-        example_output = self.forward(example_obs_dict)
-        output_shape = example_output.shape[1:]
-        return output_shape

lerobot/common/policies/diffusion/policy.py

@@ -1,199 +0,0 @@
-import copy
-import time
-
-import hydra
-import torch
-import torch.nn as nn
-from diffusion_policy.model.common.lr_scheduler import get_scheduler
-
-from .diffusion_unet_image_policy import DiffusionUnetImagePolicy
-from .multi_image_obs_encoder import MultiImageObsEncoder
-
-
-class DiffusionPolicy(nn.Module):
-    def __init__(
-        self,
-        cfg,
-        cfg_device,
-        cfg_noise_scheduler,
-        cfg_rgb_model,
-        cfg_obs_encoder,
-        cfg_optimizer,
-        cfg_ema,
-        shape_meta: dict,
-        horizon,
-        n_action_steps,
-        n_obs_steps,
-        num_inference_steps=None,
-        obs_as_global_cond=True,
-        diffusion_step_embed_dim=256,
-        down_dims=(256, 512, 1024),
-        kernel_size=5,
-        n_groups=8,
-        cond_predict_scale=True,
-        # parameters passed to step
-        **kwargs,
-    ):
-        super().__init__()
-        self.cfg = cfg
-
-        noise_scheduler = hydra.utils.instantiate(cfg_noise_scheduler)
-        rgb_model = hydra.utils.instantiate(cfg_rgb_model)
-        obs_encoder = MultiImageObsEncoder(
-            rgb_model=rgb_model,
-            **cfg_obs_encoder,
-        )
-
-        self.diffusion = DiffusionUnetImagePolicy(
-            shape_meta=shape_meta,
-            noise_scheduler=noise_scheduler,
-            obs_encoder=obs_encoder,
-            horizon=horizon,
-            n_action_steps=n_action_steps,
-            n_obs_steps=n_obs_steps,
-            num_inference_steps=num_inference_steps,
-            obs_as_global_cond=obs_as_global_cond,
-            diffusion_step_embed_dim=diffusion_step_embed_dim,
-            down_dims=down_dims,
-            kernel_size=kernel_size,
-            n_groups=n_groups,
-            cond_predict_scale=cond_predict_scale,
-            # parameters passed to step
-            **kwargs,
-        )
-
-        self.device = torch.device(cfg_device)
-        if torch.cuda.is_available() and cfg_device == "cuda":
-            self.diffusion.cuda()
-
-        self.ema = None
-        if self.cfg.use_ema:
-            self.ema = hydra.utils.instantiate(
-                cfg_ema,
-                model=copy.deepcopy(self.diffusion),
-            )
-
-        self.optimizer = hydra.utils.instantiate(
-            cfg_optimizer,
-            params=self.diffusion.parameters(),
-        )
-
-        # TODO(rcadene): modify lr scheduler so that it doesnt depend on epochs but steps
-        self.global_step = 0
-
-        # configure lr scheduler
-        self.lr_scheduler = get_scheduler(
-            cfg.lr_scheduler,
-            optimizer=self.optimizer,
-            num_warmup_steps=cfg.lr_warmup_steps,
-            num_training_steps=cfg.offline_steps,
-            # pytorch assumes stepping LRScheduler every epoch
-            # however huggingface diffusers steps it every batch
-            last_epoch=self.global_step - 1,
-        )
-
-    @torch.no_grad()
-    def forward(self, observation, step_count):
-        # TODO(rcadene): remove unused step_count
-        del step_count
-
-        # TODO(rcadene): remove unsqueeze hack to add bsize=1
-        observation["image"] = observation["image"].unsqueeze(0)
-        observation["state"] = observation["state"].unsqueeze(0)
-
-        obs_dict = {
-            "image": observation["image"],
-            "agent_pos": observation["state"],
-        }
-        out = self.diffusion.predict_action(obs_dict)
-
-        action = out["action"].squeeze(0)
-        return action
-
-    def update(self, replay_buffer, step):
-        start_time = time.time()
-
-        self.diffusion.train()
-
-        num_slices = self.cfg.batch_size
-        batch_size = self.cfg.horizon * num_slices
-
-        assert batch_size % self.cfg.horizon == 0
-        assert batch_size % num_slices == 0
-
-        def process_batch(batch, horizon, num_slices):
-            # trajectory t = 64, horizon h = 16
-            # (t h) ... -> t h ...
-            batch = batch.reshape(num_slices, horizon)  # .transpose(1, 0).contiguous()
-
-            # |-1|0|1|2|3|4|5|6|7|8|9|10|11|12|13|14| timestamps: 16
-            #  |o|o|                                  observations: 2
-            #  | |a|a|a|a|a|a|a|a|                    actions executed: 8
-            #  |p|p|p|p|p|p|p|p|p|p|p| p| p| p| p| p| actions predicted: 16
-            # note: we predict the action needed to go from t=-1 to t=0 similarly to an inverse kinematic model
-
-            image = batch["observation", "image"]
-            state = batch["observation", "state"]
-            action = batch["action"]
-            assert image.shape[1] == horizon
-            assert state.shape[1] == horizon
-            assert action.shape[1] == horizon
-
-            if not (horizon == 16 and self.cfg.n_obs_steps == 2):
-                raise NotImplementedError()
-
-            # keep first 2 observations of the slice corresponding to t=[-1,0]
-            image = image[:, : self.cfg.n_obs_steps]
-            state = state[:, : self.cfg.n_obs_steps]
-
-            out = {
-                "obs": {
-                    "image": image.to(self.device, non_blocking=True),
-                    "agent_pos": state.to(self.device, non_blocking=True),
-                },
-                "action": action.to(self.device, non_blocking=True),
-            }
-            return out
-
-        batch = replay_buffer.sample(batch_size)
-        batch = process_batch(batch, self.cfg.horizon, num_slices)
-
-        data_s = time.time() - start_time
-
-        loss = self.diffusion.compute_loss(batch)
-        loss.backward()
-
-        grad_norm = torch.nn.utils.clip_grad_norm_(
-            self.diffusion.parameters(),
-            self.cfg.grad_clip_norm,
-            error_if_nonfinite=False,
-        )
-
-        self.optimizer.step()
-        self.optimizer.zero_grad()
-        self.lr_scheduler.step()
-
-        if self.ema is not None:
-            self.ema.step(self.diffusion)
-
-        info = {
-            "loss": loss.item(),
-            "grad_norm": float(grad_norm),
-            "lr": self.lr_scheduler.get_last_lr()[0],
-            "data_s": data_s,
-            "update_s": time.time() - start_time,
-        }
-
-        # TODO(rcadene): remove hardcoding
-        # in diffusion_policy, len(dataloader) is 168 for a batch_size of 64
-        if step % 168 == 0:
-            self.global_step += 1
-
-        return info
-
-    def save(self, fp):
-        torch.save(self.state_dict(), fp)
-
-    def load(self, fp):
-        d = torch.load(fp)
-        self.load_state_dict(d)

lerobot/common/policies/factory.py

@@ -1,34 +1,157 @@
-def make_policy(cfg):
-    if cfg.policy.name == "tdmpc":
-        from lerobot.common.policies.tdmpc import TDMPC
+#!/usr/bin/env python
 
-        policy = TDMPC(cfg.policy, cfg.device)
-    elif cfg.policy.name == "diffusion":
-        from lerobot.common.policies.diffusion.policy import DiffusionPolicy
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
-        policy = DiffusionPolicy(
-            cfg=cfg.policy,
-            cfg_device=cfg.device,
-            cfg_noise_scheduler=cfg.noise_scheduler,
-            cfg_rgb_model=cfg.rgb_model,
-            cfg_obs_encoder=cfg.obs_encoder,
-            cfg_optimizer=cfg.optimizer,
-            cfg_ema=cfg.ema,
-            n_action_steps=cfg.n_action_steps + cfg.n_latency_steps,
-            **cfg.policy,
-        )
+import logging
+
+from torch import nn
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.common.datasets.utils import dataset_to_policy_features
+from lerobot.common.envs.configs import EnvConfig
+from lerobot.common.envs.utils import env_to_policy_features
+from lerobot.common.policies.act.configuration_act import ACTConfig
+from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
+from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType
+
+
+def get_policy_class(name: str) -> PreTrainedPolicy:
+    """Get the policy's class and config class given a name (matching the policy class' `name` attribute)."""
+    if name == "tdmpc":
+        from lerobot.common.policies.tdmpc.modeling_tdmpc import TDMPCPolicy
+
+        return TDMPCPolicy
+    elif name == "diffusion":
+        from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+
+        return DiffusionPolicy
+    elif name == "act":
+        from lerobot.common.policies.act.modeling_act import ACTPolicy
+
+        return ACTPolicy
+    elif name == "vqbet":
+        from lerobot.common.policies.vqbet.modeling_vqbet import VQBeTPolicy
+
+        return VQBeTPolicy
+    elif name == "pi0":
+        from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
+
+        return PI0Policy
+    elif name == "pi0fast":
+        from lerobot.common.policies.pi0fast.modeling_pi0fast import PI0FASTPolicy
+
+        return PI0FASTPolicy
     else:
-        raise ValueError(cfg.policy.name)
+        raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
-    if cfg.policy.pretrained_model_path:
-        # TODO(rcadene): hack for old pretrained models from fowm
-        if cfg.policy.name == "tdmpc" and "fowm" in cfg.policy.pretrained_model_path:
-            if "offline" in cfg.pretrained_model_path:
-                policy.step[0] = 25000
-            elif "final" in cfg.pretrained_model_path:
-                policy.step[0] = 100000
-            else:
-                raise NotImplementedError()
-        policy.load(cfg.policy.pretrained_model_path)
+
+def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
+    if policy_type == "tdmpc":
+        return TDMPCConfig(**kwargs)
+    elif policy_type == "diffusion":
+        return DiffusionConfig(**kwargs)
+    elif policy_type == "act":
+        return ACTConfig(**kwargs)
+    elif policy_type == "vqbet":
+        return VQBeTConfig(**kwargs)
+    elif policy_type == "pi0":
+        return PI0Config(**kwargs)
+    elif policy_type == "pi0fast":
+        return PI0FASTConfig(**kwargs)
+    else:
+        raise ValueError(f"Policy type '{policy_type}' is not available.")
+
+
+def make_policy(
+    cfg: PreTrainedConfig,
+    ds_meta: LeRobotDatasetMetadata | None = None,
+    env_cfg: EnvConfig | None = None,
+) -> PreTrainedPolicy:
+    """Make an instance of a policy class.
+
+    This function exists because (for now) we need to parse features from either a dataset or an environment
+    in order to properly dimension and instantiate a policy for that dataset or environment.
+
+    Args:
+        cfg (PreTrainedConfig): The config of the policy to make. If `pretrained_path` is set, the policy will
+            be loaded with the weights from that path.
+        ds_meta (LeRobotDatasetMetadata | None, optional): Dataset metadata to take input/output shapes and
+            statistics to use for (un)normalization of inputs/outputs in the policy. Defaults to None.
+        env_cfg (EnvConfig | None, optional): The config of a gym environment to parse features from. Must be
+            provided if ds_meta is not. Defaults to None.
+
+    Raises:
+        ValueError: Either ds_meta or env and env_cfg must be provided.
+        NotImplementedError: if the policy.type is 'vqbet' and the policy device 'mps' (due to an incompatibility)
+
+    Returns:
+        PreTrainedPolicy: _description_
+    """
+    if bool(ds_meta) == bool(env_cfg):
+        raise ValueError("Either one of a dataset metadata or a sim env must be provided.")
+
+    # NOTE: Currently, if you try to run vqbet with mps backend, you'll get this error.
+    # TODO(aliberts, rcadene): Implement a check_backend_compatibility in policies?
+    # NotImplementedError: The operator 'aten::unique_dim' is not currently implemented for the MPS device. If
+    # you want this op to be added in priority during the prototype phase of this feature, please comment on
+    # https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment
+    # variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be
+    # slower than running natively on MPS.
+    if cfg.type == "vqbet" and cfg.device == "mps":
+        raise NotImplementedError(
+            "Current implementation of VQBeT does not support `mps` backend. "
+            "Please use `cpu` or `cuda` backend."
+        )
+
+    policy_cls = get_policy_class(cfg.type)
+
+    kwargs = {}
+    if ds_meta is not None:
+        features = dataset_to_policy_features(ds_meta.features)
+        kwargs["dataset_stats"] = ds_meta.stats
+    else:
+        if not cfg.pretrained_path:
+            logging.warning(
+                "You are instantiating a policy from scratch and its features are parsed from an environment "
+                "rather than a dataset. Normalization modules inside the policy will have infinite values "
+                "by default without stats from a dataset."
+            )
+        features = env_to_policy_features(env_cfg)
+
+    cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
+    kwargs["config"] = cfg
+
+    if cfg.pretrained_path:
+        # Load a pretrained policy and override the config if needed (for example, if there are inference-time
+        # hyperparameters that we want to vary).
+        kwargs["pretrained_name_or_path"] = cfg.pretrained_path
+        policy = policy_cls.from_pretrained(**kwargs)
+    else:
+        # Make a fresh policy.
+        policy = policy_cls(**kwargs)
+
+    policy.to(cfg.device)
+    assert isinstance(policy, nn.Module)
+
+    # policy = torch.compile(policy, mode="reduce-overhead")
 
     return policy

lerobot/common/policies/normalize.py

@@ -0,0 +1,254 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import numpy as np
+import torch
+from torch import Tensor, nn
+
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+
+
+def create_stats_buffers(
+    features: dict[str, PolicyFeature],
+    norm_map: dict[str, NormalizationMode],
+    stats: dict[str, dict[str, Tensor]] | None = None,
+) -> dict[str, dict[str, nn.ParameterDict]]:
+    """
+    Create buffers per modality (e.g. "observation.image", "action") containing their mean, std, min, max
+    statistics.
+
+    Args: (see Normalize and Unnormalize)
+
+    Returns:
+        dict: A dictionary where keys are modalities and values are `nn.ParameterDict` containing
+            `nn.Parameters` set to `requires_grad=False`, suitable to not be updated during backpropagation.
+    """
+    stats_buffers = {}
+
+    for key, ft in features.items():
+        norm_mode = norm_map.get(ft.type, NormalizationMode.IDENTITY)
+        if norm_mode is NormalizationMode.IDENTITY:
+            continue
+
+        assert isinstance(norm_mode, NormalizationMode)
+
+        shape = tuple(ft.shape)
+
+        if ft.type is FeatureType.VISUAL:
+            # sanity checks
+            assert len(shape) == 3, f"number of dimensions of {key} != 3 ({shape=}"
+            c, h, w = shape
+            assert c < h and c < w, f"{key} is not channel first ({shape=})"
+            # override image shape to be invariant to height and width
+            shape = (c, 1, 1)
+
+        # Note: we initialize mean, std, min, max to infinity. They should be overwritten
+        # downstream by `stats` or `policy.load_state_dict`, as expected. During forward,
+        # we assert they are not infinity anymore.
+
+        buffer = {}
+        if norm_mode is NormalizationMode.MEAN_STD:
+            mean = torch.ones(shape, dtype=torch.float32) * torch.inf
+            std = torch.ones(shape, dtype=torch.float32) * torch.inf
+            buffer = nn.ParameterDict(
+                {
+                    "mean": nn.Parameter(mean, requires_grad=False),
+                    "std": nn.Parameter(std, requires_grad=False),
+                }
+            )
+        elif norm_mode is NormalizationMode.MIN_MAX:
+            min = torch.ones(shape, dtype=torch.float32) * torch.inf
+            max = torch.ones(shape, dtype=torch.float32) * torch.inf
+            buffer = nn.ParameterDict(
+                {
+                    "min": nn.Parameter(min, requires_grad=False),
+                    "max": nn.Parameter(max, requires_grad=False),
+                }
+            )
+
+        # TODO(aliberts, rcadene): harmonize this to only use one framework (np or torch)
+        if stats:
+            if isinstance(stats[key]["mean"], np.ndarray):
+                if norm_mode is NormalizationMode.MEAN_STD:
+                    buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(dtype=torch.float32)
+                    buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(dtype=torch.float32)
+                elif norm_mode is NormalizationMode.MIN_MAX:
+                    buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=torch.float32)
+                    buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=torch.float32)
+            elif isinstance(stats[key]["mean"], torch.Tensor):
+                # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
+                # tensors anywhere (for example, when we use the same stats for normalization and
+                # unnormalization). See the logic here
+                # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
+                if norm_mode is NormalizationMode.MEAN_STD:
+                    buffer["mean"].data = stats[key]["mean"].clone().to(dtype=torch.float32)
+                    buffer["std"].data = stats[key]["std"].clone().to(dtype=torch.float32)
+                elif norm_mode is NormalizationMode.MIN_MAX:
+                    buffer["min"].data = stats[key]["min"].clone().to(dtype=torch.float32)
+                    buffer["max"].data = stats[key]["max"].clone().to(dtype=torch.float32)
+            else:
+                type_ = type(stats[key]["mean"])
+                raise ValueError(f"np.ndarray or torch.Tensor expected, but type is '{type_}' instead.")
+
+        stats_buffers[key] = buffer
+    return stats_buffers
+
+
+def _no_stats_error_str(name: str) -> str:
+    return (
+        f"`{name}` is infinity. You should either initialize with `stats` as an argument, or use a "
+        "pretrained model."
+    )
+
+
+class Normalize(nn.Module):
+    """Normalizes data (e.g. "observation.image") for more stable and faster convergence during training."""
+
+    def __init__(
+        self,
+        features: dict[str, PolicyFeature],
+        norm_map: dict[str, NormalizationMode],
+        stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            shapes (dict): A dictionary where keys are input modalities (e.g. "observation.image") and values
+            are their shapes (e.g. `[3,96,96]`]). These shapes are used to create the tensor buffer containing
+            mean, std, min, max statistics. If the provided `shapes` contain keys related to images, the shape
+            is adjusted to be invariant to height and width, assuming a channel-first (c, h, w) format.
+            modes (dict): A dictionary where keys are output modalities (e.g. "observation.image") and values
+                are their normalization modes among:
+                    - "mean_std": subtract the mean and divide by standard deviation.
+                    - "min_max": map to [-1, 1] range.
+            stats (dict, optional): A dictionary where keys are output modalities (e.g. "observation.image")
+                and values are dictionaries of statistic types and their values (e.g.
+                `{"mean": torch.randn(3,1,1)}, "std": torch.randn(3,1,1)}`). If provided, as expected for
+                training the model for the first time, these statistics will overwrite the default buffers. If
+                not provided, as expected for finetuning or evaluation, the default buffers should to be
+                overwritten by a call to `policy.load_state_dict(state_dict)`. That way, initializing the
+                dataset is not needed to get the stats, since they are already in the policy state_dict.
+        """
+        super().__init__()
+        self.features = features
+        self.norm_map = norm_map
+        self.stats = stats
+        stats_buffers = create_stats_buffers(features, norm_map, stats)
+        for key, buffer in stats_buffers.items():
+            setattr(self, "buffer_" + key.replace(".", "_"), buffer)
+
+    # TODO(rcadene): should we remove torch.no_grad?
+    @torch.no_grad
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        batch = dict(batch)  # shallow copy avoids mutating the input batch
+        for key, ft in self.features.items():
+            if key not in batch:
+                # FIXME(aliberts, rcadene): This might lead to silent fail!
+                continue
+
+            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
+            if norm_mode is NormalizationMode.IDENTITY:
+                continue
+
+            buffer = getattr(self, "buffer_" + key.replace(".", "_"))
+
+            if norm_mode is NormalizationMode.MEAN_STD:
+                mean = buffer["mean"]
+                std = buffer["std"]
+                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
+                assert not torch.isinf(std).any(), _no_stats_error_str("std")
+                batch[key] = (batch[key] - mean) / (std + 1e-8)
+            elif norm_mode is NormalizationMode.MIN_MAX:
+                min = buffer["min"]
+                max = buffer["max"]
+                assert not torch.isinf(min).any(), _no_stats_error_str("min")
+                assert not torch.isinf(max).any(), _no_stats_error_str("max")
+                # normalize to [0,1]
+                batch[key] = (batch[key] - min) / (max - min + 1e-8)
+                # normalize to [-1, 1]
+                batch[key] = batch[key] * 2 - 1
+            else:
+                raise ValueError(norm_mode)
+        return batch
+
+
+class Unnormalize(nn.Module):
+    """
+    Similar to `Normalize` but unnormalizes output data (e.g. `{"action": torch.randn(b,c)}`) in their
+    original range used by the environment.
+    """
+
+    def __init__(
+        self,
+        features: dict[str, PolicyFeature],
+        norm_map: dict[str, NormalizationMode],
+        stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            shapes (dict): A dictionary where keys are input modalities (e.g. "observation.image") and values
+            are their shapes (e.g. `[3,96,96]`]). These shapes are used to create the tensor buffer containing
+            mean, std, min, max statistics. If the provided `shapes` contain keys related to images, the shape
+            is adjusted to be invariant to height and width, assuming a channel-first (c, h, w) format.
+            modes (dict): A dictionary where keys are output modalities (e.g. "observation.image") and values
+                are their normalization modes among:
+                    - "mean_std": subtract the mean and divide by standard deviation.
+                    - "min_max": map to [-1, 1] range.
+            stats (dict, optional): A dictionary where keys are output modalities (e.g. "observation.image")
+                and values are dictionaries of statistic types and their values (e.g.
+                `{"mean": torch.randn(3,1,1)}, "std": torch.randn(3,1,1)}`). If provided, as expected for
+                training the model for the first time, these statistics will overwrite the default buffers. If
+                not provided, as expected for finetuning or evaluation, the default buffers should to be
+                overwritten by a call to `policy.load_state_dict(state_dict)`. That way, initializing the
+                dataset is not needed to get the stats, since they are already in the policy state_dict.
+        """
+        super().__init__()
+        self.features = features
+        self.norm_map = norm_map
+        self.stats = stats
+        # `self.buffer_observation_state["mean"]` contains `torch.tensor(state_dim)`
+        stats_buffers = create_stats_buffers(features, norm_map, stats)
+        for key, buffer in stats_buffers.items():
+            setattr(self, "buffer_" + key.replace(".", "_"), buffer)
+
+    # TODO(rcadene): should we remove torch.no_grad?
+    @torch.no_grad
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        batch = dict(batch)  # shallow copy avoids mutating the input batch
+        for key, ft in self.features.items():
+            if key not in batch:
+                continue
+
+            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
+            if norm_mode is NormalizationMode.IDENTITY:
+                continue
+
+            buffer = getattr(self, "buffer_" + key.replace(".", "_"))
+
+            if norm_mode is NormalizationMode.MEAN_STD:
+                mean = buffer["mean"]
+                std = buffer["std"]
+                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
+                assert not torch.isinf(std).any(), _no_stats_error_str("std")
+                batch[key] = batch[key] * std + mean
+            elif norm_mode is NormalizationMode.MIN_MAX:
+                min = buffer["min"]
+                max = buffer["max"]
+                assert not torch.isinf(min).any(), _no_stats_error_str("min")
+                assert not torch.isinf(max).any(), _no_stats_error_str("max")
+                batch[key] = (batch[key] + 1) / 2
+                batch[key] = batch[key] * (max - min) + min
+            else:
+                raise ValueError(norm_mode)
+        return batch

lerobot/common/policies/pi0/configuration_pi0.py

@@ -0,0 +1,149 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.common.optim.optimizers import AdamWConfig
+from lerobot.common.optim.schedulers import (
+    CosineDecayWithWarmupSchedulerConfig,
+)
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+
+
+@PreTrainedConfig.register_subclass("pi0")
+@dataclass
+class PI0Config(PreTrainedConfig):
+    # Input / output structure.
+    n_obs_steps: int = 1
+    chunk_size: int = 50
+    n_action_steps: int = 50
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MEAN_STD,
+        }
+    )
+
+    # Shorter state and action vectors will be padded
+    max_state_dim: int = 32
+    max_action_dim: int = 32
+
+    # Image preprocessing
+    resize_imgs_with_padding: tuple[int, int] = (224, 224)
+
+    # Add empty images. Used by pi0_aloha_sim which adds the empty
+    # left and right wrist cameras in addition to the top camera.
+    empty_cameras: int = 0
+
+    # Converts the joint and gripper values from the standard Aloha space to
+    # the space used by the pi internal runtime which was used to train the base model.
+    adapt_to_pi_aloha: bool = False
+
+    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
+    # Gripper dimensions will remain in absolute values.
+    use_delta_joint_actions_aloha: bool = False
+
+    # Tokenizer
+    tokenizer_max_length: int = 48
+
+    # Projector
+    proj_width: int = 1024
+
+    # Decoding
+    num_steps: int = 10
+
+    # Attention utils
+    use_cache: bool = True
+    attention_implementation: str = "eager"  # or fa2, flex
+
+    # Finetuning settings
+    freeze_vision_encoder: bool = True
+    train_expert_only: bool = False
+    train_state_proj: bool = True
+
+    # Training presets
+    optimizer_lr: float = 2.5e-5
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-10
+
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    # TODO: Add EMA
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        # TODO(Steven): Validate device and amp? in all policy configs?
+        """Input validation (not exhaustive)."""
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
+                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
+            )
+        if self.n_obs_steps != 1:
+            raise ValueError(
+                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
+            )
+
+        if self.use_delta_joint_actions_aloha:
+            raise NotImplementedError(
+                "`use_delta_joint_actions_aloha` is used by pi0 for aloha real models. It is not ported yet in LeRobot."
+            )
+
+    def validate_features(self) -> None:
+        # TODO: implement value error
+        # if not self.image_features and not self.env_state_feature:
+        #     raise ValueError("You must provide at least one image or the environment state among the inputs.")
+
+        for i in range(self.empty_cameras):
+            key = f"observation.images.empty_camera_{i}"
+            empty_camera = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(3, 480, 640),
+            )
+            self.input_features[key] = empty_camera
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+        )
+
+    def get_scheduler_preset(self):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

lerobot/common/policies/pi0/conversion_scripts/benchmark.py

@@ -0,0 +1,82 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.policies.factory import make_policy
+from lerobot.configs.policies import PreTrainedConfig
+
+torch.backends.cudnn.benchmark = True
+
+
+def main():
+    device = "cuda"
+    dataset_repo_id = "danaaubakirova/koch_test"
+    # model_name = "pi0_base"
+    # ckpt_torch_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}_pytorch"
+    ckpt_torch_dir = "lerobot/pi0"
+
+    dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
+
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=0,
+        batch_size=1,
+    )
+
+    batch = next(iter(dataloader))
+
+    # To device
+    for k in batch:
+        if isinstance(batch[k], torch.Tensor):
+            batch[k] = batch[k].to(device=device, dtype=torch.float32)
+
+    cfg = PreTrainedConfig.from_pretrained(ckpt_torch_dir)
+    cfg.pretrained_path = ckpt_torch_dir
+    policy = make_policy(cfg, ds_meta=dataset.meta)
+
+    # policy = torch.compile(policy, mode="reduce-overhead")
+
+    warmup_iters = 10
+    benchmark_iters = 30
+
+    # Warmup
+    for _ in range(warmup_iters):
+        torch.cuda.synchronize()
+        policy.select_action(batch)
+        policy.reset()
+        torch.cuda.synchronize()
+
+    # Benchmark
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    start_event.record()
+    for _ in range(benchmark_iters):
+        policy.select_action(batch)
+        policy.reset()
+    end_event.record()
+
+    # Synchronize and measure time
+    torch.cuda.synchronize()
+    elapsed_time_ms = start_event.elapsed_time(end_event)
+
+    avg_time_per_iter = elapsed_time_ms / benchmark_iters
+    print(f"Average execution time per iteration: {avg_time_per_iter:.3f} ms")
+
+
+if __name__ == "__main__":
+    with torch.inference_mode():
+        main()

lerobot/common/policies/pi0/conversion_scripts/compare_with_jax.py

@@ -0,0 +1,131 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import pickle
+from pathlib import Path
+
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.common.policies.factory import make_policy
+from lerobot.configs.policies import PreTrainedConfig
+
+
+def display(tensor: torch.Tensor):
+    if tensor.dtype == torch.bool:
+        tensor = tensor.float()
+    print(f"Shape: {tensor.shape}")
+    print(f"Mean: {tensor.mean().item()}")
+    print(f"Std: {tensor.std().item()}")
+    print(f"Min: {tensor.min().item()}")
+    print(f"Max: {tensor.max().item()}")
+
+
+def main():
+    num_motors = 14
+    device = "cuda"
+    # model_name = "pi0_aloha_towel"
+    model_name = "pi0_aloha_sim"
+
+    if model_name == "pi0_aloha_towel":
+        dataset_repo_id = "lerobot/aloha_static_towel"
+    else:
+        dataset_repo_id = "lerobot/aloha_sim_transfer_cube_human"
+
+    ckpt_torch_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}_pytorch"
+    ckpt_jax_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}"
+    save_dir = Path(f"../openpi/data/{model_name}/save")
+
+    with open(save_dir / "example.pkl", "rb") as f:
+        example = pickle.load(f)
+    with open(save_dir / "outputs.pkl", "rb") as f:
+        outputs = pickle.load(f)
+    with open(save_dir / "noise.pkl", "rb") as f:
+        noise = pickle.load(f)
+
+    with open(ckpt_jax_dir / "assets/norm_stats.json") as f:
+        norm_stats = json.load(f)
+
+    # Override stats
+    dataset_meta = LeRobotDatasetMetadata(dataset_repo_id)
+    dataset_meta.stats["observation.state"]["mean"] = torch.tensor(
+        norm_stats["norm_stats"]["state"]["mean"][:num_motors], dtype=torch.float32
+    )
+    dataset_meta.stats["observation.state"]["std"] = torch.tensor(
+        norm_stats["norm_stats"]["state"]["std"][:num_motors], dtype=torch.float32
+    )
+
+    # Create LeRobot batch from Jax
+    batch = {}
+    for cam_key, uint_chw_array in example["images"].items():
+        batch[f"observation.images.{cam_key}"] = torch.from_numpy(uint_chw_array) / 255.0
+    batch["observation.state"] = torch.from_numpy(example["state"])
+    batch["action"] = torch.from_numpy(outputs["actions"])
+    batch["task"] = example["prompt"]
+
+    if model_name == "pi0_aloha_towel":
+        del batch["observation.images.cam_low"]
+    elif model_name == "pi0_aloha_sim":
+        batch["observation.images.top"] = batch["observation.images.cam_high"]
+        del batch["observation.images.cam_high"]
+
+    # Batchify
+    for key in batch:
+        if isinstance(batch[key], torch.Tensor):
+            batch[key] = batch[key].unsqueeze(0)
+        elif isinstance(batch[key], str):
+            batch[key] = [batch[key]]
+        else:
+            raise ValueError(f"{key}, {batch[key]}")
+
+    # To device
+    for k in batch:
+        if isinstance(batch[k], torch.Tensor):
+            batch[k] = batch[k].to(device=device, dtype=torch.float32)
+
+    noise = torch.from_numpy(noise).to(device=device, dtype=torch.float32)
+
+    from lerobot.common import policies  # noqa
+
+    cfg = PreTrainedConfig.from_pretrained(ckpt_torch_dir)
+    cfg.pretrained_path = ckpt_torch_dir
+    policy = make_policy(cfg, dataset_meta)
+
+    # loss_dict = policy.forward(batch, noise=noise, time=time_beta)
+    # loss_dict["loss"].backward()
+    # print("losses")
+    # display(loss_dict["losses_after_forward"])
+    # print("pi_losses")
+    # display(pi_losses)
+
+    actions = []
+    for _ in range(50):
+        action = policy.select_action(batch, noise=noise)
+        actions.append(action)
+
+    actions = torch.stack(actions, dim=1)
+    pi_actions = batch["action"]
+    print("actions")
+    display(actions)
+    print()
+    print("pi_actions")
+    display(pi_actions)
+    print("atol=3e-2", torch.allclose(actions, pi_actions, atol=3e-2))
+    print("atol=2e-2", torch.allclose(actions, pi_actions, atol=2e-2))
+    print("atol=1e-2", torch.allclose(actions, pi_actions, atol=1e-2))
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/policies/pi0/conversion_scripts/conversion_utils.py

@@ -0,0 +1,84 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from transformers import GemmaConfig, PaliGemmaConfig
+
+
+def get_paligemma_config(precision: str):
+    config = {
+        "image_token_index": None,
+        "pad_token_id": 0,
+        "bos_token_id": 2,
+        "eos_token_id": 1,
+    }
+
+    # image_sizes = {"2b-test": 224, "3b-224px": 224, "3b-448px": 448, "3b-896px": 896}
+
+    image_size = 224  # image_sizes[variant]
+    patch_size = 14
+    num_image_tokens = (image_size**2) // (patch_size**2)
+
+    config["image_token_index"] = 257152
+    text_config = {
+        "vocab_size": 257152,
+        "num_hidden_layers": 18,
+        "num_key_value_heads": 1,
+        "head_dim": 256,
+        "torch_dtype": precision,
+        "hidden_size": 2048,
+        "hidden_activation": "gelu_pytorch_tanh",
+        "num_attention_heads": 8,
+        "intermediate_size": 16384,
+        "is_encoder_decoder": False,
+    }
+    vision_config = {
+        "torch_dtype": precision,
+        "image_size": image_size,
+        "patch_size": patch_size,
+        "num_image_tokens": num_image_tokens,
+        "hidden_size": 1152,
+        "intermediate_size": 4304,
+        "num_hidden_layers": 27,
+        "num_attention_heads": 16,
+        "projector_hidden_act": "gelu_fast",
+        "vision_use_head": False,
+    }
+    final_config = PaliGemmaConfig(text_config=text_config, vision_config=vision_config, **config)
+    return final_config
+
+
+def get_gemma_config(precision: str):
+    config = {
+        "image_token_index": None,
+        "pad_token_id": 0,
+        "bos_token_id": 2,
+        "eos_token_id": 1,
+    }
+
+    config["image_token_index"] = 257152
+    text_config = {
+        "vocab_size": 257152,
+        "num_hidden_layers": 18,
+        "num_key_value_heads": 1,
+        "head_dim": 256,
+        "torch_dtype": precision,
+        "hidden_size": 1024,
+        "hidden_activation": "gelu_pytorch_tanh",
+        "num_attention_heads": 8,
+        "intermediate_size": 4096,
+        "is_encoder_decoder": False,
+    }
+    final_config = GemmaConfig()
+    final_config.update(text_config)
+    return final_config

lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py

@@ -0,0 +1,437 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Convert pi0 parameters from Jax to Pytorch
+
+Follow [README of openpi](https://github.com/Physical-Intelligence/openpi) to create a new environment
+and install the required libraries.
+
+```bash
+cd ~/code/openpi
+source .venv/bin/activate
+```
+
+Example downloading parameters:
+```bash
+python
+>>> import openpi.shared.download as download
+>>> path='s3://openpi-assets/checkpoints/pi0_base/params'
+>>> download.maybe_download(path)
+```
+
+Converting pi0_base:
+```python
+python lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py \
+    --checkpoint_dir /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_base/params \
+    --output_path /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_base_pytorch
+```
+
+```python
+python lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py \
+    --checkpoint_dir /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim/params \
+    --output_path /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim_pytorch
+```
+"""
+
+import argparse
+import pathlib
+
+import jax
+import numpy as np
+import orbax.checkpoint as ocp
+import torch
+from jax.sharding import SingleDeviceSharding
+
+from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.common.policies.pi0.conversion_scripts.conversion_utils import (
+    get_gemma_config,
+    get_paligemma_config,
+)
+from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
+
+PRECISIONS = {"bfloat16": torch.bfloat16, "float32": torch.float32, "float16": torch.float16}
+
+
+def slice_paligemma_state_dict(state_dict, config):
+    suffix = "/value" if "img/embedding/kernel/value" in state_dict else ""
+
+    # fmt: off
+    # patch embeddings
+    state_dict["paligemma.vision_tower.vision_model.embeddings.patch_embedding.weight"] = state_dict.pop(f"img/embedding/kernel{suffix}").transpose(
+        3, 2, 0, 1
+    )
+    state_dict["paligemma.vision_tower.vision_model.embeddings.patch_embedding.bias"] = state_dict.pop(f"img/embedding/bias{suffix}")
+    # positional embeddings
+    state_dict["paligemma.vision_tower.vision_model.embeddings.position_embedding.weight"] = state_dict.pop(f"img/pos_embedding{suffix}").reshape(
+        -1, config.vision_config.hidden_size
+    )
+
+    # extract vision layers to be sliced at index 0. There are 27 layers in the base model.
+    encoderblock_layernorm0_scale = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_0/scale{suffix}")
+    encoderblock_layernorm0_bias = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_0/bias{suffix}")
+    encoderblock_layernorm1_scale = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_1/scale{suffix}")
+    encoderblock_layernorm1_bias = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_1/bias{suffix}")
+
+    encoderblock_mlp_dense0_kernel= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_0/kernel{suffix}")
+    encoderblock_mlp_dense0_bias= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_0/bias{suffix}")
+    encoderblock_mlp_dense1_kernel= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_1/kernel{suffix}")
+    encoderblock_mlp_dense1_bias= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_1/bias{suffix}")
+
+    encoderblock_attention_0_key_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/key/kernel{suffix}")
+    encoderblock_attention_0_key_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/key/bias{suffix}")
+    encoderblock_attention_0_value_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/value/kernel{suffix}")
+    encoderblock_attention_0_value_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/value/bias{suffix}")
+    encoderblock_attention_0_query_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/query/kernel{suffix}")
+    encoderblock_attention_0_query_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/query/bias{suffix}")
+    encoderblock_attention_0_out_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/out/kernel{suffix}")
+    encoderblock_attention_0_out_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/out/bias{suffix}")
+
+    for i in range(config.vision_config.num_hidden_layers):
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm1.weight"] = encoderblock_layernorm0_scale[i].transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm1.bias"] = encoderblock_layernorm0_bias[i]
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm2.weight"] = encoderblock_layernorm1_scale[i].transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm2.bias"] = encoderblock_layernorm1_bias[i]
+
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc1.weight"] = encoderblock_mlp_dense0_kernel[i].transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc1.bias"] = encoderblock_mlp_dense0_bias[i]
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc2.weight"] = encoderblock_mlp_dense1_kernel[i].transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc2.bias"] = encoderblock_mlp_dense1_bias[i]
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.k_proj.weight"] = encoderblock_attention_0_key_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.k_proj.bias"] = encoderblock_attention_0_key_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.v_proj.weight"] = encoderblock_attention_0_value_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.v_proj.bias"] = encoderblock_attention_0_value_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.q_proj.weight"] = encoderblock_attention_0_query_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.q_proj.bias"] = encoderblock_attention_0_query_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.out_proj.weight"] = encoderblock_attention_0_out_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.out_proj.bias"] = encoderblock_attention_0_out_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
+
+    state_dict["paligemma.vision_tower.vision_model.post_layernorm.weight"] = state_dict.pop(f"img/Transformer/encoder_norm/scale{suffix}").transpose()
+    state_dict["paligemma.vision_tower.vision_model.post_layernorm.bias"] = state_dict.pop(f"img/Transformer/encoder_norm/bias{suffix}")
+
+    # multimodal projector
+
+    state_dict['paligemma.multi_modal_projector.linear.weight'] = state_dict.pop(f"img/head/kernel{suffix}").transpose()
+    state_dict['paligemma.multi_modal_projector.linear.bias'] = state_dict.pop(f"img/head/bias{suffix}")
+
+    # text decoder (gemma)
+    embedding_vector = state_dict.pop(f"llm/embedder/input_embedding{suffix}")
+    state_dict["paligemma.language_model.model.embed_tokens.weight"] = embedding_vector
+
+    # pop the einsum attention + mlp representations. There are 18 layers in gemma-2b.
+
+    llm_attention_attn_vec_einsum = state_dict.pop(f"llm/layers/attn/attn_vec_einsum/w{suffix}")
+    llm_attention_kv_einsum = state_dict.pop(f"llm/layers/attn/kv_einsum/w{suffix}")
+    llm_attention_q_einsum = state_dict.pop(f"llm/layers/attn/q_einsum/w{suffix}")
+
+    llm_mlp_gating_einsum = state_dict.pop(f"llm/layers/mlp/gating_einsum{suffix}")
+    llm_mlp_linear = state_dict.pop(f"llm/layers/mlp/linear{suffix}")
+    # TODO verify correctness of layer norm loading
+
+    llm_input_layernorm = state_dict.pop(f"llm/layers/pre_attention_norm/scale{suffix}")
+    llm_post_attention_layernorm = state_dict.pop(f"llm/layers/pre_ffw_norm/scale{suffix}")
+
+    for i in range(config.text_config.num_hidden_layers):
+        # llm_attention_q_einsum[i].shape = (8, 2048, 256)
+        q_proj_weight_reshaped = llm_attention_q_einsum[i].transpose(0, 2, 1).reshape(config.text_config.num_attention_heads * config.text_config.head_dim, config.text_config.hidden_size)
+
+        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.q_proj.weight"] = q_proj_weight_reshaped
+
+        # llm_attention_kv_einsum[i, 0, 0].shape = (2048, 256)
+        k_proj_weight_reshaped = llm_attention_kv_einsum[i, 0, 0].transpose()
+        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.k_proj.weight"] = k_proj_weight_reshaped
+        # llm_attention_kv_einsum[i, 1, 0].shape = (2048, 256)
+        v_proj_weight_reshaped = llm_attention_kv_einsum[i, 1, 0].transpose()
+        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.v_proj.weight"] = v_proj_weight_reshaped
+
+        # output projection.
+
+        # llm_attention_attn_vec_einsum[i].shape = (8, 256, 2048)
+        o_proj_weight_reshaped = llm_attention_attn_vec_einsum[i].transpose(2, 0, 1).reshape(config.text_config.num_attention_heads * config.text_config.head_dim, config.text_config.hidden_size)
+
+        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.o_proj.weight"] = o_proj_weight_reshaped
+        # mlp layers
+        gate_proj_weight = llm_mlp_gating_einsum[i, 0]
+        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.gate_proj.weight"] = gate_proj_weight.transpose()
+        up_proj_weight = llm_mlp_gating_einsum[i, 1]
+        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.up_proj.weight"] = up_proj_weight.transpose()
+        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.down_proj.weight"] = llm_mlp_linear[i].transpose()
+        state_dict[f"paligemma.language_model.model.layers.{i}.input_layernorm.weight"] = llm_input_layernorm[i]
+        state_dict[f"paligemma.language_model.model.layers.{i}.post_attention_layernorm.weight"] = llm_post_attention_layernorm[i]
+
+    state_dict["paligemma.language_model.model.norm.weight"] = state_dict.pop(f"llm/final_norm/scale{suffix}")
+    state_dict["paligemma.language_model.lm_head.weight"] = embedding_vector # weights are tied.
+
+    # fmt: on
+    expert_dict = {}
+    final_state_dict = {}
+    for key, value in state_dict.items():
+        if key not in [
+            f"llm/final_norm_1/scale{suffix}",
+            f"llm/layers/attn/attn_vec_einsum_1/w{suffix}",
+            f"llm/layers/attn/kv_einsum_1/w{suffix}",
+            f"llm/layers/attn/q_einsum_1/w{suffix}",
+            f"llm/layers/mlp_1/gating_einsum{suffix}",
+            f"llm/layers/mlp_1/linear{suffix}",
+            f"llm/layers/pre_attention_norm_1/scale{suffix}",
+            f"llm/layers/pre_ffw_norm_1/scale{suffix}",
+        ]:
+            final_state_dict[key] = torch.from_numpy(value)
+        else:
+            expert_dict[key] = value
+
+    return final_state_dict, expert_dict
+
+
+def slice_gemma_state_dict(state_dict, config, num_expert=1):
+    # fmt: off
+    # text decoder (gemma)
+    # no embedding vector, the expert just has the decoder layers
+
+    embedding_vector = torch.zeros([config.vocab_size, config.hidden_size])
+    state_dict["gemma_expert.model.embed_tokens.weight"] = embedding_vector
+
+    # pop the einsum attention + mlp representations. There are 18 layers in gemma-2b.
+
+    suffix = "/value" if f"llm/layers/attn/attn_vec_einsum_{num_expert}/w/value" in state_dict else ""
+
+    llm_attention_attn_vec_einsum = state_dict.pop(f"llm/layers/attn/attn_vec_einsum_{num_expert}/w{suffix}")
+    llm_attention_kv_einsum = state_dict.pop(f"llm/layers/attn/kv_einsum_{num_expert}/w{suffix}")
+    llm_attention_q_einsum = state_dict.pop(f"llm/layers/attn/q_einsum_{num_expert}/w{suffix}")
+
+    llm_mlp_gating_einsum = state_dict.pop(f"llm/layers/mlp_{num_expert}/gating_einsum{suffix}")
+    llm_mlp_linear = state_dict.pop(f"llm/layers/mlp_{num_expert}/linear{suffix}")
+    # TODO verify correctness of layer norm loading
+
+    llm_input_layernorm = state_dict.pop(f"llm/layers/pre_attention_norm_{num_expert}/scale{suffix}")
+    llm_post_attention_layernorm = state_dict.pop(f"llm/layers/pre_ffw_norm_{num_expert}/scale{suffix}")
+
+    for i in range(config.num_hidden_layers):
+        q_proj_weight_reshaped = llm_attention_q_einsum[i].transpose(0, 2, 1).reshape(config.num_attention_heads * config.head_dim, config.hidden_size)
+
+        state_dict[f"gemma_expert.model.layers.{i}.self_attn.q_proj.weight"] = q_proj_weight_reshaped
+
+        k_proj_weight_reshaped = llm_attention_kv_einsum[i, 0, 0].transpose()
+        state_dict[f"gemma_expert.model.layers.{i}.self_attn.k_proj.weight"] = k_proj_weight_reshaped
+        v_proj_weight_reshaped = llm_attention_kv_einsum[i, 1, 0].transpose()
+        state_dict[f"gemma_expert.model.layers.{i}.self_attn.v_proj.weight"] = v_proj_weight_reshaped
+
+        # output projection.
+
+        # llm_attention_attn_vec_einsum[i].shape = (8, 256, 1024)
+        o_proj_weight_reshaped = llm_attention_attn_vec_einsum[i].reshape(config.num_attention_heads * config.head_dim, config.hidden_size).transpose(1,0)# .transpose(2, 0, 1).reshape(config.num_attention_heads * config.head_dim, config.hidden_size).transpose(1, 0)
+
+        state_dict[f"gemma_expert.model.layers.{i}.self_attn.o_proj.weight"] = o_proj_weight_reshaped
+        # mlp layers
+        gate_proj_weight = llm_mlp_gating_einsum[i, 0]
+        state_dict[f"gemma_expert.model.layers.{i}.mlp.gate_proj.weight"] = gate_proj_weight.transpose()
+        up_proj_weight = llm_mlp_gating_einsum[i, 1]
+        state_dict[f"gemma_expert.model.layers.{i}.mlp.up_proj.weight"] = up_proj_weight.transpose()
+        state_dict[f"gemma_expert.model.layers.{i}.mlp.down_proj.weight"] = llm_mlp_linear[i].transpose()
+        state_dict[f"gemma_expert.model.layers.{i}.input_layernorm.weight"] = llm_input_layernorm[i]
+        state_dict[f"gemma_expert.model.layers.{i}.post_attention_layernorm.weight"] = llm_post_attention_layernorm[i]
+
+    state_dict["gemma_expert.model.norm.weight"] = state_dict.pop(f"llm/final_norm_{num_expert}/scale{suffix}")
+    state_dict["gemma_expert.lm_head.weight"] = embedding_vector # weights are tied. (and zeros here)
+
+    # fmt: on
+    final_state_dict = {}
+    for key, value in state_dict.items():
+        if not isinstance(value, torch.Tensor):
+            final_state_dict[key] = torch.from_numpy(value)
+        else:
+            final_state_dict[key] = value
+    return final_state_dict
+
+
+def flatten_for_memory(tree, parent_key=""):
+    out = {}
+    for k, v in tree.items():
+        new_key = f"{parent_key}/{k}" if parent_key else k
+        if isinstance(v, dict):
+            out.update(flatten_for_memory(v, new_key))
+        else:
+            out[new_key] = np.array(v)  # Ensure conversion to np.array for consistency
+    return out
+
+
+def flatten_for_npz(tree, parent_key=""):
+    out = {}
+    for k, v in tree.items():
+        new_key = f"{parent_key}/{k}" if parent_key else k
+        if isinstance(v, dict):
+            out.update(flatten_for_npz(v, new_key))
+        else:
+            # bf16/f32 here?
+            out[new_key] = np.array(v)
+    return out
+
+
+def slice_initial_orbax_checkpoint(checkpoint_dir: str):
+    params_path = pathlib.Path(checkpoint_dir).resolve()
+    checkpointer = ocp.PyTreeCheckpointer()
+
+    metadata = checkpointer.metadata(params_path)
+    print("Metadata keys:", list(metadata.keys()))
+
+    params_name = "params"
+
+    item = {params_name: metadata[params_name]}
+    device = jax.local_devices()[0]  # Use the first local device
+    sharding = SingleDeviceSharding(device)
+    restored = checkpointer.restore(
+        params_path,
+        ocp.args.PyTreeRestore(
+            item=item,
+            restore_args=jax.tree_util.tree_map(
+                lambda _: ocp.ArrayRestoreArgs(
+                    restore_type=jax.Array,  # or np.ndarray, but bf16 is annoying about it
+                    sharding=sharding,
+                ),
+                item,
+            ),
+            transforms={},
+        ),
+    )
+    params = restored[params_name]
+
+    # get params for PaliGemma
+    pali_params = params["PaliGemma"]
+    del params["PaliGemma"]
+    pali_params_flat = flatten_for_npz(pali_params)
+    return {"paligemma_params": pali_params_flat, "projection_params": params}
+
+
+def update_keys_with_prefix(d: dict, prefix: str) -> dict:
+    """Update dictionary keys by adding a prefix."""
+    return {f"{prefix}{key}": value for key, value in d.items()}
+
+
+def convert_pi0_checkpoint(checkpoint_dir: str, precision: str, tokenizer_id: str, output_path: str):
+    # Break down orbax ckpts - they are in OCDBT
+    initial_params = slice_initial_orbax_checkpoint(checkpoint_dir=checkpoint_dir)
+    # process projection params
+    keys = [
+        "state_proj",
+        "action_in_proj",
+        "action_out_proj",
+        "action_time_mlp_in",
+        "action_time_mlp_out",
+    ]
+
+    projection_params = {}
+    for key in keys:
+        kernel_params = initial_params["projection_params"][key]["kernel"]
+        bias_params = initial_params["projection_params"][key]["bias"]
+        if isinstance(kernel_params, dict):
+            weight = kernel_params["value"]
+            bias = bias_params["value"]
+        else:
+            weight = kernel_params
+            bias = bias_params
+        projection_params[f"{key}.weight"] = torch.from_numpy(np.array(weight)).T
+        projection_params[f"{key}.bias"] = torch.from_numpy(np.array(bias))
+
+    # Process PaliGemma weights
+    paligemma_config = get_paligemma_config(precision)
+    paligemma_params, gemma_raw_dictionary = slice_paligemma_state_dict(
+        initial_params["paligemma_params"], paligemma_config
+    )
+
+    # Process Gemma  weights (at this stage they are unused)
+    gemma_config = get_gemma_config(precision)
+    gemma_params = slice_gemma_state_dict(gemma_raw_dictionary, config=gemma_config)
+
+    # Instantiate model from configs
+
+    if "pi0_aloha_sim" in checkpoint_dir:
+        pi0_config = PI0Config(
+            empty_cameras=2,
+            adapt_to_pi_aloha=True,
+            use_delta_joint_actions_aloha=False,
+        )
+    elif "pi0_aloha_towel" in checkpoint_dir:
+        pi0_config = PI0Config(
+            adapt_to_pi_aloha=True,
+            use_delta_joint_actions_aloha=True,
+        )
+    elif "pi0_base" in checkpoint_dir:
+        pi0_config = PI0Config(
+            empty_cameras=0,
+            adapt_to_pi_aloha=False,
+            use_delta_joint_actions_aloha=False,
+        )
+    else:
+        raise ValueError()
+
+    # gemma_config=gemma_config, paligemma_config=paligemma_config)
+    pi0_model = PI0Policy(pi0_config)
+
+    paligemma_params = update_keys_with_prefix(paligemma_params, "model.paligemma_with_expert.")
+    gemma_params = update_keys_with_prefix(gemma_params, "model.paligemma_with_expert.")
+    projection_params = update_keys_with_prefix(projection_params, "model.")
+
+    # load state dict
+    torch_dtype = PRECISIONS[precision]
+    pi0_model.load_state_dict({**paligemma_params, **gemma_params, **projection_params})
+    pi0_model = pi0_model.to(torch_dtype)
+    # pi0_tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
+
+    pi0_model.save_pretrained(output_path, safe_serialization=True)
+    # pi0_tokenizer.save_pretrained(output_path, dtype=torch_dtype)
+
+    # assert that model loads properly
+    del pi0_model
+    PI0Policy.from_pretrained(output_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--checkpoint_dir",
+        default="/raid/pablo/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim/params",
+        type=str,
+        help="Path to the ocdbt checkpoint",
+    )
+
+    parser.add_argument(
+        "--precision",
+        choices=["float32", "bfloat16", "float16"],
+        default="float32",
+        type=str,
+        help="Precision identifier for model conversion - should match the base checkpoint precision.",
+    )
+    # tokenizer is identical to paligemma, it appears
+
+    parser.add_argument(
+        "--tokenizer_hub_id",
+        default="google/paligemma-3b-pt-224",
+        type=str,
+        help="Hub path to the tokenizer to save",
+    )
+
+    parser.add_argument(
+        "--output_path",
+        required=True,
+        type=str,
+        help="Path to save converted weights to",
+    )
+
+    args = parser.parse_args()
+    convert_pi0_checkpoint(
+        checkpoint_dir=args.checkpoint_dir,
+        precision=args.precision,
+        tokenizer_id=args.tokenizer_hub_id,
+        output_path=args.output_path,
+    )

lerobot/common/policies/pi0/flex_attention.py

@@ -0,0 +1,141 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from packaging.version import Version
+
+if Version(torch.__version__) > Version("2.5.0"):
+    # Ffex attention is only available from torch 2.5 onwards
+    from torch.nn.attention.flex_attention import (
+        _mask_mod_signature,
+        _round_up_to_multiple,
+        create_block_mask,
+        create_mask,
+        flex_attention,
+    )
+
+
+# @torch.compile(dynamic=False)
+def flex_attention_forward(
+    attention_mask: torch.Tensor,
+    batch_size: int,
+    head_dim: int,
+    query_states: torch.Tensor,
+    key_states: torch.Tensor,
+    value_states: torch.Tensor,
+    scaling=None,
+):
+    """
+    This is defined out of classes to make compile happy.
+    """
+
+    original_dtype = query_states.dtype
+    num_att_heads = 8
+    num_key_value_heads = 1
+    num_key_value_groups = num_att_heads // num_key_value_heads
+
+    key_states = key_states[:, :, :, None, :]
+    key_states = key_states.expand(
+        batch_size, key_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
+    )
+    key_states = key_states.reshape(
+        batch_size, key_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
+    )
+
+    value_states = value_states[:, :, :, None, :]
+    value_states = value_states.expand(
+        batch_size, value_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
+    )
+    value_states = value_states.reshape(
+        batch_size, value_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
+    )
+
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+
+    query_states = query_states.to(torch.float32)
+    key_states = key_states.to(torch.float32)
+    value_states = value_states.to(torch.float32)
+
+    causal_mask = attention_mask
+    if causal_mask is not None:
+        causal_mask = causal_mask[:, None, :, : key_states.shape[2]]
+
+        if causal_mask.shape[1] == 1 and query_states.shape[1] > 1:
+            causal_mask = causal_mask.expand(-1, query_states.shape[1], -1, -1)
+
+    def precomputed_mask_factory(precomputed_mask: torch.Tensor) -> _mask_mod_signature:
+        def mask_mod(b, h, q_idx, kv_idx):
+            # Danger zone: if b,h,q_idx,kv_idx exceed the shape, device-side assert occurs.
+            return precomputed_mask[b][h][q_idx][kv_idx]
+
+        return mask_mod
+
+    b_mask, h_mask, q_len, kv_len = causal_mask.shape  # The shape of your mask
+
+    block_size = 128
+    q_len_rounded = _round_up_to_multiple(q_len, block_size)
+    kv_len_rounded = _round_up_to_multiple(kv_len, block_size)
+
+    # *CRITICAL* we do need to expand here, else we get a CUDA index error
+
+    pad_q = q_len_rounded - q_len
+    pad_k = kv_len_rounded - kv_len
+
+    padded_causal_mask = F.pad(causal_mask, (0, pad_k, 0, pad_q), value=0.0)
+    mask_mod_fn_orig = precomputed_mask_factory(padded_causal_mask)
+
+    mask_4d = create_mask(
+        mod_fn=mask_mod_fn_orig,
+        B=b_mask,
+        H=h_mask,
+        Q_LEN=q_len_rounded,
+        KV_LEN=kv_len_rounded,
+        device=causal_mask.device,
+        _compile=False,
+    )
+
+    mask_mod_fn_padded = precomputed_mask_factory(mask_4d)
+    block_mask = create_block_mask(
+        mask_mod=mask_mod_fn_padded,
+        B=b_mask,
+        H=h_mask,
+        Q_LEN=q_len_rounded,
+        KV_LEN=kv_len_rounded,
+        BLOCK_SIZE=block_size,
+        device=causal_mask.device,
+        _compile=False,
+    )
+
+    #  mask is applied inside the kernel, ideally more efficiently than score_mod.
+    attn_output, attention_weights = flex_attention(
+        query_states,
+        key_states,
+        value_states,
+        block_mask=block_mask,
+        enable_gqa=True,  # because we shaped query/key states for GQA
+        scale=head_dim**-0.5 if scaling is None else scaling,
+        return_lse=True,
+    )
+
+    attn_output = attn_output.to(dtype=original_dtype)
+    attn_output = attn_output.transpose(1, 2).contiguous()  # [B, Q_LEN, H, head_dim]
+    attn_output = attn_output.reshape(
+        batch_size,
+        -1,
+        attn_output.shape[2] * attn_output.shape[3],  # merges [H, head_dim]
+    )
+    return attn_output

lerobot/common/policies/pi0/modeling_pi0.py

@@ -0,0 +1,732 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+π0: A Vision-Language-Action Flow Model for General Robot Control
+
+[Paper](https://www.physicalintelligence.company/download/pi0.pdf)
+[Jax code](https://github.com/Physical-Intelligence/openpi)
+
+Designed by Physical Intelligence. Ported from Jax by Hugging Face.
+
+Install pi0 extra dependencies:
+```bash
+pip install -e ".[pi0]"
+```
+
+Example of finetuning the pi0 pretrained model (`pi0_base` in `openpi`):
+```bash
+python lerobot/scripts/train.py \
+--policy.path=lerobot/pi0 \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of finetuning the pi0 neural network with PaliGemma and expert Gemma
+pretrained with VLM default parameters before pi0 finetuning:
+```bash
+python lerobot/scripts/train.py \
+--policy.type=pi0 \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of using the pi0 pretrained model outside LeRobot training framework:
+```python
+policy = Pi0Policy.from_pretrained("lerobot/pi0")
+```
+
+"""
+
+import math
+from collections import deque
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor, nn
+from transformers import AutoTokenizer
+
+from lerobot.common.constants import ACTION, OBS_ROBOT
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.common.policies.pi0.paligemma_with_expert import (
+    PaliGemmaWithExpertConfig,
+    PaliGemmaWithExpertModel,
+)
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+from lerobot.common.utils.utils import get_safe_dtype
+
+
+def create_sinusoidal_pos_embedding(
+    time: torch.tensor, dimension: int, min_period: float, max_period: float, device="cpu"
+) -> Tensor:
+    """Computes sine-cosine positional embedding vectors for scalar positions."""
+    if dimension % 2 != 0:
+        raise ValueError(f"dimension ({dimension}) must be divisible by 2")
+
+    if time.ndim != 1:
+        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")
+
+    dtype = get_safe_dtype(torch.float64, device.type)
+    fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
+    period = min_period * (max_period / min_period) ** fraction
+
+    # Compute the outer product
+    scaling_factor = 1.0 / period * 2 * math.pi
+    sin_input = scaling_factor[None, :] * time[:, None]
+    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+    return pos_emb
+
+
+def sample_beta(alpha, beta, bsize, device):
+    gamma1 = torch.empty((bsize,), device=device).uniform_(0, 1).pow(1 / alpha)
+    gamma2 = torch.empty((bsize,), device=device).uniform_(0, 1).pow(1 / beta)
+    return gamma1 / (gamma1 + gamma2)
+
+
+def make_att_2d_masks(pad_masks, att_masks):
+    """Copied from big_vision.
+
+    Tokens can attend to valid inputs tokens which have a cumulative mask_ar
+    smaller or equal to theirs. This way `mask_ar` int[B, N] can be used to
+    setup several types of attention, for example:
+
+      [[1 1 1 1 1 1]]: pure causal attention.
+
+      [[0 0 0 1 1 1]]: prefix-lm attention. The first 3 tokens can attend between
+          themselves and the last 3 tokens have a causal attention. The first
+          entry could also be a 1 without changing behaviour.
+
+      [[1 0 1 0 1 0 0 1 0 0]]: causal attention between 4 blocks. Tokens of a
+          block can attend all previous blocks and all tokens on the same block.
+
+    Args:
+      input_mask: bool[B, N] true if its part of the input, false if padding.
+      mask_ar: int32[B, N] mask that's 1 where previous tokens cannot depend on
+        it and 0 where it shares the same attention mask as the previous token.
+    """
+    if att_masks.ndim != 2:
+        raise ValueError(att_masks.ndim)
+    if pad_masks.ndim != 2:
+        raise ValueError(pad_masks.ndim)
+
+    cumsum = torch.cumsum(att_masks, dim=1)
+    att_2d_masks = cumsum[:, None, :] <= cumsum[:, :, None]
+    pad_2d_masks = pad_masks[:, None, :] * pad_masks[:, :, None]
+    att_2d_masks = att_2d_masks & pad_2d_masks
+    return att_2d_masks
+
+
+def resize_with_pad(img, width, height, pad_value=-1):
+    # assume no-op when width height fits already
+    if img.ndim != 4:
+        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
+
+    cur_height, cur_width = img.shape[2:]
+
+    ratio = max(cur_width / width, cur_height / height)
+    resized_height = int(cur_height / ratio)
+    resized_width = int(cur_width / ratio)
+    resized_img = F.interpolate(
+        img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+    )
+
+    pad_height = max(0, int(height - resized_height))
+    pad_width = max(0, int(width - resized_width))
+
+    # pad on left and top of image
+    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
+    return padded_img
+
+
+def pad_vector(vector, new_dim):
+    """Can be (batch_size x sequence_length x features_dimension)
+    or (batch_size x features_dimension)
+    """
+    if vector.shape[-1] == new_dim:
+        return vector
+    shape = list(vector.shape)
+    current_dim = shape[-1]
+    shape[-1] = new_dim
+    new_vector = torch.zeros(*shape, dtype=vector.dtype, device=vector.device)
+    new_vector[..., :current_dim] = vector
+    return new_vector
+
+
+def normalize(x, min_val, max_val):
+    return (x - min_val) / (max_val - min_val)
+
+
+def unnormalize(x, min_val, max_val):
+    return x * (max_val - min_val) + min_val
+
+
+def safe_arcsin(value):
+    # This ensures that the input stays within
+    # [−1,1] to avoid invalid values for arcsin
+    return torch.arcsin(torch.clamp(value, -1.0, 1.0))
+
+
+def aloha_gripper_to_angular(value):
+    # Aloha transforms the gripper positions into a linear space. The following code
+    # reverses this transformation to be consistent with pi0 which is pretrained in
+    # angular space.
+    #
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_POSITION_OPEN, PUPPET_GRIPPER_POSITION_CLOSED
+    value = unnormalize(value, min_val=0.01844, max_val=0.05800)
+
+    # This is the inverse of the angular to linear transformation inside the Interbotix code.
+    def linear_to_radian(linear_position, arm_length, horn_radius):
+        value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
+        return safe_arcsin(value)
+
+    # The constants are taken from the Interbotix code.
+    value = linear_to_radian(value, arm_length=0.036, horn_radius=0.022)
+
+    # Normalize to [0, 1].
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+def aloha_gripper_from_angular(value):
+    # Convert from the gripper position used by pi0 to the gripper position that is used by Aloha.
+    # Note that the units are still angular but the range is different.
+
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    value = unnormalize(value, min_val=0.4, max_val=1.5)
+
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_JOINT_OPEN, PUPPET_GRIPPER_JOINT_CLOSE
+    return normalize(value, min_val=-0.6213, max_val=1.4910)
+
+
+def aloha_gripper_from_angular_inv(value):
+    # Directly inverts the gripper_from_angular function.
+    value = unnormalize(value, min_val=-0.6213, max_val=1.4910)
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+class PI0Policy(PreTrainedPolicy):
+    """Wrapper class around PI0FlowMatching model to train and run inference within LeRobot."""
+
+    config_class = PI0Config
+    name = "pi0"
+
+    def __init__(
+        self,
+        config: PI0Config,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                    the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_targets = Normalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+
+        self.language_tokenizer = AutoTokenizer.from_pretrained("google/paligemma-3b-pt-224")
+        self.model = PI0FlowMatching(config)
+
+        self.reset()
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    def get_optim_params(self) -> dict:
+        return self.parameters()
+
+    @torch.no_grad
+    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method wraps `select_actions` in order to return one action at a time for execution in the
+        environment. It works by managing the actions in a queue and only calling `select_actions` when the
+        queue is empty.
+        """
+        self.eval()
+
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+
+        batch = self.normalize_inputs(batch)
+
+        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
+        # querying the policy.
+        if len(self._action_queue) == 0:
+            images, img_masks = self.prepare_images(batch)
+            state = self.prepare_state(batch)
+            lang_tokens, lang_masks = self.prepare_language(batch)
+
+            actions = self.model.sample_actions(
+                images, img_masks, lang_tokens, lang_masks, state, noise=noise
+            )
+
+            # Unpad actions
+            original_action_dim = self.config.action_feature.shape[0]
+            actions = actions[:, :, :original_action_dim]
+
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.adapt_to_pi_aloha:
+                actions = self._pi_aloha_encode_actions(actions)
+
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> tuple[Tensor, dict[str, Tensor]]:
+        """Do a full training forward pass to compute the loss"""
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
+
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
+
+        images, img_masks = self.prepare_images(batch)
+        state = self.prepare_state(batch)
+        lang_tokens, lang_masks = self.prepare_language(batch)
+        actions = self.prepare_action(batch)
+        actions_is_pad = batch.get("action_is_pad")
+
+        loss_dict = {}
+        losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
+        loss_dict["losses_after_forward"] = losses.clone()
+
+        if actions_is_pad is not None:
+            in_episode_bound = ~actions_is_pad
+            losses = losses * in_episode_bound.unsqueeze(-1)
+            loss_dict["losses_after_in_ep_bound"] = losses.clone()
+
+        # Remove padding
+        losses = losses[:, :, : self.config.max_action_dim]
+        loss_dict["losses_after_rm_padding"] = losses.clone()
+
+        # For backward pass
+        loss = losses.mean()
+        # For logging
+        loss_dict["l2_loss"] = loss.item()
+
+        return loss, loss_dict
+
+    def prepare_images(self, batch):
+        """Apply Pi0 preprocessing to the images, like resizing to 224x224 and padding to keep aspect ratio, and
+        convert pixel range from [0.0, 1.0] to [-1.0, 1.0] as requested by SigLIP.
+        """
+        images = []
+        img_masks = []
+
+        present_img_keys = [key for key in self.config.image_features if key in batch]
+        missing_img_keys = [key for key in self.config.image_features if key not in batch]
+
+        if len(present_img_keys) == 0:
+            raise ValueError(
+                f"All image features are missing from the batch. At least one expected. (batch: {batch.keys()}) (image_features:{self.config.image_features})"
+            )
+
+        # Preprocess image features present in the batch
+        for key in present_img_keys:
+            img = batch[key]
+
+            if self.config.resize_imgs_with_padding is not None:
+                img = resize_with_pad(img, *self.config.resize_imgs_with_padding, pad_value=0)
+
+            # Normalize from range [0,1] to [-1,1] as expacted by siglip
+            img = img * 2.0 - 1.0
+
+            bsize = img.shape[0]
+            device = img.device
+            mask = torch.ones(bsize, dtype=torch.bool, device=device)
+            images.append(img)
+            img_masks.append(mask)
+
+        # Create image features not present in the batch
+        # as fully 0 padded images.
+        for num_empty_cameras in range(len(missing_img_keys)):
+            if num_empty_cameras >= self.config.empty_cameras:
+                break
+            img = torch.ones_like(img) * -1
+            mask = torch.zeros_like(mask)
+            images.append(img)
+            img_masks.append(mask)
+
+        return images, img_masks
+
+    def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
+        """Tokenize the text input"""
+        device = batch[OBS_ROBOT].device
+        tasks = batch["task"]
+
+        # PaliGemma prompt has to end with a new line
+        tasks = [task if task.endswith("\n") else f"{task}\n" for task in tasks]
+
+        tokenized_prompt = self.language_tokenizer.__call__(
+            tasks,
+            padding="max_length",
+            padding_side="right",
+            max_length=self.config.tokenizer_max_length,
+            return_tensors="pt",
+        )
+        lang_tokens = tokenized_prompt["input_ids"].to(device=device)
+        lang_masks = tokenized_prompt["attention_mask"].to(device=device, dtype=torch.bool)
+
+        return lang_tokens, lang_masks
+
+    def _pi_aloha_decode_state(self, state):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            state[:, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            state[:, motor_idx] = aloha_gripper_to_angular(state[:, motor_idx])
+        return state
+
+    def _pi_aloha_encode_actions(self, actions):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
+        return actions
+
+    def _pi_aloha_encode_actions_inv(self, actions):
+        # Flip the joints again.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
+        return actions
+
+    def prepare_state(self, batch):
+        """Pad state"""
+        state = pad_vector(batch[OBS_ROBOT], self.config.max_state_dim)
+        return state
+
+    def prepare_action(self, batch):
+        """Pad action"""
+        actions = pad_vector(batch[ACTION], self.config.max_action_dim)
+        return actions
+
+
+class PI0FlowMatching(nn.Module):
+    """
+    π0: A Vision-Language-Action Flow Model for General Robot Control
+
+    [Paper](https://www.physicalintelligence.company/download/pi0.pdf)
+    [Jax code](https://github.com/Physical-Intelligence/openpi)
+
+    Designed by Physical Intelligence. Ported from Jax by Hugging Face.
+    ┌──────────────────────────────┐
+    │               actions        │
+    │               ▲              │
+    │              ┌┴─────┐        │
+    │  kv cache    │Gemma │        │
+    │  ┌──────────►│Expert│        │
+    │  │           │      │        │
+    │ ┌┴────────┐  │x 10  │        │
+    │ │         │  └▲──▲──┘        │
+    │ │PaliGemma│   │  │           │
+    │ │         │   │  robot state │
+    │ │         │   noise          │
+    │ └▲──▲─────┘                  │
+    │  │  │                        │
+    │  │  image(s)                 │
+    │  language tokens             │
+    └──────────────────────────────┘
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        paligemma_with_export_config = PaliGemmaWithExpertConfig(
+            freeze_vision_encoder=self.config.freeze_vision_encoder,
+            train_expert_only=self.config.train_expert_only,
+            attention_implementation=self.config.attention_implementation,
+        )
+        self.paligemma_with_expert = PaliGemmaWithExpertModel(paligemma_with_export_config)
+
+        # Projections are float32
+        self.state_proj = nn.Linear(self.config.max_state_dim, self.config.proj_width)
+        self.action_in_proj = nn.Linear(self.config.max_action_dim, self.config.proj_width)
+        self.action_out_proj = nn.Linear(self.config.proj_width, self.config.max_action_dim)
+
+        self.action_time_mlp_in = nn.Linear(self.config.proj_width * 2, self.config.proj_width)
+        self.action_time_mlp_out = nn.Linear(self.config.proj_width, self.config.proj_width)
+
+        self.set_requires_grad()
+
+    def set_requires_grad(self):
+        for params in self.state_proj.parameters():
+            params.requires_grad = self.config.train_state_proj
+
+    def sample_noise(self, shape, device):
+        noise = torch.normal(
+            mean=0.0,
+            std=1.0,
+            size=shape,
+            dtype=torch.float32,
+            device=device,
+        )
+        return noise
+
+    def sample_time(self, bsize, device):
+        time_beta = sample_beta(1.5, 1.0, bsize, device)
+        time = time_beta * 0.999 + 0.001
+        return time.to(dtype=torch.float32, device=device)
+
+    def embed_prefix(
+        self, images, img_masks, lang_tokens, lang_masks
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Embed images with SigLIP and language tokens with embedding layer to prepare
+        for PaliGemma transformer processing.
+        """
+        # TODO: avoid list in python and torch.cat ; prefer pre-allocation with torch.empty
+        embs = []
+        pad_masks = []
+        att_masks = []
+
+        # TODO: remove for loop
+        for (
+            img,
+            img_mask,
+        ) in zip(images, img_masks, strict=False):
+            img_emb = self.paligemma_with_expert.embed_image(img)
+            img_emb = img_emb.to(dtype=torch.bfloat16)
+
+            # Normalize image embeddings
+            img_emb_dim = img_emb.shape[-1]
+            img_emb = img_emb * torch.tensor(img_emb_dim**0.5, dtype=img_emb.dtype, device=img_emb.device)
+
+            bsize, num_img_embs = img_emb.shape[:2]
+            img_mask = img_mask[:, None].expand(bsize, num_img_embs)
+
+            embs.append(img_emb)
+            pad_masks.append(img_mask)
+
+            # Create attention masks so that image tokens attend to each other
+            att_masks += [0] * num_img_embs
+
+        lang_emb = self.paligemma_with_expert.embed_language_tokens(lang_tokens)
+
+        # Normalize language embeddings
+        lang_emb_dim = lang_emb.shape[-1]
+        lang_emb = lang_emb * math.sqrt(lang_emb_dim)
+
+        embs.append(lang_emb)
+        pad_masks.append(lang_masks)
+
+        # full attention between image and language inputs
+        num_lang_embs = lang_emb.shape[1]
+        att_masks += [0] * num_lang_embs
+
+        embs = torch.cat(embs, dim=1)
+        pad_masks = torch.cat(pad_masks, dim=1)
+        att_masks = torch.tensor(att_masks, dtype=torch.bool, device=pad_masks.device)
+        att_masks = att_masks[None, :].expand(bsize, len(att_masks))
+
+        return embs, pad_masks, att_masks
+
+    def embed_suffix(self, state, noisy_actions, timestep):
+        """Embed state, noisy_actions, timestep to prepare for Expert Gemma processing."""
+        embs = []
+        pad_masks = []
+        att_masks = []
+
+        # Embed state
+        state_emb = self.state_proj(state)
+        state_emb = state_emb.to(dtype=torch.bfloat16)
+        embs.append(state_emb[:, None, :])
+        bsize = state_emb.shape[0]
+        dtype = state_emb.dtype
+        device = state_emb.device
+
+        state_mask = torch.ones(bsize, 1, dtype=torch.bool, device=device)
+        pad_masks.append(state_mask)
+
+        # Set attention masks so that image and language inputs do not attend to state or actions
+        att_masks += [1]
+
+        # Embed timestep using sine-cosine positional encoding with sensitivity in the range [0, 1]
+        time_emb = create_sinusoidal_pos_embedding(
+            timestep, self.config.proj_width, min_period=4e-3, max_period=4.0, device=device
+        )
+        time_emb = time_emb.type(dtype=dtype)
+
+        # Fuse timestep + action information using an MLP
+        action_emb = self.action_in_proj(noisy_actions)
+
+        time_emb = time_emb[:, None, :].expand_as(action_emb)
+        action_time_emb = torch.cat([action_emb, time_emb], dim=2)
+
+        action_time_emb = self.action_time_mlp_in(action_time_emb)
+        action_time_emb = F.silu(action_time_emb)  # swish == silu
+        action_time_emb = self.action_time_mlp_out(action_time_emb)
+
+        # Add to input tokens
+        embs.append(action_time_emb)
+
+        bsize, action_time_dim = action_time_emb.shape[:2]
+        action_time_mask = torch.ones(bsize, action_time_dim, dtype=torch.bool, device=device)
+        pad_masks.append(action_time_mask)
+
+        # Set attention masks so that image, language and state inputs do not attend to action tokens
+        att_masks += [1] + ([0] * (self.config.n_action_steps - 1))
+
+        embs = torch.cat(embs, dim=1)
+        pad_masks = torch.cat(pad_masks, dim=1)
+        att_masks = torch.tensor(att_masks, dtype=embs.dtype, device=embs.device)
+        att_masks = att_masks[None, :].expand(bsize, len(att_masks))
+
+        return embs, pad_masks, att_masks
+
+    def forward(
+        self, images, img_masks, lang_tokens, lang_masks, state, actions, noise=None, time=None
+    ) -> Tensor:
+        """Do a full training forward pass and compute the loss (batch_size x num_steps x num_motors)"""
+        if noise is None:
+            noise = self.sample_noise(actions.shape, actions.device)
+
+        if time is None:
+            time = self.sample_time(actions.shape[0], actions.device)
+
+        time_expanded = time[:, None, None]
+        x_t = time_expanded * noise + (1 - time_expanded) * actions
+        u_t = noise - actions
+
+        prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
+            images, img_masks, lang_tokens, lang_masks
+        )
+        suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(state, x_t, time)
+
+        pad_masks = torch.cat([prefix_pad_masks, suffix_pad_masks], dim=1)
+        att_masks = torch.cat([prefix_att_masks, suffix_att_masks], dim=1)
+
+        att_2d_masks = make_att_2d_masks(pad_masks, att_masks)
+        position_ids = torch.cumsum(pad_masks, dim=1) - 1
+
+        (_, suffix_out), _ = self.paligemma_with_expert.forward(
+            attention_mask=att_2d_masks,
+            position_ids=position_ids,
+            past_key_values=None,
+            inputs_embeds=[prefix_embs, suffix_embs],
+            use_cache=False,
+            fill_kv_cache=False,
+        )
+        suffix_out = suffix_out[:, -self.config.n_action_steps :]
+        # Original openpi code, upcast attention output
+        suffix_out = suffix_out.to(dtype=torch.float32)
+        v_t = self.action_out_proj(suffix_out)
+
+        losses = F.mse_loss(u_t, v_t, reduction="none")
+        return losses
+
+    def sample_actions(self, images, img_masks, lang_tokens, lang_masks, state, noise=None) -> Tensor:
+        """Do a full inference forward and compute the action (batch_size x num_steps x num_motors)"""
+        bsize = state.shape[0]
+        device = state.device
+
+        if noise is None:
+            actions_shape = (bsize, self.config.n_action_steps, self.config.max_action_dim)
+            noise = self.sample_noise(actions_shape, device)
+
+        prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
+            images, img_masks, lang_tokens, lang_masks
+        )
+        prefix_att_2d_masks = make_att_2d_masks(prefix_pad_masks, prefix_att_masks)
+        prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
+
+        # Compute image and language key value cache
+        _, past_key_values = self.paligemma_with_expert.forward(
+            attention_mask=prefix_att_2d_masks,
+            position_ids=prefix_position_ids,
+            past_key_values=None,
+            inputs_embeds=[prefix_embs, None],
+            use_cache=self.config.use_cache,
+            fill_kv_cache=True,
+        )
+
+        dt = -1.0 / self.config.num_steps
+        dt = torch.tensor(dt, dtype=torch.float32, device=device)
+
+        x_t = noise
+        time = torch.tensor(1.0, dtype=torch.float32, device=device)
+        while time >= -dt / 2:
+            expanded_time = time.expand(bsize)
+            v_t = self.denoise_step(
+                state,
+                prefix_pad_masks,
+                past_key_values,
+                x_t,
+                expanded_time,
+            )
+
+            # Euler step
+            x_t += dt * v_t
+            time += dt
+        return x_t
+
+    def denoise_step(
+        self,
+        state,
+        prefix_pad_masks,
+        past_key_values,
+        x_t,
+        timestep,
+    ):
+        """Apply one denoising step of the noise `x_t` at a given timestep."""
+        suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(state, x_t, timestep)
+
+        suffix_len = suffix_pad_masks.shape[1]
+        batch_size = prefix_pad_masks.shape[0]
+        prefix_len = prefix_pad_masks.shape[1]
+        prefix_pad_2d_masks = prefix_pad_masks[:, None, :].expand(batch_size, suffix_len, prefix_len)
+
+        suffix_att_2d_masks = make_att_2d_masks(suffix_pad_masks, suffix_att_masks)
+
+        full_att_2d_masks = torch.cat([prefix_pad_2d_masks, suffix_att_2d_masks], dim=2)
+
+        prefix_offsets = torch.sum(prefix_pad_masks, dim=-1)[:, None]
+        position_ids = prefix_offsets + torch.cumsum(suffix_pad_masks, dim=1) - 1
+
+        outputs_embeds, _ = self.paligemma_with_expert.forward(
+            attention_mask=full_att_2d_masks,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=[None, suffix_embs],
+            use_cache=self.config.use_cache,
+            fill_kv_cache=False,
+        )
+        suffix_out = outputs_embeds[1]
+        suffix_out = suffix_out[:, -self.config.n_action_steps :]
+        suffix_out = suffix_out.to(dtype=torch.float32)
+        v_t = self.action_out_proj(suffix_out)
+        return v_t

lerobot/common/policies/pi0/paligemma_with_expert.py

@@ -0,0 +1,417 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Union
+
+import torch
+import torch.version
+from pytest import Cache
+from torch import nn
+from transformers import (
+    AutoConfig,
+    GemmaForCausalLM,
+    PaliGemmaForConditionalGeneration,
+    PretrainedConfig,
+    PreTrainedModel,
+)
+from transformers.models.auto import CONFIG_MAPPING
+
+from lerobot.common.policies.pi0.flex_attention import flex_attention_forward
+
+
+def apply_rope(x, positions, max_wavelength=10_000):
+    """
+    Applies RoPE positions [B, L] to x [B, L, H, D].
+    """
+    d_half = x.shape[-1] // 2
+    device = x.device
+    dtype = x.dtype
+    x = x.to(torch.float32)
+
+    freq_exponents = (2.0 / x.shape[-1]) * torch.arange(d_half, dtype=torch.float32, device=device)
+    timescale = max_wavelength**freq_exponents
+    radians = positions[..., None].to(torch.float32) / timescale[None, None, :].to(torch.float32)
+
+    radians = radians[..., None, :]
+
+    sin = torch.sin(radians)  # .to(dtype=dtype)
+    cos = torch.cos(radians)  # .to(dtype=dtype)
+
+    x1, x2 = x.split(d_half, dim=-1)
+    res = torch.empty_like(x)
+    res[..., :d_half] = x1 * cos - x2 * sin
+    res[..., d_half:] = x2 * cos + x1 * sin
+
+    return res.to(dtype)
+
+
+class PaliGemmaWithExpertConfig(PretrainedConfig):
+    model_type = "PaliGemmaWithExpertModel"
+    sub_configs = {"paligemma_config": AutoConfig, "gemma_expert_config": AutoConfig}
+
+    def __init__(
+        self,
+        paligemma_config: dict | None = None,
+        gemma_expert_config: dict | None = None,
+        freeze_vision_encoder: bool = True,
+        train_expert_only: bool = True,
+        attention_implementation: str = "eager",
+        **kwargs,
+    ):
+        self.freeze_vision_encoder = freeze_vision_encoder
+        self.train_expert_only = train_expert_only
+        self.attention_implementation = attention_implementation
+
+        if paligemma_config is None:
+            # Default config from Pi0
+            self.paligemma_config = CONFIG_MAPPING["paligemma"](
+                transformers_version="4.48.1",
+                _vocab_size=257152,
+                bos_token_id=2,
+                eos_token_id=1,
+                hidden_size=2048,
+                image_token_index=257152,
+                model_type="paligemma",
+                pad_token_id=0,
+                projection_dim=2048,
+                text_config={
+                    "hidden_activation": "gelu_pytorch_tanh",
+                    "hidden_size": 2048,
+                    "intermediate_size": 16384,
+                    "model_type": "gemma",
+                    "num_attention_heads": 8,
+                    "num_hidden_layers": 18,
+                    "num_image_tokens": 256,
+                    "num_key_value_heads": 1,
+                    "torch_dtype": "float32",
+                    "vocab_size": 257152,
+                },
+                vision_config={
+                    "hidden_size": 1152,
+                    "intermediate_size": 4304,
+                    "model_type": "siglip_vision_model",
+                    "num_attention_heads": 16,
+                    "num_hidden_layers": 27,
+                    "num_image_tokens": 256,
+                    "patch_size": 14,
+                    "projection_dim": 2048,
+                    "projector_hidden_act": "gelu_fast",
+                    "torch_dtype": "float32",
+                    "vision_use_head": False,
+                },
+            )
+        elif isinstance(self.paligemma_config, dict):
+            # Override Pi0 default config for PaliGemma
+            if "model_type" not in gemma_expert_config:
+                paligemma_config["model_type"] = "paligemma"
+
+            cfg_cls = CONFIG_MAPPING[paligemma_config["model_type"]]
+            self.paligemma_config = cfg_cls(**paligemma_config)
+
+        if gemma_expert_config is None:
+            # Default config from Pi0
+            self.gemma_expert_config = CONFIG_MAPPING["gemma"](
+                attention_bias=False,
+                attention_dropout=0.0,
+                bos_token_id=2,
+                eos_token_id=1,
+                head_dim=256,
+                hidden_act="gelu_pytorch_tanh",
+                hidden_activation="gelu_pytorch_tanh",
+                hidden_size=1024,
+                initializer_range=0.02,
+                intermediate_size=4096,
+                max_position_embeddings=8192,
+                model_type="gemma",
+                num_attention_heads=8,
+                num_hidden_layers=18,
+                num_key_value_heads=1,
+                pad_token_id=0,
+                rms_norm_eps=1e-06,
+                rope_theta=10000.0,
+                torch_dtype="float32",
+                transformers_version="4.48.1",
+                use_cache=True,
+                vocab_size=257152,
+            )
+        elif isinstance(self.gemma_expert_config, dict):
+            # Override Pi0 default config for Gemma Expert
+            if "model_type" not in gemma_expert_config:
+                gemma_expert_config["model_type"] = "gemma"
+
+            cfg_cls = CONFIG_MAPPING[paligemma_config["model_type"]]
+            self.gemma_expert_config = cfg_cls(**gemma_expert_config)
+
+        super().__init__(**kwargs)
+
+    def __post_init__(self):
+        super().__post_init__()
+        if self.train_expert_only and not self.freeze_vision_encoder:
+            raise ValueError(
+                "You set `freeze_vision_encoder=False` and `train_expert_only=True` which are not compatible."
+            )
+
+        if self.attention_implementation not in ["eager", "fa2", "flex"]:
+            raise ValueError(
+                f"Wrong value provided for `attention_implementation` ({self.attention_implementation}). Expected 'eager', 'fa2' or 'flex'."
+            )
+
+
+class PaliGemmaWithExpertModel(PreTrainedModel):
+    config_class = PaliGemmaWithExpertConfig
+
+    def __init__(self, config: PaliGemmaWithExpertConfig):
+        super().__init__(config=config)
+        self.config = config
+        self.paligemma = PaliGemmaForConditionalGeneration(config=config.paligemma_config)
+        self.gemma_expert = GemmaForCausalLM(config=config.gemma_expert_config)
+        # Remove unused embed_tokens
+        self.gemma_expert.model.embed_tokens = None
+
+        self.to_bfloat16_like_physical_intelligence()
+        self.set_requires_grad()
+
+    def set_requires_grad(self):
+        if self.config.freeze_vision_encoder:
+            self.paligemma.vision_tower.eval()
+            for params in self.paligemma.vision_tower.parameters():
+                params.requires_grad = False
+
+        if self.config.train_expert_only:
+            self.paligemma.eval()
+            for params in self.paligemma.parameters():
+                params.requires_grad = False
+
+    def train(self, mode: bool = True):
+        super().train(mode)
+
+        if self.config.freeze_vision_encoder:
+            self.paligemma.vision_tower.eval()
+
+        if self.config.train_expert_only:
+            self.paligemma.eval()
+
+    def to_bfloat16_like_physical_intelligence(self):
+        self.paligemma = self.paligemma.to(dtype=torch.bfloat16)
+
+        params_to_change_dtype = [
+            "language_model.model.layers",
+            "gemma_expert.model.layers",
+            "vision_tower",
+            "multi_modal",
+        ]
+        for name, param in self.named_parameters():
+            if any(selector in name for selector in params_to_change_dtype):
+                param.data = param.data.to(dtype=torch.bfloat16)
+
+    def embed_image(self, image: torch.Tensor):
+        return self.paligemma.get_image_features(image)
+
+    def embed_language_tokens(self, tokens: torch.Tensor):
+        return self.paligemma.language_model.model.embed_tokens(tokens)
+
+    # TODO: break down this huge forward into modules or functions
+    def forward(
+        self,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None,
+        inputs_embeds: List[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        fill_kv_cache: Optional[bool] = None,
+    ):
+        models = [self.paligemma.language_model.model, self.gemma_expert.model]
+
+        for hidden_states in inputs_embeds:
+            # TODO this is very inefficient
+            # dtype is always the same, batch size too (if > 1 len)
+            # device could be trickier in multi gpu edge cases but that's it
+            if hidden_states is None:
+                continue
+            batch_size = hidden_states.shape[0]
+
+        # RMSNorm
+        num_layers = self.paligemma.config.text_config.num_hidden_layers
+        head_dim = self.paligemma.config.text_config.head_dim
+        for layer_idx in range(num_layers):
+            query_states = []
+            key_states = []
+            value_states = []
+            for i, hidden_states in enumerate(inputs_embeds):
+                if hidden_states is None:
+                    continue
+                layer = models[i].layers[layer_idx]
+                # normalizer = torch.tensor(models[i].config.hidden_size**0.5, dtype=hidden_states.dtype)
+                # hidden_states = hidden_states * normalizer
+                hidden_states = layer.input_layernorm(hidden_states)
+
+                input_shape = hidden_states.shape[:-1]
+                hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
+
+                hidden_states = hidden_states.to(dtype=torch.bfloat16)
+                query_state = layer.self_attn.q_proj(hidden_states).view(hidden_shape)
+                key_state = layer.self_attn.k_proj(hidden_states).view(hidden_shape)
+                value_state = layer.self_attn.v_proj(hidden_states).view(hidden_shape)
+
+                query_states.append(query_state)
+                key_states.append(key_state)
+                value_states.append(value_state)
+
+            # B,L,H,D with L sequence length, H number of heads, D head dim
+            # concatenate on the number of embeddings/tokens
+            query_states = torch.cat(query_states, dim=1)
+            key_states = torch.cat(key_states, dim=1)
+            value_states = torch.cat(value_states, dim=1)
+
+            query_states = apply_rope(query_states, position_ids)
+            key_states = apply_rope(key_states, position_ids)
+
+            if use_cache and past_key_values is None:
+                past_key_values = {}
+
+            if use_cache:
+                if fill_kv_cache:
+                    past_key_values[layer_idx] = {
+                        "key_states": key_states,
+                        "value_states": value_states,
+                    }
+                else:
+                    # TODO here, some optimization can be done - similar to a `StaticCache` we can declare the `max_len` before.
+                    # so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
+                    # the max len, then we (for instance) double the cache size. This implementation already exists
+                    # in `transformers`. (molbap)
+                    key_states = torch.cat([past_key_values[layer_idx]["key_states"], key_states], dim=1)
+                    value_states = torch.cat(
+                        [past_key_values[layer_idx]["value_states"], value_states], dim=1
+                    )
+
+            attention_interface = self.get_attention_interface()
+            att_output = attention_interface(
+                attention_mask, batch_size, head_dim, query_states, key_states, value_states
+            )
+            att_output = att_output.to(dtype=torch.bfloat16)
+
+            # first part of att_output is prefix (up to sequence length, [:, 0:prefix_seq_len])
+            outputs_embeds = []
+            start = 0
+            for i, hidden_states in enumerate(inputs_embeds):
+                layer = models[i].layers[layer_idx]
+
+                if hidden_states is not None:
+                    end = start + hidden_states.shape[1]
+
+                    if att_output.dtype != layer.self_attn.o_proj.weight.dtype:
+                        att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
+                    out_emb = layer.self_attn.o_proj(att_output[:, start:end])
+
+                    # TODO: first dropout (by default 0.0)
+
+                    # first residual
+                    out_emb += hidden_states
+                    after_first_residual = out_emb.clone()
+
+                    out_emb = layer.post_attention_layernorm(out_emb)
+                    out_emb = layer.mlp(out_emb)
+
+                    # TODO: second dropout (by default 0.0)
+
+                    # second residual
+                    out_emb += after_first_residual
+
+                    outputs_embeds.append(out_emb)
+
+                    start = end
+                else:
+                    outputs_embeds.append(None)
+
+            inputs_embeds = outputs_embeds
+
+        # final norm
+        outputs_embeds = []
+        for i, hidden_states in enumerate(inputs_embeds):
+            if hidden_states is not None:
+                out_emb = models[i].norm(hidden_states)
+                outputs_embeds.append(out_emb)
+            else:
+                outputs_embeds.append(None)
+
+        return outputs_embeds, past_key_values
+
+    def get_attention_interface(self):
+        if self.config.attention_implementation == "fa2":
+            attention_interface = self.flash_attention_forward
+        elif self.config.attention_implementation == "flex":
+            attention_interface = flex_attention_forward
+        else:
+            attention_interface = self.eager_attention_forward
+        return attention_interface
+
+    def flash_attention_forward(
+        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
+    ):
+        raise NotImplementedError("FA2 is not implemented (yet)")
+
+    def eager_attention_forward(
+        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
+    ):
+        num_att_heads = self.config.paligemma_config.text_config.num_attention_heads
+        num_key_value_heads = self.config.paligemma_config.text_config.num_key_value_heads
+        num_key_value_groups = num_att_heads // num_key_value_heads
+
+        # query_states: batch_size, sequence_length, num_att_head, head_dim
+        # key_states: batch_size, sequence_length, num_key_value_head, head_dim
+        # value_states: batch_size, sequence_length, num_key_value_head, head_dim
+        sequence_length = key_states.shape[1]
+
+        key_states = key_states[:, :, :, None, :].expand(
+            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
+        )
+        key_states = key_states.reshape(
+            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
+        )
+
+        value_states = value_states[:, :, :, None, :].expand(
+            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
+        )
+        value_states = value_states.reshape(
+            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
+        )
+
+        # Attention here is upcasted to float32 to match the original eager implementation.
+
+        query_states = query_states.to(dtype=torch.float32)
+        key_states = key_states.to(dtype=torch.float32)
+
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+
+        att_weights = torch.matmul(query_states, key_states.transpose(2, 3))
+        att_weights *= head_dim**-0.5
+        big_neg = -2.3819763e38  # See gemma/modules.py
+
+        masked_att_weights = torch.where(attention_mask[:, None, :, :], att_weights, big_neg)
+
+        probs = nn.functional.softmax(masked_att_weights, dim=-1)
+        probs = probs.to(dtype=value_states.dtype)
+
+        # probs: batch_size, num_key_value_head, num_att_head, sequence_length, sequence_length
+        # value_states: batch_size, sequence_length, num_att_heads, head_dim
+
+        att_output = torch.matmul(probs, value_states.permute(0, 2, 1, 3))
+
+        att_output = att_output.permute(0, 2, 1, 3)
+        # we use -1 because sequence length can change
+        att_output = att_output.reshape(batch_size, -1, num_key_value_heads * num_key_value_groups * head_dim)
+
+        return att_output

lerobot/common/policies/pi0fast/configuration_pi0fast.py

@@ -0,0 +1,136 @@
+from dataclasses import dataclass, field
+
+from lerobot.common.optim.optimizers import AdamWConfig
+from lerobot.common.optim.schedulers import (
+    CosineDecayWithWarmupSchedulerConfig,
+)
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+
+
+@PreTrainedConfig.register_subclass("pi0fast")
+@dataclass
+class PI0FASTConfig(PreTrainedConfig):
+    # Input / output structure.
+    n_obs_steps: int = 1
+    chunk_size: int = 10
+    n_action_steps: int = 5
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MEAN_STD,
+        }
+    )
+
+    # Shorter state and action vectors will be padded
+    max_state_dim: int = 32  # 32
+    max_action_dim: int = 32  # 32
+
+    # Image preprocessing
+    resize_imgs_with_padding: tuple[int, int] = (224, 224)
+    interpolate_like_pi: bool = False
+
+    # Add empty images. Used by pi0_aloha_sim which adds the empty
+    # left and right wrist cameras in addition to the top camera.
+    empty_cameras: int = 0
+
+    # Converts the joint and gripper values from the standard Aloha space to
+    # the space used by the pi internal runtime which was used to train the base model.
+    adapt_to_pi_aloha: bool = False
+
+    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
+    # Gripper dimensions will remain in absolute values.
+    use_delta_joint_actions_aloha: bool = False
+
+    # Tokenizer
+    tokenizer_max_length: int = 48
+
+    # Projector
+    proj_width: int = 1024
+
+    # Decoding
+    max_decoding_steps: int = 256
+    fast_skip_tokens: int = 128  # Skip last 128 tokens in PaliGemma vocab since they are special tokens
+    max_input_seq_len: int = 256  # 512
+
+    # Utils
+    use_cache: bool = True
+
+    # Frozen parameters
+    freeze_vision_encoder: bool = True
+    freeze_lm_head: bool = True
+
+    # Training presets
+    optimizer_lr: float = 1e-4
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-5
+
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    checkpoint_path: str = None
+
+    padding_side: str = "right"
+
+    precision: str = "bfloat16"
+    grad_clip_norm: float = 1
+
+    # Allows padding/truncation of generated action tokens during detokenization to ensure decoding.
+    # In the original version, tensors of 0s were generated if shapes didn't match for stable decoding.
+    relaxed_action_decoding: bool = True
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        """Input validation (not exhaustive)."""
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
+                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
+            )
+        if self.n_obs_steps != 1:
+            raise ValueError(
+                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
+            )
+
+    def validate_features(self) -> None:
+        for i in range(self.empty_cameras):
+            key = f"observation.images.empty_camera_{i}"
+            empty_camera = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(3, 480, 640),
+            )
+            self.input_features[key] = empty_camera
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

lerobot/common/policies/pi0fast/modeling_pi0fast.py

@@ -0,0 +1,973 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+π0+FAST: Efficient Action Tokenization for Vision-Language-Action Models
+
+[Paper](https://arxiv.org/abs/2501.09747)
+[Jax code](https://github.com/Physical-Intelligence/openpi)
+
+Designed by Physical Intelligence. Ported from Jax by Hugging Face.
+
+Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
+```bash
+python lerobot/scripts/train.py \
+--policy.path=lerobot/pi0fast_base \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of training the pi0+FAST neural network with from scratch:
+```bash
+python lerobot/scripts/train.py \
+--policy.type=pi0fast \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of using the pi0 pretrained model outside LeRobot training framework:
+```python
+policy = PI0FASTPolicy.from_pretrained("lerobot/pi0fast_base")
+```
+
+"""
+
+from collections import deque
+from functools import partial
+
+import numpy as np
+import torch
+import torch.nn.functional as F  # noqa: N812
+from PIL import Image
+from scipy.fft import idct
+from torch import Tensor, nn
+from transformers import AutoProcessor, AutoTokenizer, PaliGemmaForConditionalGeneration
+from transformers.cache_utils import HybridCache, StaticCache
+from transformers.models.auto import CONFIG_MAPPING
+
+from lerobot.common.constants import ACTION, OBS_ROBOT
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+
+PRECISION = {
+    "float16": torch.float16,
+    "float32": torch.float32,
+    "bfloat16": torch.bfloat16,
+}
+
+
+def normalize(x, min_val, max_val):
+    return (x - min_val) / (max_val - min_val)
+
+
+def unnormalize(x, min_val, max_val):
+    return x * (max_val - min_val) + min_val
+
+
+def safe_arcsin(value):
+    # This ensures that the input stays within
+    # [−1,1] to avoid invalid values for arcsin
+    return torch.arcsin(torch.clamp(value, -1.0, 1.0))
+
+
+def aloha_gripper_to_angular(value):
+    # Aloha transforms the gripper positions into a linear space. The following code
+    # reverses this transformation to be consistent with pi0 which is pretrained in
+    # angular space.
+    #
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_POSITION_OPEN, PUPPET_GRIPPER_POSITION_CLOSED
+    value = unnormalize(value, min_val=0.01844, max_val=0.05800)
+
+    # This is the inverse of the angular to linear transformation inside the Interbotix code.
+    def linear_to_radian(linear_position, arm_length, horn_radius):
+        value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
+        return safe_arcsin(value)
+
+    # The constants are taken from the Interbotix code.
+    value = linear_to_radian(value, arm_length=0.036, horn_radius=0.022)
+
+    # Normalize to [0, 1].
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+def aloha_gripper_from_angular(value):
+    # Convert from the gripper position used by pi0 to the gripper position that is used by Aloha.
+    # Note that the units are still angular but the range is different.
+
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    value = unnormalize(value, min_val=0.4, max_val=1.5)
+
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_JOINT_OPEN, PUPPET_GRIPPER_JOINT_CLOSE
+    return normalize(value, min_val=-0.6213, max_val=1.4910)
+
+
+def aloha_gripper_from_angular_inv(value):
+    # Directly inverts the gripper_from_angular function.
+    value = unnormalize(value, min_val=-0.6213, max_val=1.4910)
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+class PI0FASTPolicy(PreTrainedPolicy):
+    """Wrapper class around PI0FAST tokenizer and model to train and run inference within LeRobot."""
+
+    config_class = PI0FASTConfig
+    name = "pi0fast"
+
+    def __init__(
+        self,
+        config: PI0FASTConfig,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                    the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_targets = Normalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+
+        self.language_tokenizer = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
+        self.model = PI0FAST(config)
+
+        self.reset()
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    def get_optim_params(self) -> dict:
+        return self.parameters()
+
+    def _pi_aloha_decode_state(self, state):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            state[:, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            state[:, motor_idx] = aloha_gripper_to_angular(state[:, motor_idx])
+        return state
+
+    def _pi_aloha_encode_actions(self, actions):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
+        return actions
+
+    def _pi_aloha_encode_actions_inv(self, actions):
+        # Flip the joints again.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
+        return actions
+
+    @torch.no_grad
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method wraps `select_actions` in order to return one action at a time for execution in the
+        environment. It works by managing the actions in a queue and only calling `select_actions` when the
+        queue is empty.
+        """
+        self.eval()
+
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+
+        batch = self.normalize_inputs(batch)
+
+        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
+        # querying the policy.
+        if len(self._action_queue) == 0:
+            actions = self.model.generate_actions(batch)
+
+            actions = actions[:, : self.config.n_action_steps]
+
+            original_action_dim = self.config.action_feature.shape[
+                0
+            ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
+            actions = actions[:, :, :original_action_dim]
+
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.adapt_to_pi_aloha:
+                actions = self._pi_aloha_encode_actions(actions)
+
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
+        loss_dict = self.model.forward(batch)
+        return loss_dict["loss"], loss_dict
+
+
+def block_causal_update_causal_mask(
+    attention_mask,
+    token_type_ids=None,
+    past_key_values=None,
+    cache_position=None,
+    input_tensor=None,
+    attn_implementation: str = "eager",
+    dtype: torch.dtype = "float32",
+):
+    """
+    Update the causal mask during training and generation. It can be customized to different attention masks.
+    """
+    if attn_implementation == "flash_attention_2":
+        if attention_mask is not None and 0.0 in attention_mask:
+            return attention_mask
+        return None
+    using_static_cache = isinstance(past_key_values, StaticCache)
+    min_dtype = torch.finfo(dtype).min
+
+    if input_tensor is None:
+        input_tensor = attention_mask
+
+    inputs_lead_dim, sequence_length = input_tensor.shape[:2]
+
+    if using_static_cache or isinstance(past_key_values, HybridCache):
+        target_length = past_key_values.get_max_cache_shape()
+    else:
+        target_length = (
+            attention_mask.shape[-1]
+            if isinstance(attention_mask, torch.Tensor)
+            else cache_position[0] + sequence_length + 1
+        )
+
+    # Handle precomputed attention masks
+    if attention_mask is not None and attention_mask.dim() == 4:
+        return attention_mask
+
+    # Causal mask initialization
+    causal_mask = torch.full(
+        (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
+    )
+
+    # Standard causal masking (triu ensures tokens can only attend to past)
+    if sequence_length != 1:
+        causal_mask = torch.triu(causal_mask, diagonal=1)
+
+        # Apply block causal mask
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.to(causal_mask.device).bool()
+            cumsum = torch.cumsum(token_type_ids, dim=1)
+            block_causal_mask = cumsum[:, None, :] <= cumsum[:, :, None]
+
+            # Combine causal_mask with block-wise attention mask
+            causal_mask = torch.where(block_causal_mask, 0.0, causal_mask)
+            causal_mask = causal_mask[:, None, :, :]
+        else:
+            # Apply past cache position constraint
+            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
+                -1, 1
+            )
+            causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
+    else:
+        # Apply past cache position constraint
+        causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
+            -1, 1
+        )
+        causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
+
+    if attention_mask is not None:
+        causal_mask = causal_mask.clone()  # Copy to contiguous memory for in-place edits
+        mask_length = attention_mask.shape[-1]
+
+        # Apply padding mask
+        padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+            causal_mask.device
+        )
+        padding_mask = padding_mask == 0
+        causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+            padding_mask, min_dtype
+        )
+
+    return causal_mask
+
+
+def prepare_inputs_for_generation(
+    # self,
+    input_ids,
+    past_key_values=None,
+    inputs_embeds=None,
+    cache_position=None,
+    position_ids=None,
+    pixel_values=None,
+    attention_mask=None,
+    token_type_ids=None,
+    use_cache=True,
+    num_logits_to_keep=None,
+    labels=None,
+    self=None,
+    **kwargs,
+):
+    # create block causal attention
+    if cache_position[0] > 0 and input_ids.shape[1] > 0:
+        input_tensor = input_ids[:, -1:]
+        new_positions = (
+            torch.ones(
+                (position_ids.shape[0], input_ids.shape[1]),
+                dtype=position_ids.dtype,
+                device=position_ids.device,
+            ).cumsum(-1)
+            + position_ids[:, -1:]
+        )
+        position_ids = torch.cat([position_ids, new_positions], dim=-1)
+    else:
+        input_tensor = inputs_embeds
+    attention_mask = block_causal_update_causal_mask(
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        cache_position=cache_position,
+        input_tensor=input_tensor,
+        token_type_ids=token_type_ids,
+        dtype=self.dtype,
+        attn_implementation=self.config.text_config._attn_implementation,
+    )
+    # Overwritten -- custom `position_ids` and `pixel_values` handling
+    model_inputs = self.language_model.prepare_inputs_for_generation(
+        input_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        cache_position=cache_position,
+        use_cache=use_cache,
+        num_logits_to_keep=num_logits_to_keep,
+        token_type_ids=token_type_ids,
+        **kwargs,
+    )
+
+    # Position_ids in Paligemma are 1-indexed
+    if model_inputs.get("position_ids") is not None:
+        model_inputs["position_ids"] += 1
+    # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+    # Otherwise we need pixel values to be passed to model. NOTE: use_cache=False needs pixel_values always
+    if cache_position[0] == 0:
+        model_inputs["pixel_values"] = pixel_values
+    is_training = token_type_ids is not None and labels is not None
+    if cache_position[0] == 0 and isinstance(past_key_values, HybridCache):
+        input_tensor = inputs_embeds if inputs_embeds is not None else input_ids
+        causal_mask = self._update_causal_mask(
+            attention_mask, token_type_ids, past_key_values, cache_position, input_tensor, is_training
+        )
+        model_inputs["attention_mask"] = causal_mask
+
+    return model_inputs
+
+
+class PI0FAST(nn.Module):
+    def __init__(self, config: PI0FASTConfig):
+        super().__init__()
+        self.config = config
+
+        # TODO: move tokenizers in Policy
+        fast_tokenizer_path = "physical-intelligence/fast"
+        pi0_paligemma_path = "google/paligemma-3b-pt-224"
+        self.paligemma_tokenizer = AutoTokenizer.from_pretrained(pi0_paligemma_path)
+        self.processor = AutoProcessor.from_pretrained(pi0_paligemma_path)
+        self.fast_tokenizer = AutoProcessor.from_pretrained(fast_tokenizer_path, trust_remote_code=True)
+        self.fast_skip_tokens = self.config.fast_skip_tokens
+        self.max_input_seq_len = self.config.max_input_seq_len
+        self.action_horizon = self.config.chunk_size
+        self.action_dim = self.config.action_feature.shape[
+            0
+        ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
+        precision = config.precision
+        torch_precision = PRECISION.get(precision, torch.float32)
+        self.pad_token_id = (
+            self.paligemma_tokenizer.pad_token_id
+            if hasattr(self.paligemma_tokenizer, "pad_token_id")
+            else self.paligemma_tokenizer.eos_token_id
+        )
+
+        paligemma_config = CONFIG_MAPPING["paligemma"](
+            transformers_version="4.48.1",
+            _vocab_size=257152,
+            bos_token_id=2,
+            eos_token_id=1,
+            hidden_size=2048,
+            image_token_index=257152,
+            model_type="paligemma",
+            pad_token_id=0,
+            projection_dim=2048,
+            text_config={
+                "hidden_activation": "gelu_pytorch_tanh",
+                "hidden_size": 2048,
+                "intermediate_size": 16384,
+                "model_type": "gemma",
+                "num_attention_heads": 8,
+                "num_hidden_layers": 18,
+                "num_image_tokens": 256,
+                "num_key_value_heads": 1,
+                "torch_dtype": precision,
+                "vocab_size": 257152,
+                "_attn_implementation": "eager",
+            },
+            vision_config={
+                "hidden_size": 1152,
+                "intermediate_size": 4304,
+                "model_type": "siglip_vision_model",
+                "num_attention_heads": 16,
+                "num_hidden_layers": 27,
+                "num_image_tokens": 256,
+                "patch_size": 14,
+                "projection_dim": 2048,
+                "projector_hidden_act": "gelu_pytorch_tanh",
+                "torch_dtype": precision,
+                "vision_use_head": False,
+            },
+        )
+        self.pi0_paligemma = PaliGemmaForConditionalGeneration(config=paligemma_config)
+
+        self.pi0_paligemma.prepare_inputs_for_generation = partial(
+            prepare_inputs_for_generation, self=self.pi0_paligemma
+        )
+        # change important stuff in bf16
+        params_to_change_dtype = [
+            "language_model",
+            "vision_tower",
+            "multi_modal",
+        ]
+        for name, param in self.pi0_paligemma.named_parameters():
+            if any(selector in name for selector in params_to_change_dtype):
+                param.data = param.data.to(dtype=torch_precision)
+        self.set_requires_grad()
+        self.image_keys = self.config.image_features.keys()
+        self.ignore_index = self.pi0_paligemma.config.ignore_index
+        self.padding_side = self.config.padding_side
+
+    def set_requires_grad(self):
+        if self.config.freeze_vision_encoder:
+            self.pi0_paligemma.vision_tower.eval()
+            for params in self.pi0_paligemma.vision_tower.parameters():
+                params.requires_grad = False
+        # To avoid unused params issue with distributed training
+        if self.config.freeze_lm_head:
+            for name, params in self.pi0_paligemma.named_parameters():
+                if "embed_tokens" in name:  # lm heads and embedding layer are tied
+                    params.requires_grad = False
+
+    def embed_tokens(self, tokens: torch.Tensor):
+        return self.pi0_paligemma.language_model.model.embed_tokens(tokens)
+
+    def prepare_inputs_for_generation(self, *args, **kwargs):
+        return self.pi0_paligemma.prepare_inputs_for_generation(*args, **kwargs)
+
+    def prepare_images(self, batch):
+        """Preprocess LeRobot batch into Pi0 inputs"""
+        images = []
+        img_masks = []
+        present_img_keys = [key for key in self.image_keys if key in batch]
+        if len(present_img_keys) == 0:
+            raise ValueError(
+                f"All image features are missing from the batch. At least one expected. (batch: {batch.keys()}) (image_features:{self.config.image_features})"
+            )
+
+        # Preprocess image features present in the batch
+        num_empty_cameras = 0
+        for key in self.image_keys:
+            if key in present_img_keys:
+                img = batch[key]
+
+                if self.config.resize_imgs_with_padding is not None:
+                    img = resize_with_pad(
+                        img,
+                        *self.config.resize_imgs_with_padding,
+                        pad_value=0,
+                        interpolate_like_pi=self.config.interpolate_like_pi,
+                    )
+
+                # Normalize from range [0,1] to [-1,1] as expacted by siglip
+                img = img * 2.0 - 1.0
+
+                bsize = img.shape[0]
+                device = img.device
+                mask = torch.ones(bsize, dtype=torch.bool, device=device)
+            else:
+                if num_empty_cameras >= self.config.empty_cameras:
+                    continue
+                img = torch.ones_like(img) * -1
+                bsize = img.shape[0]
+                device = img.device
+                mask = torch.ones(bsize, dtype=torch.bool, device=device)
+                num_empty_cameras += 1
+
+            images.append(img)
+            img_masks.append(mask)
+        return images, img_masks
+
+    def normalize_actions(self, actions: torch.Tensor) -> torch.Tensor:
+        mins = actions.amin(dim=(1, 2), keepdim=True)  # [0]
+        maxs = actions.amax(dim=(1, 2), keepdim=True)  # [0]
+        return 2 * (actions - mins) / (maxs - mins + 1e-8) - 1
+
+    def _act_tokens_to_paligemma_tokens(self, tokens: torch.Tensor) -> torch.Tensor:
+        out = self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens - tokens
+        return out
+
+    def fast_tokenizer_wrapper(self, actions_norm):
+        """
+        A wrapper for self.fast_tokenizer that ensures batch processing,
+        conversion to PyTorch tensors, and returns a dictionary without padding.
+        """
+        batch_tokens = self.fast_tokenizer(actions_norm)
+        fast_out = self.processor.tokenizer.pad({"input_ids": batch_tokens}, return_tensors="pt")
+
+        return fast_out
+
+    def create_token_type_ids(self, padded_mask: torch.Tensor, prefix_len: int) -> torch.Tensor:
+        token_type_ids = torch.zeros_like(padded_mask, dtype=torch.bool)
+        # Compute cumulative sum mask
+        cumsum_mask = (padded_mask != 0).cumsum(dim=1)
+        # Suffix block (everything after prefix_len)
+        suffix_mask = cumsum_mask > prefix_len
+        token_type_ids = suffix_mask
+        return token_type_ids
+
+    def create_input_tokens(self, state, lang_text, actions=None):
+        bsize = state.shape[0]
+        device = state.device
+        bins = torch.linspace(-1, 1, 256 + 1, device=device)[:-1]
+        discretized = torch.bucketize(state, bins) - 1
+        discretized = discretized[:, :32]
+
+        prefix_texts = []
+        state_text = []
+        for txt, disc in zip(lang_text, discretized, strict=False):
+            cleaned = txt.lower().strip().replace("_", " ")
+            state_str = " ".join(str(val.item()) for val in disc)
+            prefix_texts.append(f"Task: {cleaned}, State: {state_str};\n")
+            state_text.append(f"State: {state_str};\n")
+
+        prefix_out = self.paligemma_tokenizer(
+            prefix_texts, add_special_tokens=True, return_tensors="pt", padding="longest", truncation=False
+        )
+        prefix_ids = prefix_out["input_ids"].to(device)
+        prefix_mask = prefix_out["attention_mask"].to(device)
+        prefix_lens = prefix_mask.sum(dim=1)[:, None].cpu()
+
+        if actions is not None:
+            actions_norm = self.normalize_actions(actions)
+            actions_pad = F.pad(
+                actions_norm, (0, max(0, self.config.max_action_dim - actions_norm.shape[2])), value=0
+            )[:, :, : self.config.max_action_dim]
+            fast_out = self.fast_tokenizer_wrapper(
+                actions_pad.cpu(),
+            )
+            act_ids = fast_out["input_ids"]
+            act_mask = fast_out["attention_mask"].to(device)
+
+            act_ids = self._act_tokens_to_paligemma_tokens(act_ids).to(device)
+            # Replace action with 0 to pad tokens
+            act_ids = torch.where(
+                act_ids == self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens,
+                self.pad_token_id,
+                act_ids,
+            )
+
+            eos_token = torch.tensor(
+                [self.paligemma_tokenizer.eos_token_id], dtype=torch.long, device=device
+            ).expand(bsize, -1)
+            eos_mask = torch.tensor([1], dtype=torch.long, device=device).expand(bsize, -1)
+            bos = self.paligemma_tokenizer("Action: ", add_special_tokens=False, return_tensors="pt")
+            bos_token = bos["input_ids"].expand(act_ids.shape[0], -1).to(device)
+            bos_mask = bos["attention_mask"].expand(act_ids.shape[0], -1).to(device)
+            act_ids = torch.cat([bos_token, act_ids, eos_token], dim=1)
+            act_mask = torch.cat([bos_mask, act_mask, eos_mask], dim=1)
+            act_mask = act_mask.to(device)
+        else:
+            act_ids = torch.empty(bsize, self.pad_token_id, dtype=torch.long, device=device)
+            act_mask = torch.empty(bsize, 0, dtype=torch.long, device=device)
+        final_ids = torch.cat([prefix_ids, act_ids], dim=1)
+
+        final_mask = torch.cat([prefix_mask, act_mask], dim=1)
+        batch_inputs = {"input_ids": final_ids.tolist(), "attention_mask": final_mask.tolist()}
+
+        # Use tokenizer pad function
+        padded_output = self.paligemma_tokenizer.pad(
+            batch_inputs, padding="longest", max_length=180, return_tensors="pt"
+        )
+        padded_mask = padded_output["attention_mask"]
+
+        # define tensor of padding lengths
+        att_mask = (padded_mask != 0).cumsum(dim=1) > prefix_lens
+
+        token_type_ids = self.create_token_type_ids(padded_mask=padded_mask, prefix_len=prefix_lens)
+
+        padded_output["padded_mask"] = padded_output.pop("attention_mask")
+        padded_output["attention_mask"] = att_mask
+        # loss is computed not on prefix, and not on padding
+        padded_output["loss_mask"] = att_mask & padded_output["padded_mask"]
+        padded_output["token_type_ids"] = token_type_ids
+        return padded_output
+
+    def shift_padding_side(
+        self,
+        tokens: torch.Tensor,
+        ar_mask: torch.Tensor,
+        padding_mask: torch.Tensor,
+        loss_mask: torch.Tensor,
+        targets: torch.Tensor,
+        token_type_ids: torch.Tensor,
+        padding_side: str = "right",
+    ) -> tuple[torch.Tensor]:
+        if padding_side not in ["right", "left"]:
+            return tokens, ar_mask, padding_mask, loss_mask, targets, token_type_ids
+
+        new_tokens = torch.empty_like(tokens)
+        new_ar_masks = torch.empty_like(ar_mask)
+        new_padding_mask = torch.empty_like(padding_mask)
+        new_loss_mask = torch.empty_like(loss_mask)
+        new_targets = torch.empty_like(targets)
+        new_token_type_ids = torch.empty_like(token_type_ids)
+        batch_size = tokens.shape[0]
+        for i in range(batch_size):
+            padding_indices = torch.where(padding_mask[i] == 0)[0]
+            non_padding_indices = torch.where(padding_mask[i] == 1)[0]
+            if padding_side == "left":
+                new_indices = torch.cat((padding_indices, non_padding_indices), dim=0)
+            else:
+                new_indices = torch.cat((non_padding_indices, padding_indices), dim=0)
+            new_tokens[i] = tokens[i].index_select(0, new_indices)
+            new_ar_masks[i] = ar_mask[i].index_select(0, new_indices)
+            new_padding_mask[i] = padding_mask[i].index_select(0, new_indices)
+            new_loss_mask[i] = loss_mask[i].index_select(0, new_indices)
+            new_targets[i] = targets[i].index_select(0, new_indices)
+            new_token_type_ids[i] = token_type_ids[i].index_select(0, new_indices)
+
+        return new_tokens, new_ar_masks, new_padding_mask, new_loss_mask, new_targets, new_token_type_ids
+
+    def forward(self, batch: dict[str, Tensor]):
+        device = batch[OBS_ROBOT].device
+        # TODO: keep like this or move to the policy .forward
+        images, img_masks = self.prepare_images(batch)
+
+        padded_outs = self.create_input_tokens(
+            state=batch[OBS_ROBOT],
+            lang_text=batch["task"],
+            actions=batch[ACTION],
+        )
+
+        embs, pad_masks, _, targets, loss_mask, token_type_ids = self.embed_inputs(
+            images,
+            img_masks,
+            padded_outs["input_ids"],
+            padded_outs["padded_mask"],
+            padded_outs["attention_mask"],
+            padded_outs["loss_mask"],
+            padded_outs["token_type_ids"],
+            padding_side=self.padding_side,
+        )
+        position_ids = torch.cumsum(pad_masks, dim=1) - 1
+        token_type_ids = token_type_ids.to(dtype=torch.int64)
+        past_seen_tokens = 0
+        cache_position = torch.arange(past_seen_tokens, past_seen_tokens + embs.shape[1], device=embs.device)
+        pad_masks = block_causal_update_causal_mask(
+            attention_mask=pad_masks,
+            past_key_values=None,
+            cache_position=cache_position,
+            input_tensor=embs,
+            token_type_ids=token_type_ids,
+            dtype=self.pi0_paligemma.dtype,
+            attn_implementation=self.pi0_paligemma.config.text_config._attn_implementation,
+        )
+        outputs = self.pi0_paligemma.forward(
+            input_ids=None,
+            token_type_ids=None,
+            attention_mask=pad_masks,
+            position_ids=position_ids,
+            past_key_values=None,
+            inputs_embeds=embs,
+            use_cache=False,
+            labels=None,
+        )
+
+        logits = outputs.logits
+
+        loss_fct = nn.CrossEntropyLoss(reduction="none")
+
+        # Shift left for next-step prediction
+        logits = logits[:, :-1, :]
+        targets = targets[:, 1:].to(device)  # Shift targets
+        loss_mask = loss_mask[:, 1:].to(device)  # Ensure correct shape
+
+        # Compute per-token loss
+        token_loss = loss_fct(logits.reshape(-1, logits.shape[-1]), targets.reshape(-1))
+
+        # Apply loss mask
+        token_loss = token_loss * loss_mask.reshape(-1)
+
+        # Compute final loss
+        loss = token_loss.sum() / torch.clamp(loss_mask.sum(), min=1)
+
+        # Return loss dictionary
+        loss_dict = {"ce_loss": loss.item(), "loss": loss}
+        return loss_dict
+
+    def decode_actions_with_fast(
+        self,
+        tokens: list[list[int]],
+        *,
+        time_horizon: int | None = None,
+        action_dim: int | None = None,
+        relaxed_decoding: bool = True,
+    ) -> np.array:
+        """
+        Adapt original decoding in FAST to always return actions instead of zeros.
+        """
+        self.time_horizon = (
+            time_horizon or self.fast_tokenizer.time_horizon or self.fast_tokenizer.called_time_horizon
+        )
+        self.action_dim = (
+            action_dim or self.fast_tokenizer.action_dim or self.fast_tokenizer.called_action_dim
+        )
+
+        # Cache the time horizon and action dimension for the next call
+        self.called_time_horizon = self.time_horizon
+        self.called_action_dim = self.action_dim
+
+        assert self.time_horizon is not None and self.action_dim is not None, (
+            "Tokenizer not initialized, call encode() once or pass in time_horizon and action_dim."
+        )
+
+        decoded_actions = []
+        for token in tokens:
+            try:
+                decoded_tokens = self.fast_tokenizer.bpe_tokenizer.decode(token)
+                decoded_dct_coeff = np.array(list(map(ord, decoded_tokens))) + self.fast_tokenizer.min_token
+                if relaxed_decoding:
+                    # Expected sequence length
+                    expected_seq_len = self.time_horizon * self.action_dim
+                    diff = expected_seq_len - decoded_dct_coeff.shape[0]
+                    # Apply truncation if too long
+                    if diff < 0:
+                        decoded_dct_coeff = decoded_dct_coeff[:expected_seq_len]  # Truncate on the right
+                    # Apply padding if too short
+                    elif diff > 0:
+                        decoded_dct_coeff = np.pad(
+                            decoded_dct_coeff, (0, diff), mode="constant", constant_values=0
+                        )
+
+                decoded_dct_coeff = decoded_dct_coeff.reshape(-1, self.action_dim)
+                assert decoded_dct_coeff.shape == (
+                    self.time_horizon,
+                    self.action_dim,
+                ), (
+                    f"Decoded DCT coefficients have shape {decoded_dct_coeff.shape}, expected ({self.time_horizon}, {self.action_dim})"
+                )
+            except Exception as e:
+                print(f"Error decoding tokens: {e}")
+                print(f"Tokens: {token}")
+                decoded_dct_coeff = np.zeros((self.time_horizon, self.action_dim))
+            decoded_actions.append(idct(decoded_dct_coeff / self.fast_tokenizer.scale, axis=0, norm="ortho"))
+        return np.stack(decoded_actions)
+
+    def extract_actions(self, tokens: torch.Tensor, action_horizon: int, action_dim: int) -> torch.Tensor:
+        """
+        Extracts actions from predicted output tokens using the FAST model.
+
+        Args:
+            tokens (torch.Tensor): The input tensor of tokenized outputs.
+            action_horizon (int): The number of timesteps for actions.
+            action_dim (int): The dimensionality of each action.
+
+        Returns:
+            torch.Tensor: The extracted actions as a tensor of shape (action_horizon, action_dim).
+        """
+        # Decode predicted output tokens
+        decoded_tokens = self.paligemma_tokenizer.batch_decode(tokens, skip_special_tokens=True)
+        cleaned_tokens = [
+            tokens_sequence.replace("Action:", "").replace(":", "").strip().split("|")[0].strip()
+            for tokens_sequence in decoded_tokens
+        ]
+        raw_action_tokens = [
+            self.processor.tokenizer.encode(sample_tokens, return_tensors="pt", padding=False)
+            for sample_tokens in cleaned_tokens
+        ]  # something like this should be robust #looks good
+        action_tokens = [
+            self._act_tokens_to_paligemma_tokens(raw_action_token) for raw_action_token in raw_action_tokens
+        ]
+        # returns the tensor of decoded actions per sample in a list
+        decoded_actions = [
+            torch.tensor(
+                self.decode_actions_with_fast(
+                    tok.tolist(),
+                    time_horizon=action_horizon,
+                    action_dim=action_dim,
+                    relaxed_decoding=self.config.relaxed_action_decoding,
+                ),
+                device=tokens.device,
+            ).squeeze(0)
+            for tok in action_tokens
+        ]
+
+        return torch.stack(
+            decoded_actions,
+            dim=0,
+        )
+
+    def generate_actions(self, batch: dict[str, Tensor]):
+        # TODO: keep like this or move to the policy .forward
+        images, img_masks = self.prepare_images(batch)
+
+        padded_outs = self.create_input_tokens(state=batch[OBS_ROBOT], lang_text=batch["task"], actions=None)
+        embs, pad_masks, att_masks2, targets, loss_mask, token_type_ids = self.embed_inputs(
+            images,
+            img_masks,
+            padded_outs["input_ids"],
+            padded_outs["padded_mask"],
+            padded_outs["attention_mask"],
+            padded_outs["loss_mask"],
+            padded_outs["token_type_ids"],
+            padding_side="left",
+        )
+        token_type_ids = token_type_ids.to(dtype=torch.int64)
+        prefix_position_ids = torch.cumsum(pad_masks, dim=1) - 1
+        output_tokens = self.pi0_paligemma.generate(
+            input_ids=None,
+            attention_mask=pad_masks,
+            position_ids=prefix_position_ids,
+            past_key_values=None,
+            inputs_embeds=embs,
+            use_cache=self.config.use_cache,
+            max_new_tokens=self.config.max_decoding_steps,
+            do_sample=False,
+            num_beams=1,
+            token_type_ids=token_type_ids,
+        )
+        actions = self.extract_actions(output_tokens, self.action_horizon, self.action_dim)
+        return actions
+
+    def embed_image(self, image: torch.Tensor):
+        return self.pi0_paligemma.get_image_features(image)
+
+    def embed_inputs(
+        self,
+        images,
+        img_masks,
+        tokens,
+        pad_mask,
+        ar_mask,
+        loss_mask,
+        token_type_ids,
+        padding_side: str = "right",
+    ):
+        # TODO: avoid list in python and torch.cat ; prefer pre-allocation with torch.empty
+        # images are a list of same size
+        # vectorizing everything!
+        device = images[0].device
+        image_embedding_dim = images[0].shape[-1]  # TODO should be from self.config
+        all_images = torch.stack(images, dim=1).to(device)
+        b, n, c, h, w = all_images.shape
+        all_images = all_images.view(b * n, c, h, w)
+        embedded = self.embed_image(all_images).to(device)
+        b_n, p, image_embedding_dim = embedded.shape  # Extract current dimensions
+        m = b_n // b  # Compute the number of images per sample dynamically
+
+        # Reshape dynamically
+        embedded = embedded.view(b, m, p, image_embedding_dim)
+        tokens_embs = self.embed_tokens(tokens.to(device))
+
+        img_masks = torch.stack(img_masks, dim=1).unsqueeze(-1).to(device)
+        num_img_emb = embedded.shape[2]
+        img_pad_masks = img_masks.repeat(1, 1, num_img_emb).view(b, -1)
+        img_att_masks = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
+
+        image_target_tokens = (
+            torch.ones((b, n, num_img_emb), dtype=torch.long, device=device) * self.pad_token_id
+        ).reshape(b, -1)
+        image_loss_mask = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
+
+        embedded = embedded.reshape(b, n * num_img_emb, image_embedding_dim)  # Shape: (B, N*P, D)
+
+        embs = torch.cat([embedded, tokens_embs], dim=1).to(device)
+        pad_masks = torch.cat([img_pad_masks, pad_mask.to(device)], dim=1)
+        att_masks = torch.cat([img_att_masks, ar_mask.to(device)], dim=1)
+        loss_masks = torch.cat([image_loss_mask, loss_mask.to(device)], dim=1)
+        targets = torch.cat([image_target_tokens, tokens.to(device)], dim=1)
+        token_type_ids = torch.cat([img_att_masks, token_type_ids.to(device)], dim=1)
+
+        # Shift pad tokens to the left (.generate()) or right (.train())
+        embs, att_masks, pad_masks, loss_masks, targets, token_type_ids = self.shift_padding_side(
+            embs, att_masks, pad_masks, loss_masks, targets, token_type_ids, padding_side=padding_side
+        )
+
+        targets = torch.where(targets == self.pad_token_id, self.ignore_index, targets)
+        return embs, pad_masks, att_masks, targets, loss_masks, token_type_ids
+
+
+def resize_with_pad(img, width, height, pad_value=0, interpolate_like_pi=True):
+    # assume no-op when width height fits already
+    if img.ndim != 4:
+        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
+
+    cur_height, cur_width = img.shape[2:]
+
+    ratio = max(cur_width / width, cur_height / height)
+    resized_height = int(cur_height / ratio)
+    resized_width = int(cur_width / ratio)
+
+    if interpolate_like_pi:
+        img = (img * 255.0).to(dtype=torch.uint8)
+        img = img.permute(0, 2, 3, 1)
+        original_device = img.device
+        img = img.to(device="cpu").numpy()
+        imgs = []
+        for sub_img in img:
+            sub_img = Image.fromarray(sub_img)
+            resized_img = sub_img.resize((resized_width, resized_height), resample=2)
+            resized_img = torch.from_numpy(np.array(resized_img))
+            imgs.append(resized_img)
+        img = torch.stack(imgs, dim=0)
+        img = img.permute(0, 3, 1, 2)
+        resized_img = img.to(device=original_device, dtype=torch.float32) / 255.0
+    else:
+        resized_img = F.interpolate(
+            img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+        )
+
+    pad_height = max(0, int(height - resized_height))
+    pad_width = max(0, int(width - resized_width))
+
+    # pad on left and top of image
+    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
+    return padded_img