save

Co-authored-by: Remi <re.cadene@gmail.com>

.dockerignore

@@ -0,0 +1,142 @@
+# 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/data
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# 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 +1,6 @@
+*.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 filter=lfs diff=lfs merge=lfs -text

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,54 @@
+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
+DATA_DIR=tests/data 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 +1,139 @@
+# 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 * * *"
+
+env:
+  PYTHON_VERSION: "3.10"
+#   CI_SLACK_CHANNEL: ${{ secrets.CI_DOCKER_CHANNEL }}
+
+jobs:
+  latest-cpu:
+    name: CPU
+    runs-on: ubuntu-latest
+    steps:
+      - name: Cleanup disk
+        run: |
+          sudo df -h
+          # sudo ls -l /usr/local/lib/
+          # sudo ls -l /usr/share/
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /usr/share/dotnet
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo df -h
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+
+      - 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 }}
+
+    #   - name: Post to a Slack channel
+    #     id: slack
+    #     #uses: slackapi/slack-github-action@v1.25.0
+    #     uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
+    #     with:
+    #       # Slack channel id, channel name, or user id to post message.
+    #       # See also: https://api.slack.com/methods/chat.postMessage#channels
+    #       channel-id: ${{ env.CI_SLACK_CHANNEL }}
+    #       # For posting a rich message using Block Kit
+    #       payload: |
+    #         {
+    #           "text": "lerobot-cpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}",
+    #           "blocks": [
+    #             {
+    #               "type": "section",
+    #               "text": {
+    #                 "type": "mrkdwn",
+    #                 "text": "lerobot-cpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}"
+    #               }
+    #             }
+    #           ]
+    #         }
+    #     env:
+    #       SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+  latest-cuda:
+    name: GPU
+    runs-on: ubuntu-latest
+    steps:
+      - name: Cleanup disk
+        run: |
+          sudo df -h
+          # sudo ls -l /usr/local/lib/
+          # sudo ls -l /usr/share/
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /usr/share/dotnet
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo df -h
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+
+      - 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 }}
+
+      # - name: Post to a Slack channel
+      #   id: slack
+      #   #uses: slackapi/slack-github-action@v1.25.0
+      #   uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
+      #   with:
+      #     # Slack channel id, channel name, or user id to post message.
+      #     # See also: https://api.slack.com/methods/chat.postMessage#channels
+      #     channel-id: ${{ env.CI_SLACK_CHANNEL }}
+      #     # For posting a rich message using Block Kit
+      #     payload: |
+      #       {
+      #         "text": "lerobot-gpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}",
+      #         "blocks": [
+      #           {
+      #             "type": "section",
+      #             "text": {
+      #               "type": "mrkdwn",
+      #               "text": "lerobot-gpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}"
+      #             }
+      #           }
+      #         ]
+      #       }
+      #   env:
+      #     SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}

.github/workflows/nightly-tests.yml

@@ -0,0 +1,79 @@
+# Inspired by
+# https://github.com/huggingface/peft/blob/main/.github/workflows/nightly.yml
+name: Nightly
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "0 2 * * *"
+
+env:
+  DATA_DIR: tests/data
+  # SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
+
+jobs:
+  run_all_tests_cpu:
+    name: CPU
+    strategy:
+      fail-fast: false
+    runs-on: ubuntu-latest
+    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
+        env:
+          DATA_DIR: tests/data
+        run: pytest -v --cov=./lerobot --disable-warnings tests
+
+      - name: Tests end-to-end
+        env:
+          DATA_DIR: tests/data
+        run: make test-end-to-end
+
+
+  run_all_tests_single_gpu:
+    name: GPU
+    strategy:
+      fail-fast: false
+    runs-on: [single-gpu, nvidia-gpu, t4, ci]
+    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
+        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 +1,56 @@
+name: Quality
+
+on:
+  workflow_dispatch:
+  workflow_call:
+  pull_request:
+    branches:
+      - main
+  push:
+    branches:
+      - main
+
+env:
+  PYTHON_VERSION: "3.10"
+
+jobs:
+  style:
+    name: Style
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - 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}" >> $GITHUB_ENV
+
+      - name: Install Ruff
+        run: python -m pip install "ruff==${{ env.RUFF_VERSION }}"
+
+      - name: Ruff check
+        run: ruff check
+
+      - name: Ruff format
+        run: ruff format --diff
+
+
+  poetry_check:
+    name: Poetry check
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Install poetry
+        run: pipx install poetry
+
+      - name: Poetry check
+        run: poetry check

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

@@ -0,0 +1,77 @@
+# Inspired by
+# https://github.com/huggingface/peft/blob/main/.github/workflows/test-docker-build.yml
+name: Test Dockerfiles
+
+on:
+  pull_request:
+    branches:
+      - main
+    paths:
+      # Run only when DockerFile files are modified
+      - "docker/**"
+
+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
+
+      - name: Get changed files
+        id: changed-files
+        uses: tj-actions/changed-files@v44
+        with:
+          files: docker/**
+          json: "true"
+
+      - name: Run step if only the files listed above change
+        if: steps.changed-files.outputs.any_changed == 'true'
+        id: set-matrix
+        env:
+          ALL_CHANGED_FILES: ${{ steps.changed-files.outputs.all_changed_files }}
+        run: |
+          echo "matrix=${{ steps.changed-files.outputs.all_changed_files}}" >> $GITHUB_OUTPUT
+
+
+  build_modified_dockerfiles:
+    name: Build modified Docker images
+    needs: get_changed_files
+    runs-on: ubuntu-latest
+    if: ${{ needs.get_changed_files.outputs.matrix }} != ''
+    strategy:
+      fail-fast: false
+      matrix:
+        docker-file: ${{ fromJson(needs.get_changed_files.outputs.matrix) }}
+    steps:
+      - name: Cleanup disk
+        run: |
+          sudo df -h
+          # sudo ls -l /usr/local/lib/
+          # sudo ls -l /usr/share/
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /usr/share/dotnet
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo df -h
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+
+      - 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 +1,128 @@
-name: Test
+name: Tests
 
 on:
   pull_request:
     branches:
       - main
-    types: [opened, synchronize, reopened, labeled]
+    paths:
+      - "lerobot/**"
+      - "tests/**"
+      - "examples/**"
+      - ".github/**"
+      - "poetry.lock"
   push:
     branches:
       - main
+    paths:
+      - "lerobot/**"
+      - "tests/**"
+      - "examples/**"
+      - ".github/**"
+      - "poetry.lock"
 
 jobs:
-  test:
-    if: |
-      ${{ github.event_name == 'pull_request' && contains(github.event.pull_request.labels.*.name, 'CI') }} ||
-      ${{ github.event_name == 'push' }}
+  pytest:
+    name: Pytest
     runs-on: ubuntu-latest
     env:
-      POETRY_VERSION: 1.8.1
+      DATA_DIR: tests/data
+      MUJOCO_GL: egl
     steps:
-      #----------------------------------------------
-      #       check-out repo and set-up python
-      #----------------------------------------------
-      - name: Check out repository
-        uses: actions/checkout@v4
-      - name: Set up python
-        id: setup-python
+      - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
+
+      - name: Install EGL
+        run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
+
+      - name: Install poetry
+        run: |
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+
+      - name: Set up Python 3.10
         uses: actions/setup-python@v5
         with:
-          python-version: '3.10'
-      #----------------------------------------------
-      #         install & configure poetry
-      #----------------------------------------------
-      - name: Load cached Poetry installation
-        id: restore-poetry-cache
-        uses: actions/cache/restore@v3
-        with:
-          path: ~/.local  # the path depends on the OS
-          key: poetry-${{ 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'
+          python-version: "3.10"
+          cache: "poetry"
+
+      - name: Install poetry dependencies
         run: |
-          poetry install --no-interaction --no-root
-          git clone https://github.com/real-stanford/diffusion_policy
-          cp -r diffusion_policy/diffusion_policy $(poetry env info -p)/lib/python3.10/site-packages/
-      - name: Save cached venv
-        if: |
-            steps.restore-dependencies-cache.outputs.cache-hit != 'true' &&
-            github.ref_name == 'main'
-        id: save-dependencies-cache
-        uses: actions/cache/save@v3
+          poetry install --all-extras
+
+      - name: Test with pytest
+        run: |
+          pytest tests -v --cov=./lerobot --durations=0 \
+            -W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
+            -W ignore::UserWarning:torch.utils.data.dataloader:558 \
+            -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
+            && rm -rf tests/outputs outputs
+
+
+  pytest-minimal:
+    name: Pytest (minimal install)
+    runs-on: ubuntu-latest
+    env:
+      DATA_DIR: tests/data
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
         with:
-          path: .venv
-          key: venv-${{ 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
+          lfs: true  # Ensure LFS files are pulled
+
+      - name: Install poetry
         run: |
-          source .venv/bin/activate
-          pytest tests
-      - name: Test train pusht end-to-end
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+
+      - name: Set up Python 3.10
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Install poetry dependencies
         run: |
-          source .venv/bin/activate
-          python lerobot/scripts/train.py \
-            hydra.job.name=pusht \
-            env=pusht \
-            wandb.enable=False \
-            offline_steps=1 \
-            online_steps=0 \
-            device=cpu
-      # TODO(rcadene, aliberts): Add end-to-end test of eval checkpoint post training
-      # - name: Test eval pusht end-to-end
-      #   run: |
-      #     source .venv/bin/activate
-      #     python lerobot/scripts/eval.py
-      #       hydra.job.name=pusht \
-      #       env=pusht \
-      #       wandb.enable=False \
-      #       eval_episodes=1 \
-      #       device=cpu
-      #----------------------------------------------
-      #              cleanup
-      #----------------------------------------------
-      - name: Cleanup
-        run: rm -rf diffusion_policy data
+          poetry install --extras "test"
+
+      - name: Test with pytest
+        run: |
+          pytest tests -v --cov=./lerobot --durations=0 \
+            -W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
+            -W ignore::UserWarning:torch.utils.data.dataloader:558 \
+            -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
+            && rm -rf tests/outputs outputs
+
+
+  end-to-end:
+    name: End-to-end
+    runs-on: ubuntu-latest
+    env:
+      DATA_DIR: tests/data
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
+
+      - name: Install EGL
+        run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
+
+      - name: Install poetry
+        run: |
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+
+      - name: Set up Python 3.10
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+          cache: "poetry"
+
+      - name: Install poetry dependencies
+        run: |
+          poetry install --all-extras
+
+      - name: Test end-to-end
+        run: |
+          make test-end-to-end \
+            && rm -rf outputs

.gitignore

@@ -1,6 +1,3 @@
-# Custom
-diffusion_policy
-
 # Logging
 logs
 tmp
@@ -9,11 +6,15 @@ data
 outputs
 .vscode
 rl
+.DS_Store
 
 # HPC
 nautilus/*.yaml
 *.key
 
+# Slurm
+sbatch*.sh
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]
@@ -54,6 +55,7 @@ pip-log.txt
 pip-delete-this-directory.txt
 
 # Unit test / coverage reports
+!tests/data
 htmlcov/
 .tox/
 .nox/

.pre-commit-config.yaml

@@ -1,9 +1,9 @@
-exclude: ^(data/|tests/|diffusion_policy/)
+exclude: ^(tests/data)
 default_language_version:
     python: python3.10
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.5.0
+    rev: v4.6.0
     hooks:
       - id: check-added-large-files
       - id: debug-statements
@@ -14,11 +14,11 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.15.1
+    rev: v3.15.2
     hooks:
     -   id: pyupgrade
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.2.2
+    rev: v0.4.3
     hooks:
       - id: ruff
         args: [--fix]

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 +1,275 @@
+# 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](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 use `poetry` instead of just `pip` to easily track our dependencies.
+   If you don't have it already, follow the [instructions](https://python-poetry.org/docs/#installation) to install it.
+
+   Set up a development environment with conda or miniconda:
+   ```bash
+   conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev
+   ```
+
+   To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
+   ```bash
+   poetry install --sync --extras "dev test"
+   ```
+
+   You can also install the project with all its dependencies (including environments):
+   ```bash
+   poetry install --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 install --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:
+   ```bash
+   poetry add some-package
+   ```
+
+   When making changes to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
+   ```bash
+   poetry lock --no-update
+   ```
+
+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 commited 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/data](tests/data)
+```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
+DATA_DIR="tests/data" 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 +1,144 @@
+.PHONY: tests
+
+PYTHON_PATH := $(shell which python)
+
+# If Poetry is installed, redefine PYTHON_PATH to use the Poetry-managed Python
+POETRY_CHECK := $(shell command -v poetry)
+ifneq ($(POETRY_CHECK),)
+    PYTHON_PATH := $(shell poetry run which python)
+endif
+
+export PATH := $(dir $(PYTHON_PATH)):$(PATH)
+
+
+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} test-act-ete-train
+	${MAKE} test-act-ete-eval
+	${MAKE} test-act-ete-train-amp
+	${MAKE} test-act-ete-eval-amp
+	${MAKE} test-diffusion-ete-train
+	${MAKE} test-diffusion-ete-eval
+	${MAKE} test-tdmpc-ete-train
+	${MAKE} test-tdmpc-ete-eval
+	${MAKE} test-default-ete-eval
+
+test-act-ete-train:
+	python lerobot/scripts/train.py \
+		policy=act \
+		policy.dim_model=64 \
+		env=aloha \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=cpu \
+		training.save_model=true \
+		training.save_freq=2 \
+		policy.n_action_steps=20 \
+		policy.chunk_size=20 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/act/
+
+test-act-ete-eval:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/act/checkpoints/000002 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=cpu \
+
+test-act-ete-train-amp:
+	python lerobot/scripts/train.py \
+		policy=act \
+		policy.dim_model=64 \
+		env=aloha \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=cpu \
+		training.save_model=true \
+		training.save_freq=2 \
+		policy.n_action_steps=20 \
+		policy.chunk_size=20 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/act/ \
+		use_amp=true
+
+test-act-ete-eval-amp:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/act/checkpoints/000002 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=cpu \
+		use_amp=true
+
+test-diffusion-ete-train:
+	python lerobot/scripts/train.py \
+		policy=diffusion \
+		policy.down_dims=\[64,128,256\] \
+		policy.diffusion_step_embed_dim=32 \
+		policy.num_inference_steps=10 \
+		env=pusht \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=cpu \
+		training.save_model=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/diffusion/
+
+test-diffusion-ete-eval:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/diffusion/checkpoints/000002 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=cpu \
+
+# TODO(alexander-soare): Restore online_steps to 2 when it is reinstated.
+test-tdmpc-ete-train:
+	python lerobot/scripts/train.py \
+		policy=tdmpc \
+		env=xarm \
+		env.task=XarmLift-v0 \
+		dataset_repo_id=lerobot/xarm_lift_medium \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=2 \
+		device=cpu \
+		training.save_model=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/tdmpc/
+
+test-tdmpc-ete-eval:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/tdmpc/checkpoints/000002 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=cpu \
+
+test-default-ete-eval:
+	python lerobot/scripts/eval.py \
+		--config lerobot/configs/default.yaml \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=cpu \

README.md

@@ -1,83 +1,236 @@
-# 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>
+
+<h3 align="center">
+    <p>State-of-the-art Machine Learning 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="http://remicadene.com/assets/gif/aloha_act.gif" width="100%" alt="ACT policy on ALOHA env"/></td>
+    <td><img src="http://remicadene.com/assets/gif/simxarm_tdmpc.gif" width="100%" alt="TDMPC policy on SimXArm env"/></td>
+    <td><img src="http://remicadene.com/assets/gif/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 Zaho, 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.
+
 
 ## Installation
 
-Create a virtual environment with python 3.10, e.g. using `conda`:
-```
-conda create -y -n lerobot python=3.10
-conda activate lerobot
+Download our source code:
+```bash
+git clone https://github.com/huggingface/lerobot.git && cd lerobot
 ```
 
-[Install `poetry`](https://python-poetry.org/docs/#installation) (if you don't have it already)
-```
-curl -sSL https://install.python-poetry.org | python -
+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 dependencies
-```
-poetry install
+Install 🤗 LeRobot:
+```bash
+pip install .
 ```
 
-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'
+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 ".[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
+├── lerobot
+|   ├── configs          # contains hydra yaml files with all options that you can override in the command line
+|   |   ├── default.yaml   # selected by default, it loads pusht environment and diffusion policy
+|   |   ├── env            # various sim environments and their datasets: aloha.yaml, pusht.yaml, xarm.yaml
+|   |   └── policy         # various policies: act.yaml, diffusion.yaml, tdmpc.yaml
+|   ├── 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
+|   |   └── 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
+|       ├── 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 download data from the Hugging Face hub.
+
+You can also locally visualize episodes from a dataset 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
+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.
+
+### 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
+    -p lerobot/diffusion_pusht \
+    eval.n_episodes=10 \
+    eval.batch_size=10
 ```
 
+Note: After training your own policy, you can re-evaluate the checkpoints with:
+```bash
+python lerobot/scripts/eval.py \
+    -p PATH/TO/TRAIN/OUTPUT/FOLDER
+```
 
-## TODO
+See `python lerobot/scripts/eval.py --help` for more instructions.
 
-- [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
+### Train your own policy
 
-## Profile
+Check out [example 3](./examples/3_train_policy.py) that illustrates how to start training a model.
 
-**Example**
+In general, you can use our training script to easily train any policy. Here is an example of training the ACT policy on trajectories collected by humans on the Aloha simulation environment for the insertion task:
+```bash
+python lerobot/scripts/train.py \
+    policy=act \
+    env=aloha \
+    env.task=AlohaInsertion-v0 \
+    dataset_repo_id=lerobot/aloha_sim_insertion_human \
+```
+
+The experiment directory is automatically generated and will show up in yellow in your terminal. It looks like `outputs/train/2024-05-05/20-21-12_aloha_act_default`. You can manually specify an experiment directory by adding this argument to the `train.py` python command:
+```bash
+    hydra.run.dir=your/new/experiment/dir
+```
+
+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:
+
+```bash
+    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:
+
+![](media/wandb.png)
+
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. 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.
+
+## 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 ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Then move your dataset folder in `data` directory (e.g. `data/aloha_static_pingpong_test`), and push your dataset to the hub with:
+```bash
+python lerobot/scripts/push_dataset_to_hub.py \
+--data-dir data \
+--dataset-id aloha_static_pingpong_test \
+--raw-format aloha_hdf5 \
+--community-id lerobot
+```
+
+See `python lerobot/scripts/push_dataset_to_hub.py --help` for more instructions.
+
+If your dataset format is not supported, implement your own in `lerobot/common/datasets/push_dataset_to_hub/${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 located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). It 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.
+- `config.yaml`: A consolidated Hydra training configuration containing the policy, environment, and dataset configs. The policy configuration should match `config.json` exactly. The environment config is useful for anyone who wants to evaluate your policy. The dataset config just serves as a paper trail for reproducibility.
+
+To upload these to the hub, run the following:
+```bash
+huggingface-cli upload ${hf_user}/${repo_name} path/to/checkpoint/dir
+```
+
+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 +249,17 @@ 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
 
-## Contribute
-
-**Style**
+If you want, you can cite this work with:
 ```
-# install if needed
-pre-commit install
-# apply style and linter checks before git commit
-pre-commit run -a
-```
-
-**Tests**
-```
-pytest -sx tests
+@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}
+}
 ```

docker/lerobot-cpu/Dockerfile

@@ -0,0 +1,31 @@
+# Configure image
+ARG PYTHON_VERSION=3.10
+
+FROM python:${PYTHON_VERSION}-slim
+ARG PYTHON_VERSION
+ARG DEBIAN_FRONTEND=noninteractive
+
+# Install apt dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Create virtual environment
+RUN ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python
+RUN python -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
+
+# Install LeRobot
+COPY . /lerobot
+WORKDIR /lerobot
+RUN pip install --upgrade --no-cache-dir pip
+RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht]" \
+    --extra-index-url https://download.pytorch.org/whl/cpu
+
+# Set EGL as the rendering backend for MuJoCo
+ENV MUJOCO_GL="egl"
+
+# Execute in bash shell rather than python
+CMD ["/bin/bash"]

docker/lerobot-gpu/Dockerfile

@@ -0,0 +1,33 @@
+FROM nvidia/cuda:12.4.1-base-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 ffmpeg \
+    htop atop nvtop \
+    sed gawk grep curl wget \
+    tcpdump sysstat screen \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
+    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Create virtual environment
+RUN ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python
+RUN python -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
+
+# Install LeRobot
+RUN git lfs install
+RUN git clone https://github.com/huggingface/lerobot.git
+WORKDIR /lerobot
+RUN pip install --upgrade --no-cache-dir pip
+RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht]"
+
+# Set EGL as the rendering backend for MuJoCo
+ENV MUJOCO_GL="egl"

examples/1_load_lerobot_dataset.py

@@ -0,0 +1,91 @@
+"""
+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:
+- Loading a dataset and accessing its properties.
+- Filtering data by episode number.
+- Converting tensor data for visualization.
+- Saving video files from dataset frames.
+- 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 pathlib import Path
+from pprint import pprint
+
+import imageio
+import torch
+
+import lerobot
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+print("List of available datasets:")
+pprint(lerobot.available_datasets)
+
+# Let's take one for this example
+repo_id = "lerobot/pusht"
+
+# You can easily load a dataset from a Hugging Face repository
+dataset = LeRobotDataset(repo_id)
+
+# LeRobotDataset is actually a thin wrapper around an underlying Hugging Face dataset
+# (see https://huggingface.co/docs/datasets/index for more information).
+print(dataset)
+print(dataset.hf_dataset)
+
+# And provides additional utilities for robotics and compatibility with Pytorch
+print(f"\naverage number of frames per episode: {dataset.num_samples / dataset.num_episodes:.3f}")
+print(f"frames per second used during data collection: {dataset.fps=}")
+print(f"keys to access images from cameras: {dataset.camera_keys=}\n")
+
+# Access frame indexes associated to 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()
+
+# LeRobot datasets actually subclass PyTorch datasets so you can do everything you know and love from working
+# with the latter, like iterating through the dataset. Here we grab all the image frames.
+frames = [dataset[idx]["observation.image"] for idx in range(from_idx, to_idx)]
+
+# Video frames are now float32 in range [0,1] channel first (c,h,w) to follow pytorch convention. To visualize
+# them, we convert to uint8 in range [0,255]
+frames = [(frame * 255).type(torch.uint8) for frame in frames]
+# and to channel last (h,w,c).
+frames = [frame.permute((1, 2, 0)).numpy() for frame in frames]
+
+# Finally, we save the frames to a mp4 video for visualization.
+Path("outputs/examples/1_load_lerobot_dataset").mkdir(parents=True, exist_ok=True)
+imageio.mimsave("outputs/examples/1_load_lerobot_dataset/episode_0.mp4", frames, fps=dataset.fps)
+
+# 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
+    "observation.image": [-1, -0.5, -0.20, 0],
+    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 20 ms, 10 ms, and current frame
+    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, -0.02, -0.01, 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)],
+}
+dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
+print(f"\n{dataset[0]['observation.image'].shape=}")  # (4,c,h,w)
+print(f"{dataset[0]['observation.state'].shape=}")  # (8,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['observation.image'].shape=}")  # (32,4,c,h,w)
+    print(f"{batch['observation.state'].shape=}")  # (32,8,c)
+    print(f"{batch['action'].shape=}")  # (32,64,c)
+    break

examples/2_evaluate_pretrained_policy.py

@@ -0,0 +1,112 @@
+"""
+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.
+"""
+
+from pathlib import Path
+
+import gym_pusht  # noqa: F401
+import gymnasium as gym
+import imageio
+import numpy
+import torch
+from huggingface_hub import snapshot_download
+
+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)
+
+device = torch.device("cuda")
+
+# Download the diffusion policy for pusht environment
+pretrained_policy_path = Path(snapshot_download("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)
+
+# 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,
+)
+
+# Reset the policy and environmens 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 +1,79 @@
+"""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
+from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+
+# Create a directory to store the training checkpoint.
+output_directory = Path("outputs/train/example_pusht_diffusion")
+output_directory.mkdir(parents=True, exist_ok=True)
+
+# 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
+device = torch.device("cuda")
+log_freq = 250
+
+# 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 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],
+}
+dataset = LeRobotDataset("lerobot/pusht", delta_timestamps=delta_timestamps)
+
+# Set up the the policy.
+# Policies are initialized with a configuration class, in this case `DiffusionConfig`.
+# For this example, no arguments need to be passed because the defaults are set up for PushT.
+# If you're doing something different, you will likely need to change at least some of the defaults.
+cfg = DiffusionConfig()
+policy = DiffusionPolicy(cfg, dataset_stats=dataset.stats)
+policy.train()
+policy.to(device)
+
+optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
+
+# Create dataloader for offline training.
+dataloader = torch.utils.data.DataLoader(
+    dataset,
+    num_workers=4,
+    batch_size=64,
+    shuffle=True,
+    pin_memory=device != torch.device("cpu"),
+    drop_last=True,
+)
+
+# Run training loop.
+step = 0
+done = False
+while not done:
+    for batch in dataloader:
+        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+        output_dict = policy.forward(batch)
+        loss = output_dict["loss"]
+        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)

examples/4_calculate_validation_loss.py

@@ -0,0 +1,90 @@
+"""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
+from pathlib import Path
+
+import torch
+from huggingface_hub import snapshot_download
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+
+device = torch.device("cuda")
+
+# Download the diffusion policy for pusht environment
+pretrained_policy_path = Path(snapshot_download("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 full dataset
+full_dataset = LeRobotDataset("lerobot/pusht", split="train")
+# - Calculate train and val subsets
+num_train_episodes = math.floor(full_dataset.num_episodes * 90 / 100)
+num_val_episodes = full_dataset.num_episodes - num_train_episodes
+print(f"Number of episodes in full dataset: {full_dataset.num_episodes}")
+print(f"Number of episodes in training dataset (90% subset): {num_train_episodes}")
+print(f"Number of episodes in validation dataset (10% subset): {num_val_episodes}")
+# - Get first frame index of the validation set
+first_val_frame_index = full_dataset.episode_data_index["from"][num_train_episodes].item()
+# - Load frames subset belonging to validation set using the `split` argument.
+#   It utilizes the `datasets` library's syntax for slicing datasets.
+#   For more information on the Slice API, please see:
+#   https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
+train_dataset = LeRobotDataset(
+    "lerobot/pusht", split=f"train[:{first_val_frame_index}]", delta_timestamps=delta_timestamps
+)
+val_dataset = LeRobotDataset(
+    "lerobot/pusht", split=f"train[{first_val_frame_index}:]", 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()}
+    output_dict = policy.forward(batch)
+
+    loss_cumsum += output_dict["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}")

lerobot/__init__.py

@@ -0,0 +1,132 @@
+#!/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)
+    ```
+
+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
+
+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",
+    ],
+    "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",
+]
+
+available_datasets = list(
+    itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets)
+)
+
+available_policies = [
+    "act",
+    "diffusion",
+    "tdmpc",
+]
+
+available_policies_per_env = {
+    "aloha": ["act"],
+    "pusht": ["diffusion"],
+    "xarm": ["tdmpc"],
+}
+
+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/datasets/_video_benchmark/README.md

@@ -0,0 +1,334 @@
+# 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?
+
+How to encode videos?
+- How much compression (`-crf`)? Low compression with `0`, normal compression with `20` or extreme with `56`?
+- What pixel format to use (`-pix_fmt`)? `yuv444p` or `yuv420p`?
+- How many key frames (`-g`)? A key frame every `10` frames?
+
+How to decode videos?
+- Which `decoder`? `torchvision`, `torchaudio`, `ffmpegio`, `decord`, or `nvc`?
+
+## Metrics
+
+**Percentage of data compression (higher is better)**
+`compression_factor` is the ratio of the memory space on disk taken by the original images to encode, to the memory space taken by the encoded video. For instance, `compression_factor=4` means that the video takes 4 times less memory space on disk compared to the original images.
+
+**Percentage of loading time (higher is better)**
+`load_time_factor` is the ratio of the time it takes to load original images at given timestamps, to the time it takes to decode the exact same frames from the video. Higher is better. For instance, `load_time_factor=0.5` means that decoding from video is 2 times slower than loading the original images.
+
+**Average L2 error per pixel (lower is better)**
+`avg_per_pixel_l2_error` is the average L2 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.
+
+**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.
+
+
+## Variables
+
+**Image content**
+We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an appartment, or in a factory, or outdoor, etc. Hence, we run this benchmark on two datasets: `pusht` (simulation) and `umi` (real-world outdoor).
+
+**Requested timestamps**
+In this benchmark, we focus on the loading time of random access, so we are not interested in sequentially loading all frames of a video like in a movie. However, the number of consecutive timestamps requested and their spacing can greatly affect the `load_time_factor`. In fact, it is expected to get faster loading time by decoding a large number of consecutive frames from a video, than to load the same data from individual images. To reflect our robotics use case, we consider a few settings:
+- `single_frame`: 1 frame,
+- `2_frames`: 2 consecutive frames (e.g. `[t, t + 1 / fps]`),
+- `2_frames_4_space`: 2 consecutive frames with 4 frames of spacing (e.g `[t, t + 4 / fps]`),
+
+**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.).
+
+
+## Results
+
+**`decoder`**
+| repo_id | decoder | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | <span style="color: #32CD32;">torchvision</span> | 0.166 | 0.0000119 |
+| lerobot/pusht | ffmpegio | 0.009 | 0.0001182 |
+| lerobot/pusht | torchaudio | 0.138 | 0.0000359 |
+| lerobot/umi_cup_in_the_wild | <span style="color: #32CD32;">torchvision</span> | 0.174 | 0.0000174 |
+| lerobot/umi_cup_in_the_wild | ffmpegio | 0.010 | 0.0000735 |
+| lerobot/umi_cup_in_the_wild | torchaudio | 0.154 | 0.0000340 |
+
+### `1_frame`
+
+**`pix_fmt`**
+| repo_id | pix_fmt | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | yuv420p | 3.788 | 0.224 | 0.0000760 |
+| lerobot/pusht | yuv444p | 3.646 | 0.185 | 0.0000443 |
+| lerobot/umi_cup_in_the_wild | yuv420p | 14.391 | 0.388 | 0.0000469 |
+| lerobot/umi_cup_in_the_wild | yuv444p | 14.932 | 0.329 | 0.0000397 |
+
+**`g`**
+| repo_id | g | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 1 | 2.543 | 0.204 | 0.0000556 |
+| lerobot/pusht | 2 | 3.646 | 0.182 | 0.0000443 |
+| lerobot/pusht | 3 | 4.431 | 0.174 | 0.0000450 |
+| lerobot/pusht | 4 | 5.103 | 0.163 | 0.0000448 |
+| lerobot/pusht | 5 | 5.625 | 0.163 | 0.0000436 |
+| lerobot/pusht | 6 | 5.974 | 0.155 | 0.0000427 |
+| lerobot/pusht | 10 | 6.814 | 0.130 | 0.0000410 |
+| lerobot/pusht | 15 | 7.431 | 0.105 | 0.0000406 |
+| lerobot/pusht | 20 | 7.662 | 0.097 | 0.0000400 |
+| lerobot/pusht | 40 | 8.163 | 0.061 | 0.0000405 |
+| lerobot/pusht | 100 | 8.761 | 0.039 | 0.0000422 |
+| lerobot/pusht | None | 8.909 | 0.024 | 0.0000431 |
+| lerobot/umi_cup_in_the_wild | 1 | 14.411 | 0.444 | 0.0000601 |
+| lerobot/umi_cup_in_the_wild | 2 | 14.932 | 0.345 | 0.0000397 |
+| lerobot/umi_cup_in_the_wild | 3 | 20.174 | 0.282 | 0.0000416 |
+| lerobot/umi_cup_in_the_wild | 4 | 24.889 | 0.271 | 0.0000415 |
+| lerobot/umi_cup_in_the_wild | 5 | 28.825 | 0.260 | 0.0000415 |
+| lerobot/umi_cup_in_the_wild | 6 | 31.635 | 0.249 | 0.0000415 |
+| lerobot/umi_cup_in_the_wild | 10 | 39.418 | 0.195 | 0.0000399 |
+| lerobot/umi_cup_in_the_wild | 15 | 44.577 | 0.169 | 0.0000394 |
+| lerobot/umi_cup_in_the_wild | 20 | 47.907 | 0.140 | 0.0000390 |
+| lerobot/umi_cup_in_the_wild | 40 | 52.554 | 0.096 | 0.0000384 |
+| lerobot/umi_cup_in_the_wild | 100 | 58.241 | 0.046 | 0.0000390 |
+| lerobot/umi_cup_in_the_wild | None | 60.530 | 0.022 | 0.0000400 |
+
+**`crf`**
+| repo_id | crf | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 0 | 1.699 | 0.175 | 0.0000035 |
+| lerobot/pusht | 5 | 1.409 | 0.181 | 0.0000080 |
+| lerobot/pusht | 10 | 1.842 | 0.172 | 0.0000123 |
+| lerobot/pusht | 15 | 2.322 | 0.187 | 0.0000211 |
+| lerobot/pusht | 20 | 3.050 | 0.181 | 0.0000346 |
+| lerobot/pusht | None | 3.646 | 0.189 | 0.0000443 |
+| lerobot/pusht | 25 | 3.969 | 0.186 | 0.0000521 |
+| lerobot/pusht | 30 | 5.687 | 0.184 | 0.0000850 |
+| lerobot/pusht | 40 | 10.818 | 0.193 | 0.0001726 |
+| lerobot/pusht | 50 | 18.185 | 0.183 | 0.0002606 |
+| lerobot/umi_cup_in_the_wild | 0 | 1.918 | 0.165 | 0.0000056 |
+| lerobot/umi_cup_in_the_wild | 5 | 3.207 | 0.171 | 0.0000111 |
+| lerobot/umi_cup_in_the_wild | 10 | 4.818 | 0.212 | 0.0000153 |
+| lerobot/umi_cup_in_the_wild | 15 | 7.329 | 0.261 | 0.0000218 |
+| lerobot/umi_cup_in_the_wild | 20 | 11.361 | 0.312 | 0.0000317 |
+| lerobot/umi_cup_in_the_wild | None | 14.932 | 0.339 | 0.0000397 |
+| lerobot/umi_cup_in_the_wild | 25 | 17.741 | 0.297 | 0.0000452 |
+| lerobot/umi_cup_in_the_wild | 30 | 27.983 | 0.406 | 0.0000629 |
+| lerobot/umi_cup_in_the_wild | 40 | 82.449 | 0.468 | 0.0001184 |
+| lerobot/umi_cup_in_the_wild | 50 | 186.145 | 0.515 | 0.0001879 |
+
+**best**
+| repo_id | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | 3.646 | 0.188 | 0.0000443 |
+| lerobot/umi_cup_in_the_wild | 14.932 | 0.339 | 0.0000397 |
+
+### `2_frames`
+
+**`pix_fmt`**
+| repo_id | pix_fmt | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | yuv420p | 3.788 | 0.314 | 0.0000799 |
+| lerobot/pusht | yuv444p | 3.646 | 0.303 | 0.0000496 |
+| lerobot/umi_cup_in_the_wild | yuv420p | 14.391 | 0.642 | 0.0000503 |
+| lerobot/umi_cup_in_the_wild | yuv444p | 14.932 | 0.529 | 0.0000436 |
+
+**`g`**
+| repo_id | g | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 1 | 2.543 | 0.308 | 0.0000599 |
+| lerobot/pusht | 2 | 3.646 | 0.279 | 0.0000496 |
+| lerobot/pusht | 3 | 4.431 | 0.259 | 0.0000498 |
+| lerobot/pusht | 4 | 5.103 | 0.243 | 0.0000501 |
+| lerobot/pusht | 5 | 5.625 | 0.235 | 0.0000492 |
+| lerobot/pusht | 6 | 5.974 | 0.230 | 0.0000481 |
+| lerobot/pusht | 10 | 6.814 | 0.194 | 0.0000468 |
+| lerobot/pusht | 15 | 7.431 | 0.152 | 0.0000460 |
+| lerobot/pusht | 20 | 7.662 | 0.151 | 0.0000455 |
+| lerobot/pusht | 40 | 8.163 | 0.095 | 0.0000454 |
+| lerobot/pusht | 100 | 8.761 | 0.062 | 0.0000472 |
+| lerobot/pusht | None | 8.909 | 0.037 | 0.0000479 |
+| lerobot/umi_cup_in_the_wild | 1 | 14.411 | 0.638 | 0.0000625 |
+| lerobot/umi_cup_in_the_wild | 2 | 14.932 | 0.537 | 0.0000436 |
+| lerobot/umi_cup_in_the_wild | 3 | 20.174 | 0.493 | 0.0000437 |
+| lerobot/umi_cup_in_the_wild | 4 | 24.889 | 0.458 | 0.0000446 |
+| lerobot/umi_cup_in_the_wild | 5 | 28.825 | 0.438 | 0.0000445 |
+| lerobot/umi_cup_in_the_wild | 6 | 31.635 | 0.424 | 0.0000444 |
+| lerobot/umi_cup_in_the_wild | 10 | 39.418 | 0.345 | 0.0000435 |
+| lerobot/umi_cup_in_the_wild | 15 | 44.577 | 0.313 | 0.0000417 |
+| lerobot/umi_cup_in_the_wild | 20 | 47.907 | 0.264 | 0.0000421 |
+| lerobot/umi_cup_in_the_wild | 40 | 52.554 | 0.185 | 0.0000414 |
+| lerobot/umi_cup_in_the_wild | 100 | 58.241 | 0.090 | 0.0000420 |
+| lerobot/umi_cup_in_the_wild | None | 60.530 | 0.042 | 0.0000424 |
+
+**`crf`**
+| repo_id | crf | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 0 | 1.699 | 0.302 | 0.0000097 |
+| lerobot/pusht | 5 | 1.409 | 0.287 | 0.0000142 |
+| lerobot/pusht | 10 | 1.842 | 0.283 | 0.0000184 |
+| lerobot/pusht | 15 | 2.322 | 0.305 | 0.0000268 |
+| lerobot/pusht | 20 | 3.050 | 0.285 | 0.0000402 |
+| lerobot/pusht | None | 3.646 | 0.285 | 0.0000496 |
+| lerobot/pusht | 25 | 3.969 | 0.293 | 0.0000572 |
+| lerobot/pusht | 30 | 5.687 | 0.293 | 0.0000893 |
+| lerobot/pusht | 40 | 10.818 | 0.319 | 0.0001762 |
+| lerobot/pusht | 50 | 18.185 | 0.304 | 0.0002626 |
+| lerobot/umi_cup_in_the_wild | 0 | 1.918 | 0.235 | 0.0000112 |
+| lerobot/umi_cup_in_the_wild | 5 | 3.207 | 0.261 | 0.0000166 |
+| lerobot/umi_cup_in_the_wild | 10 | 4.818 | 0.333 | 0.0000207 |
+| lerobot/umi_cup_in_the_wild | 15 | 7.329 | 0.406 | 0.0000267 |
+| lerobot/umi_cup_in_the_wild | 20 | 11.361 | 0.489 | 0.0000361 |
+| lerobot/umi_cup_in_the_wild | None | 14.932 | 0.537 | 0.0000436 |
+| lerobot/umi_cup_in_the_wild | 25 | 17.741 | 0.578 | 0.0000487 |
+| lerobot/umi_cup_in_the_wild | 30 | 27.983 | 0.453 | 0.0000655 |
+| lerobot/umi_cup_in_the_wild | 40 | 82.449 | 0.767 | 0.0001192 |
+| lerobot/umi_cup_in_the_wild | 50 | 186.145 | 0.816 | 0.0001881 |
+
+**best**
+| repo_id | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | 3.646 | 0.283 | 0.0000496 |
+| lerobot/umi_cup_in_the_wild | 14.932 | 0.543 | 0.0000436 |
+
+### `2_frames_4_space`
+
+**`pix_fmt`**
+| repo_id | pix_fmt | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | yuv420p | 3.788 | 0.257 | 0.0000855 |
+| lerobot/pusht | yuv444p | 3.646 | 0.261 | 0.0000556 |
+| lerobot/umi_cup_in_the_wild | yuv420p | 14.391 | 0.493 | 0.0000476 |
+| lerobot/umi_cup_in_the_wild | yuv444p | 14.932 | 0.371 | 0.0000404 |
+
+**`g`**
+| repo_id | g | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 1 | 2.543 | 0.226 | 0.0000670 |
+| lerobot/pusht | 2 | 3.646 | 0.222 | 0.0000556 |
+| lerobot/pusht | 3 | 4.431 | 0.217 | 0.0000567 |
+| lerobot/pusht | 4 | 5.103 | 0.204 | 0.0000555 |
+| lerobot/pusht | 5 | 5.625 | 0.179 | 0.0000556 |
+| lerobot/pusht | 6 | 5.974 | 0.188 | 0.0000544 |
+| lerobot/pusht | 10 | 6.814 | 0.160 | 0.0000531 |
+| lerobot/pusht | 15 | 7.431 | 0.150 | 0.0000521 |
+| lerobot/pusht | 20 | 7.662 | 0.123 | 0.0000519 |
+| lerobot/pusht | 40 | 8.163 | 0.092 | 0.0000519 |
+| lerobot/pusht | 100 | 8.761 | 0.053 | 0.0000533 |
+| lerobot/pusht | None | 8.909 | 0.034 | 0.0000541 |
+| lerobot/umi_cup_in_the_wild | 1 | 14.411 | 0.409 | 0.0000607 |
+| lerobot/umi_cup_in_the_wild | 2 | 14.932 | 0.381 | 0.0000404 |
+| lerobot/umi_cup_in_the_wild | 3 | 20.174 | 0.355 | 0.0000418 |
+| lerobot/umi_cup_in_the_wild | 4 | 24.889 | 0.346 | 0.0000425 |
+| lerobot/umi_cup_in_the_wild | 5 | 28.825 | 0.354 | 0.0000419 |
+| lerobot/umi_cup_in_the_wild | 6 | 31.635 | 0.336 | 0.0000419 |
+| lerobot/umi_cup_in_the_wild | 10 | 39.418 | 0.314 | 0.0000402 |
+| lerobot/umi_cup_in_the_wild | 15 | 44.577 | 0.269 | 0.0000397 |
+| lerobot/umi_cup_in_the_wild | 20 | 47.907 | 0.246 | 0.0000395 |
+| lerobot/umi_cup_in_the_wild | 40 | 52.554 | 0.171 | 0.0000390 |
+| lerobot/umi_cup_in_the_wild | 100 | 58.241 | 0.091 | 0.0000399 |
+| lerobot/umi_cup_in_the_wild | None | 60.530 | 0.043 | 0.0000409 |
+
+**`crf`**
+| repo_id | crf | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 0 | 1.699 | 0.212 | 0.0000193 |
+| lerobot/pusht | 5 | 1.409 | 0.211 | 0.0000232 |
+| lerobot/pusht | 10 | 1.842 | 0.199 | 0.0000270 |
+| lerobot/pusht | 15 | 2.322 | 0.198 | 0.0000347 |
+| lerobot/pusht | 20 | 3.050 | 0.211 | 0.0000469 |
+| lerobot/pusht | None | 3.646 | 0.206 | 0.0000556 |
+| lerobot/pusht | 25 | 3.969 | 0.210 | 0.0000626 |
+| lerobot/pusht | 30 | 5.687 | 0.223 | 0.0000927 |
+| lerobot/pusht | 40 | 10.818 | 0.227 | 0.0001763 |
+| lerobot/pusht | 50 | 18.185 | 0.223 | 0.0002625 |
+| lerobot/umi_cup_in_the_wild | 0 | 1.918 | 0.147 | 0.0000071 |
+| lerobot/umi_cup_in_the_wild | 5 | 3.207 | 0.182 | 0.0000125 |
+| lerobot/umi_cup_in_the_wild | 10 | 4.818 | 0.222 | 0.0000166 |
+| lerobot/umi_cup_in_the_wild | 15 | 7.329 | 0.270 | 0.0000229 |
+| lerobot/umi_cup_in_the_wild | 20 | 11.361 | 0.325 | 0.0000326 |
+| lerobot/umi_cup_in_the_wild | None | 14.932 | 0.362 | 0.0000404 |
+| lerobot/umi_cup_in_the_wild | 25 | 17.741 | 0.390 | 0.0000459 |
+| lerobot/umi_cup_in_the_wild | 30 | 27.983 | 0.437 | 0.0000633 |
+| lerobot/umi_cup_in_the_wild | 40 | 82.449 | 0.499 | 0.0001186 |
+| lerobot/umi_cup_in_the_wild | 50 | 186.145 | 0.564 | 0.0001879 |
+
+**best**
+| repo_id | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | 3.646 | 0.224 | 0.0000556 |
+| lerobot/umi_cup_in_the_wild | 14.932 | 0.368 | 0.0000404 |
+
+### `6_frames`
+
+**`pix_fmt`**
+| repo_id | pix_fmt | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | yuv420p | 3.788 | 0.660 | 0.0000839 |
+| lerobot/pusht | yuv444p | 3.646 | 0.546 | 0.0000542 |
+| lerobot/umi_cup_in_the_wild | yuv420p | 14.391 | 1.225 | 0.0000497 |
+| lerobot/umi_cup_in_the_wild | yuv444p | 14.932 | 0.908 | 0.0000428 |
+
+**`g`**
+| repo_id | g | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 1 | 2.543 | 0.552 | 0.0000646 |
+| lerobot/pusht | 2 | 3.646 | 0.534 | 0.0000542 |
+| lerobot/pusht | 3 | 4.431 | 0.563 | 0.0000546 |
+| lerobot/pusht | 4 | 5.103 | 0.537 | 0.0000545 |
+| lerobot/pusht | 5 | 5.625 | 0.477 | 0.0000532 |
+| lerobot/pusht | 6 | 5.974 | 0.515 | 0.0000530 |
+| lerobot/pusht | 10 | 6.814 | 0.410 | 0.0000512 |
+| lerobot/pusht | 15 | 7.431 | 0.405 | 0.0000503 |
+| lerobot/pusht | 20 | 7.662 | 0.345 | 0.0000500 |
+| lerobot/pusht | 40 | 8.163 | 0.247 | 0.0000496 |
+| lerobot/pusht | 100 | 8.761 | 0.147 | 0.0000510 |
+| lerobot/pusht | None | 8.909 | 0.100 | 0.0000519 |
+| lerobot/umi_cup_in_the_wild | 1 | 14.411 | 0.997 | 0.0000620 |
+| lerobot/umi_cup_in_the_wild | 2 | 14.932 | 0.911 | 0.0000428 |
+| lerobot/umi_cup_in_the_wild | 3 | 20.174 | 0.869 | 0.0000433 |
+| lerobot/umi_cup_in_the_wild | 4 | 24.889 | 0.874 | 0.0000438 |
+| lerobot/umi_cup_in_the_wild | 5 | 28.825 | 0.864 | 0.0000439 |
+| lerobot/umi_cup_in_the_wild | 6 | 31.635 | 0.834 | 0.0000440 |
+| lerobot/umi_cup_in_the_wild | 10 | 39.418 | 0.781 | 0.0000421 |
+| lerobot/umi_cup_in_the_wild | 15 | 44.577 | 0.679 | 0.0000411 |
+| lerobot/umi_cup_in_the_wild | 20 | 47.907 | 0.652 | 0.0000410 |
+| lerobot/umi_cup_in_the_wild | 40 | 52.554 | 0.465 | 0.0000404 |
+| lerobot/umi_cup_in_the_wild | 100 | 58.241 | 0.245 | 0.0000413 |
+| lerobot/umi_cup_in_the_wild | None | 60.530 | 0.116 | 0.0000417 |
+
+**`crf`**
+| repo_id | crf | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 0 | 1.699 | 0.534 | 0.0000163 |
+| lerobot/pusht | 5 | 1.409 | 0.524 | 0.0000205 |
+| lerobot/pusht | 10 | 1.842 | 0.510 | 0.0000245 |
+| lerobot/pusht | 15 | 2.322 | 0.512 | 0.0000324 |
+| lerobot/pusht | 20 | 3.050 | 0.508 | 0.0000452 |
+| lerobot/pusht | None | 3.646 | 0.518 | 0.0000542 |
+| lerobot/pusht | 25 | 3.969 | 0.534 | 0.0000616 |
+| lerobot/pusht | 30 | 5.687 | 0.530 | 0.0000927 |
+| lerobot/pusht | 40 | 10.818 | 0.552 | 0.0001777 |
+| lerobot/pusht | 50 | 18.185 | 0.564 | 0.0002644 |
+| lerobot/umi_cup_in_the_wild | 0 | 1.918 | 0.401 | 0.0000101 |
+| lerobot/umi_cup_in_the_wild | 5 | 3.207 | 0.499 | 0.0000156 |
+| lerobot/umi_cup_in_the_wild | 10 | 4.818 | 0.599 | 0.0000197 |
+| lerobot/umi_cup_in_the_wild | 15 | 7.329 | 0.704 | 0.0000258 |
+| lerobot/umi_cup_in_the_wild | 20 | 11.361 | 0.834 | 0.0000352 |
+| lerobot/umi_cup_in_the_wild | None | 14.932 | 0.925 | 0.0000428 |
+| lerobot/umi_cup_in_the_wild | 25 | 17.741 | 0.978 | 0.0000480 |
+| lerobot/umi_cup_in_the_wild | 30 | 27.983 | 1.088 | 0.0000648 |
+| lerobot/umi_cup_in_the_wild | 40 | 82.449 | 1.324 | 0.0001190 |
+| lerobot/umi_cup_in_the_wild | 50 | 186.145 | 1.436 | 0.0001880 |
+
+**best**
+| repo_id | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | 3.646 | 0.546 | 0.0000542 |
+| lerobot/umi_cup_in_the_wild | 14.932 | 0.934 | 0.0000428 |

lerobot/common/datasets/_video_benchmark/run_video_benchmark.py

@@ -0,0 +1,372 @@
+#!/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 json
+import random
+import shutil
+import subprocess
+import time
+from pathlib import Path
+
+import einops
+import numpy
+import PIL
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.video_utils import (
+    decode_video_frames_torchvision,
+)
+
+
+def get_directory_size(directory):
+    total_size = 0
+    # Iterate over all files and subdirectories recursively
+    for item in directory.rglob("*"):
+        if item.is_file():
+            # Add the file size to the total
+            total_size += item.stat().st_size
+    return total_size
+
+
+def run_video_benchmark(
+    output_dir,
+    cfg,
+    timestamps_mode,
+    seed=1337,
+):
+    output_dir = Path(output_dir)
+    if output_dir.exists():
+        shutil.rmtree(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    repo_id = cfg["repo_id"]
+
+    # TODO(rcadene): rewrite with hardcoding of original images and episodes
+    dataset = LeRobotDataset(repo_id)
+
+    # Get fps
+    fps = dataset.fps
+
+    # we only load first episode
+    ep_num_images = dataset.episode_data_index["to"][0].item()
+
+    # Save/Load image directory for the first episode
+    imgs_dir = Path(f"tmp/data/images/{repo_id}/observation.image_episode_000000")
+    if not imgs_dir.exists():
+        imgs_dir.mkdir(parents=True, exist_ok=True)
+        hf_dataset = dataset.hf_dataset.with_format(None)
+        imgs_dataset = hf_dataset.select_columns("observation.image")
+
+        for i, item in enumerate(imgs_dataset):
+            img = item["observation.image"]
+            img.save(str(imgs_dir / f"frame_{i:06d}.png"), quality=100)
+
+            if i >= ep_num_images - 1:
+                break
+
+    sum_original_frames_size_bytes = get_directory_size(imgs_dir)
+
+    # Encode images into video
+    video_path = output_dir / "episode_0.mp4"
+
+    g = cfg.get("g")
+    crf = cfg.get("crf")
+    pix_fmt = cfg["pix_fmt"]
+
+    cmd = f"ffmpeg -r {fps} "
+    cmd += "-f image2 "
+    cmd += "-loglevel error "
+    cmd += f"-i {str(imgs_dir / 'frame_%06d.png')} "
+    cmd += "-vcodec libx264 "
+    if g is not None:
+        cmd += f"-g {g} "  # ensures at least 1 keyframe every 10 frames
+    # cmd += "-keyint_min 10 " set a minimum of 10 frames between 2 key frames
+    # cmd += "-sc_threshold 0 " disable scene change detection to lower the number of key frames
+    if crf is not None:
+        cmd += f"-crf {crf} "
+    cmd += f"-pix_fmt {pix_fmt} "
+    cmd += f"{str(video_path)}"
+    subprocess.run(cmd.split(" "), check=True)
+
+    video_size_bytes = video_path.stat().st_size
+
+    # Set decoder
+
+    decoder = cfg["decoder"]
+    decoder_kwgs = cfg["decoder_kwgs"]
+    device = cfg["device"]
+
+    if decoder == "torchvision":
+        decode_frames_fn = decode_video_frames_torchvision
+    else:
+        raise ValueError(decoder)
+
+    # Estimate average loading time
+
+    def load_original_frames(imgs_dir, timestamps):
+        frames = []
+        for ts in timestamps:
+            idx = int(ts * fps)
+            frame = PIL.Image.open(imgs_dir / f"frame_{idx:06d}.png")
+            frame = torch.from_numpy(numpy.array(frame))
+            frame = frame.type(torch.float32) / 255
+            frame = einops.rearrange(frame, "h w c -> c h w")
+            frames.append(frame)
+        return frames
+
+    list_avg_load_time = []
+    list_avg_load_time_from_images = []
+    per_pixel_l2_errors = []
+
+    random.seed(seed)
+
+    for t in range(50):
+        # test loading 2 frames that are 4 frames appart, which might be a common setting
+        ts = random.randint(fps, ep_num_images - fps) / fps
+
+        if timestamps_mode == "1_frame":
+            timestamps = [ts]
+        elif timestamps_mode == "2_frames":
+            timestamps = [ts - 1 / fps, ts]
+        elif timestamps_mode == "2_frames_4_space":
+            timestamps = [ts - 4 / fps, ts]
+        elif timestamps_mode == "6_frames":
+            timestamps = [ts - i / fps for i in range(6)][::-1]
+        else:
+            raise ValueError(timestamps_mode)
+
+        num_frames = len(timestamps)
+
+        start_time_s = time.monotonic()
+        frames = decode_frames_fn(
+            video_path, timestamps=timestamps, tolerance_s=1e-4, device=device, **decoder_kwgs
+        )
+        avg_load_time = (time.monotonic() - start_time_s) / num_frames
+        list_avg_load_time.append(avg_load_time)
+
+        start_time_s = time.monotonic()
+        original_frames = load_original_frames(imgs_dir, timestamps)
+        avg_load_time_from_images = (time.monotonic() - start_time_s) / num_frames
+        list_avg_load_time_from_images.append(avg_load_time_from_images)
+
+        # Estimate average L2 error between original frames and decoded frames
+        for i, ts in enumerate(timestamps):
+            # are_close = torch.allclose(frames[i], original_frames[i], atol=0.02)
+            num_pixels = original_frames[i].numel()
+            per_pixel_l2_error = torch.norm(frames[i] - original_frames[i], p=2).item() / num_pixels
+
+            # save decoded frames
+            if t == 0:
+                frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
+                PIL.Image.fromarray(frame_hwc).save(output_dir / f"frame_{i:06d}.png")
+
+            # save original_frames
+            idx = int(ts * fps)
+            if t == 0:
+                original_frame = PIL.Image.open(imgs_dir / f"frame_{idx:06d}.png")
+                original_frame.save(output_dir / f"original_frame_{i:06d}.png")
+
+            per_pixel_l2_errors.append(per_pixel_l2_error)
+
+    avg_load_time = float(numpy.array(list_avg_load_time).mean())
+    avg_load_time_from_images = float(numpy.array(list_avg_load_time_from_images).mean())
+    avg_per_pixel_l2_error = float(numpy.array(per_pixel_l2_errors).mean())
+
+    # Save benchmark info
+
+    info = {
+        "sum_original_frames_size_bytes": sum_original_frames_size_bytes,
+        "video_size_bytes": video_size_bytes,
+        "avg_load_time_from_images": avg_load_time_from_images,
+        "avg_load_time": avg_load_time,
+        "compression_factor": sum_original_frames_size_bytes / video_size_bytes,
+        "load_time_factor": avg_load_time_from_images / avg_load_time,
+        "avg_per_pixel_l2_error": avg_per_pixel_l2_error,
+    }
+
+    with open(output_dir / "info.json", "w") as f:
+        json.dump(info, f)
+
+    return info
+
+
+def display_markdown_table(headers, rows):
+    for i, row in enumerate(rows):
+        new_row = []
+        for col in row:
+            if col is None:
+                new_col = "None"
+            elif isinstance(col, float):
+                new_col = f"{col:.3f}"
+                if new_col == "0.000":
+                    new_col = f"{col:.7f}"
+            elif isinstance(col, int):
+                new_col = f"{col}"
+            else:
+                new_col = col
+            new_row.append(new_col)
+        rows[i] = new_row
+
+    header_line = "| " + " | ".join(headers) + " |"
+    separator_line = "| " + " | ".join(["---" for _ in headers]) + " |"
+    body_lines = ["| " + " | ".join(row) + " |" for row in rows]
+    markdown_table = "\n".join([header_line, separator_line] + body_lines)
+    print(markdown_table)
+    print()
+
+
+def load_info(out_dir):
+    with open(out_dir / "info.json") as f:
+        info = json.load(f)
+    return info
+
+
+def main():
+    out_dir = Path("tmp/run_video_benchmark")
+    dry_run = False
+    repo_ids = ["lerobot/pusht", "lerobot/umi_cup_in_the_wild"]
+    timestamps_modes = [
+        "1_frame",
+        "2_frames",
+        "2_frames_4_space",
+        "6_frames",
+    ]
+    for timestamps_mode in timestamps_modes:
+        bench_dir = out_dir / timestamps_mode
+
+        print(f"### `{timestamps_mode}`")
+        print()
+
+        print("**`pix_fmt`**")
+        headers = ["repo_id", "pix_fmt", "compression_factor", "load_time_factor", "avg_per_pixel_l2_error"]
+        rows = []
+        for repo_id in repo_ids:
+            for pix_fmt in ["yuv420p", "yuv444p"]:
+                cfg = {
+                    "repo_id": repo_id,
+                    # video encoding
+                    "g": 2,
+                    "crf": None,
+                    "pix_fmt": pix_fmt,
+                    # video decoding
+                    "device": "cpu",
+                    "decoder": "torchvision",
+                    "decoder_kwgs": {},
+                }
+                if not dry_run:
+                    run_video_benchmark(bench_dir / repo_id / f"torchvision_{pix_fmt}", cfg, timestamps_mode)
+                info = load_info(bench_dir / repo_id / f"torchvision_{pix_fmt}")
+                rows.append(
+                    [
+                        repo_id,
+                        pix_fmt,
+                        info["compression_factor"],
+                        info["load_time_factor"],
+                        info["avg_per_pixel_l2_error"],
+                    ]
+                )
+        display_markdown_table(headers, rows)
+
+        print("**`g`**")
+        headers = ["repo_id", "g", "compression_factor", "load_time_factor", "avg_per_pixel_l2_error"]
+        rows = []
+        for repo_id in repo_ids:
+            for g in [1, 2, 3, 4, 5, 6, 10, 15, 20, 40, 100, None]:
+                cfg = {
+                    "repo_id": repo_id,
+                    # video encoding
+                    "g": g,
+                    "pix_fmt": "yuv444p",
+                    # video decoding
+                    "device": "cpu",
+                    "decoder": "torchvision",
+                    "decoder_kwgs": {},
+                }
+                if not dry_run:
+                    run_video_benchmark(bench_dir / repo_id / f"torchvision_g_{g}", cfg, timestamps_mode)
+                info = load_info(bench_dir / repo_id / f"torchvision_g_{g}")
+                rows.append(
+                    [
+                        repo_id,
+                        g,
+                        info["compression_factor"],
+                        info["load_time_factor"],
+                        info["avg_per_pixel_l2_error"],
+                    ]
+                )
+        display_markdown_table(headers, rows)
+
+        print("**`crf`**")
+        headers = ["repo_id", "crf", "compression_factor", "load_time_factor", "avg_per_pixel_l2_error"]
+        rows = []
+        for repo_id in repo_ids:
+            for crf in [0, 5, 10, 15, 20, None, 25, 30, 40, 50]:
+                cfg = {
+                    "repo_id": repo_id,
+                    # video encoding
+                    "g": 2,
+                    "crf": crf,
+                    "pix_fmt": "yuv444p",
+                    # video decoding
+                    "device": "cpu",
+                    "decoder": "torchvision",
+                    "decoder_kwgs": {},
+                }
+                if not dry_run:
+                    run_video_benchmark(bench_dir / repo_id / f"torchvision_crf_{crf}", cfg, timestamps_mode)
+                info = load_info(bench_dir / repo_id / f"torchvision_crf_{crf}")
+                rows.append(
+                    [
+                        repo_id,
+                        crf,
+                        info["compression_factor"],
+                        info["load_time_factor"],
+                        info["avg_per_pixel_l2_error"],
+                    ]
+                )
+        display_markdown_table(headers, rows)
+
+        print("**best**")
+        headers = ["repo_id", "compression_factor", "load_time_factor", "avg_per_pixel_l2_error"]
+        rows = []
+        for repo_id in repo_ids:
+            cfg = {
+                "repo_id": repo_id,
+                # video encoding
+                "g": 2,
+                "crf": None,
+                "pix_fmt": "yuv444p",
+                # video decoding
+                "device": "cpu",
+                "decoder": "torchvision",
+                "decoder_kwgs": {},
+            }
+            if not dry_run:
+                run_video_benchmark(bench_dir / repo_id / "torchvision_best", cfg, timestamps_mode)
+            info = load_info(bench_dir / repo_id / "torchvision_best")
+            rows.append(
+                [
+                    repo_id,
+                    info["compression_factor"],
+                    info["load_time_factor"],
+                    info["avg_per_pixel_l2_error"],
+                ]
+            )
+        display_markdown_table(headers, rows)
+
+
+if __name__ == "__main__":
+    main()

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/factory.py

@@ -1,114 +1,55 @@
+#!/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
 
 import torch
-from torchrl.data.replay_buffers import PrioritizedSliceSampler, SliceSampler
+from omegaconf import OmegaConf
 
-from lerobot.common.envs.transforms import NormalizeTransform
-
-DATA_DIR = Path(os.environ.get("DATA_DIR", "data"))
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
 
-def make_offline_buffer(
-    cfg, overwrite_sampler=None, normalize=True, overwrite_batch_size=None, overwrite_prefetch=None
+def make_dataset(
+    cfg,
+    split="train",
 ):
-    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
+    if cfg.env.name not in cfg.dataset_repo_id:
+        logging.warning(
+            f"There might be a mismatch between your training dataset ({cfg.dataset_repo_id=}) and your "
+            f"environment ({cfg.env.name=})."
+        )
 
-    if overwrite_batch_size is not None:
-        batch_size = overwrite_batch_size
+    delta_timestamps = cfg.training.get("delta_timestamps")
+    if delta_timestamps is not None:
+        for key in delta_timestamps:
+            if isinstance(delta_timestamps[key], str):
+                delta_timestamps[key] = eval(delta_timestamps[key])
 
-    if overwrite_prefetch is not None:
-        prefetch = overwrite_prefetch
+    # TODO(rcadene): add data augmentations
 
-    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 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
-
-    if cfg.env.name == "simxarm":
-        from lerobot.common.datasets.simxarm import SimxarmExperienceReplay
-
-        clsfunc = SimxarmExperienceReplay
-        dataset_id = f"xarm_{cfg.env.task}_medium"
-
-    elif cfg.env.name == "pusht":
-        from lerobot.common.datasets.pusht import PushtExperienceReplay
-
-        clsfunc = PushtExperienceReplay
-        dataset_id = "pusht"
-
-    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,
+    dataset = LeRobotDataset(
+        cfg.dataset_repo_id,
+        split=split,
+        delta_timestamps=delta_timestamps,
     )
 
-    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 cfg.get("override_dataset_stats"):
+        for key, stats_dict in cfg.override_dataset_stats.items():
+            for stats_type, listconfig in stats_dict.items():
+                # example of stats_type: min, max, mean, std
+                stats = OmegaConf.to_container(listconfig, resolve=True)
+                dataset.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
-        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.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/lerobot_dataset.py

@@ -0,0 +1,200 @@
+#!/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 os
+from pathlib import Path
+
+import datasets
+import torch
+
+from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
+    load_episode_data_index,
+    load_hf_dataset,
+    load_info,
+    load_previous_and_future_frames,
+    load_stats,
+    load_videos,
+    reset_episode_index,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, load_from_videos
+
+DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
+CODEBASE_VERSION = "v1.4"
+
+
+class LeRobotDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        repo_id: str,
+        version: str | None = CODEBASE_VERSION,
+        root: Path | None = DATA_DIR,
+        split: str = "train",
+        transform: callable = None,
+        delta_timestamps: dict[list[float]] | None = None,
+    ):
+        super().__init__()
+        self.repo_id = repo_id
+        self.version = version
+        self.root = root
+        self.split = split
+        self.transform = transform
+        self.delta_timestamps = delta_timestamps
+        # load data from hub or locally when root is provided
+        # TODO(rcadene, aliberts): implement faster transfer
+        # https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads
+        self.hf_dataset = load_hf_dataset(repo_id, version, root, split)
+        if split == "train":
+            self.episode_data_index = load_episode_data_index(repo_id, version, root)
+        else:
+            self.episode_data_index = calculate_episode_data_index(self.hf_dataset)
+            self.hf_dataset = reset_episode_index(self.hf_dataset)
+        self.stats = load_stats(repo_id, version, root)
+        self.info = load_info(repo_id, version, root)
+        if self.video:
+            self.videos_dir = load_videos(repo_id, version, root)
+
+    @property
+    def fps(self) -> int:
+        """Frames per second used during data collection."""
+        return self.info["fps"]
+
+    @property
+    def video(self) -> bool:
+        """Returns True if this dataset loads video frames from mp4 files.
+        Returns False if it only loads images from png files.
+        """
+        return self.info.get("video", False)
+
+    @property
+    def features(self) -> datasets.Features:
+        return self.hf_dataset.features
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """Keys to access image and video stream from cameras."""
+        keys = []
+        for key, feats in self.hf_dataset.features.items():
+            if isinstance(feats, (datasets.Image, VideoFrame)):
+                keys.append(key)
+        return keys
+
+    @property
+    def video_frame_keys(self) -> list[str]:
+        """Keys to access video frames that requires to be decoded into images.
+
+        Note: It is empty if the dataset contains images only,
+        or equal to `self.cameras` if the dataset contains videos only,
+        or can even be a subset of `self.cameras` in a case of a mixed image/video dataset.
+        """
+        video_frame_keys = []
+        for key, feats in self.hf_dataset.features.items():
+            if isinstance(feats, VideoFrame):
+                video_frame_keys.append(key)
+        return video_frame_keys
+
+    @property
+    def num_samples(self) -> int:
+        """Number of samples/frames."""
+        return len(self.hf_dataset)
+
+    @property
+    def num_episodes(self) -> int:
+        """Number of episodes."""
+        return len(self.hf_dataset.unique("episode_index"))
+
+    @property
+    def tolerance_s(self) -> float:
+        """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 or when loading video frames from mp4 files.
+        """
+        # 1e-4 to account for possible numerical error
+        return 1 / self.fps - 1e-4
+
+    def __len__(self):
+        return self.num_samples
+
+    def __getitem__(self, idx):
+        item = self.hf_dataset[idx]
+
+        if self.delta_timestamps is not None:
+            item = load_previous_and_future_frames(
+                item,
+                self.hf_dataset,
+                self.episode_data_index,
+                self.delta_timestamps,
+                self.tolerance_s,
+            )
+
+        if self.video:
+            item = load_from_videos(
+                item,
+                self.video_frame_keys,
+                self.videos_dir,
+                self.tolerance_s,
+            )
+
+        if self.transform is not None:
+            item = self.transform(item)
+
+        return item
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}(\n"
+            f"  Repository ID: '{self.repo_id}',\n"
+            f"  Version: '{self.version}',\n"
+            f"  Split: '{self.split}',\n"
+            f"  Number of Samples: {self.num_samples},\n"
+            f"  Number of Episodes: {self.num_episodes},\n"
+            f"  Type: {'video (.mp4)' if self.video else 'image (.png)'},\n"
+            f"  Recorded Frames per Second: {self.fps},\n"
+            f"  Camera Keys: {self.camera_keys},\n"
+            f"  Video Frame Keys: {self.video_frame_keys if self.video else 'N/A'},\n"
+            f"  Transformations: {self.transform},\n"
+            f")"
+        )
+
+    @classmethod
+    def from_preloaded(
+        cls,
+        repo_id: str,
+        version: str | None = CODEBASE_VERSION,
+        root: Path | None = None,
+        split: str = "train",
+        transform: callable = None,
+        delta_timestamps: dict[list[float]] | None = None,
+        # additional preloaded attributes
+        hf_dataset=None,
+        episode_data_index=None,
+        stats=None,
+        info=None,
+        videos_dir=None,
+    ):
+        # create an empty object of type LeRobotDataset
+        obj = cls.__new__(cls)
+        obj.repo_id = repo_id
+        obj.version = version
+        obj.root = root
+        obj.split = split
+        obj.transform = transform
+        obj.delta_timestamps = delta_timestamps
+        obj.hf_dataset = hf_dataset
+        obj.episode_data_index = episode_data_index
+        obj.stats = stats
+        obj.info = info
+        obj.videos_dir = videos_dir
+        return obj

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_cabinet.txt

@@ -0,0 +1,85 @@
+https://drive.google.com/file/d/1_SOJkgfP5yZyVjMhTt3nwhvyUjcnlI51/view?usp=drive_link
+https://drive.google.com/file/d/1rmgN8UUzph1qwJnzG1d-uOafodn-gLvb/view?usp=drive_link
+https://drive.google.com/file/d/1NYQ-XxsBVinB6dUoZmVWweT83367P3i2/view?usp=drive_link
+https://drive.google.com/file/d/1oAv_j74zxxCJieMG7r5Vl2BeHK1__3s3/view?usp=drive_link
+https://drive.google.com/file/d/1wFUJQROsrTJt64YRuIeExhFjr2wnK5uu/view?usp=drive_link
+https://drive.google.com/file/d/1KzL3Tt0Le7jVl58XVRUcmigmXjyiuhbK/view?usp=drive_link
+https://drive.google.com/file/d/1qy_YBladeHtianSSGtgAPSHtMin7msvf/view?usp=drive_link
+https://drive.google.com/file/d/1rA_F0V_qL_nyuC_0aBKCisF4-0TIkF2Y/view?usp=drive_link
+https://drive.google.com/file/d/1hw-8qMpz9VgSt62XoASqNRuPECpCwJQP/view?usp=drive_link
+https://drive.google.com/file/d/1BpHOl9rKMzdvNGka6js7C0s40hH6vnDA/view?usp=drive_link
+https://drive.google.com/file/d/1PazhkhiDnJ-OUMyDVDFxEZNKQQqHiNWS/view?usp=drive_link
+https://drive.google.com/file/d/1lZ665R6ATl57dypxH4dGJ2NSt6XYnbuz/view?usp=drive_link
+https://drive.google.com/file/d/1V9HzLaf-tlG15wUzT7KrTDCS_z1vi5NV/view?usp=drive_link
+https://drive.google.com/file/d/1aKauWiXoKqbNwn_2xs4MrmLlaNYlVNmO/view?usp=drive_link
+https://drive.google.com/file/d/1WVD5DFhriO1YmmOgiVHhacR6HWoTPxav/view?usp=drive_link
+https://drive.google.com/file/d/1_X43WgeBAsfkhH9EmpyPki8U9joMeAGC/view?usp=drive_link
+https://drive.google.com/file/d/1t8x0GqWoNKWtnBsB7_D40Z34nL9ak4kf/view?usp=drive_link
+https://drive.google.com/file/d/15V_f26WaKOXjKnq2T3HRWAmtQUi4lbu2/view?usp=drive_link
+https://drive.google.com/file/d/11VFIAsiSDsMOBANgrOcZBpKB9AFWnLy7/view?usp=drive_link
+https://drive.google.com/file/d/1M0NS7vVaxJv3FHnuRYtdwTFYF7We4LxP/view?usp=drive_link
+https://drive.google.com/file/d/1mR0OItTNqFnVLoczcyKYlm6drAy778lO/view?usp=drive_link
+https://drive.google.com/file/d/1NbVFWDQAh-z4JJ4D-Zw6Lps9kdvpqh2j/view?usp=drive_link
+https://drive.google.com/file/d/1JQoZGBzl4W3QG26-n39tefcGN0fDRMbB/view?usp=drive_link
+https://drive.google.com/file/d/1VBjHl-TvZpncopvasIP5G9gecbB2a5f6/view?usp=drive_link
+https://drive.google.com/file/d/1VzSf6zaB21nahm7MsPwroXbJ84NIwq0b/view?usp=drive_link
+https://drive.google.com/file/d/1OtNnfMEydNtZOcivs4k6E_uJSpf8PkGy/view?usp=drive_link
+https://drive.google.com/file/d/14nVvpvsrFr_03Pa_N7MKzwnRwibOUYM6/view?usp=drive_link
+https://drive.google.com/file/d/1M8li6duiO2r3lv_9HhF_XJn0oZUIEK5F/view?usp=drive_link
+https://drive.google.com/file/d/1Cpzea6fO14lxAaNfSBifqoa4ekhCiLD1/view?usp=drive_link
+https://drive.google.com/file/d/1mbxRTm5vlbsY9UJ0jfjM6j9D7kPJjBpG/view?usp=drive_link
+https://drive.google.com/file/d/1RXD1i6IfWsHRlCxVmG04h2h5Ycm_WwZN/view?usp=drive_link
+https://drive.google.com/file/d/1QFqFSwDGOk1BkgGmqgCcc2BRWnJ6R3MA/view?usp=drive_link
+https://drive.google.com/file/d/1bFqWR8DQM0ZUxxtS2bl-RANQvukeFLzp/view?usp=drive_link
+https://drive.google.com/file/d/1pR-rH3yNGoyPdD4hJ6-3lXQ-PstBx9du/view?usp=drive_link
+https://drive.google.com/file/d/107OAwLY-hva9HeQLIK7VCh-ytdDabVjr/view?usp=drive_link
+https://drive.google.com/file/d/1Tpl08QOaSZ37GTO4awFWSdD8wBR9xdlT/view?usp=drive_link
+https://drive.google.com/file/d/1MR164AOM-0S1T6RX8xKTV2IHyaCvpqAW/view?usp=drive_link
+https://drive.google.com/file/d/1_wknJfVnStIhJ82lU_QtcrwahsqYIsr8/view?usp=drive_link
+https://drive.google.com/file/d/1ZuEktWrbYkTx0l5pj3WiZ2CJrfbDOHNo/view?usp=drive_link
+https://drive.google.com/file/d/15G_10hkkkq6yxvyI5NGZirlF-RzduR2F/view?usp=drive_link
+https://drive.google.com/file/d/1DBKxg3ONqh7dhLuX6oh1Yyo2x383V1Hp/view?usp=drive_link
+https://drive.google.com/file/d/1B5iDBkTUr5vopDddV_fHud18SqAHhauS/view?usp=drive_link
+https://drive.google.com/file/d/1acwFV0eenRkki1QcjSKH5xqOtys-P3Pr/view?usp=drive_link
+https://drive.google.com/file/d/1S47BI83xyrh-FKXsvAQqer98Biu_p8XK/view?usp=drive_link
+https://drive.google.com/file/d/1JL6DmBZl3uyq9dyLfgSqtGF06e7E9JwM/view?usp=drive_link
+https://drive.google.com/file/d/16WvRS4Kjog8Pxgr0E3sGGnI01YwL9Uql/view?usp=drive_link
+https://drive.google.com/file/d/12ttGqL33IPWg0-s1SD44rr22M6LiSQBr/view?usp=drive_link
+https://drive.google.com/file/d/1OyZqqnldTU_DliRbr6x0C4a_iWPwIN7j/view?usp=drive_link
+https://drive.google.com/file/d/1oYk00IpLnR9fesLfD15Ebe7nVBffEbcS/view?usp=drive_link
+https://drive.google.com/file/d/1eyE2-MQduCEqCd-5_kl5zsoOEERAzpZD/view?usp=drive_link
+https://drive.google.com/file/d/1ir1Ya-vO0d97pfvbePlUeuKTTRc0qIMU/view?usp=drive_link
+https://drive.google.com/file/d/1hOi-JnqlMt47gVnLZHMTqeojyYVErohl/view?usp=drive_link
+https://drive.google.com/file/d/1NFFw5_PqigQ7xGqsL-MNq2B1r5yAscCf/view?usp=drive_link
+https://drive.google.com/file/d/1uftq1-Zlh8d2sNLWrlVcKYQUwZTD7o24/view?usp=drive_link
+https://drive.google.com/file/d/1-ax19dSLPacVgk000T-m3l4flPcg07pM/view?usp=drive_link
+https://drive.google.com/file/d/126y-lgn86-ZmCz8hooF1THKJGGObw3OB/view?usp=drive_link
+https://drive.google.com/file/d/1JiDniK0VmDIkk92AbBILb8J2Ba59PWML/view?usp=drive_link
+https://drive.google.com/file/d/1kr8nPIRljiU0R4J9SMgj80o1FPQxzu9z/view?usp=drive_link
+https://drive.google.com/file/d/1bbThWRij1pKBh_kFgV8FwK0sXtTHBoLX/view?usp=drive_link
+https://drive.google.com/file/d/1WenzDW6lxk1xkOFm-OiGFfc0ROskAuKU/view?usp=drive_link
+https://drive.google.com/file/d/1MiKRzuzUn1yN-k_6kPJJzIGy7dT-nnsD/view?usp=drive_link
+https://drive.google.com/file/d/17rRg2tcmB-gNhQ0KoZJQmNfyFeoij1jH/view?usp=drive_link
+https://drive.google.com/file/d/11mokBpvrY3ld6sY5WztREtJ1jgqfQV70/view?usp=drive_link
+https://drive.google.com/file/d/1Il_6IOx9NDp1bX_KHizJfBwzTufTmn86/view?usp=drive_link
+https://drive.google.com/file/d/1KswtJGsxJ7eeBDAmNA_aeLjOxcH6MIxa/view?usp=drive_link
+https://drive.google.com/file/d/1gzMhi5uWu4C3Y6WbQ3L-08V96GxTZrRR/view?usp=drive_link
+https://drive.google.com/file/d/1nRQFtaBxfUCYc2W90Qibh0kHCt6YQCfc/view?usp=drive_link
+https://drive.google.com/file/d/1vs-gyW-KheqHbUATwAhA2mmR9GOGw7f_/view?usp=drive_link
+https://drive.google.com/file/d/1MuxzGOA2fgLaHryq82KkQumtuRJGcUOC/view?usp=drive_link
+https://drive.google.com/file/d/1IIwxZnGlqrXLUXqG6yMO0r7uhCvhpk9e/view?usp=drive_link
+https://drive.google.com/file/d/1vE7XPyaFcXP4DtTY5Y9WKIt7zWgmX-Cr/view?usp=drive_link
+https://drive.google.com/file/d/1j-bIV09gr21RC3-x1N_pK4RPLV3fmWKz/view?usp=drive_link
+https://drive.google.com/file/d/1t3nW1rD3S-EL0Oymb5U7ZAj5UMkydkln/view?usp=drive_link
+https://drive.google.com/file/d/14hbfHCdMKtJZ41F9CQReMec2jeRFTOqR/view?usp=drive_link
+https://drive.google.com/file/d/1x-hUyOSne5BW0AzQ3W6_Pf4g5yXQWi9M/view?usp=drive_link
+https://drive.google.com/file/d/1sw9JqRg6E-3P84I3ZhzTrJMu0vuiaMmP/view?usp=drive_link
+https://drive.google.com/file/d/1LuqhQlL4MGZhB_6THmkovRxrlP26BbdC/view?usp=drive_link
+https://drive.google.com/file/d/15C5K6v_lkjnMSmUvVyqHQKwh2N166e7K/view?usp=drive_link
+https://drive.google.com/file/d/1ns_9eSsQeeoZ10nlbkLy8tu0GmJFSnkt/view?usp=drive_link
+https://drive.google.com/file/d/1NpzWJeK6CqjxzjIMYe6aYdX8xGsQwD4o/view?usp=drive_link
+https://drive.google.com/file/d/1NMLezwufKJ9_8xTc9KQThSzVVD71B9Ui/view?usp=drive_link
+https://drive.google.com/file/d/1aa71DCUqs6oXlIxX35jgsmsgm-NlDxPV/view?usp=drive_link
+https://drive.google.com/file/d/1UJzkIZzAL0j-D5YQBnoq7mHvttASy12O/view?usp=drive_link
+https://drive.google.com/file/d/1nPgx36HIJFb7oI94VbRzWjpPP2GANxzG/view?usp=drive_link
+https://drive.google.com/file/d/1NovAP-KVJjqcuvWy3d6G4ptGGAIDqcCx/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_chair.txt

@@ -0,0 +1,55 @@
+https://drive.google.com/file/d/11M3Ye0r5agMaaicPbVGD0q2Hb3rGklbb/view?usp=drive_link
+https://drive.google.com/file/d/1-tx7SvYYgSvXCvnf_EI2OVdwK-CkFY6S/view?usp=drive_link
+https://drive.google.com/file/d/1EWJunmOpMHaU1hE106wwpbkGYcjQXYAF/view?usp=drive_link
+https://drive.google.com/file/d/1IDn95Z7FSiCckrSENtGV4u3RyFHNQSDY/view?usp=drive_link
+https://drive.google.com/file/d/1CwzvWj1i7QOtqrZvsCZ6BdZaKNDfpN32/view?usp=drive_link
+https://drive.google.com/file/d/1HvAvlhm77nAD3Td24QPSeq8lw-Rl_aOh/view?usp=drive_link
+https://drive.google.com/file/d/1t-suKYOPhXH666RpAYNRp2QU_DOy3AeM/view?usp=drive_link
+https://drive.google.com/file/d/18xpKgWh7RWyjMN5PkLTOo-AxsAadAuRw/view?usp=drive_link
+https://drive.google.com/file/d/1oci5Eto-ztv-AQNz8EnwZveBIhxvk-xJ/view?usp=drive_link
+https://drive.google.com/file/d/1Y-t_4vxdE6NpHO0DLJR8f3mD0Q-Wj5-c/view?usp=drive_link
+https://drive.google.com/file/d/1lylRqbbbB8bgtpsBWMPACmHJreuKmllv/view?usp=drive_link
+https://drive.google.com/file/d/1yliSyMig_NXShWfQx6qyW7Ijf2Y5lFK6/view?usp=drive_link
+https://drive.google.com/file/d/1XXhwJsJbeb7KXAooGvJapnm9bjnGUmxS/view?usp=drive_link
+https://drive.google.com/file/d/1_xs1f3hW2JArKyvfF7UWubWjyROGTLs6/view?usp=drive_link
+https://drive.google.com/file/d/1WVEHpr6EqKCZbkHapQSTXJq4xE4SWFT-/view?usp=drive_link
+https://drive.google.com/file/d/1RqOHv9pEQGvW8NUA7ynffFmG999TL_Az/view?usp=drive_link
+https://drive.google.com/file/d/1cu5AgD2gh-uA3PFJmzxxzNaF3qOSlYY1/view?usp=drive_link
+https://drive.google.com/file/d/1SsrXqiPclNrnYToPZ9Uq-k3y0C4qdHT1/view?usp=drive_link
+https://drive.google.com/file/d/1-J7EXf0vjkLIfSqT8ICEsP6CTjzSLBop/view?usp=drive_link
+https://drive.google.com/file/d/11O7ewUmoZXfyyKjy_6B5RW4DpjICxqBT/view?usp=drive_link
+https://drive.google.com/file/d/1iic44kZoCsjNsfAz2cMstZ9-WQvAhblF/view?usp=drive_link
+https://drive.google.com/file/d/1yLV1lVX-2WnWQldGlnQZ0x7QBuDiVkL3/view?usp=drive_link
+https://drive.google.com/file/d/1Tybp9ru98TTbGn4eyROpUQwDFuALWXmk/view?usp=drive_link
+https://drive.google.com/file/d/13E9OTMiipVJByDs5-J19oWwAz7l94LTN/view?usp=drive_link
+https://drive.google.com/file/d/1EeTpJQdMSliw4JzSMtJ6CyTvVdexjM4M/view?usp=drive_link
+https://drive.google.com/file/d/1NHyNwoFqzeAu-1_PSpq5JfxaiD_xbpn9/view?usp=drive_link
+https://drive.google.com/file/d/1fJcS0phDp4xm_FyGaJ5wr9Pe4KqtHaxD/view?usp=drive_link
+https://drive.google.com/file/d/12AqrLUaewDPEcFRqPZeZFb_TQ0Lfi3At/view?usp=drive_link
+https://drive.google.com/file/d/1x_hd4Qsq1oJS-aj2t3qM7WbbV7KZj05b/view?usp=drive_link
+https://drive.google.com/file/d/14OUSUArmsB068hs6BuEIXQhI1Cyz8Sf0/view?usp=drive_link
+https://drive.google.com/file/d/16zlzh1T5zeUJQnFf382NXkFEKEnDub4O/view?usp=drive_link
+https://drive.google.com/file/d/1IbDltmN-NEFCNtr1TO4ILxEgQ94rtjWv/view?usp=drive_link
+https://drive.google.com/file/d/15gmlf8Gx9455pZ1AlqcCSwh3nDPxMzSr/view?usp=drive_link
+https://drive.google.com/file/d/1qHpRL1oZfIMo_vxnm8qfwQ-7l0BZIVva/view?usp=drive_link
+https://drive.google.com/file/d/1H1xskIgiFZivkYn23rMzH3xePGOh3VTC/view?usp=drive_link
+https://drive.google.com/file/d/1avls6Pv0kYiCMNVknbc1zQsgy64MUDMM/view?usp=drive_link
+https://drive.google.com/file/d/1MmWVgCj5khc8KMIifmt3EzF1o-CtPyyn/view?usp=drive_link
+https://drive.google.com/file/d/1U0kCc_xqW0WNppf4sbnK14euWKdPZtzB/view?usp=drive_link
+https://drive.google.com/file/d/16CaEyQscOuhLj23PEGDTL9DeyNkohkMn/view?usp=drive_link
+https://drive.google.com/file/d/1Iu8uM6UUJ0zW8tvN-9UiOe_4oSNzEutg/view?usp=drive_link
+https://drive.google.com/file/d/1UImqiBaIxCR-1DNJaZhHqeHhaySOtVIr/view?usp=drive_link
+https://drive.google.com/file/d/1VpU2V_leIoRIyv_lAvE7eLHBG8DxCTnp/view?usp=drive_link
+https://drive.google.com/file/d/1_Q8J27OT3Xby7QY6yHvIJauFRWEMxkRm/view?usp=drive_link
+https://drive.google.com/file/d/1bantmVo1L9Xz4tbiNw_a1UC2Z_HPO1wT/view?usp=drive_link
+https://drive.google.com/file/d/1IRIXMJMCBDkBjbaHvAlEiBogSvZ1jK_3/view?usp=drive_link
+https://drive.google.com/file/d/1mAHXKjiFbjwydypW2t5Lv8_H5x6nHegl/view?usp=drive_link
+https://drive.google.com/file/d/1SfyY796fLrBCMY39OcyuxZafqSCRZPZk/view?usp=drive_link
+https://drive.google.com/file/d/1X-44sZ8CcfzIskc0dvSx882o1yFhHaZB/view?usp=drive_link
+https://drive.google.com/file/d/1BOIWCCCk6DLD4Bmvc75ZbbLi9AQm-1ao/view?usp=drive_link
+https://drive.google.com/file/d/1RuyDtRE1kk76sw-wP8vx5SgLoPF3PA_H/view?usp=drive_link
+https://drive.google.com/file/d/1c4eoQiBbGuy3CTAQDUSkd84Ponh1roAQ/view?usp=drive_link
+https://drive.google.com/file/d/19PXB9z4Ljq6dsbf9TqcOrrP5SRbw2Tc_/view?usp=drive_link
+https://drive.google.com/file/d/1nn1VVZVoIXWdYDozR7XHXE4mPLQG80PQ/view?usp=drive_link
+https://drive.google.com/file/d/1MBdFGOKPV8GUhwoSsJ_Ky3qAMLM2Bv3K/view?usp=drive_link
+https://drive.google.com/file/d/1of3k_M-7Nh3I1TndcWedxK4ca9dn8Sc5/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_elevator.txt

@@ -0,0 +1,20 @@
+https://drive.google.com/file/d/12ctkOAdkCNGN1JLbZb5ww3XTBn2LFpGI/view?usp=drive_link
+https://drive.google.com/file/d/1G_Vd46_4fq6O64gHHjUbJX5Ld44ZZx0y/view?usp=drive_link
+https://drive.google.com/file/d/1uKgUy73B3xBogQAOUhfZjO0X5qZGsi2c/view?usp=drive_link
+https://drive.google.com/file/d/1fu9cIrfI-fE2LhdGUxbx7-8Ci_PF8Ypm/view?usp=drive_link
+https://drive.google.com/file/d/1Ygk9ZPJzx8xw2A9JF3NHbJ44TqnvSTQR/view?usp=drive_link
+https://drive.google.com/file/d/18m5xPuccNsEB20WPshm3zhxmXc6k63ED/view?usp=drive_link
+https://drive.google.com/file/d/1DiqqxC44rriviRQpqogcv0-EB-Y6nr9g/view?usp=drive_link
+https://drive.google.com/file/d/1qPdaoTVDizJXkfXLioWU7iJ8hqCXSyOQ/view?usp=drive_link
+https://drive.google.com/file/d/1Fj9kIA_mG7f67WFfACJEaZ7izcHG7vUm/view?usp=drive_link
+https://drive.google.com/file/d/1WpYehZnI2P7dUdJPfkE-ij1rqCnjZEbB/view?usp=drive_link
+https://drive.google.com/file/d/1_zwWkT4jPyzB38STWb6whlzsPzXmfA9r/view?usp=drive_link
+https://drive.google.com/file/d/1U6-J4I_fPlSFFGfhZPxS5_YzKXwXIZYp/view?usp=drive_link
+https://drive.google.com/file/d/1pRhxxcTfZp5tQo_EScvJUwfc3amiS6Vk/view?usp=drive_link
+https://drive.google.com/file/d/1lWLntqra83RlYU_gN7Vostnfydf6gutd/view?usp=drive_link
+https://drive.google.com/file/d/1vIBKo0x-NYEHV1FvRpco1lQMpRdAWAIL/view?usp=drive_link
+https://drive.google.com/file/d/1pdrLV3JTQou_XH0Aap61Ssf60iVKm1jJ/view?usp=drive_link
+https://drive.google.com/file/d/1QTsLoQ7SwmKdQHjBGVDaR2uTwfFwtrOf/view?usp=drive_link
+https://drive.google.com/file/d/1Gytai8M_12J36GY6L_TulEcOC-035jwS/view?usp=drive_link
+https://drive.google.com/file/d/14LJudNc629NT-i8xreXtzl27ce_DxOFJ/view?usp=drive_link
+https://drive.google.com/file/d/1sBvPCODbzxGAI0S3lgN5cSG9Go3lRi00/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_shrimp.txt

@@ -0,0 +1,18 @@
+https://drive.google.com/file/d/1MJn9GbC8p9lN4gC9KDMLEkTkP_gGpXj0/view?usp=drive_link
+https://drive.google.com/file/d/1-4LXgjl7ZCOgp-8GCJmFRD8OeqN5Jf7-/view?usp=drive_link
+https://drive.google.com/file/d/1Ho06Ce0SPbqU3juaMxNUwAt3zCRLGC8W/view?usp=drive_link
+https://drive.google.com/file/d/1ivHoj7_7olBSxH-Y8kqXEW7ttITK-45j/view?usp=drive_link
+https://drive.google.com/file/d/1qjY4hM_IvZ8cq2II_n9MeJbvyeuN4oBP/view?usp=drive_link
+https://drive.google.com/file/d/1rKVhO_f92-7sw13T8hTVrza3B9oAVgoy/view?usp=drive_link
+https://drive.google.com/file/d/1pcLPHO8fBkc1-CRa88tyQtEueE4xiXNi/view?usp=drive_link
+https://drive.google.com/file/d/1Vev_chCsIeEdvQ8poEYNsOJFGy_QU8kZ/view?usp=drive_link
+https://drive.google.com/file/d/1l5G4zpRkxSLCQjvGPYSN4zfCvVRQuzMz/view?usp=drive_link
+https://drive.google.com/file/d/14vgthE1eoakXkr2-DRw50E6lAqYOiUuE/view?usp=drive_link
+https://drive.google.com/file/d/17nPSmKKmgQ2B7zkzWrZYiLM3RBuFod82/view?usp=drive_link
+https://drive.google.com/file/d/1QcDsxplVvb_ID9BVrihl5FvlC-j7waXi/view?usp=drive_link
+https://drive.google.com/file/d/18pEejBpI-eEVaWAAjBCyC0vgbX3T1Esj/view?usp=drive_link
+https://drive.google.com/file/d/1H8eH6_IRODtEFT6WoM77ltR5OoOrqXmI/view?usp=drive_link
+https://drive.google.com/file/d/1IWlpFRZhoxyG4nS13CWK4leZVk5wbNx4/view?usp=drive_link
+https://drive.google.com/file/d/1PbZA8_OCGmMLxNP9xbkLRSChniL4uGxl/view?usp=drive_link
+https://drive.google.com/file/d/1p9XAdmG2f_WeflNO4DIJ_tr1rK6M9B4B/view?usp=drive_link
+https://drive.google.com/file/d/1nS59Et1cNAvKo3Y4SeSGRuZD5TvBbCF3/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_wash_pan.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1S8eFg98IaGAIKVZ8QFWG1bx4mHa-O204

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_wipe_wine.txt

@@ -0,0 +1,4 @@
+https://drive.google.com/drive/folders/1tC_g1AJ8lglBLY-fjsQrG6DMBa3Ucp-0
+https://drive.google.com/file/d/1fG_Yi2MJrFjiUVN3XoiWXLtTxHlwwaDv/view?usp=drive_link
+https://drive.google.com/file/d/1WX32VWfzzX3Blmd06DRxLwFbMJfVe7P4/view?usp=drive_link
+https://drive.google.com/file/d/18onsX3vXg3xkFwP5bVUCjdV4n9TRn0C9/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_insertion_human.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1RgyD0JgTX30H4IM5XZn8I3zSV_mr8pyF
+https://drive.google.com/file/d/18Cudl6nikDtgRolea7je8iF_gGKzynOP/view?usp=drive_link
+https://drive.google.com/file/d/1C1kZYyROzs-PrLc0SkDgUgMi4-L3lauE/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_insertion_scripted.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1TsojQQSXtHEoGnqgJ3gmpPQR2DPLtS2N
+https://drive.google.com/file/d/1wfMSZ24oOh5KR_0aaP3Cnu_c4ZCveduB/view?usp=drive_link
+https://drive.google.com/file/d/17EuCUWS6uCCr6yyNzpXdcdE-_TTNCKtf/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_transfer_cube_human.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1sc-E4QYW7A0o23m1u2VWNGVq5smAsfCo
+https://drive.google.com/file/d/18smMymtr8tIxaNUQ61gW6dG50pt3MvGq/view?usp=drive_link
+https://drive.google.com/file/d/1Nk7l53d9sJoGDBKAOnNrExX5nLacATc6/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_transfer_cube_scripted.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1aRyoOhQwxhyt1J8XgEig4s6kzaw__LXj
+https://drive.google.com/file/d/1pnGIOd-E4-rhz2P3VxpknMKRZCoKt6eI/view?usp=drive_link
+https://drive.google.com/file/d/1GKReZHrXU73NMiC5zKCq_UtqPVtYq8eo/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_battery.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/19qS_n7vKgDcPeTMnvDHQ5-n73xEbJz5D
+https://drive.google.com/file/d/1oC31By0A2bsBeHyUwBdQw1z4ng6yi9Za/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_candy.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1m5rQ6UVH8Q9RQp_6c0CxkQ88-L-ScO7q
+https://drive.google.com/file/d/1wHz2qcmwcVG0C0CZ9MjQDQcmj4OY9_a3/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_coffee.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1seQGay470nGQ-knBI5TjsTr8iL9Qws5q
+https://drive.google.com/file/d/1T89hSX5U99wLGvGTE7yUBaQPOpyj6Sai/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_coffee_new.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1t3eDc5Rg0DveyRe8oTm6Dia_FYU5mXyf
+https://drive.google.com/file/d/1TXFaduTakvS0ZWJqKCX-HIvYglum_5CY/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_cups_open.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1Z9X3DNzd6LS0FFjQemNUMoMA5yk5VQOh
+https://drive.google.com/file/d/1Wlyc0vTkjXuWB6zbaVOWhEfD7BmPgUV_/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_fork_pick_up.txt

@@ -0,0 +1,53 @@
+https://drive.google.com/drive/folders/1DYgB4ifX4uIid9m9jnC0Zdz8Nf7ZC0fc
+https://drive.google.com/file/d/1Eb-NRNk_FmVleCbU_Ng5Y4dfcjTKN7Rv/view?usp=drive_link
+https://drive.google.com/file/d/1dkhjEADakT-44l9jf-nK4x89kr4yG_qb/view?usp=drive_link
+https://drive.google.com/file/d/14hDhgcZkVqNExGb4tIXpSjMshhqZETch/view?usp=drive_link
+https://drive.google.com/file/d/1zVMEHpHbuNyP5A_lYU7RPSLB-4V0yfZw/view?usp=drive_link
+https://drive.google.com/file/d/1JtgDjBvy7FnRpFzrx_foC3quorYQFAR-/view?usp=drive_link
+https://drive.google.com/file/d/1EHdneB6F-PP0dQlX8qPaXbxmKoBy_YwO/view?usp=drive_link
+https://drive.google.com/file/d/17Z0jjVBy1OPKREPu77_n_rQzorDiapji/view?usp=drive_link
+https://drive.google.com/file/d/1F4i23qPJ_qTf5jWjfLo4ARGJChznYWt3/view?usp=drive_link
+https://drive.google.com/file/d/1kZtXWM3uS0-rLblydBfJ0mMcVnMMXw9w/view?usp=drive_link
+https://drive.google.com/file/d/1mNODox87xFfY5Z_o5mcLsr8SHb39jDik/view?usp=drive_link
+https://drive.google.com/file/d/1Ob44VdmEUA93FKDECiRb5Ogz2xQg5IWp/view?usp=drive_link
+https://drive.google.com/file/d/1fdQLdjj3Cwv33R1wZhfrLz9Del8mqgHb/view?usp=drive_link
+https://drive.google.com/file/d/1Yu3L3ft21zP__XL8pCfhb788ZleuW1n5/view?usp=drive_link
+https://drive.google.com/file/d/1ozBBWXVZ9hXDh9ooHUNroHdYm8UDqnhJ/view?usp=drive_link
+https://drive.google.com/file/d/1o0TGqvfWw_Lunxb5ubKDS21Lr_WC0h75/view?usp=drive_link
+https://drive.google.com/file/d/1jZnd5eP5L6BH5l98BPN6OnoQx3fu8e9n/view?usp=drive_link
+https://drive.google.com/file/d/1S5sYbz8wcLYp0V67v13i4PRcBxodn4Hg/view?usp=drive_link
+https://drive.google.com/file/d/1rFeg_x6ftJYwPtBv34D3h2L2cpDLeR4G/view?usp=drive_link
+https://drive.google.com/file/d/1GvS3lcm4o6nm_scUk0XxKeVFNmzjucDZ/view?usp=drive_link
+https://drive.google.com/file/d/1-9i0riphC7NhhDahcQfD1QoBXP5gF90A/view?usp=drive_link
+https://drive.google.com/file/d/15p_IqGsMbKuvzMS872THAZr-3SBtb1Fr/view?usp=drive_link
+https://drive.google.com/file/d/1ToyYcBfJL8gbQn0q_59zPLsFmm7dmMJo/view?usp=drive_link
+https://drive.google.com/file/d/1e_7PNH7CYafE4pAebP7ZdI7XFbmEcy_i/view?usp=drive_link
+https://drive.google.com/file/d/1JoabvGVsIQdug2xOhUIhetEIyDM91y_Y/view?usp=drive_link
+https://drive.google.com/file/d/1kOMw1y0lmnVaCjwZICfzCsx6e0Z8MNGR/view?usp=drive_link
+https://drive.google.com/file/d/16it_wd1JOevUQTK2_CvF_pBACTgpIPgM/view?usp=drive_link
+https://drive.google.com/file/d/1IRcCj9HnJSfbyMgr5XEERGlEnWeZQwOc/view?usp=drive_link
+https://drive.google.com/file/d/1Z2dIJfq_S3liGmPN9Rphvkmucnmw7tlb/view?usp=drive_link
+https://drive.google.com/file/d/1J3NoAjzndGx9yNyaBOJHdNny1epzUoBt/view?usp=drive_link
+https://drive.google.com/file/d/18nOvxV1k8FSmBrhT4TPo2sKKSZXougyx/view?usp=drive_link
+https://drive.google.com/file/d/1CT8FxclafFMjSd7gCWVw3VSeryeiF04i/view?usp=drive_link
+https://drive.google.com/file/d/16M9KVqQMFfSsXfypK0bocFft8Nz3j2Rt/view?usp=drive_link
+https://drive.google.com/file/d/18QPVkw6bj6HW8LTPrQLWrrUX4R6RcF42/view?usp=drive_link
+https://drive.google.com/file/d/1hQTVtA5hBTE_StXpJafTZJ3tgt2VQQ_t/view?usp=drive_link
+https://drive.google.com/file/d/1Dn-d5g69H6EgAWgsFdrcbJKtz7ySsCQ8/view?usp=drive_link
+https://drive.google.com/file/d/13hMr16483P7ALYv73yMRUN37fJdVQM62/view?usp=drive_link
+https://drive.google.com/file/d/1848yN3XMN5zJMEgApt6KzrWgfRPfimtv/view?usp=drive_link
+https://drive.google.com/file/d/1oAD9kSnS0fTgj-CjD4u9VdZ5X67IOIMa/view?usp=drive_link
+https://drive.google.com/file/d/1ilzIWLCCG5b_KgF5s0wdN2I5-lFNpwC1/view?usp=drive_link
+https://drive.google.com/file/d/1rjsT2YBjnidxod1s9s-myAYz8boHr-WB/view?usp=drive_link
+https://drive.google.com/file/d/18Gg48HTub15bd8qzbhiCUufbVy0fbN5G/view?usp=drive_link
+https://drive.google.com/file/d/1WsSnQSqmMTVSRwrhT1Y-v782My2zcjLm/view?usp=drive_link
+https://drive.google.com/file/d/1ea9ZCvoyc-xqiFXgeDcA_mOWsw7VUuoi/view?usp=drive_link
+https://drive.google.com/file/d/1wv1v3-XhPgbNzp62BXbJTDzMPu2tlDUc/view?usp=drive_link
+https://drive.google.com/file/d/18-ikzt8LoZ83Gi3goKCELs4U4z8hrRoF/view?usp=drive_link
+https://drive.google.com/file/d/16Bjhp7JNCXkGuLvyNcZowAx3W-Y-15DV/view?usp=drive_link
+https://drive.google.com/file/d/1Gc-KRI-xwcp1fMR55ugbrLg_5y3SPde-/view?usp=drive_link
+https://drive.google.com/file/d/1oP72Q386Z4Sy5MMm-t5yNogIe5Van_9k/view?usp=drive_link
+https://drive.google.com/file/d/112T90eDUDVH-SyOV7UnZl5bscAH2hcfq/view?usp=drive_link
+https://drive.google.com/file/d/1y-uKOesRRhjgDtFbG_j65f4SGg0v8XDg/view?usp=drive_link
+https://drive.google.com/file/d/1LOP05OagoI3km-ZKQBrS204A85UVk7Ok/view?usp=drive_link
+https://drive.google.com/file/d/1QkHQKgasVzWsmdPvkXgGhWyQ84d93_Az/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_pingpong_test.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1Ut2cv6o6Pkfgg46DgwVUM7Z5PkNG8eJ-

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_pro_pencil.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1FqxPV0PgvgIu8XFjtvZSPSExuNcxVVAY

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_screw_driver.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1SKtG0ct9q0nVdYssJNMWSOjikcXliT58
+https://drive.google.com/file/d/1nchD21O30B3i3LDoqramo1zgW5YvpJIN/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_tape.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1_4DHf2cma0xsChLQFghwigX6Ukti5-zQ
+https://drive.google.com/file/d/1_8vS4hDNDgUQY-SmekrNaa7dF67QJYU-/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_thread_velcro.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1_4DHf2cma0xsChLQFghwigX6Ukti5-zQ
+https://drive.google.com/file/d/1_8vS4hDNDgUQY-SmekrNaa7dF67QJYU-/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_towel.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1fAD7vkyTGTFB_nGXIKofCU1U05oE3MFv
+https://drive.google.com/file/d/1XzyQ2B6LLvcurIonOpEu4nij2qwNWshH/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_vinh_cup.txt

@@ -0,0 +1,53 @@
+https://drive.google.com/drive/folders/13EQsVsnxT86K20QAoyE_YpsFbQ7fZQdu
+https://drive.google.com/file/d/1-W_JHghZG65FNTVhw1SXhtQrazdLL3Ue/view?usp=drive_link
+https://drive.google.com/file/d/1VwRJgdWUo-2nQaNM7Bs77-fsm8iwUxEo/view?usp=drive_link
+https://drive.google.com/file/d/1wFzGRo5iYA13WLi6IV1ry64RyahQBFio/view?usp=drive_link
+https://drive.google.com/file/d/1IKtQzQ-n-UTv64hYpReu2R4cqUvmNQqD/view?usp=drive_link
+https://drive.google.com/file/d/1GicVci9OiuuZZH79i5Mg7AtWod94MzwT/view?usp=drive_link
+https://drive.google.com/file/d/1JVnIoR7EIQp70T4eAf9RX65JcTrzsjQc/view?usp=drive_link
+https://drive.google.com/file/d/1W2xr4h23ucjPrc-mBEeqnACsfaImpc0p/view?usp=drive_link
+https://drive.google.com/file/d/10xj_0V7A07o3uCa7v5omUrTC0YlPW8H3/view?usp=drive_link
+https://drive.google.com/file/d/1FOc3EMaCy8Mb0_a7PuXLAwKwvxkbKmwU/view?usp=drive_link
+https://drive.google.com/file/d/143PgDXBcf2GQ0Q07ZPMVMfBgZDd5sLJG/view?usp=drive_link
+https://drive.google.com/file/d/1pE5Tyj0LlGbGWvUzuhixp86Ibu55Ez3I/view?usp=drive_link
+https://drive.google.com/file/d/141668b1VzX80ncrVJPzhkoAeIFB4MEK9/view?usp=drive_link
+https://drive.google.com/file/d/1bw12lo37p1ZvRvErHsll7cEYi2OxscvZ/view?usp=drive_link
+https://drive.google.com/file/d/1zfnMFvbgBjl6SzYhksbaOzfbwLrCN6tb/view?usp=drive_link
+https://drive.google.com/file/d/1-GIszA6mUJMaNB-tdh9r9skc77SWA0VX/view?usp=drive_link
+https://drive.google.com/file/d/1fTB0zWFYU6zh4IIUFT2zX_OkwYqmElwY/view?usp=drive_link
+https://drive.google.com/file/d/1gPIPNKGmrO9c7gKF7SP0SuUYbIBBq8z1/view?usp=drive_link
+https://drive.google.com/file/d/12JeJ-dQd5lYyn6PlDOGdE-ChVeiZ-Uv0/view?usp=drive_link
+https://drive.google.com/file/d/100_20cgCqerU6qoh3TfTbwLy9mlDAFEG/view?usp=drive_link
+https://drive.google.com/file/d/111oAGJ76ku_pYgbBoIdZAC1_XEQcPI__/view?usp=drive_link
+https://drive.google.com/file/d/1UhC8L-354ZQ2gblPFGI35EMsVwfpuKa0/view?usp=drive_link
+https://drive.google.com/file/d/1sIXQSgUR_xdrNtGrL6QGBnkLMKErsIp1/view?usp=drive_link
+https://drive.google.com/file/d/16Ax77bDSIXnsn4GFL8XYKKT1P6bPpfMd/view?usp=drive_link
+https://drive.google.com/file/d/1pgRVYwwVIsWq_qsWqZpe1UBzZfF5Fa9D/view?usp=drive_link
+https://drive.google.com/file/d/1jtimaZkWsY1P5gC2bbS64H_WCUU7HXN2/view?usp=drive_link
+https://drive.google.com/file/d/1N6Bh02P-RiTEgtx1YH1Db_X3TGpP-X_r/view?usp=drive_link
+https://drive.google.com/file/d/14Fy8EwJ8d9Vh97Yt1VOvUChSCrfIjBij/view?usp=drive_link
+https://drive.google.com/file/d/1IRuv42dvIMPuKhcMZmuXaBjJ-lPFOmQd/view?usp=drive_link
+https://drive.google.com/file/d/16XWzNY2D8ucVVn5geBgsVdhm3ppO4que/view?usp=drive_link
+https://drive.google.com/file/d/1xsVOoQgthK_L_SDrmq_JvQgUpAvPEAY8/view?usp=drive_link
+https://drive.google.com/file/d/1bZbw66DyEMvnJnzkdUUNbKjvNKg8KFYM/view?usp=drive_link
+https://drive.google.com/file/d/1CyTVkdrNGGpouCXr4CfhKbMzE6Ah3oo3/view?usp=drive_link
+https://drive.google.com/file/d/1hDRyeM-XEDpHXpptbT8LvNnlQUR3PWOh/view?usp=drive_link
+https://drive.google.com/file/d/1XhHWxbra8Iy5irQZ83IvxwaJqHq9x4s1/view?usp=drive_link
+https://drive.google.com/file/d/1haZcn6aM1o4JlmP9tJj3x2enrxiPaDSD/view?usp=drive_link
+https://drive.google.com/file/d/1ypDyuUTbljaBZ34f-t7lj3O_0bRmyX2n/view?usp=drive_link
+https://drive.google.com/file/d/1ILEEZo_tA9_ChIAprr2mPaNVKZi5vXsO/view?usp=drive_link
+https://drive.google.com/file/d/1U7nVYFaGE8vVTfLCW33D74xOjDcqfgyJ/view?usp=drive_link
+https://drive.google.com/file/d/1rZ93_rmCov5SMDxPkfM3qthcRELZrQX6/view?usp=drive_link
+https://drive.google.com/file/d/1mYO1b_csddtyE3qT6cwLiw-m2w2_1Lxh/view?usp=drive_link
+https://drive.google.com/file/d/1xz7Q5x2jikY8wJQjMRQpRws6AnfWlHm5/view?usp=drive_link
+https://drive.google.com/file/d/1OO8GaO-0FrSZRd1kxMYwBmubyiLOWnbl/view?usp=drive_link
+https://drive.google.com/file/d/1EXn4NVDmf-4_HCy34mYwT-vwK2CFI9ev/view?usp=drive_link
+https://drive.google.com/file/d/10hH70XhXRL9C5SnAG4toHtfHqfJUJo4H/view?usp=drive_link
+https://drive.google.com/file/d/18tiBcxea0guUai4lwsXQvt0q2LZ8ZnnJ/view?usp=drive_link
+https://drive.google.com/file/d/1Q8R8qv37vk5PQ5kQ2ibx6BFLOySD0VpX/view?usp=drive_link
+https://drive.google.com/file/d/17aNriHzjhdibCyuUjQoMFZqjybJZtggG/view?usp=drive_link
+https://drive.google.com/file/d/1LVjEYHSdeKm6CotU1QguIeNEPaIaFl_1/view?usp=drive_link
+https://drive.google.com/file/d/1ufAhE_EkgJ85slg2EW8aW_grOzE_Lmxd/view?usp=drive_link
+https://drive.google.com/file/d/1wtzLtXrkw9eXRGESTPIOlpl1tInu-b2m/view?usp=drive_link
+https://drive.google.com/file/d/1Mk5qvVtD_QHwGOUApRq76TUw2T5THu6f/view?usp=drive_link
+https://drive.google.com/file/d/1y1WQ3hboWVJ68KEYQQ3OhreGuaUpSgwc/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_vinh_cup_left.txt

@@ -0,0 +1,52 @@
+https://drive.google.com/drive/folders/1dxWh6YFZUDt6qXIoxgD9bla3CiFjZ11C
+https://drive.google.com/file/d/1hNBJN00SCAlOl0ZEgm7RRGbAGDjyBs0p/view?usp=drive_link
+https://drive.google.com/file/d/17He0CVwXGeoMmXg4SHKo-osNn7YPKVL7/view?usp=drive_link
+https://drive.google.com/file/d/1laNKUVID1x2CV6a2O2WQjwFewKu4lidL/view?usp=drive_link
+https://drive.google.com/file/d/1pNf36xbZJGRArYLmNAvRj5y6CoqdC6kB/view?usp=drive_link
+https://drive.google.com/file/d/1_4E1-y3JXk5I0ebycLYM70YDPK9g52gZ/view?usp=drive_link
+https://drive.google.com/file/d/1PHfzhGPdbolKyOpS3FnR2w7Q8zUlJXSk/view?usp=drive_link
+https://drive.google.com/file/d/17ls2PPN-Pi3tEuK059cwV2_iDT8aGhOO/view?usp=drive_link
+https://drive.google.com/file/d/1LWsg6PmCT00Kv_N_slrmcwKmQPGoBT3k/view?usp=drive_link
+https://drive.google.com/file/d/12LckrchoHTUVH7rxi8J7zD9dA19GXvoW/view?usp=drive_link
+https://drive.google.com/file/d/1VqrJKjAIkj5gtFXL69grdSeu9CyaqnSw/view?usp=drive_link
+https://drive.google.com/file/d/1g5rQYDBZvW-kUtYPeyF3qmd53v6k7kXu/view?usp=drive_link
+https://drive.google.com/file/d/10kUgaSJ0TS7teaG83G3Rf_DG4XGrBt6A/view?usp=drive_link
+https://drive.google.com/file/d/1je9XmneZQZvTma5adMJICUPDovW3ppei/view?usp=drive_link
+https://drive.google.com/file/d/1v28r6bedwZGbUPVVTVImXhK-42XdtGfj/view?usp=drive_link
+https://drive.google.com/file/d/1-TEEx9sGVvzMMaNXYfQMtY2JJ6cvl0dT/view?usp=drive_link
+https://drive.google.com/file/d/1YdBKdJFP9rJWBUX7qrOYL_gfUA8o6J9M/view?usp=drive_link
+https://drive.google.com/file/d/1X9vffwQHNUSKLXr2RlYNtbWDIFCIDfdF/view?usp=drive_link
+https://drive.google.com/file/d/11hqesqa5kvEe5FABUnZRcvmOhR373cYM/view?usp=drive_link
+https://drive.google.com/file/d/1ltTTECjEcbQPgS3UPRgMzaE2x9n6H7dC/view?usp=drive_link
+https://drive.google.com/file/d/1Zxqfa29JdwT-bfMpivi6IG2vz34d21dD/view?usp=drive_link
+https://drive.google.com/file/d/11LQlVxS5hz494dYUJ_PNRPx2NHIJbQns/view?usp=drive_link
+https://drive.google.com/file/d/1i1JhNtnZpO_E8rAv8gxBP3ZTZRvcvsZi/view?usp=drive_link
+https://drive.google.com/file/d/11jOXAr2EULUO4Qkm748634lg4UUFho5U/view?usp=drive_link
+https://drive.google.com/file/d/1rj67wur8DdB_Pipwx24bY43xu4X1eQ5e/view?usp=drive_link
+https://drive.google.com/file/d/15ZTm6lO6f_JQy_4SNfrOu3iPYn1Ro8mh/view?usp=drive_link
+https://drive.google.com/file/d/1q4gBtqWPJtCwXEvknGgN0WHGp7Vfn1b9/view?usp=drive_link
+https://drive.google.com/file/d/1t17keyre47AYqm8GgXiQ7EcvcUkeSiDQ/view?usp=drive_link
+https://drive.google.com/file/d/1OYUPGxtZgOF86Ng_BEOTXm_XOYpuQPsO/view?usp=drive_link
+https://drive.google.com/file/d/1cBjbGHi3dwWHtx6r9EQJi0JT_CE3LuHt/view?usp=drive_link
+https://drive.google.com/file/d/14qaMyF0mcbCB-fCYKNyo5_2NahSC6D5u/view?usp=drive_link
+https://drive.google.com/file/d/12FgX86eA7Y5co9ULBVK80XMsiKQSs-Ri/view?usp=drive_link
+https://drive.google.com/file/d/1yvoHWidf-jdBVw6qCCXOFfkVwKj_2hPk/view?usp=drive_link
+https://drive.google.com/file/d/1a2SugsSDlC8UtUrFzp-_KAwyZckQOvdQ/view?usp=drive_link
+https://drive.google.com/file/d/1l8pILBFSAosypWJMza2K09Vm7rug9axm/view?usp=drive_link
+https://drive.google.com/file/d/1hfPQ8dBCk97PnOhq6_MIISm3IEzcOxJG/view?usp=drive_link
+https://drive.google.com/file/d/1PPAUwlJCFKpms8cqF_k1v2_fCgDBOc3S/view?usp=drive_link
+https://drive.google.com/file/d/1lVKQZeqFfK3amEmLuFhYLUFQ2eyE8rOW/view?usp=drive_link
+https://drive.google.com/file/d/1K9iPMLfDowcIFoyzpvgn88dQ6x6kVwNG/view?usp=drive_link
+https://drive.google.com/file/d/1PNvMqG9tL7QxeLaYBGHiWYR6SYb5iIct/view?usp=drive_link
+https://drive.google.com/file/d/1xkRtzbvIkUsylx9hrFLGQsJn0h1EYu-5/view?usp=drive_link
+https://drive.google.com/file/d/1nxMRrJlSayjDIfr5CmHO1NzAw3COhsLi/view?usp=drive_link
+https://drive.google.com/file/d/1Qs3WEyMGrmagiHIkkFEueWNnJhkUeR1s/view?usp=drive_link
+https://drive.google.com/file/d/1D-G2_Q0SS3M8zyJbg_XzkF2ANPw1HTuX/view?usp=drive_link
+https://drive.google.com/file/d/1mdmJsDGO-YtJAOF_yPKl6lq4PJOIbQhT/view?usp=drive_link
+https://drive.google.com/file/d/11m9bwfop_sPmnQr_8amB6EEsrbAeG_z5/view?usp=drive_link
+https://drive.google.com/file/d/19tyYt5FMn5kru0g9o2nMJhKPnsDqkIZv/view?usp=drive_link
+https://drive.google.com/file/d/1XvTpUdsVTZ-vydvdYYmynbma--HfUGSl/view?usp=drive_link
+https://drive.google.com/file/d/1MO3hFu68J6NohTzr9aB_fY02VA6QSOqj/view?usp=drive_link
+https://drive.google.com/file/d/1Lh-UjwAk__04YOTWINF_QGVU8SjetVaY/view?usp=drive_link
+https://drive.google.com/file/d/1jkSOUwZV5GJ7rZlVeErjcu0DBQs8Np0d/view?usp=drive_link
+https://drive.google.com/file/d/1VIN1eLI-93WrVQwCjsv6XQr353DqqBYA/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_ziploc_slide.txt

@@ -0,0 +1,8 @@
+https://drive.google.com/drive/folders/1EgKar7rWBmTIRmeJYZciSwjZx3uP2mHO
+https://drive.google.com/file/d/12eYWQO15atK2hBjXhynPJd9MKAj_42pz/view?usp=drive_link
+https://drive.google.com/file/d/1Ul4oEeICJDjgfYTl4H1uaisTzVYIM6wd/view?usp=drive_link
+https://drive.google.com/file/d/1WSF-OG8lKSe2wVYCv5D1aJNipxpgddk-/view?usp=drive_link
+https://drive.google.com/file/d/1_ppD5j5sFh26aWW0JmhLzJMeNB-lCArk/view?usp=drive_link
+https://drive.google.com/file/d/1WUp846dgWXYhu4oJfhHxiU6YL_7N6s4W/view?usp=drive_link
+https://drive.google.com/file/d/1HRZNAIoAQw_uYiPwnBvtBioQoqiqoXdA/view?usp=drive_link
+https://drive.google.com/file/d/1hedGq-QDMnIn8GlXXBC3GiEJ_Y-LTxyt/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_diffusion_policy_replay_buffer.py

@@ -0,0 +1,634 @@
+#!/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.
+"""Helper code for loading PushT dataset from Diffusion Policy (https://diffusion-policy.cs.columbia.edu/)
+
+Copied from the original Diffusion Policy repository and used in our `download_and_upload_dataset.py` script.
+"""
+
+from __future__ import annotations
+
+import math
+import numbers
+import os
+from functools import cached_property
+
+import numcodecs
+import numpy as np
+import zarr
+
+
+def check_chunks_compatible(chunks: tuple, shape: tuple):
+    assert len(shape) == len(chunks)
+    for c in chunks:
+        assert isinstance(c, numbers.Integral)
+        assert c > 0
+
+
+def rechunk_recompress_array(group, name, chunks=None, chunk_length=None, compressor=None, tmp_key="_temp"):
+    old_arr = group[name]
+    if chunks is None:
+        chunks = (chunk_length,) + old_arr.chunks[1:] if chunk_length is not None else old_arr.chunks
+    check_chunks_compatible(chunks, old_arr.shape)
+
+    if compressor is None:
+        compressor = old_arr.compressor
+
+    if (chunks == old_arr.chunks) and (compressor == old_arr.compressor):
+        # no change
+        return old_arr
+
+    # rechunk recompress
+    group.move(name, tmp_key)
+    old_arr = group[tmp_key]
+    n_copied, n_skipped, n_bytes_copied = zarr.copy(
+        source=old_arr,
+        dest=group,
+        name=name,
+        chunks=chunks,
+        compressor=compressor,
+    )
+    del group[tmp_key]
+    arr = group[name]
+    return arr
+
+
+def get_optimal_chunks(shape, dtype, target_chunk_bytes=2e6, max_chunk_length=None):
+    """
+    Common shapes
+    T,D
+    T,N,D
+    T,H,W,C
+    T,N,H,W,C
+    """
+    itemsize = np.dtype(dtype).itemsize
+    # reversed
+    rshape = list(shape[::-1])
+    if max_chunk_length is not None:
+        rshape[-1] = int(max_chunk_length)
+    split_idx = len(shape) - 1
+    for i in range(len(shape) - 1):
+        this_chunk_bytes = itemsize * np.prod(rshape[:i])
+        next_chunk_bytes = itemsize * np.prod(rshape[: i + 1])
+        if this_chunk_bytes <= target_chunk_bytes and next_chunk_bytes > target_chunk_bytes:
+            split_idx = i
+
+    rchunks = rshape[:split_idx]
+    item_chunk_bytes = itemsize * np.prod(rshape[:split_idx])
+    this_max_chunk_length = rshape[split_idx]
+    next_chunk_length = min(this_max_chunk_length, math.ceil(target_chunk_bytes / item_chunk_bytes))
+    rchunks.append(next_chunk_length)
+    len_diff = len(shape) - len(rchunks)
+    rchunks.extend([1] * len_diff)
+    chunks = tuple(rchunks[::-1])
+    # print(np.prod(chunks) * itemsize / target_chunk_bytes)
+    return chunks
+
+
+class ReplayBuffer:
+    """
+    Zarr-based temporal datastructure.
+    Assumes first dimension to be time. Only chunk in time dimension.
+    """
+
+    def __init__(self, root: zarr.Group | dict[str, dict]):
+        """
+        Dummy constructor. Use copy_from* and create_from* class methods instead.
+        """
+        assert "data" in root
+        assert "meta" in root
+        assert "episode_ends" in root["meta"]
+        for value in root["data"].values():
+            assert value.shape[0] == root["meta"]["episode_ends"][-1]
+        self.root = root
+
+    # ============= create constructors ===============
+    @classmethod
+    def create_empty_zarr(cls, storage=None, root=None):
+        if root is None:
+            if storage is None:
+                storage = zarr.MemoryStore()
+            root = zarr.group(store=storage)
+        root.require_group("data", overwrite=False)
+        meta = root.require_group("meta", overwrite=False)
+        if "episode_ends" not in meta:
+            meta.zeros("episode_ends", shape=(0,), dtype=np.int64, compressor=None, overwrite=False)
+        return cls(root=root)
+
+    @classmethod
+    def create_empty_numpy(cls):
+        root = {"data": {}, "meta": {"episode_ends": np.zeros((0,), dtype=np.int64)}}
+        return cls(root=root)
+
+    @classmethod
+    def create_from_group(cls, group, **kwargs):
+        if "data" not in group:
+            # create from stratch
+            buffer = cls.create_empty_zarr(root=group, **kwargs)
+        else:
+            # already exist
+            buffer = cls(root=group, **kwargs)
+        return buffer
+
+    @classmethod
+    def create_from_path(cls, zarr_path, mode="r", **kwargs):
+        """
+        Open a on-disk zarr directly (for dataset larger than memory).
+        Slower.
+        """
+        group = zarr.open(os.path.expanduser(zarr_path), mode)
+        return cls.create_from_group(group, **kwargs)
+
+    # ============= copy constructors ===============
+    @classmethod
+    def copy_from_store(
+        cls,
+        src_store,
+        store=None,
+        keys=None,
+        chunks: dict[str, tuple] | None = None,
+        compressors: dict | str | numcodecs.abc.Codec | None = None,
+        if_exists="replace",
+        **kwargs,
+    ):
+        """
+        Load to memory.
+        """
+        src_root = zarr.group(src_store)
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        root = None
+        if store is None:
+            # numpy backend
+            meta = {}
+            for key, value in src_root["meta"].items():
+                if len(value.shape) == 0:
+                    meta[key] = np.array(value)
+                else:
+                    meta[key] = value[:]
+
+            if keys is None:
+                keys = src_root["data"].keys()
+            data = {}
+            for key in keys:
+                arr = src_root["data"][key]
+                data[key] = arr[:]
+
+            root = {"meta": meta, "data": data}
+        else:
+            root = zarr.group(store=store)
+            # copy without recompression
+            n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
+                source=src_store, dest=store, source_path="/meta", dest_path="/meta", if_exists=if_exists
+            )
+            data_group = root.create_group("data", overwrite=True)
+            if keys is None:
+                keys = src_root["data"].keys()
+            for key in keys:
+                value = src_root["data"][key]
+                cks = cls._resolve_array_chunks(chunks=chunks, key=key, array=value)
+                cpr = cls._resolve_array_compressor(compressors=compressors, key=key, array=value)
+                if cks == value.chunks and cpr == value.compressor:
+                    # copy without recompression
+                    this_path = "/data/" + key
+                    n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
+                        source=src_store,
+                        dest=store,
+                        source_path=this_path,
+                        dest_path=this_path,
+                        if_exists=if_exists,
+                    )
+                else:
+                    # copy with recompression
+                    n_copied, n_skipped, n_bytes_copied = zarr.copy(
+                        source=value,
+                        dest=data_group,
+                        name=key,
+                        chunks=cks,
+                        compressor=cpr,
+                        if_exists=if_exists,
+                    )
+        buffer = cls(root=root)
+        return buffer
+
+    @classmethod
+    def copy_from_path(
+        cls,
+        zarr_path,
+        backend=None,
+        store=None,
+        keys=None,
+        chunks: dict[str, tuple] | None = None,
+        compressors: dict | str | numcodecs.abc.Codec | None = None,
+        if_exists="replace",
+        **kwargs,
+    ):
+        """
+        Copy a on-disk zarr to in-memory compressed.
+        Recommended
+        """
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        if backend == "numpy":
+            print("backend argument is deprecated!")
+            store = None
+        group = zarr.open(os.path.expanduser(zarr_path), "r")
+        return cls.copy_from_store(
+            src_store=group.store,
+            store=store,
+            keys=keys,
+            chunks=chunks,
+            compressors=compressors,
+            if_exists=if_exists,
+            **kwargs,
+        )
+
+    # ============= save methods ===============
+    def save_to_store(
+        self,
+        store,
+        chunks: dict[str, tuple] | None = None,
+        compressors: str | numcodecs.abc.Codec | dict | None = None,
+        if_exists="replace",
+        **kwargs,
+    ):
+        root = zarr.group(store)
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        if self.backend == "zarr":
+            # recompression free copy
+            n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
+                source=self.root.store,
+                dest=store,
+                source_path="/meta",
+                dest_path="/meta",
+                if_exists=if_exists,
+            )
+        else:
+            meta_group = root.create_group("meta", overwrite=True)
+            # save meta, no chunking
+            for key, value in self.root["meta"].items():
+                _ = meta_group.array(name=key, data=value, shape=value.shape, chunks=value.shape)
+
+        # save data, chunk
+        data_group = root.create_group("data", overwrite=True)
+        for key, value in self.root["data"].items():
+            cks = self._resolve_array_chunks(chunks=chunks, key=key, array=value)
+            cpr = self._resolve_array_compressor(compressors=compressors, key=key, array=value)
+            if isinstance(value, zarr.Array):
+                if cks == value.chunks and cpr == value.compressor:
+                    # copy without recompression
+                    this_path = "/data/" + key
+                    n_copied, n_skipped, n_bytes_copied = zarr.copy_store(
+                        source=self.root.store,
+                        dest=store,
+                        source_path=this_path,
+                        dest_path=this_path,
+                        if_exists=if_exists,
+                    )
+                else:
+                    # copy with recompression
+                    n_copied, n_skipped, n_bytes_copied = zarr.copy(
+                        source=value,
+                        dest=data_group,
+                        name=key,
+                        chunks=cks,
+                        compressor=cpr,
+                        if_exists=if_exists,
+                    )
+            else:
+                # numpy
+                _ = data_group.array(name=key, data=value, chunks=cks, compressor=cpr)
+        return store
+
+    def save_to_path(
+        self,
+        zarr_path,
+        chunks: dict[str, tuple] | None = None,
+        compressors: str | numcodecs.abc.Codec | dict | None = None,
+        if_exists="replace",
+        **kwargs,
+    ):
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        store = zarr.DirectoryStore(os.path.expanduser(zarr_path))
+        return self.save_to_store(
+            store, chunks=chunks, compressors=compressors, if_exists=if_exists, **kwargs
+        )
+
+    @staticmethod
+    def resolve_compressor(compressor="default"):
+        if compressor == "default":
+            compressor = numcodecs.Blosc(cname="lz4", clevel=5, shuffle=numcodecs.Blosc.NOSHUFFLE)
+        elif compressor == "disk":
+            compressor = numcodecs.Blosc("zstd", clevel=5, shuffle=numcodecs.Blosc.BITSHUFFLE)
+        return compressor
+
+    @classmethod
+    def _resolve_array_compressor(cls, compressors: dict | str | numcodecs.abc.Codec, key, array):
+        # allows compressor to be explicitly set to None
+        cpr = "nil"
+        if isinstance(compressors, dict):
+            if key in compressors:
+                cpr = cls.resolve_compressor(compressors[key])
+            elif isinstance(array, zarr.Array):
+                cpr = array.compressor
+        else:
+            cpr = cls.resolve_compressor(compressors)
+        # backup default
+        if cpr == "nil":
+            cpr = cls.resolve_compressor("default")
+        return cpr
+
+    @classmethod
+    def _resolve_array_chunks(cls, chunks: dict | tuple, key, array):
+        cks = None
+        if isinstance(chunks, dict):
+            if key in chunks:
+                cks = chunks[key]
+            elif isinstance(array, zarr.Array):
+                cks = array.chunks
+        elif isinstance(chunks, tuple):
+            cks = chunks
+        else:
+            raise TypeError(f"Unsupported chunks type {type(chunks)}")
+        # backup default
+        if cks is None:
+            cks = get_optimal_chunks(shape=array.shape, dtype=array.dtype)
+        # check
+        check_chunks_compatible(chunks=cks, shape=array.shape)
+        return cks
+
+    # ============= properties =================
+    @cached_property
+    def data(self):
+        return self.root["data"]
+
+    @cached_property
+    def meta(self):
+        return self.root["meta"]
+
+    def update_meta(self, data):
+        # sanitize data
+        np_data = {}
+        for key, value in data.items():
+            if isinstance(value, np.ndarray):
+                np_data[key] = value
+            else:
+                arr = np.array(value)
+                if arr.dtype == object:
+                    raise TypeError(f"Invalid value type {type(value)}")
+                np_data[key] = arr
+
+        meta_group = self.meta
+        if self.backend == "zarr":
+            for key, value in np_data.items():
+                _ = meta_group.array(
+                    name=key, data=value, shape=value.shape, chunks=value.shape, overwrite=True
+                )
+        else:
+            meta_group.update(np_data)
+
+        return meta_group
+
+    @property
+    def episode_ends(self):
+        return self.meta["episode_ends"]
+
+    def get_episode_idxs(self):
+        import numba
+
+        numba.jit(nopython=True)
+
+        def _get_episode_idxs(episode_ends):
+            result = np.zeros((episode_ends[-1],), dtype=np.int64)
+            for i in range(len(episode_ends)):
+                start = 0
+                if i > 0:
+                    start = episode_ends[i - 1]
+                end = episode_ends[i]
+                for idx in range(start, end):
+                    result[idx] = i
+            return result
+
+        return _get_episode_idxs(self.episode_ends)
+
+    @property
+    def backend(self):
+        backend = "numpy"
+        if isinstance(self.root, zarr.Group):
+            backend = "zarr"
+        return backend
+
+    # =========== dict-like API ==============
+    def __repr__(self) -> str:
+        if self.backend == "zarr":
+            return str(self.root.tree())
+        else:
+            return super().__repr__()
+
+    def keys(self):
+        return self.data.keys()
+
+    def values(self):
+        return self.data.values()
+
+    def items(self):
+        return self.data.items()
+
+    def __getitem__(self, key):
+        return self.data[key]
+
+    def __contains__(self, key):
+        return key in self.data
+
+    # =========== our API ==============
+    @property
+    def n_steps(self):
+        if len(self.episode_ends) == 0:
+            return 0
+        return self.episode_ends[-1]
+
+    @property
+    def n_episodes(self):
+        return len(self.episode_ends)
+
+    @property
+    def chunk_size(self):
+        if self.backend == "zarr":
+            return next(iter(self.data.arrays()))[-1].chunks[0]
+        return None
+
+    @property
+    def episode_lengths(self):
+        ends = self.episode_ends[:]
+        ends = np.insert(ends, 0, 0)
+        lengths = np.diff(ends)
+        return lengths
+
+    def add_episode(
+        self,
+        data: dict[str, np.ndarray],
+        chunks: dict[str, tuple] | None = None,
+        compressors: str | numcodecs.abc.Codec | dict | None = None,
+    ):
+        if chunks is None:
+            chunks = {}
+        if compressors is None:
+            compressors = {}
+        assert len(data) > 0
+        is_zarr = self.backend == "zarr"
+
+        curr_len = self.n_steps
+        episode_length = None
+        for value in data.values():
+            assert len(value.shape) >= 1
+            if episode_length is None:
+                episode_length = len(value)
+            else:
+                assert episode_length == len(value)
+        new_len = curr_len + episode_length
+
+        for key, value in data.items():
+            new_shape = (new_len,) + value.shape[1:]
+            # create array
+            if key not in self.data:
+                if is_zarr:
+                    cks = self._resolve_array_chunks(chunks=chunks, key=key, array=value)
+                    cpr = self._resolve_array_compressor(compressors=compressors, key=key, array=value)
+                    arr = self.data.zeros(
+                        name=key, shape=new_shape, chunks=cks, dtype=value.dtype, compressor=cpr
+                    )
+                else:
+                    # copy data to prevent modify
+                    arr = np.zeros(shape=new_shape, dtype=value.dtype)
+                    self.data[key] = arr
+            else:
+                arr = self.data[key]
+                assert value.shape[1:] == arr.shape[1:]
+                # same method for both zarr and numpy
+                if is_zarr:
+                    arr.resize(new_shape)
+                else:
+                    arr.resize(new_shape, refcheck=False)
+            # copy data
+            arr[-value.shape[0] :] = value
+
+        # append to episode ends
+        episode_ends = self.episode_ends
+        if is_zarr:
+            episode_ends.resize(episode_ends.shape[0] + 1)
+        else:
+            episode_ends.resize(episode_ends.shape[0] + 1, refcheck=False)
+        episode_ends[-1] = new_len
+
+        # rechunk
+        if is_zarr and episode_ends.chunks[0] < episode_ends.shape[0]:
+            rechunk_recompress_array(self.meta, "episode_ends", chunk_length=int(episode_ends.shape[0] * 1.5))
+
+    def drop_episode(self):
+        is_zarr = self.backend == "zarr"
+        episode_ends = self.episode_ends[:].copy()
+        assert len(episode_ends) > 0
+        start_idx = 0
+        if len(episode_ends) > 1:
+            start_idx = episode_ends[-2]
+        for value in self.data.values():
+            new_shape = (start_idx,) + value.shape[1:]
+            if is_zarr:
+                value.resize(new_shape)
+            else:
+                value.resize(new_shape, refcheck=False)
+        if is_zarr:
+            self.episode_ends.resize(len(episode_ends) - 1)
+        else:
+            self.episode_ends.resize(len(episode_ends) - 1, refcheck=False)
+
+    def pop_episode(self):
+        assert self.n_episodes > 0
+        episode = self.get_episode(self.n_episodes - 1, copy=True)
+        self.drop_episode()
+        return episode
+
+    def extend(self, data):
+        self.add_episode(data)
+
+    def get_episode(self, idx, copy=False):
+        idx = list(range(len(self.episode_ends)))[idx]
+        start_idx = 0
+        if idx > 0:
+            start_idx = self.episode_ends[idx - 1]
+        end_idx = self.episode_ends[idx]
+        result = self.get_steps_slice(start_idx, end_idx, copy=copy)
+        return result
+
+    def get_episode_slice(self, idx):
+        start_idx = 0
+        if idx > 0:
+            start_idx = self.episode_ends[idx - 1]
+        end_idx = self.episode_ends[idx]
+        return slice(start_idx, end_idx)
+
+    def get_steps_slice(self, start, stop, step=None, copy=False):
+        _slice = slice(start, stop, step)
+
+        result = {}
+        for key, value in self.data.items():
+            x = value[_slice]
+            if copy and isinstance(value, np.ndarray):
+                x = x.copy()
+            result[key] = x
+        return result
+
+    # =========== chunking =============
+    def get_chunks(self) -> dict:
+        assert self.backend == "zarr"
+        chunks = {}
+        for key, value in self.data.items():
+            chunks[key] = value.chunks
+        return chunks
+
+    def set_chunks(self, chunks: dict):
+        assert self.backend == "zarr"
+        for key, value in chunks.items():
+            if key in self.data:
+                arr = self.data[key]
+                if value != arr.chunks:
+                    check_chunks_compatible(chunks=value, shape=arr.shape)
+                    rechunk_recompress_array(self.data, key, chunks=value)
+
+    def get_compressors(self) -> dict:
+        assert self.backend == "zarr"
+        compressors = {}
+        for key, value in self.data.items():
+            compressors[key] = value.compressor
+        return compressors
+
+    def set_compressors(self, compressors: dict):
+        assert self.backend == "zarr"
+        for key, value in compressors.items():
+            if key in self.data:
+                arr = self.data[key]
+                compressor = self.resolve_compressor(value)
+                if compressor != arr.compressor:
+                    rechunk_recompress_array(self.data, key, compressor=compressor)

lerobot/common/datasets/push_dataset_to_hub/_download_raw.py

@@ -0,0 +1,169 @@
+#!/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 all obsolete download scripts. They are centralized here to not have to load
+useless dependencies when using datasets.
+"""
+
+import io
+import logging
+import shutil
+from pathlib import Path
+
+import tqdm
+from huggingface_hub import snapshot_download
+
+
+def download_raw(raw_dir, dataset_id):
+    if "aloha" in dataset_id or "image" in dataset_id:
+        download_hub(raw_dir, dataset_id)
+    elif "pusht" in dataset_id:
+        download_pusht(raw_dir)
+    elif "xarm" in dataset_id:
+        download_xarm(raw_dir)
+    elif "umi" in dataset_id:
+        download_umi(raw_dir)
+    else:
+        raise ValueError(dataset_id)
+
+
+def download_and_extract_zip(url: str, destination_folder: Path) -> bool:
+    import zipfile
+
+    import requests
+
+    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)
+
+        zip_file = io.BytesIO()
+        for chunk in response.iter_content(chunk_size=1024):
+            if chunk:
+                zip_file.write(chunk)
+                progress_bar.update(len(chunk))
+
+        progress_bar.close()
+
+        zip_file.seek(0)
+
+        with zipfile.ZipFile(zip_file, "r") as zip_ref:
+            zip_ref.extractall(destination_folder)
+
+
+def download_pusht(raw_dir: str):
+    pusht_url = "https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip"
+
+    raw_dir = Path(raw_dir)
+    raw_dir.mkdir(parents=True, exist_ok=True)
+    download_and_extract_zip(pusht_url, raw_dir)
+    # file is created inside a useful "pusht" directory, so we move it out and delete the dir
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    shutil.move(raw_dir / "pusht" / "pusht_cchi_v7_replay.zarr", zarr_path)
+    shutil.rmtree(raw_dir / "pusht")
+
+
+def download_xarm(raw_dir: Path):
+    """Download all xarm datasets at once"""
+    import zipfile
+
+    import gdown
+
+    raw_dir = Path(raw_dir)
+    raw_dir.mkdir(parents=True, exist_ok=True)
+    # from https://github.com/fyhMer/fowm/blob/main/scripts/download_datasets.py
+    url = "https://drive.google.com/uc?id=1nhxpykGtPDhmQKm-_B8zBSywVRdgeVya"
+    zip_path = raw_dir / "data.zip"
+    gdown.download(url, str(zip_path), quiet=False)
+    print("Extracting...")
+    with zipfile.ZipFile(str(zip_path), "r") as zip_f:
+        for pkl_path in zip_f.namelist():
+            if pkl_path.startswith("data/xarm") and pkl_path.endswith(".pkl"):
+                zip_f.extract(member=pkl_path)
+                # move to corresponding raw directory
+                extract_dir = pkl_path.replace("/buffer.pkl", "")
+                raw_pkl_path = raw_dir / "buffer.pkl"
+                shutil.move(pkl_path, raw_pkl_path)
+                shutil.rmtree(extract_dir)
+    zip_path.unlink()
+
+
+def download_hub(raw_dir: Path, dataset_id: str):
+    raw_dir = Path(raw_dir)
+    raw_dir.mkdir(parents=True, exist_ok=True)
+
+    logging.info(f"Start downloading from huggingface.co/cadene for {dataset_id}")
+    snapshot_download(f"cadene/{dataset_id}_raw", repo_type="dataset", local_dir=raw_dir)
+    logging.info(f"Finish downloading from huggingface.co/cadene for {dataset_id}")
+
+
+def download_umi(raw_dir: Path):
+    url_cup_in_the_wild = "https://real.stanford.edu/umi/data/zarr_datasets/cup_in_the_wild.zarr.zip"
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+
+    raw_dir = Path(raw_dir)
+    raw_dir.mkdir(parents=True, exist_ok=True)
+    download_and_extract_zip(url_cup_in_the_wild, zarr_path)
+
+
+if __name__ == "__main__":
+    data_dir = Path("data")
+    dataset_ids = [
+        "pusht_image",
+        "xarm_lift_medium_image",
+        "xarm_lift_medium_replay_image",
+        "xarm_push_medium_image",
+        "xarm_push_medium_replay_image",
+        "aloha_sim_insertion_human_image",
+        "aloha_sim_insertion_scripted_image",
+        "aloha_sim_transfer_cube_human_image",
+        "aloha_sim_transfer_cube_scripted_image",
+        "pusht",
+        "xarm_lift_medium",
+        "xarm_lift_medium_replay",
+        "xarm_push_medium",
+        "xarm_push_medium_replay",
+        "aloha_sim_insertion_human",
+        "aloha_sim_insertion_scripted",
+        "aloha_sim_transfer_cube_human",
+        "aloha_sim_transfer_cube_scripted",
+        "aloha_mobile_cabinet",
+        "aloha_mobile_chair",
+        "aloha_mobile_elevator",
+        "aloha_mobile_shrimp",
+        "aloha_mobile_wash_pan",
+        "aloha_mobile_wipe_wine",
+        "aloha_static_battery",
+        "aloha_static_candy",
+        "aloha_static_coffee",
+        "aloha_static_coffee_new",
+        "aloha_static_cups_open",
+        "aloha_static_fork_pick_up",
+        "aloha_static_pingpong_test",
+        "aloha_static_pro_pencil",
+        "aloha_static_screw_driver",
+        "aloha_static_tape",
+        "aloha_static_thread_velcro",
+        "aloha_static_towel",
+        "aloha_static_vinh_cup",
+        "aloha_static_vinh_cup_left",
+        "aloha_static_ziploc_slide",
+        "umi_cup_in_the_wild",
+    ]
+    for dataset_id in dataset_ids:
+        raw_dir = data_dir / f"{dataset_id}_raw"
+        download_raw(raw_dir, dataset_id)

lerobot/common/datasets/push_dataset_to_hub/_umi_imagecodecs_numcodecs.py

@@ -0,0 +1,326 @@
+#!/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.
+# imagecodecs/numcodecs.py
+
+# Copyright (c) 2021-2022, Christoph Gohlke
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# 1. Redistributions of source code must retain the above copyright notice,
+#    this list of conditions and the following disclaimer.
+#
+# 2. Redistributions in binary form must reproduce the above copyright notice,
+#    this list of conditions and the following disclaimer in the documentation
+#    and/or other materials provided with the distribution.
+#
+# 3. Neither the name of the copyright holder nor the names of its
+#    contributors may be used to endorse or promote products derived from
+#    this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+# POSSIBILITY OF SUCH DAMAGE.
+
+# Copied from: https://github.com/real-stanford/universal_manipulation_interface/blob/298776ce251f33b6b3185a98d6e7d1f9ad49168b/diffusion_policy/codecs/imagecodecs_numcodecs.py#L1
+"""Additional numcodecs implemented using imagecodecs."""
+
+__version__ = "2022.9.26"
+
+__all__ = ("register_codecs",)
+
+import imagecodecs
+import numpy
+from numcodecs.abc import Codec
+from numcodecs.registry import get_codec, register_codec
+
+# TODO (azouitine): Remove useless codecs
+
+
+def protective_squeeze(x: numpy.ndarray):
+    """
+    Squeeze dim only if it's not the last dim.
+    Image dim expected to be *, H, W, C
+    """
+    img_shape = x.shape[-3:]
+    if len(x.shape) > 3:
+        n_imgs = numpy.prod(x.shape[:-3])
+        if n_imgs > 1:
+            img_shape = (-1,) + img_shape
+    return x.reshape(img_shape)
+
+
+def get_default_image_compressor(**kwargs):
+    if imagecodecs.JPEGXL:
+        # has JPEGXL
+        this_kwargs = {
+            "effort": 3,
+            "distance": 0.3,
+            # bug in libjxl, invalid codestream for non-lossless
+            # when decoding speed > 1
+            "decodingspeed": 1,
+        }
+        this_kwargs.update(kwargs)
+        return JpegXl(**this_kwargs)
+    else:
+        this_kwargs = {"level": 50}
+        this_kwargs.update(kwargs)
+        return Jpeg2k(**this_kwargs)
+
+
+class Jpeg2k(Codec):
+    """JPEG 2000 codec for numcodecs."""
+
+    codec_id = "imagecodecs_jpeg2k"
+
+    def __init__(
+        self,
+        level=None,
+        codecformat=None,
+        colorspace=None,
+        tile=None,
+        reversible=None,
+        bitspersample=None,
+        resolutions=None,
+        numthreads=None,
+        verbose=0,
+    ):
+        self.level = level
+        self.codecformat = codecformat
+        self.colorspace = colorspace
+        self.tile = None if tile is None else tuple(tile)
+        self.reversible = reversible
+        self.bitspersample = bitspersample
+        self.resolutions = resolutions
+        self.numthreads = numthreads
+        self.verbose = verbose
+
+    def encode(self, buf):
+        buf = protective_squeeze(numpy.asarray(buf))
+        return imagecodecs.jpeg2k_encode(
+            buf,
+            level=self.level,
+            codecformat=self.codecformat,
+            colorspace=self.colorspace,
+            tile=self.tile,
+            reversible=self.reversible,
+            bitspersample=self.bitspersample,
+            resolutions=self.resolutions,
+            numthreads=self.numthreads,
+            verbose=self.verbose,
+        )
+
+    def decode(self, buf, out=None):
+        return imagecodecs.jpeg2k_decode(buf, verbose=self.verbose, numthreads=self.numthreads, out=out)
+
+
+class JpegXl(Codec):
+    """JPEG XL codec for numcodecs."""
+
+    codec_id = "imagecodecs_jpegxl"
+
+    def __init__(
+        self,
+        # encode
+        level=None,
+        effort=None,
+        distance=None,
+        lossless=None,
+        decodingspeed=None,
+        photometric=None,
+        planar=None,
+        usecontainer=None,
+        # decode
+        index=None,
+        keeporientation=None,
+        # both
+        numthreads=None,
+    ):
+        """
+        Return JPEG XL image from numpy array.
+        Float must be in nominal range 0..1.
+
+        Currently L, LA, RGB, RGBA images are supported in contig mode.
+        Extra channels are only supported for grayscale images in planar mode.
+
+        Parameters
+        ----------
+        level : Default to None, i.e. not overwriting lossess and decodingspeed options.
+            When < 0: Use lossless compression
+            When in [0,1,2,3,4]: Sets the decoding speed tier for the provided options.
+                Minimum is 0 (slowest to decode, best quality/density), and maximum
+                is 4 (fastest to decode, at the cost of some quality/density).
+        effort : Default to 3.
+            Sets encoder effort/speed level without affecting decoding speed.
+            Valid values are, from faster to slower speed: 1:lightning 2:thunder
+                3:falcon 4:cheetah 5:hare 6:wombat 7:squirrel 8:kitten 9:tortoise.
+            Speed: lightning, thunder, falcon, cheetah, hare, wombat, squirrel, kitten, tortoise
+            control the encoder effort in ascending order.
+            This also affects memory usage: using lower effort will typically reduce memory
+            consumption during encoding.
+            lightning and thunder are fast modes useful for lossless mode (modular).
+            falcon disables all of the following tools.
+            cheetah enables coefficient reordering, context clustering, and heuristics for selecting DCT sizes and quantization steps.
+            hare enables Gaborish filtering, chroma from luma, and an initial estimate of quantization steps.
+            wombat enables error diffusion quantization and full DCT size selection heuristics.
+            squirrel (default) enables dots, patches, and spline detection, and full context clustering.
+            kitten optimizes the adaptive quantization for a psychovisual metric.
+            tortoise enables a more thorough adaptive quantization search.
+        distance : Default to 1.0
+            Sets the distance level for lossy compression: target max butteraugli distance,
+            lower = higher quality. Range: 0 .. 15. 0.0 = mathematically lossless
+            (however, use JxlEncoderSetFrameLossless instead to use true lossless,
+            as setting distance to 0 alone is not the only requirement).
+            1.0 = visually lossless. Recommended range: 0.5 .. 3.0.
+        lossess : Default to False.
+            Use lossess encoding.
+        decodingspeed : Default to 0.
+            Duplicate to level. [0,4]
+        photometric : Return JxlColorSpace value.
+            Default logic is quite complicated but works most of the time.
+            Accepted value:
+                int: [-1,3]
+                str: ['RGB',
+                    'WHITEISZERO', 'MINISWHITE',
+                    'BLACKISZERO', 'MINISBLACK', 'GRAY',
+                    'XYB', 'KNOWN']
+        planar : Enable multi-channel mode.
+            Default to false.
+        usecontainer :
+            Forces the encoder to use the box-based container format (BMFF)
+            even when not necessary.
+            When using JxlEncoderUseBoxes, JxlEncoderStoreJPEGMetadata or
+            JxlEncoderSetCodestreamLevel with level 10, the encoder will
+            automatically also use the container format, it is not necessary
+            to use JxlEncoderUseContainer for those use cases.
+            By default this setting is disabled.
+        index : Selectively decode frames for animation.
+            Default to 0, decode all frames.
+            When set to > 0, decode that frame index only.
+        keeporientation :
+            Enables or disables preserving of as-in-bitstream pixeldata orientation.
+            Some images are encoded with an Orientation tag indicating that the
+            decoder must perform a rotation and/or mirroring to the encoded image data.
+
+            If skip_reorientation is JXL_FALSE (the default): the decoder will apply
+            the transformation from the orientation setting, hence rendering the image
+            according to its specified intent. When producing a JxlBasicInfo, the decoder
+            will always set the orientation field to JXL_ORIENT_IDENTITY (matching the
+            returned pixel data) and also align xsize and ysize so that they correspond
+            to the width and the height of the returned pixel data.
+
+            If skip_reorientation is JXL_TRUE: the decoder will skip applying the
+            transformation from the orientation setting, returning the image in
+            the as-in-bitstream pixeldata orientation. This may be faster to decode
+            since the decoder doesnt have to apply the transformation, but can
+            cause wrong display of the image if the orientation tag is not correctly
+            taken into account by the user.
+
+            By default, this option is disabled, and the returned pixel data is
+            re-oriented according to the images Orientation setting.
+        threads : Default to 1.
+            If <= 0, use all cores.
+            If > 32, clipped to 32.
+        """
+
+        self.level = level
+        self.effort = effort
+        self.distance = distance
+        self.lossless = bool(lossless)
+        self.decodingspeed = decodingspeed
+        self.photometric = photometric
+        self.planar = planar
+        self.usecontainer = usecontainer
+        self.index = index
+        self.keeporientation = keeporientation
+        self.numthreads = numthreads
+
+    def encode(self, buf):
+        # TODO: only squeeze all but last dim
+        buf = protective_squeeze(numpy.asarray(buf))
+        return imagecodecs.jpegxl_encode(
+            buf,
+            level=self.level,
+            effort=self.effort,
+            distance=self.distance,
+            lossless=self.lossless,
+            decodingspeed=self.decodingspeed,
+            photometric=self.photometric,
+            planar=self.planar,
+            usecontainer=self.usecontainer,
+            numthreads=self.numthreads,
+        )
+
+    def decode(self, buf, out=None):
+        return imagecodecs.jpegxl_decode(
+            buf,
+            index=self.index,
+            keeporientation=self.keeporientation,
+            numthreads=self.numthreads,
+            out=out,
+        )
+
+
+def _flat(out):
+    """Return numpy array as contiguous view of bytes if possible."""
+    if out is None:
+        return None
+    view = memoryview(out)
+    if view.readonly or not view.contiguous:
+        return None
+    return view.cast("B")
+
+
+def register_codecs(codecs=None, force=False, verbose=True):
+    """Register codecs in this module with numcodecs."""
+    for name, cls in globals().items():
+        if not hasattr(cls, "codec_id") or name == "Codec":
+            continue
+        if codecs is not None and cls.codec_id not in codecs:
+            continue
+        try:
+            try:  # noqa: SIM105
+                get_codec({"id": cls.codec_id})
+            except TypeError:
+                # registered, but failed
+                pass
+        except ValueError:
+            # not registered yet
+            pass
+        else:
+            if not force:
+                if verbose:
+                    log_warning(f"numcodec {cls.codec_id!r} already registered")
+                continue
+            if verbose:
+                log_warning(f"replacing registered numcodec {cls.codec_id!r}")
+        register_codec(cls)
+
+
+def log_warning(msg, *args, **kwargs):
+    """Log message with level WARNING."""
+    import logging
+
+    logging.getLogger(__name__).warning(msg, *args, **kwargs)

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -0,0 +1,215 @@
+#!/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.
+"""
+Contains utilities to process raw data format of HDF5 files like in: https://github.com/tonyzhaozh/act
+"""
+
+import gc
+import shutil
+from pathlib import Path
+
+import h5py
+import numpy as np
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def get_cameras(hdf5_data):
+    # ignore depth channel, not currently handled
+    # TODO(rcadene): add depth
+    rgb_cameras = [key for key in hdf5_data["/observations/images"].keys() if "depth" not in key]  # noqa: SIM118
+    return rgb_cameras
+
+
+def check_format(raw_dir) -> bool:
+    compressed_images = None
+
+    hdf5_paths = list(raw_dir.glob("episode_*.hdf5"))
+    assert len(hdf5_paths) != 0
+    for hdf5_path in hdf5_paths:
+        with h5py.File(hdf5_path, "r") as data:
+            assert "/action" in data
+            assert "/observations/qpos" in data
+
+            assert data["/action"].ndim == 2
+            assert data["/observations/qpos"].ndim == 2
+
+            num_frames = data["/action"].shape[0]
+            assert num_frames == data["/observations/qpos"].shape[0]
+
+            for camera in get_cameras(data):
+                assert num_frames == data[f"/observations/images/{camera}"].shape[0]
+
+                assert data[f"/observations/images/{camera}"].ndim in [2, 4]
+                if data[f"/observations/images/{camera}"].ndim == 2:
+                    assert compressed_images is None or compressed_images
+                    compressed_images = True
+                else:
+                    assert compressed_images is None or not compressed_images
+                    compressed_images = False
+                    assert data[f"/observations/images/{camera}"].ndim == 4
+                    b, h, w, c = data[f"/observations/images/{camera}"].shape
+                    assert c < h and c < w, f"Expect (h,w,c) image format but ({h=},{w=},{c=}) provided."
+    return compressed_images
+
+
+def load_from_raw(raw_dir, out_dir, fps, video, debug, compressed_images):
+    hdf5_files = list(raw_dir.glob("*.hdf5"))
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+    for ep_idx, ep_path in tqdm.tqdm(enumerate(hdf5_files), total=len(hdf5_files)):
+        with h5py.File(ep_path, "r") as ep:
+            num_frames = ep["/action"].shape[0]
+
+            # last step of demonstration is considered done
+            done = torch.zeros(num_frames, dtype=torch.bool)
+            done[-1] = True
+
+            state = torch.from_numpy(ep["/observations/qpos"][:])
+            action = torch.from_numpy(ep["/action"][:])
+            if "/observations/qvel" in ep:
+                velocity = torch.from_numpy(ep["/observations/qvel"][:])
+            if "/observations/effort" in ep:
+                effort = torch.from_numpy(ep["/observations/effort"][:])
+
+            ep_dict = {}
+
+            for camera in get_cameras(ep):
+                img_key = f"observation.images.{camera}"
+
+                if compressed_images:
+                    import cv2
+
+                    # load one compressed image after the other in RAM and uncompress
+                    imgs_array = []
+                    for data in ep[f"/observations/images/{camera}"]:
+                        imgs_array.append(cv2.imdecode(data, 1))
+                    imgs_array = np.array(imgs_array)
+
+                else:
+                    # load all images in RAM
+                    imgs_array = ep[f"/observations/images/{camera}"][:]
+
+                if video:
+                    # save png images in temporary directory
+                    tmp_imgs_dir = out_dir / "tmp_images"
+                    save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                    # encode images to a mp4 video
+                    fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                    video_path = out_dir / "videos" / fname
+                    encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+                    # clean temporary images directory
+                    shutil.rmtree(tmp_imgs_dir)
+
+                    # store the reference to the video frame
+                    ep_dict[img_key] = [
+                        {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                    ]
+                else:
+                    ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+            ep_dict["observation.state"] = state
+            if "/observations/velocity" in ep:
+                ep_dict["observation.velocity"] = velocity
+            if "/observations/effort" in ep:
+                ep_dict["observation.effort"] = effort
+            ep_dict["action"] = action
+            ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+            ep_dict["next.done"] = done
+            # TODO(rcadene): add reward and success by computing them in sim
+
+            assert isinstance(ep_idx, int)
+            ep_dicts.append(ep_dict)
+
+            episode_data_index["from"].append(id_from)
+            episode_data_index["to"].append(id_from + num_frames)
+
+        id_from += num_frames
+
+        gc.collect()
+
+        # process first episode only
+        if debug:
+            break
+
+    data_dict = concatenate_episodes(ep_dicts)
+    return data_dict, episode_data_index
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        if video:
+            features[key] = VideoFrame()
+        else:
+            features[key] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if "observation.velocity" in data_dict:
+        features["observation.velocity"] = Sequence(
+            length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    if "observation.effort" in data_dict:
+        features["observation.effort"] = Sequence(
+            length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    compressed_images = check_format(raw_dir)
+
+    if fps is None:
+        fps = 50
+
+    data_dir, episode_data_index = load_from_raw(raw_dir, out_dir, fps, video, debug, compressed_images)
+    hf_dataset = to_hf_dataset(data_dir, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/compute_stats.py

@@ -0,0 +1,161 @@
+#!/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 copy import deepcopy
+from math import ceil
+
+import datasets
+import einops
+import torch
+import tqdm
+from datasets import Image
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def get_stats_einops_patterns(dataset: LeRobotDataset | datasets.Dataset, num_workers=0):
+    """These einops patterns will be used to aggregate batches and compute statistics.
+
+    Note: We assume the images are in channel first format
+    """
+
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=num_workers,
+        batch_size=2,
+        shuffle=False,
+    )
+    batch = next(iter(dataloader))
+
+    stats_patterns = {}
+    for key, feats_type in dataset.features.items():
+        # sanity check that tensors are not float64
+        assert batch[key].dtype != torch.float64
+
+        if isinstance(feats_type, (VideoFrame, Image)):
+            # sanity check that images are channel first
+            _, c, h, w = batch[key].shape
+            assert c < h and c < w, f"expect channel first images, but instead {batch[key].shape}"
+
+            # sanity check that images are float32 in range [0,1]
+            assert batch[key].dtype == torch.float32, f"expect torch.float32, but instead {batch[key].dtype=}"
+            assert batch[key].max() <= 1, f"expect pixels lower than 1, but instead {batch[key].max()=}"
+            assert batch[key].min() >= 0, f"expect pixels greater than 1, but instead {batch[key].min()=}"
+
+            stats_patterns[key] = "b c h w -> c 1 1"
+        elif batch[key].ndim == 2:
+            stats_patterns[key] = "b c -> c "
+        elif batch[key].ndim == 1:
+            stats_patterns[key] = "b -> 1"
+        else:
+            raise ValueError(f"{key}, {feats_type}, {batch[key].shape}")
+
+    return stats_patterns
+
+
+def compute_stats(
+    dataset: LeRobotDataset | datasets.Dataset, batch_size=32, num_workers=16, max_num_samples=None
+):
+    if max_num_samples is None:
+        max_num_samples = len(dataset)
+
+    # for more info on why we need to set the same number of workers, see `load_from_videos`
+    stats_patterns = get_stats_einops_patterns(dataset, num_workers)
+
+    # mean and std will be computed incrementally while max and min will track the running value.
+    mean, std, max, min = {}, {}, {}, {}
+    for key in stats_patterns:
+        mean[key] = torch.tensor(0.0).float()
+        std[key] = torch.tensor(0.0).float()
+        max[key] = torch.tensor(-float("inf")).float()
+        min[key] = torch.tensor(float("inf")).float()
+
+    def create_seeded_dataloader(dataset, batch_size, seed):
+        generator = torch.Generator()
+        generator.manual_seed(seed)
+        dataloader = torch.utils.data.DataLoader(
+            dataset,
+            num_workers=num_workers,
+            batch_size=batch_size,
+            shuffle=True,
+            drop_last=False,
+            generator=generator,
+        )
+        return dataloader
+
+    # Note: Due to be refactored soon. The point of storing `first_batch` is to make sure we don't get
+    # surprises when rerunning the sampler.
+    first_batch = None
+    running_item_count = 0  # for online mean computation
+    dataloader = create_seeded_dataloader(dataset, batch_size, seed=1337)
+    for i, batch in enumerate(
+        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute mean, min, max")
+    ):
+        this_batch_size = len(batch["index"])
+        running_item_count += this_batch_size
+        if first_batch is None:
+            first_batch = deepcopy(batch)
+        for key, pattern in stats_patterns.items():
+            batch[key] = batch[key].float()
+            # Numerically stable update step for mean computation.
+            batch_mean = einops.reduce(batch[key], pattern, "mean")
+            # Hint: to update the mean we need x̄ₙ = (Nₙ₋₁x̄ₙ₋₁ + Bₙxₙ) / Nₙ, where the subscript represents
+            # the update step, N is the running item count, B is this batch size, x̄ is the running mean,
+            # and x is the current batch mean. Some rearrangement is then required to avoid risking
+            # numerical overflow. Another hint: Nₙ₋₁ = Nₙ - Bₙ. Rearrangement yields
+            # x̄ₙ = x̄ₙ₋₁ + Bₙ * (xₙ - x̄ₙ₋₁) / Nₙ
+            mean[key] = mean[key] + this_batch_size * (batch_mean - mean[key]) / running_item_count
+            max[key] = torch.maximum(max[key], einops.reduce(batch[key], pattern, "max"))
+            min[key] = torch.minimum(min[key], einops.reduce(batch[key], pattern, "min"))
+
+        if i == ceil(max_num_samples / batch_size) - 1:
+            break
+
+    first_batch_ = None
+    running_item_count = 0  # for online std computation
+    dataloader = create_seeded_dataloader(dataset, batch_size, seed=1337)
+    for i, batch in enumerate(
+        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute std")
+    ):
+        this_batch_size = len(batch["index"])
+        running_item_count += this_batch_size
+        # Sanity check to make sure the batches are still in the same order as before.
+        if first_batch_ is None:
+            first_batch_ = deepcopy(batch)
+            for key in stats_patterns:
+                assert torch.equal(first_batch_[key], first_batch[key])
+        for key, pattern in stats_patterns.items():
+            batch[key] = batch[key].float()
+            # Numerically stable update step for mean computation (where the mean is over squared
+            # residuals).See notes in the mean computation loop above.
+            batch_std = einops.reduce((batch[key] - mean[key]) ** 2, pattern, "mean")
+            std[key] = std[key] + this_batch_size * (batch_std - std[key]) / running_item_count
+
+        if i == ceil(max_num_samples / batch_size) - 1:
+            break
+
+    for key in stats_patterns:
+        std[key] = torch.sqrt(std[key])
+
+    stats = {}
+    for key in stats_patterns:
+        stats[key] = {
+            "mean": mean[key],
+            "std": std[key],
+            "max": max[key],
+            "min": min[key],
+        }
+    return stats

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -0,0 +1,229 @@
+#!/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.
+"""Process zarr files formatted like in: https://github.com/real-stanford/diffusion_policy"""
+
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir):
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    required_datasets = {
+        "data/action",
+        "data/img",
+        "data/keypoint",
+        "data/n_contacts",
+        "data/state",
+        "meta/episode_ends",
+    }
+    for dataset in required_datasets:
+        assert dataset in zarr_data
+    nb_frames = zarr_data["data/img"].shape[0]
+
+    required_datasets.remove("meta/episode_ends")
+
+    assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
+
+
+def load_from_raw(raw_dir, out_dir, fps, video, debug):
+    try:
+        import pymunk
+        from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
+
+        from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
+            ReplayBuffer as DiffusionPolicyReplayBuffer,
+        )
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+    # as define in gmy-pusht env: https://github.com/huggingface/gym-pusht/blob/e0684ff988d223808c0a9dcfaba9dc4991791370/gym_pusht/envs/pusht.py#L174
+    success_threshold = 0.95  # 95% coverage,
+
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
+
+    episode_ids = torch.from_numpy(zarr_data.get_episode_idxs())
+    num_episodes = zarr_data.meta["episode_ends"].shape[0]
+    assert len(
+        {zarr_data[key].shape[0] for key in zarr_data.keys()}  # noqa: SIM118
+    ), "Some data type dont have the same number of total frames."
+
+    # TODO(rcadene): 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 = PushTEnv.get_goal_pose_body(goal_pos_angle)
+
+    imgs = torch.from_numpy(zarr_data["img"])  # b h w c
+    states = torch.from_numpy(zarr_data["state"])
+    actions = torch.from_numpy(zarr_data["action"])
+
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+    for ep_idx in tqdm.tqdm(range(num_episodes)):
+        id_to = zarr_data.meta["episode_ends"][ep_idx]
+        num_frames = id_to - id_from
+
+        # sanity check
+        assert (episode_ids[id_from:id_to] == ep_idx).all()
+
+        # get image
+        image = imgs[id_from:id_to]
+        assert image.min() >= 0.0
+        assert image.max() <= 255.0
+        image = image.type(torch.uint8)
+
+        # get state
+        state = states[id_from:id_to]
+        agent_pos = state[:, :2]
+        block_pos = state[:, 2:4]
+        block_angle = state[:, 4]
+
+        # get reward, success, done
+        reward = torch.zeros(num_frames)
+        success = torch.zeros(num_frames, dtype=torch.bool)
+        done = torch.zeros(num_frames, dtype=torch.bool)
+        for i in range(num_frames):
+            space = pymunk.Space()
+            space.gravity = 0, 0
+            space.damping = 0
+
+            # Add walls.
+            walls = [
+                PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
+                PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
+                PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
+                PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
+            ]
+            space.add(*walls)
+
+            block_body = PushTEnv.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_dict = {}
+
+        imgs_array = [x.numpy() for x in image]
+        img_key = "observation.image"
+        if video:
+            # save png images in temporary directory
+            tmp_imgs_dir = out_dir / "tmp_images"
+            save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+            # encode images to a mp4 video
+            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+            video_path = out_dir / "videos" / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+            # clean temporary images directory
+            shutil.rmtree(tmp_imgs_dir)
+
+            # store the reference to the video frame
+            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+        else:
+            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = agent_pos
+        ep_dict["action"] = actions[id_from:id_to]
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        # ep_dict["next.observation.image"] = image[1:],
+        # ep_dict["next.observation.state"] = agent_pos[1:],
+        # TODO(rcadene)] = verify that reward and done are aligned with image and agent_pos
+        ep_dict["next.reward"] = torch.cat([reward[1:], reward[[-1]]])
+        ep_dict["next.done"] = torch.cat([done[1:], done[[-1]]])
+        ep_dict["next.success"] = torch.cat([success[1:], success[[-1]]])
+        ep_dicts.append(ep_dict)
+
+        episode_data_index["from"].append(id_from)
+        episode_data_index["to"].append(id_from + num_frames)
+
+        id_from += num_frames
+
+        # process first episode only
+        if debug:
+            break
+
+    data_dict = concatenate_episodes(ep_dicts)
+    return data_dict, episode_data_index
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["next.success"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 10
+
+    data_dict, episode_data_index = load_from_raw(raw_dir, out_dir, fps, video, debug)
+    hf_dataset = to_hf_dataset(data_dict, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -0,0 +1,222 @@
+#!/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.
+"""Process UMI (Universal Manipulation Interface) data stored in Zarr format like in: https://github.com/real-stanford/universal_manipulation_interface"""
+
+import logging
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.push_dataset_to_hub._umi_imagecodecs_numcodecs import register_codecs
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir) -> bool:
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    required_datasets = {
+        "data/robot0_demo_end_pose",
+        "data/robot0_demo_start_pose",
+        "data/robot0_eef_pos",
+        "data/robot0_eef_rot_axis_angle",
+        "data/robot0_gripper_width",
+        "meta/episode_ends",
+        "data/camera0_rgb",
+    }
+    for dataset in required_datasets:
+        if dataset not in zarr_data:
+            return False
+
+    # mandatory to access zarr_data
+    register_codecs()
+    nb_frames = zarr_data["data/camera0_rgb"].shape[0]
+
+    required_datasets.remove("meta/episode_ends")
+    assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
+
+
+def get_episode_idxs(episode_ends: np.ndarray) -> np.ndarray:
+    # Optimized and simplified version of this function: https://github.com/real-stanford/universal_manipulation_interface/blob/298776ce251f33b6b3185a98d6e7d1f9ad49168b/diffusion_policy/common/replay_buffer.py#L374
+    from numba import jit
+
+    @jit(nopython=True)
+    def _get_episode_idxs(episode_ends):
+        result = np.zeros((episode_ends[-1],), dtype=np.int64)
+        start_idx = 0
+        for episode_number, end_idx in enumerate(episode_ends):
+            result[start_idx:end_idx] = episode_number
+            start_idx = end_idx
+        return result
+
+    return _get_episode_idxs(episode_ends)
+
+
+def load_from_raw(raw_dir, out_dir, fps, video, debug):
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    # We process the image data separately because it is too large to fit in memory
+    end_pose = torch.from_numpy(zarr_data["data/robot0_demo_end_pose"][:])
+    start_pos = torch.from_numpy(zarr_data["data/robot0_demo_start_pose"][:])
+    eff_pos = torch.from_numpy(zarr_data["data/robot0_eef_pos"][:])
+    eff_rot_axis_angle = torch.from_numpy(zarr_data["data/robot0_eef_rot_axis_angle"][:])
+    gripper_width = torch.from_numpy(zarr_data["data/robot0_gripper_width"][:])
+
+    states_pos = torch.cat([eff_pos, eff_rot_axis_angle], dim=1)
+    states = torch.cat([states_pos, gripper_width], dim=1)
+
+    episode_ends = zarr_data["meta/episode_ends"][:]
+    num_episodes = episode_ends.shape[0]
+
+    episode_ids = torch.from_numpy(get_episode_idxs(episode_ends))
+
+    # We convert it in torch tensor later because the jit function does not support torch tensors
+    episode_ends = torch.from_numpy(episode_ends)
+
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+    for ep_idx in tqdm.tqdm(range(num_episodes)):
+        id_to = episode_ends[ep_idx]
+        num_frames = id_to - id_from
+
+        # sanity heck
+        assert (episode_ids[id_from:id_to] == ep_idx).all()
+
+        # TODO(rcadene): save temporary images of the episode?
+
+        state = states[id_from:id_to]
+
+        ep_dict = {}
+
+        # load 57MB of images in RAM (400x224x224x3 uint8)
+        imgs_array = zarr_data["data/camera0_rgb"][id_from:id_to]
+        img_key = "observation.image"
+        if video:
+            # save png images in temporary directory
+            tmp_imgs_dir = out_dir / "tmp_images"
+            save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+            # encode images to a mp4 video
+            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+            video_path = out_dir / "videos" / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+            # clean temporary images directory
+            shutil.rmtree(tmp_imgs_dir)
+
+            # store the reference to the video frame
+            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+        else:
+            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = state
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        ep_dict["episode_data_index_from"] = torch.tensor([id_from] * num_frames)
+        ep_dict["episode_data_index_to"] = torch.tensor([id_from + num_frames] * num_frames)
+        ep_dict["end_pose"] = end_pose[id_from:id_to]
+        ep_dict["start_pos"] = start_pos[id_from:id_to]
+        ep_dict["gripper_width"] = gripper_width[id_from:id_to]
+        ep_dicts.append(ep_dict)
+
+        episode_data_index["from"].append(id_from)
+        episode_data_index["to"].append(id_from + num_frames)
+        id_from += num_frames
+
+        # process first episode only
+        if debug:
+            break
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = id_from
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+
+    return data_dict, episode_data_index
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+    features["episode_data_index_from"] = Value(dtype="int64", id=None)
+    features["episode_data_index_to"] = Value(dtype="int64", id=None)
+    # `start_pos` and `end_pos` respectively represent the positions of the end-effector
+    # at the beginning and the end of the episode.
+    # `gripper_width` indicates the distance between the grippers, and this value is included
+    # in the state vector, which comprises the concatenation of the end-effector position
+    # and gripper width.
+    features["end_pose"] = Sequence(
+        length=data_dict["end_pose"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["start_pos"] = Sequence(
+        length=data_dict["start_pos"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["gripper_width"] = Sequence(
+        length=data_dict["gripper_width"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        # For umi cup in the wild: https://arxiv.org/pdf/2402.10329#table.caption.16
+        fps = 10
+
+    if not video:
+        logging.warning(
+            "Generating UMI dataset without `video=True` creates ~150GB on disk and requires ~80GB in RAM."
+        )
+
+    data_dict, episode_data_index = load_from_raw(raw_dir, out_dir, fps, video, debug)
+    hf_dataset = to_hf_dataset(data_dict, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -0,0 +1,53 @@
+#!/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 concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+import numpy
+import PIL
+import torch
+
+
+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)]

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.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.
+"""Process pickle files formatted like in: https://github.com/fyhMer/fowm"""
+
+import pickle
+import shutil
+from pathlib import Path
+
+import einops
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir):
+    keys = {"actions", "rewards", "dones"}
+    nested_keys = {"observations": {"rgb", "state"}, "next_observations": {"rgb", "state"}}
+
+    xarm_files = list(raw_dir.glob("*.pkl"))
+    assert len(xarm_files) > 0
+
+    with open(xarm_files[0], "rb") as f:
+        dataset_dict = pickle.load(f)
+
+    assert isinstance(dataset_dict, dict)
+    assert all(k in dataset_dict for k in keys)
+
+    # Check for consistent lengths in nested keys
+    expected_len = len(dataset_dict["actions"])
+    assert all(len(dataset_dict[key]) == expected_len for key in keys if key in dataset_dict)
+
+    for key, subkeys in nested_keys.items():
+        nested_dict = dataset_dict.get(key, {})
+        assert all(len(nested_dict[subkey]) == expected_len for subkey in subkeys if subkey in nested_dict)
+
+
+def load_from_raw(raw_dir, out_dir, fps, video, debug):
+    pkl_path = raw_dir / "buffer.pkl"
+
+    with open(pkl_path, "rb") as f:
+        pkl_data = pickle.load(f)
+
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+    id_to = 0
+    ep_idx = 0
+    total_frames = pkl_data["actions"].shape[0]
+    for i in tqdm.tqdm(range(total_frames)):
+        id_to += 1
+
+        if not pkl_data["dones"][i]:
+            continue
+
+        num_frames = id_to - id_from
+
+        image = torch.tensor(pkl_data["observations"]["rgb"][id_from:id_to])
+        image = einops.rearrange(image, "b c h w -> b h w c")
+        state = torch.tensor(pkl_data["observations"]["state"][id_from:id_to])
+        action = torch.tensor(pkl_data["actions"][id_from:id_to])
+        # TODO(rcadene): we have a missing last frame which is the observation when the env is done
+        # it is critical to have this frame for tdmpc to predict a "done observation/state"
+        # next_image = torch.tensor(pkl_data["next_observations"]["rgb"][id_from:id_to])
+        # next_state = torch.tensor(pkl_data["next_observations"]["state"][id_from:id_to])
+        next_reward = torch.tensor(pkl_data["rewards"][id_from:id_to])
+        next_done = torch.tensor(pkl_data["dones"][id_from:id_to])
+
+        ep_dict = {}
+
+        imgs_array = [x.numpy() for x in image]
+        img_key = "observation.image"
+        if video:
+            # save png images in temporary directory
+            tmp_imgs_dir = out_dir / "tmp_images"
+            save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+            # encode images to a mp4 video
+            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+            video_path = out_dir / "videos" / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+            # clean temporary images directory
+            shutil.rmtree(tmp_imgs_dir)
+
+            # store the reference to the video frame
+            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+        else:
+            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = state
+        ep_dict["action"] = action
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        # ep_dict["next.observation.image"] = next_image
+        # ep_dict["next.observation.state"] = next_state
+        ep_dict["next.reward"] = next_reward
+        ep_dict["next.done"] = next_done
+        ep_dicts.append(ep_dict)
+
+        episode_data_index["from"].append(id_from)
+        episode_data_index["to"].append(id_from + num_frames)
+
+        id_from = id_to
+        ep_idx += 1
+
+        # process first episode only
+        if debug:
+            break
+
+    data_dict = concatenate_episodes(ep_dicts)
+    return data_dict, episode_data_index
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+    # TODO(rcadene): add success
+    # features["next.success"] = Value(dtype='bool', id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 15
+
+    data_dict, episode_data_index = load_from_raw(raw_dir, out_dir, fps, video, debug)
+    hf_dataset = to_hf_dataset(data_dict, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

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/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/utils.py

@@ -1,30 +1,354 @@
-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 json
+import re
 from pathlib import Path
+from typing import Dict
 
-import requests
-import tqdm
+import datasets
+import torch
+from datasets import load_dataset, load_from_disk
+from huggingface_hub import hf_hub_download, snapshot_download
+from PIL import Image as PILImage
+from safetensors.torch import load_file
+from torchvision import transforms
 
 
-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, parent_key="", sep="/"):
+    """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, sep="/"):
+    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)
-        return True
+
+def hf_transform_to_torch(items_dict):
+    """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 isinstance(first_item, dict) and "path" in first_item and "timestamp" in first_item:
+            # video frame will be processed downstream
+            pass
+        else:
+            items_dict[key] = [torch.tensor(x) for x in items_dict[key]]
+    return items_dict
+
+
+def load_hf_dataset(repo_id, version, root, split) -> datasets.Dataset:
+    """hf_dataset contains all the observations, states, actions, rewards, etc."""
+    if root is not None:
+        hf_dataset = load_from_disk(str(Path(root) / repo_id / "train"))
+        # TODO(rcadene): clean this which enables getting a subset of dataset
+        if split != "train":
+            if "%" in split:
+                raise NotImplementedError(f"We dont support splitting based on percentage for now ({split}).")
+            match_from = re.search(r"train\[(\d+):\]", split)
+            match_to = re.search(r"train\[:(\d+)\]", split)
+            if match_from:
+                from_frame_index = int(match_from.group(1))
+                hf_dataset = hf_dataset.select(range(from_frame_index, len(hf_dataset)))
+            elif match_to:
+                to_frame_index = int(match_to.group(1))
+                hf_dataset = hf_dataset.select(range(to_frame_index))
+            else:
+                raise ValueError(
+                    f'`split` ({split}) should either be "train", "train[INT:]", or "train[:INT]"'
+                )
     else:
-        return False
+        hf_dataset = load_dataset(repo_id, revision=version, split=split)
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def load_episode_data_index(repo_id, version, root) -> dict[str, torch.Tensor]:
+    """episode_data_index contains the range of indices for each episode
+
+    Example:
+    ```python
+    from_id = episode_data_index["from"][episode_id].item()
+    to_id = episode_data_index["to"][episode_id].item()
+    episode_frames = [dataset[i] for i in range(from_id, to_id)]
+    ```
+    """
+    if root is not None:
+        path = Path(root) / repo_id / "meta_data" / "episode_data_index.safetensors"
+    else:
+        path = hf_hub_download(
+            repo_id, "meta_data/episode_data_index.safetensors", repo_type="dataset", revision=version
+        )
+
+    return load_file(path)
+
+
+def load_stats(repo_id, version, root) -> dict[str, dict[str, torch.Tensor]]:
+    """stats contains the statistics per modality computed over the full dataset, such as max, min, mean, std
+
+    Example:
+    ```python
+    normalized_action = (action - stats["action"]["mean"]) / stats["action"]["std"]
+    ```
+    """
+    if root is not None:
+        path = Path(root) / repo_id / "meta_data" / "stats.safetensors"
+    else:
+        path = hf_hub_download(repo_id, "meta_data/stats.safetensors", repo_type="dataset", revision=version)
+
+    stats = load_file(path)
+    return unflatten_dict(stats)
+
+
+def load_info(repo_id, version, root) -> dict:
+    """info contains useful information regarding the dataset that are not stored elsewhere
+
+    Example:
+    ```python
+    print("frame per second used to collect the video", info["fps"])
+    ```
+    """
+    if root is not None:
+        path = Path(root) / repo_id / "meta_data" / "info.json"
+    else:
+        path = hf_hub_download(repo_id, "meta_data/info.json", repo_type="dataset", revision=version)
+
+    with open(path) as f:
+        info = json.load(f)
+    return info
+
+
+def load_videos(repo_id, version, root) -> Path:
+    if root is not None:
+        path = Path(root) / repo_id / "videos"
+    else:
+        # TODO(rcadene): we download the whole repo here. see if we can avoid this
+        repo_dir = snapshot_download(repo_id, repo_type="dataset", revision=version)
+        path = Path(repo_dir) / "videos"
+
+    return path
+
+
+def load_previous_and_future_frames(
+    item: dict[str, torch.Tensor],
+    hf_dataset: datasets.Dataset,
+    episode_data_index: dict[str, torch.Tensor],
+    delta_timestamps: dict[str, list[float]],
+    tolerance_s: float,
+) -> dict[torch.Tensor]:
+    """
+    Given a current item in the dataset containing a timestamp (e.g. 0.6 seconds), and a list of time differences of
+    some modalities (e.g. delta_timestamps={"observation.image": [-0.8, -0.2, 0, 0.2]}), this function computes for each
+    given modality (e.g. "observation.image") a list of query timestamps (e.g. [-0.2, 0.4, 0.6, 0.8]) and loads the closest
+    frames in the dataset.
+
+    Importantly, when no frame can be found around a query timestamp within a specified tolerance window, this function
+    raises an AssertionError. When a timestamp is queried before the first available timestamp of the episode or after
+    the last available timestamp, the violation of the tolerance doesnt raise an AssertionError, and the function
+    populates a boolean array indicating which frames are outside of the episode range. For instance, this boolean array
+    is useful during batched training to not supervise actions associated to timestamps coming after the end of the
+    episode, or to pad the observations in a specific way. Note that by default the observation frames before the start
+    of the episode are the same as the first frame of the episode.
+
+    Parameters:
+    - item (dict): A dictionary containing all the data related to a frame. It is the result of `dataset[idx]`. Each key
+      corresponds to a different modality (e.g., "timestamp", "observation.image", "action").
+    - hf_dataset (datasets.Dataset): A dictionary containing the full dataset. Each key corresponds to a different
+      modality (e.g., "timestamp", "observation.image", "action").
+    - episode_data_index (dict): A dictionary containing two keys ("from" and "to") associated to dataset indices.
+      They indicate the start index and end index of each episode in the dataset.
+    - delta_timestamps (dict): A dictionary containing lists of delta timestamps for each possible modality to be
+      retrieved. These deltas are added to the item timestamp to form the query timestamps.
+    - tolerance_s (float, optional): The tolerance level (in seconds) used to determine if a data point is close enough to the query
+      timestamp by asserting `tol > difference`. It is suggested to set `tol` to a smaller value than the
+      smallest expected inter-frame period, but large enough to account for jitter.
+
+    Returns:
+    - The same item with the queried frames for each modality specified in delta_timestamps, with an additional key for
+      each modality (e.g. "observation.image_is_pad").
+
+    Raises:
+    - AssertionError: If any of the frames unexpectedly violate the tolerance level. This could indicate synchronization
+      issues with timestamps during data collection.
+    """
+    # get indices of the frames associated to the episode, and their timestamps
+    ep_id = item["episode_index"].item()
+    ep_data_id_from = episode_data_index["from"][ep_id].item()
+    ep_data_id_to = episode_data_index["to"][ep_id].item()
+    ep_data_ids = torch.arange(ep_data_id_from, ep_data_id_to, 1)
+
+    # load timestamps
+    ep_timestamps = hf_dataset.select_columns("timestamp")[ep_data_id_from:ep_data_id_to]["timestamp"]
+    ep_timestamps = torch.stack(ep_timestamps)
+
+    # we make the assumption that the timestamps are sorted
+    ep_first_ts = ep_timestamps[0]
+    ep_last_ts = ep_timestamps[-1]
+    current_ts = item["timestamp"].item()
+
+    for key in delta_timestamps:
+        # get timestamps used as query to retrieve data of previous/future frames
+        delta_ts = delta_timestamps[key]
+        query_ts = current_ts + torch.tensor(delta_ts)
+
+        # compute distances between each query timestamp and all timestamps of all the frames belonging to the episode
+        dist = torch.cdist(query_ts[:, None], ep_timestamps[:, None], p=1)
+        min_, argmin_ = dist.min(1)
+
+        # TODO(rcadene): synchronize timestamps + interpolation if needed
+
+        is_pad = min_ > tolerance_s
+
+        # check violated query timestamps are all outside the episode range
+        assert ((query_ts[is_pad] < ep_first_ts) | (ep_last_ts < query_ts[is_pad])).all(), (
+            f"One or several timestamps unexpectedly violate the tolerance ({min_} > {tolerance_s=}) inside episode range."
+            "This might be due to synchronization issues with timestamps during data collection."
+        )
+
+        # get dataset indices corresponding to frames to be loaded
+        data_ids = ep_data_ids[argmin_]
+
+        # load frames modality
+        item[key] = hf_dataset.select_columns(key)[data_ids][key]
+
+        if isinstance(item[key][0], dict) and "path" in item[key][0]:
+            # video mode where frame are expressed as dict of path and timestamp
+            item[key] = item[key]
+        else:
+            item[key] = torch.stack(item[key])
+
+        item[f"{key}_is_pad"] = is_pad
+
+    return item
+
+
+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
+
+
+def reset_episode_index(hf_dataset: datasets.Dataset) -> datasets.Dataset:
+    """
+    Reset the `episode_index` of the provided HuggingFace Dataset.
+
+    `episode_data_index` (and related functionality such as `load_previous_and_future_frames`) requires the
+    `episode_index` to be sorted, continuous (1,1,1 and not 1,2,1) and start at 0.
+
+    This brings the `episode_index` to the required format.
+    """
+    if len(hf_dataset) == 0:
+        return hf_dataset
+    unique_episode_idxs = torch.stack(hf_dataset["episode_index"]).unique().tolist()
+    episode_idx_to_reset_idx_mapping = {
+        ep_id: reset_ep_id for reset_ep_id, ep_id in enumerate(unique_episode_idxs)
+    }
+
+    def modify_ep_idx_func(example):
+        example["episode_index"] = episode_idx_to_reset_idx_mapping[example["episode_index"].item()]
+        return example
+
+    hf_dataset = hf_dataset.map(modify_ep_idx_func)
+    return hf_dataset
+
+
+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)

lerobot/common/datasets/video_utils.py

@@ -0,0 +1,202 @@
+#!/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 subprocess
+import warnings
+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
+
+
+def load_from_videos(
+    item: dict[str, torch.Tensor], video_frame_keys: list[str], videos_dir: Path, tolerance_s: float
+):
+    """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
+    in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a Segmentation Fault.
+    This probably happens because a memory reference to the video loader is created in the main process and a
+    subprocess fails to access it.
+    """
+    # since video path already contains "videos" (e.g. videos_dir="data/videos", path="videos/episode_0.mp4")
+    data_dir = videos_dir.parent
+
+    for key in video_frame_keys:
+        if isinstance(item[key], list):
+            # load multiple frames at once (expected when delta_timestamps is not None)
+            timestamps = [frame["timestamp"] for frame in item[key]]
+            paths = [frame["path"] for frame in item[key]]
+            if len(set(paths)) > 1:
+                raise NotImplementedError("All video paths are expected to be the same for now.")
+            video_path = data_dir / paths[0]
+
+            frames = decode_video_frames_torchvision(video_path, timestamps, tolerance_s)
+            item[key] = frames
+        else:
+            # load one frame
+            timestamps = [item[key]["timestamp"]]
+            video_path = data_dir / item[key]["path"]
+
+            frames = decode_video_frames_torchvision(video_path, timestamps, tolerance_s)
+            item[key] = frames[0]
+
+    return item
+
+
+def decode_video_frames_torchvision(
+    video_path: str,
+    timestamps: list[float],
+    tolerance_s: float,
+    device: str = "cpu",
+    log_loaded_timestamps: bool = False,
+):
+    """Loads frames associated to the requested timestamps of a video
+
+    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
+    if device == "cpu":
+        # explicitely use pyav
+        torchvision.set_video_backend("pyav")
+        keyframes_only = True  # pyav doesnt support accuracte seek
+    elif device == "cuda":
+        # TODO(rcadene, aliberts): implement video decoding with GPU
+        # torchvision.set_video_backend("cuda")
+        # torchvision.set_video_backend("video_reader")
+        # requires installing torchvision from source, see: https://github.com/pytorch/vision/blob/main/torchvision/csrc/io/decoder/gpu/README.rst
+        # check possible bug: https://github.com/pytorch/vision/issues/7745
+        raise NotImplementedError(
+            "Video decoding on gpu with cuda is currently not supported. Use `device='cpu'`."
+        )
+    else:
+        raise ValueError(device)
+
+    # 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 = timestamps[0]
+    last_ts = timestamps[-1]
+
+    # access closest key frame of the first requested frame
+    # Note: closest key frame timestamp is usally 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
+
+    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."
+    )
+
+    # 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 encode_video_frames(imgs_dir: Path, video_path: Path, fps: int):
+    """More info on ffmpeg arguments tuning on `lerobot/common/datasets/_video_benchmark/README.md`"""
+    video_path = Path(video_path)
+    video_path.parent.mkdir(parents=True, exist_ok=True)
+
+    ffmpeg_cmd = (
+        f"ffmpeg -r {fps} "
+        "-f image2 "
+        "-loglevel error "
+        f"-i {str(imgs_dir / 'frame_%06d.png')} "
+        "-vcodec libx264 "
+        "-g 2 "
+        "-pix_fmt yuv444p "
+        f"{str(video_path)}"
+    )
+    subprocess.run(ffmpeg_cmd.split(" "), check=True)
+
+
+@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")

lerobot/common/envs/factory.py

@@ -1,62 +1,59 @@
-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 omegaconf import DictConfig
 
 
-def make_env(cfg, transform=None):
+def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv:
+    """Makes a gym vector environment according to the evaluation config.
+
+    n_envs can be used to override eval.batch_size in the configuration. Must be at least 1.
+    """
+    if n_envs is not None and n_envs < 1:
+        raise ValueError("`n_envs must be at least 1")
+
     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,
+        "obs_type": "pixels_agent_pos",
+        "render_mode": "rgb_array",
+        "max_episode_steps": cfg.env.episode_length,
+        "visualization_width": 384,
+        "visualization_height": 384,
     }
 
-    if cfg.env.name == "simxarm":
-        from lerobot.common.envs.simxarm import SimxarmEnv
+    package_name = f"gym_{cfg.env.name}"
 
-        kwargs["task"] = cfg.env.task
-        clsfunc = SimxarmEnv
-    elif cfg.env.name == "pusht":
-        from lerobot.common.envs.pusht import PushtEnv
+    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.env.name}]'`"
+        )
+        raise e
 
-        # 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."
+    gym_handle = f"{package_name}/{cfg.env.task}"
 
-        clsfunc = PushtEnv
-    else:
-        raise ValueError(cfg.env.name)
-
-    env = clsfunc(**kwargs)
-
-    # limit rollout to max_steps
-    env = TransformedEnv(env, StepCounter(max_steps=cfg.env.episode_length))
-
-    if transform is not None:
-        # useful to add normalization
-        env.append_transform(transform)
+    # batched version of the env that returns an observation of shape (b, c)
+    env_cls = gym.vector.AsyncVectorEnv if cfg.eval.use_async_envs else gym.vector.SyncVectorEnv
+    env = env_cls(
+        [
+            lambda: gym.make(gym_handle, disable_env_checker=True, **kwargs)
+            for _ in range(n_envs if n_envs is not None else cfg.eval.batch_size)
+        ]
+    )
 
     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 +1,58 @@
+#!/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 einops
+import numpy as np
+import torch
+from torch import Tensor
+
+
+def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
+    """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 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():
+        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 first images, but instead {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
+
+    # 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

lerobot/common/logger.py

@@ -1,19 +1,41 @@
+#!/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.
+# TODO(rcadene, alexander-soare): clean this file
+"""Borrowed from https://github.com/fyhMer/fowm/blob/main/src/logger.py"""
+
 import logging
 import os
 from pathlib import Path
 
+from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from omegaconf import OmegaConf
 from termcolor import colored
 
+from lerobot.common.policies.policy_protocol import Policy
+
 
 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)]
+    """Return a group name for logging. Optionally returns group name as list."""
     lst = [
+        f"policy:{cfg.policy.name}",
+        f"dataset:{cfg.dataset_repo_id}",
         f"env:{cfg.env.name}",
         f"seed:{cfg.seed}",
     ]
@@ -27,10 +49,11 @@ class Logger:
         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._model_dir = self._log_dir / "checkpoints"
         self._buffer_dir = self._log_dir / "buffers"
-        self._save_model = cfg.save_model
-        self._save_buffer = cfg.save_buffer
+        self._save_model = cfg.training.save_model
+        self._disable_wandb_artifact = cfg.wandb.disable_artifact
+        self._save_buffer = cfg.training.get("save_buffer", False)
         self._group = cfg_to_group(cfg)
         self._seed = cfg.seed
         self._cfg = cfg
@@ -38,7 +61,7 @@ class Logger:
         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
+        run_offline = not enable_wandb or not project
         if run_offline:
             logging.info(colored("Logs will be saved locally.", "yellow", attrs=["bold"]))
             self._wandb = None
@@ -63,28 +86,33 @@ class Logger:
                 resume=None,
             )
             print(colored("Logs will be synced with wandb.", "blue", attrs=["bold"]))
+            logging.info(f"Track this run --> {colored(wandb.run.get_url(), 'yellow', attrs=['bold'])}")
             self._wandb = wandb
 
-    def save_model(self, policy, identifier):
+    def save_model(self, policy: 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:
+            save_dir = self._model_dir / str(identifier)
+            policy.save_pretrained(save_dir)
+            # Also save the full Hydra config for the env configuration.
+            OmegaConf.save(self._cfg, save_dir / "config.yaml")
+            if self._wandb and not self._disable_wandb_artifact:
+                # note wandb artifact does not accept ":" or "/" in its name
                 artifact = self._wandb.Artifact(
-                    self._group + "-" + str(self._seed) + "-" + str(identifier),
+                    f"{self._group.replace(':', '_').replace('/', '_')}-{self._seed}-{identifier}",
                     type="model",
                 )
-                artifact.add_file(fp)
+                artifact.add_file(save_dir / SAFETENSORS_SINGLE_FILE)
                 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:
+        if self._wandb and not self._disable_wandb_artifact:
+            # note wandb artifact does not accept ":" or "/" in its name
             artifact = self._wandb.Artifact(
-                self._group + "-" + str(self._seed) + "-" + str(identifier),
+                f"{self._group.replace(':', '_').replace('/', '_')}-{self._seed}-{identifier}",
                 type="buffer",
             )
             artifact.add_file(fp)
@@ -102,9 +130,14 @@ class Logger:
         assert mode in {"train", "eval"}
         if self._wandb is not None:
             for k, v in d.items():
+                if not isinstance(v, (int, float, str)):
+                    logging.warning(
+                        f'WandB logging of key "{k}" was ignored as its type is not handled by this wrapper.'
+                    )
+                    continue
                 self._wandb.log({f"{mode}/{k}": v}, step=step)
 
-    def log_video(self, video, step, mode="train"):
+    def log_video(self, video_path: str, step: int, mode: str = "train"):
         assert mode in {"train", "eval"}
-        wandb_video = self._wandb.Video(video, fps=self._cfg.fps, format="mp4")
+        wandb_video = self._wandb.Video(video_path, fps=self._cfg.fps, format="mp4")
         self._wandb.log({f"{mode}/video": wandb_video}, step=step)

lerobot/common/policies/act/configuration_act.py

@@ -0,0 +1,157 @@
+#!/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
+
+
+@dataclass
+class ACTConfig:
+    """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`.
+
+    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.images.top" refers to an input from the
+            "top" camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
+            Importantly, 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, 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 initalize 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_momentum: Exponential moving average (EMA) momentum parameter (α) for ensembling
+            actions for a given time step over multiple policy invocations. Updates are calculated as:
+            x⁻ₙ = αx⁻ₙ₋₁ + (1-α)xₙ. Note that the ACT paper and original ACT code describes a different
+            parameter here: they refer to a weighting scheme wᵢ = exp(-m⋅i) and set m = 0.01. With our
+            formulation, this is equivalent to α = exp(-0.01) ≈ 0.99. When this parameter is provided, we
+            require `n_action_steps == 1` (since we need to query the policy every step anyway).
+        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
+
+    input_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "observation.images.top": [3, 480, 640],
+            "observation.state": [14],
+        }
+    )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [14],
+        }
+    )
+
+    # Normalization / Unnormalization
+    input_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {
+            "observation.images.top": "mean_std",
+            "observation.state": "mean_std",
+        }
+    )
+    output_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {
+            "action": "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.
+    temporal_ensemble_momentum: float | None = None
+
+    # Training and loss computation.
+    dropout: float = 0.1
+    kl_weight: float = 10.0
+
+    def __post_init__(self):
+        """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_momentum 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}`"
+            )

lerobot/common/policies/act/modeling_act.py

@@ -0,0 +1,608 @@
+#!/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 huggingface_hub import PyTorchModelHubMixin
+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
+
+
+class ACTPolicy(nn.Module, PyTorchModelHubMixin):
+    """
+    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)
+    """
+
+    name = "act"
+
+    def __init__(
+        self,
+        config: ACTConfig | None = None,
+        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__()
+        if config is None:
+            config = ACTConfig()
+        self.config: ACTConfig = config
+
+        self.normalize_inputs = Normalize(
+            config.input_shapes, config.input_normalization_modes, dataset_stats
+        )
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+
+        self.model = ACT(config)
+
+        self.expected_image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+
+        self.reset()
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        if self.config.temporal_ensemble_momentum is not None:
+            self._ensembled_actions = None
+        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)
+        batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+
+        # If we are doing temporal ensembling, keep track of the exponential moving average (EMA), and return
+        # the first action.
+        if self.config.temporal_ensemble_momentum is not None:
+            actions = self.model(batch)[0]  # (batch_size, chunk_size, action_dim)
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+            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()
+            else:
+                # self._ensembled_actions will have shape (batch_size, chunk_size - 1, action_dim). Compute
+                # the EMA update for those entries.
+                alpha = self.config.temporal_ensemble_momentum
+                self._ensembled_actions = alpha * self._ensembled_actions + (1 - alpha) * actions[:, :-1]
+                # The last action, which has no prior moving average, needs to get concatenated onto the end.
+                self._ensembled_actions = torch.cat([self._ensembled_actions, actions[:, -1:]], dim=1)
+            # "Consume" the first action.
+            action, self._ensembled_actions = self._ensembled_actions[:, 0], self._ensembled_actions[:, 1:]
+            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]) -> dict[str, Tensor]:
+        """Run the batch through the model and compute the loss for training or validation."""
+        batch = self.normalize_inputs(batch)
+        batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+        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_dict["loss"] = l1_loss + mean_kld * self.config.kl_weight
+        else:
+            loss_dict["loss"] = l1_loss
+
+        return loss_dict
+
+
+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):
+        super().__init__()
+        self.config = config
+        # BERT style VAE encoder with input [cls, *joint_space_configuration, *action_sequence].
+        # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
+        if self.config.use_vae:
+            self.vae_encoder = ACTEncoder(config)
+            self.vae_encoder_cls_embed = nn.Embedding(1, config.dim_model)
+            # Projection layer for joint-space configuration to hidden dimension.
+            self.vae_encoder_robot_state_input_proj = nn.Linear(
+                config.input_shapes["observation.state"][0], config.dim_model
+            )
+            # Projection layer for action (joint-space target) to hidden dimension.
+            self.vae_encoder_action_input_proj = nn.Linear(
+                config.input_shapes["observation.state"][0], config.dim_model
+            )
+            self.latent_dim = config.latent_dim
+            # 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, self.latent_dim * 2)
+            # Fixed sinusoidal positional embedding the whole input to the VAE encoder. Unsqueeze for batch
+            # dimension.
+            self.register_buffer(
+                "vae_encoder_pos_enc",
+                create_sinusoidal_pos_embedding(1 + 1 + config.chunk_size, config.dim_model).unsqueeze(0),
+            )
+
+        # Backbone for image feature extraction.
+        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, image_feature_map_pixels].
+        self.encoder_robot_state_input_proj = nn.Linear(
+            config.input_shapes["observation.state"][0], config.dim_model
+        )
+        self.encoder_latent_input_proj = nn.Linear(self.latent_dim, config.dim_model)
+        self.encoder_img_feat_input_proj = nn.Conv2d(
+            backbone_model.fc.in_features, config.dim_model, kernel_size=1
+        )
+        # Transformer encoder positional embeddings.
+        self.encoder_robot_and_latent_pos_embed = nn.Embedding(2, config.dim_model)
+        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, config.output_shapes["action"][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:
+
+        {
+            "observation.state": (B, state_dim) batch of robot states.
+            "observation.images": (B, n_cameras, C, H, W) batch of images.
+            "action" (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."
+
+        batch_size = batch["observation.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)
+            robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"]).unsqueeze(
+                1
+            )  # (B, 1, D)
+            action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
+            vae_encoder_input = torch.cat([cls_embed, robot_state_embed, action_embed], axis=1)  # (B, S+2, D)
+
+            # 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)
+
+            # 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)
+            )[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.latent_dim]
+            # This is 2log(sigma). Done this way to match the original implementation.
+            log_sigma_x2 = latent_pdf_params[:, self.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
+            latent_sample = torch.zeros([batch_size, self.latent_dim], dtype=torch.float32).to(
+                batch["observation.state"].device
+            )
+
+        # Prepare all other transformer encoder inputs.
+        # Camera observation features and positional embeddings.
+        all_cam_features = []
+        all_cam_pos_embeds = []
+        images = batch["observation.images"]
+        for cam_index in range(images.shape[-4]):
+            cam_features = self.backbone(images[:, cam_index])["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)  # (B, C, h, w)
+            all_cam_features.append(cam_features)
+            all_cam_pos_embeds.append(cam_pos_embed)
+        # Concatenate camera observation feature maps and positional embeddings along the width dimension.
+        encoder_in = torch.cat(all_cam_features, axis=-1)
+        cam_pos_embed = torch.cat(all_cam_pos_embeds, axis=-1)
+
+        # Get positional embeddings for robot state and latent.
+        robot_state_embed = self.encoder_robot_state_input_proj(batch["observation.state"])  # (B, C)
+        latent_embed = self.encoder_latent_input_proj(latent_sample)  # (B, C)
+
+        # Stack encoder input and positional embeddings moving to (S, B, C).
+        encoder_in = torch.cat(
+            [
+                torch.stack([latent_embed, robot_state_embed], axis=0),
+                einops.rearrange(encoder_in, "b c h w -> (h w) b c"),
+            ]
+        )
+        pos_embed = torch.cat(
+            [
+                self.encoder_robot_and_latent_pos_embed.weight.unsqueeze(1),
+                cam_pos_embed.flatten(2).permute(2, 0, 1),
+            ],
+            axis=0,
+        )
+
+        # Forward pass through the transformer modules.
+        encoder_out = self.encoder(encoder_in, pos_embed=pos_embed)
+        decoder_in = torch.zeros(
+            (self.config.chunk_size, batch_size, self.config.dim_model),
+            dtype=pos_embed.dtype,
+            device=pos_embed.device,
+        )
+        decoder_out = self.decoder(
+            decoder_in,
+            encoder_out,
+            encoder_pos_embed=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):
+        super().__init__()
+        self.layers = nn.ModuleList([ACTEncoderLayer(config) for _ in range(config.n_encoder_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) -> Tensor:
+        for layer in self.layers:
+            x = layer(x, pos_embed=pos_embed)
+        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) -> 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)[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.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 +1,177 @@
+#!/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
+
+
+@dataclass
+class DiffusionConfig:
+    """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`.
+
+    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, shapes doesnt 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, shapes doesnt 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 initalize 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.
+        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 mor 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
+
+    input_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "observation.image": [3, 96, 96],
+            "observation.state": [2],
+        }
+    )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [2],
+        }
+    )
+
+    # Normalization / Unnormalization
+    input_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {
+            "observation.image": "mean_std",
+            "observation.state": "min_max",
+        }
+    )
+    output_normalization_modes: dict[str, str] = field(default_factory=lambda: {"action": "min_max"})
+
+    # 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
+    # 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
+
+    def __post_init__(self):
+        """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}."
+            )
+        # There should only be one image key.
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
+        if len(image_keys) != 1:
+            raise ValueError(
+                f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
+            )
+        image_key = next(iter(image_keys))
+        if (
+            self.crop_shape[0] > self.input_shapes[image_key][1]
+            or self.crop_shape[1] > self.input_shapes[image_key][2]
+        ):
+            raise ValueError(
+                f"`crop_shape` should fit within `input_shapes[{image_key}]`. Got {self.crop_shape} "
+                f"for `crop_shape` and {self.input_shapes[image_key]} for "
+                "`input_shapes[{image_key}]`."
+            )
+        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}."
+            )

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 +1,708 @@
+#!/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.
+  - Make compatible with multiple image keys.
+"""
+
+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 huggingface_hub import PyTorchModelHubMixin
+from torch import Tensor, nn
+
+from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.utils import (
+    get_device_from_parameters,
+    get_dtype_from_parameters,
+    populate_queues,
+)
+
+
+class DiffusionPolicy(nn.Module, PyTorchModelHubMixin):
+    """
+    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).
+    """
+
+    name = "diffusion"
+
+    def __init__(
+        self,
+        config: DiffusionConfig | None = None,
+        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__()
+        if config is None:
+            config = DiffusionConfig()
+        self.config = config
+        self.normalize_inputs = Normalize(
+            config.input_shapes, config.input_normalization_modes, dataset_stats
+        )
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_shapes, config.output_normalization_modes, 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)
+
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        # Note: This check is covered in the post-init of the config but have a sanity check just in case.
+        if len(image_keys) != 1:
+            raise NotImplementedError(
+                f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
+            )
+        self.input_image_key = image_keys[0]
+
+        self.reset()
+
+    def reset(self):
+        """Clear observation and action queues. Should be called on `env.reset()`"""
+        self._queues = {
+            "observation.image": deque(maxlen=self.config.n_obs_steps),
+            "observation.state": deque(maxlen=self.config.n_obs_steps),
+            "action": 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 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+1+h|
+            |observation is used | YES   | YES   | YES   | NO    | 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)
+        batch["observation.image"] = batch[self.input_image_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]) -> dict[str, Tensor]:
+        """Run the batch through the model and compute the loss for training or validation."""
+        batch = self.normalize_inputs(batch)
+        batch["observation.image"] = batch[self.input_image_key]
+        batch = self.normalize_targets(batch)
+        loss = self.diffusion.compute_loss(batch)
+        return {"loss": loss}
+
+
+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
+
+        self.rgb_encoder = DiffusionRgbEncoder(config)
+        self.unet = DiffusionConditionalUnet1d(
+            config,
+            global_cond_dim=(config.output_shapes["action"][0] + self.rgb_encoder.feature_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.output_shapes["action"][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 generate_actions(self, batch: dict[str, Tensor]) -> Tensor:
+        """
+        This function expects `batch` to have:
+        {
+            "observation.state": (B, n_obs_steps, state_dim)
+            "observation.image": (B, n_obs_steps, C, H, W)
+        }
+        """
+        batch_size, n_obs_steps = batch["observation.state"].shape[:2]
+        assert n_obs_steps == self.config.n_obs_steps
+
+        # Extract image feature (first combine batch and sequence dims).
+        img_features = self.rgb_encoder(einops.rearrange(batch["observation.image"], "b n ... -> (b n) ..."))
+        # Separate batch and sequence dims.
+        img_features = einops.rearrange(img_features, "(b n) ... -> b n ...", b=batch_size)
+        # Concatenate state and image features then flatten to (B, global_cond_dim).
+        global_cond = torch.cat([batch["observation.state"], img_features], dim=-1).flatten(start_dim=1)
+
+        # run sampling
+        sample = self.conditional_sample(batch_size, global_cond=global_cond)
+
+        # `horizon` steps worth of actions (from the first observation).
+        actions = sample[..., : self.config.output_shapes["action"][0]]
+        # 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.image": (B, n_obs_steps, C, H, W)
+            "action": (B, horizon, action_dim)
+            "action_is_pad": (B, horizon)
+        }
+        """
+        # Input validation.
+        assert set(batch).issuperset({"observation.state", "observation.image", "action", "action_is_pad"})
+        batch_size, n_obs_steps = batch["observation.state"].shape[:2]
+        horizon = batch["action"].shape[1]
+        assert horizon == self.config.horizon
+        assert n_obs_steps == self.config.n_obs_steps
+
+        # Extract image feature (first combine batch and sequence dims).
+        img_features = self.rgb_encoder(einops.rearrange(batch["observation.image"], "b n ... -> (b n) ..."))
+        # Separate batch and sequence dims.
+        img_features = einops.rearrange(img_features, "(b n) ... -> b n ...", b=batch_size)
+        # Concatenate state and image features then flatten to (B, global_cond_dim).
+        global_cond = torch.cat([batch["observation.state"], img_features], dim=-1).flatten(start_dim=1)
+
+        trajectory = batch["action"]
+
+        # Forward diffusion.
+        # 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 and "action_is_pad" in batch:
+            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):
+    """Encoder 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.input_shapes` and it should
+        # use the height and width from `config.crop_shape`.
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        assert len(image_keys) == 1
+        image_key = image_keys[0]
+        dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *config.crop_shape))
+        with torch.inference_mode():
+            dummy_feature_map = self.backbone(dummy_input)
+        feature_map_shape = tuple(dummy_feature_map.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.output_shapes["action"][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.output_shapes["action"][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 +1,97 @@
-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.
+import inspect
+import logging
 
-        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,
+from omegaconf import DictConfig, OmegaConf
+
+from lerobot.common.policies.policy_protocol import Policy
+from lerobot.common.utils.utils import get_safe_torch_device
+
+
+def _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg):
+    expected_kwargs = set(inspect.signature(policy_cfg_class).parameters)
+    if not set(hydra_cfg.policy).issuperset(expected_kwargs):
+        logging.warning(
+            f"Hydra config is missing arguments: {set(expected_kwargs).difference(hydra_cfg.policy)}"
         )
-    else:
-        raise ValueError(cfg.policy.name)
+    policy_cfg = policy_cfg_class(
+        **{
+            k: v
+            for k, v in OmegaConf.to_container(hydra_cfg.policy, resolve=True).items()
+            if k in expected_kwargs
+        }
+    )
+    return policy_cfg
 
-    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 get_policy_and_config_classes(name: str) -> tuple[Policy, object]:
+    """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.configuration_tdmpc import TDMPCConfig
+        from lerobot.common.policies.tdmpc.modeling_tdmpc import TDMPCPolicy
+
+        return TDMPCPolicy, TDMPCConfig
+    elif name == "diffusion":
+        from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
+        from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+
+        return DiffusionPolicy, DiffusionConfig
+    elif name == "act":
+        from lerobot.common.policies.act.configuration_act import ACTConfig
+        from lerobot.common.policies.act.modeling_act import ACTPolicy
+
+        return ACTPolicy, ACTConfig
+    else:
+        raise NotImplementedError(f"Policy with name {name} is not implemented.")
+
+
+def make_policy(
+    hydra_cfg: DictConfig, pretrained_policy_name_or_path: str | None = None, dataset_stats=None
+) -> Policy:
+    """Make an instance of a policy class.
+
+    Args:
+        hydra_cfg: A parsed Hydra configuration (see scripts). If `pretrained_policy_name_or_path` is
+            provided, only `hydra_cfg.policy.name` is used while everything else is ignored.
+        pretrained_policy_name_or_path: Either the repo ID of a model hosted on the Hub or a path to a
+            directory containing weights saved using `Policy.save_pretrained`. Note that providing this
+            argument overrides everything in `hydra_cfg.policy` apart from `hydra_cfg.policy.name`.
+        dataset_stats: Dataset statistics to use for (un)normalization of inputs/outputs in the policy. Must
+            be provided when initializing a new policy, and must not be provided when loading a pretrained
+            policy. Therefore, this argument is mutually exclusive with `pretrained_policy_name_or_path`.
+    """
+    if not (pretrained_policy_name_or_path is None) ^ (dataset_stats is None):
+        raise ValueError("Only one of `pretrained_policy_name_or_path` and `dataset_stats` may be provided.")
+
+    policy_cls, policy_cfg_class = get_policy_and_config_classes(hydra_cfg.policy.name)
+
+    policy_cfg = _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg)
+    if pretrained_policy_name_or_path is None:
+        # Make a fresh policy.
+        policy = policy_cls(policy_cfg, dataset_stats)
+    else:
+        # Load a pretrained policy and override the config if needed (for example, if there are inference-time
+        # hyperparameters that we want to vary).
+        # TODO(alexander-soare): This hack makes use of huggingface_hub's tooling to load the policy with, pretrained
+        # weights which are then loaded into a fresh policy with the desired config. This PR in huggingface_hub should
+        # make it possible to avoid the hack: https://github.com/huggingface/huggingface_hub/pull/2274.
+        policy = policy_cls(policy_cfg)
+        policy.load_state_dict(policy_cls.from_pretrained(pretrained_policy_name_or_path).state_dict())
+
+    policy.to(get_safe_torch_device(hydra_cfg.device))
 
     return policy

lerobot/common/policies/normalize.py

@@ -0,0 +1,218 @@
+#!/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 torch
+from torch import Tensor, nn
+
+
+def create_stats_buffers(
+    shapes: dict[str, list[int]],
+    modes: dict[str, str],
+    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, mode in modes.items():
+        assert mode in ["mean_std", "min_max"]
+
+        shape = tuple(shapes[key])
+
+        if "image" in key:
+            # 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 mode == "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 mode == "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),
+                }
+            )
+
+        if stats is not None:
+            # 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 mode == "mean_std":
+                buffer["mean"].data = stats[key]["mean"].clone()
+                buffer["std"].data = stats[key]["std"].clone()
+            elif mode == "min_max":
+                buffer["min"].data = stats[key]["min"].clone()
+                buffer["max"].data = stats[key]["max"].clone()
+
+        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,
+        shapes: dict[str, list[int]],
+        modes: dict[str, str],
+        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.shapes = shapes
+        self.modes = modes
+        self.stats = stats
+        stats_buffers = create_stats_buffers(shapes, modes, 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]:
+        for key, mode in self.modes.items():
+            buffer = getattr(self, "buffer_" + key.replace(".", "_"))
+
+            if mode == "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 mode == "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)
+                # normalize to [-1, 1]
+                batch[key] = batch[key] * 2 - 1
+            else:
+                raise ValueError(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,
+        shapes: dict[str, list[int]],
+        modes: dict[str, str],
+        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.shapes = shapes
+        self.modes = modes
+        self.stats = stats
+        # `self.buffer_observation_state["mean"]` contains `torch.tensor(state_dim)`
+        stats_buffers = create_stats_buffers(shapes, modes, 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]:
+        for key, mode in self.modes.items():
+            buffer = getattr(self, "buffer_" + key.replace(".", "_"))
+
+            if mode == "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 mode == "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(mode)
+        return batch

lerobot/common/policies/policy_protocol.py

@@ -0,0 +1,75 @@
+#!/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.
+"""A protocol that all policies should follow.
+
+This provides a mechanism for type-hinting and isinstance checks without requiring the policies classes
+subclass a base class.
+
+The protocol structure, method signatures, and docstrings should be used by developers as a reference for
+how to implement new policies.
+"""
+
+from typing import Protocol, runtime_checkable
+
+from torch import Tensor
+
+
+@runtime_checkable
+class Policy(Protocol):
+    """The required interface for implementing a policy.
+
+    We also expect all policies to subclass torch.nn.Module and PyTorchModelHubMixin.
+    """
+
+    name: str
+
+    def __init__(self, cfg, dataset_stats: dict[str, dict[str, Tensor]] | None = None):
+        """
+        Args:
+            cfg: 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.
+        """
+
+    def reset(self):
+        """To be called whenever the environment is reset.
+
+        Does things like clearing caches.
+        """
+
+    def forward(self, batch: dict[str, Tensor]) -> dict:
+        """Run the batch through the model and compute the loss for training or validation.
+
+        Returns a dictionary with "loss" and potentially other information. Apart from "loss" which is a Tensor, all
+        other items should be logging-friendly, native Python types.
+        """
+
+    def select_action(self, batch: dict[str, Tensor]):
+        """Return one action to run in the environment (potentially in batch mode).
+
+        When the model uses a history of observations, or outputs a sequence of actions, this method deals
+        with caching.
+        """
+
+
+@runtime_checkable
+class PolicyWithUpdate(Policy, Protocol):
+    def update(self):
+        """An update method that is to be called after a training optimization step.
+
+        Implements an additional updates the model parameters may need (for example, doing an EMA step for a
+        target model, or incrementing an internal buffer).
+        """

lerobot/common/policies/tdmpc.py

@@ -1,512 +0,0 @@
-# ruff: noqa: N806
-
-import time
-from copy import deepcopy
-
-import einops
-import numpy as np
-import torch
-import torch.nn as nn
-
-import lerobot.common.policies.tdmpc_helper as h
-
-FIRST_FRAME = 0
-
-
-class TOLD(nn.Module):
-    """Task-Oriented Latent Dynamics (TOLD) model used in TD-MPC."""
-
-    def __init__(self, cfg):
-        super().__init__()
-        action_dim = cfg.action_dim
-
-        self.cfg = cfg
-        self._encoder = h.enc(cfg)
-        self._dynamics = h.dynamics(cfg.latent_dim + action_dim, cfg.mlp_dim, cfg.latent_dim)
-        self._reward = h.mlp(cfg.latent_dim + action_dim, cfg.mlp_dim, 1)
-        self._pi = h.mlp(cfg.latent_dim, cfg.mlp_dim, action_dim)
-        self._Qs = nn.ModuleList([h.q(cfg) for _ in range(cfg.num_q)])
-        self._V = h.v(cfg)
-        self.apply(h.orthogonal_init)
-        for m in [self._reward, *self._Qs]:
-            m[-1].weight.data.fill_(0)
-            m[-1].bias.data.fill_(0)
-
-    def track_q_grad(self, enable=True):
-        """Utility function. Enables/disables gradient tracking of Q-networks."""
-        for m in self._Qs:
-            h.set_requires_grad(m, enable)
-
-    def track_v_grad(self, enable=True):
-        """Utility function. Enables/disables gradient tracking of Q-networks."""
-        if hasattr(self, "_V"):
-            h.set_requires_grad(self._V, enable)
-
-    def encode(self, obs):
-        """Encodes an observation into its latent representation."""
-        out = self._encoder(obs)
-        if isinstance(obs, dict):
-            # fusion
-            out = torch.stack([v for k, v in out.items()]).mean(dim=0)
-        return out
-
-    def next(self, z, a):
-        """Predicts next latent state (d) and single-step reward (R)."""
-        x = torch.cat([z, a], dim=-1)
-        return self._dynamics(x), self._reward(x)
-
-    def next_dynamics(self, z, a):
-        """Predicts next latent state (d)."""
-        x = torch.cat([z, a], dim=-1)
-        return self._dynamics(x)
-
-    def pi(self, z, std=0):
-        """Samples an action from the learned policy (pi)."""
-        mu = torch.tanh(self._pi(z))
-        if std > 0:
-            std = torch.ones_like(mu) * std
-            return h.TruncatedNormal(mu, std).sample(clip=0.3)
-        return mu
-
-    def V(self, z):  # noqa: N802
-        """Predict state value (V)."""
-        return self._V(z)
-
-    def Q(self, z, a, return_type):  # noqa: N802
-        """Predict state-action value (Q)."""
-        assert return_type in {"min", "avg", "all"}
-        x = torch.cat([z, a], dim=-1)
-
-        if return_type == "all":
-            return torch.stack([q(x) for q in self._Qs], dim=0)
-
-        idxs = np.random.choice(self.cfg.num_q, 2, replace=False)
-        Q1, Q2 = self._Qs[idxs[0]](x), self._Qs[idxs[1]](x)
-        return torch.min(Q1, Q2) if return_type == "min" else (Q1 + Q2) / 2
-
-
-class TDMPC(nn.Module):
-    """Implementation of TD-MPC learning + inference."""
-
-    def __init__(self, cfg, device):
-        super().__init__()
-        self.action_dim = cfg.action_dim
-
-        self.cfg = cfg
-        self.device = torch.device(device)
-        self.std = h.linear_schedule(cfg.std_schedule, 0)
-        self.model = TOLD(cfg).cuda() if torch.cuda.is_available() and device == "cuda" else TOLD(cfg)
-        self.model_target = deepcopy(self.model)
-        self.optim = torch.optim.Adam(self.model.parameters(), lr=self.cfg.lr)
-        self.pi_optim = torch.optim.Adam(self.model._pi.parameters(), lr=self.cfg.lr)
-        # self.bc_optim = torch.optim.Adam(self.model.parameters(), lr=self.cfg.lr)
-        self.model.eval()
-        self.model_target.eval()
-        self.batch_size = cfg.batch_size
-
-        self.register_buffer("step", torch.zeros(1))
-
-    def state_dict(self):
-        """Retrieve state dict of TOLD model, including slow-moving target network."""
-        return {
-            "model": self.model.state_dict(),
-            "model_target": self.model_target.state_dict(),
-        }
-
-    def save(self, fp):
-        """Save state dict of TOLD model to filepath."""
-        torch.save(self.state_dict(), fp)
-
-    def load(self, fp):
-        """Load a saved state dict from filepath into current agent."""
-        d = torch.load(fp)
-        self.model.load_state_dict(d["model"])
-        self.model_target.load_state_dict(d["model_target"])
-
-    @torch.no_grad()
-    def forward(self, observation, step_count):
-        t0 = step_count.item() == 0
-
-        # TODO(rcadene): remove unsqueeze hack...
-        if observation["image"].ndim == 3:
-            observation["image"] = observation["image"].unsqueeze(0)
-            observation["state"] = observation["state"].unsqueeze(0)
-
-        obs = {
-            # TODO(rcadene): remove contiguous hack...
-            "rgb": observation["image"].contiguous(),
-            "state": observation["state"].contiguous(),
-        }
-        action = self.act(obs, t0=t0, step=self.step.item())
-        return action
-
-    @torch.no_grad()
-    def act(self, obs, t0=False, step=None):
-        """Take an action. Uses either MPC or the learned policy, depending on the self.cfg.mpc flag."""
-        obs = {k: o.detach() for k, o in obs.items()} if isinstance(obs, dict) else obs.detach()
-        z = self.model.encode(obs)
-        if self.cfg.mpc:
-            a = self.plan(z, t0=t0, step=step)
-        else:
-            a = self.model.pi(z, self.cfg.min_std * self.model.training).squeeze(0)
-        return a
-
-    @torch.no_grad()
-    def estimate_value(self, z, actions, horizon):
-        """Estimate value of a trajectory starting at latent state z and executing given actions."""
-        G, discount = 0, 1
-        for t in range(horizon):
-            if self.cfg.uncertainty_cost > 0:
-                G -= (
-                    discount
-                    * self.cfg.uncertainty_cost
-                    * self.model.Q(z, actions[t], return_type="all").std(dim=0)
-                )
-            z, reward = self.model.next(z, actions[t])
-            G += discount * reward
-            discount *= self.cfg.discount
-        pi = self.model.pi(z, self.cfg.min_std)
-        G += discount * self.model.Q(z, pi, return_type="min")
-        if self.cfg.uncertainty_cost > 0:
-            G -= discount * self.cfg.uncertainty_cost * self.model.Q(z, pi, return_type="all").std(dim=0)
-        return G
-
-    @torch.no_grad()
-    def plan(self, z, step=None, t0=True):
-        """
-        Plan next action using TD-MPC inference.
-        z: latent state.
-        step: current time step. determines e.g. planning horizon.
-        t0: whether current step is the first step of an episode.
-        """
-        # during eval: eval_mode: uniform sampling and action noise is disabled during evaluation.
-
-        assert step is not None
-        # Seed steps
-        if step < self.cfg.seed_steps and self.model.training:
-            return torch.empty(self.action_dim, dtype=torch.float32, device=self.device).uniform_(-1, 1)
-
-        # Sample policy trajectories
-        horizon = int(min(self.cfg.horizon, h.linear_schedule(self.cfg.horizon_schedule, step)))
-        num_pi_trajs = int(self.cfg.mixture_coef * self.cfg.num_samples)
-        if num_pi_trajs > 0:
-            pi_actions = torch.empty(horizon, num_pi_trajs, self.action_dim, device=self.device)
-            _z = z.repeat(num_pi_trajs, 1)
-            for t in range(horizon):
-                pi_actions[t] = self.model.pi(_z, self.cfg.min_std)
-                _z = self.model.next_dynamics(_z, pi_actions[t])
-
-        # Initialize state and parameters
-        z = z.repeat(self.cfg.num_samples + num_pi_trajs, 1)
-        mean = torch.zeros(horizon, self.action_dim, device=self.device)
-        std = self.cfg.max_std * torch.ones(horizon, self.action_dim, device=self.device)
-        if not t0 and hasattr(self, "_prev_mean"):
-            mean[:-1] = self._prev_mean[1:]
-
-        # Iterate CEM
-        for _ in range(self.cfg.iterations):
-            actions = torch.clamp(
-                mean.unsqueeze(1)
-                + std.unsqueeze(1)
-                * torch.randn(horizon, self.cfg.num_samples, self.action_dim, device=std.device),
-                -1,
-                1,
-            )
-            if num_pi_trajs > 0:
-                actions = torch.cat([actions, pi_actions], dim=1)
-
-            # Compute elite actions
-            value = self.estimate_value(z, actions, horizon).nan_to_num_(0)
-            elite_idxs = torch.topk(value.squeeze(1), self.cfg.num_elites, dim=0).indices
-            elite_value, elite_actions = value[elite_idxs], actions[:, elite_idxs]
-
-            # Update parameters
-            max_value = elite_value.max(0)[0]
-            score = torch.exp(self.cfg.temperature * (elite_value - max_value))
-            score /= score.sum(0)
-            _mean = torch.sum(score.unsqueeze(0) * elite_actions, dim=1) / (score.sum(0) + 1e-9)
-            _std = torch.sqrt(
-                torch.sum(
-                    score.unsqueeze(0) * (elite_actions - _mean.unsqueeze(1)) ** 2,
-                    dim=1,
-                )
-                / (score.sum(0) + 1e-9)
-            )
-            _std = _std.clamp_(self.std, self.cfg.max_std)
-            mean, std = self.cfg.momentum * mean + (1 - self.cfg.momentum) * _mean, _std
-
-        # Outputs
-        # TODO(rcadene): remove numpy with
-        # # Convert score tensor to probabilities using softmax
-        # probabilities = torch.softmax(score, dim=0)
-        # # Generate a random sample index based on the probabilities
-        # sample_index = torch.multinomial(probabilities, 1).item()
-        score = score.squeeze(1).cpu().numpy()
-        actions = elite_actions[:, np.random.choice(np.arange(score.shape[0]), p=score)]
-        self._prev_mean = mean
-        mean, std = actions[0], _std[0]
-        a = mean
-        if self.model.training:
-            a += std * torch.randn(self.action_dim, device=std.device)
-        return torch.clamp(a, -1, 1)
-
-    def update_pi(self, zs, acts=None):
-        """Update policy using a sequence of latent states."""
-        self.pi_optim.zero_grad(set_to_none=True)
-        self.model.track_q_grad(False)
-        self.model.track_v_grad(False)
-
-        info = {}
-        # Advantage Weighted Regression
-        assert acts is not None
-        vs = self.model.V(zs)
-        qs = self.model_target.Q(zs, acts, return_type="min")
-        adv = qs - vs
-        exp_a = torch.exp(adv * self.cfg.A_scaling)
-        exp_a = torch.clamp(exp_a, max=100.0)
-        log_probs = h.gaussian_logprob(self.model.pi(zs) - acts, 0)
-        rho = torch.pow(self.cfg.rho, torch.arange(len(qs), device=self.device))
-        pi_loss = -((exp_a * log_probs).mean(dim=(1, 2)) * rho).mean()
-        info["adv"] = adv[0]
-
-        pi_loss.backward()
-        torch.nn.utils.clip_grad_norm_(
-            self.model._pi.parameters(),
-            self.cfg.grad_clip_norm,
-            error_if_nonfinite=False,
-        )
-        self.pi_optim.step()
-        self.model.track_q_grad(True)
-        self.model.track_v_grad(True)
-
-        info["pi_loss"] = pi_loss.item()
-        return pi_loss.item(), info
-
-    @torch.no_grad()
-    def _td_target(self, next_z, reward, mask):
-        """Compute the TD-target from a reward and the observation at the following time step."""
-        next_v = self.model.V(next_z)
-        td_target = reward + self.cfg.discount * mask * next_v
-        return td_target
-
-    def update(self, replay_buffer, step, demo_buffer=None):
-        """Main update function. Corresponds to one iteration of the model learning."""
-        start_time = time.time()
-
-        num_slices = self.cfg.batch_size
-        batch_size = self.cfg.horizon * num_slices
-
-        if demo_buffer is None:
-            demo_batch_size = 0
-        else:
-            # Update oversampling ratio
-            demo_pc_batch = h.linear_schedule(self.cfg.demo_schedule, step)
-            demo_num_slices = int(demo_pc_batch * self.batch_size)
-            demo_batch_size = self.cfg.horizon * demo_num_slices
-            batch_size -= demo_batch_size
-            num_slices -= demo_num_slices
-            replay_buffer._sampler.num_slices = num_slices
-            demo_buffer._sampler.num_slices = demo_num_slices
-
-            assert demo_batch_size % self.cfg.horizon == 0
-            assert demo_batch_size % demo_num_slices == 0
-
-        assert batch_size % self.cfg.horizon == 0
-        assert batch_size % num_slices == 0
-
-        # Sample from interaction dataset
-
-        def process_batch(batch, horizon, num_slices):
-            # trajectory t = 256, horizon h = 5
-            # (t h) ... -> h t ...
-            batch = batch.reshape(num_slices, horizon).transpose(1, 0).contiguous()
-
-            obs = {
-                "rgb": batch["observation", "image"][FIRST_FRAME].to(self.device, non_blocking=True),
-                "state": batch["observation", "state"][FIRST_FRAME].to(self.device, non_blocking=True),
-            }
-            action = batch["action"].to(self.device, non_blocking=True)
-            next_obses = {
-                "rgb": batch["next", "observation", "image"].to(self.device, non_blocking=True),
-                "state": batch["next", "observation", "state"].to(self.device, non_blocking=True),
-            }
-            reward = batch["next", "reward"].to(self.device, non_blocking=True)
-
-            idxs = batch["index"][FIRST_FRAME].to(self.device, non_blocking=True)
-            weights = batch["_weight"][FIRST_FRAME, :, None].to(self.device, non_blocking=True)
-
-            # TODO(rcadene): rearrange directly in offline dataset
-            if reward.ndim == 2:
-                reward = einops.rearrange(reward, "h t -> h t 1")
-
-            assert reward.ndim == 3
-            assert reward.shape == (horizon, num_slices, 1)
-            # We dont use `batch["next", "done"]` since it only indicates the end of an
-            # episode, but not the end of the trajectory of an episode.
-            # Neither does `batch["next", "terminated"]`
-            done = torch.zeros_like(reward, dtype=torch.bool, device=reward.device)
-            mask = torch.ones_like(reward, dtype=torch.bool, device=reward.device)
-            return obs, action, next_obses, reward, mask, done, idxs, weights
-
-        batch = replay_buffer.sample(batch_size) if self.cfg.balanced_sampling else replay_buffer.sample()
-
-        obs, action, next_obses, reward, mask, done, idxs, weights = process_batch(
-            batch, self.cfg.horizon, num_slices
-        )
-
-        # Sample from demonstration dataset
-        if demo_batch_size > 0:
-            demo_batch = demo_buffer.sample(demo_batch_size)
-            (
-                demo_obs,
-                demo_action,
-                demo_next_obses,
-                demo_reward,
-                demo_mask,
-                demo_done,
-                demo_idxs,
-                demo_weights,
-            ) = process_batch(demo_batch, self.cfg.horizon, demo_num_slices)
-
-            if isinstance(obs, dict):
-                obs = {k: torch.cat([obs[k], demo_obs[k]]) for k in obs}
-                next_obses = {k: torch.cat([next_obses[k], demo_next_obses[k]], dim=1) for k in next_obses}
-            else:
-                obs = torch.cat([obs, demo_obs])
-                next_obses = torch.cat([next_obses, demo_next_obses], dim=1)
-            action = torch.cat([action, demo_action], dim=1)
-            reward = torch.cat([reward, demo_reward], dim=1)
-            mask = torch.cat([mask, demo_mask], dim=1)
-            done = torch.cat([done, demo_done], dim=1)
-            idxs = torch.cat([idxs, demo_idxs])
-            weights = torch.cat([weights, demo_weights])
-
-        # Apply augmentations
-        aug_tf = h.aug(self.cfg)
-        obs = aug_tf(obs)
-
-        for k in next_obses:
-            next_obses[k] = einops.rearrange(next_obses[k], "h t ... -> (h t) ...")
-        next_obses = aug_tf(next_obses)
-        for k in next_obses:
-            next_obses[k] = einops.rearrange(
-                next_obses[k],
-                "(h t) ... -> h t ...",
-                h=self.cfg.horizon,
-                t=self.cfg.batch_size,
-            )
-
-        horizon = self.cfg.horizon
-        loss_mask = torch.ones_like(mask, device=self.device)
-        for t in range(1, horizon):
-            loss_mask[t] = loss_mask[t - 1] * (~done[t - 1])
-
-        self.optim.zero_grad(set_to_none=True)
-        self.std = h.linear_schedule(self.cfg.std_schedule, step)
-        self.model.train()
-
-        data_s = time.time() - start_time
-
-        # Compute targets
-        with torch.no_grad():
-            next_z = self.model.encode(next_obses)
-            z_targets = self.model_target.encode(next_obses)
-            td_targets = self._td_target(next_z, reward, mask)
-
-        # Latent rollout
-        zs = torch.empty(horizon + 1, self.batch_size, self.cfg.latent_dim, device=self.device)
-        reward_preds = torch.empty_like(reward, device=self.device)
-        assert reward.shape[0] == horizon
-        z = self.model.encode(obs)
-        zs[0] = z
-        value_info = {"Q": 0.0, "V": 0.0}
-        for t in range(horizon):
-            z, reward_pred = self.model.next(z, action[t])
-            zs[t + 1] = z
-            reward_preds[t] = reward_pred
-
-        with torch.no_grad():
-            v_target = self.model_target.Q(zs[:-1].detach(), action, return_type="min")
-
-        # Predictions
-        qs = self.model.Q(zs[:-1], action, return_type="all")
-        value_info["Q"] = qs.mean().item()
-        v = self.model.V(zs[:-1])
-        value_info["V"] = v.mean().item()
-
-        # Losses
-        rho = torch.pow(self.cfg.rho, torch.arange(horizon, device=self.device)).view(-1, 1, 1)
-        consistency_loss = (rho * torch.mean(h.mse(zs[1:], z_targets), dim=2, keepdim=True) * loss_mask).sum(
-            dim=0
-        )
-        reward_loss = (rho * h.mse(reward_preds, reward) * loss_mask).sum(dim=0)
-        q_value_loss, priority_loss = 0, 0
-        for q in range(self.cfg.num_q):
-            q_value_loss += (rho * h.mse(qs[q], td_targets) * loss_mask).sum(dim=0)
-            priority_loss += (rho * h.l1(qs[q], td_targets) * loss_mask).sum(dim=0)
-
-        expectile = h.linear_schedule(self.cfg.expectile, step)
-        v_value_loss = (rho * h.l2_expectile(v_target - v, expectile=expectile) * loss_mask).sum(dim=0)
-
-        total_loss = (
-            self.cfg.consistency_coef * consistency_loss
-            + self.cfg.reward_coef * reward_loss
-            + self.cfg.value_coef * q_value_loss
-            + self.cfg.value_coef * v_value_loss
-        )
-
-        weighted_loss = (total_loss.squeeze(1) * weights).mean()
-        weighted_loss.register_hook(lambda grad: grad * (1 / self.cfg.horizon))
-        has_nan = torch.isnan(weighted_loss).item()
-        if has_nan:
-            print(f"weighted_loss has nan: {total_loss=} {weights=}")
-        else:
-            weighted_loss.backward()
-
-        grad_norm = torch.nn.utils.clip_grad_norm_(
-            self.model.parameters(), self.cfg.grad_clip_norm, error_if_nonfinite=False
-        )
-        self.optim.step()
-
-        if self.cfg.per:
-            # Update priorities
-            priorities = priority_loss.clamp(max=1e4).detach()
-            has_nan = torch.isnan(priorities).any().item()
-            if has_nan:
-                print(f"priorities has nan: {priorities=}")
-            else:
-                replay_buffer.update_priority(
-                    idxs[:num_slices],
-                    priorities[:num_slices],
-                )
-                if demo_batch_size > 0:
-                    demo_buffer.update_priority(demo_idxs, priorities[num_slices:])
-
-        # Update policy + target network
-        _, pi_update_info = self.update_pi(zs[:-1].detach(), acts=action)
-
-        if step % self.cfg.update_freq == 0:
-            h.ema(self.model._encoder, self.model_target._encoder, self.cfg.tau)
-            h.ema(self.model._Qs, self.model_target._Qs, self.cfg.tau)
-
-        self.model.eval()
-
-        info = {
-            "consistency_loss": float(consistency_loss.mean().item()),
-            "reward_loss": float(reward_loss.mean().item()),
-            "Q_value_loss": float(q_value_loss.mean().item()),
-            "V_value_loss": float(v_value_loss.mean().item()),
-            "sum_loss": float(total_loss.mean().item()),
-            "loss": float(weighted_loss.mean().item()),
-            "grad_norm": float(grad_norm),
-            "lr": self.cfg.lr,
-            "data_s": data_s,
-            "update_s": time.time() - start_time,
-        }
-        info["demo_batch_size"] = demo_batch_size
-        info["expectile"] = expectile
-        info.update(value_info)
-        info.update(pi_update_info)
-
-        self.step[0] = step
-        return info

lerobot/common/policies/tdmpc/configuration_tdmpc.py

@@ -0,0 +1,172 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Nicklas Hansen, Xiaolong Wang, Hao Su,
+# 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
+
+
+@dataclass
+class TDMPCConfig:
+    """Configuration class for TDMPCPolicy.
+
+    Defaults are configured for training with xarm_lift_medium_replay 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`, `output_shapes`, and perhaps `max_random_shift`.
+
+    Args:
+        n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
+            action repeats in Q-learning or ask your favorite chatbot)
+        horizon: Horizon for model predictive control.
+        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. Note that here this defaults to None meaning inputs are not normalized. This is to
+            match the original implementation.
+        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. NOTE: Clipping
+            to [-1, +1] is used during MPPI/CEM. Therefore, it is recommended that you stick with "min_max"
+            normalization mode here.
+        image_encoder_hidden_dim: Number of channels for the convolutional layers used for image encoding.
+        state_encoder_hidden_dim: Hidden dimension for MLP used for state vector encoding.
+        latent_dim: Observation's latent embedding dimension.
+        q_ensemble_size: Number of Q function estimators to use in an ensemble for uncertainty estimation.
+        mlp_dim: Hidden dimension of MLPs used for modelling the dynamics encoder, reward function, policy
+            (π), Q ensemble, and V.
+        discount: Discount factor (γ) to use for the reinforcement learning formalism.
+        use_mpc: Whether to use model predictive control. The alternative is to just sample the policy model
+            (π) for each step.
+        cem_iterations: Number of iterations for the MPPI/CEM loop in MPC.
+        max_std: Maximum standard deviation for actions sampled from the gaussian PDF in CEM.
+        min_std: Minimum standard deviation for noise applied to actions sampled from the policy model (π).
+            Doubles up as the minimum standard deviation for actions sampled from the gaussian PDF in CEM.
+        n_gaussian_samples: Number of samples to draw from the gaussian distribution every CEM iteration. Must
+            be non-zero.
+        n_pi_samples: Number of samples to draw from the policy / world model rollout every CEM iteration. Can
+            be zero.
+        uncertainty_regularizer_coeff: Coefficient for the uncertainty regularization used when estimating
+            trajectory values (this is the λ coeffiecient in eqn 4 of FOWM).
+        n_elites: The number of elite samples to use for updating the gaussian parameters every CEM iteration.
+        elite_weighting_temperature: The temperature to use for softmax weighting (by trajectory value) of the
+            elites, when updating the gaussian parameters for CEM.
+        gaussian_mean_momentum: Momentum (α) used for EMA updates of the mean parameter μ of the gaussian
+            parameters optimized in CEM. Updates are calculated as μ⁻ ← αμ⁻ + (1-α)μ.
+        max_random_shift_ratio: Maximum random shift (as a proportion of the image size) to apply to the
+            image(s) (in units of pixels) for training-time augmentation. If set to 0, no such augmentation
+            is applied. Note that the input images are assumed to be square for this augmentation.
+        reward_coeff: Loss weighting coefficient for the reward regression loss.
+        expectile_weight: Weighting (τ) used in expectile regression for the state value function (V).
+            v_pred < v_target is weighted by τ and v_pred >= v_target is weighted by (1-τ). τ is expected to
+            be in [0, 1]. Setting τ closer to 1 results in a more "optimistic" V. This is sensible to do
+            because v_target is obtained by evaluating the learned state-action value functions (Q) with
+            in-sample actions that may not be always optimal.
+        value_coeff: Loss weighting coefficient for both the state-action value (Q) TD loss, and the state
+            value (V) expectile regression loss.
+        consistency_coeff: Loss weighting coefficient for the consistency loss.
+        advantage_scaling: A factor by which the advantages are scaled prior to exponentiation for advantage
+            weighted regression of the policy (π) estimator parameters. Note that the exponentiated advantages
+            are clamped at 100.0.
+        pi_coeff: Loss weighting coefficient for the action regression loss.
+        temporal_decay_coeff: Exponential decay coefficient for decaying the loss coefficient for future time-
+            steps. Hint: each loss computation involves `horizon` steps worth of actions starting from the
+            current time step.
+        target_model_momentum: Momentum (α) used for EMA updates of the target models. Updates are calculated
+            as ϕ ← αϕ + (1-α)θ where ϕ are the parameters of the target model and θ are the parameters of the
+            model being trained.
+    """
+
+    # Input / output structure.
+    n_action_repeats: int = 2
+    horizon: int = 5
+
+    input_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "observation.image": [3, 84, 84],
+            "observation.state": [4],
+        }
+    )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [4],
+        }
+    )
+
+    # Normalization / Unnormalization
+    input_normalization_modes: dict[str, str] | None = None
+    output_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {"action": "min_max"},
+    )
+
+    # Architecture / modeling.
+    # Neural networks.
+    image_encoder_hidden_dim: int = 32
+    state_encoder_hidden_dim: int = 256
+    latent_dim: int = 50
+    q_ensemble_size: int = 5
+    mlp_dim: int = 512
+    # Reinforcement learning.
+    discount: float = 0.9
+
+    # Inference.
+    use_mpc: bool = True
+    cem_iterations: int = 6
+    max_std: float = 2.0
+    min_std: float = 0.05
+    n_gaussian_samples: int = 512
+    n_pi_samples: int = 51
+    uncertainty_regularizer_coeff: float = 1.0
+    n_elites: int = 50
+    elite_weighting_temperature: float = 0.5
+    gaussian_mean_momentum: float = 0.1
+
+    # Training and loss computation.
+    max_random_shift_ratio: float = 0.0476
+    # Loss coefficients.
+    reward_coeff: float = 0.5
+    expectile_weight: float = 0.9
+    value_coeff: float = 0.1
+    consistency_coeff: float = 20.0
+    advantage_scaling: float = 3.0
+    pi_coeff: float = 0.5
+    temporal_decay_coeff: float = 0.5
+    # Target model.
+    target_model_momentum: float = 0.995
+
+    def __post_init__(self):
+        """Input validation (not exhaustive)."""
+        # There should only be one image key.
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
+        if len(image_keys) != 1:
+            raise ValueError(
+                f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
+            )
+        image_key = next(iter(image_keys))
+        if self.input_shapes[image_key][-2] != self.input_shapes[image_key][-1]:
+            # TODO(alexander-soare): This limitation is solely because of code in the random shift
+            # augmentation. It should be able to be removed.
+            raise ValueError(
+                f"Only square images are handled now. Got image shape {self.input_shapes[image_key]}."
+            )
+        if self.n_gaussian_samples <= 0:
+            raise ValueError(
+                f"The number of guassian samples for CEM should be non-zero. Got `{self.n_gaussian_samples=}`"
+            )
+        if self.output_normalization_modes != {"action": "min_max"}:
+            raise ValueError(
+                "TD-MPC assumes the action space dimensions to all be in [-1, 1]. Therefore it is strongly "
+                f"advised that you stick with the default. See {self.__class__.__name__} docstring for more "
+                "information."
+            )

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -0,0 +1,810 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Nicklas Hansen, Xiaolong Wang, Hao Su,
+# 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.
+"""Implementation of Finetuning Offline World Models in the Real World.
+
+The comments in this code may sometimes refer to these references:
+    TD-MPC paper: Temporal Difference Learning for Model Predictive Control (https://arxiv.org/abs/2203.04955)
+    FOWM paper: Finetuning Offline World Models in the Real World (https://arxiv.org/abs/2310.16029)
+
+TODO(alexander-soare): Make rollout work for batch sizes larger than 1.
+TODO(alexander-soare): Use batch-first throughout.
+"""
+
+# ruff: noqa: N806
+
+import logging
+from collections import deque
+from copy import deepcopy
+from functools import partial
+from typing import Callable
+
+import einops
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from huggingface_hub import PyTorchModelHubMixin
+from torch import Tensor
+
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
+from lerobot.common.policies.utils import get_device_from_parameters, populate_queues
+
+
+class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
+    """Implementation of TD-MPC learning + inference.
+
+    Please note several warnings for this policy.
+        - Evaluation of pretrained weights created with the original FOWM code
+            (https://github.com/fyhMer/fowm) works as expected. To be precise: we trained and evaluated a
+            model with the FOWM code for the xarm_lift_medium_replay dataset. We ported the weights across
+            to LeRobot, and were able to evaluate with the same success metric. BUT, we had to use inter-
+            process communication to use the xarm environment from FOWM. This is because our xarm
+            environment uses newer dependencies and does not match the environment in FOWM. See
+            https://github.com/huggingface/lerobot/pull/103 for implementation details.
+        - We have NOT checked that training on LeRobot reproduces SOTA results. This is a TODO.
+        - Our current xarm datasets were generated using the environment from FOWM. Therefore they do not
+            match our xarm environment.
+    """
+
+    name = "tdmpc"
+
+    def __init__(
+        self, config: TDMPCConfig | None = None, 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__()
+        logging.warning(
+            """
+            Please note several warnings for this policy.
+
+            - Evaluation of pretrained weights created with the original FOWM code
+              (https://github.com/fyhMer/fowm) works as expected. To be precise: we trained and evaluated a
+              model with the FOWM code for the xarm_lift_medium_replay dataset. We ported the weights across
+              to LeRobot, and were able to evaluate with the same success metric. BUT, we had to use inter-
+              process communication to use the xarm environment from FOWM. This is because our xarm
+              environment uses newer dependencies and does not match the environment in FOWM. See
+              https://github.com/huggingface/lerobot/pull/103 for implementation details.
+            - We have NOT checked that training on LeRobot reproduces SOTA results. This is a TODO.
+            - Our current xarm datasets were generated using the environment from FOWM. Therefore they do not
+              match our xarm environment.
+            """
+        )
+
+        if config is None:
+            config = TDMPCConfig()
+        self.config = config
+        self.model = TDMPCTOLD(config)
+        self.model_target = deepcopy(self.model)
+        for param in self.model_target.parameters():
+            param.requires_grad = False
+
+        if config.input_normalization_modes is not None:
+            self.normalize_inputs = Normalize(
+                config.input_shapes, config.input_normalization_modes, dataset_stats
+            )
+        else:
+            self.normalize_inputs = nn.Identity()
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        # Note: This check is covered in the post-init of the config but have a sanity check just in case.
+        assert len(image_keys) == 1
+        self.input_image_key = image_keys[0]
+
+        self.reset()
+
+    def reset(self):
+        """
+        Clear observation and action queues. Clear previous means for warm starting of MPPI/CEM. Should be
+        called on `env.reset()`
+        """
+        self._queues = {
+            "observation.image": deque(maxlen=1),
+            "observation.state": deque(maxlen=1),
+            "action": deque(maxlen=self.config.n_action_repeats),
+        }
+        # Previous mean obtained from the cross-entropy method (CEM) used during MPC. It is used to warm start
+        # CEM for the next step.
+        self._prev_mean: torch.Tensor | None = None
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]):
+        """Select a single action given environment observations."""
+        batch = self.normalize_inputs(batch)
+        batch["observation.image"] = batch[self.input_image_key]
+
+        self._queues = populate_queues(self._queues, batch)
+
+        # When the action queue is depleted, populate it again by querying the policy.
+        if len(self._queues["action"]) == 0:
+            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
+
+            # Remove the time dimensions as it is not handled yet.
+            for key in batch:
+                assert batch[key].shape[1] == 1
+                batch[key] = batch[key][:, 0]
+
+            # NOTE: Order of observations matters here.
+            z = self.model.encode({k: batch[k] for k in ["observation.image", "observation.state"]})
+            if self.config.use_mpc:
+                batch_size = batch["observation.image"].shape[0]
+                # Batch processing is not handled in MPC mode, so process the batch in a loop.
+                action = []  # will be a batch of actions for one step
+                for i in range(batch_size):
+                    # Note: self.plan does not handle batches, hence the squeeze.
+                    action.append(self.plan(z[i]))
+                action = torch.stack(action)
+            else:
+                # Plan with the policy (π) alone.
+                action = self.model.pi(z)
+
+            self.unnormalize_outputs({"action": action})["action"]
+
+            for _ in range(self.config.n_action_repeats):
+                self._queues["action"].append(action)
+
+        action = self._queues["action"].popleft()
+        return torch.clamp(action, -1, 1)
+
+    @torch.no_grad()
+    def plan(self, z: Tensor) -> Tensor:
+        """Plan next action using TD-MPC inference.
+
+        Args:
+            z: (latent_dim,) tensor for the initial state.
+        Returns:
+            (action_dim,) tensor for the next action.
+
+        TODO(alexander-soare) Extend this to be able to work with batches.
+        """
+        device = get_device_from_parameters(self)
+
+        # Sample Nπ trajectories from the policy.
+        pi_actions = torch.empty(
+            self.config.horizon,
+            self.config.n_pi_samples,
+            self.config.output_shapes["action"][0],
+            device=device,
+        )
+        if self.config.n_pi_samples > 0:
+            _z = einops.repeat(z, "d -> n d", n=self.config.n_pi_samples)
+            for t in range(self.config.horizon):
+                # Note: Adding a small amount of noise here doesn't hurt during inference and may even be
+                # helpful for CEM.
+                pi_actions[t] = self.model.pi(_z, self.config.min_std)
+                _z = self.model.latent_dynamics(_z, pi_actions[t])
+
+        # In the CEM loop we will need this for a call to estimate_value with the gaussian sampled
+        # trajectories.
+        z = einops.repeat(z, "d -> n d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)
+
+        # Model Predictive Path Integral (MPPI) with the cross-entropy method (CEM) as the optimization
+        # algorithm.
+        # The initial mean and standard deviation for the cross-entropy method (CEM).
+        mean = torch.zeros(self.config.horizon, self.config.output_shapes["action"][0], device=device)
+        # Maybe warm start CEM with the mean from the previous step.
+        if self._prev_mean is not None:
+            mean[:-1] = self._prev_mean[1:]
+        std = self.config.max_std * torch.ones_like(mean)
+
+        for _ in range(self.config.cem_iterations):
+            # Randomly sample action trajectories for the gaussian distribution.
+            std_normal_noise = torch.randn(
+                self.config.horizon,
+                self.config.n_gaussian_samples,
+                self.config.output_shapes["action"][0],
+                device=std.device,
+            )
+            gaussian_actions = torch.clamp(mean.unsqueeze(1) + std.unsqueeze(1) * std_normal_noise, -1, 1)
+
+            # Compute elite actions.
+            actions = torch.cat([gaussian_actions, pi_actions], dim=1)
+            value = self.estimate_value(z, actions).nan_to_num_(0)
+            elite_idxs = torch.topk(value, self.config.n_elites, dim=0).indices
+            elite_value, elite_actions = value[elite_idxs], actions[:, elite_idxs]
+
+            # Update guassian PDF parameters to be the (weighted) mean and standard deviation of the elites.
+            max_value = elite_value.max(0)[0]
+            # The weighting is a softmax over trajectory values. Note that this is not the same as the usage
+            # of Ω in eqn 4 of the TD-MPC paper. Instead it is the normalized version of it: s = Ω/ΣΩ. This
+            # makes the equations: μ = Σ(s⋅Γ), σ = Σ(s⋅(Γ-μ)²).
+            score = torch.exp(self.config.elite_weighting_temperature * (elite_value - max_value))
+            score /= score.sum()
+            _mean = torch.sum(einops.rearrange(score, "n -> n 1") * elite_actions, dim=1)
+            _std = torch.sqrt(
+                torch.sum(
+                    einops.rearrange(score, "n -> n 1")
+                    * (elite_actions - einops.rearrange(_mean, "h d -> h 1 d")) ** 2,
+                    dim=1,
+                )
+            )
+            # Update mean with an exponential moving average, and std with a direct replacement.
+            mean = (
+                self.config.gaussian_mean_momentum * mean + (1 - self.config.gaussian_mean_momentum) * _mean
+            )
+            std = _std.clamp_(self.config.min_std, self.config.max_std)
+
+        # Keep track of the mean for warm-starting subsequent steps.
+        self._prev_mean = mean
+
+        # Randomly select one of the elite actions from the last iteration of MPPI/CEM using the softmax
+        # scores from the last iteration.
+        actions = elite_actions[:, torch.multinomial(score, 1).item()]
+
+        # Select only the first action
+        action = actions[0]
+        return action
+
+    @torch.no_grad()
+    def estimate_value(self, z: Tensor, actions: Tensor):
+        """Estimates the value of a trajectory as per eqn 4 of the FOWM paper.
+
+        Args:
+            z: (batch, latent_dim) tensor of initial latent states.
+            actions: (horizon, batch, action_dim) tensor of action trajectories.
+        Returns:
+            (batch,) tensor of values.
+        """
+        # Initialize return and running discount factor.
+        G, running_discount = 0, 1
+        # Iterate over the actions in the trajectory to simulate the trajectory using the latent dynamics
+        # model. Keep track of return.
+        for t in range(actions.shape[0]):
+            # We will compute the reward in a moment. First compute the uncertainty regularizer from eqn 4
+            # of the FOWM paper.
+            if self.config.uncertainty_regularizer_coeff > 0:
+                regularization = -(
+                    self.config.uncertainty_regularizer_coeff * self.model.Qs(z, actions[t]).std(0)
+                )
+            else:
+                regularization = 0
+            # Estimate the next state (latent) and reward.
+            z, reward = self.model.latent_dynamics_and_reward(z, actions[t])
+            # Update the return and running discount.
+            G += running_discount * (reward + regularization)
+            running_discount *= self.config.discount
+        # Add the estimated value of the final state (using the minimum for a conservative estimate).
+        # Do so by predicting the next action, then taking a minimum over the ensemble of state-action value
+        # estimators.
+        # Note: This small amount of added noise seems to help a bit at inference time as observed by success
+        # metrics over 50 episodes of xarm_lift_medium_replay.
+        next_action = self.model.pi(z, self.config.min_std)  # (batch, action_dim)
+        terminal_values = self.model.Qs(z, next_action)  # (ensemble, batch)
+        # Randomly choose 2 of the Qs for terminal value estimation (as in App C. of the FOWM paper).
+        if self.config.q_ensemble_size > 2:
+            G += (
+                running_discount
+                * torch.min(terminal_values[torch.randint(0, self.config.q_ensemble_size, size=(2,))], dim=0)[
+                    0
+                ]
+            )
+        else:
+            G += running_discount * torch.min(terminal_values, dim=0)[0]
+        # Finally, also regularize the terminal value.
+        if self.config.uncertainty_regularizer_coeff > 0:
+            G -= running_discount * self.config.uncertainty_regularizer_coeff * terminal_values.std(0)
+        return G
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        """Run the batch through the model and compute the loss."""
+        device = get_device_from_parameters(self)
+
+        batch = self.normalize_inputs(batch)
+        batch["observation.image"] = batch[self.input_image_key]
+        batch = self.normalize_targets(batch)
+
+        info = {}
+
+        # (b, t) -> (t, b)
+        for key in batch:
+            if batch[key].ndim > 1:
+                batch[key] = batch[key].transpose(1, 0)
+
+        action = batch["action"]  # (t, b)
+        reward = batch["next.reward"]  # (t,)
+        observations = {k: v for k, v in batch.items() if k.startswith("observation.")}
+
+        # Apply random image augmentations.
+        if self.config.max_random_shift_ratio > 0:
+            observations["observation.image"] = flatten_forward_unflatten(
+                partial(random_shifts_aug, max_random_shift_ratio=self.config.max_random_shift_ratio),
+                observations["observation.image"],
+            )
+
+        # Get the current observation for predicting trajectories, and all future observations for use in
+        # the latent consistency loss and TD loss.
+        current_observation, next_observations = {}, {}
+        for k in observations:
+            current_observation[k] = observations[k][0]
+            next_observations[k] = observations[k][1:]
+        horizon = next_observations["observation.image"].shape[0]
+
+        # Run latent rollout using the latent dynamics model and policy model.
+        # Note this has shape `horizon+1` because there are `horizon` actions and a current `z`. Each action
+        # gives us a next `z`.
+        batch_size = batch["index"].shape[0]
+        z_preds = torch.empty(horizon + 1, batch_size, self.config.latent_dim, device=device)
+        z_preds[0] = self.model.encode(current_observation)
+        reward_preds = torch.empty_like(reward, device=device)
+        for t in range(horizon):
+            z_preds[t + 1], reward_preds[t] = self.model.latent_dynamics_and_reward(z_preds[t], action[t])
+
+        # Compute Q and V value predictions based on the latent rollout.
+        q_preds_ensemble = self.model.Qs(z_preds[:-1], action)  # (ensemble, horizon, batch)
+        v_preds = self.model.V(z_preds[:-1])
+        info.update({"Q": q_preds_ensemble.mean().item(), "V": v_preds.mean().item()})
+
+        # Compute various targets with stopgrad.
+        with torch.no_grad():
+            # Latent state consistency targets.
+            z_targets = self.model_target.encode(next_observations)
+            # State-action value targets (or TD targets) as in eqn 3 of the FOWM. Unlike TD-MPC which uses the
+            # learned state-action value function in conjunction with the learned policy: Q(z, π(z)), FOWM
+            # uses a learned state value function: V(z). This means the TD targets only depend on in-sample
+            # actions (not actions estimated by π).
+            # Note: Here we do not use self.model_target, but self.model. This is to follow the original code
+            # and the FOWM paper.
+            q_targets = reward + self.config.discount * self.model.V(self.model.encode(next_observations))
+            # From eqn 3 of FOWM. These appear as Q(z, a). Here we call them v_targets to emphasize that we
+            # are using them to compute loss for V.
+            v_targets = self.model_target.Qs(z_preds[:-1].detach(), action, return_min=True)
+
+        # Compute losses.
+        # Exponentially decay the loss weight with respect to the timestep. Steps that are more distant in the
+        # future have less impact on the loss. Note: unsqueeze will let us broadcast to (seq, batch).
+        temporal_loss_coeffs = torch.pow(
+            self.config.temporal_decay_coeff, torch.arange(horizon, device=device)
+        ).unsqueeze(-1)
+        # Compute consistency loss as MSE loss between latents predicted from the rollout and latents
+        # predicted from the (target model's) observation encoder.
+        consistency_loss = (
+            (
+                temporal_loss_coeffs
+                * F.mse_loss(z_preds[1:], z_targets, reduction="none").mean(dim=-1)
+                # `z_preds` depends on the current observation and the actions.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+                # `z_targets` depends on the next observation.
+                * ~batch["observation.state_is_pad"][1:]
+            )
+            .sum(0)
+            .mean()
+        )
+        # Compute the reward loss as MSE loss between rewards predicted from the rollout and the dataset
+        # rewards.
+        reward_loss = (
+            (
+                temporal_loss_coeffs
+                * F.mse_loss(reward_preds, reward, reduction="none")
+                * ~batch["next.reward_is_pad"]
+                # `reward_preds` depends on the current observation and the actions.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+            )
+            .sum(0)
+            .mean()
+        )
+        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
+        q_value_loss = (
+            (
+                F.mse_loss(
+                    q_preds_ensemble,
+                    einops.repeat(q_targets, "t b -> e t b", e=q_preds_ensemble.shape[0]),
+                    reduction="none",
+                ).sum(0)  # sum over ensemble
+                # `q_preds_ensemble` depends on the first observation and the actions.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+                # q_targets depends on the reward and the next observations.
+                * ~batch["next.reward_is_pad"]
+                * ~batch["observation.state_is_pad"][1:]
+            )
+            .sum(0)
+            .mean()
+        )
+        # Compute state value loss as in eqn 3 of FOWM.
+        diff = v_targets - v_preds
+        # Expectile loss penalizes:
+        #   - `v_preds <  v_targets` with weighting `expectile_weight`
+        #   - `v_preds >= v_targets` with weighting `1 - expectile_weight`
+        raw_v_value_loss = torch.where(
+            diff > 0, self.config.expectile_weight, (1 - self.config.expectile_weight)
+        ) * (diff**2)
+        v_value_loss = (
+            (
+                temporal_loss_coeffs
+                * raw_v_value_loss
+                # `v_targets` depends on the first observation and the actions, as does `v_preds`.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+            )
+            .sum(0)
+            .mean()
+        )
+
+        # Calculate the advantage weighted regression loss for π as detailed in FOWM 3.1.
+        # We won't need these gradients again so detach.
+        z_preds = z_preds.detach()
+        # Use stopgrad for the advantage calculation.
+        with torch.no_grad():
+            advantage = self.model_target.Qs(z_preds[:-1], action, return_min=True) - self.model.V(
+                z_preds[:-1]
+            )
+            info["advantage"] = advantage[0]
+            # (t, b)
+            exp_advantage = torch.clamp(torch.exp(advantage * self.config.advantage_scaling), max=100.0)
+        action_preds = self.model.pi(z_preds[:-1])  # (t, b, a)
+        # Calculate the MSE between the actions and the action predictions.
+        # Note: FOWM's original code calculates the log probability (wrt to a unit standard deviation
+        # gaussian) and sums over the action dimension. Computing the log probability amounts to multiplying
+        # the MSE by 0.5 and adding a constant offset (the log(2*pi) term) . Here we drop the constant offset
+        # as it doesn't change the optimization step, and we drop the 0.5 as we instead make a configuration
+        # parameter for it (see below where we compute the total loss).
+        mse = F.mse_loss(action_preds, action, reduction="none").sum(-1)  # (t, b)
+        # NOTE: The original implementation does not take the sum over the temporal dimension like with the
+        # other losses.
+        # TODO(alexander-soare): Take the sum over the temporal dimension and check that training still works
+        # as well as expected.
+        pi_loss = (
+            exp_advantage
+            * mse
+            * temporal_loss_coeffs
+            # `action_preds` depends on the first observation and the actions.
+            * ~batch["observation.state_is_pad"][0]
+            * ~batch["action_is_pad"]
+        ).mean()
+
+        loss = (
+            self.config.consistency_coeff * consistency_loss
+            + self.config.reward_coeff * reward_loss
+            + self.config.value_coeff * q_value_loss
+            + self.config.value_coeff * v_value_loss
+            + self.config.pi_coeff * pi_loss
+        )
+
+        info.update(
+            {
+                "consistency_loss": consistency_loss.item(),
+                "reward_loss": reward_loss.item(),
+                "Q_value_loss": q_value_loss.item(),
+                "V_value_loss": v_value_loss.item(),
+                "pi_loss": pi_loss.item(),
+                "loss": loss,
+                "sum_loss": loss.item() * self.config.horizon,
+            }
+        )
+
+        # Undo (b, t) -> (t, b).
+        for key in batch:
+            if batch[key].ndim > 1:
+                batch[key] = batch[key].transpose(1, 0)
+
+        return info
+
+    def update(self):
+        """Update the target model's parameters with an EMA step."""
+        # Note a minor variation with respect to the original FOWM code. Here they do this based on an EMA
+        # update frequency parameter which is set to 2 (every 2 steps an update is done). To simplify the code
+        # we update every step and adjust the decay parameter `alpha` accordingly (0.99 -> 0.995)
+        update_ema_parameters(self.model_target, self.model, self.config.target_model_momentum)
+
+
+class TDMPCTOLD(nn.Module):
+    """Task-Oriented Latent Dynamics (TOLD) model used in TD-MPC."""
+
+    def __init__(self, config: TDMPCConfig):
+        super().__init__()
+        self.config = config
+        self._encoder = TDMPCObservationEncoder(config)
+        self._dynamics = nn.Sequential(
+            nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.latent_dim),
+            nn.LayerNorm(config.latent_dim),
+            nn.Sigmoid(),
+        )
+        self._reward = nn.Sequential(
+            nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, 1),
+        )
+        self._pi = nn.Sequential(
+            nn.Linear(config.latent_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.output_shapes["action"][0]),
+        )
+        self._Qs = nn.ModuleList(
+            [
+                nn.Sequential(
+                    nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
+                    nn.LayerNorm(config.mlp_dim),
+                    nn.Tanh(),
+                    nn.Linear(config.mlp_dim, config.mlp_dim),
+                    nn.ELU(),
+                    nn.Linear(config.mlp_dim, 1),
+                )
+                for _ in range(config.q_ensemble_size)
+            ]
+        )
+        self._V = nn.Sequential(
+            nn.Linear(config.latent_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Tanh(),
+            nn.Linear(config.mlp_dim, config.mlp_dim),
+            nn.ELU(),
+            nn.Linear(config.mlp_dim, 1),
+        )
+        self._init_weights()
+
+    def _init_weights(self):
+        """Initialize model weights.
+
+        Orthogonal initialization for all linear and convolutional layers' weights (apart from final layers
+        of reward network and Q networks which get zero initialization).
+        Zero initialization for all linear and convolutional layers' biases.
+        """
+
+        def _apply_fn(m):
+            if isinstance(m, nn.Linear):
+                nn.init.orthogonal_(m.weight.data)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Conv2d):
+                gain = nn.init.calculate_gain("relu")
+                nn.init.orthogonal_(m.weight.data, gain)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+        self.apply(_apply_fn)
+        for m in [self._reward, *self._Qs]:
+            assert isinstance(
+                m[-1], nn.Linear
+            ), "Sanity check. The last linear layer needs 0 initialization on weights."
+            nn.init.zeros_(m[-1].weight)
+            nn.init.zeros_(m[-1].bias)  # this has already been done, but keep this line here for good measure
+
+    def encode(self, obs: dict[str, Tensor]) -> Tensor:
+        """Encodes an observation into its latent representation."""
+        return self._encoder(obs)
+
+    def latent_dynamics_and_reward(self, z: Tensor, a: Tensor) -> tuple[Tensor, Tensor]:
+        """Predict the next state's latent representation and the reward given a current latent and action.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+        Returns:
+            A tuple containing:
+                - (*, latent_dim) tensor for the next state's latent representation.
+                - (*,) tensor for the estimated reward.
+        """
+        x = torch.cat([z, a], dim=-1)
+        return self._dynamics(x), self._reward(x).squeeze(-1)
+
+    def latent_dynamics(self, z: Tensor, a: Tensor) -> Tensor:
+        """Predict the next state's latent representation given a current latent and action.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+        Returns:
+            (*, latent_dim) tensor for the next state's latent representation.
+        """
+        x = torch.cat([z, a], dim=-1)
+        return self._dynamics(x)
+
+    def pi(self, z: Tensor, std: float = 0.0) -> Tensor:
+        """Samples an action from the learned policy.
+
+        The policy can also have added (truncated) Gaussian noise injected for encouraging exploration when
+        generating rollouts for online training.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            std: The standard deviation of the injected noise.
+        Returns:
+            (*, action_dim) tensor for the sampled action.
+        """
+        action = torch.tanh(self._pi(z))
+        if std > 0:
+            std = torch.ones_like(action) * std
+            action += torch.randn_like(action) * std
+        return action
+
+    def V(self, z: Tensor) -> Tensor:  # noqa: N802
+        """Predict state value (V).
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+        Returns:
+            (*,) tensor of estimated state values.
+        """
+        return self._V(z).squeeze(-1)
+
+    def Qs(self, z: Tensor, a: Tensor, return_min: bool = False) -> Tensor:  # noqa: N802
+        """Predict state-action value for all of the learned Q functions.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+            return_min: Set to true for implementing the detail in App. C of the FOWM paper: randomly select
+                2 of the Qs and return the minimum
+        Returns:
+            (q_ensemble, *) tensor for the value predictions of each learned Q function in the ensemble OR
+            (*,) tensor if return_min=True.
+        """
+        x = torch.cat([z, a], dim=-1)
+        if not return_min:
+            return torch.stack([q(x).squeeze(-1) for q in self._Qs], dim=0)
+        else:
+            if len(self._Qs) > 2:  # noqa: SIM108
+                Qs = [self._Qs[i] for i in np.random.choice(len(self._Qs), size=2)]
+            else:
+                Qs = self._Qs
+            return torch.stack([q(x).squeeze(-1) for q in Qs], dim=0).min(dim=0)[0]
+
+
+class TDMPCObservationEncoder(nn.Module):
+    """Encode image and/or state vector observations."""
+
+    def __init__(self, config: TDMPCConfig):
+        """
+        Creates encoders for pixel and/or state modalities.
+        TODO(alexander-soare): The original work allows for multiple images by concatenating them along the
+            channel dimension. Re-implement this capability.
+        """
+        super().__init__()
+        self.config = config
+
+        if "observation.image" in config.input_shapes:
+            self.image_enc_layers = nn.Sequential(
+                nn.Conv2d(
+                    config.input_shapes["observation.image"][0], config.image_encoder_hidden_dim, 7, stride=2
+                ),
+                nn.ReLU(),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 5, stride=2),
+                nn.ReLU(),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
+                nn.ReLU(),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
+                nn.ReLU(),
+            )
+            dummy_batch = torch.zeros(1, *config.input_shapes["observation.image"])
+            with torch.inference_mode():
+                out_shape = self.image_enc_layers(dummy_batch).shape[1:]
+            self.image_enc_layers.extend(
+                nn.Sequential(
+                    nn.Flatten(),
+                    nn.Linear(np.prod(out_shape), config.latent_dim),
+                    nn.LayerNorm(config.latent_dim),
+                    nn.Sigmoid(),
+                )
+            )
+        if "observation.state" in config.input_shapes:
+            self.state_enc_layers = nn.Sequential(
+                nn.Linear(config.input_shapes["observation.state"][0], config.state_encoder_hidden_dim),
+                nn.ELU(),
+                nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
+                nn.LayerNorm(config.latent_dim),
+                nn.Sigmoid(),
+            )
+
+    def forward(self, obs_dict: dict[str, Tensor]) -> Tensor:
+        """Encode the image and/or state vector.
+
+        Each modality is encoded into a feature vector of size (latent_dim,) and then a uniform mean is taken
+        over all features.
+        """
+        feat = []
+        if "observation.image" in self.config.input_shapes:
+            feat.append(flatten_forward_unflatten(self.image_enc_layers, obs_dict["observation.image"]))
+        if "observation.state" in self.config.input_shapes:
+            feat.append(self.state_enc_layers(obs_dict["observation.state"]))
+        return torch.stack(feat, dim=0).mean(0)
+
+
+def random_shifts_aug(x: Tensor, max_random_shift_ratio: float) -> Tensor:
+    """Randomly shifts images horizontally and vertically.
+
+    Adapted from https://github.com/facebookresearch/drqv2
+    """
+    b, _, h, w = x.size()
+    assert h == w, "non-square images not handled yet"
+    pad = int(round(max_random_shift_ratio * h))
+    x = F.pad(x, tuple([pad] * 4), "replicate")
+    eps = 1.0 / (h + 2 * pad)
+    arange = torch.linspace(
+        -1.0 + eps,
+        1.0 - eps,
+        h + 2 * pad,
+        device=x.device,
+        dtype=torch.float32,
+    )[:h]
+    arange = einops.repeat(arange, "w -> h w 1", h=h)
+    base_grid = torch.cat([arange, arange.transpose(1, 0)], dim=2)
+    base_grid = einops.repeat(base_grid, "h w c -> b h w c", b=b)
+    # A random shift in units of pixels and within the boundaries of the padding.
+    shift = torch.randint(
+        0,
+        2 * pad + 1,
+        size=(b, 1, 1, 2),
+        device=x.device,
+        dtype=torch.float32,
+    )
+    shift *= 2.0 / (h + 2 * pad)
+    grid = base_grid + shift
+    return F.grid_sample(x, grid, padding_mode="zeros", align_corners=False)
+
+
+def update_ema_parameters(ema_net: nn.Module, net: nn.Module, alpha: float):
+    """Update EMA parameters in place with ema_param <- alpha * ema_param + (1 - alpha) * param."""
+    for ema_module, module in zip(ema_net.modules(), net.modules(), strict=True):
+        for (n_p_ema, p_ema), (n_p, p) in zip(
+            ema_module.named_parameters(recurse=False), module.named_parameters(recurse=False), strict=True
+        ):
+            assert n_p_ema == n_p, "Parameter names don't match for EMA model update"
+            if isinstance(p, dict):
+                raise RuntimeError("Dict parameter not supported")
+            if isinstance(module, nn.modules.batchnorm._BatchNorm) or not p.requires_grad:
+                # Copy BatchNorm parameters, and non-trainable parameters directly.
+                p_ema.copy_(p.to(dtype=p_ema.dtype).data)
+            with torch.no_grad():
+                p_ema.mul_(alpha)
+                p_ema.add_(p.to(dtype=p_ema.dtype).data, alpha=1 - alpha)
+
+
+def flatten_forward_unflatten(fn: Callable[[Tensor], Tensor], image_tensor: Tensor) -> Tensor:
+    """Helper to temporarily flatten extra dims at the start of the image tensor.
+
+    Args:
+        fn: Callable that the image tensor will be passed to. It should accept (B, C, H, W) and return
+            (B, *), where * is any number of dimensions.
+        image_tensor: An image tensor of shape (**, C, H, W), where ** is any number of dimensions, generally
+            different from *.
+    Returns:
+        A return value from the callable reshaped to (**, *).
+    """
+    if image_tensor.ndim == 4:
+        return fn(image_tensor)
+    start_dims = image_tensor.shape[:-3]
+    inp = torch.flatten(image_tensor, end_dim=-4)
+    flat_out = fn(inp)
+    return torch.reshape(flat_out, (*start_dims, *flat_out.shape[1:]))

lerobot/common/policies/tdmpc_helper.py

@@ -1,576 +0,0 @@
-import os
-import pickle
-import re
-
-import numpy as np
-import torch
-import torch.nn as nn
-import torch.nn.functional as F  # noqa: N812
-from torch import distributions as pyd
-from torch.distributions.utils import _standard_normal
-
-DEFAULT_ACT_FN = nn.Mish()
-
-
-def __REDUCE__(b):  # noqa: N802, N807
-    return "mean" if b else "none"
-
-
-def l1(pred, target, reduce=False):
-    """Computes the L1-loss between predictions and targets."""
-    return F.l1_loss(pred, target, reduction=__REDUCE__(reduce))
-
-
-def mse(pred, target, reduce=False):
-    """Computes the MSE loss between predictions and targets."""
-    return F.mse_loss(pred, target, reduction=__REDUCE__(reduce))
-
-
-def l2_expectile(diff, expectile=0.7, reduce=False):
-    weight = torch.where(diff > 0, expectile, (1 - expectile))
-    loss = weight * (diff**2)
-    reduction = __REDUCE__(reduce)
-    if reduction == "mean":
-        return torch.mean(loss)
-    elif reduction == "sum":
-        return torch.sum(loss)
-    return loss
-
-
-def _get_out_shape(in_shape, layers):
-    """Utility function. Returns the output shape of a network for a given input shape."""
-    x = torch.randn(*in_shape).unsqueeze(0)
-    return (nn.Sequential(*layers) if isinstance(layers, list) else layers)(x).squeeze(0).shape
-
-
-def gaussian_logprob(eps, log_std):
-    """Compute Gaussian log probability."""
-    residual = (-0.5 * eps.pow(2) - log_std).sum(-1, keepdim=True)
-    return residual - 0.5 * np.log(2 * np.pi) * eps.size(-1)
-
-
-def squash(mu, pi, log_pi):
-    """Apply squashing function."""
-    mu = torch.tanh(mu)
-    pi = torch.tanh(pi)
-    log_pi -= torch.log(F.relu(1 - pi.pow(2)) + 1e-6).sum(-1, keepdim=True)
-    return mu, pi, log_pi
-
-
-def orthogonal_init(m):
-    """Orthogonal layer initialization."""
-    if isinstance(m, nn.Linear):
-        nn.init.orthogonal_(m.weight.data)
-        if m.bias is not None:
-            nn.init.zeros_(m.bias)
-    elif isinstance(m, nn.Conv2d):
-        gain = nn.init.calculate_gain("relu")
-        nn.init.orthogonal_(m.weight.data, gain)
-        if m.bias is not None:
-            nn.init.zeros_(m.bias)
-
-
-def ema(m, m_target, tau):
-    """Update slow-moving average of online network (target network) at rate tau."""
-    with torch.no_grad():
-        # TODO(rcadene, aliberts): issue with strict=False
-        # for p, p_target in zip(m.parameters(), m_target.parameters(), strict=False):
-        #     p_target.data.lerp_(p.data, tau)
-        m_params_iter = iter(m.parameters())
-        m_target_params_iter = iter(m_target.parameters())
-
-        while True:
-            try:
-                p = next(m_params_iter)
-                p_target = next(m_target_params_iter)
-                p_target.data.lerp_(p.data, tau)
-            except StopIteration:
-                # If any iterator is exhausted, exit the loop
-                break
-
-
-def set_requires_grad(net, value):
-    """Enable/disable gradients for a given (sub)network."""
-    for param in net.parameters():
-        param.requires_grad_(value)
-
-
-class TruncatedNormal(pyd.Normal):
-    """Utility class implementing the truncated normal distribution."""
-
-    default_sample_shape = torch.Size()
-
-    def __init__(self, loc, scale, low=-1.0, high=1.0, eps=1e-6):
-        super().__init__(loc, scale, validate_args=False)
-        self.low = low
-        self.high = high
-        self.eps = eps
-
-    def _clamp(self, x):
-        clamped_x = torch.clamp(x, self.low + self.eps, self.high - self.eps)
-        x = x - x.detach() + clamped_x.detach()
-        return x
-
-    def sample(self, clip=None, sample_shape=default_sample_shape):
-        shape = self._extended_shape(sample_shape)
-        eps = _standard_normal(shape, dtype=self.loc.dtype, device=self.loc.device)
-        eps *= self.scale
-        if clip is not None:
-            eps = torch.clamp(eps, -clip, clip)
-        x = self.loc + eps
-        return self._clamp(x)
-
-
-class NormalizeImg(nn.Module):
-    """Normalizes pixel observations to [0,1) range."""
-
-    def __init__(self):
-        super().__init__()
-
-    def forward(self, x):
-        return x.div(255.0)
-
-
-class Flatten(nn.Module):
-    """Flattens its input to a (batched) vector."""
-
-    def __init__(self):
-        super().__init__()
-
-    def forward(self, x):
-        return x.view(x.size(0), -1)
-
-
-def enc(cfg):
-    obs_shape = {
-        "rgb": (3, cfg.img_size, cfg.img_size),
-        "state": (cfg.state_dim,),
-    }
-
-    """Returns a TOLD encoder."""
-    pixels_enc_layers, state_enc_layers = None, None
-    if cfg.modality in {"pixels", "all"}:
-        C = int(3 * cfg.frame_stack)  # noqa: N806
-        pixels_enc_layers = [
-            NormalizeImg(),
-            nn.Conv2d(C, cfg.num_channels, 7, stride=2),
-            nn.ReLU(),
-            nn.Conv2d(cfg.num_channels, cfg.num_channels, 5, stride=2),
-            nn.ReLU(),
-            nn.Conv2d(cfg.num_channels, cfg.num_channels, 3, stride=2),
-            nn.ReLU(),
-            nn.Conv2d(cfg.num_channels, cfg.num_channels, 3, stride=2),
-            nn.ReLU(),
-        ]
-        out_shape = _get_out_shape((C, cfg.img_size, cfg.img_size), pixels_enc_layers)
-        pixels_enc_layers.extend(
-            [
-                Flatten(),
-                nn.Linear(np.prod(out_shape), cfg.latent_dim),
-                nn.LayerNorm(cfg.latent_dim),
-                nn.Sigmoid(),
-            ]
-        )
-        if cfg.modality == "pixels":
-            return ConvExt(nn.Sequential(*pixels_enc_layers))
-    if cfg.modality in {"state", "all"}:
-        state_dim = obs_shape[0] if cfg.modality == "state" else obs_shape["state"][0]
-        state_enc_layers = [
-            nn.Linear(state_dim, cfg.enc_dim),
-            nn.ELU(),
-            nn.Linear(cfg.enc_dim, cfg.latent_dim),
-            nn.LayerNorm(cfg.latent_dim),
-            nn.Sigmoid(),
-        ]
-        if cfg.modality == "state":
-            return nn.Sequential(*state_enc_layers)
-    else:
-        raise NotImplementedError
-
-    encoders = {}
-    for k in obs_shape:
-        if k == "state":
-            encoders[k] = nn.Sequential(*state_enc_layers)
-        elif k.endswith("rgb"):
-            encoders[k] = ConvExt(nn.Sequential(*pixels_enc_layers))
-        else:
-            raise NotImplementedError
-    return Multiplexer(nn.ModuleDict(encoders))
-
-
-def mlp(in_dim, mlp_dim, out_dim, act_fn=DEFAULT_ACT_FN):
-    """Returns an MLP."""
-    if isinstance(mlp_dim, int):
-        mlp_dim = [mlp_dim, mlp_dim]
-    return nn.Sequential(
-        nn.Linear(in_dim, mlp_dim[0]),
-        nn.LayerNorm(mlp_dim[0]),
-        act_fn,
-        nn.Linear(mlp_dim[0], mlp_dim[1]),
-        nn.LayerNorm(mlp_dim[1]),
-        act_fn,
-        nn.Linear(mlp_dim[1], out_dim),
-    )
-
-
-def dynamics(in_dim, mlp_dim, out_dim, act_fn=DEFAULT_ACT_FN):
-    """Returns a dynamics network."""
-    return nn.Sequential(
-        mlp(in_dim, mlp_dim, out_dim, act_fn),
-        nn.LayerNorm(out_dim),
-        nn.Sigmoid(),
-    )
-
-
-def q(cfg):
-    action_dim = cfg.action_dim
-    """Returns a Q-function that uses Layer Normalization."""
-    return nn.Sequential(
-        nn.Linear(cfg.latent_dim + action_dim, cfg.mlp_dim),
-        nn.LayerNorm(cfg.mlp_dim),
-        nn.Tanh(),
-        nn.Linear(cfg.mlp_dim, cfg.mlp_dim),
-        nn.ELU(),
-        nn.Linear(cfg.mlp_dim, 1),
-    )
-
-
-def v(cfg):
-    """Returns a state value function that uses Layer Normalization."""
-    return nn.Sequential(
-        nn.Linear(cfg.latent_dim, cfg.mlp_dim),
-        nn.LayerNorm(cfg.mlp_dim),
-        nn.Tanh(),
-        nn.Linear(cfg.mlp_dim, cfg.mlp_dim),
-        nn.ELU(),
-        nn.Linear(cfg.mlp_dim, 1),
-    )
-
-
-def aug(cfg):
-    obs_shape = {
-        "rgb": (3, cfg.img_size, cfg.img_size),
-        "state": (4,),
-    }
-
-    """Multiplex augmentation"""
-    if cfg.modality == "state":
-        return nn.Identity()
-    elif cfg.modality == "pixels":
-        return RandomShiftsAug(cfg)
-    else:
-        augs = {}
-        for k in obs_shape:
-            if k == "state":
-                augs[k] = nn.Identity()
-            elif k.endswith("rgb"):
-                augs[k] = RandomShiftsAug(cfg)
-            else:
-                raise NotImplementedError
-        return Multiplexer(nn.ModuleDict(augs))
-
-
-class ConvExt(nn.Module):
-    """Auxiliary conv net accommodating high-dim input"""
-
-    def __init__(self, conv):
-        super().__init__()
-        self.conv = conv
-
-    def forward(self, x):
-        if x.ndim > 4:
-            batch_shape = x.shape[:-3]
-            out = self.conv(x.view(-1, *x.shape[-3:]))
-            out = out.view(*batch_shape, *out.shape[1:])
-        else:
-            out = self.conv(x)
-        return out
-
-
-class Multiplexer(nn.Module):
-    """Model multiplexer"""
-
-    def __init__(self, choices):
-        super().__init__()
-        self.choices = choices
-
-    def forward(self, x, key=None):
-        if isinstance(x, dict):
-            if key is not None:
-                return self.choices[key](x)
-            return {k: self.choices[k](_x) for k, _x in x.items()}
-        return self.choices(x)
-
-
-class RandomShiftsAug(nn.Module):
-    """
-    Random shift image augmentation.
-    Adapted from https://github.com/facebookresearch/drqv2
-    """
-
-    def __init__(self, cfg):
-        super().__init__()
-        assert cfg.modality in {"pixels", "all"}
-        self.pad = int(cfg.img_size / 21)
-
-    def forward(self, x):
-        n, c, h, w = x.size()
-        assert h == w
-        padding = tuple([self.pad] * 4)
-        x = F.pad(x, padding, "replicate")
-        eps = 1.0 / (h + 2 * self.pad)
-        arange = torch.linspace(
-            -1.0 + eps,
-            1.0 - eps,
-            h + 2 * self.pad,
-            device=x.device,
-            dtype=torch.float32,
-        )[:h]
-        arange = arange.unsqueeze(0).repeat(h, 1).unsqueeze(2)
-        base_grid = torch.cat([arange, arange.transpose(1, 0)], dim=2)
-        base_grid = base_grid.unsqueeze(0).repeat(n, 1, 1, 1)
-        shift = torch.randint(
-            0,
-            2 * self.pad + 1,
-            size=(n, 1, 1, 2),
-            device=x.device,
-            dtype=torch.float32,
-        )
-        shift *= 2.0 / (h + 2 * self.pad)
-        grid = base_grid + shift
-        return F.grid_sample(x, grid, padding_mode="zeros", align_corners=False)
-
-
-# TODO(aliberts): remove class
-# class Episode:
-#     """Storage object for a single episode."""
-
-#     def __init__(self, cfg, init_obs):
-#         action_dim = cfg.action_dim
-
-#         self.cfg = cfg
-#         self.device = torch.device(cfg.buffer_device)
-#         if cfg.modality in {"pixels", "state"}:
-#             dtype = torch.float32 if cfg.modality == "state" else torch.uint8
-#             self.obses = torch.empty(
-#                 (cfg.episode_length + 1, *init_obs.shape),
-#                 dtype=dtype,
-#                 device=self.device,
-#             )
-#             self.obses[0] = torch.tensor(init_obs, dtype=dtype, device=self.device)
-#         elif cfg.modality == "all":
-#             self.obses = {}
-#             for k, v in init_obs.items():
-#                 assert k in {"rgb", "state"}
-#                 dtype = torch.float32 if k == "state" else torch.uint8
-#                 self.obses[k] = torch.empty(
-#                     (cfg.episode_length + 1, *v.shape), dtype=dtype, device=self.device
-#                 )
-#                 self.obses[k][0] = torch.tensor(v, dtype=dtype, device=self.device)
-#         else:
-#             raise ValueError
-#         self.actions = torch.empty((cfg.episode_length, action_dim), dtype=torch.float32, device=self.device)
-#         self.rewards = torch.empty((cfg.episode_length,), dtype=torch.float32, device=self.device)
-#         self.dones = torch.empty((cfg.episode_length,), dtype=torch.bool, device=self.device)
-#         self.masks = torch.empty((cfg.episode_length,), dtype=torch.float32, device=self.device)
-#         self.cumulative_reward = 0
-#         self.done = False
-#         self.success = False
-#         self._idx = 0
-
-#     def __len__(self):
-#         return self._idx
-
-#     @classmethod
-#     def from_trajectory(cls, cfg, obses, actions, rewards, dones=None, masks=None):
-#         """Constructs an episode from a trajectory."""
-
-#         if cfg.modality in {"pixels", "state"}:
-#             episode = cls(cfg, obses[0])
-#             episode.obses[1:] = torch.tensor(obses[1:], dtype=episode.obses.dtype, device=episode.device)
-#         elif cfg.modality == "all":
-#             episode = cls(cfg, {k: v[0] for k, v in obses.items()})
-#             for k in obses:
-#                 episode.obses[k][1:] = torch.tensor(
-#                     obses[k][1:], dtype=episode.obses[k].dtype, device=episode.device
-#                 )
-#         else:
-#             raise NotImplementedError
-#         episode.actions = torch.tensor(actions, dtype=episode.actions.dtype, device=episode.device)
-#         episode.rewards = torch.tensor(rewards, dtype=episode.rewards.dtype, device=episode.device)
-#         episode.dones = (
-#             torch.tensor(dones, dtype=episode.dones.dtype, device=episode.device)
-#             if dones is not None
-#             else torch.zeros_like(episode.dones)
-#         )
-#         episode.masks = (
-#             torch.tensor(masks, dtype=episode.masks.dtype, device=episode.device)
-#             if masks is not None
-#             else torch.ones_like(episode.masks)
-#         )
-#         episode.cumulative_reward = torch.sum(episode.rewards)
-#         episode.done = True
-#         episode._idx = cfg.episode_length
-#         return episode
-
-#     @property
-#     def first(self):
-#         return len(self) == 0
-
-#     def __add__(self, transition):
-#         self.add(*transition)
-#         return self
-
-#     def add(self, obs, action, reward, done, mask=1.0, success=False):
-#         """Add a transition into the episode."""
-#         if isinstance(obs, dict):
-#             for k, v in obs.items():
-#                 self.obses[k][self._idx + 1] = torch.tensor(
-#                     v, dtype=self.obses[k].dtype, device=self.obses[k].device
-#                 )
-#         else:
-#             self.obses[self._idx + 1] = torch.tensor(obs, dtype=self.obses.dtype, device=self.obses.device)
-#         self.actions[self._idx] = action
-#         self.rewards[self._idx] = reward
-#         self.dones[self._idx] = done
-#         self.masks[self._idx] = mask
-#         self.cumulative_reward += reward
-#         self.done = done
-#         self.success = self.success or success
-#         self._idx += 1
-
-
-def get_dataset_dict(cfg, env, return_reward_normalizer=False):
-    """Construct a dataset for env"""
-    required_keys = [
-        "observations",
-        "next_observations",
-        "actions",
-        "rewards",
-        "dones",
-        "masks",
-    ]
-
-    if cfg.task.startswith("xarm"):
-        dataset_path = os.path.join(cfg.dataset_dir, "buffer.pkl")
-        print(f"Using offline dataset '{dataset_path}'")
-        with open(dataset_path, "rb") as f:
-            dataset_dict = pickle.load(f)
-        for k in required_keys:
-            if k not in dataset_dict and k[:-1] in dataset_dict:
-                dataset_dict[k] = dataset_dict.pop(k[:-1])
-    elif cfg.task.startswith("legged"):
-        dataset_path = os.path.join(cfg.dataset_dir, "buffer.pkl")
-        print(f"Using offline dataset '{dataset_path}'")
-        with open(dataset_path, "rb") as f:
-            dataset_dict = pickle.load(f)
-        dataset_dict["actions"] /= env.unwrapped.clip_actions
-        print(f"clip_actions={env.unwrapped.clip_actions}")
-    else:
-        import d4rl
-
-        dataset_dict = d4rl.qlearning_dataset(env)
-        dones = np.full_like(dataset_dict["rewards"], False, dtype=bool)
-
-        for i in range(len(dones) - 1):
-            if (
-                np.linalg.norm(dataset_dict["observations"][i + 1] - dataset_dict["next_observations"][i])
-                > 1e-6
-                or dataset_dict["terminals"][i] == 1.0
-            ):
-                dones[i] = True
-
-        dones[-1] = True
-
-        dataset_dict["masks"] = 1.0 - dataset_dict["terminals"]
-        del dataset_dict["terminals"]
-
-        for k, v in dataset_dict.items():
-            dataset_dict[k] = v.astype(np.float32)
-
-        dataset_dict["dones"] = dones
-
-    if cfg.is_data_clip:
-        lim = 1 - cfg.data_clip_eps
-        dataset_dict["actions"] = np.clip(dataset_dict["actions"], -lim, lim)
-    reward_normalizer = get_reward_normalizer(cfg, dataset_dict)
-    dataset_dict["rewards"] = reward_normalizer(dataset_dict["rewards"])
-
-    for key in required_keys:
-        assert key in dataset_dict, f"Missing `{key}` in dataset."
-
-    if return_reward_normalizer:
-        return dataset_dict, reward_normalizer
-    return dataset_dict
-
-
-def get_trajectory_boundaries_and_returns(dataset):
-    """
-    Split dataset into trajectories and compute returns
-    """
-    episode_starts = [0]
-    episode_ends = []
-
-    episode_return = 0
-    episode_returns = []
-
-    n_transitions = len(dataset["rewards"])
-
-    for i in range(n_transitions):
-        episode_return += dataset["rewards"][i]
-
-        if dataset["dones"][i]:
-            episode_returns.append(episode_return)
-            episode_ends.append(i + 1)
-            if i + 1 < n_transitions:
-                episode_starts.append(i + 1)
-            episode_return = 0.0
-
-    return episode_starts, episode_ends, episode_returns
-
-
-def normalize_returns(dataset, scaling=1000):
-    """
-    Normalize returns in the dataset
-    """
-    (_, _, episode_returns) = get_trajectory_boundaries_and_returns(dataset)
-    dataset["rewards"] /= np.max(episode_returns) - np.min(episode_returns)
-    dataset["rewards"] *= scaling
-    return dataset
-
-
-def get_reward_normalizer(cfg, dataset):
-    """
-    Get a reward normalizer for the dataset
-    """
-    if cfg.task.startswith("xarm"):
-        return lambda x: x
-    elif "maze" in cfg.task:
-        return lambda x: x - 1.0
-    elif cfg.task.split("-")[0] in ["hopper", "halfcheetah", "walker2d"]:
-        (_, _, episode_returns) = get_trajectory_boundaries_and_returns(dataset)
-        return lambda x: x / (np.max(episode_returns) - np.min(episode_returns)) * 1000.0
-    elif hasattr(cfg, "reward_scale"):
-        return lambda x: x * cfg.reward_scale
-    return lambda x: x
-
-
-def linear_schedule(schdl, step):
-    """
-    Outputs values following a linear decay schedule.
-    Adapted from https://github.com/facebookresearch/drqv2
-    """
-    try:
-        return float(schdl)
-    except ValueError:
-        match = re.match(r"linear\((.+),(.+),(.+),(.+)\)", schdl)
-        if match:
-            init, final, start, end = (float(g) for g in match.groups())
-            mix = np.clip((step - start) / (end - start), 0.0, 1.0)
-            return (1.0 - mix) * init + mix * final
-        match = re.match(r"linear\((.+),(.+),(.+)\)", schdl)
-        if match:
-            init, final, duration = (float(g) for g in match.groups())
-            mix = np.clip(step / duration, 0.0, 1.0)
-            return (1.0 - mix) * init + mix * final
-    raise NotImplementedError(schdl)

lerobot/common/policies/utils.py

@@ -0,0 +1,49 @@
+#!/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 torch
+from torch import nn
+
+
+def populate_queues(queues, batch):
+    for key in batch:
+        # Ignore keys not in the queues already (leaving the responsibility to the caller to make sure the
+        # queues have the keys they want).
+        if key not in queues:
+            continue
+        if len(queues[key]) != queues[key].maxlen:
+            # initialize by copying the first observation several times until the queue is full
+            while len(queues[key]) != queues[key].maxlen:
+                queues[key].append(batch[key])
+        else:
+            # add latest observation to the queue
+            queues[key].append(batch[key])
+    return queues
+
+
+def get_device_from_parameters(module: nn.Module) -> torch.device:
+    """Get a module's device by checking one of its parameters.
+
+    Note: assumes that all parameters have the same device
+    """
+    return next(iter(module.parameters())).device
+
+
+def get_dtype_from_parameters(module: nn.Module) -> torch.dtype:
+    """Get a module's parameter dtype by checking one of its parameters.
+
+    Note: assumes that all parameters have the same dtype.
+    """
+    return next(iter(module.parameters())).dtype

lerobot/common/utils.py

@@ -1,45 +0,0 @@
-import logging
-import random
-from datetime import datetime
-
-import numpy as np
-import torch
-
-
-def set_seed(seed):
-    """Set seed for reproducibility."""
-    random.seed(seed)
-    np.random.seed(seed)
-    torch.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)
-
-
-def init_logging():
-    def custom_format(record):
-        dt = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-        fnameline = f"{record.pathname}:{record.lineno}"
-        message = f"{record.levelname} {dt} {fnameline[-15:]:>15} {record.msg}"
-        return message
-
-    logging.basicConfig(level=logging.INFO)
-
-    for handler in logging.root.handlers[:]:
-        logging.root.removeHandler(handler)
-
-    formatter = logging.Formatter()
-    formatter.format = custom_format
-    console_handler = logging.StreamHandler()
-    console_handler.setFormatter(formatter)
-    logging.getLogger().addHandler(console_handler)
-
-
-def format_big_number(num):
-    suffixes = ["", "K", "M", "B", "T", "Q"]
-    divisor = 1000.0
-
-    for suffix in suffixes:
-        if abs(num) < divisor:
-            return f"{num:.0f}{suffix}"
-        num /= divisor
-
-    return num

lerobot/common/utils/import_utils.py

@@ -0,0 +1,59 @@
+#!/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 logging
+
+
+def is_package_available(pkg_name: str, return_version: bool = False) -> tuple[bool, str] | bool:
+    """Copied from https://github.com/huggingface/transformers/blob/main/src/transformers/utils/import_utils.py
+    Check if the package spec exists and grab its version to avoid importing a local directory.
+    **Note:** this doesn't work for all packages.
+    """
+    package_exists = importlib.util.find_spec(pkg_name) is not None
+    package_version = "N/A"
+    if package_exists:
+        try:
+            # Primary method to get the package version
+            package_version = importlib.metadata.version(pkg_name)
+        except importlib.metadata.PackageNotFoundError:
+            # Fallback method: Only for "torch" and versions containing "dev"
+            if pkg_name == "torch":
+                try:
+                    package = importlib.import_module(pkg_name)
+                    temp_version = getattr(package, "__version__", "N/A")
+                    # Check if the version contains "dev"
+                    if "dev" in temp_version:
+                        package_version = temp_version
+                        package_exists = True
+                    else:
+                        package_exists = False
+                except ImportError:
+                    # If the package can't be imported, it's not available
+                    package_exists = False
+            else:
+                # For packages other than "torch", don't attempt the fallback and set as not available
+                package_exists = False
+        logging.debug(f"Detected {pkg_name} version: {package_version}")
+    if return_version:
+        return package_exists, package_version
+    else:
+        return package_exists
+
+
+_torch_available, _torch_version = is_package_available("torch", return_version=True)
+_gym_xarm_available = is_package_available("gym_xarm")
+_gym_aloha_available = is_package_available("gym_aloha")
+_gym_pusht_available = is_package_available("gym_pusht")

lerobot/common/utils/io_utils.py

@@ -0,0 +1,27 @@
+#!/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
+
+import imageio
+
+
+def write_video(video_path, stacked_frames, fps):
+    # Filter out DeprecationWarnings raised from pkg_resources
+    with warnings.catch_warnings():
+        warnings.filterwarnings(
+            "ignore", "pkg_resources is deprecated as an API", category=DeprecationWarning
+        )
+        imageio.mimsave(video_path, stacked_frames, fps=fps)

lerobot/common/utils/utils.py

@@ -0,0 +1,154 @@
+#!/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.path as osp
+import random
+from contextlib import contextmanager
+from datetime import datetime
+from pathlib import Path
+from typing import Generator
+
+import hydra
+import numpy as np
+import torch
+from omegaconf import DictConfig
+
+
+def get_safe_torch_device(cfg_device: str, log: bool = False) -> torch.device:
+    """Given a string, return a torch.device with checks on whether the device is available."""
+    match cfg_device:
+        case "cuda":
+            assert torch.cuda.is_available()
+            device = torch.device("cuda")
+        case "mps":
+            assert torch.backends.mps.is_available()
+            device = torch.device("mps")
+        case "cpu":
+            device = torch.device("cpu")
+            if log:
+                logging.warning("Using CPU, this will be slow.")
+        case _:
+            device = torch.device(cfg_device)
+            if log:
+                logging.warning(f"Using custom {cfg_device} device.")
+
+    return device
+
+
+def set_global_seed(seed):
+    """Set seed for reproducibility."""
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+
+
+@contextmanager
+def seeded_context(seed: int) -> Generator[None, None, None]:
+    """Set the seed when entering a context, and restore the prior random state at exit.
+
+    Example usage:
+
+    ```
+    a = random.random()  # produces some random number
+    with seeded_context(1337):
+        b = random.random()  # produces some other random number
+    c = random.random()  # produces yet another random number, but the same it would have if we never made `b`
+    ```
+    """
+    random_state = random.getstate()
+    np_random_state = np.random.get_state()
+    torch_random_state = torch.random.get_rng_state()
+    torch_cuda_random_state = torch.cuda.random.get_rng_state()
+    set_global_seed(seed)
+    yield None
+    random.setstate(random_state)
+    np.random.set_state(np_random_state)
+    torch.random.set_rng_state(torch_random_state)
+    torch.cuda.random.set_rng_state(torch_cuda_random_state)
+
+
+def init_logging():
+    def custom_format(record):
+        dt = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        fnameline = f"{record.pathname}:{record.lineno}"
+        message = f"{record.levelname} {dt} {fnameline[-15:]:>15} {record.msg}"
+        return message
+
+    logging.basicConfig(level=logging.INFO)
+
+    for handler in logging.root.handlers[:]:
+        logging.root.removeHandler(handler)
+
+    formatter = logging.Formatter()
+    formatter.format = custom_format
+    console_handler = logging.StreamHandler()
+    console_handler.setFormatter(formatter)
+    logging.getLogger().addHandler(console_handler)
+
+
+def format_big_number(num):
+    suffixes = ["", "K", "M", "B", "T", "Q"]
+    divisor = 1000.0
+
+    for suffix in suffixes:
+        if abs(num) < divisor:
+            return f"{num:.0f}{suffix}"
+        num /= divisor
+
+    return num
+
+
+def _relative_path_between(path1: Path, path2: Path) -> Path:
+    """Returns path1 relative to path2."""
+    path1 = path1.absolute()
+    path2 = path2.absolute()
+    try:
+        return path1.relative_to(path2)
+    except ValueError:  # most likely because path1 is not a subpath of path2
+        common_parts = Path(osp.commonpath([path1, path2])).parts
+        return Path(
+            "/".join([".."] * (len(path2.parts) - len(common_parts)) + list(path1.parts[len(common_parts) :]))
+        )
+
+
+def init_hydra_config(config_path: str, overrides: list[str] | None = None) -> DictConfig:
+    """Initialize a Hydra config given only the path to the relevant config file.
+
+    For config resolution, it is assumed that the config file's parent is the Hydra config dir.
+    """
+    # TODO(alexander-soare): Resolve configs without Hydra initialization.
+    hydra.core.global_hydra.GlobalHydra.instance().clear()
+    # Hydra needs a path relative to this file.
+    hydra.initialize(
+        str(_relative_path_between(Path(config_path).absolute().parent, Path(__file__).absolute().parent)),
+        version_base="1.2",
+    )
+    cfg = hydra.compose(Path(config_path).stem, overrides)
+    return cfg
+
+
+def print_cuda_memory_usage():
+    """Use this function to locate and debug memory leak."""
+    import gc
+
+    gc.collect()
+    # Also clear the cache if you want to fully release the memory
+    torch.cuda.empty_cache()
+    print("Current GPU Memory Allocated: {:.2f} MB".format(torch.cuda.memory_allocated(0) / 1024**2))
+    print("Maximum GPU Memory Allocated: {:.2f} MB".format(torch.cuda.max_memory_allocated(0) / 1024**2))
+    print("Current GPU Memory Reserved: {:.2f} MB".format(torch.cuda.memory_reserved(0) / 1024**2))
+    print("Maximum GPU Memory Reserved: {:.2f} MB".format(torch.cuda.max_memory_reserved(0) / 1024**2))

lerobot/configs/default.yaml

@@ -5,31 +5,42 @@ defaults:
 
 hydra:
   run:
-    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}_${env.name}_${policy.name}_${hydra.job.name}
+    dir: outputs/train/${now:%Y-%m-%d}/${now:%H-%M-%S}_${env.name}_${policy.name}_${hydra.job.name}
   job:
     name: default
 
-seed: 1337
 device: cuda  # cpu
-prefetch: 4
-eval_freq: ???
-save_freq: ???
-eval_episodes: ???
-save_video: false
-save_model: false
-save_buffer: false
-train_steps: ???
-fps: ???
+# `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
+# automatic gradient scaling is used.
+use_amp: false
+# `seed` is used for training (eg: model initialization, dataset shuffling)
+# AND for the evaluation environments.
+seed: ???
+dataset_repo_id: lerobot/pusht
 
-offline_prioritized_sampler: true
+training:
+  offline_steps: ???
+  # NOTE: `online_steps` is not implemented yet. It's here as a placeholder.
+  online_steps: ???
+  online_steps_between_rollouts: ???
+  online_sampling_ratio: 0.5
+  # `online_env_seed` is used for environments for online training data rollouts.
+  online_env_seed: ???
+  eval_freq: ???
+  save_freq: ???
+  log_freq: 250
+  save_model: true
 
-n_action_steps: ???
-env: ???
-
-policy: ???
+eval:
+  n_episodes: 1
+  # `batch_size` specifies the number of environments to use in a gym.vector.VectorEnv.
+  batch_size: 1
+  # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
+  use_async_envs: false
 
 wandb:
-  enable: true
+  enable: false
+  # Set to true to disable saving an artifact despite save_model == True
+  disable_artifact: false
   project: lerobot
-  entity: rcadene  # insert your own
   notes: ""

lerobot/configs/env/aloha.yaml

@@ -1,25 +1,14 @@
 # @package _global_
 
-eval_episodes: 50
-eval_freq: 7500
-save_freq: 75000
-log_freq: 250
-# TODO: same as simxarm, need to adjust
-offline_steps: 25000
-online_steps: 25000
-
 fps: 50
 
 env:
   name: aloha
-  task: sim_insertion_human
+  task: AlohaInsertion-v0
   from_pixels: True
   pixels_only: False
-  image_size: 96
-  action_repeat: 1
-  episode_length: 300
+  image_size: [3, 480, 640]
+  episode_length: 400
   fps: ${fps}
-
-policy:
-  state_dim: 2
-  action_dim: 2
+  state_dim: 14
+  action_dim: 14

lerobot/configs/env/aloha_thom.yaml

@@ -0,0 +1,14 @@
+# @package _global_
+
+fps: 50
+
+env:
+  name: aloha
+  task: AlohaInsertion-v0
+  from_pixels: True
+  pixels_only: False
+  image_size: [3, 480, 640]
+  episode_length: 500
+  fps: ${fps}
+  state_dim: 6
+  action_dim: 6