fix

Co-authored-by: Alexander Soare <alexander.soare159@gmail.com>
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

@@ -1,2 +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/PULL_REQUEST_TEMPLATE.md

@@ -1,15 +1,34 @@
-# What does this PR do?
+## What this does
+Explain what this PR does. Feel free to tag your PR with the appropriate label(s).
 
-Example: Fixes # (issue)
+Examples:
+|  Title               | Label           |
+|----------------------|-----------------|
+| Fixes #[issue]       | (🐛 Bug)        |
+| Adds new dataset     | (🗃️ Dataset)    |
+| Optimizes something  | (⚡️ Performance) |
 
+## How it was tested
+Explain/show how you tested your changes.
 
-## Before submitting
-- Read the [contributor guideline](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md#submitting-a-pull-request-pr).
-- Provide a minimal code example for the reviewer to checkout & try.
-- Explain 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.
 
-## Who can review?
+Examples:
+```bash
+DATA_DIR=tests/data pytest -sx tests/test_stuff.py::test_something
+```
+```bash
+python lerobot/scripts/train.py --some.option=true
+```
 
-Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
+## 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/poetry/cpu/pyproject.toml

@@ -1,107 +0,0 @@
-[tool.poetry]
-name = "lerobot"
-version = "0.1.0"
-description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
-authors = [
-    "Rémi Cadène <re.cadene@gmail.com>",
-    "Alexander Soare <alexander.soare159@gmail.com>",
-    "Quentin Gallouédec <quentin.gallouedec@ec-lyon.fr>",
-    "Simon Alibert <alibert.sim@gmail.com>",
-    "Thomas Wolf <thomaswolfcontact@gmail.com>",
-]
-repository = "https://github.com/huggingface/lerobot"
-readme = "README.md"
-license = "Apache-2.0"
-classifiers=[
-    "Development Status :: 3 - Alpha",
-    "Intended Audience :: Developers",
-    "Intended Audience :: Education",
-    "Intended Audience :: Science/Research",
-    "Topic :: Software Development :: Build Tools",
-    "Topic :: Scientific/Engineering :: Artificial Intelligence",
-    "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3.10",
-]
-packages = [{include = "lerobot"}]
-
-
-[tool.poetry.dependencies]
-python = "^3.10"
-termcolor = "^2.4.0"
-omegaconf = "^2.3.0"
-wandb = "^0.16.3"
-imageio = {extras = ["ffmpeg"], version = "^2.34.0"}
-gdown = "^5.1.0"
-hydra-core = "^1.3.2"
-einops = "^0.7.0"
-pymunk = "^6.6.0"
-zarr = "^2.17.0"
-numba = "^0.59.0"
-torch = {version = "^2.2.1", source = "torch-cpu"}
-opencv-python = "^4.9.0.80"
-diffusers = "^0.26.3"
-torchvision = {version = "^0.17.1", source = "torch-cpu"}
-h5py = "^3.10.0"
-huggingface-hub = "^0.21.4"
-robomimic = "0.2.0"
-gymnasium = "^0.29.1"
-cmake = "^3.29.0.1"
-gym-pusht = { git = "git@github.com:huggingface/gym-pusht.git", optional = true}
-gym-xarm = { git = "git@github.com:huggingface/gym-xarm.git", optional = true}
-gym-aloha = { git = "git@github.com:huggingface/gym-aloha.git", optional = true}
-pre-commit = {version = "^3.7.0", optional = true}
-debugpy = {version = "^1.8.1", optional = true}
-pytest = {version = "^8.1.0", optional = true}
-pytest-cov = {version = "^5.0.0", optional = true}
-datasets = "^2.19.0"
-
-
-[tool.poetry.extras]
-pusht = ["gym-pusht"]
-xarm = ["gym-xarm"]
-aloha = ["gym-aloha"]
-dev = ["pre-commit", "debugpy"]
-test = ["pytest", "pytest-cov"]
-
-
-[[tool.poetry.source]]
-name = "torch-cpu"
-url = "https://download.pytorch.org/whl/cpu"
-priority = "supplemental"
-
-
-[tool.ruff]
-line-length = 110
-target-version = "py310"
-exclude = [
-    ".bzr",
-    ".direnv",
-    ".eggs",
-    ".git",
-    ".git-rewrite",
-    ".hg",
-    ".mypy_cache",
-    ".nox",
-    ".pants.d",
-    ".pytype",
-    ".ruff_cache",
-    ".svn",
-    ".tox",
-    ".venv",
-    "__pypackages__",
-    "_build",
-    "buck-out",
-    "build",
-    "dist",
-    "node_modules",
-    "venv",
-]
-
-
-[tool.ruff.lint]
-select = ["E4", "E7", "E9", "F", "I", "N", "B", "C4", "SIM"]
-
-
-[build-system]
-requires = ["poetry-core>=1.5.0"]
-build-backend = "poetry.core.masonry.api"

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

@@ -0,0 +1,109 @@
+# 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"
+
+jobs:
+  latest-cpu:
+    name: CPU
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Install Git LFS
+        run: |
+          sudo apt-get update
+          sudo apt-get install git-lfs
+          git lfs install
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          lfs: true
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push CPU
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-cpu/Dockerfile
+          push: true
+          tags: huggingface/lerobot-cpu
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
+
+
+  latest-cuda:
+    name: GPU
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Install Git LFS
+        run: |
+          sudo apt-get update
+          sudo apt-get install git-lfs
+          git lfs install
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          lfs: true
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-gpu/Dockerfile
+          push: true
+          tags: huggingface/lerobot-gpu
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
+
+
+  latest-cuda-dev:
+    name: GPU Dev
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU dev
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-gpu-dev/Dockerfile
+          push: true
+          tags: huggingface/lerobot-gpu:dev
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}

.github/workflows/nightly-tests.yml

@@ -0,0 +1,83 @@
+# 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:
+      group: aws-general-8-plus
+    container:
+      image: huggingface/lerobot-cpu:latest
+      options: --shm-size "16gb"
+      credentials:
+        username: ${{ secrets.DOCKERHUB_USERNAME }}
+        password: ${{ secrets.DOCKERHUB_PASSWORD }}
+    defaults:
+      run:
+        shell: bash
+        working-directory: /lerobot
+    steps:
+      - name: Tests
+        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:
+      group: aws-g6-4xlarge-plus
+    env:
+      CUDA_VISIBLE_DEVICES: "0"
+      TEST_TYPE: "single_gpu"
+    container:
+      image: huggingface/lerobot-gpu:latest
+      options: --gpus all --shm-size "16gb"
+      credentials:
+        username: ${{ secrets.DOCKERHUB_USERNAME }}
+        password: ${{ secrets.DOCKERHUB_PASSWORD }}
+    defaults:
+      run:
+        shell: bash
+        working-directory: /lerobot
+    steps:
+      - name: Nvidia-smi
+        run: nvidia-smi
+
+      - name: Test
+        run: pytest -v --cov=./lerobot --cov-report=xml --disable-warnings tests
+      #   TODO(aliberts): Link with HF Codecov account
+      # - name: Upload coverage reports to Codecov with GitHub Action
+      #   uses: codecov/codecov-action@v4
+      #   with:
+      #     files: ./coverage.xml
+      #     verbose: true
+      - name: Tests end-to-end
+        env:
+          DEVICE: cuda
+        run: make test-end-to-end
+
+    #   - name: Generate Report
+    #     if: always()
+    #     run: |
+    #       pip install slack_sdk tabulate
+    #       python scripts/log_reports.py >> $GITHUB_STEP_SUMMARY

.github/workflows/quality.yml

@@ -0,0 +1,84 @@
+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
+
+
+  poetry_relax:
+    name: Poetry relax
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Install poetry
+        run: pipx install poetry
+
+      - name: Install poetry-relax
+        run: poetry self add poetry-relax
+
+      - name: Poetry relax
+        id: poetry_relax
+        run: |
+          output=$(poetry relax --check 2>&1)
+          if echo "$output" | grep -q "Proposing updates"; then
+            echo "$output"
+            echo ""
+            echo "Some dependencies have caret '^' version requirement added by poetry by default."
+            echo "Please replace them with '>='. You can do this by hand or use poetry-relax to do this."
+            exit 1
+          else
+            echo "$output"
+          fi

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

@@ -0,0 +1,65 @@
+# 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:
+      group: aws-general-8-plus
+    if: ${{ needs.get_changed_files.outputs.matrix }} != ''
+    strategy:
+      fail-fast: false
+      matrix:
+        docker-file: ${{ fromJson(needs.get_changed_files.outputs.matrix) }}
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - 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

@@ -4,210 +4,130 @@ on:
   pull_request:
     branches:
       - main
-    types: [opened, synchronize, reopened, labeled]
+    paths:
+      - "lerobot/**"
+      - "tests/**"
+      - "examples/**"
+      - ".github/**"
+      - "poetry.lock"
+      - "Makefile"
   push:
     branches:
       - main
+    paths:
+      - "lerobot/**"
+      - "tests/**"
+      - "examples/**"
+      - ".github/**"
+      - "poetry.lock"
+      - "Makefile"
 
 jobs:
-  tests:
-    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: 1.8.2
       DATA_DIR: tests/data
       MUJOCO_GL: egl
     steps:
-      #----------------------------------------------
-      #       check-out repo and set-up python
-      #----------------------------------------------
-      - name: Check out repository
-        uses: actions/checkout@v4
+      - uses: actions/checkout@v4
         with:
-          lfs: true
+          lfs: true  # Ensure LFS files are pulled
 
-      - name: Set up python
-        id: setup-python
+      - name: Install apt dependencies
+        run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev ffmpeg
+
+      - 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'
+          python-version: "3.10"
+          cache: "poetry"
 
-      - name: Add SSH key for installing envs
-        uses: webfactory/ssh-agent@v0.9.0
-        with:
-          ssh-private-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      #----------------------------------------------
-      #         install & configure poetry
-      #----------------------------------------------
-      - name: Load cached Poetry installation
-        id: restore-poetry-cache
-        uses: actions/cache/restore@v3
-        with:
-          path: ~/.local
-          key: poetry-${{ env.POETRY_VERSION }}
-
-      - name: Install Poetry
-        if: steps.restore-poetry-cache.outputs.cache-hit != 'true'
-        uses: snok/install-poetry@v1
-        with:
-          version: ${{ env.POETRY_VERSION }}
-          virtualenvs-create: true
-          installer-parallel: true
-
-      - name: Save cached Poetry installation
-        if: |
-          steps.restore-poetry-cache.outputs.cache-hit != 'true' &&
-          github.ref_name == 'main'
-        id: save-poetry-cache
-        uses: actions/cache/save@v3
-        with:
-          path: ~/.local
-          key: poetry-${{ env.POETRY_VERSION }}
-
-      - name: Configure Poetry
-        run: poetry config virtualenvs.in-project true
-
-      #----------------------------------------------
-      #           install dependencies
-      #----------------------------------------------
-      # TODO(aliberts): move to gpu runners
-      - name: Select cpu dependencies  # HACK
-        run: cp -t . .github/poetry/cpu/pyproject.toml .github/poetry/cpu/poetry.lock
-
-      - 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'
-        env:
-          TMPDIR: ~/tmp
-          TEMP: ~/tmp
-          TMP: ~/tmp
+      - name: Install poetry dependencies
         run: |
-          mkdir ~/tmp
-          poetry install --no-interaction --no-root --all-extras
+          poetry install --all-extras
 
-      - 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
+      - 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') }}
+          lfs: true  # Ensure LFS files are pulled
 
-      - name: Install libegl1-mesa-dev (to use MUJOCO_GL=egl)
+      - name: Install apt dependencies
+        run: sudo apt-get update && sudo apt-get install -y ffmpeg
+
+      - 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"
+
+      - name: Install poetry dependencies
+        run: |
+          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 apt dependencies
         run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
 
-      #----------------------------------------------
-      #             install project
-      #----------------------------------------------
-      - name: Install project
-        run: poetry install --no-interaction --all-extras
-
-      #----------------------------------------------
-      #            run tests & coverage
-      #----------------------------------------------
-      - name: Run tests
+      - name: Install poetry
         run: |
-          source .venv/bin/activate
-          pytest -v --cov=./lerobot --cov-report=xml tests
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
 
-      #   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: Set up Python 3.10
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+          cache: "poetry"
 
-      #----------------------------------------------
-      #            run end-to-end tests
-      #----------------------------------------------
-      - name: Test train ACT on ALOHA end-to-end
+      - name: Install poetry dependencies
         run: |
-          source .venv/bin/activate
-          python lerobot/scripts/train.py \
-            policy=act \
-            env=aloha \
-            wandb.enable=False \
-            offline_steps=2 \
-            online_steps=0 \
-            eval_episodes=1 \
-            device=cpu \
-            save_model=true \
-            save_freq=2 \
-            policy.n_action_steps=20 \
-            policy.chunk_size=20 \
-            policy.batch_size=2 \
-            hydra.run.dir=tests/outputs/act/
+          poetry install --all-extras
 
-      - name: Test eval ACT on ALOHA end-to-end
+      - name: Test end-to-end
         run: |
-          source .venv/bin/activate
-          python lerobot/scripts/eval.py \
-            --config tests/outputs/act/.hydra/config.yaml \
-            eval_episodes=1 \
-            env.episode_length=8 \
-            device=cpu \
-            policy.pretrained_model_path=tests/outputs/act/models/2.pt
-
-      - name: Test train Diffusion on PushT end-to-end
-        run: |
-          source .venv/bin/activate
-          python lerobot/scripts/train.py \
-            policy=diffusion \
-            env=pusht \
-            wandb.enable=False \
-            offline_steps=2 \
-            online_steps=0 \
-            eval_episodes=1 \
-            device=cpu \
-            save_model=true \
-            save_freq=2 \
-            policy.batch_size=2 \
-            hydra.run.dir=tests/outputs/diffusion/
-
-      - name: Test eval Diffusion on PushT end-to-end
-        run: |
-          source .venv/bin/activate
-          python lerobot/scripts/eval.py \
-            --config tests/outputs/diffusion/.hydra/config.yaml \
-            eval_episodes=1 \
-            env.episode_length=8 \
-            device=cpu \
-            policy.pretrained_model_path=tests/outputs/diffusion/models/2.pt
-
-      - name: Test train TDMPC on Simxarm end-to-end
-        run: |
-          source .venv/bin/activate
-          python lerobot/scripts/train.py \
-            policy=tdmpc \
-            env=xarm \
-            wandb.enable=False \
-            offline_steps=1 \
-            online_steps=2 \
-            eval_episodes=1 \
-            env.episode_length=2 \
-            device=cpu \
-            save_model=true \
-            save_freq=2 \
-            policy.batch_size=2 \
-            hydra.run.dir=tests/outputs/tdmpc/
-
-      - name: Test eval TDMPC on Simxarm end-to-end
-        run: |
-          source .venv/bin/activate
-          python lerobot/scripts/eval.py \
-            --config tests/outputs/tdmpc/.hydra/config.yaml \
-            eval_episodes=1 \
-            env.episode_length=8 \
-            device=cpu \
-            policy.pretrained_model_path=tests/outputs/tdmpc/models/2.pt
+          make test-end-to-end \
+            && rm -rf outputs

.github/workflows/trufflehog.yml

@@ -0,0 +1,20 @@
+on:
+  push:
+
+name: Secret Leaks
+
+permissions:
+  contents: read
+
+jobs:
+  trufflehog:
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+      with:
+        fetch-depth: 0
+    - name: Secret Scanning
+      uses: trufflesecurity/trufflehog@main
+      with:
+        extra_args: --only-verified

.gitignore

@@ -2,10 +2,16 @@
 logs
 tmp
 wandb
+
+# Data
 data
 outputs
+
+# Apple
+.DS_Store
+
+# VS Code
 .vscode
-rl
 
 # HPC
 nautilus/*.yaml
@@ -89,6 +95,7 @@ instance/
 docs/_build/
 
 # PyBuilder
+.pybuilder/
 target/
 
 # Jupyter Notebook
@@ -101,13 +108,6 @@ ipython_config.py
 # pyenv
 .python-version
 
-# pipenv
-#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
-#   However, in case of collaboration, if having platform-specific dependencies or dependencies
-#   having no cross-platform support, pipenv may install dependencies that don't work, or not
-#   install all needed dependencies.
-#Pipfile.lock
-
 # PEP 582; used by e.g. github.com/David-OConnor/pyflow
 __pypackages__/
 
@@ -118,6 +118,14 @@ celerybeat.pid
 # SageMath parsed files
 *.sage.py
 
+# Environments
+.env
+.venv
+env/
+venv/
+env.bak/
+venv.bak/
+
 # Spyder project settings
 .spyderproject
 .spyproject
@@ -135,3 +143,9 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/

.pre-commit-config.yaml

@@ -1,4 +1,4 @@
-exclude: ^(data/|tests/data)
+exclude: ^(tests/data)
 default_language_version:
     python: python3.10
 repos:
@@ -14,11 +14,11 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.15.2
+    rev: v3.16.0
     hooks:
     -   id: pyupgrade
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.3.7
+    rev: v0.5.2
     hooks:
       - id: ruff
         args: [--fix]
@@ -31,3 +31,7 @@ repos:
         args:
           - "--check"
           - "--no-update"
+  - repo: https://github.com/gitleaks/gitleaks
+    rev: v8.18.4
+    hooks:
+      - id: gitleaks

CONTRIBUTING.md

@@ -129,41 +129,38 @@ Follow these steps to start contributing:
 
    🚨 **Do not** work on the `main` branch.
 
-4. Instead of using `pip` directly, we use `poetry` for development purposes to easily track our dependencies.
+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 by running the following command in a conda or a virtual environment you've created for working on this library:
-   Install the project with dev dependencies and all environments:
-   ```bash
-   poetry install --sync --with dev --all-extras
-   ```
-   This command should be run when pulling code with and updated version of `pyproject.toml` and `poetry.lock` in order to synchronize your virtual environment with the dependencies.
 
-   To selectively install environments (for example aloha and pusht) use:
+   Set up a development environment with conda or miniconda:
    ```bash
-   poetry install --sync --with dev --extras "aloha pusht"
+   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 changes are made to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
+   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
    ```
 
-   **NOTE:** Currently, to ensure the CI works properly, any new package must also be added in the    CPU-only environment dedicated to the CI. To do this, you should create a separate environment and    add the new package there as well. For example:
-   ```bash
-   # Add the new package to your main poetry env
-   poetry add some-package
-   # Add the same package to the CPU-only env dedicated to CI
-   conda create -y -n lerobot-ci python=3.10
-   conda activate lerobot-ci
-   cd .github/poetry/cpu
-   poetry add some-package
-   ```
-
 5. Develop the features on your branch.
 
    As you work on the features, you should make sure that the test suite
@@ -198,6 +195,11 @@ Follow these steps to start contributing:
    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

Makefile

@@ -0,0 +1,190 @@
+.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)
+
+DEVICE ?= cpu
+
+build-cpu:
+	docker build -t lerobot:latest -f docker/lerobot-cpu/Dockerfile .
+
+build-gpu:
+	docker build -t lerobot:latest -f docker/lerobot-gpu/Dockerfile .
+
+test-end-to-end:
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train-amp
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-eval-amp
+	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train-with-online
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-default-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-pusht-tutorial
+
+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=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		policy.n_action_steps=20 \
+		policy.chunk_size=20 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/act/
+
+test-act-ete-eval:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/act/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+
+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=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		policy.n_action_steps=20 \
+		policy.chunk_size=20 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/act_amp/ \
+		training.image_transforms.enable=true \
+		use_amp=true
+
+test-act-ete-eval-amp:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/act_amp/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+		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=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/diffusion/
+
+test-diffusion-ete-eval:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/diffusion/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+
+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=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/tdmpc/
+
+test-tdmpc-ete-train-with-online:
+	python lerobot/scripts/train.py \
+		env=pusht \
+		env.gym.obs_type=environment_state_agent_pos \
+		policy=tdmpc_pusht_keypoints \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=10 \
+		device=$(DEVICE) \
+		training.offline_steps=2 \
+		training.online_steps=20 \
+		training.save_checkpoint=false \
+		training.save_freq=10 \
+		training.batch_size=2 \
+		training.online_rollout_n_episodes=2 \
+		training.online_rollout_batch_size=2 \
+		training.online_steps_between_rollouts=10 \
+		training.online_buffer_capacity=15 \
+		eval.use_async_envs=true \
+		hydra.run.dir=tests/outputs/tdmpc_online/
+
+
+test-tdmpc-ete-eval:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+
+test-default-ete-eval:
+	python lerobot/scripts/eval.py \
+		--config lerobot/configs/default.yaml \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+
+test-act-pusht-tutorial:
+	cp examples/advanced/1_train_act_pusht/act_pusht.yaml lerobot/configs/policy/created_by_Makefile.yaml
+	python lerobot/scripts/train.py \
+		policy=created_by_Makefile.yaml \
+		env=pusht \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=2 \
+		device=$(DEVICE) \
+		training.save_model=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/act_pusht/
+	rm lerobot/configs/policy/created_by_Makefile.yaml

README.md

@@ -10,7 +10,7 @@
 
 <div align="center">
 
-[![Tests](https://github.com/huggingface/lerobot/actions/workflows/test.yml/badge.svg?branch=main)](https://github.com/huggingface/lerobot/actions/workflows/test.yml?query=branch%3Amain)
+[![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)
@@ -22,22 +22,36 @@
 
 </div>
 
+<h2 align="center">
+    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md">Hot new tutorial: Getting started with real-world robots</a></p>
+</h2>
+
+<div align="center">
+    <img src="media/tutorial/koch_v1_1_leader_follower.webp?raw=true" alt="Koch v1.1 leader and follower arms" title="Koch v1.1 leader and follower arms" width="50%">
+    <p>We just dropped an in-depth tutorial on how to build your own robot!</p>
+    <p>Teach it new skills by showing it a few moves with just a laptop.</p>
+    <p>Then watch your homemade robot act autonomously 🤯</p>
+    <p>For more info, see <a href="https://x.com/RemiCadene/status/1825455895561859185">our thread on X</a> or <a href="https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md">our tutorial page</a>.</p>
+</div>
+
+<br/>
+
 <h3 align="center">
-    <p>State-of-the-art Machine Learning for real-world robotics</p>
+    <p>LeRobot: State-of-the-art AI for real-world robotics</p>
 </h3>
 
 ---
 
 
-🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier for entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
+🤗 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 simulated environments so that everyone can get started. 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 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 HuggingFace community page: [huggingface.co/lerobot](https://huggingface.co/lerobot)
+🤗 LeRobot hosts pretrained models and datasets on this Hugging Face community page: [huggingface.co/lerobot](https://huggingface.co/lerobot)
 
-#### Examples of pretrained models and environments
+#### Examples of pretrained models on simulation environments
 
 <table>
   <tr>
@@ -54,28 +68,36 @@
 
 ### Acknowledgment
 
-- ACT policy and ALOHA environment are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha/)
-- Diffusion policy and Pusht environment are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu/)
-- TDMPC policy and Simxarm environment are adapted from [FOWM](https://www.yunhaifeng.com/FOWM/)
-- Abstractions and utilities for Reinforcement Learning come from [TorchRL](https://github.com/pytorch/rl)
+- 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.
+- Thanks to [Seungjae (Jay) Lee](https://sjlee.cc/), [Mahi Shafiullah](https://mahis.life/) and colleagues for open sourcing [VQ-BeT](https://sjlee.cc/vq-bet/) policy and helping us adapt the codebase to our repository. The policy is adapted from [VQ-BeT repo](https://github.com/jayLEE0301/vq_bet_official).
+
 
 ## Installation
 
 Download our source code:
 ```bash
-git clone https://github.com/huggingface/lerobot.git && cd lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
 ```
 
 Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/free/miniconda/index.html):
 ```bash
-conda create -y -n lerobot python=3.10 && conda activate lerobot
+conda create -y -n lerobot python=3.10
+conda activate lerobot
 ```
 
 Install 🤗 LeRobot:
 ```bash
-python -m pip install .
+pip install -e .
 ```
 
+> **NOTE:** Depending on your platform, If you encounter any build errors during this step
+you may need to install `cmake` and `build-essential` for building some of our dependencies.
+On linux: `sudo apt-get install cmake build-essential`
+
 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)
@@ -83,18 +105,22 @@ For simulations, 🤗 LeRobot comes with gymnasium environments that can be inst
 
 For instance, to install 🤗 LeRobot with aloha and pusht, use:
 ```bash
-python -m pip install ".[aloha, pusht]"
+pip install -e ".[aloha, pusht]"
 ```
 
-To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiments tracking, log in with
+To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
 ```bash
 wandb login
 ```
 
+(note: you will also need to enable WandB in the configuration. See below.)
+
 ## Walkthrough
 
 ```
 .
+├── examples             # contains demonstration examples, start here to learn about LeRobot
+|   └── advanced         # contains even more examples for those who have mastered the basics
 ├── lerobot
 |   ├── configs          # contains 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
@@ -103,69 +129,174 @@ wandb login
 |   ├── 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
-|   └── scripts                  # contains functions to execute via command line
-|       ├── visualize_dataset.py  # load a dataset and render its demonstrations
-|       ├── eval.py               # load policy and evaluate it on an environment
-|       └── train.py              # train a policy via imitation learning and/or reinforcement learning
+|   |   ├── policies       # various policies: act, diffusion, tdmpc
+|   |   ├── robot_devices  # various real devices: dynamixel motors, opencv cameras, koch robots
+|   |   └── utils          # various utilities
+|   └── scripts          # contains functions to execute via command line
+|       ├── eval.py                 # load policy and evaluate it on an environment
+|       ├── train.py                # train a policy via imitation learning and/or reinforcement learning
+|       ├── control_robot.py        # teleoperate a real robot, record data, run a policy
+|       ├── push_dataset_to_hub.py  # convert your dataset into LeRobot dataset format and upload it to the Hugging Face hub
+|       └── visualize_dataset.py    # load a dataset and render its demonstrations
 ├── outputs               # contains results of scripts execution: logs, videos, model checkpoints
-├── .github
-|   └── workflows
-|       └── test.yml      # defines install settings for continuous integration and specifies end-to-end tests
 └── tests                 # contains pytest utilities for continuous integration
-
 ```
 
 ### Visualize datasets
 
-Check out [examples](./examples) to see how you can import our dataset class, download the data from the HuggingFace hub and use our rendering utilities.
+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.
 
-Or you can achieve the same result by executing our script from the command line:
+You can also locally visualize episodes from a dataset on the hub by executing our script from the command line:
 ```bash
 python lerobot/scripts/visualize_dataset.py \
-env=pusht \
-hydra.run.dir=outputs/visualize_dataset/example
-# >>> ['outputs/visualize_dataset/example/episode_0.mp4']
+    --repo-id lerobot/pusht \
+    --episode-index 0
 ```
 
+or from a dataset in a local folder with the root `DATA_DIR` environment variable (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
+```bash
+DATA_DIR='./my_local_data_dir' python lerobot/scripts/visualize_dataset.py \
+    --repo-id lerobot/pusht \
+    --episode-index 0
+```
+
+
+It will open `rerun.io` and display the camera streams, robot states and actions, like this:
+
+https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-fd46b787-b532-47e2-bb6f-fd536a55a7ed.mov?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20240505%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240505T172924Z&X-Amz-Expires=300&X-Amz-Signature=d680b26c532eeaf80740f08af3320d22ad0b8a4e4da1bcc4f33142c15b509eda&X-Amz-SignedHeaders=host&actor_id=24889239&key_id=0&repo_id=748713144
+
+
+Our script can also visualize datasets stored on a distant server. See `python lerobot/scripts/visualize_dataset.py --help` for more instructions.
+
+### The `LeRobotDataset` format
+
+A dataset in `LeRobotDataset` format is very simple to use. It can be loaded from a repository on the Hugging Face hub or a local folder simply with e.g. `dataset = LeRobotDataset("lerobot/aloha_static_coffee")` and can be indexed into like any Hugging Face and PyTorch dataset. For instance `dataset[0]` will retrieve a single temporal frame from the dataset containing observation(s) and an action as PyTorch tensors ready to be fed to a model.
+
+A specificity of `LeRobotDataset` is that, rather than retrieving a single frame by its index, we can retrieve several frames based on their temporal relationship with the indexed frame, by setting `delta_timestamps` to a list of relative times with respect to the indexed frame. For example, with `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}`  one can retrieve, for a given index, 4 frames: 3 "previous" frames 1 second, 0.5 seconds, and 0.2 seconds before the indexed frame, and the indexed frame itself (corresponding to the 0 entry). See example [1_load_lerobot_dataset.py](examples/1_load_lerobot_dataset.py) for more details on `delta_timestamps`.
+
+Under the hood, the `LeRobotDataset` format makes use of several ways to serialize data which can be useful to understand if you plan to work more closely with this format. We tried to make a flexible yet simple dataset format that would cover most type of features and specificities present in reinforcement learning and robotics, in simulation and in real-world, with a focus on cameras and robot states but easily extended to other types of sensory inputs as long as they can be represented by a tensor.
+
+Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
+
+```
+dataset attributes:
+  ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
+  │  ├ observation.images.cam_high (VideoFrame):
+  │  │   VideoFrame = {'path': path to a mp4 video, 'timestamp' (float32): timestamp in the video}
+  │  ├ observation.state (list of float32): position of an arm joints (for instance)
+  │  ... (more observations)
+  │  ├ action (list of float32): goal position of an arm joints (for instance)
+  │  ├ episode_index (int64): index of the episode for this sample
+  │  ├ frame_index (int64): index of the frame for this sample in the episode ; starts at 0 for each episode
+  │  ├ timestamp (float32): timestamp in the episode
+  │  ├ next.done (bool): indicates the end of en episode ; True for the last frame in each episode
+  │  └ index (int64): general index in the whole dataset
+  ├ episode_data_index: contains 2 tensors with the start and end indices of each episode
+  │  ├ from (1D int64 tensor): first frame index for each episode — shape (num episodes,) starts with 0
+  │  └ to: (1D int64 tensor): last frame index for each episode — shape (num episodes,)
+  ├ stats: a dictionary of statistics (max, mean, min, std) for each feature in the dataset, for instance
+  │  ├ observation.images.cam_high: {'max': tensor with same number of dimensions (e.g. `(c, 1, 1)` for images, `(c,)` for states), etc.}
+  │  ...
+  ├ info: a dictionary of metadata on the dataset
+  │  ├ codebase_version (str): this is to keep track of the codebase version the dataset was created with
+  │  ├ fps (float): frame per second the dataset is recorded/synchronized to
+  │  ├ video (bool): indicates if frames are encoded in mp4 video files to save space or stored as png files
+  │  └ encoding (dict): if video, this documents the main options that were used with ffmpeg to encode the videos
+  ├ videos_dir (Path): where the mp4 videos or png images are stored/accessed
+  └ camera_keys (list of string): the keys to access camera features in the item returned by the dataset (e.g. `["observation.images.cam_high", ...]`)
+```
+
+A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
+- hf_dataset stored using Hugging Face datasets library serialization to parquet
+- videos are stored in mp4 format to save space or png files
+- episode_data_index saved using `safetensor` tensor serialization format
+- stats saved using `safetensor` tensor serialization format
+- info are saved using JSON
+
+Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can set the `DATA_DIR` environment variable to your root dataset folder as illustrated in the above section on dataset visualization.
+
 ### Evaluate a pretrained policy
 
-Check out [examples](./examples) to see how you can load a pretrained policy from HuggingFace hub, load up the corresponding environment and model, and run an evaluation.
+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.
 
-Or you can achieve the same result by executing our script from the command line:
+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 \
---hub-id lerobot/diffusion_policy_pusht_image \
-eval_episodes=10 \
-hydra.run.dir=outputs/eval/example_hub
+    -p lerobot/diffusion_pusht \
+    eval.n_episodes=10 \
+    eval.batch_size=10
 ```
 
-After training your own policy, you can also re-evaluate the checkpoints with:
+Note: After training your own policy, you can re-evaluate the checkpoints with:
+
 ```bash
-python lerobot/scripts/eval.py \
---config PATH/TO/FOLDER/config.yaml \
-policy.pretrained_model_path=PATH/TO/FOLDER/weights.pth \
-eval_episodes=10 \
-hydra.run.dir=outputs/eval/example_dir
+python lerobot/scripts/eval.py -p {OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```
 
 See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-Check out [examples](./examples) to see how you can start training a model on a dataset, which will be automatically downloaded if needed.
+Check out [example 3](./examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
+
+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:
 
-In general, you can use our training script to easily train any policy on any environment:
 ```bash
 python lerobot/scripts/train.py \
-env=aloha \
-task=sim_insertion \
-repo_id=lerobot/aloha_sim_insertion_scripted \
-policy=act \
-hydra.run.dir=outputs/train/aloha_act
+    policy=act \
+    env=aloha \
+    env.task=AlohaInsertion-v0 \
+    dataset_repo_id=lerobot/aloha_sim_insertion_human \
 ```
 
-After training, you may want to revisit model evaluation to change the evaluation settings. In fact, during training every checkpoint is already evaluated but on a low number of episodes for efficiency. Check out [example](./examples) to evaluate any model checkpoint on more episodes to increase statistical significance.
+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
+```
+
+In the experiment directory there will be a folder called `checkpoints` which will have the following structure:
+
+```bash
+checkpoints
+├── 000250  # checkpoint_dir for training step 250
+│   ├── pretrained_model  # Hugging Face pretrained model dir
+│   │   ├── config.json  # Hugging Face pretrained model config
+│   │   ├── config.yaml  # consolidated Hydra config
+│   │   ├── model.safetensors  # model weights
+│   │   └── README.md  # Hugging Face model card
+│   └── training_state.pth  # optimizer/scheduler/rng state and training step
+```
+
+To resume training from a checkpoint, you can add these to the `train.py` python command:
+```bash
+    hydra.run.dir=your/original/experiment/dir resume=true
+```
+
+It will load the pretrained model, optimizer and scheduler states for training. For more information please see our tutorial on training resumption [here](https://github.com/huggingface/lerobot/blob/main/examples/5_resume_training.md).
+
+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. Please also check [here](https://github.com/huggingface/lerobot/blob/main/examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explaination of some commonly used metrics in logs.
+
+![](media/wandb.png)
+
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
+
+#### Reproduce state-of-the-art (SOTA)
+
+We have organized our configuration files (found under [`lerobot/configs`](./lerobot/configs)) such that they reproduce SOTA results from a given model variant in their respective original works. Simply running:
+
+```bash
+python lerobot/scripts/train.py policy=diffusion env=pusht
+```
+
+reproduces SOTA results for Diffusion Policy on the PushT task.
+
+Pretrained policies, along with reproduction details, can be found under the "Models" section of https://huggingface.co/lerobot.
 
 ## Contribute
 
@@ -173,105 +304,40 @@ If you would like to contribute to 🤗 LeRobot, please check out our [contribut
 
 ### Add a new dataset
 
-```python
-# TODO(rcadene, AdilZouitine): rewrite this section
-```
-
-To add a dataset to the hub, first login and use a token generated from [huggingface settings](https://huggingface.co/settings/tokens) with write access:
+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 you can upload it to the hub with:
+Then point to your raw dataset folder (e.g. `data/aloha_static_pingpong_test_raw`), and push your dataset to the hub with:
 ```bash
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload $HF_USER/$DATASET data/$DATASET \
---repo-type dataset  \
---revision v1.0
+python lerobot/scripts/push_dataset_to_hub.py \
+--raw-dir data/aloha_static_pingpong_test_raw \
+--out-dir data \
+--repo-id lerobot/aloha_static_pingpong_test \
+--raw-format aloha_hdf5
 ```
 
-You will need to set the corresponding version as a default argument in your dataset class:
-```python
-  version: str | None = "v1.1",
-```
-See: [`lerobot/common/datasets/pusht.py`](https://github.com/Cadene/lerobot/blob/main/lerobot/common/datasets/pusht.py)
+See `python lerobot/scripts/push_dataset_to_hub.py --help` for more instructions.
 
-For instance, for [lerobot/pusht](https://huggingface.co/datasets/lerobot/pusht), we used:
-```bash
-HF_USER=lerobot
-DATASET=pusht
-```
+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).
 
-If you want to improve an existing dataset, you can download it locally with:
-```bash
-mkdir -p data/$DATASET
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download ${HF_USER}/$DATASET \
---repo-type dataset \
---local-dir data/$DATASET \
---local-dir-use-symlinks=False \
---revision v1.0
-```
-
-Iterate on your code and dataset with:
-```bash
-DATA_DIR=data python train.py
-```
-
-Upload a new version (v2.0 or v1.1 if the changes are respectively more or less significant):
-```bash
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload $HF_USER/$DATASET data/$DATASET \
---repo-type dataset \
---revision v1.1 \
---delete "*"
-```
-
-Then you will need to set the corresponding version as a default argument in your dataset class:
-```python
-  version: str | None = "v1.1",
-```
-See: [`lerobot/common/datasets/pusht.py`](https://github.com/Cadene/lerobot/blob/main/lerobot/common/datasets/pusht.py)
-
-
-Finally, you might want to mock the dataset if you need to update the unit tests as well:
-```bash
-python tests/scripts/mock_dataset.py --in-data-dir data/$DATASET --out-data-dir tests/data/$DATASET
-```
 
 ### Add a pretrained policy
 
-```python
-# TODO(rcadene, alexander-soare): rewrite this section
-```
+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)).
 
-Once you have trained a policy you may upload it to the HuggingFace hub.
-
-Firstly, make sure you have a model repository set up on the hub. The hub ID looks like HF_USER/REPO_NAME.
-
-Secondly, assuming you have trained a policy, you need:
-
-- `config.yaml` which you can get from the `.hydra` directory of your training output folder.
-- `model.pt` which should be one of the saved models in the `models` directory of your training output folder (they won't be named `model.pt` but you will need to choose one).
-
-To upload these to the hub, prepare a folder with the following structure (you can use symlinks rather than copying):
-
-```
-to_upload
-    ├── config.yaml
-    └── model.pt
-```
-
-With the folder prepared, run the following with a desired revision ID.
+You first need to find the checkpoint folder located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). Within that there is a `pretrained_model` directory which should contain:
+- `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
+- `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
+- `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 $HUB_ID to_upload --revision $REVISION_ID
+huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
 ```
 
-If you want this to be the default revision also run the following (don't worry, it won't upload the files again; it will just adjust the file pointers):
-
-```bash
-huggingface-cli upload $HUB_ID to_upload
-```
-
-See `eval.py` for an example of how a user may use your policy.
+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
@@ -298,9 +364,56 @@ with profile(
             # insert code to profile, potentially whole body of eval_policy function
 ```
 
-```bash
-python lerobot/scripts/eval.py \
---config outputs/pusht/.hydra/config.yaml \
-pretrained_model_path=outputs/pusht/model.pt \
-eval_episodes=7
+## Citation
+
+If you want, you can cite this work with:
+```bibtex
+@misc{cadene2024lerobot,
+    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Wolf, Thomas},
+    title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
+    howpublished = "\url{https://github.com/huggingface/lerobot}",
+    year = {2024}
+}
+```
+
+Additionally, if you are using any of the particular policy architecture, pretrained models, or datasets, it is recommended to cite the original authors of the work as they appear below:
+
+- [Diffusion Policy](https://diffusion-policy.cs.columbia.edu)
+```bibtex
+@article{chi2024diffusionpolicy,
+	author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
+	title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+	journal = {The International Journal of Robotics Research},
+	year = {2024},
+}
+```
+- [ACT or ALOHA](https://tonyzhaozh.github.io/aloha)
+```bibtex
+@article{zhao2023learning,
+  title={Learning fine-grained bimanual manipulation with low-cost hardware},
+  author={Zhao, Tony Z and Kumar, Vikash and Levine, Sergey and Finn, Chelsea},
+  journal={arXiv preprint arXiv:2304.13705},
+  year={2023}
+}
+```
+
+- [TDMPC](https://www.nicklashansen.com/td-mpc/)
+
+```bibtex
+@inproceedings{Hansen2022tdmpc,
+	title={Temporal Difference Learning for Model Predictive Control},
+	author={Nicklas Hansen and Xiaolong Wang and Hao Su},
+	booktitle={ICML},
+	year={2022}
+}
+```
+
+- [VQ-BeT](https://sjlee.cc/vq-bet/)
+```bibtex
+@article{lee2024behavior,
+  title={Behavior generation with latent actions},
+  author={Lee, Seungjae and Wang, Yibin and Etukuru, Haritheja and Kim, H Jin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},
+  journal={arXiv preprint arXiv:2403.03181},
+  year={2024}
+}
 ```

benchmarks/video/README.md

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

benchmarks/video/capture_camera_feed.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Capture video feed from a camera as raw images."""
+
+import argparse
+import datetime as dt
+from pathlib import Path
+
+import cv2
+
+
+def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int):
+    now = dt.datetime.now()
+    capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
+    if not capture_dir.exists():
+        capture_dir.mkdir(parents=True, exist_ok=True)
+
+    # Opens the default webcam
+    cap = cv2.VideoCapture(0)
+    if not cap.isOpened():
+        print("Error: Could not open video stream.")
+        return
+
+    cap.set(cv2.CAP_PROP_FPS, fps)
+    cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
+    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
+
+    frame_index = 0
+    while True:
+        ret, frame = cap.read()
+
+        if not ret:
+            print("Error: Could not read frame.")
+            break
+
+        cv2.imshow("Video Stream", frame)
+        cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
+        frame_index += 1
+
+        # Break the loop on 'q' key press
+        if cv2.waitKey(1) & 0xFF == ord("q"):
+            break
+
+    # Release the capture and destroy all windows
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path("outputs/cam_capture/"),
+        help="Directory where the capture images are written. A subfolder named with the current date & time will be created inside it for each capture.",
+    )
+    parser.add_argument(
+        "--fps",
+        type=int,
+        default=30,
+        help="Frames Per Second of the capture.",
+    )
+    parser.add_argument(
+        "--width",
+        type=int,
+        default=1280,
+        help="Width of the captured images.",
+    )
+    parser.add_argument(
+        "--height",
+        type=int,
+        default=720,
+        help="Height of the captured images.",
+    )
+    args = parser.parse_args()
+    display_and_save_video_stream(**vars(args))

benchmarks/video/run_video_benchmark.py

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

docker/lerobot-cpu/Dockerfile

@@ -0,0 +1,32 @@
+# Configure image
+ARG PYTHON_VERSION=3.10
+
+FROM python:${PYTHON_VERSION}-slim
+ARG PYTHON_VERSION
+ARG DEBIAN_FRONTEND=noninteractive
+
+# Install apt dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    speech-dispatcher \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Create virtual environment
+RUN ln -s /usr/bin/python${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, koch]" \
+    --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-dev/Dockerfile

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

docker/lerobot-gpu/Dockerfile

@@ -0,0 +1,30 @@
+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 \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    speech-dispatcher \
+    python${PYTHON_VERSION}-dev python${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
+COPY . /lerobot
+WORKDIR /lerobot
+RUN pip install --upgrade --no-cache-dir pip
+RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, koch]"
+
+# Set EGL as the rendering backend for MuJoCo
+ENV MUJOCO_GL="egl"

download_and_upload_dataset.py

@@ -1,550 +0,0 @@
-"""
-This file contains all obsolete download scripts. They are centralized here to not have to load
-useless dependencies when using datasets.
-"""
-
-import io
-import json
-import pickle
-import shutil
-from pathlib import Path
-
-import einops
-import h5py
-import numpy as np
-import torch
-import tqdm
-from datasets import Dataset, Features, Image, Sequence, Value
-from huggingface_hub import HfApi
-from PIL import Image as PILImage
-from safetensors.torch import save_file
-
-from lerobot.common.datasets.utils import compute_stats, flatten_dict, hf_transform_to_torch
-
-
-def download_and_upload(root, revision, dataset_id):
-    # TODO(rcadene, adilzouitine): add community_id/user_id (e.g. "lerobot", "cadene") or repo_id (e.g. "lerobot/pusht")
-    if "pusht" in dataset_id:
-        download_and_upload_pusht(root, revision, dataset_id)
-    elif "xarm" in dataset_id:
-        download_and_upload_xarm(root, revision, dataset_id)
-    elif "aloha" in dataset_id:
-        download_and_upload_aloha(root, revision, dataset_id)
-    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)
-        return True
-    else:
-        return False
-
-
-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 push_to_hub(hf_dataset, episode_data_index, info, stats, root, revision, dataset_id):
-    # push to main to indicate latest version
-    hf_dataset.push_to_hub(f"lerobot/{dataset_id}", token=True)
-
-    # push to version branch
-    hf_dataset.push_to_hub(f"lerobot/{dataset_id}", token=True, revision=revision)
-
-    # create and store meta_data
-    meta_data_dir = root / dataset_id / "meta_data"
-    meta_data_dir.mkdir(parents=True, exist_ok=True)
-
-    api = HfApi()
-
-    # info
-    info_path = meta_data_dir / "info.json"
-    with open(str(info_path), "w") as f:
-        json.dump(info, f, indent=4)
-    api.upload_file(
-        path_or_fileobj=info_path,
-        path_in_repo=str(info_path).replace(f"{root}/{dataset_id}", ""),
-        repo_id=f"lerobot/{dataset_id}",
-        repo_type="dataset",
-    )
-    api.upload_file(
-        path_or_fileobj=info_path,
-        path_in_repo=str(info_path).replace(f"{root}/{dataset_id}", ""),
-        repo_id=f"lerobot/{dataset_id}",
-        repo_type="dataset",
-        revision=revision,
-    )
-
-    # stats
-    stats_path = meta_data_dir / "stats.safetensors"
-    save_file(flatten_dict(stats), stats_path)
-    api.upload_file(
-        path_or_fileobj=stats_path,
-        path_in_repo=str(stats_path).replace(f"{root}/{dataset_id}", ""),
-        repo_id=f"lerobot/{dataset_id}",
-        repo_type="dataset",
-    )
-    api.upload_file(
-        path_or_fileobj=stats_path,
-        path_in_repo=str(stats_path).replace(f"{root}/{dataset_id}", ""),
-        repo_id=f"lerobot/{dataset_id}",
-        repo_type="dataset",
-        revision=revision,
-    )
-
-    # episode_data_index
-    episode_data_index = {key: torch.tensor(episode_data_index[key]) for key in episode_data_index}
-    ep_data_idx_path = meta_data_dir / "episode_data_index.safetensors"
-    save_file(episode_data_index, ep_data_idx_path)
-    api.upload_file(
-        path_or_fileobj=ep_data_idx_path,
-        path_in_repo=str(ep_data_idx_path).replace(f"{root}/{dataset_id}", ""),
-        repo_id=f"lerobot/{dataset_id}",
-        repo_type="dataset",
-    )
-    api.upload_file(
-        path_or_fileobj=ep_data_idx_path,
-        path_in_repo=str(ep_data_idx_path).replace(f"{root}/{dataset_id}", ""),
-        repo_id=f"lerobot/{dataset_id}",
-        repo_type="dataset",
-        revision=revision,
-    )
-
-    # copy in tests folder, the first episode and the meta_data directory
-    num_items_first_ep = episode_data_index["to"][0] - episode_data_index["from"][0]
-    hf_dataset.select(range(num_items_first_ep)).with_format("torch").save_to_disk(
-        f"tests/data/lerobot/{dataset_id}/train"
-    )
-    if Path(f"tests/data/lerobot/{dataset_id}/meta_data").exists():
-        shutil.rmtree(f"tests/data/lerobot/{dataset_id}/meta_data")
-    shutil.copytree(meta_data_dir, f"tests/data/lerobot/{dataset_id}/meta_data")
-
-
-def download_and_upload_pusht(root, revision, dataset_id="pusht", fps=10):
-    try:
-        import pymunk
-        from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
-
-        from lerobot.common.datasets._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 env
-    success_threshold = 0.95  # 95% coverage,
-
-    pusht_url = "https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip"
-    pusht_zarr = Path("pusht/pusht_cchi_v7_replay.zarr")
-
-    root = Path(root)
-    raw_dir = root / f"{dataset_id}_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]
-    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 = PushTEnv.get_goal_pose_body(goal_pos_angle)
-
-    imgs = torch.from_numpy(dataset_dict["img"])  # b h w c
-    states = torch.from_numpy(dataset_dict["state"])
-    actions = torch.from_numpy(dataset_dict["action"])
-
-    ep_dicts = []
-    episode_data_index = {"from": [], "to": []}
-
-    id_from = 0
-    for episode_id in tqdm.tqdm(range(num_episodes)):
-        id_to = dataset_dict.meta["episode_ends"][episode_id]
-
-        num_frames = id_to - id_from
-
-        assert (episode_ids[id_from:id_to] == episode_id).all()
-
-        image = imgs[id_from:id_to]
-        assert image.min() >= 0.0
-        assert image.max() <= 255.0
-        image = image.type(torch.uint8)
-
-        state = states[id_from:id_to]
-        agent_pos = state[:, :2]
-        block_pos = state[:, 2:4]
-        block_angle = state[:, 4]
-
-        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 = {
-            "observation.image": [PILImage.fromarray(x.numpy()) for x in image],
-            "observation.state": agent_pos,
-            "action": actions[id_from:id_to],
-            "episode_index": torch.tensor([episode_id] * num_frames, dtype=torch.int),
-            "frame_index": torch.arange(0, num_frames, 1),
-            "timestamp": torch.arange(0, num_frames, 1) / fps,
-            # "next.observation.image": image[1:],
-            # "next.observation.state": agent_pos[1:],
-            # TODO(rcadene): verify that reward and done are aligned with image and agent_pos
-            "next.reward": torch.cat([reward[1:], reward[[-1]]]),
-            "next.done": torch.cat([done[1:], done[[-1]]]),
-            "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
-
-    data_dict = concatenate_episodes(ep_dicts)
-
-    features = {
-        "observation.image": Image(),
-        "observation.state": Sequence(
-            length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
-        ),
-        "action": Sequence(length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)),
-        "episode_index": Value(dtype="int64", id=None),
-        "frame_index": Value(dtype="int64", id=None),
-        "timestamp": Value(dtype="float32", id=None),
-        "next.reward": Value(dtype="float32", id=None),
-        "next.done": Value(dtype="bool", id=None),
-        "next.success": Value(dtype="bool", id=None),
-        "index": Value(dtype="int64", id=None),
-    }
-    features = Features(features)
-    hf_dataset = Dataset.from_dict(data_dict, features=features)
-    hf_dataset.set_transform(hf_transform_to_torch)
-
-    info = {
-        "fps": fps,
-    }
-    stats = compute_stats(hf_dataset)
-    push_to_hub(hf_dataset, episode_data_index, info, stats, root, revision, dataset_id)
-
-
-def download_and_upload_xarm(root, revision, dataset_id, fps=15):
-    root = Path(root)
-    raw_dir = root / "xarm_datasets_raw"
-    if not raw_dir.exists():
-        import zipfile
-
-        import gdown
-
-        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 member in zip_f.namelist():
-                if member.startswith("data/xarm") and member.endswith(".pkl"):
-                    print(member)
-                    zip_f.extract(member=member)
-        zip_path.unlink()
-
-    dataset_path = root / f"{dataset_id}" / "buffer.pkl"
-    print(f"Using offline dataset '{dataset_path}'")
-    with open(dataset_path, "rb") as f:
-        dataset_dict = pickle.load(f)
-
-    ep_dicts = []
-    episode_data_index = {"from": [], "to": []}
-
-    id_from = 0
-    id_to = 0
-    episode_id = 0
-    total_frames = dataset_dict["actions"].shape[0]
-    for i in tqdm.tqdm(range(total_frames)):
-        id_to += 1
-
-        if not dataset_dict["dones"][i]:
-            continue
-
-        num_frames = id_to - id_from
-
-        image = torch.tensor(dataset_dict["observations"]["rgb"][id_from:id_to])
-        image = einops.rearrange(image, "b c h w -> b h w c")
-        state = torch.tensor(dataset_dict["observations"]["state"][id_from:id_to])
-        action = torch.tensor(dataset_dict["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(dataset_dict["next_observations"]["rgb"][id_from:id_to])
-        # next_state = torch.tensor(dataset_dict["next_observations"]["state"][id_from:id_to])
-        next_reward = torch.tensor(dataset_dict["rewards"][id_from:id_to])
-        next_done = torch.tensor(dataset_dict["dones"][id_from:id_to])
-
-        ep_dict = {
-            "observation.image": [PILImage.fromarray(x.numpy()) for x in image],
-            "observation.state": state,
-            "action": action,
-            "episode_index": torch.tensor([episode_id] * num_frames, dtype=torch.int),
-            "frame_index": torch.arange(0, num_frames, 1),
-            "timestamp": torch.arange(0, num_frames, 1) / fps,
-            # "next.observation.image": next_image,
-            # "next.observation.state": next_state,
-            "next.reward": next_reward,
-            "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
-        episode_id += 1
-
-    data_dict = concatenate_episodes(ep_dicts)
-
-    features = {
-        "observation.image": Image(),
-        "observation.state": Sequence(
-            length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
-        ),
-        "action": Sequence(length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)),
-        "episode_index": Value(dtype="int64", id=None),
-        "frame_index": Value(dtype="int64", id=None),
-        "timestamp": Value(dtype="float32", id=None),
-        "next.reward": Value(dtype="float32", id=None),
-        "next.done": Value(dtype="bool", id=None),
-        #'next.success': Value(dtype='bool', id=None),
-        "index": Value(dtype="int64", id=None),
-    }
-    features = Features(features)
-    hf_dataset = Dataset.from_dict(data_dict, features=features)
-    hf_dataset.set_transform(hf_transform_to_torch)
-
-    info = {
-        "fps": fps,
-    }
-    stats = compute_stats(hf_dataset)
-    push_to_hub(hf_dataset, episode_data_index, info, stats, root, revision, dataset_id)
-
-
-def download_and_upload_aloha(root, revision, dataset_id, fps=50):
-    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"],
-    }
-
-    root = Path(root)
-    raw_dir = root / f"{dataset_id}_raw"
-    if not raw_dir.is_dir():
-        import gdown
-
-        assert dataset_id in folder_urls
-        assert dataset_id in ep48_urls
-        assert dataset_id in ep49_urls
-
-        raw_dir.mkdir(parents=True, exist_ok=True)
-
-        gdown.download_folder(folder_urls[dataset_id], output=str(raw_dir))
-
-        # because of the 50 files limit per directory, two files episode 48 and 49 were missing
-        gdown.download(ep48_urls[dataset_id], output=str(raw_dir / "episode_48.hdf5"), fuzzy=True)
-        gdown.download(ep49_urls[dataset_id], output=str(raw_dir / "episode_49.hdf5"), fuzzy=True)
-
-    ep_dicts = []
-    episode_data_index = {"from": [], "to": []}
-
-    id_from = 0
-    for ep_id in tqdm.tqdm(range(num_episodes[dataset_id])):
-        ep_path = raw_dir / f"episode_{ep_id}.hdf5"
-        with h5py.File(ep_path, "r") as ep:
-            num_frames = ep["/action"].shape[0]
-            assert episode_len[dataset_id] == num_frames
-
-            # 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"][:])
-
-            ep_dict = {}
-
-            for cam in cameras[dataset_id]:
-                image = torch.from_numpy(ep[f"/observations/images/{cam}"][:])  # b h w c
-                # image = einops.rearrange(image, "b h w c -> b c h w").contiguous()
-                ep_dict[f"observation.images.{cam}"] = [PILImage.fromarray(x.numpy()) for x in image]
-                # ep_dict[f"next.observation.images.{cam}"] = image
-
-            ep_dict.update(
-                {
-                    "observation.state": state,
-                    "action": action,
-                    "episode_index": torch.tensor([ep_id] * num_frames),
-                    "frame_index": torch.arange(0, num_frames, 1),
-                    "timestamp": torch.arange(0, num_frames, 1) / fps,
-                    # "next.observation.state": state,
-                    # TODO(rcadene): compute reward and success
-                    # "next.reward": reward,
-                    "next.done": done,
-                    # "next.success": success,
-                }
-            )
-
-            assert isinstance(ep_id, 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
-
-    data_dict = concatenate_episodes(ep_dicts)
-
-    features = {
-        "observation.images.top": Image(),
-        "observation.state": Sequence(
-            length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
-        ),
-        "action": Sequence(length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)),
-        "episode_index": Value(dtype="int64", id=None),
-        "frame_index": Value(dtype="int64", id=None),
-        "timestamp": Value(dtype="float32", id=None),
-        #'next.reward': Value(dtype='float32', id=None),
-        "next.done": Value(dtype="bool", id=None),
-        #'next.success': Value(dtype='bool', id=None),
-        "index": Value(dtype="int64", id=None),
-    }
-    features = Features(features)
-    hf_dataset = Dataset.from_dict(data_dict, features=features)
-    hf_dataset.set_transform(hf_transform_to_torch)
-
-    info = {
-        "fps": fps,
-    }
-    stats = compute_stats(hf_dataset)
-    push_to_hub(hf_dataset, episode_data_index, info, stats, root, revision, dataset_id)
-
-
-if __name__ == "__main__":
-    root = "data"
-    revision = "v1.1"
-
-    dataset_ids = [
-        "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",
-    ]
-    for dataset_id in dataset_ids:
-        download_and_upload(root, revision, dataset_id)

examples/1_load_hugging_face_dataset.py

@@ -1,69 +0,0 @@
-"""
-This script demonstrates the visualization of various robotic datasets from Hugging Face hub.
-It covers the steps from loading the datasets, filtering specific episodes, and converting the frame data to MP4 videos.
-Importantly, the dataset format is agnostic to any deep learning library and doesn't require using `lerobot` functions.
-It is compatible with pytorch, jax, numpy, etc.
-
-As an example, this script saves frames of episode number 5 of the PushT dataset to a mp4 video and saves the result here:
-`outputs/examples/1_visualize_hugging_face_datasets/episode_5.mp4`
-
-This script supports several Hugging Face datasets, among which:
-1. [Pusht](https://huggingface.co/datasets/lerobot/pusht)
-2. [Xarm Lift Medium](https://huggingface.co/datasets/lerobot/xarm_lift_medium)
-3. [Xarm Lift Medium Replay](https://huggingface.co/datasets/lerobot/xarm_lift_medium_replay)
-4. [Xarm Push Medium](https://huggingface.co/datasets/lerobot/xarm_push_medium)
-5. [Xarm Push Medium Replay](https://huggingface.co/datasets/lerobot/xarm_push_medium_replay)
-6. [Aloha Sim Insertion Human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human)
-7. [Aloha Sim Insertion Scripted](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_scripted)
-8. [Aloha Sim Transfer Cube Human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human)
-9. [Aloha Sim Transfer Cube Scripted](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_scripted)
-
-To try a different Hugging Face dataset, you can replace this line:
-```python
-hf_dataset, fps = load_dataset("lerobot/pusht", split="train"), 10
-```
-by one of these:
-```python
-hf_dataset, fps = load_dataset("lerobot/xarm_lift_medium", split="train"), 15
-hf_dataset, fps = load_dataset("lerobot/xarm_lift_medium_replay", split="train"), 15
-hf_dataset, fps = load_dataset("lerobot/xarm_push_medium", split="train"), 15
-hf_dataset, fps = load_dataset("lerobot/xarm_push_medium_replay", split="train"), 15
-hf_dataset, fps = load_dataset("lerobot/aloha_sim_insertion_human", split="train"), 50
-hf_dataset, fps = load_dataset("lerobot/aloha_sim_insertion_scripted", split="train"), 50
-hf_dataset, fps = load_dataset("lerobot/aloha_sim_transfer_cube_human", split="train"), 50
-hf_dataset, fps = load_dataset("lerobot/aloha_sim_transfer_cube_scripted", split="train"), 50
-```
-"""
-# TODO(rcadene): remove this example file of using hf_dataset
-
-from pathlib import Path
-
-import imageio
-from datasets import load_dataset
-
-# TODO(rcadene): list available datasets on lerobot page using `datasets`
-
-# download/load hugging face dataset in pyarrow format
-hf_dataset, fps = load_dataset("lerobot/pusht", split="train", revision="v1.1"), 10
-
-# display name of dataset and its features
-# TODO(rcadene): update to make the print pretty
-print(f"{hf_dataset=}")
-print(f"{hf_dataset.features=}")
-
-# display useful statistics about frames and episodes, which are sequences of frames from the same video
-print(f"number of frames: {len(hf_dataset)=}")
-print(f"number of episodes: {len(hf_dataset.unique('episode_index'))=}")
-print(
-    f"average number of frames per episode: {len(hf_dataset) / len(hf_dataset.unique('episode_index')):.3f}"
-)
-
-# select the frames belonging to episode number 5
-hf_dataset = hf_dataset.filter(lambda frame: frame["episode_index"] == 5)
-
-# load all frames of episode 5 in RAM in PIL format
-frames = hf_dataset["observation.image"]
-
-# save episode frames to a mp4 video
-Path("outputs/examples/1_load_hugging_face_dataset").mkdir(parents=True, exist_ok=True)
-imageio.mimsave("outputs/examples/1_load_hugging_face_dataset/episode_5.mp4", frames, fps=fps)

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 +1,121 @@
+"""
+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)
+
+# 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()
+
+# Check if GPU is available
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+    print("GPU is available. Device set to:", device)
+else:
+    device = torch.device("cpu")
+    print(f"GPU is not available. Device set to: {device}. Inference will be slower than on GPU.")
+    # Decrease the number of reverse-diffusion steps (trades off a bit of quality for 10x speed)
+    policy.diffusion.num_inference_steps = 10
+
+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/2_load_lerobot_dataset.py

@@ -1,91 +0,0 @@
-"""
-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
-
-import imageio
-import torch
-
-import lerobot
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-
-print("List of available datasets", lerobot.available_datasets)
-# # >>> ['lerobot/aloha_sim_insertion_human', 'lerobot/aloha_sim_insertion_scripted',
-# #     'lerobot/aloha_sim_transfer_cube_human', 'lerobot/aloha_sim_transfer_cube_scripted',
-# #     'lerobot/pusht', 'lerobot/xarm_lift_medium']
-
-repo_id = "lerobot/pusht"
-
-# You can easily load a dataset from a Hugging Face repositery
-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).
-# TODO(rcadene): update to make the print pretty
-print(f"{dataset=}")
-print(f"{dataset.hf_dataset=}")
-
-# and provides additional utilities for robotics and compatibility with pytorch
-print(f"number of samples/frames: {dataset.num_samples=}")
-print(f"number of episodes: {dataset.num_episodes=}")
-print(f"average 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.image_keys=}")
-
-# While the LeRobotDataset adds helpers for working within our library, we still expose the underling Hugging Face dataset.
-# It may be freely replaced or modified in place. Here we use the filtering to keep only frames from episode 5.
-# TODO(rcadene): remove this example of accessing hf_dataset
-dataset.hf_dataset = dataset.hf_dataset.filter(lambda frame: frame["episode_index"] == 5)
-
-# LeRobot datsets actually subclass PyTorch datasets. So you can do everything you know and love from working with the latter, for example: iterating through the dataset. Here we grab all the image frames.
-frames = [sample["observation.image"] for sample in dataset]
-
-# but frames are now float32 range [0,1] channel first (c,h,w) to follow pytorch convention,
-# to view them, we convert to uint8 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]
-
-# and finally save them to a mp4 video
-Path("outputs/examples/2_load_lerobot_dataset").mkdir(parents=True, exist_ok=True)
-imageio.mimsave("outputs/examples/2_load_lerobot_dataset/episode_5.mp4", frames, fps=dataset.fps)
-
-# For many machine learning applications we need to load histories of past observations, or trajectorys 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"{dataset[0]['observation.image'].shape=}")  # (4,c,h,w)
-print(f"{dataset[0]['observation.state'].shape=}")  # (8,c)
-print(f"{dataset[0]['action'].shape=}")  # (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/3_evaluate_pretrained_policy.py

@@ -1,38 +0,0 @@
-"""
-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
-
-from huggingface_hub import snapshot_download
-
-from lerobot.common.utils.utils import init_hydra_config
-from lerobot.scripts.eval import eval
-
-# Get a pretrained policy from the hub.
-# TODO(alexander-soare): This no longer works until we upload a new model that uses the current configs.
-hub_id = "lerobot/diffusion_policy_pusht_image"
-folder = Path(snapshot_download(hub_id))
-# OR uncomment the following to evaluate a policy from the local outputs/train folder.
-# folder = Path("outputs/train/example_pusht_diffusion")
-
-config_path = folder / "config.yaml"
-weights_path = folder / "model.pt"
-
-# Override some config parameters to do with evaluation.
-overrides = [
-    f"policy.pretrained_model_path={weights_path}",
-    "eval_episodes=10",
-    "rollout_batch_size=10",
-    "device=cuda",
-]
-
-# Create a Hydra config.
-cfg = init_hydra_config(config_path, overrides)
-
-# Evaluate the policy and save the outputs including metrics and videos.
-eval(
-    cfg,
-    out_dir=f"outputs/eval/example_{cfg.env.name}_{cfg.policy.name}",
-)

examples/3_train_policy.py

@@ -4,45 +4,53 @@ Once you have trained a model with this script, you can try to evaluate it on
 examples/2_evaluate_pretrained_policy.py
 """
 
-import os
 from pathlib import Path
 
 import torch
-from omegaconf import OmegaConf
 
-from lerobot.common.datasets.factory import make_dataset
+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
-from lerobot.common.utils.utils import init_hydra_config
 
+# Create a directory to store the training checkpoint.
 output_directory = Path("outputs/train/example_pusht_diffusion")
-os.makedirs(output_directory, exist_ok=True)
+output_directory.mkdir(parents=True, exist_ok=True)
 
-# Number of offline training steps (we'll only do offline training for this example.
+# 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.
-hydra_cfg = init_hydra_config("lerobot/configs/default.yaml", overrides=["env=pusht"])
-dataset = make_dataset(hydra_cfg)
+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()
-# TODO(alexander-soare): Remove LR scheduler from the policy.
-policy = DiffusionPolicy(cfg, lr_scheduler_num_training_steps=training_steps, dataset_stats=dataset.stats)
+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=cfg.batch_size,
+    batch_size=64,
     shuffle=True,
     pin_memory=device != torch.device("cpu"),
     drop_last=True,
@@ -54,14 +62,18 @@ done = False
 while not done:
     for batch in dataloader:
         batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
-        info = policy.update(batch)
+        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: {info['loss']:.3f} update_time: {info['update_s']:.3f} (seconds)")
+            print(f"step: {step} loss: {loss.item():.3f}")
         step += 1
         if step >= training_steps:
             done = True
             break
 
-# Save the policy and configuration for later use.
-policy.save(output_directory / "model.pt")
-OmegaConf.save(hydra_cfg, output_directory / "config.yaml")
+# Save a policy checkpoint.
+policy.save_pretrained(output_directory)

examples/4_train_policy_with_script.md

@@ -0,0 +1,213 @@
+This tutorial will explain the training script, how to use it, and particularly the use of Hydra to configure everything needed for the training run.
+
+## The training script
+
+LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
+
+- Loads a Hydra configuration file for the following steps (more on Hydra in a moment).
+- Makes a simulation environment.
+- Makes a dataset corresponding to that simulation environment.
+- Makes a policy.
+- Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
+
+## Basics of how we use Hydra
+
+Explaining the ins and outs of [Hydra](https://hydra.cc/docs/intro/) is beyond the scope of this document, but here we'll share the main points you need to know.
+
+First, `lerobot/configs` has a directory structure like this:
+
+```
+.
+├── default.yaml
+├── env
+│   ├── aloha.yaml
+│   ├── pusht.yaml
+│   └── xarm.yaml
+└── policy
+    ├── act.yaml
+    ├── diffusion.yaml
+    └── tdmpc.yaml
+```
+
+**_For brevity, in the rest of this document we'll drop the leading `lerobot/configs` path. So `default.yaml` really refers to `lerobot/configs/default.yaml`._**
+
+When you run the training script with
+
+```python
+python lerobot/scripts/train.py
+```
+
+Hydra is set up to read `default.yaml` (via the `@hydra.main` decorator). If you take a look at the `@hydra.main`'s arguments you will see `config_path="../configs", config_name="default"`. At the top of `default.yaml`, is a `defaults` section which looks likes this:
+
+```yaml
+defaults:
+  - _self_
+  - env: pusht
+  - policy: diffusion
+```
+
+This logic tells Hydra to incorporate configuration parameters from `env/pusht.yaml` and `policy/diffusion.yaml`. _Note: Be aware of the order as any configuration parameters with the same name will be overidden. Thus, `default.yaml` is overridden by `env/pusht.yaml`  which is overidden by `policy/diffusion.yaml`_.
+
+Then, `default.yaml` also contains common configuration parameters such as `device: cuda` or `use_amp: false` (for enabling fp16 training). Some other parameters are set to `???` which indicates that they are expected to be set in additional yaml files. For instance, `training.offline_steps: ???` in `default.yaml` is set to `200000` in `diffusion.yaml`.
+
+Thanks to this `defaults` section in `default.yaml`, if you want to train Diffusion Policy with PushT, you really only need to run:
+
+```bash
+python lerobot/scripts/train.py
+```
+
+However, you can be more explicit and launch the exact same Diffusion Policy training on PushT with:
+
+```bash
+python lerobot/scripts/train.py policy=diffusion env=pusht
+```
+
+This way of overriding defaults via the CLI is especially useful when you want to change the policy and/or environment. For instance, you can train ACT on the default Aloha environment with:
+
+```bash
+python lerobot/scripts/train.py policy=act env=aloha
+```
+
+There are two things to note here:
+- Config overrides are passed as `param_name=param_value`.
+- Here we have overridden the defaults section. `policy=act` tells Hydra to use `policy/act.yaml`, and `env=aloha` tells Hydra to use `env/aloha.yaml`.
+
+_As an aside: we've set up all of our configurations so that they reproduce state-of-the-art results from papers in the literature._
+
+## Overriding configuration parameters in the CLI
+
+Now let's say that we want to train on a different task in the Aloha environment. If you look in `env/aloha.yaml` you will see something like:
+
+```yaml
+# lerobot/configs/env/aloha.yaml
+env:
+  task: AlohaInsertion-v0
+```
+
+And if you look in `policy/act.yaml` you will see something like:
+
+```yaml
+# lerobot/configs/policy/act.yaml
+dataset_repo_id: lerobot/aloha_sim_insertion_human
+```
+
+But our Aloha environment actually supports a cube transfer task as well. To train for this task, you could manually modify the two yaml configuration files respectively.
+
+First, we'd need to switch to using the cube transfer task for the ALOHA environment.
+
+```diff
+# lerobot/configs/env/aloha.yaml
+env:
+-  task: AlohaInsertion-v0
++  task: AlohaTransferCube-v0
+```
+
+Then, we'd also need to switch to using the cube transfer dataset.
+
+```diff
+# lerobot/configs/policy/act.yaml
+-dataset_repo_id: lerobot/aloha_sim_insertion_human
++dataset_repo_id: lerobot/aloha_sim_transfer_cube_human
+```
+
+Then, you'd be able to run:
+
+```bash
+python lerobot/scripts/train.py policy=act env=aloha
+```
+
+and you'd be training and evaluating on the cube transfer task.
+
+An alternative approach to editing the yaml configuration files, would be to override the defaults via the command line:
+
+```bash
+python lerobot/scripts/train.py \
+    policy=act \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    env=aloha \
+    env.task=AlohaTransferCube-v0
+```
+
+There's something new here. Notice the `.` delimiter used to traverse the configuration hierarchy. _But be aware that the `defaults` section is an exception. As you saw above, we didn't need to write `defaults.policy=act` in the CLI. `policy=act` was enough._
+
+Putting all that knowledge together, here's the command that was used to train https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human.
+
+```bash
+python lerobot/scripts/train.py \
+    hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human \
+    device=cuda
+    env=aloha \
+    env.task=AlohaTransferCube-v0 \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    policy=act \
+    training.eval_freq=10000 \
+    training.log_freq=250 \
+    training.offline_steps=100000 \
+    training.save_model=true \
+    training.save_freq=25000 \
+    eval.n_episodes=50 \
+    eval.batch_size=50 \
+    wandb.enable=false \
+```
+
+There's one new thing here: `hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human`, which specifies where to save the training output.
+
+## Using a configuration file not in `lerobot/configs`
+
+Above we discusses the our training script is set up such that Hydra looks for `default.yaml` in `lerobot/configs`. But, if you have a configuration file elsewhere in your filesystem you may use:
+
+```bash
+python lerobot/scripts/train.py --config-dir PARENT/PATH --config-name FILE_NAME_WITHOUT_EXTENSION
+```
+
+Note: here we use regular syntax for providing CLI arguments to a Python script, not Hydra's `param_name=param_value` syntax.
+
+As a concrete example, this becomes particularly handy when you have a folder with training outputs, and would like to re-run the training. For example, say you previously ran the training script with one of the earlier commands and have `outputs/train/my_experiment/checkpoints/pretrained_model/config.yaml`. This `config.yaml` file will have the full set of configuration parameters within it. To run the training with the same configuration again, do:
+
+```bash
+python lerobot/scripts/train.py --config-dir outputs/train/my_experiment/checkpoints/last/pretrained_model --config-name config
+```
+
+Note that you may still use the regular syntax for config parameter overrides (eg: by adding `training.offline_steps=200000`).
+
+## Typical logs and metrics
+
+When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
+
+After that, you will see training log like this one:
+
+```
+INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
+```
+
+or evaluation log like:
+
+```
+INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
+```
+
+These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
+
+- `smpl`: number of samples seen during training.
+- `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
+- `epch`: number of time all unique samples are seen (epoch).
+- `grdn`: gradient norm.
+- `∑rwrd`: compute the sum of rewards in every evaluation episode and then take an average of them.
+- `success`: average success rate of eval episodes. Reward and success are usually different except for the sparsing reward setting, where reward=1 only when the task is completed successfully.
+- `eval_s`: time to evaluate the policy in the environment, in second.
+- `updt_s`: time to update the network parameters, in second.
+- `data_s`: time to load a batch of data, in second. 
+
+Some metrics are useful for initial performance profiling. For example, if you find the current GPU utilization is low via the `nvidia-smi` command and `data_s` sometimes is too high, you may need to modify batch size or number of dataloading workers to accelerate dataloading. We also recommend [pytorch profiler](https://github.com/huggingface/lerobot?tab=readme-ov-file#improve-your-code-with-profiling) for detailed performance probing.
+
+---
+
+So far we've seen how to train Diffusion Policy for PushT and ACT for ALOHA. Now, what if we want to train ACT for PushT? Well, there are aspects of the ACT configuration that are specific to the ALOHA environments, and these happen to be incompatible with PushT. Therefore, trying to run the following will almost certainly raise an exception of sorts (eg: feature dimension mismatch):
+
+```bash
+python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
+```
+
+Please, head on over to our [advanced tutorial on adapting policy configuration to various environments](./advanced/train_act_pusht/train_act_pusht.md) to learn more.
+
+Or in the meantime, happy coding! 🤗

examples/5_resume_training.md

@@ -0,0 +1,37 @@
+This tutorial explains how to resume a training run that you've started with the training script. If you don't know how our training script and configuration system works, please read [4_train_policy_with_script.md](./4_train_policy_with_script.md) first.
+
+## Basic training resumption
+
+Let's consider the example of training ACT for one of the ALOHA tasks. Here's a command that can achieve that:
+
+```bash
+python lerobot/scripts/train.py \
+    hydra.run.dir=outputs/train/run_resumption \
+    policy=act \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    env=aloha \
+    env.task=AlohaTransferCube-v0 \
+    training.log_freq=25 \
+    training.save_checkpoint=true \
+    training.save_freq=100
+```
+
+Here we're using the default dataset and environment for ACT, and we've taken care to set up the log frequency and checkpointing frequency to low numbers so we can test resumption. You should be able to see some logging and have a first checkpoint within 1 minute. Please interrupt the training after the first checkpoint.
+
+To resume, all that we have to do is run the training script, providing the run directory, and the resume option:
+
+```bash
+python lerobot/scripts/train.py \
+    hydra.run.dir=outputs/train/run_resumption \
+    resume=true
+```
+
+You should see from the logging that your training picks up from where it left off.
+
+Note that with `resume=true`, the configuration file from the last checkpoint in the training output directory is loaded. So it doesn't matter that we haven't provided all the other configuration parameters from our previous command (although there may be warnings to notify you that your command has a different configuration than than the checkpoint).
+
+---
+
+Now you should know how to resume your training run in case it gets interrupted or you want to extend a finished training run.
+
+Happy coding! 🤗

examples/6_add_image_transforms.py

@@ -0,0 +1,52 @@
+"""
+This script demonstrates how to use torchvision's image transformation with LeRobotDataset for data
+augmentation purposes. The transformations are passed to the dataset as an argument upon creation, and
+transforms are applied to the observation images before they are returned in the dataset's __get_item__.
+"""
+
+from pathlib import Path
+
+from torchvision.transforms import ToPILImage, v2
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+dataset_repo_id = "lerobot/aloha_static_tape"
+
+# Create a LeRobotDataset with no transformations
+dataset = LeRobotDataset(dataset_repo_id)
+# This is equivalent to `dataset = LeRobotDataset(dataset_repo_id, image_transforms=None)`
+
+# Get the index of the first observation in the first episode
+first_idx = dataset.episode_data_index["from"][0].item()
+
+# Get the frame corresponding to the first camera
+frame = dataset[first_idx][dataset.camera_keys[0]]
+
+
+# Define the transformations
+transforms = v2.Compose(
+    [
+        v2.ColorJitter(brightness=(0.5, 1.5)),
+        v2.ColorJitter(contrast=(0.5, 1.5)),
+        v2.RandomAdjustSharpness(sharpness_factor=2, p=1),
+    ]
+)
+
+# Create another LeRobotDataset with the defined transformations
+transformed_dataset = LeRobotDataset(dataset_repo_id, image_transforms=transforms)
+
+# Get a frame from the transformed dataset
+transformed_frame = transformed_dataset[first_idx][transformed_dataset.camera_keys[0]]
+
+# Create a directory to store output images
+output_dir = Path("outputs/image_transforms")
+output_dir.mkdir(parents=True, exist_ok=True)
+
+# Save the original frame
+to_pil = ToPILImage()
+to_pil(frame).save(output_dir / "original_frame.png", quality=100)
+print(f"Original frame saved to {output_dir / 'original_frame.png'}.")
+
+# Save the transformed frame
+to_pil(transformed_frame).save(output_dir / "transformed_frame.png", quality=100)
+print(f"Transformed frame saved to {output_dir / 'transformed_frame.png'}.")

examples/advanced/1_train_act_pusht/act_pusht.yaml

@@ -0,0 +1,87 @@
+# @package _global_
+
+# Change the seed to match what PushT eval uses
+# (to avoid evaluating on seeds used for generating the training data).
+seed: 100000
+# Change the dataset repository to the PushT one.
+dataset_repo_id: lerobot/pusht
+
+override_dataset_stats:
+  observation.image:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+training:
+  offline_steps: 80000
+  online_steps: 0
+  eval_freq: 10000
+  save_freq: 100000
+  log_freq: 250
+  save_model: true
+
+  batch_size: 8
+  lr: 1e-5
+  lr_backbone: 1e-5
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
+
+  delta_timestamps:
+    action: "[i / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    observation.image: [3, 96, 96]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.image: mean_std
+    # Use min_max normalization just because it's more standard.
+    observation.state: min_max
+  output_normalization_modes:
+    # Use min_max normalization just because it's more standard.
+    action: min_max
+
+  # Architecture.
+  # Vision backbone.
+  vision_backbone: resnet18
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
+  replace_final_stride_with_dilation: false
+  # Transformer layers.
+  pre_norm: false
+  dim_model: 512
+  n_heads: 8
+  dim_feedforward: 3200
+  feedforward_activation: relu
+  n_encoder_layers: 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: 1
+  # VAE.
+  use_vae: true
+  latent_dim: 32
+  n_vae_encoder_layers: 4
+
+  # Inference.
+  temporal_ensemble_coeff: null
+
+  # Training and loss computation.
+  dropout: 0.1
+  kl_weight: 10.0

examples/advanced/1_train_act_pusht/train_act_pusht.md

@@ -0,0 +1,70 @@
+In this tutorial we will learn how to adapt a policy configuration to be compatible with a new environment and dataset. As a concrete example, we will adapt the default configuration for ACT to be compatible with the PushT environment and dataset.
+
+If you haven't already read our tutorial on the [training script and configuration tooling](../4_train_policy_with_script.md) please do so prior to tackling this tutorial.
+
+Let's get started!
+
+Suppose we want to train ACT for PushT. Well, there are aspects of the ACT configuration that are specific to the ALOHA environments, and these happen to be incompatible with PushT. Therefore, trying to run the following will almost certainly raise an exception of sorts (eg: feature dimension mismatch):
+
+```bash
+python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
+```
+
+We need to adapt the parameters of the ACT policy configuration to the PushT environment. The most important ones are the image keys.
+
+ALOHA's datasets and environments typically use a variable number of cameras. In `lerobot/configs/policy/act.yaml` you may notice two relevant sections. Here we show you the minimal diff needed to adjust to PushT:
+
+```diff
+override_dataset_stats:
+-  observation.images.top:
++  observation.image:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+policy:
+  input_shapes:
+-    observation.images.top: [3, 480, 640]
++    observation.image: [3, 96, 96]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  input_normalization_modes:
+-    observation.images.top: mean_std
++    observation.image: mean_std
+     observation.state: min_max
+  output_normalization_modes:
+    action: min_max
+```
+
+Here we've accounted for the following:
+- PushT uses "observation.image" for its image key.
+- PushT provides smaller images.
+
+_Side note: technically we could override these via the CLI, but with many changes it gets a bit messy, and we also have a bit of a challenge in that we're using `.` in our observation keys which is treated by Hydra as a hierarchical separator_.
+
+For your convenience, we provide [`act_pusht.yaml`](./act_pusht.yaml) in this directory. It contains the diff above, plus some other (optional) ones that are explained within. Please copy it into `lerobot/configs/policy` with:
+
+```bash
+cp examples/advanced/1_train_act_pusht/act_pusht.yaml lerobot/configs/policy/act_pusht.yaml
+```
+
+(remember from a [previous tutorial](../4_train_policy_with_script.md) that Hydra will look in the `lerobot/configs` directory). Now try running the following.
+
+<!-- Note to contributor: are you changing this command? Note that it's tested in `Makefile`, so change it there too! -->
+```bash
+python lerobot/scripts/train.py policy=act_pusht env=pusht
+```
+
+Notice that this is much the same as the command that failed at the start of the tutorial, only:
+- Now we are using `policy=act_pusht` to point to our new configuration file.
+- We can drop `dataset_repo_id=lerobot/pusht` as the change is incorporated in our new configuration file.
+
+Hurrah! You're now training ACT for the PushT environment.
+
+---
+
+The bottom line of this tutorial is that when training policies for different environments and datasets you will need to understand what parts of the policy configuration are specific to those and make changes accordingly.
+
+Happy coding! 🤗

examples/advanced/2_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

@@ -1,3 +1,18 @@
+#!/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.
@@ -9,6 +24,7 @@ Example:
         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)
     ```
@@ -25,8 +41,13 @@ When implementing a new policy class (e.g. `DiffusionPolicy`) follow these steps
 - Update variables in `tests/test_available.py` by importing your new Policy class
 """
 
+import itertools
+
 from lerobot.__version__ import __version__  # noqa: F401
 
+# TODO(rcadene): Improve policies and envs. As of now, an item in `available_policies`
+# refers to a yaml file AND a modeling name. Same for `available_envs` which refers to
+# a yaml file AND a environment name. The difference should be more obvious.
 available_tasks_per_env = {
     "aloha": [
         "AlohaInsertion-v0",
@@ -34,6 +55,7 @@ available_tasks_per_env = {
     ],
     "pusht": ["PushT-v0"],
     "xarm": ["XarmLift-v0"],
+    "dora_aloha_real": ["DoraAloha-v0", "DoraKoch-v0", "DoraReachy2-v0"],
 }
 available_envs = list(available_tasks_per_env.keys())
 
@@ -43,27 +65,90 @@ available_datasets_per_env = {
         "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"],
+    # TODO(alexander-soare): Add "lerobot/pusht_keypoints". Right now we can't because this is too tightly
+    # coupled with tests.
+    "pusht": ["lerobot/pusht", "lerobot/pusht_image"],
     "xarm": [
         "lerobot/xarm_lift_medium",
         "lerobot/xarm_lift_medium_replay",
         "lerobot/xarm_push_medium",
         "lerobot/xarm_push_medium_replay",
+        "lerobot/xarm_lift_medium_image",
+        "lerobot/xarm_lift_medium_replay_image",
+        "lerobot/xarm_push_medium_image",
+        "lerobot/xarm_push_medium_replay_image",
+    ],
+    "dora_aloha_real": [
+        "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",
     ],
 }
-available_datasets = [dataset for datasets in available_datasets_per_env.values() for dataset in datasets]
 
+available_real_world_datasets = [
+    "lerobot/aloha_mobile_cabinet",
+    "lerobot/aloha_mobile_chair",
+    "lerobot/aloha_mobile_elevator",
+    "lerobot/aloha_mobile_shrimp",
+    "lerobot/aloha_mobile_wash_pan",
+    "lerobot/aloha_mobile_wipe_wine",
+    "lerobot/aloha_static_battery",
+    "lerobot/aloha_static_candy",
+    "lerobot/aloha_static_coffee",
+    "lerobot/aloha_static_coffee_new",
+    "lerobot/aloha_static_cups_open",
+    "lerobot/aloha_static_fork_pick_up",
+    "lerobot/aloha_static_pingpong_test",
+    "lerobot/aloha_static_pro_pencil",
+    "lerobot/aloha_static_screw_driver",
+    "lerobot/aloha_static_tape",
+    "lerobot/aloha_static_thread_velcro",
+    "lerobot/aloha_static_towel",
+    "lerobot/aloha_static_vinh_cup",
+    "lerobot/aloha_static_vinh_cup_left",
+    "lerobot/aloha_static_ziploc_slide",
+    "lerobot/umi_cup_in_the_wild",
+    "lerobot/unitreeh1_fold_clothes",
+    "lerobot/unitreeh1_rearrange_objects",
+    "lerobot/unitreeh1_two_robot_greeting",
+    "lerobot/unitreeh1_warehouse",
+]
+
+available_datasets = list(
+    itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets)
+)
+
+# lists all available policies from `lerobot/common/policies` by their class attribute: `name`.
 available_policies = [
     "act",
     "diffusion",
     "tdmpc",
+    "vqbet",
 ]
 
+# keys and values refer to yaml files
 available_policies_per_env = {
     "aloha": ["act"],
-    "pusht": ["diffusion"],
+    "pusht": ["diffusion", "vqbet"],
     "xarm": ["tdmpc"],
+    "dora_aloha_real": ["act_real"],
 }
 
 env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]

lerobot/__version__.py

@@ -1,3 +1,18 @@
+#!/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

lerobot/common/datasets/compute_stats.py

@@ -0,0 +1,209 @@
+#!/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 einops
+import torch
+import tqdm
+from datasets import Image
+
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def get_stats_einops_patterns(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, batch_size=32, num_workers=16, max_num_samples=None):
+    """Compute mean/std and min/max statistics of all data keys in a LeRobotDataset."""
+    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
+
+
+def aggregate_stats(ls_datasets) -> dict[str, torch.Tensor]:
+    """Aggregate stats of multiple LeRobot datasets into one set of stats without recomputing from scratch.
+
+    The final stats will have the union of all data keys from each of the datasets.
+
+    The final stats will have the union of all data keys from each of the datasets. For instance:
+    - new_max = max(max_dataset_0, max_dataset_1, ...)
+    - new_min = min(min_dataset_0, min_dataset_1, ...)
+    - new_mean = (mean of all data)
+    - new_std = (std of all data)
+    """
+    data_keys = set()
+    for dataset in ls_datasets:
+        data_keys.update(dataset.stats.keys())
+    stats = {k: {} for k in data_keys}
+    for data_key in data_keys:
+        for stat_key in ["min", "max"]:
+            # compute `max(dataset_0["max"], dataset_1["max"], ...)`
+            stats[data_key][stat_key] = einops.reduce(
+                torch.stack([d.stats[data_key][stat_key] for d in ls_datasets if data_key in d.stats], dim=0),
+                "n ... -> ...",
+                stat_key,
+            )
+        total_samples = sum(d.num_samples for d in ls_datasets if data_key in d.stats)
+        # Compute the "sum" statistic by multiplying each mean by the number of samples in the respective
+        # dataset, then divide by total_samples to get the overall "mean".
+        # NOTE: the brackets around (d.num_samples / total_samples) are needed tor minimize the risk of
+        # numerical overflow!
+        stats[data_key]["mean"] = sum(
+            d.stats[data_key]["mean"] * (d.num_samples / total_samples)
+            for d in ls_datasets
+            if data_key in d.stats
+        )
+        # The derivation for standard deviation is a little more involved but is much in the same spirit as
+        # the computation of the mean.
+        # Given two sets of data where the statistics are known:
+        # σ_combined = sqrt[ (n1 * (σ1^2 + d1^2) + n2 * (σ2^2 + d2^2)) / (n1 + n2) ]
+        # where d1 = μ1 - μ_combined, d2 = μ2 - μ_combined
+        # NOTE: the brackets around (d.num_samples / total_samples) are needed tor minimize the risk of
+        # numerical overflow!
+        stats[data_key]["std"] = torch.sqrt(
+            sum(
+                (d.stats[data_key]["std"] ** 2 + (d.stats[data_key]["mean"] - stats[data_key]["mean"]) ** 2)
+                * (d.num_samples / total_samples)
+                for d in ls_datasets
+                if data_key in d.stats
+            )
+        )
+    return stats

lerobot/common/datasets/factory.py

@@ -1,38 +1,111 @@
+#!/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 omegaconf import OmegaConf
+from omegaconf import ListConfig, OmegaConf
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-
-DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
+from lerobot.common.datasets.transforms import get_image_transforms
 
 
-def make_dataset(
-    cfg,
-    split="train",
-):
-    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 environment ({cfg.env.name=})."
-        )
+def resolve_delta_timestamps(cfg):
+    """Resolves delta_timestamps config key (in-place) by using `eval`.
 
-    delta_timestamps = cfg.policy.get("delta_timestamps")
+    Doesn't do anything if delta_timestamps is not specified or has already been resolve (as evidenced by
+    the data type of its values).
+    """
+    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])
+                # TODO(rcadene, alexander-soare): remove `eval` to avoid exploit
+                cfg.training.delta_timestamps[key] = eval(delta_timestamps[key])
 
-    # TODO(rcadene): add data augmentations
 
-    dataset = LeRobotDataset(
-        cfg.dataset.repo_id,
-        split=split,
-        root=DATA_DIR,
-        delta_timestamps=delta_timestamps,
-    )
+def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotDataset:
+    """
+    Args:
+        cfg: A Hydra config as per the LeRobot config scheme.
+        split: Select the data subset used to create an instance of LeRobotDataset.
+            All datasets hosted on [lerobot](https://huggingface.co/lerobot) contain only one subset: "train".
+            Thus, by default, `split="train"` selects all the available data. `split` aims to work like the
+            slicer in the hugging face datasets:
+            https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
+            As of now, it only supports `split="train[:n]"` to load the first n frames of the dataset or
+            `split="train[n:]"` to load the last n frames. For instance `split="train[:1000]"`.
+    Returns:
+        The LeRobotDataset.
+    """
+    if not isinstance(cfg.dataset_repo_id, (str, ListConfig)):
+        raise ValueError(
+            "Expected cfg.dataset_repo_id to be either a single string to load one dataset or a list of "
+            "strings to load multiple datasets."
+        )
+
+    # A soft check to warn if the environment matches the dataset. Don't check if we are using a real world env (dora).
+    if cfg.env.name != "dora":
+        if isinstance(cfg.dataset_repo_id, str):
+            dataset_repo_ids = [cfg.dataset_repo_id]  # single dataset
+        else:
+            dataset_repo_ids = cfg.dataset_repo_id  # multiple datasets
+
+        for dataset_repo_id in dataset_repo_ids:
+            if cfg.env.name not in dataset_repo_id:
+                logging.warning(
+                    f"There might be a mismatch between your training dataset ({dataset_repo_id=}) and your "
+                    f"environment ({cfg.env.name=})."
+                )
+
+    resolve_delta_timestamps(cfg)
+
+    image_transforms = None
+    if cfg.training.image_transforms.enable:
+        cfg_tf = cfg.training.image_transforms
+        image_transforms = get_image_transforms(
+            brightness_weight=cfg_tf.brightness.weight,
+            brightness_min_max=cfg_tf.brightness.min_max,
+            contrast_weight=cfg_tf.contrast.weight,
+            contrast_min_max=cfg_tf.contrast.min_max,
+            saturation_weight=cfg_tf.saturation.weight,
+            saturation_min_max=cfg_tf.saturation.min_max,
+            hue_weight=cfg_tf.hue.weight,
+            hue_min_max=cfg_tf.hue.min_max,
+            sharpness_weight=cfg_tf.sharpness.weight,
+            sharpness_min_max=cfg_tf.sharpness.min_max,
+            max_num_transforms=cfg_tf.max_num_transforms,
+            random_order=cfg_tf.random_order,
+        )
+
+    if isinstance(cfg.dataset_repo_id, str):
+        dataset = LeRobotDataset(
+            cfg.dataset_repo_id,
+            split=split,
+            delta_timestamps=cfg.training.get("delta_timestamps"),
+            image_transforms=image_transforms,
+            video_backend=cfg.video_backend,
+        )
+    else:
+        dataset = MultiLeRobotDataset(
+            cfg.dataset_repo_id,
+            split=split,
+            delta_timestamps=cfg.training.get("delta_timestamps"),
+            image_transforms=image_transforms,
+            video_backend=cfg.video_backend,
+        )
 
     if cfg.get("override_dataset_stats"):
         for key, stats_dict in cfg.override_dataset_stats.items():

lerobot/common/datasets/lerobot_dataset.py

@@ -1,56 +1,134 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+import os
 from pathlib import Path
+from typing import Callable
 
 import datasets
 import torch
+import torch.utils
 
+from lerobot.common.datasets.compute_stats import aggregate_stats
 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
+
+# For maintainers, see lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md
+CODEBASE_VERSION = "v1.6"
+DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
 
 
 class LeRobotDataset(torch.utils.data.Dataset):
     def __init__(
         self,
         repo_id: str,
-        version: str | None = "v1.1",
-        root: Path | None = None,
+        root: Path | None = DATA_DIR,
         split: str = "train",
-        transform: callable = None,
+        image_transforms: Callable | None = None,
         delta_timestamps: dict[list[float]] | None = None,
+        video_backend: str | None = None,
     ):
         super().__init__()
         self.repo_id = repo_id
-        self.version = version
         self.root = root
         self.split = split
-        self.transform = transform
+        self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps
         # load data from hub or locally when root is provided
-        self.hf_dataset = load_hf_dataset(repo_id, version, root, split)
-        self.episode_data_index = load_episode_data_index(repo_id, version, root)
-        self.stats = load_stats(repo_id, version, root)
-        self.info = load_info(repo_id, version, root)
+        # 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, CODEBASE_VERSION, root, split)
+        if split == "train":
+            self.episode_data_index = load_episode_data_index(repo_id, CODEBASE_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, CODEBASE_VERSION, root)
+        self.info = load_info(repo_id, CODEBASE_VERSION, root)
+        if self.video:
+            self.videos_dir = load_videos(repo_id, CODEBASE_VERSION, root)
+            self.video_backend = video_backend if video_backend is not None else "pyav"
 
     @property
     def fps(self) -> int:
+        """Frames per second used during data collection."""
         return self.info["fps"]
 
     @property
-    def image_keys(self) -> list[str]:
-        return [key for key, feats in self.hf_dataset.features.items() if isinstance(feats, datasets.Image)]
+    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
 
@@ -63,10 +141,261 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 self.hf_dataset,
                 self.episode_data_index,
                 self.delta_timestamps,
-                tol=1 / self.fps - 1e-4,  # 1e-4 to account for possible numerical error
+                self.tolerance_s,
             )
 
-        if self.transform is not None:
-            item = self.transform(item)
+        if self.video:
+            item = load_from_videos(
+                item,
+                self.video_frame_keys,
+                self.videos_dir,
+                self.tolerance_s,
+                self.video_backend,
+            )
+
+        if self.image_transforms is not None:
+            for cam in self.camera_keys:
+                item[cam] = self.image_transforms(item[cam])
 
         return item
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}(\n"
+            f"  Repository ID: '{self.repo_id}',\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.image_transforms},\n"
+            f"  Codebase Version: {self.info.get('codebase_version', '< v1.6')},\n"
+            f")"
+        )
+
+    @classmethod
+    def from_preloaded(
+        cls,
+        repo_id: str = "from_preloaded",
+        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,
+        video_backend=None,
+    ) -> "LeRobotDataset":
+        """Create a LeRobot Dataset from existing data and attributes instead of loading from the filesystem.
+
+        It is especially useful when converting raw data into LeRobotDataset before saving the dataset
+        on the filesystem or uploading to the hub.
+
+        Note: Meta-data attributes like `repo_id`, `version`, `root`, etc are optional and potentially
+        meaningless depending on the downstream usage of the return dataset.
+        """
+        # create an empty object of type LeRobotDataset
+        obj = cls.__new__(cls)
+        obj.repo_id = repo_id
+        obj.root = root
+        obj.split = split
+        obj.image_transforms = transform
+        obj.delta_timestamps = delta_timestamps
+        obj.hf_dataset = hf_dataset
+        obj.episode_data_index = episode_data_index
+        obj.stats = stats
+        obj.info = info if info is not None else {}
+        obj.videos_dir = videos_dir
+        obj.video_backend = video_backend if video_backend is not None else "pyav"
+        return obj
+
+
+class MultiLeRobotDataset(torch.utils.data.Dataset):
+    """A dataset consisting of multiple underlying `LeRobotDataset`s.
+
+    The underlying `LeRobotDataset`s are effectively concatenated, and this class adopts much of the API
+    structure of `LeRobotDataset`.
+    """
+
+    def __init__(
+        self,
+        repo_ids: list[str],
+        root: Path | None = DATA_DIR,
+        split: str = "train",
+        image_transforms: Callable | None = None,
+        delta_timestamps: dict[list[float]] | None = None,
+        video_backend: str | None = None,
+    ):
+        super().__init__()
+        self.repo_ids = repo_ids
+        # Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
+        # are handled by this class.
+        self._datasets = [
+            LeRobotDataset(
+                repo_id,
+                root=root,
+                split=split,
+                delta_timestamps=delta_timestamps,
+                image_transforms=image_transforms,
+                video_backend=video_backend,
+            )
+            for repo_id in repo_ids
+        ]
+        # Check that some properties are consistent across datasets. Note: We may relax some of these
+        # consistency requirements in future iterations of this class.
+        for repo_id, dataset in zip(self.repo_ids, self._datasets, strict=True):
+            if dataset.info != self._datasets[0].info:
+                raise ValueError(
+                    f"Detected a mismatch in dataset info between {self.repo_ids[0]} and {repo_id}. This is "
+                    "not yet supported."
+                )
+        # Disable any data keys that are not common across all of the datasets. Note: we may relax this
+        # restriction in future iterations of this class. For now, this is necessary at least for being able
+        # to use PyTorch's default DataLoader collate function.
+        self.disabled_data_keys = set()
+        intersection_data_keys = set(self._datasets[0].hf_dataset.features)
+        for dataset in self._datasets:
+            intersection_data_keys.intersection_update(dataset.hf_dataset.features)
+        if len(intersection_data_keys) == 0:
+            raise RuntimeError(
+                "Multiple datasets were provided but they had no keys common to all of them. The "
+                "multi-dataset functionality currently only keeps common keys."
+            )
+        for repo_id, dataset in zip(self.repo_ids, self._datasets, strict=True):
+            extra_keys = set(dataset.hf_dataset.features).difference(intersection_data_keys)
+            logging.warning(
+                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                "other datasets."
+            )
+            self.disabled_data_keys.update(extra_keys)
+
+        self.root = root
+        self.split = split
+        self.image_transforms = image_transforms
+        self.delta_timestamps = delta_timestamps
+        self.stats = aggregate_stats(self._datasets)
+
+    @property
+    def repo_id_to_index(self):
+        """Return a mapping from dataset repo_id to a dataset index automatically created by this class.
+
+        This index is incorporated as a data key in the dictionary returned by `__getitem__`.
+        """
+        return {repo_id: i for i, repo_id in enumerate(self.repo_ids)}
+
+    @property
+    def repo_index_to_id(self):
+        """Return the inverse mapping if repo_id_to_index."""
+        return {v: k for k, v in self.repo_id_to_index}
+
+    @property
+    def fps(self) -> int:
+        """Frames per second used during data collection.
+
+        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
+        """
+        return self._datasets[0].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.
+
+        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
+        """
+        return self._datasets[0].info.get("video", False)
+
+    @property
+    def features(self) -> datasets.Features:
+        features = {}
+        for dataset in self._datasets:
+            features.update({k: v for k, v in dataset.features.items() if k not in self.disabled_data_keys})
+        return features
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """Keys to access image and video stream from cameras."""
+        keys = []
+        for key, feats in self.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.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 sum(d.num_samples for d in self._datasets)
+
+    @property
+    def num_episodes(self) -> int:
+        """Number of episodes."""
+        return sum(d.num_episodes for d in self._datasets)
+
+    @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: int) -> dict[str, torch.Tensor]:
+        if idx >= len(self):
+            raise IndexError(f"Index {idx} out of bounds.")
+        # Determine which dataset to get an item from based on the index.
+        start_idx = 0
+        dataset_idx = 0
+        for dataset in self._datasets:
+            if idx >= start_idx + dataset.num_samples:
+                start_idx += dataset.num_samples
+                dataset_idx += 1
+                continue
+            break
+        else:
+            raise AssertionError("We expect the loop to break out as long as the index is within bounds.")
+        item = self._datasets[dataset_idx][idx - start_idx]
+        item["dataset_index"] = torch.tensor(dataset_idx)
+        for data_key in self.disabled_data_keys:
+            if data_key in item:
+                del item[data_key]
+
+        return item
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}(\n"
+            f"  Repository IDs: '{self.repo_ids}',\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.image_transforms},\n"
+            f")"
+        )

lerobot/common/datasets/online_buffer.py

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

lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md

@@ -0,0 +1,56 @@
+## Using / Updating `CODEBASE_VERSION` (for maintainers)
+
+Since our dataset pushed to the hub are decoupled with the evolution of this repo, we ensure compatibility of
+the datasets with our code, we use a `CODEBASE_VERSION` (defined in
+lerobot/common/datasets/lerobot_dataset.py) variable.
+
+For instance, [`lerobot/pusht`](https://huggingface.co/datasets/lerobot/pusht) has many versions to maintain backward compatibility between LeRobot codebase versions:
+- [v1.0](https://huggingface.co/datasets/lerobot/pusht/tree/v1.0)
+- [v1.1](https://huggingface.co/datasets/lerobot/pusht/tree/v1.1)
+- [v1.2](https://huggingface.co/datasets/lerobot/pusht/tree/v1.2)
+- [v1.3](https://huggingface.co/datasets/lerobot/pusht/tree/v1.3)
+- [v1.4](https://huggingface.co/datasets/lerobot/pusht/tree/v1.4)
+- [v1.5](https://huggingface.co/datasets/lerobot/pusht/tree/v1.5)
+- [v1.6](https://huggingface.co/datasets/lerobot/pusht/tree/v1.6) <-- last version
+- [main](https://huggingface.co/datasets/lerobot/pusht/tree/main) <-- points to the last version
+
+Starting with v1.6, every dataset pushed to the hub or saved locally also have this version number in their
+`info.json` metadata.
+
+### Uploading a new dataset
+If you are pushing a new dataset, you don't need to worry about any of the instructions below, nor to be
+compatible with previous codebase versions. The `push_dataset_to_hub.py` script will automatically tag your
+dataset with the current `CODEBASE_VERSION`.
+
+### Updating an existing dataset
+If you want to update an existing dataset, you need to change the `CODEBASE_VERSION` from `lerobot_dataset.py`
+before running `push_dataset_to_hub.py`. This is especially useful if you introduce a breaking change
+intentionally or not (i.e. something not backward compatible such as modifying the reward functions used,
+deleting some frames at the end of an episode, etc.). That way, people running a previous version of the
+codebase won't be affected by your change and backward compatibility is maintained.
+
+However, you will need to update the version of ALL the other datasets so that they have the new
+`CODEBASE_VERSION` as a branch in their hugging face dataset repository. Don't worry, there is an easy way
+that doesn't require to run `push_dataset_to_hub.py`. You can just "branch-out" from the `main` branch on HF
+dataset repo by running this script which corresponds to a `git checkout -b` (so no copy or upload needed):
+
+```python
+from huggingface_hub import HfApi
+
+from lerobot import available_datasets
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+
+api = HfApi()
+
+for repo_id in available_datasets:
+    dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
+    branches = [b.name for b in dataset_info.branches]
+    if CODEBASE_VERSION in branches:
+        print(f"{repo_id} already @{CODEBASE_VERSION}, skipping.")
+        continue
+    else:
+        # Now create a branch named after the new version by branching out from "main"
+        # which is expected to be the preceding version
+        api.create_branch(repo_id, repo_type="dataset", branch=CODEBASE_VERSION, revision="main")
+        print(f"{repo_id} successfully updated @{CODEBASE_VERSION}")
+```

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

@@ -1,3 +1,18 @@
+#!/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.

lerobot/common/datasets/push_dataset_to_hub/_download_raw.py

@@ -0,0 +1,134 @@
+#!/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 download scripts for raw datasets.
+
+Example of usage:
+```
+python lerobot/common/datasets/push_dataset_to_hub/_download_raw.py \
+--raw-dir data/lerobot-raw/pusht_raw \
+--repo-id lerobot-raw/pusht_raw
+```
+"""
+
+import argparse
+import logging
+import warnings
+from pathlib import Path
+
+from huggingface_hub import snapshot_download
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
+
+# {raw_repo_id: raw_format}
+AVAILABLE_RAW_REPO_IDS = {
+    "lerobot-raw/aloha_mobile_cabinet_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_chair_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_elevator_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_shrimp_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_wash_pan_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_wipe_wine_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_insertion_human_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_insertion_scripted_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_transfer_cube_human_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_transfer_cube_scripted_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_battery_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_candy_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_coffee_new_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_coffee_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_cups_open_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_fork_pick_up_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_pingpong_test_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_pro_pencil_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_screw_driver_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_tape_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_thread_velcro_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_towel_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_vinh_cup_left_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_vinh_cup_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_ziploc_slide_raw": "aloha_hdf5",
+    "lerobot-raw/pusht_raw": "pusht_zarr",
+    "lerobot-raw/umi_cup_in_the_wild_raw": "umi_zarr",
+    "lerobot-raw/unitreeh1_fold_clothes_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_rearrange_objects_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_two_robot_greeting_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_warehouse_raw": "aloha_hdf5",
+    "lerobot-raw/xarm_lift_medium_raw": "xarm_pkl",
+    "lerobot-raw/xarm_lift_medium_replay_raw": "xarm_pkl",
+    "lerobot-raw/xarm_push_medium_raw": "xarm_pkl",
+    "lerobot-raw/xarm_push_medium_replay_raw": "xarm_pkl",
+}
+
+
+def download_raw(raw_dir: Path, repo_id: str):
+    check_repo_id(repo_id)
+    user_id, dataset_id = repo_id.split("/")
+
+    if not dataset_id.endswith("_raw"):
+        warnings.warn(
+            f"""`dataset_id` ({dataset_id}) doesn't end with '_raw' (e.g. 'lerobot/pusht_raw'). Following this
+             naming convention by renaming your repository is advised, but not mandatory.""",
+            stacklevel=1,
+        )
+
+    # Send warning if raw_dir isn't well formated
+    if raw_dir.parts[-2] != user_id or raw_dir.parts[-1] != dataset_id:
+        warnings.warn(
+            f"""`raw_dir` ({raw_dir}) doesn't contain a community or user id `/` the name of the dataset that
+             match the `repo_id` (e.g. 'data/lerobot/pusht_raw'). Following this naming convention is advised,
+             but not mandatory.""",
+            stacklevel=1,
+        )
+    raw_dir.mkdir(parents=True, exist_ok=True)
+
+    logging.info(f"Start downloading from huggingface.co/{user_id} for {dataset_id}")
+    snapshot_download(repo_id, repo_type="dataset", local_dir=raw_dir)
+    logging.info(f"Finish downloading from huggingface.co/{user_id} for {dataset_id}")
+
+
+def download_all_raw_datasets(data_dir: Path | None = None):
+    if data_dir is None:
+        data_dir = Path("data")
+    for repo_id in AVAILABLE_RAW_REPO_IDS:
+        raw_dir = data_dir / repo_id
+        download_raw(raw_dir, repo_id)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=f"""A script to download raw datasets from Hugging Face hub to a local directory. Here is a
+            non exhaustive list of available repositories to use in `--repo-id`: {AVAILABLE_RAW_REPO_IDS}""",
+    )
+
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        required=True,
+        help="Directory containing input raw datasets (e.g. `data/aloha_mobile_chair_raw` or `data/pusht_raw).",
+    )
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="""Repositery identifier on Hugging Face: a community or a user name `/` the name of
+        the dataset (e.g. `lerobot/pusht_raw`, `cadene/aloha_sim_insertion_human_raw`).""",
+    )
+    args = parser.parse_args()
+    download_raw(**vars(args))
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py

@@ -0,0 +1,184 @@
+#!/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.
+"""
+Use this script to batch encode lerobot dataset from their raw format to LeRobotDataset and push their updated
+version to the hub. Under the hood, this script reuses 'push_dataset_to_hub.py'. It assumes that you already
+downloaded raw datasets, which you can do with the related '_download_raw.py' script.
+
+For instance, for codebase_version = 'v1.6', the following command was run, assuming raw datasets from
+lerobot-raw were downloaded in 'raw/datasets/directory':
+```bash
+python lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py \
+  --raw-dir raw/datasets/directory \
+  --raw-repo-ids lerobot-raw \
+  --local-dir push/datasets/directory \
+  --tests-data-dir tests/data \
+  --push-repo lerobot \
+  --vcodec libsvtav1 \
+  --pix-fmt yuv420p \
+  --g 2 \
+  --crf 30
+```
+"""
+
+import argparse
+from pathlib import Path
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub._download_raw import AVAILABLE_RAW_REPO_IDS
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
+from lerobot.scripts.push_dataset_to_hub import push_dataset_to_hub
+
+
+def get_push_repo_id_from_raw(raw_repo_id: str, push_repo: str) -> str:
+    dataset_id_raw = raw_repo_id.split("/")[1]
+    dataset_id = dataset_id_raw.removesuffix("_raw")
+    return f"{push_repo}/{dataset_id}"
+
+
+def encode_datasets(
+    raw_dir: Path,
+    raw_repo_ids: list[str],
+    push_repo: str,
+    vcodec: str,
+    pix_fmt: str,
+    g: int,
+    crf: int,
+    local_dir: Path | None = None,
+    tests_data_dir: Path | None = None,
+    raw_format: str | None = None,
+    dry_run: bool = False,
+) -> None:
+    if len(raw_repo_ids) == 1 and raw_repo_ids[0].lower() == "lerobot-raw":
+        raw_repo_ids_format = AVAILABLE_RAW_REPO_IDS
+    else:
+        if raw_format is None:
+            raise ValueError(raw_format)
+        raw_repo_ids_format = {id_: raw_format for id_ in raw_repo_ids}
+
+    for raw_repo_id, repo_raw_format in raw_repo_ids_format.items():
+        check_repo_id(raw_repo_id)
+        dataset_repo_id_push = get_push_repo_id_from_raw(raw_repo_id, push_repo)
+        dataset_raw_dir = raw_dir / raw_repo_id
+        dataset_dir = local_dir / dataset_repo_id_push if local_dir is not None else None
+        encoding = {
+            "vcodec": vcodec,
+            "pix_fmt": pix_fmt,
+            "g": g,
+            "crf": crf,
+        }
+
+        if not (dataset_raw_dir).is_dir():
+            raise NotADirectoryError(dataset_raw_dir)
+
+        if not dry_run:
+            push_dataset_to_hub(
+                dataset_raw_dir,
+                raw_format=repo_raw_format,
+                repo_id=dataset_repo_id_push,
+                local_dir=dataset_dir,
+                resume=True,
+                encoding=encoding,
+                tests_data_dir=tests_data_dir,
+            )
+        else:
+            print(
+                f"DRY RUN: {dataset_raw_dir}  -->  {dataset_dir}  -->  {dataset_repo_id_push}@{CODEBASE_VERSION}"
+            )
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        default=Path("data"),
+        help="Directory where raw datasets are located.",
+    )
+    parser.add_argument(
+        "--raw-repo-ids",
+        type=str,
+        nargs="*",
+        default=["lerobot-raw"],
+        help="""Raw dataset repo ids. if 'lerobot-raw', the keys from `AVAILABLE_RAW_REPO_IDS` will be
+            used and raw datasets will be fetched from the 'lerobot-raw/' repo and pushed with their
+            associated format. It is assumed that each dataset is located at `raw_dir / raw_repo_id` """,
+    )
+    parser.add_argument(
+        "--raw-format",
+        type=str,
+        default=None,
+        help="""Raw format to use for the raw repo-ids. Must be specified if --raw-repo-ids is not
+            'lerobot-raw'""",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="""When provided, writes the dataset converted to LeRobotDataset format in this directory
+        (e.g. `data/lerobot/aloha_mobile_chair`).""",
+    )
+    parser.add_argument(
+        "--push-repo",
+        type=str,
+        default="lerobot",
+        help="Repo to upload datasets to",
+    )
+    parser.add_argument(
+        "--vcodec",
+        type=str,
+        default="libsvtav1",
+        help="Codec to use for encoding videos",
+    )
+    parser.add_argument(
+        "--pix-fmt",
+        type=str,
+        default="yuv420p",
+        help="Pixel formats (chroma subsampling) to be used for encoding",
+    )
+    parser.add_argument(
+        "--g",
+        type=int,
+        default=2,
+        help="Group of pictures sizes to be used for encoding.",
+    )
+    parser.add_argument(
+        "--crf",
+        type=int,
+        default=30,
+        help="Constant rate factors to be used for encoding.",
+    )
+    parser.add_argument(
+        "--tests-data-dir",
+        type=Path,
+        default=None,
+        help=(
+            "When provided, save tests artifacts into the given directory "
+            "(e.g. `--tests-data-dir tests/data` will save to tests/data/{--repo-id})."
+        ),
+    )
+    parser.add_argument(
+        "--dry-run",
+        type=int,
+        default=0,
+        help="If not set to 0, this script won't download or upload anything.",
+    )
+    args = parser.parse_args()
+    encode_datasets(**vars(args))
+
+
+if __name__ == "__main__":
+    main()

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 +1,233 @@
+#!/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.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
+    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:
+    # only frames from simulation are uncompressed
+    compressed_images = "sim" not in raw_dir.name
+
+    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]
+
+                if compressed_images:
+                    assert data[f"/observations/images/{camera}"].ndim == 2
+                else:
+                    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."
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # only frames from simulation are uncompressed
+    compressed_images = "sim" not in raw_dir.name
+
+    hdf5_files = sorted(raw_dir.glob("episode_*.hdf5"))
+    num_episodes = len(hdf5_files)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx in tqdm.tqdm(ep_ids):
+        ep_path = hdf5_files[ep_idx]
+        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 = videos_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 = videos_dir / fname
+                    encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                    # 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)
+
+        gc.collect()
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+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,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 50
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/cam_png_format.py

@@ -0,0 +1,104 @@
+#!/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 png images files recorded with capture_camera_feed.py
+"""
+
+from pathlib import Path
+
+import torch
+from datasets import Dataset, Features, Image, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes
+from lerobot.common.datasets.utils import calculate_episode_data_index, hf_transform_to_torch
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def check_format(raw_dir: Path) -> bool:
+    image_paths = list(raw_dir.glob("frame_*.png"))
+    if len(image_paths) == 0:
+        raise ValueError
+
+
+def load_from_raw(raw_dir: Path, fps: int, episodes: list[int] | None = None):
+    if episodes is not None:
+        # TODO(aliberts): add support for multi-episodes.
+        raise NotImplementedError()
+
+    ep_dict = {}
+    ep_idx = 0
+
+    image_paths = sorted(raw_dir.glob("frame_*.png"))
+    num_frames = len(image_paths)
+
+    ep_dict["observation.image"] = [PILImage.open(x) for x in image_paths]
+    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_dicts = [ep_dict]
+    data_dict = concatenate_episodes(ep_dicts)
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    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)
+
+    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,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    if video or episodes or encoding is not None:
+        # TODO(aliberts): support this
+        raise NotImplementedError
+
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/dora_parquet_format.py

@@ -0,0 +1,233 @@
+#!/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 from dora-record
+"""
+
+import re
+import warnings
+from pathlib import Path
+
+import pandas as pd
+import torch
+from datasets import Dataset, Features, Image, Sequence, Value
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def check_format(raw_dir) -> bool:
+    assert raw_dir.exists()
+
+    leader_file = list(raw_dir.glob("*.parquet"))
+    if len(leader_file) == 0:
+        raise ValueError(f"Missing parquet files in '{raw_dir}'")
+    return True
+
+
+def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episodes: list[int] | None = None):
+    # Load data stream that will be used as reference for the timestamps synchronization
+    reference_files = list(raw_dir.glob("observation.images.cam_*.parquet"))
+    if len(reference_files) == 0:
+        raise ValueError(f"Missing reference files for camera, starting with  in '{raw_dir}'")
+    # select first camera in alphanumeric order
+    reference_key = sorted(reference_files)[0].stem
+    reference_df = pd.read_parquet(raw_dir / f"{reference_key}.parquet")
+    reference_df = reference_df[["timestamp_utc", reference_key]]
+
+    # Merge all data stream using nearest backward strategy
+    df = reference_df
+    for path in raw_dir.glob("*.parquet"):
+        key = path.stem  # action or observation.state or ...
+        if key == reference_key:
+            continue
+        if "failed_episode_index" in key:
+            # TODO(rcadene): add support for removing episodes that are tagged as "failed"
+            continue
+        modality_df = pd.read_parquet(path)
+        modality_df = modality_df[["timestamp_utc", key]]
+        df = pd.merge_asof(
+            df,
+            modality_df,
+            on="timestamp_utc",
+            # "nearest" is the best option over "backward", since the latter can desynchronizes camera timestamps by
+            # matching timestamps that are too far appart, in order to fit the backward constraints. It's not the case for "nearest".
+            # However, note that "nearest" might synchronize the reference camera with other cameras on slightly future timestamps.
+            # are too far appart.
+            direction="nearest",
+            tolerance=pd.Timedelta(f"{1/fps} seconds"),
+        )
+    # Remove rows with episode_index -1 which indicates data that correspond to in-between episodes
+    df = df[df["episode_index"] != -1]
+
+    image_keys = [key for key in df if "observation.images." in key]
+
+    def get_episode_index(row):
+        episode_index_per_cam = {}
+        for key in image_keys:
+            path = row[key][0]["path"]
+            match = re.search(r"_(\d{6}).mp4", path)
+            if not match:
+                raise ValueError(path)
+            episode_index = int(match.group(1))
+            episode_index_per_cam[key] = episode_index
+        if len(set(episode_index_per_cam.values())) != 1:
+            raise ValueError(
+                f"All cameras are expected to belong to the same episode, but getting {episode_index_per_cam}"
+            )
+        return episode_index
+
+    df["episode_index"] = df.apply(get_episode_index, axis=1)
+
+    # dora only use arrays, so single values are encapsulated into a list
+    df["frame_index"] = df.groupby("episode_index").cumcount()
+    df = df.reset_index()
+    df["index"] = df.index
+
+    # set 'next.done' to True for the last frame of each episode
+    df["next.done"] = False
+    df.loc[df.groupby("episode_index").tail(1).index, "next.done"] = True
+
+    df["timestamp"] = df["timestamp_utc"].map(lambda x: x.timestamp())
+    # each episode starts with timestamp 0 to match the ones from the video
+    df["timestamp"] = df.groupby("episode_index")["timestamp"].transform(lambda x: x - x.iloc[0])
+
+    del df["timestamp_utc"]
+
+    # sanity check
+    has_nan = df.isna().any().any()
+    if has_nan:
+        raise ValueError("Dataset contains Nan values.")
+
+    # sanity check episode indices go from 0 to n-1
+    ep_ids = [ep_idx for ep_idx, _ in df.groupby("episode_index")]
+    expected_ep_ids = list(range(df["episode_index"].max() + 1))
+    if ep_ids != expected_ep_ids:
+        raise ValueError(f"Episodes indices go from {ep_ids} instead of {expected_ep_ids}")
+
+    # Create symlink to raw videos directory (that needs to be absolute not relative)
+    videos_dir.parent.mkdir(parents=True, exist_ok=True)
+    videos_dir.symlink_to((raw_dir / "videos").absolute())
+
+    # sanity check the video paths are well formated
+    for key in df:
+        if "observation.images." not in key:
+            continue
+        for ep_idx in ep_ids:
+            video_path = videos_dir / f"{key}_episode_{ep_idx:06d}.mp4"
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
+
+    data_dict = {}
+    for key in df:
+        # is video frame
+        if "observation.images." in key:
+            # we need `[0] because dora only use arrays, so single values are encapsulated into a list.
+            # it is the case for video_frame dictionary = [{"path": ..., "timestamp": ...}]
+            data_dict[key] = [video_frame[0] for video_frame in df[key].values]
+
+            # sanity check the video path is well formated
+            video_path = videos_dir.parent / data_dict[key][0]["path"]
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
+        # is number
+        elif df[key].iloc[0].ndim == 0 or df[key].iloc[0].shape[0] == 1:
+            data_dict[key] = torch.from_numpy(df[key].values)
+        # is vector
+        elif df[key].iloc[0].shape[0] > 1:
+            data_dict[key] = torch.stack([torch.from_numpy(x.copy()) for x in df[key].values])
+        else:
+            raise ValueError(key)
+
+    return data_dict
+
+
+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,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+    else:
+        raise NotImplementedError()
+
+    if not video:
+        raise NotImplementedError()
+
+    if encoding is not None:
+        warnings.warn(
+            "Video encoding is currently done outside of LeRobot for the dora_parquet format.",
+            stacklevel=1,
+        )
+
+    data_df = load_from_raw(raw_dir, videos_dir, fps, episodes)
+    hf_dataset = to_hf_dataset(data_df, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = "unknown"
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -0,0 +1,275 @@
+#!/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.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
+    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: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    keypoints_instead_of_image: bool = False,
+    encoding: dict | None = None,
+):
+    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())
+    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"])
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx = 0
+    for to_idx in zarr_data.meta["episode_ends"]:
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    num_episodes = len(from_ids)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        from_idx = from_ids[selected_ep_idx]
+        to_idx = to_ids[selected_ep_idx]
+        num_frames = to_idx - from_idx
+
+        # sanity check
+        assert (episode_ids[from_idx:to_idx] == ep_idx).all()
+
+        # get image
+        if not keypoints_instead_of_image:
+            image = imgs[from_idx:to_idx]
+            assert image.min() >= 0.0
+            assert image.max() <= 255.0
+            image = image.type(torch.uint8)
+
+        # get state
+        state = states[from_idx:to_idx]
+        agent_pos = state[:, :2]
+        block_pos = state[:, 2:4]
+        block_angle = state[:, 4]
+
+        # get reward, success, done, and (maybe) keypoints
+        reward = torch.zeros(num_frames)
+        success = torch.zeros(num_frames, dtype=torch.bool)
+        if keypoints_instead_of_image:
+            keypoints = torch.zeros(num_frames, 16)  # 8 keypoints each with 2 coords
+        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, block_shapes = 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
+            if keypoints_instead_of_image:
+                keypoints[i] = torch.from_numpy(PushTEnv.get_keypoints(block_shapes).flatten())
+
+        # last step of demonstration is considered done
+        done[-1] = True
+
+        ep_dict = {}
+
+        if not keypoints_instead_of_image:
+            imgs_array = [x.numpy() for x in image]
+            img_key = "observation.image"
+            if video:
+                # save png images in temporary directory
+                tmp_imgs_dir = videos_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 = videos_dir / fname
+                encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                # 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
+        if keypoints_instead_of_image:
+            ep_dict["observation.environment_state"] = keypoints
+        ep_dict["action"] = actions[from_idx:to_idx]
+        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)
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video, keypoints_instead_of_image: bool = False):
+    features = {}
+
+    if not keypoints_instead_of_image:
+        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)
+    )
+    if keypoints_instead_of_image:
+        features["observation.environment_state"] = Sequence(
+            length=data_dict["observation.environment_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,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # Manually change this to True to use keypoints of the T instead of an image observation (but don't merge
+    # with True). Also make sure to use video = 0 in the `push_dataset_to_hub.py` script.
+    keypoints_instead_of_image = False
+
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 10
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, keypoints_instead_of_image, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video, keypoints_instead_of_image)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video if not keypoints_instead_of_image else 0,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -0,0 +1,234 @@
+#!/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 torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+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,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
+    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 load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    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]
+
+    # We convert it in torch tensor later because the jit function does not support torch tensors
+    episode_ends = torch.from_numpy(episode_ends)
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx = 0
+    for to_idx in episode_ends:
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    ep_dicts_dir = videos_dir / "ep_dicts"
+    ep_dicts_dir.mkdir(exist_ok=True, parents=True)
+    ep_dicts = []
+
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        ep_dict_path = ep_dicts_dir / f"{ep_idx}"
+        if not ep_dict_path.is_file():
+            from_idx = from_ids[selected_ep_idx]
+            to_idx = to_ids[selected_ep_idx]
+            num_frames = to_idx - from_idx
+
+            # TODO(rcadene): save temporary images of the episode?
+
+            state = states[from_idx:to_idx]
+
+            ep_dict = {}
+
+            # load 57MB of images in RAM (400x224x224x3 uint8)
+            imgs_array = zarr_data["data/camera0_rgb"][from_idx:to_idx]
+            img_key = "observation.image"
+            if video:
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                if not video_path.is_file():
+                    # save png images in temporary directory
+                    tmp_imgs_dir = videos_dir / "tmp_images"
+                    save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                    # encode images to a mp4 video
+                    encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                    # 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([from_idx] * num_frames)
+            ep_dict["episode_data_index_to"] = torch.tensor([from_idx + num_frames] * num_frames)
+            ep_dict["end_pose"] = end_pose[from_idx:to_idx]
+            ep_dict["start_pos"] = start_pos[from_idx:to_idx]
+            ep_dict["gripper_width"] = gripper_width[from_idx:to_idx]
+            torch.save(ep_dict, ep_dict_path)
+        else:
+            ep_dict = torch.load(ep_dict_path)
+
+        ep_dicts.append(ep_dict)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+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,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # 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 = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -0,0 +1,74 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import inspect
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+import numpy
+import PIL
+import torch
+
+from lerobot.common.datasets.video_utils import encode_video_frames
+
+
+def concatenate_episodes(ep_dicts):
+    data_dict = {}
+
+    keys = ep_dicts[0].keys()
+    for key in keys:
+        if torch.is_tensor(ep_dicts[0][key][0]):
+            data_dict[key] = torch.cat([ep_dict[key] for ep_dict in ep_dicts])
+        else:
+            if key not in data_dict:
+                data_dict[key] = []
+            for ep_dict in ep_dicts:
+                for x in ep_dict[key]:
+                    data_dict[key].append(x)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def save_images_concurrently(imgs_array: numpy.array, out_dir: Path, max_workers: int = 4):
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    def save_image(img_array, i, out_dir):
+        img = PIL.Image.fromarray(img_array)
+        img.save(str(out_dir / f"frame_{i:06d}.png"), quality=100)
+
+    num_images = len(imgs_array)
+    with ThreadPoolExecutor(max_workers=max_workers) as executor:
+        [executor.submit(save_image, imgs_array[i], i, out_dir) for i in range(num_images)]
+
+
+def get_default_encoding() -> dict:
+    """Returns the default ffmpeg encoding parameters used by `encode_video_frames`."""
+    signature = inspect.signature(encode_video_frames)
+    return {
+        k: v.default
+        for k, v in signature.parameters.items()
+        if v.default is not inspect.Parameter.empty and k in ["vcodec", "pix_fmt", "g", "crf"]
+    }
+
+
+def check_repo_id(repo_id: str) -> None:
+    if len(repo_id.split("/")) != 2:
+        raise ValueError(
+            f"""`repo_id` is expected to contain a community or user id `/` the name of the dataset
+            (e.g. 'lerobot/pusht'), but contains '{repo_id}'."""
+        )

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.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.
+"""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.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
+    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: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    pkl_path = raw_dir / "buffer.pkl"
+
+    with open(pkl_path, "rb") as f:
+        pkl_data = pickle.load(f)
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx, to_idx = 0, 0
+    for done in pkl_data["dones"]:
+        to_idx += 1
+        if not done:
+            continue
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    num_episodes = len(from_ids)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        from_idx = from_ids[selected_ep_idx]
+        to_idx = to_ids[selected_ep_idx]
+        num_frames = to_idx - from_idx
+
+        image = torch.tensor(pkl_data["observations"]["rgb"][from_idx:to_idx])
+        image = einops.rearrange(image, "b c h w -> b h w c")
+        state = torch.tensor(pkl_data["observations"]["state"][from_idx:to_idx])
+        action = torch.tensor(pkl_data["actions"][from_idx:to_idx])
+        # 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"][from_idx:to_idx])
+        # next_state = torch.tensor(pkl_data["next_observations"]["state"][from_idx:to_idx])
+        next_reward = torch.tensor(pkl_data["rewards"][from_idx:to_idx])
+        next_done = torch.tensor(pkl_data["dones"][from_idx:to_idx])
+
+        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 = videos_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 = videos_dir / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+            # 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)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+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,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 15
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/sampler.py

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

lerobot/common/datasets/transforms.py

@@ -0,0 +1,197 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import collections
+from typing import Any, Callable, Dict, Sequence
+
+import torch
+from torchvision.transforms import v2
+from torchvision.transforms.v2 import Transform
+from torchvision.transforms.v2 import functional as F  # noqa: N812
+
+
+class RandomSubsetApply(Transform):
+    """Apply a random subset of N transformations from a list of transformations.
+
+    Args:
+        transforms: list of transformations.
+        p: represents the multinomial probabilities (with no replacement) used for sampling the transform.
+            If the sum of the weights is not 1, they will be normalized. If ``None`` (default), all transforms
+            have the same probability.
+        n_subset: number of transformations to apply. If ``None``, all transforms are applied.
+            Must be in [1, len(transforms)].
+        random_order: apply transformations in a random order.
+    """
+
+    def __init__(
+        self,
+        transforms: Sequence[Callable],
+        p: list[float] | None = None,
+        n_subset: int | None = None,
+        random_order: bool = False,
+    ) -> None:
+        super().__init__()
+        if not isinstance(transforms, Sequence):
+            raise TypeError("Argument transforms should be a sequence of callables")
+        if p is None:
+            p = [1] * len(transforms)
+        elif len(p) != len(transforms):
+            raise ValueError(
+                f"Length of p doesn't match the number of transforms: {len(p)} != {len(transforms)}"
+            )
+
+        if n_subset is None:
+            n_subset = len(transforms)
+        elif not isinstance(n_subset, int):
+            raise TypeError("n_subset should be an int or None")
+        elif not (1 <= n_subset <= len(transforms)):
+            raise ValueError(f"n_subset should be in the interval [1, {len(transforms)}]")
+
+        self.transforms = transforms
+        total = sum(p)
+        self.p = [prob / total for prob in p]
+        self.n_subset = n_subset
+        self.random_order = random_order
+
+    def forward(self, *inputs: Any) -> Any:
+        needs_unpacking = len(inputs) > 1
+
+        selected_indices = torch.multinomial(torch.tensor(self.p), self.n_subset)
+        if not self.random_order:
+            selected_indices = selected_indices.sort().values
+
+        selected_transforms = [self.transforms[i] for i in selected_indices]
+
+        for transform in selected_transforms:
+            outputs = transform(*inputs)
+            inputs = outputs if needs_unpacking else (outputs,)
+
+        return outputs
+
+    def extra_repr(self) -> str:
+        return (
+            f"transforms={self.transforms}, "
+            f"p={self.p}, "
+            f"n_subset={self.n_subset}, "
+            f"random_order={self.random_order}"
+        )
+
+
+class SharpnessJitter(Transform):
+    """Randomly change the sharpness of an image or video.
+
+    Similar to a v2.RandomAdjustSharpness with p=1 and a sharpness_factor sampled randomly.
+    While v2.RandomAdjustSharpness applies — with a given probability — a fixed sharpness_factor to an image,
+    SharpnessJitter applies a random sharpness_factor each time. This is to have a more diverse set of
+    augmentations as a result.
+
+    A sharpness_factor of 0 gives a blurred image, 1 gives the original image while 2 increases the sharpness
+    by a factor of 2.
+
+    If the input is a :class:`torch.Tensor`,
+    it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions.
+
+    Args:
+        sharpness: How much to jitter sharpness. sharpness_factor is chosen uniformly from
+            [max(0, 1 - sharpness), 1 + sharpness] or the given
+            [min, max]. Should be non negative numbers.
+    """
+
+    def __init__(self, sharpness: float | Sequence[float]) -> None:
+        super().__init__()
+        self.sharpness = self._check_input(sharpness)
+
+    def _check_input(self, sharpness):
+        if isinstance(sharpness, (int, float)):
+            if sharpness < 0:
+                raise ValueError("If sharpness is a single number, it must be non negative.")
+            sharpness = [1.0 - sharpness, 1.0 + sharpness]
+            sharpness[0] = max(sharpness[0], 0.0)
+        elif isinstance(sharpness, collections.abc.Sequence) and len(sharpness) == 2:
+            sharpness = [float(v) for v in sharpness]
+        else:
+            raise TypeError(f"{sharpness=} should be a single number or a sequence with length 2.")
+
+        if not 0.0 <= sharpness[0] <= sharpness[1]:
+            raise ValueError(f"sharpnesss values should be between (0., inf), but got {sharpness}.")
+
+        return float(sharpness[0]), float(sharpness[1])
+
+    def _generate_value(self, left: float, right: float) -> float:
+        return torch.empty(1).uniform_(left, right).item()
+
+    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+        sharpness_factor = self._generate_value(self.sharpness[0], self.sharpness[1])
+        return self._call_kernel(F.adjust_sharpness, inpt, sharpness_factor=sharpness_factor)
+
+
+def get_image_transforms(
+    brightness_weight: float = 1.0,
+    brightness_min_max: tuple[float, float] | None = None,
+    contrast_weight: float = 1.0,
+    contrast_min_max: tuple[float, float] | None = None,
+    saturation_weight: float = 1.0,
+    saturation_min_max: tuple[float, float] | None = None,
+    hue_weight: float = 1.0,
+    hue_min_max: tuple[float, float] | None = None,
+    sharpness_weight: float = 1.0,
+    sharpness_min_max: tuple[float, float] | None = None,
+    max_num_transforms: int | None = None,
+    random_order: bool = False,
+):
+    def check_value(name, weight, min_max):
+        if min_max is not None:
+            if len(min_max) != 2:
+                raise ValueError(
+                    f"`{name}_min_max` is expected to be a tuple of 2 dimensions, but {min_max} provided."
+                )
+            if weight < 0.0:
+                raise ValueError(
+                    f"`{name}_weight` is expected to be 0 or positive, but is negative ({weight})."
+                )
+
+    check_value("brightness", brightness_weight, brightness_min_max)
+    check_value("contrast", contrast_weight, contrast_min_max)
+    check_value("saturation", saturation_weight, saturation_min_max)
+    check_value("hue", hue_weight, hue_min_max)
+    check_value("sharpness", sharpness_weight, sharpness_min_max)
+
+    weights = []
+    transforms = []
+    if brightness_min_max is not None and brightness_weight > 0.0:
+        weights.append(brightness_weight)
+        transforms.append(v2.ColorJitter(brightness=brightness_min_max))
+    if contrast_min_max is not None and contrast_weight > 0.0:
+        weights.append(contrast_weight)
+        transforms.append(v2.ColorJitter(contrast=contrast_min_max))
+    if saturation_min_max is not None and saturation_weight > 0.0:
+        weights.append(saturation_weight)
+        transforms.append(v2.ColorJitter(saturation=saturation_min_max))
+    if hue_min_max is not None and hue_weight > 0.0:
+        weights.append(hue_weight)
+        transforms.append(v2.ColorJitter(hue=hue_min_max))
+    if sharpness_min_max is not None and sharpness_weight > 0.0:
+        weights.append(sharpness_weight)
+        transforms.append(SharpnessJitter(sharpness=sharpness_min_max))
+
+    n_subset = len(transforms)
+    if max_num_transforms is not None:
+        n_subset = min(n_subset, max_num_transforms)
+
+    if n_subset == 0:
+        return v2.Identity()
+    else:
+        # TODO(rcadene, aliberts): add v2.ToDtype float16?
+        return RandomSubsetApply(transforms, p=weights, n_subset=n_subset, random_order=random_order)

lerobot/common/datasets/utils.py

@@ -1,18 +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.
 import json
-from copy import deepcopy
-from math import ceil
+import re
+import warnings
+from functools import cache
 from pathlib import Path
+from typing import Dict
 
 import datasets
-import einops
 import torch
-import tqdm
-from datasets import Image, load_dataset, load_from_disk
-from huggingface_hub import hf_hub_download
+from datasets import load_dataset, load_from_disk
+from huggingface_hub import DatasetCard, HfApi, hf_hub_download, snapshot_download
 from PIL import Image as PILImage
 from safetensors.torch import load_file
 from torchvision import transforms
 
+DATASET_CARD_TEMPLATE = """
+---
+# Metadata will go there
+---
+This dataset was created using [🤗 LeRobot](https://github.com/huggingface/lerobot).
+
+"""
+
 
 def flatten_dict(d, parent_key="", sep="/"):
     """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
@@ -46,7 +69,7 @@ def unflatten_dict(d, sep="/"):
     return outdict
 
 
-def hf_transform_to_torch(items_dict):
+def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     """Get a transform function that convert items from Hugging Face dataset (pyarrow)
     to torch tensors. Importantly, images are converted from PIL, which corresponds to
     a channel last representation (h w c) of uint8 type, to a torch image representation
@@ -57,17 +80,61 @@ def hf_transform_to_torch(items_dict):
         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
+        elif first_item is None:
+            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:
+@cache
+def get_hf_dataset_safe_version(repo_id: str, version: str) -> str:
+    api = HfApi()
+    dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
+    branches = [b.name for b in dataset_info.branches]
+    if version not in branches:
+        warnings.warn(
+            f"""You are trying to load a dataset from {repo_id} created with a previous version of the
+            codebase. The following versions are available: {branches}.
+            The requested version ('{version}') is not found. You should be fine since
+            backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
+            Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
+            stacklevel=1,
+        )
+        if "main" not in branches:
+            raise ValueError(f"Version 'main' not found on {repo_id}")
+        return "main"
+    else:
+        return version
+
+
+def load_hf_dataset(repo_id: str, version: str, root: Path, split: str) -> 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 / split))
+        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:
-        hf_dataset = load_dataset(repo_id, revision=version, split=split)
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        hf_dataset = load_dataset(repo_id, revision=safe_version, split=split)
+
     hf_dataset.set_transform(hf_transform_to_torch)
     return hf_dataset
 
@@ -85,8 +152,9 @@ def load_episode_data_index(repo_id, version, root) -> dict[str, torch.Tensor]:
     if root is not None:
         path = Path(root) / repo_id / "meta_data" / "episode_data_index.safetensors"
     else:
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
         path = hf_hub_download(
-            repo_id, "meta_data/episode_data_index.safetensors", repo_type="dataset", revision=version
+            repo_id, "meta_data/episode_data_index.safetensors", repo_type="dataset", revision=safe_version
         )
 
     return load_file(path)
@@ -103,7 +171,10 @@ def load_stats(repo_id, version, root) -> dict[str, dict[str, torch.Tensor]]:
     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)
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        path = hf_hub_download(
+            repo_id, "meta_data/stats.safetensors", repo_type="dataset", revision=safe_version
+        )
 
     stats = load_file(path)
     return unflatten_dict(stats)
@@ -120,24 +191,38 @@ def load_info(repo_id, version, root) -> dict:
     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)
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        path = hf_hub_download(repo_id, "meta_data/info.json", repo_type="dataset", revision=safe_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
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        repo_dir = snapshot_download(repo_id, repo_type="dataset", revision=safe_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]],
-    tol: 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 a list of query timestamps (e.g. [-0.2, 0.4, 0.6, 0.8]) and loads the closest frames in the dataset.
+    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
@@ -156,7 +241,7 @@ def load_previous_and_future_frames(
       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.
-    - tol (float, optional): The tolerance level used to determine if a data point is close enough to the query
+    - 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.
 
@@ -194,11 +279,11 @@ def load_previous_and_future_frames(
 
         # TODO(rcadene): synchronize timestamps + interpolation if needed
 
-        is_pad = min_ > tol
+        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_} > {tol=}) inside episode range."
+            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."
         )
 
@@ -207,143 +292,94 @@ def load_previous_and_future_frames(
 
         # load frames modality
         item[key] = hf_dataset.select_columns(key)[data_ids][key]
-        item[key] = torch.stack(item[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 get_stats_einops_patterns(hf_dataset):
-    """These einops patterns will be used to aggregate batches and compute statistics.
-
-    Note: We assume the images of `hf_dataset` are in channel first format
+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.
 
-    dataloader = torch.utils.data.DataLoader(
-        hf_dataset,
-        num_workers=0,
-        batch_size=2,
-        shuffle=False,
-    )
-    batch = next(iter(dataloader))
+    Parameters:
+    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
 
-    stats_patterns = {}
-    for key, feats_type in hf_dataset.features.items():
-        # sanity check that tensors are not float64
-        assert batch[key].dtype != torch.float64
+    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": []}
 
-        if isinstance(feats_type, 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}"
+    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]
 
-            # 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(hf_dataset, batch_size=32, max_num_samples=None):
-    if max_num_samples is None:
-        max_num_samples = len(hf_dataset)
-
-    stats_patterns = get_stats_einops_patterns(hf_dataset)
-
-    # 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(hf_dataset, batch_size, seed):
-        generator = torch.Generator()
-        generator.manual_seed(seed)
-        dataloader = torch.utils.data.DataLoader(
-            hf_dataset,
-            num_workers=4,
-            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(hf_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(hf_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],
+    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)
 
-    return stats
+    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):
@@ -357,3 +393,29 @@ def cycle(iterable):
             yield next(iterator)
         except StopIteration:
             iterator = iter(iterable)
+
+
+def create_branch(repo_id, *, branch: str, repo_type: str | None = None):
+    """Create a branch on a existing Hugging Face repo. Delete the branch if it already
+    exists before creating it.
+    """
+    api = HfApi()
+
+    branches = api.list_repo_refs(repo_id, repo_type=repo_type).branches
+    refs = [branch.ref for branch in branches]
+    ref = f"refs/heads/{branch}"
+    if ref in refs:
+        api.delete_branch(repo_id, repo_type=repo_type, branch=branch)
+
+    api.create_branch(repo_id, repo_type=repo_type, branch=branch)
+
+
+def create_lerobot_dataset_card(tags: list | None = None, text: str | None = None) -> DatasetCard:
+    card = DatasetCard(DATASET_CARD_TEMPLATE)
+    card.data.task_categories = ["robotics"]
+    card.data.tags = ["LeRobot"]
+    if tags is not None:
+        card.data.tags += tags
+    if text is not None:
+        card.text += text
+    return card

lerobot/common/datasets/video_utils.py

@@ -0,0 +1,249 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+import subprocess
+import warnings
+from collections import OrderedDict
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, ClassVar
+
+import pyarrow as pa
+import torch
+import torchvision
+from datasets.features.features import register_feature
+
+
+def load_from_videos(
+    item: dict[str, torch.Tensor],
+    video_frame_keys: list[str],
+    videos_dir: Path,
+    tolerance_s: float,
+    backend: str = "pyav",
+):
+    """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, backend)
+            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, backend)
+            item[key] = frames[0]
+
+    return item
+
+
+def decode_video_frames_torchvision(
+    video_path: str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str = "pyav",
+    log_loaded_timestamps: bool = False,
+) -> torch.Tensor:
+    """Loads frames associated to the requested timestamps of a video
+
+    The backend can be either "pyav" (default) or "video_reader".
+    "video_reader" requires installing torchvision from source, see:
+    https://github.com/pytorch/vision/blob/main/torchvision/csrc/io/decoder/gpu/README.rst
+    (note that you need to compile against ffmpeg<4.3)
+
+    While both use cpu, "video_reader" is supposedly faster than "pyav" but requires additional setup.
+    For more info on video decoding, see `benchmark/video/README.md`
+
+    See torchvision doc for more info on these two backends:
+    https://pytorch.org/vision/0.18/index.html?highlight=backend#torchvision.set_video_backend
+
+    Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
+    the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
+    that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
+    and all subsequent frames until reaching the requested frame. The number of key frames in a video
+    can be adjusted during encoding to take into account decoding time and video size in bytes.
+    """
+    video_path = str(video_path)
+
+    # set backend
+    keyframes_only = False
+    torchvision.set_video_backend(backend)
+    if backend == "pyav":
+        keyframes_only = True  # pyav doesnt support accuracte seek
+
+    # set a video stream reader
+    # TODO(rcadene): also load audio stream at the same time
+    reader = torchvision.io.VideoReader(video_path, "video")
+
+    # set the first and last requested timestamps
+    # Note: previous timestamps are usually loaded, since we need to access the previous key frame
+    first_ts = 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
+
+    if backend == "pyav":
+        reader.container.close()
+
+    reader = None
+
+    query_ts = torch.tensor(timestamps)
+    loaded_ts = torch.tensor(loaded_ts)
+
+    # compute distances between each query timestamp and timestamps of all loaded frames
+    dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
+    min_, argmin_ = dist.min(1)
+
+    is_within_tol = min_ < tolerance_s
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
+        f"\nbackend: {backend}"
+    )
+
+    # get closest frames to the query timestamps
+    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
+    closest_ts = loaded_ts[argmin_]
+
+    if log_loaded_timestamps:
+        logging.info(f"{closest_ts=}")
+
+    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
+    closest_frames = closest_frames.type(torch.float32) / 255
+
+    assert len(timestamps) == len(closest_frames)
+    return closest_frames
+
+
+def encode_video_frames(
+    imgs_dir: Path,
+    video_path: Path,
+    fps: int,
+    vcodec: str = "libsvtav1",
+    pix_fmt: str = "yuv420p",
+    g: int | None = 2,
+    crf: int | None = 30,
+    fast_decode: int = 0,
+    log_level: str | None = "error",
+    overwrite: bool = False,
+) -> None:
+    """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
+    video_path = Path(video_path)
+    video_path.parent.mkdir(parents=True, exist_ok=True)
+
+    ffmpeg_args = OrderedDict(
+        [
+            ("-f", "image2"),
+            ("-r", str(fps)),
+            ("-i", str(imgs_dir / "frame_%06d.png")),
+            ("-vcodec", vcodec),
+            ("-pix_fmt", pix_fmt),
+        ]
+    )
+
+    if g is not None:
+        ffmpeg_args["-g"] = str(g)
+
+    if crf is not None:
+        ffmpeg_args["-crf"] = str(crf)
+
+    if fast_decode:
+        key = "-svtav1-params" if vcodec == "libsvtav1" else "-tune"
+        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
+        ffmpeg_args[key] = value
+
+    if log_level is not None:
+        ffmpeg_args["-loglevel"] = str(log_level)
+
+    ffmpeg_args = [item for pair in ffmpeg_args.items() for item in pair]
+    if overwrite:
+        ffmpeg_args.append("-y")
+
+    ffmpeg_cmd = ["ffmpeg"] + ffmpeg_args + [str(video_path)]
+    # redirect stdin to subprocess.DEVNULL to prevent reading random keyboard inputs from terminal
+    subprocess.run(ffmpeg_cmd, check=True, stdin=subprocess.DEVNULL)
+
+    if not video_path.exists():
+        raise OSError(
+            f"Video encoding did not work. File not found: {video_path}. "
+            f"Try running the command manually to debug: `{''.join(ffmpeg_cmd)}`"
+        )
+
+
+@dataclass
+class VideoFrame:
+    # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
+    """
+    Provides a type for a dataset containing video frames.
+
+    Example:
+
+    ```python
+    data_dict = [{"image": {"path": "videos/episode_0.mp4", "timestamp": 0.3}}]
+    features = {"image": VideoFrame()}
+    Dataset.from_dict(data_dict, features=Features(features))
+    ```
+    """
+
+    pa_type: ClassVar[Any] = pa.struct({"path": pa.string(), "timestamp": pa.float32()})
+    _type: str = field(default="VideoFrame", init=False, repr=False)
+
+    def __call__(self):
+        return self.pa_type
+
+
+with warnings.catch_warnings():
+    warnings.filterwarnings(
+        "ignore",
+        "'register_feature' is experimental and might be subject to breaking changes in the future.",
+        category=UserWarning,
+    )
+    # to make VideoFrame available in HuggingFace `datasets`
+    register_feature(VideoFrame, "VideoFrame")

lerobot/common/envs/factory.py

@@ -1,20 +1,34 @@
+#!/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, num_parallel_envs=0) -> gym.Env | gym.vector.SyncVectorEnv:
+def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv | None:
+    """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.
     """
-    Note: When `num_parallel_envs > 0`, this function returns a `SyncVectorEnv` which takes batched action as input and
-    returns batched observation, reward, terminated, truncated of `num_parallel_envs` items.
-    """
-    kwargs = {
-        "obs_type": "pixels_agent_pos",
-        "render_mode": "rgb_array",
-        "max_episode_steps": cfg.env.episode_length,
-        "visualization_width": 384,
-        "visualization_height": 384,
-    }
+    if n_envs is not None and n_envs < 1:
+        raise ValueError("`n_envs must be at least 1")
+
+    if cfg.env.name == "real_world":
+        return
 
     package_name = f"gym_{cfg.env.name}"
 
@@ -27,17 +41,18 @@ def make_env(cfg, num_parallel_envs=0) -> gym.Env | gym.vector.SyncVectorEnv:
         raise e
 
     gym_handle = f"{package_name}/{cfg.env.task}"
+    gym_kwgs = dict(cfg.env.get("gym", {}))
 
-    if num_parallel_envs == 0:
-        # non-batched version of the env that returns an observation of shape (c)
-        env = gym.make(gym_handle, disable_env_checker=True, **kwargs)
-    else:
-        # batched version of the env that returns an observation of shape (b, c)
-        env = gym.vector.SyncVectorEnv(
-            [
-                lambda: gym.make(gym_handle, disable_env_checker=True, **kwargs)
-                for _ in range(num_parallel_envs)
-            ]
-        )
+    if cfg.env.get("episode_length"):
+        gym_kwgs["max_episode_steps"] = cfg.env.episode_length
+
+    # 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, **gym_kwgs)
+            for _ in range(n_envs if n_envs is not None else cfg.eval.batch_size)
+        ]
+    )
 
     return env

lerobot/common/envs/utils.py

@@ -1,42 +1,62 @@
+#!/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(observation):
+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
-    obs = {}
+    return_observations = {}
+    if "pixels" in observations:
+        if isinstance(observations["pixels"], dict):
+            imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
+        else:
+            imgs = {"observation.image": observations["pixels"]}
 
-    if isinstance(observation["pixels"], dict):
-        imgs = {f"observation.images.{key}": img for key, img in observation["pixels"].items()}
-    else:
-        imgs = {"observation.image": observation["pixels"]}
+        for imgkey, img in imgs.items():
+            img = torch.from_numpy(img)
 
-    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 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=}"
 
-        # 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
 
-        # 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
 
-        obs[imgkey] = img
+    if "environment_state" in observations:
+        return_observations["observation.environment_state"] = torch.from_numpy(
+            observations["environment_state"]
+        ).float()
 
-    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing requirement for "agent_pos"
-    obs["observation.state"] = torch.from_numpy(observation["agent_pos"]).float()
-
-    return obs
-
-
-def postprocess_action(action):
-    action = action.to("cpu").numpy()
-    assert (
-        action.ndim == 2
-    ), "we assume dimensions are respectively the number of parallel envs, action dimensions"
-    return action
+    # 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

@@ -1,41 +1,103 @@
+#!/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.
+"""Borrowed from https://github.com/fyhMer/fowm/blob/main/src/logger.py
+
+# TODO(rcadene, alexander-soare): clean this file
+"""
+
 import logging
 import os
+import re
+from glob import glob
 from pathlib import Path
 
-from omegaconf import OmegaConf
+import torch
+from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
+from omegaconf import DictConfig, OmegaConf
 from termcolor import colored
+from torch.optim import Optimizer
+from torch.optim.lr_scheduler import LRScheduler
+
+from lerobot.common.policies.policy_protocol import Policy
+from lerobot.common.utils.utils import get_global_random_state, set_global_random_state
 
 
 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)]
+def cfg_to_group(cfg: DictConfig, return_list: bool = False) -> list[str] | str:
+    """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}",
     ]
     return lst if return_list else "-".join(lst)
 
 
-class Logger:
-    """Primary logger object. Logs either locally or using wandb."""
+def get_wandb_run_id_from_filesystem(checkpoint_dir: Path) -> str:
+    # Get the WandB run ID.
+    paths = glob(str(checkpoint_dir / "../wandb/latest-run/run-*"))
+    if len(paths) != 1:
+        raise RuntimeError("Couldn't get the previous WandB run ID for run resumption.")
+    match = re.search(r"run-([^\.]+).wandb", paths[0].split("/")[-1])
+    if match is None:
+        raise RuntimeError("Couldn't get the previous WandB run ID for run resumption.")
+    wandb_run_id = match.groups(0)[0]
+    return wandb_run_id
 
-    def __init__(self, log_dir, job_name, cfg):
-        self._log_dir = Path(log_dir)
-        self._log_dir.mkdir(parents=True, exist_ok=True)
-        self._job_name = job_name
-        self._model_dir = self._log_dir / "models"
-        self._buffer_dir = self._log_dir / "buffers"
-        self._save_model = cfg.save_model
-        self._disable_wandb_artifact = cfg.wandb.disable_artifact
-        self._save_buffer = cfg.save_buffer
-        self._group = cfg_to_group(cfg)
-        self._seed = cfg.seed
+
+class Logger:
+    """Primary logger object. Logs either locally or using wandb.
+
+    The logger creates the following directory structure:
+
+    provided_log_dir
+    ├── .hydra  # hydra's configuration cache
+    ├── checkpoints
+    │   ├── specific_checkpoint_name
+    │   │   ├── pretrained_model  # Hugging Face pretrained model directory
+    │   │   │   ├── ...
+    │   │   └── training_state.pth  # optimizer, scheduler, and random states + training step
+    |   ├── another_specific_checkpoint_name
+    │   │   ├── ...
+    |   ├── ...
+    │   └── last  # a softlink to the last logged checkpoint
+    """
+
+    pretrained_model_dir_name = "pretrained_model"
+    training_state_file_name = "training_state.pth"
+
+    def __init__(self, cfg: DictConfig, log_dir: str, wandb_job_name: str | None = None):
+        """
+        Args:
+            log_dir: The directory to save all logs and training outputs to.
+            job_name: The WandB job name.
+        """
         self._cfg = cfg
-        self._eval = []
+        self.log_dir = Path(log_dir)
+        self.log_dir.mkdir(parents=True, exist_ok=True)
+        self.checkpoints_dir = self.get_checkpoints_dir(log_dir)
+        self.last_checkpoint_dir = self.get_last_checkpoint_dir(log_dir)
+        self.last_pretrained_model_dir = self.get_last_pretrained_model_dir(log_dir)
+
+        # Set up WandB.
+        self._group = cfg_to_group(cfg)
         project = cfg.get("wandb", {}).get("project")
         entity = cfg.get("wandb", {}).get("entity")
         enable_wandb = cfg.get("wandb", {}).get("enable", False)
@@ -47,67 +109,138 @@ class Logger:
             os.environ["WANDB_SILENT"] = "true"
             import wandb
 
+            wandb_run_id = None
+            if cfg.resume:
+                wandb_run_id = get_wandb_run_id_from_filesystem(self.checkpoints_dir)
+
             wandb.init(
+                id=wandb_run_id,
                 project=project,
                 entity=entity,
-                name=job_name,
+                name=wandb_job_name,
                 notes=cfg.get("wandb", {}).get("notes"),
-                # group=self._group,
                 tags=cfg_to_group(cfg, return_list=True),
-                dir=self._log_dir,
+                dir=log_dir,
                 config=OmegaConf.to_container(cfg, resolve=True),
                 # TODO(rcadene): try set to True
                 save_code=False,
                 # TODO(rcadene): split train and eval, and run async eval with job_type="eval"
                 job_type="train_eval",
-                # TODO(rcadene): add resume option
-                resume=None,
+                resume="must" if cfg.resume else 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):
-        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 and not self._disable_wandb_artifact:
-                # note wandb artifact does not accept ":" in its name
-                artifact = self._wandb.Artifact(
-                    self._group.replace(":", "_") + "-" + str(self._seed) + "-" + str(identifier),
-                    type="model",
-                )
-                artifact.add_file(fp)
-                self._wandb.log_artifact(artifact)
+    @classmethod
+    def get_checkpoints_dir(cls, log_dir: str | Path) -> Path:
+        """Given the log directory, get the sub-directory in which checkpoints will be saved."""
+        return Path(log_dir) / "checkpoints"
 
-    def save_buffer(self, buffer, identifier):
-        self._buffer_dir.mkdir(parents=True, exist_ok=True)
-        fp = self._buffer_dir / f"{str(identifier)}.pkl"
-        buffer.save(fp)
-        if self._wandb:
-            artifact = self._wandb.Artifact(
-                self._group + "-" + str(self._seed) + "-" + str(identifier),
-                type="buffer",
-            )
-            artifact.add_file(fp)
+    @classmethod
+    def get_last_checkpoint_dir(cls, log_dir: str | Path) -> Path:
+        """Given the log directory, get the sub-directory in which the last checkpoint will be saved."""
+        return cls.get_checkpoints_dir(log_dir) / "last"
+
+    @classmethod
+    def get_last_pretrained_model_dir(cls, log_dir: str | Path) -> Path:
+        """
+        Given the log directory, get the sub-directory in which the last checkpoint's pretrained weights will
+        be saved.
+        """
+        return cls.get_last_checkpoint_dir(log_dir) / cls.pretrained_model_dir_name
+
+    def save_model(self, save_dir: Path, policy: Policy, wandb_artifact_name: str | None = None):
+        """Save the weights of the Policy model using PyTorchModelHubMixin.
+
+        The weights are saved in a folder called "pretrained_model" under the checkpoint directory.
+
+        Optionally also upload the model to WandB.
+        """
+        self.checkpoints_dir.mkdir(parents=True, exist_ok=True)
+        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._cfg.wandb.disable_artifact:
+            # note wandb artifact does not accept ":" or "/" in its name
+            artifact = self._wandb.Artifact(wandb_artifact_name, type="model")
+            artifact.add_file(save_dir / SAFETENSORS_SINGLE_FILE)
             self._wandb.log_artifact(artifact)
+        if self.last_checkpoint_dir.exists():
+            os.remove(self.last_checkpoint_dir)
 
-    def finish(self, agent, buffer):
-        if self._save_model:
-            self.save_model(agent, identifier="final")
-        if self._save_buffer:
-            self.save_buffer(buffer, identifier="buffer")
-        if self._wandb:
-            self._wandb.finish()
+    def save_training_state(
+        self,
+        save_dir: Path,
+        train_step: int,
+        optimizer: Optimizer,
+        scheduler: LRScheduler | None,
+    ):
+        """Checkpoint the global training_step, optimizer state, scheduler state, and random state.
+
+        All of these are saved as "training_state.pth" under the checkpoint directory.
+        """
+        training_state = {
+            "step": train_step,
+            "optimizer": optimizer.state_dict(),
+            **get_global_random_state(),
+        }
+        if scheduler is not None:
+            training_state["scheduler"] = scheduler.state_dict()
+        torch.save(training_state, save_dir / self.training_state_file_name)
+
+    def save_checkpont(
+        self,
+        train_step: int,
+        policy: Policy,
+        optimizer: Optimizer,
+        scheduler: LRScheduler | None,
+        identifier: str,
+    ):
+        """Checkpoint the model weights and the training state."""
+        checkpoint_dir = self.checkpoints_dir / str(identifier)
+        wandb_artifact_name = (
+            None
+            if self._wandb is None
+            else f"{self._group.replace(':', '_').replace('/', '_')}-{self._cfg.seed}-{identifier}"
+        )
+        self.save_model(
+            checkpoint_dir / self.pretrained_model_dir_name, policy, wandb_artifact_name=wandb_artifact_name
+        )
+        self.save_training_state(checkpoint_dir, train_step, optimizer, scheduler)
+        os.symlink(checkpoint_dir.absolute(), self.last_checkpoint_dir)
+
+    def load_last_training_state(self, optimizer: Optimizer, scheduler: LRScheduler | None) -> int:
+        """
+        Given the last checkpoint in the logging directory, load the optimizer state, scheduler state, and
+        random state, and return the global training step.
+        """
+        training_state = torch.load(self.last_checkpoint_dir / self.training_state_file_name)
+        optimizer.load_state_dict(training_state["optimizer"])
+        if scheduler is not None:
+            scheduler.load_state_dict(training_state["scheduler"])
+        elif "scheduler" in training_state:
+            raise ValueError(
+                "The checkpoint contains a scheduler state_dict, but no LRScheduler was provided."
+            )
+        # Small hack to get the expected keys: use `get_global_random_state`.
+        set_global_random_state({k: training_state[k] for k in get_global_random_state()})
+        return training_state["step"]
 
     def log_dict(self, d, step, mode="train"):
         assert mode in {"train", "eval"}
+        # TODO(alexander-soare): Add local text log.
         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")
+        assert self._wandb is not None
+        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

@@ -1,8 +1,23 @@
+#!/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 ActionChunkingTransformerConfig:
+class ACTConfig:
     """Configuration class for the Action Chunking Transformers policy.
 
     Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
@@ -10,6 +25,16 @@ class ActionChunkingTransformerConfig:
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
     Those are: `input_shapes` and 'output_shapes`.
 
+    Notes on the inputs and outputs:
+        - Either:
+            - At least one key starting with "observation.image is required as an input.
+              AND/OR
+            - The key "observation.environment_state" is required as input.
+        - If there are multiple keys beginning with "observation.images." they are treated as multiple camera
+          views. Right now we only support all images having the same shape.
+        - May optionally work without an "observation.state" key for the proprioceptive robot state.
+        - "action" is required as an output key.
+
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
@@ -18,27 +43,28 @@ class ActionChunkingTransformerConfig:
             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 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.
-        normalize_input_modes: A dictionary with key represents the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two availables
-            modes are "mean_std" which substracts the mean and divide by the standard
-            deviation and "min_max" which rescale in a [-1, 1] range.
-        unnormalize_output_modes: Similar dictionary as `normalize_input_modes`, but to unormalize in original scale.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
-        use_pretrained_backbone: Whether the backbone should be initialized with pretrained weights from
-            torchvision.
+        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.
-        d_model: The transformer blocks' main hidden dimension.
+        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.
@@ -50,8 +76,10 @@ class ActionChunkingTransformerConfig:
             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.
-        use_temporal_aggregation: Whether to blend the actions of multiple policy invocations for any given
-            environment step.
+        temporal_ensemble_coeff: Coefficient for the exponential weighting scheme to apply for temporal
+            ensembling. Defaults to None which means temporal ensembling is not used. `n_action_steps` must be
+            1 when using this feature, as inference needs to happen at every step to form an ensemble. For
+            more information on how ensembling works, please see `ACTTemporalEnsembler`.
         dropout: Dropout to use in the transformer layers (see code for details).
         kl_weight: The weight to use for the KL-divergence component of the loss if the variational objective
             is enabled. Loss is then calculated as: `reconstruction_loss + kl_weight * kld_loss`.
@@ -62,26 +90,26 @@ class ActionChunkingTransformerConfig:
     chunk_size: int = 100
     n_action_steps: int = 100
 
-    input_shapes: dict[str, list[str]] = field(
+    input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
             "observation.images.top": [3, 480, 640],
             "observation.state": [14],
         }
     )
-    output_shapes: dict[str, list[str]] = field(
+    output_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
             "action": [14],
         }
     )
 
     # Normalization / Unnormalization
-    normalize_input_modes: dict[str, str] = field(
+    input_normalization_modes: dict[str, str] = field(
         default_factory=lambda: {
-            "observation.image": "mean_std",
+            "observation.images.top": "mean_std",
             "observation.state": "mean_std",
         }
     )
-    unnormalize_output_modes: dict[str, str] = field(
+    output_normalization_modes: dict[str, str] = field(
         default_factory=lambda: {
             "action": "mean_std",
         }
@@ -90,15 +118,18 @@ class ActionChunkingTransformerConfig:
     # Architecture.
     # Vision backbone.
     vision_backbone: str = "resnet18"
-    use_pretrained_backbone: bool = True
+    pretrained_backbone_weights: str | None = "ResNet18_Weights.IMAGENET1K_V1"
     replace_final_stride_with_dilation: int = False
     # Transformer layers.
     pre_norm: bool = False
-    d_model: int = 512
+    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
@@ -106,29 +137,24 @@ class ActionChunkingTransformerConfig:
     n_vae_encoder_layers: int = 4
 
     # Inference.
-    use_temporal_aggregation: bool = False
+    # Note: the value used in ACT when temporal ensembling is enabled is 0.01.
+    temporal_ensemble_coeff: float | None = None
 
     # Training and loss computation.
     dropout: float = 0.1
     kl_weight: float = 10.0
 
-    # ---
-    # TODO(alexander-soare): Remove these from the policy config.
-    batch_size: int = 8
-    lr: float = 1e-5
-    lr_backbone: float = 1e-5
-    weight_decay: float = 1e-4
-    grad_clip_norm: float = 10
-    utd: int = 1
-
     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.use_temporal_aggregation:
-            raise NotImplementedError("Temporal aggregation is not yet implemented.")
+        if self.temporal_ensemble_coeff is not None and self.n_action_steps > 1:
+            raise NotImplementedError(
+                "`n_action_steps` must be 1 when using temporal ensembling. This is "
+                "because the policy needs to be queried every step to compute the ensembled action."
+            )
         if self.n_action_steps > self.chunk_size:
             raise ValueError(
                 f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
@@ -138,10 +164,8 @@ class ActionChunkingTransformerConfig:
             raise ValueError(
                 f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
             )
-        # Check that there is only one image.
-        # TODO(alexander-soare): generalize this to multiple images.
         if (
-            sum(k.startswith("observation.images.") for k in self.input_shapes) != 1
-            or "observation.images.top" not in self.input_shapes
+            not any(k.startswith("observation.image") for k in self.input_shapes)
+            and "observation.environment_state" not in self.input_shapes
         ):
-            raise ValueError('For now, only "observation.images.top" is accepted for an image input.')
+            raise ValueError("You must provide at least one image or the environment state among the inputs.")

lerobot/common/policies/act/modeling_act.py

@@ -1,3 +1,18 @@
+#!/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).
@@ -5,7 +20,6 @@ The majority of changes here involve removing unused code, unifying naming, and
 """
 
 import math
-import time
 from collections import deque
 from itertools import chain
 from typing import Callable
@@ -15,18 +29,231 @@ 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 ActionChunkingTransformerConfig
+from lerobot.common.policies.act.configuration_act import ACTConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 
 
-class ActionChunkingTransformerPolicy(nn.Module):
+class ACTPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "act"],
+):
     """
     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")]
+
+        if config.temporal_ensemble_coeff is not None:
+            self.temporal_ensembler = ACTTemporalEnsembler(config.temporal_ensemble_coeff, config.chunk_size)
+
+        self.reset()
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        if self.config.temporal_ensemble_coeff is not None:
+            self.temporal_ensembler.reset()
+        else:
+            self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    @torch.no_grad
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method wraps `select_actions` in order to return one action at a time for execution in the
+        environment. It works by managing the actions in a queue and only calling `select_actions` when the
+        queue is empty.
+        """
+        self.eval()
+
+        batch = self.normalize_inputs(batch)
+        if len(self.expected_image_keys) > 0:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+
+        # If we are doing temporal ensembling, do online updates where we keep track of the number of actions
+        # we are ensembling over.
+        if self.config.temporal_ensemble_coeff is not None:
+            actions = self.model(batch)[0]  # (batch_size, chunk_size, action_dim)
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+            action = self.temporal_ensembler.update(actions)
+            return action
+
+        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
+        # querying the policy.
+        if len(self._action_queue) == 0:
+            actions = self.model(batch)[0][:, : self.config.n_action_steps]
+
+            # TODO(rcadene): make _forward return output dictionary?
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        """Run the batch through the model and compute the loss for training or validation."""
+        batch = self.normalize_inputs(batch)
+        if len(self.expected_image_keys) > 0:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            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 ACTTemporalEnsembler:
+    def __init__(self, temporal_ensemble_coeff: float, chunk_size: int) -> None:
+        """Temporal ensembling as described in Algorithm 2 of https://arxiv.org/abs/2304.13705.
+
+        The weights are calculated as wᵢ = exp(-temporal_ensemble_coeff * i) where w₀ is the oldest action.
+        They are then normalized to sum to 1 by dividing by Σwᵢ. Here's some intuition around how the
+        coefficient works:
+            - Setting it to 0 uniformly weighs all actions.
+            - Setting it positive gives more weight to older actions.
+            - Setting it negative gives more weight to newer actions.
+        NOTE: The default value for `temporal_ensemble_coeff` used by the original ACT work is 0.01. This
+        results in older actions being weighed more highly than newer actions (the experiments documented in
+        https://github.com/huggingface/lerobot/pull/319 hint at why highly weighing new actions might be
+        detrimental: doing so aggressively may diminish the benefits of action chunking).
+
+        Here we use an online method for computing the average rather than caching a history of actions in
+        order to compute the average offline. For a simple 1D sequence it looks something like:
+
+        ```
+        import torch
+
+        seq = torch.linspace(8, 8.5, 100)
+        print(seq)
+
+        m = 0.01
+        exp_weights = torch.exp(-m * torch.arange(len(seq)))
+        print(exp_weights)
+
+        # Calculate offline
+        avg = (exp_weights * seq).sum() / exp_weights.sum()
+        print("offline", avg)
+
+        # Calculate online
+        for i, item in enumerate(seq):
+            if i == 0:
+                avg = item
+                continue
+            avg *= exp_weights[:i].sum()
+            avg += item * exp_weights[i]
+            avg /= exp_weights[:i+1].sum()
+        print("online", avg)
+        ```
+        """
+        self.chunk_size = chunk_size
+        self.ensemble_weights = torch.exp(-temporal_ensemble_coeff * torch.arange(chunk_size))
+        self.ensemble_weights_cumsum = torch.cumsum(self.ensemble_weights, dim=0)
+        self.reset()
+
+    def reset(self):
+        """Resets the online computation variables."""
+        self.ensembled_actions = None
+        # (chunk_size,) count of how many actions are in the ensemble for each time step in the sequence.
+        self.ensembled_actions_count = None
+
+    def update(self, actions: Tensor) -> Tensor:
+        """
+        Takes a (batch, chunk_size, action_dim) sequence of actions, update the temporal ensemble for all
+        time steps, and pop/return the next batch of actions in the sequence.
+        """
+        self.ensemble_weights = self.ensemble_weights.to(device=actions.device)
+        self.ensemble_weights_cumsum = self.ensemble_weights_cumsum.to(device=actions.device)
+        if self.ensembled_actions is None:
+            # Initializes `self._ensembled_action` to the sequence of actions predicted during the first
+            # time step of the episode.
+            self.ensembled_actions = actions.clone()
+            # Note: The last dimension is unsqueeze to make sure we can broadcast properly for tensor
+            # operations later.
+            self.ensembled_actions_count = torch.ones(
+                (self.chunk_size, 1), dtype=torch.long, device=self.ensembled_actions.device
+            )
+        else:
+            # self.ensembled_actions will have shape (batch_size, chunk_size - 1, action_dim). Compute
+            # the online update for those entries.
+            self.ensembled_actions *= self.ensemble_weights_cumsum[self.ensembled_actions_count - 1]
+            self.ensembled_actions += actions[:, :-1] * self.ensemble_weights[self.ensembled_actions_count]
+            self.ensembled_actions /= self.ensemble_weights_cumsum[self.ensembled_actions_count]
+            self.ensembled_actions_count = torch.clamp(self.ensembled_actions_count + 1, max=self.chunk_size)
+            # The last action, which has no prior online average, needs to get concatenated onto the end.
+            self.ensembled_actions = torch.cat([self.ensembled_actions, actions[:, -1:]], dim=1)
+            self.ensembled_actions_count = torch.cat(
+                [self.ensembled_actions_count, torch.ones_like(self.ensembled_actions_count[-1:])]
+            )
+        # "Consume" the first action.
+        action, self.ensembled_actions, self.ensembled_actions_count = (
+            self.ensembled_actions[:, 0],
+            self.ensembled_actions[:, 1:],
+            self.ensembled_actions_count[1:],
+        )
+        return action
+
+
+class ACT(nn.Module):
+    """Action Chunking Transformer: The underlying neural network for ACTPolicy.
 
     Note: In this code we use the terms `vae_encoder`, 'encoder', `decoder`. The meanings are as follows.
         - The `vae_encoder` is, as per the literature around variational auto-encoders (VAE), the part of the
@@ -53,104 +280,94 @@ class ActionChunkingTransformerPolicy(nn.Module):
               │ encoder │ │     │ │Transf.│             │
               │         │ │     │ │encoder│             │
               └───▲─────┘ │     │ │       │             │
-                  │       │     │ └───▲───┘             │
-                  │       │     │     │                 │
-                inputs    └─────┼─────┘                 │
-                                │                       │
+                  │       │     │ └▲──▲─▲─┘             │
+                  │       │     │  │  │ │               │
+                inputs    └─────┼──┘  │ image emb.      │
+                                │    state emb.         │
                                 └───────────────────────┘
     """
 
-    name = "act"
-
-    def __init__(self, cfg: ActionChunkingTransformerConfig | None = None, dataset_stats=None):
-        """
-        Args:
-            cfg: Policy configuration class instance or None, in which case the default instantiation of the
-                 configuration class is used.
-        """
+    def __init__(self, config: ACTConfig):
         super().__init__()
-        if cfg is None:
-            cfg = ActionChunkingTransformerConfig()
-        self.cfg = cfg
-        self.normalize_inputs = Normalize(cfg.input_shapes, cfg.normalize_input_modes, dataset_stats)
-        self.unnormalize_outputs = Unnormalize(cfg.output_shapes, cfg.unnormalize_output_modes, dataset_stats)
-
-        # BERT style VAE encoder with input [cls, *joint_space_configuration, *action_sequence].
+        self.config = config
+        # BERT style VAE encoder with input tokens [cls, robot_state, *action_sequence].
         # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
-        if self.cfg.use_vae:
-            self.vae_encoder = _TransformerEncoder(cfg)
-            self.vae_encoder_cls_embed = nn.Embedding(1, cfg.d_model)
+        self.use_robot_state = "observation.state" in config.input_shapes
+        self.use_images = any(k.startswith("observation.image") for k in config.input_shapes)
+        self.use_env_state = "observation.environment_state" in config.input_shapes
+        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(
-                cfg.input_shapes["observation.state"][0], cfg.d_model
-            )
+            if self.use_robot_state:
+                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(
-                cfg.input_shapes["observation.state"][0], cfg.d_model
+                config.output_shapes["action"][0], config.dim_model
             )
-            self.latent_dim = cfg.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(cfg.d_model, self.latent_dim * 2)
-            # Fixed sinusoidal positional embedding the whole input to the VAE encoder. Unsqueeze for batch
+            self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, config.latent_dim * 2)
+            # Fixed sinusoidal positional embedding for the input to the VAE encoder. Unsqueeze for batch
             # dimension.
+            num_input_token_encoder = 1 + config.chunk_size
+            if self.use_robot_state:
+                num_input_token_encoder += 1
             self.register_buffer(
                 "vae_encoder_pos_enc",
-                _create_sinusoidal_position_embedding(1 + 1 + cfg.chunk_size, cfg.d_model).unsqueeze(0),
+                create_sinusoidal_pos_embedding(num_input_token_encoder, config.dim_model).unsqueeze(0),
             )
 
         # Backbone for image feature extraction.
-        backbone_model = getattr(torchvision.models, cfg.vision_backbone)(
-            replace_stride_with_dilation=[False, False, cfg.replace_final_stride_with_dilation],
-            pretrained=cfg.use_pretrained_backbone,
-            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"})
+        if self.use_images:
+            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 = _TransformerEncoder(cfg)
-        self.decoder = _TransformerDecoder(cfg)
+        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(cfg.input_shapes["observation.state"][0], cfg.d_model)
-        self.encoder_latent_input_proj = nn.Linear(self.latent_dim, cfg.d_model)
-        self.encoder_img_feat_input_proj = nn.Conv2d(
-            backbone_model.fc.in_features, cfg.d_model, kernel_size=1
-        )
+        # [latent, (robot_state), (env_state), (image_feature_map_pixels)].
+        if self.use_robot_state:
+            self.encoder_robot_state_input_proj = nn.Linear(
+                config.input_shapes["observation.state"][0], config.dim_model
+            )
+        if self.use_env_state:
+            self.encoder_env_state_input_proj = nn.Linear(
+                config.input_shapes["observation.environment_state"][0], config.dim_model
+            )
+        self.encoder_latent_input_proj = nn.Linear(config.latent_dim, config.dim_model)
+        if self.use_images:
+            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, cfg.d_model)
-        self.encoder_cam_feat_pos_embed = _SinusoidalPositionEmbedding2D(cfg.d_model // 2)
+        n_1d_tokens = 1  # for the latent
+        if self.use_robot_state:
+            n_1d_tokens += 1
+        if self.use_env_state:
+            n_1d_tokens += 1
+        self.encoder_1d_feature_pos_embed = nn.Embedding(n_1d_tokens, config.dim_model)
+        if self.use_images:
+            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(cfg.chunk_size, cfg.d_model)
+        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(cfg.d_model, cfg.output_shapes["action"][0])
+        self.action_head = nn.Linear(config.dim_model, config.output_shapes["action"][0])
 
         self._reset_parameters()
-        self._create_optimizer()
-
-    def _create_optimizer(self):
-        optimizer_params_dicts = [
-            {
-                "params": [
-                    p for n, p in self.named_parameters() if not n.startswith("backbone") and p.requires_grad
-                ]
-            },
-            {
-                "params": [
-                    p for n, p in self.named_parameters() if n.startswith("backbone") and p.requires_grad
-                ],
-                "lr": self.cfg.lr_backbone,
-            },
-        ]
-        self.optimizer = torch.optim.AdamW(
-            optimizer_params_dicts, lr=self.cfg.lr, weight_decay=self.cfg.weight_decay
-        )
 
     def _reset_parameters(self):
         """Xavier-uniform initialization of the transformer parameters as in the original code."""
@@ -158,109 +375,17 @@ class ActionChunkingTransformerPolicy(nn.Module):
             if p.dim() > 1:
                 nn.init.xavier_uniform_(p)
 
-    def reset(self):
-        """This should be called whenever the environment is reset."""
-        if self.cfg.n_action_steps is not None:
-            self._action_queue = deque([], maxlen=self.cfg.n_action_steps)
-
-    @torch.no_grad
-    def select_action(self, batch: dict[str, Tensor], **_) -> Tensor:
-        """Select a single action given environment observations.
-
-        This method wraps `select_actions` in order to return one action at a time for execution in the
-        environment. It works by managing the actions in a queue and only calling `select_actions` when the
-        queue is empty.
-        """
-        self.eval()
-
-        batch = self.normalize_inputs(batch)
-
-        if len(self._action_queue) == 0:
-            # `_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.
-            actions = self._forward(batch)[0][: self.cfg.n_action_steps]
-
-            # TODO(rcadene): make _forward return output dictionary?
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
-            self._action_queue.extend(actions.transpose(0, 1))
-        return self._action_queue.popleft()
-
-    def forward(self, batch, **_) -> dict[str, Tensor]:
-        """Run the batch through the model and compute the loss for training or validation."""
-        actions_hat, (mu_hat, log_sigma_x2_hat) = self._forward(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}
-        if self.cfg.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
-            loss_dict["loss"] = l1_loss + mean_kld * self.cfg.kl_weight
-        else:
-            loss_dict["loss"] = l1_loss
-
-        return loss_dict
-
-    def update(self, batch, **_) -> dict:
-        """Run the model in train mode, compute the loss, and do an optimization step."""
-        start_time = time.time()
-        self.train()
-
-        batch = self.normalize_inputs(batch)
-
-        loss_dict = self.forward(batch)
-        # TODO(rcadene): self.unnormalize_outputs(out_dict)
-        loss = loss_dict["loss"]
-        loss.backward()
-
-        grad_norm = torch.nn.utils.clip_grad_norm_(
-            self.parameters(), self.cfg.grad_clip_norm, error_if_nonfinite=False
-        )
-
-        self.optimizer.step()
-        self.optimizer.zero_grad()
-
-        info = {
-            "loss": loss.item(),
-            "grad_norm": float(grad_norm),
-            "lr": self.cfg.lr,
-            "update_s": time.time() - start_time,
-        }
-
-        return info
-
-    def _stack_images(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        """Stacks all the images in a batch and puts them in a new key: "observation.images".
-
-        This function expects `batch` to have (at least):
-        {
-            "observation.state": (B, state_dim) batch of robot states.
-            "observation.images.{name}": (B, C, H, W) tensor of images.
-        }
-        """
-        # Stack images in the order dictated by input_shapes.
-        batch["observation.images"] = torch.stack(
-            [batch[k] for k in self.cfg.input_shapes if k.startswith("observation.images.")],
-            dim=-4,
-        )
-
-    def _forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, tuple[Tensor, Tensor] | tuple[None, None]]:
+    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.state" (optional): (B, state_dim) batch of robot states.
+
             "observation.images": (B, n_cameras, C, H, W) batch of images.
+                AND/OR
+            "observation.environment_state": (B, env_dim) batch of environment states.
+
             "action" (optional, only if training with VAE): (B, chunk_size, action dim) batch of actions.
         }
 
@@ -269,94 +394,119 @@ class ActionChunkingTransformerPolicy(nn.Module):
             Tuple containing the latent PDF's parameters (mean, log(σ²)) both as (B, L) tensors where L is the
             latent dimension.
         """
-        if self.cfg.use_vae and self.training:
+        if self.config.use_vae and self.training:
             assert (
                 "action" in batch
             ), "actions must be provided when using the variational objective in training mode."
 
-        self._stack_images(batch)
-
-        batch_size = batch["observation.state"].shape[0]
+        batch_size = (
+            batch["observation.images"]
+            if "observation.images" in batch
+            else batch["observation.environment_state"]
+        ).shape[0]
 
         # Prepare the latent for input to the transformer encoder.
-        if self.cfg.use_vae and "action" in batch:
+        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)
+            if self.use_robot_state:
+                robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
+                robot_state_embed = robot_state_embed.unsqueeze(1)  # (B, 1, D)
             action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
-            vae_encoder_input = torch.cat([cls_embed, robot_state_embed, action_embed], axis=1)  # (B, S+2, D)
+
+            if self.use_robot_state:
+                vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
+            else:
+                vae_encoder_input = [cls_embed, action_embed]
+            vae_encoder_input = torch.cat(vae_encoder_input, axis=1)
 
             # Prepare fixed positional embedding.
             # Note: detach() shouldn't be necessary but leaving it the same as the original code just in case.
             pos_embed = self.vae_encoder_pos_enc.clone().detach()  # (1, S+2, D)
 
+            # Prepare key padding mask for the transformer encoder. We have 1 or 2 extra tokens at the start of the
+            # sequence depending whether we use the input states or not (cls and robot state)
+            # False means not a padding token.
+            cls_joint_is_pad = torch.full(
+                (batch_size, 2 if self.use_robot_state else 1),
+                False,
+                device=batch["observation.state"].device,
+            )
+            key_padding_mask = torch.cat(
+                [cls_joint_is_pad, batch["action_is_pad"]], axis=1
+            )  # (bs, seq+1 or 2)
+
             # Forward pass through VAE encoder to get the latent PDF parameters.
             cls_token_out = self.vae_encoder(
-                vae_encoder_input.permute(1, 0, 2), pos_embed=pos_embed.permute(1, 0, 2)
+                vae_encoder_input.permute(1, 0, 2),
+                pos_embed=pos_embed.permute(1, 0, 2),
+                key_padding_mask=key_padding_mask,
             )[0]  # select the class token, with shape (B, D)
             latent_pdf_params = self.vae_encoder_latent_output_proj(cls_token_out)
-            mu = latent_pdf_params[:, : self.latent_dim]
+            mu = latent_pdf_params[:, : self.config.latent_dim]
             # This is 2log(sigma). Done this way to match the original implementation.
-            log_sigma_x2 = latent_pdf_params[:, self.latent_dim :]
+            log_sigma_x2 = latent_pdf_params[:, self.config.latent_dim :]
 
             # Sample the latent with the reparameterization trick.
             latent_sample = mu + log_sigma_x2.div(2).exp() * torch.randn_like(mu)
         else:
             # When not using the VAE encoder, we set the latent to be all zeros.
             mu = log_sigma_x2 = None
-            latent_sample = torch.zeros([batch_size, self.latent_dim], dtype=torch.float32).to(
+            # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
+            latent_sample = torch.zeros([batch_size, self.config.latent_dim], dtype=torch.float32).to(
                 batch["observation.state"].device
             )
 
-        # Prepare all other transformer encoder inputs.
+        # Prepare transformer encoder inputs.
+        encoder_in_tokens = [self.encoder_latent_input_proj(latent_sample)]
+        encoder_in_pos_embed = list(self.encoder_1d_feature_pos_embed.weight.unsqueeze(1))
+        # Robot state token.
+        if self.use_robot_state:
+            encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch["observation.state"]))
+        # Environment state token.
+        if self.use_env_state:
+            encoder_in_tokens.append(
+                self.encoder_env_state_input_proj(batch["observation.environment_state"])
+            )
+
         # 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=3)
-        cam_pos_embed = torch.cat(all_cam_pos_embeds, axis=3)
+        if self.use_images:
+            all_cam_features = []
+            all_cam_pos_embeds = []
 
-        # Get positional embeddings for robot state and latent.
-        robot_state_embed = self.encoder_robot_state_input_proj(batch["observation.state"])
-        latent_embed = self.encoder_latent_input_proj(latent_sample)
+            for cam_index in range(batch["observation.images"].shape[-4]):
+                cam_features = self.backbone(batch["observation.images"][:, cam_index])["feature_map"]
+                # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use
+                # buffer
+                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,
+            # and move to (sequence, batch, dim).
+            all_cam_features = torch.cat(all_cam_features, axis=-1)
+            encoder_in_tokens.extend(einops.rearrange(all_cam_features, "b c h w -> (h w) b c"))
+            all_cam_pos_embeds = torch.cat(all_cam_pos_embeds, axis=-1)
+            encoder_in_pos_embed.extend(einops.rearrange(all_cam_pos_embeds, "b c h w -> (h w) 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),
-                encoder_in.flatten(2).permute(2, 0, 1),
-            ]
-        )
-        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,
-        )
+        # Stack all tokens along the sequence dimension.
+        encoder_in_tokens = torch.stack(encoder_in_tokens, axis=0)
+        encoder_in_pos_embed = torch.stack(encoder_in_pos_embed, axis=0)
 
         # Forward pass through the transformer modules.
-        encoder_out = self.encoder(encoder_in, pos_embed=pos_embed)
+        encoder_out = self.encoder(encoder_in_tokens, pos_embed=encoder_in_pos_embed)
+        # TODO(rcadene, alexander-soare): remove call to `device` ; precompute and use buffer
         decoder_in = torch.zeros(
-            (self.cfg.chunk_size, batch_size, self.cfg.d_model),
-            dtype=pos_embed.dtype,
-            device=pos_embed.device,
+            (self.config.chunk_size, batch_size, self.config.dim_model),
+            dtype=encoder_in_pos_embed.dtype,
+            device=encoder_in_pos_embed.device,
         )
         decoder_out = self.decoder(
             decoder_in,
             encoder_out,
-            encoder_pos_embed=pos_embed,
+            encoder_pos_embed=encoder_in_pos_embed,
             decoder_pos_embed=self.decoder_pos_embed.weight.unsqueeze(1),
         )
 
@@ -367,53 +517,49 @@ class ActionChunkingTransformerPolicy(nn.Module):
 
         return actions, (mu, log_sigma_x2)
 
-    def save(self, fp):
-        torch.save(self.state_dict(), fp)
 
-    def load(self, fp):
-        d = torch.load(fp)
-        self.load_state_dict(d)
-
-
-class _TransformerEncoder(nn.Module):
+class ACTEncoder(nn.Module):
     """Convenience module for running multiple encoder layers, maybe followed by normalization."""
 
-    def __init__(self, cfg: ActionChunkingTransformerConfig):
+    def __init__(self, config: ACTConfig):
         super().__init__()
-        self.layers = nn.ModuleList([_TransformerEncoderLayer(cfg) for _ in range(cfg.n_encoder_layers)])
-        self.norm = nn.LayerNorm(cfg.d_model) if cfg.pre_norm else nn.Identity()
+        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:
+    def forward(
+        self, x: Tensor, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None
+    ) -> Tensor:
         for layer in self.layers:
-            x = layer(x, pos_embed=pos_embed)
+            x = layer(x, pos_embed=pos_embed, key_padding_mask=key_padding_mask)
         x = self.norm(x)
         return x
 
 
-class _TransformerEncoderLayer(nn.Module):
-    def __init__(self, cfg: ActionChunkingTransformerConfig):
+class ACTEncoderLayer(nn.Module):
+    def __init__(self, config: ACTConfig):
         super().__init__()
-        self.self_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
+        self.self_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
 
         # Feed forward layers.
-        self.linear1 = nn.Linear(cfg.d_model, cfg.dim_feedforward)
-        self.dropout = nn.Dropout(cfg.dropout)
-        self.linear2 = nn.Linear(cfg.dim_feedforward, cfg.d_model)
+        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(cfg.d_model)
-        self.norm2 = nn.LayerNorm(cfg.d_model)
-        self.dropout1 = nn.Dropout(cfg.dropout)
-        self.dropout2 = nn.Dropout(cfg.dropout)
+        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(cfg.feedforward_activation)
-        self.pre_norm = cfg.pre_norm
+        self.activation = get_activation_fn(config.feedforward_activation)
+        self.pre_norm = config.pre_norm
 
-    def forward(self, x, pos_embed: Tensor | None = None) -> Tensor:
+    def forward(self, x, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None) -> Tensor:
         skip = x
         if self.pre_norm:
             x = self.norm1(x)
         q = k = x if pos_embed is None else x + pos_embed
-        x = self.self_attn(q, k, value=x)[0]  # select just the output, not the attention weights
+        x = self.self_attn(q, k, value=x, key_padding_mask=key_padding_mask)
+        x = x[0]  # note: [0] to select just the output, not the attention weights
         x = skip + self.dropout1(x)
         if self.pre_norm:
             skip = x
@@ -428,12 +574,12 @@ class _TransformerEncoderLayer(nn.Module):
         return x
 
 
-class _TransformerDecoder(nn.Module):
-    def __init__(self, cfg: ActionChunkingTransformerConfig):
+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([_TransformerDecoderLayer(cfg) for _ in range(cfg.n_decoder_layers)])
-        self.norm = nn.LayerNorm(cfg.d_model)
+        self.layers = nn.ModuleList([ACTDecoderLayer(config) for _ in range(config.n_decoder_layers)])
+        self.norm = nn.LayerNorm(config.dim_model)
 
     def forward(
         self,
@@ -451,26 +597,26 @@ class _TransformerDecoder(nn.Module):
         return x
 
 
-class _TransformerDecoderLayer(nn.Module):
-    def __init__(self, cfg: ActionChunkingTransformerConfig):
+class ACTDecoderLayer(nn.Module):
+    def __init__(self, config: ACTConfig):
         super().__init__()
-        self.self_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
-        self.multihead_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
+        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(cfg.d_model, cfg.dim_feedforward)
-        self.dropout = nn.Dropout(cfg.dropout)
-        self.linear2 = nn.Linear(cfg.dim_feedforward, cfg.d_model)
+        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(cfg.d_model)
-        self.norm2 = nn.LayerNorm(cfg.d_model)
-        self.norm3 = nn.LayerNorm(cfg.d_model)
-        self.dropout1 = nn.Dropout(cfg.dropout)
-        self.dropout2 = nn.Dropout(cfg.dropout)
-        self.dropout3 = nn.Dropout(cfg.dropout)
+        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(cfg.feedforward_activation)
-        self.pre_norm = cfg.pre_norm
+        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
@@ -523,7 +669,7 @@ class _TransformerDecoderLayer(nn.Module):
         return x
 
 
-def _create_sinusoidal_position_embedding(num_positions: int, dimension: int) -> Tensor:
+def create_sinusoidal_pos_embedding(num_positions: int, dimension: int) -> Tensor:
     """1D sinusoidal positional embeddings as in Attention is All You Need.
 
     Args:
@@ -541,7 +687,7 @@ def _create_sinusoidal_position_embedding(num_positions: int, dimension: int) ->
     return torch.from_numpy(sinusoid_table).float()
 
 
-class _SinusoidalPositionEmbedding2D(nn.Module):
+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
@@ -595,7 +741,7 @@ class _SinusoidalPositionEmbedding2D(nn.Module):
         return pos_embed
 
 
-def _get_activation_fn(activation: str) -> Callable:
+def get_activation_fn(activation: str) -> Callable:
     """Return an activation function given a string."""
     if activation == "relu":
         return F.relu

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -1,42 +1,69 @@
+#!/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 Diffusion Policy.
+    """Configuration class for DiffusionPolicy.
 
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
     Those are: `input_shapes` and `output_shapes`.
 
+    Notes on the inputs and outputs:
+        - "observation.state" is required as an input key.
+        - Either:
+            - At least one key starting with "observation.image is required as an input.
+              AND/OR
+            - The key "observation.environment_state" is required as input.
+        - If there are multiple keys beginning with "observation.image" they are treated as multiple camera
+          views. Right now we only support all images having the same shape.
+        - "action" is required as an output key.
+
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
         horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
         n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
             See `DiffusionPolicy.select_action` for more details.
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-            The key represents the input data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "observation.image" refers to an input from
-            a camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
-            Importantly, 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.
-        normalize_input_modes: A dictionary with key represents the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two availables
-            modes are "mean_std" which substracts the mean and divide by the standard
-            deviation and "min_max" which rescale in a [-1, 1] range.
-        unnormalize_output_modes: Similar dictionary as `normalize_input_modes`, but to unormalize in original scale.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
             within the image size. If None, no cropping is done.
         crop_is_random: Whether the crop should be random at training time (it's always a center crop in eval
             mode).
-        use_pretrained_backbone: Whether the backbone should be initialized with pretrained weights from
-            torchvision.
+        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.
@@ -50,6 +77,7 @@ class DiffusionConfig:
         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.
@@ -63,6 +91,9 @@ class DiffusionConfig:
         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.
@@ -70,37 +101,33 @@ class DiffusionConfig:
     horizon: int = 16
     n_action_steps: int = 8
 
-    input_shapes: dict[str, list[str]] = field(
+    input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
             "observation.image": [3, 96, 96],
             "observation.state": [2],
         }
     )
-    output_shapes: dict[str, list[str]] = field(
+    output_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
             "action": [2],
         }
     )
 
     # Normalization / Unnormalization
-    normalize_input_modes: dict[str, str] = field(
+    input_normalization_modes: dict[str, str] = field(
         default_factory=lambda: {
             "observation.image": "mean_std",
             "observation.state": "min_max",
         }
     )
-    unnormalize_output_modes: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": "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
-    use_pretrained_backbone: bool = False
+    pretrained_backbone_weights: str | None = None
     use_group_norm: bool = True
     spatial_softmax_num_keypoints: int = 32
     # Unet.
@@ -110,6 +137,7 @@ class DiffusionConfig:
     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
@@ -121,23 +149,8 @@ class DiffusionConfig:
     # Inference
     num_inference_steps: int | None = None
 
-    # ---
-    # TODO(alexander-soare): Remove these from the policy config.
-    batch_size: int = 64
-    grad_clip_norm: int = 10
-    lr: float = 1.0e-4
-    lr_scheduler: str = "cosine"
-    lr_warmup_steps: int = 500
-    adam_betas: tuple[float, float] = (0.95, 0.999)
-    adam_eps: float = 1.0e-8
-    adam_weight_decay: float = 1.0e-6
-    utd: int = 1
-    use_ema: bool = True
-    ema_update_after_step: int = 0
-    ema_min_alpha: float = 0.0
-    ema_max_alpha: float = 0.9999
-    ema_inv_gamma: float = 1.0
-    ema_power: float = 0.75
+    # Loss computation
+    do_mask_loss_for_padding: bool = False
 
     def __post_init__(self):
         """Input validation (not exhaustive)."""
@@ -145,17 +158,41 @@ class DiffusionConfig:
             raise ValueError(
                 f"`vision_backbone` must be one of the ResNet variants. Got {self.vision_backbone}."
             )
-        if (
-            self.crop_shape[0] > self.input_shapes["observation.image"][1]
-            or self.crop_shape[1] > self.input_shapes["observation.image"][2]
-        ):
-            raise ValueError(
-                f'`crop_shape` should fit within `input_shapes["observation.image"]`. Got {self.crop_shape} '
-                f'for `crop_shape` and {self.input_shapes["observation.image"]} for '
-                '`input_shapes["observation.image"]`.'
-            )
+
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
+
+        if len(image_keys) == 0 and "observation.environment_state" not in self.input_shapes:
+            raise ValueError("You must provide at least one image or the environment state among the inputs.")
+
+        if len(image_keys) > 0:
+            if self.crop_shape is not None:
+                for image_key in 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}]`."
+                        )
+            # Check that all input images have the same shape.
+            first_image_key = next(iter(image_keys))
+            for image_key in image_keys:
+                if self.input_shapes[image_key] != self.input_shapes[first_image_key]:
+                    raise ValueError(
+                        f"`input_shapes[{image_key}]` does not match `input_shapes[{first_image_key}]`, but we "
+                        "expect all image shapes to match."
+                    )
+
         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/modeling_diffusion.py

@@ -1,29 +1,38 @@
+#!/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 Robomimic for SpatialSoftmax.
   - Remove reliance on diffusers for DDPMScheduler and LR scheduler.
-  - Move EMA out of policy.
-  - Consolidate _DiffusionUnetImagePolicy into DiffusionPolicy.
-  - One more pass on comments and documentation.
 """
 
-import copy
-import logging
 import math
-import time
 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.optimization import get_scheduler
+from diffusers.schedulers.scheduling_ddim import DDIMScheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
-from robomimic.models.base_nets import SpatialSoftmax
+from huggingface_hub import PyTorchModelHubMixin
 from torch import Tensor, nn
-from torch.nn.modules.batchnorm import _BatchNorm
 
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
@@ -34,7 +43,13 @@ from lerobot.common.policies.utils import (
 )
 
 
-class DiffusionPolicy(nn.Module):
+class DiffusionPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "diffusion-policy"],
+):
     """
     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).
@@ -43,66 +58,54 @@ class DiffusionPolicy(nn.Module):
     name = "diffusion"
 
     def __init__(
-        self, cfg: DiffusionConfig | None, lr_scheduler_num_training_steps: int = 0, dataset_stats=None
+        self,
+        config: DiffusionConfig | None = None,
+        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.
+            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__()
-        # TODO(alexander-soare): LR scheduler will be removed.
-        assert lr_scheduler_num_training_steps > 0
-        if cfg is None:
-            cfg = DiffusionConfig()
-        self.cfg = cfg
-        self.normalize_inputs = Normalize(cfg.input_shapes, cfg.normalize_input_modes, dataset_stats)
-        self.unnormalize_outputs = Unnormalize(cfg.output_shapes, cfg.unnormalize_output_modes, dataset_stats)
+        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 = _DiffusionUnetImagePolicy(cfg)
+        self.diffusion = DiffusionModel(config)
 
-        # TODO(alexander-soare): This should probably be managed outside of the policy class.
-        self.ema_diffusion = None
-        self.ema = None
-        if self.cfg.use_ema:
-            self.ema_diffusion = copy.deepcopy(self.diffusion)
-            self.ema = _EMA(cfg, model=self.ema_diffusion)
+        self.expected_image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        self.use_env_state = "observation.environment_state" in config.input_shapes
 
-        # TODO(alexander-soare): Move optimizer out of policy.
-        self.optimizer = torch.optim.Adam(
-            self.diffusion.parameters(), cfg.lr, cfg.adam_betas, cfg.adam_eps, cfg.adam_weight_decay
-        )
-
-        # TODO(alexander-soare): Move LR scheduler out of policy.
-        # TODO(rcadene): modify lr scheduler so that it doesn't 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=lr_scheduler_num_training_steps,
-            # pytorch assumes stepping LRScheduler every epoch
-            # however huggingface diffusers steps it every batch
-            last_epoch=self.global_step - 1,
-        )
+        self.reset()
 
     def reset(self):
-        """
-        Clear observation and action queues. Should be called on `env.reset()`
-        """
+        """Clear observation and action queues. Should be called on `env.reset()`"""
         self._queues = {
-            "observation.image": deque(maxlen=self.cfg.n_obs_steps),
-            "observation.state": deque(maxlen=self.cfg.n_obs_steps),
-            "action": deque(maxlen=self.cfg.n_action_steps),
+            "observation.state": deque(maxlen=self.config.n_obs_steps),
+            "action": deque(maxlen=self.config.n_action_steps),
         }
+        if len(self.expected_image_keys) > 0:
+            self._queues["observation.images"] = deque(maxlen=self.config.n_obs_steps)
+        if self.use_env_state:
+            self._queues["observation.environment_state"] = deque(maxlen=self.config.n_obs_steps)
 
     @torch.no_grad
-    def select_action(self, batch: dict[str, Tensor], **_) -> Tensor:
+    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
@@ -114,33 +117,26 @@ class DiffusionPolicy(nn.Module):
         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    |
+            |timestep            | n-o+1 | n-o+2 | ..... | n     | ..... | n+a-1 | n+a   | ..... | n-o+h |
+            |observation is used | YES   | YES   | YES   | YES   | NO    | NO    | NO    | NO    | NO    |
             |action is generated | YES   | YES   | YES   | YES   | YES   | YES   | YES   | YES   | YES   |
             |action is used      | NO    | NO    | NO    | YES   | YES   | YES   | NO    | NO    | NO    |
             ----------------------------------------------------------------------------------------------
-        Note that this means we require: `n_action_steps < horizon - n_obs_steps + 1`. Also, note that
+        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.
-
-        Note: this method uses the ema model weights if self.training == False, otherwise the non-ema model
-        weights.
         """
-        assert "observation.image" in batch
-        assert "observation.state" in batch
-        assert len(batch) == 2
-
         batch = self.normalize_inputs(batch)
-
+        if len(self.expected_image_keys) > 0:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+        # Note: It's important that this happens after stacking the images into a single key.
         self._queues = populate_queues(self._queues, batch)
 
         if len(self._queues["action"]) == 0:
             # stack n latest observations from the queue
-            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
-            if not self.training and self.ema_diffusion is not None:
-                actions = self.ema_diffusion.generate_actions(batch)
-            else:
-                actions = self.diffusion.generate_actions(batch)
+            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"]
@@ -150,86 +146,65 @@ class DiffusionPolicy(nn.Module):
         action = self._queues["action"].popleft()
         return action
 
-    def forward(self, batch: dict[str, Tensor], **_) -> dict[str, Tensor]:
+    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)
+        if len(self.expected_image_keys) > 0:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+        batch = self.normalize_targets(batch)
         loss = self.diffusion.compute_loss(batch)
         return {"loss": loss}
 
-    def update(self, batch: dict[str, Tensor], **_) -> dict:
-        """Run the model in train mode, compute the loss, and do an optimization step."""
-        start_time = time.time()
 
-        self.diffusion.train()
-
-        batch = self.normalize_inputs(batch)
-
-        loss = self.forward(batch)["loss"]
-        loss.backward()
-
-        # TODO(rcadene): self.unnormalize_outputs(out_dict)
-
-        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],
-            "update_s": time.time() - start_time,
-        }
-
-        return info
-
-    def save(self, fp):
-        torch.save(self.state_dict(), fp)
-
-    def load(self, fp):
-        d = torch.load(fp)
-        missing_keys, unexpected_keys = self.load_state_dict(d, strict=False)
-        if len(missing_keys) > 0:
-            assert all(k.startswith("ema_diffusion.") for k in missing_keys)
-            logging.warning(
-                "DiffusionPolicy.load expected ema parameters in loaded state dict but none were found."
-            )
-        assert len(unexpected_keys) == 0
+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 _DiffusionUnetImagePolicy(nn.Module):
-    def __init__(self, cfg: DiffusionConfig):
+class DiffusionModel(nn.Module):
+    def __init__(self, config: DiffusionConfig):
         super().__init__()
-        self.cfg = cfg
+        self.config = config
 
-        self.rgb_encoder = _RgbEncoder(cfg)
-        self.unet = _ConditionalUnet1D(
-            cfg,
-            global_cond_dim=(cfg.output_shapes["action"][0] + self.rgb_encoder.feature_dim) * cfg.n_obs_steps,
+        # Build observation encoders (depending on which observations are provided).
+        global_cond_dim = config.input_shapes["observation.state"][0]
+        num_images = len([k for k in config.input_shapes if k.startswith("observation.image")])
+        self._use_images = False
+        self._use_env_state = False
+        if num_images > 0:
+            self._use_images = True
+            self.rgb_encoder = DiffusionRgbEncoder(config)
+            global_cond_dim += self.rgb_encoder.feature_dim * num_images
+        if "observation.environment_state" in config.input_shapes:
+            self._use_env_state = True
+            global_cond_dim += config.input_shapes["observation.environment_state"][0]
+
+        self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
+
+        self.noise_scheduler = _make_noise_scheduler(
+            config.noise_scheduler_type,
+            num_train_timesteps=config.num_train_timesteps,
+            beta_start=config.beta_start,
+            beta_end=config.beta_end,
+            beta_schedule=config.beta_schedule,
+            clip_sample=config.clip_sample,
+            clip_sample_range=config.clip_sample_range,
+            prediction_type=config.prediction_type,
         )
 
-        self.noise_scheduler = DDPMScheduler(
-            num_train_timesteps=cfg.num_train_timesteps,
-            beta_start=cfg.beta_start,
-            beta_end=cfg.beta_end,
-            beta_schedule=cfg.beta_schedule,
-            variance_type="fixed_small",
-            clip_sample=cfg.clip_sample,
-            clip_sample_range=cfg.clip_sample_range,
-            prediction_type=cfg.prediction_type,
-        )
-
-        if cfg.num_inference_steps is None:
+        if config.num_inference_steps is None:
             self.num_inference_steps = self.noise_scheduler.config.num_train_timesteps
         else:
-            self.num_inference_steps = cfg.num_inference_steps
+            self.num_inference_steps = config.num_inference_steps
 
     # ========= inference  ============
     def conditional_sample(
@@ -240,7 +215,7 @@ class _DiffusionUnetImagePolicy(nn.Module):
 
         # Sample prior.
         sample = torch.randn(
-            size=(batch_size, self.cfg.horizon, self.cfg.output_shapes["action"][0]),
+            size=(batch_size, self.config.horizon, self.config.output_shapes["action"][0]),
             dtype=dtype,
             device=device,
             generator=generator,
@@ -260,33 +235,51 @@ class _DiffusionUnetImagePolicy(nn.Module):
 
         return sample
 
+    def _prepare_global_conditioning(self, batch: dict[str, Tensor]) -> Tensor:
+        """Encode image features and concatenate them all together along with the state vector."""
+        batch_size, n_obs_steps = batch["observation.state"].shape[:2]
+        global_cond_feats = [batch["observation.state"]]
+        # Extract image feature (first combine batch, sequence, and camera index dims).
+        if self._use_images:
+            img_features = self.rgb_encoder(
+                einops.rearrange(batch["observation.images"], "b s n ... -> (b s n) ...")
+            )
+            # Separate batch dim and sequence dim back out. The camera index dim gets absorbed into the
+            # feature dim (effectively concatenating the camera features).
+            img_features = einops.rearrange(
+                img_features, "(b s n) ... -> b s (n ...)", b=batch_size, s=n_obs_steps
+            )
+            global_cond_feats.append(img_features)
+
+        if self._use_env_state:
+            global_cond_feats.append(batch["observation.environment_state"])
+
+        # Concatenate features then flatten to (B, global_cond_dim).
+        return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)
+
     def generate_actions(self, batch: dict[str, Tensor]) -> Tensor:
         """
-        This function expects `batch` to have (at least):
+        This function expects `batch` to have:
         {
             "observation.state": (B, n_obs_steps, state_dim)
-            "observation.image": (B, n_obs_steps, C, H, W)
+
+            "observation.images": (B, n_obs_steps, num_cameras, C, H, W)
+                AND/OR
+            "observation.environment_state": (B, environment_dim)
         }
         """
-        assert set(batch).issuperset({"observation.state", "observation.image"})
         batch_size, n_obs_steps = batch["observation.state"].shape[:2]
-        assert n_obs_steps == self.cfg.n_obs_steps
+        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)
+        # Encode image features and concatenate them all together along with the state vector.
+        global_cond = self._prepare_global_conditioning(batch)  # (B, global_cond_dim)
 
         # run sampling
-        sample = self.conditional_sample(batch_size, global_cond=global_cond)
+        actions = self.conditional_sample(batch_size, global_cond=global_cond)
 
-        # `horizon` steps worth of actions (from the first observation).
-        actions = sample[..., : self.cfg.output_shapes["action"][0]]
         # Extract `n_action_steps` steps worth of actions (from the current observation).
         start = n_obs_steps - 1
-        end = start + self.cfg.n_action_steps
+        end = start + self.config.n_action_steps
         actions = actions[:, start:end]
 
         return actions
@@ -296,28 +289,28 @@ class _DiffusionUnetImagePolicy(nn.Module):
         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)
+
+            "observation.images": (B, n_obs_steps, num_cameras, C, H, W)
+                AND/OR
+            "observation.environment_state": (B, environment_dim)
+
             "action": (B, horizon, action_dim)
             "action_is_pad": (B, horizon)
         }
         """
         # Input validation.
-        assert set(batch).issuperset({"observation.state", "observation.image", "action", "action_is_pad"})
-        batch_size, n_obs_steps = batch["observation.state"].shape[:2]
+        assert set(batch).issuperset({"observation.state", "action", "action_is_pad"})
+        assert "observation.images" in batch or "observation.environment_state" in batch
+        n_obs_steps = batch["observation.state"].shape[1]
         horizon = batch["action"].shape[1]
-        assert horizon == self.cfg.horizon
-        assert n_obs_steps == self.cfg.n_obs_steps
+        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"]
+        # Encode image features and concatenate them all together along with the state vector.
+        global_cond = self._prepare_global_conditioning(batch)  # (B, global_cond_dim)
 
         # Forward diffusion.
+        trajectory = batch["action"]
         # Sample noise to add to the trajectory.
         eps = torch.randn(trajectory.shape, device=trajectory.device)
         # Sample a random noising timestep for each item in the batch.
@@ -335,52 +328,128 @@ class _DiffusionUnetImagePolicy(nn.Module):
 
         # Compute the loss.
         # The target is either the original trajectory, or the noise.
-        if self.cfg.prediction_type == "epsilon":
+        if self.config.prediction_type == "epsilon":
             target = eps
-        elif self.cfg.prediction_type == "sample":
+        elif self.config.prediction_type == "sample":
             target = batch["action"]
         else:
-            raise ValueError(f"Unsupported prediction type {self.cfg.prediction_type}")
+            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 "action_is_pad" in batch:
+        if self.config.do_mask_loss_for_padding:
+            if "action_is_pad" not in batch:
+                raise ValueError(
+                    "You need to provide 'action_is_pad' in the batch when "
+                    f"{self.config.do_mask_loss_for_padding=}."
+                )
             in_episode_bound = ~batch["action_is_pad"]
             loss = loss * in_episode_bound.unsqueeze(-1)
 
         return loss.mean()
 
 
-class _RgbEncoder(nn.Module):
+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, cfg: DiffusionConfig):
+    def __init__(self, config: DiffusionConfig):
         super().__init__()
         # Set up optional preprocessing.
-        if cfg.crop_shape is not None:
+        if config.crop_shape is not None:
             self.do_crop = True
             # Always use center crop for eval
-            self.center_crop = torchvision.transforms.CenterCrop(cfg.crop_shape)
-            if cfg.crop_is_random:
-                self.maybe_random_crop = torchvision.transforms.RandomCrop(cfg.crop_shape)
+            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, cfg.vision_backbone)(
-            pretrained=cfg.use_pretrained_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 cfg.use_group_norm:
-            if cfg.use_pretrained_backbone:
+        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!"
                 )
@@ -392,13 +461,22 @@ class _RgbEncoder(nn.Module):
 
         # 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` if it is provided, otherwise it should use the
+        # height and width from `config.input_shapes`.
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        # Note: we have a check in the config class to make sure all images have the same shape.
+        image_key = image_keys[0]
+        dummy_input_h_w = (
+            config.crop_shape if config.crop_shape is not None else config.input_shapes[image_key][1:]
+        )
+        dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *dummy_input_h_w))
         with torch.inference_mode():
-            feat_map_shape = tuple(
-                self.backbone(torch.zeros(size=(1, *cfg.input_shapes["observation.image"]))).shape[1:]
-            )
-        self.pool = SpatialSoftmax(feat_map_shape, num_kp=cfg.spatial_softmax_num_keypoints)
-        self.feature_dim = cfg.spatial_softmax_num_keypoints * 2
-        self.out = nn.Linear(cfg.spatial_softmax_num_keypoints * 2, self.feature_dim)
+            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:
@@ -455,7 +533,7 @@ def _replace_submodules(
     return root_module
 
 
-class _SinusoidalPosEmb(nn.Module):
+class DiffusionSinusoidalPosEmb(nn.Module):
     """1D sinusoidal positional embeddings as in Attention is All You Need."""
 
     def __init__(self, dim: int):
@@ -472,7 +550,7 @@ class _SinusoidalPosEmb(nn.Module):
         return emb
 
 
-class _Conv1dBlock(nn.Module):
+class DiffusionConv1dBlock(nn.Module):
     """Conv1d --> GroupNorm --> Mish"""
 
     def __init__(self, inp_channels, out_channels, kernel_size, n_groups=8):
@@ -488,40 +566,40 @@ class _Conv1dBlock(nn.Module):
         return self.block(x)
 
 
-class _ConditionalUnet1D(nn.Module):
+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, cfg: DiffusionConfig, global_cond_dim: int):
+    def __init__(self, config: DiffusionConfig, global_cond_dim: int):
         super().__init__()
 
-        self.cfg = cfg
+        self.config = config
 
         # Encoder for the diffusion timestep.
         self.diffusion_step_encoder = nn.Sequential(
-            _SinusoidalPosEmb(cfg.diffusion_step_embed_dim),
-            nn.Linear(cfg.diffusion_step_embed_dim, cfg.diffusion_step_embed_dim * 4),
+            DiffusionSinusoidalPosEmb(config.diffusion_step_embed_dim),
+            nn.Linear(config.diffusion_step_embed_dim, config.diffusion_step_embed_dim * 4),
             nn.Mish(),
-            nn.Linear(cfg.diffusion_step_embed_dim * 4, cfg.diffusion_step_embed_dim),
+            nn.Linear(config.diffusion_step_embed_dim * 4, config.diffusion_step_embed_dim),
         )
 
         # The FiLM conditioning dimension.
-        cond_dim = cfg.diffusion_step_embed_dim + global_cond_dim
+        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 = [(cfg.output_shapes["action"][0], cfg.down_dims[0])] + list(
-            zip(cfg.down_dims[:-1], cfg.down_dims[1:], strict=True)
+        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": cfg.kernel_size,
-            "n_groups": cfg.n_groups,
-            "use_film_scale_modulation": cfg.use_film_scale_modulation,
+            "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):
@@ -529,8 +607,8 @@ class _ConditionalUnet1D(nn.Module):
             self.down_modules.append(
                 nn.ModuleList(
                     [
-                        _ConditionalResidualBlock1D(dim_in, dim_out, **common_res_block_kwargs),
-                        _ConditionalResidualBlock1D(dim_out, dim_out, **common_res_block_kwargs),
+                        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(),
                     ]
@@ -540,8 +618,12 @@ class _ConditionalUnet1D(nn.Module):
         # Processing in the middle of the auto-encoder.
         self.mid_modules = nn.ModuleList(
             [
-                _ConditionalResidualBlock1D(cfg.down_dims[-1], cfg.down_dims[-1], **common_res_block_kwargs),
-                _ConditionalResidualBlock1D(cfg.down_dims[-1], cfg.down_dims[-1], **common_res_block_kwargs),
+                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
+                ),
             ]
         )
 
@@ -553,8 +635,8 @@ class _ConditionalUnet1D(nn.Module):
                 nn.ModuleList(
                     [
                         # dim_in * 2, because it takes the encoder's skip connection as well
-                        _ConditionalResidualBlock1D(dim_in * 2, dim_out, **common_res_block_kwargs),
-                        _ConditionalResidualBlock1D(dim_out, dim_out, **common_res_block_kwargs),
+                        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(),
                     ]
@@ -562,8 +644,8 @@ class _ConditionalUnet1D(nn.Module):
             )
 
         self.final_conv = nn.Sequential(
-            _Conv1dBlock(cfg.down_dims[0], cfg.down_dims[0], kernel_size=cfg.kernel_size),
-            nn.Conv1d(cfg.down_dims[0], cfg.output_shapes["action"][0], 1),
+            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:
@@ -611,7 +693,7 @@ class _ConditionalUnet1D(nn.Module):
         return x
 
 
-class _ConditionalResidualBlock1D(nn.Module):
+class DiffusionConditionalResidualBlock1d(nn.Module):
     """ResNet style 1D convolutional block with FiLM modulation for conditioning."""
 
     def __init__(
@@ -630,13 +712,13 @@ class _ConditionalResidualBlock1D(nn.Module):
         self.use_film_scale_modulation = use_film_scale_modulation
         self.out_channels = out_channels
 
-        self.conv1 = _Conv1dBlock(in_channels, out_channels, kernel_size, n_groups=n_groups)
+        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 = _Conv1dBlock(out_channels, out_channels, kernel_size, n_groups=n_groups)
+        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 = (
@@ -667,67 +749,3 @@ class _ConditionalResidualBlock1D(nn.Module):
         out = self.conv2(out)
         out = out + self.residual_conv(x)
         return out
-
-
-class _EMA:
-    """
-    Exponential Moving Average of models weights
-    """
-
-    def __init__(self, cfg: DiffusionConfig, model: nn.Module):
-        """
-        @crowsonkb's notes on EMA Warmup:
-            If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are good values for models you plan
-            to train for a million or more steps (reaches decay factor 0.999 at 31.6K steps, 0.9999 at 1M steps),
-            gamma=1, power=3/4 for models you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999
-            at 215.4k steps).
-        Args:
-            inv_gamma (float): Inverse multiplicative factor of EMA warmup. Default: 1.
-            power (float): Exponential factor of EMA warmup. Default: 2/3.
-            min_alpha (float): The minimum EMA decay rate. Default: 0.
-        """
-
-        self.averaged_model = model
-        self.averaged_model.eval()
-        self.averaged_model.requires_grad_(False)
-
-        self.update_after_step = cfg.ema_update_after_step
-        self.inv_gamma = cfg.ema_inv_gamma
-        self.power = cfg.ema_power
-        self.min_alpha = cfg.ema_min_alpha
-        self.max_alpha = cfg.ema_max_alpha
-
-        self.alpha = 0.0
-        self.optimization_step = 0
-
-    def get_decay(self, optimization_step):
-        """
-        Compute the decay factor for the exponential moving average.
-        """
-        step = max(0, optimization_step - self.update_after_step - 1)
-        value = 1 - (1 + step / self.inv_gamma) ** -self.power
-
-        if step <= 0:
-            return 0.0
-
-        return max(self.min_alpha, min(value, self.max_alpha))
-
-    @torch.no_grad()
-    def step(self, new_model):
-        self.alpha = self.get_decay(self.optimization_step)
-
-        for module, ema_module in zip(new_model.modules(), self.averaged_model.modules(), strict=True):
-            # Iterate over immediate parameters only.
-            for param, ema_param in zip(
-                module.parameters(recurse=False), ema_module.parameters(recurse=False), strict=True
-            ):
-                if isinstance(param, dict):
-                    raise RuntimeError("Dict parameter not supported")
-                if isinstance(module, _BatchNorm) or not param.requires_grad:
-                    # Copy BatchNorm parameters, and non-trainable parameters directly.
-                    ema_param.copy_(param.to(dtype=ema_param.dtype).data)
-                else:
-                    ema_param.mul_(self.alpha)
-                    ema_param.add_(param.data.to(dtype=ema_param.dtype), alpha=1 - self.alpha)
-
-        self.optimization_step += 1

lerobot/common/policies/factory.py

@@ -1,18 +1,42 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 import inspect
+import logging
 
 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)
-    assert set(hydra_cfg.policy).issuperset(
-        expected_kwargs
-    ), f"Hydra config is missing arguments: {set(expected_kwargs).difference(hydra_cfg.policy)}"
+    if not set(hydra_cfg.policy).issuperset(expected_kwargs):
+        logging.warning(
+            f"Hydra config is missing arguments: {set(expected_kwargs).difference(hydra_cfg.policy)}"
+        )
+
+    # OmegaConf.to_container returns lists where sequences are found, but our dataclasses use tuples to avoid
+    # issues with mutable defaults. This filter changes all lists to tuples.
+    def list_to_tuple(item):
+        return tuple(item) if isinstance(item, list) else item
+
     policy_cfg = policy_cfg_class(
         **{
-            k: v
+            k: list_to_tuple(v)
             for k, v in OmegaConf.to_container(hydra_cfg.policy, resolve=True).items()
             if k in expected_kwargs
         }
@@ -20,42 +44,68 @@ def _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg):
     return policy_cfg
 
 
-def make_policy(hydra_cfg: DictConfig, dataset_stats=None):
-    if hydra_cfg.policy.name == "tdmpc":
-        from lerobot.common.policies.tdmpc.policy import TDMPCPolicy
+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
 
-        policy = TDMPCPolicy(
-            hydra_cfg.policy,
-            n_obs_steps=hydra_cfg.n_obs_steps,
-            n_action_steps=hydra_cfg.n_action_steps,
-            device=hydra_cfg.device,
-        )
-    elif hydra_cfg.policy.name == "diffusion":
+        return TDMPCPolicy, TDMPCConfig
+    elif name == "diffusion":
         from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
         from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 
-        policy_cfg = _policy_cfg_from_hydra_cfg(DiffusionConfig, hydra_cfg)
-        policy = DiffusionPolicy(policy_cfg, hydra_cfg.offline_steps, dataset_stats)
-        policy.to(get_safe_torch_device(hydra_cfg.device))
-    elif hydra_cfg.policy.name == "act":
-        from lerobot.common.policies.act.configuration_act import ActionChunkingTransformerConfig
-        from lerobot.common.policies.act.modeling_act import ActionChunkingTransformerPolicy
+        return DiffusionPolicy, DiffusionConfig
+    elif name == "act":
+        from lerobot.common.policies.act.configuration_act import ACTConfig
+        from lerobot.common.policies.act.modeling_act import ACTPolicy
 
-        policy_cfg = _policy_cfg_from_hydra_cfg(ActionChunkingTransformerConfig, hydra_cfg)
-        policy = ActionChunkingTransformerPolicy(policy_cfg, dataset_stats)
-        policy.to(get_safe_torch_device(hydra_cfg.device))
+        return ACTPolicy, ACTConfig
+    elif name == "vqbet":
+        from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
+        from lerobot.common.policies.vqbet.modeling_vqbet import VQBeTPolicy
+
+        return VQBeTPolicy, VQBeTConfig
     else:
-        raise ValueError(hydra_cfg.policy.name)
+        raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
-    if hydra_cfg.policy.pretrained_model_path:
-        # TODO(rcadene): hack for old pretrained models from fowm
-        if hydra_cfg.policy.name == "tdmpc" and "fowm" in hydra_cfg.policy.pretrained_model_path:
-            if "offline" in hydra_cfg.policy.pretrained_model_path:
-                policy.step[0] = 25000
-            elif "final" in hydra_cfg.policy.pretrained_model_path:
-                policy.step[0] = 100000
-            else:
-                raise NotImplementedError()
-        policy.load(hydra_cfg.policy.pretrained_model_path)
+
+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(
+            "Exactly one of `pretrained_policy_name_or_path` and `dataset_stats` must 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

@@ -1,27 +1,36 @@
+#!/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
+from torch import Tensor, nn
 
 
-def create_stats_buffers(shapes, modes, stats=None):
+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.
+    Create buffers per modality (e.g. "observation.image", "action") containing their mean, std, min, max
+    statistics.
 
-    Parameters:
-        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": substract 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
-            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.
+    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.
+        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 = {}
 
@@ -63,72 +72,83 @@ def create_stats_buffers(shapes, modes, stats=None):
             )
 
         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"]
-                buffer["std"].data = stats[key]["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"]
-                buffer["max"].data = stats[key]["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 the input data (e.g. "observation.image") for more stable and faster convergence during training.
+    """Normalizes data (e.g. "observation.image") for more stable and faster convergence during training."""
 
-    Parameters:
-        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": substract 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
-            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.
-    """
-
-    def __init__(self, shapes, modes, stats=None):
+    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):
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        batch = dict(batch)  # shallow copy avoids mutating the input batch
         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(), "`mean` is infinity. You forgot to initialize with `stats` as argument, or called `policy.load_state_dict`."
-                assert not torch.isinf(
-                    std
-                ).any(), "`std` is infinity. You forgot to initialize with `stats` as argument, or called `policy.load_state_dict`."
+                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(), "`min` is infinity. You forgot to initialize with `stats` as argument, or called `policy.load_state_dict`."
-                assert not torch.isinf(
-                    max
-                ).any(), "`max` is infinity. You forgot to initialize with `stats` as argument, or called `policy.load_state_dict`."
+                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)
+                batch[key] = (batch[key] - min) / (max - min + 1e-8)
                 # normalize to [-1, 1]
                 batch[key] = batch[key] * 2 - 1
             else:
@@ -138,23 +158,34 @@ class Normalize(nn.Module):
 
 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.
-
-    Parameters:
-        shapes (dict): A dictionary where keys are output modalities (e.g. "action") and values are their shapes (e.g. [10]).
-            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. "action") and values are their unnormalization modes among:
-            - "mean_std": multiply by standard deviation and add mean
-            - "min_max": go from [-1, 1] range to original range.
-        stats (dict, optional): A dictionary where keys are output modalities (e.g. "action") and values are dictionaries of statistic types and their values
-            (e.g. `{"max": torch.tensor(1)}, "min": torch.tensor(0)}`). 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
-            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.
+    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, modes, stats=None):
+    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
@@ -166,29 +197,22 @@ class Unnormalize(nn.Module):
 
     # TODO(rcadene): should we remove torch.no_grad?
     @torch.no_grad
-    def forward(self, batch):
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        batch = dict(batch)  # shallow copy avoids mutating the input batch
         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(), "`mean` is infinity. You forgot to initialize with `stats` as argument, or called `policy.load_state_dict`."
-                assert not torch.isinf(
-                    std
-                ).any(), "`std` is infinity. You forgot to initialize with `stats` as argument, or called `policy.load_state_dict`."
+                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(), "`min` is infinity. You forgot to initialize with `stats` as argument, or called `policy.load_state_dict`."
-                assert not torch.isinf(
-                    max
-                ).any(), "`max` is infinity. You forgot to initialize with `stats` as argument, or called `policy.load_state_dict`."
+                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:

lerobot/common/policies/policy_protocol.py

@@ -1,3 +1,18 @@
+#!/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
@@ -14,10 +29,21 @@ from torch import Tensor
 
 @runtime_checkable
 class Policy(Protocol):
-    """The required interface for implementing a policy."""
+    """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.
 
@@ -27,19 +53,23 @@ class Policy(Protocol):
     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 maybe other information.
+        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]):
+    def select_action(self, batch: dict[str, Tensor]) -> 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.
         """
 
-    def update(self, batch):
-        """Does compute_loss then an optimization step.
 
-        TODO(alexander-soare): We will move the optimization step back into the training loop, so this will
-        disappear.
+@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/configuration_tdmpc.py

@@ -0,0 +1,196 @@
+#!/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_ratio`.
+
+    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.
+        n_action_steps: Number of action steps to take from the plan given by model predictive control. This
+            is an alternative to using action repeats. If this is set to more than 1, then we require
+            `n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
+            approach of using multiple steps from the plan is not in the original implementation.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range. 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
+    n_action_steps: int = 1
+
+    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__} handles at most one image for now. Got image keys {image_keys}."
+            )
+        if len(image_keys) > 0:
+            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."
+            )
+        if self.n_action_steps > 1:
+            if self.n_action_repeats != 1:
+                raise ValueError(
+                    "If `n_action_steps > 1`, `n_action_repeats` must be left to its default value of 1."
+                )
+            if not self.use_mpc:
+                raise ValueError("If `n_action_steps > 1`, `use_mpc` must be set to `True`.")
+            if self.n_action_steps > self.horizon:
+                raise ValueError("`n_action_steps` must be less than or equal to `horizon`.")

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/tdmpc/modeling_tdmpc.py

@@ -0,0 +1,843 @@
+#!/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)
+"""
+
+# ruff: noqa: N806
+
+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,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "tdmpc"],
+):
+    """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 the results from FOWM.
+        - Nevertheless, we have verified that we can train TD-MPC for PushT. See
+          `lerobot/configs/policy/tdmpc_pusht_keypoints.yaml`.
+        - 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__()
+
+        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.
+        self._use_image = False
+        self._use_env_state = False
+        if len(image_keys) > 0:
+            assert len(image_keys) == 1
+            self._use_image = True
+            self.input_image_key = image_keys[0]
+        if "observation.environment_state" in config.input_shapes:
+            self._use_env_state = True
+
+        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.state": deque(maxlen=1),
+            "action": deque(maxlen=max(self.config.n_action_steps, self.config.n_action_repeats)),
+        }
+        if self._use_image:
+            self._queues["observation.image"] = deque(maxlen=1)
+        if self._use_env_state:
+            self._queues["observation.environment_state"] = deque(maxlen=1)
+        # 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]) -> Tensor:
+        """Select a single action given environment observations."""
+        batch = self.normalize_inputs(batch)
+        if self._use_image:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            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.
+            encode_keys = []
+            if self._use_image:
+                encode_keys.append("observation.image")
+            if self._use_env_state:
+                encode_keys.append("observation.environment_state")
+            encode_keys.append("observation.state")
+            z = self.model.encode({k: batch[k] for k in encode_keys})
+            if self.config.use_mpc:  # noqa: SIM108
+                actions = self.plan(z)  # (horizon, batch, action_dim)
+            else:
+                # Plan with the policy (π) alone. This always returns one action so unsqueeze to get a
+                # sequence dimension like in the MPC branch.
+                actions = self.model.pi(z).unsqueeze(0)
+
+            actions = torch.clamp(actions, -1, +1)
+
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.n_action_repeats > 1:
+                for _ in range(self.config.n_action_repeats):
+                    self._queues["action"].append(actions[0])
+            else:
+                # Action queue is (n_action_steps, batch_size, action_dim), so we transpose the action.
+                self._queues["action"].extend(actions[: self.config.n_action_steps])
+
+        action = self._queues["action"].popleft()
+        return action
+
+    @torch.no_grad()
+    def plan(self, z: Tensor) -> Tensor:
+        """Plan sequence of actions using TD-MPC inference.
+
+        Args:
+            z: (batch, latent_dim,) tensor for the initial state.
+        Returns:
+            (horizon, batch, action_dim,) tensor for the planned trajectory of actions.
+        """
+        device = get_device_from_parameters(self)
+
+        batch_size = z.shape[0]
+
+        # Sample Nπ trajectories from the policy.
+        pi_actions = torch.empty(
+            self.config.horizon,
+            self.config.n_pi_samples,
+            batch_size,
+            self.config.output_shapes["action"][0],
+            device=device,
+        )
+        if self.config.n_pi_samples > 0:
+            _z = einops.repeat(z, "b d -> n b 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, "b d -> n b 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, batch_size, 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,
+                batch_size,
+                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  # (n_elites, batch)
+            elite_value = value.take_along_dim(elite_idxs, dim=0)  # (n_elites, batch)
+            # (horizon, n_elites, batch, action_dim)
+            elite_actions = actions.take_along_dim(einops.rearrange(elite_idxs, "n b -> 1 n b 1"), dim=1)
+
+            # Update gaussian PDF parameters to be the (weighted) mean and standard deviation of the elites.
+            max_value = elite_value.max(0, keepdim=True)[0]  # (1, batch)
+            # 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(axis=0, keepdim=True)
+            # (horizon, batch, action_dim)
+            _mean = torch.sum(einops.rearrange(score, "n b -> n b 1") * elite_actions, dim=1)
+            _std = torch.sqrt(
+                torch.sum(
+                    einops.rearrange(score, "n b -> n b 1")
+                    * (elite_actions - einops.rearrange(_mean, "h b d -> h 1 b 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.T, 1).squeeze(), torch.arange(batch_size)]
+
+        return actions
+
+    @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 | float]:
+        """Run the batch through the model and compute the loss.
+
+        Returns a dictionary with loss as a tensor, and other information as native floats.
+        """
+        device = get_device_from_parameters(self)
+
+        batch = self.normalize_inputs(batch)
+        if self._use_image:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            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, action_dim)
+        reward = batch["next.reward"]  # (t, b)
+        observations = {k: v for k, v in batch.items() if k.startswith("observation.")}
+
+        # Apply random image augmentations.
+        if self._use_image and 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, batch_size = next_observations[
+            "observation.image" if self._use_image else "observation.environment_state"
+        ].shape[:2]
+
+        # 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 = (
+            (
+                temporal_loss_coeffs
+                * 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 (negative) log probability amounts to
+        # multiplying the MSE by 0.5 and adding a constant offset (the log(2*pi)/2 term, times the action
+        # dimension). 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(),
+            )
+        if "observation.environment_state" in config.input_shapes:
+            self.env_state_enc_layers = nn.Sequential(
+                nn.Linear(
+                    config.input_shapes["observation.environment_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 = []
+        # NOTE: Order of observations matters here.
+        if "observation.image" in self.config.input_shapes:
+            feat.append(flatten_forward_unflatten(self.image_enc_layers, obs_dict["observation.image"]))
+        if "observation.environment_state" in self.config.input_shapes:
+            feat.append(self.env_state_enc_layers(obs_dict["observation.environment_state"]))
+        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/policy.py

@@ -1,495 +0,0 @@
-# ruff: noqa: N806
-
-import time
-from collections import deque
-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
-from lerobot.common.policies.utils import populate_queues
-from lerobot.common.utils.utils import get_safe_torch_device
-
-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 TDMPCPolicy(nn.Module):
-    """Implementation of TD-MPC learning + inference."""
-
-    name = "tdmpc"
-
-    def __init__(self, cfg, n_obs_steps, n_action_steps, device):
-        super().__init__()
-        self.action_dim = cfg.action_dim
-
-        self.cfg = cfg
-        self.n_obs_steps = n_obs_steps
-        self.n_action_steps = n_action_steps
-        self.device = get_safe_torch_device(device)
-        self.std = h.linear_schedule(cfg.std_schedule, 0)
-        self.model = TOLD(cfg)
-        self.model.to(self.device)
-        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.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"])
-
-    def reset(self):
-        """
-        Clear observation and action queues. Should be called on `env.reset()`
-        """
-        self._queues = {
-            "observation.image": deque(maxlen=self.n_obs_steps),
-            "observation.state": deque(maxlen=self.n_obs_steps),
-            "action": deque(maxlen=self.n_action_steps),
-        }
-
-    @torch.no_grad()
-    def select_action(self, batch, step):
-        assert "observation.image" in batch
-        assert "observation.state" in batch
-        assert len(batch) == 2
-
-        self._queues = populate_queues(self._queues, batch)
-
-        t0 = step == 0
-
-        self.eval()
-
-        if len(self._queues["action"]) == 0:
-            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
-
-            if self.n_obs_steps == 1:
-                # hack to remove the time dimension
-                for key in batch:
-                    assert batch[key].shape[1] == 1
-                    batch[key] = batch[key][:, 0]
-
-            actions = []
-            batch_size = batch["observation.image"].shape[0]
-            for i in range(batch_size):
-                obs = {
-                    "rgb": batch["observation.image"][[i]],
-                    "state": batch["observation.state"][[i]],
-                }
-                # Note: unsqueeze needed because `act` still uses non-batch logic.
-                action = self.act(obs, t0=t0, step=self.step)
-                actions.append(action)
-            action = torch.stack(actions)
-
-            # tdmpc returns an action for 1 timestep only, so we copy it over `n_action_steps` time
-            if i in range(self.n_action_steps):
-                self._queues["action"].append(action)
-
-        action = self._queues["action"].popleft()
-        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.squeeze(2)
-        return td_target
-
-    def forward(self, batch, step):
-        # TODO(alexander-soare): Refactor TDMPC and make it comply with the policy interface documentation.
-        raise NotImplementedError()
-
-    def update(self, batch, step):
-        """Main update function. Corresponds to one iteration of the model learning."""
-        start_time = time.time()
-
-        batch_size = batch["index"].shape[0]
-
-        # TODO(rcadene): convert tdmpc with (batch size, time/horizon, channels)
-        # instead of currently (time/horizon, batch size, channels) which is not the pytorch convention
-        # batch size b = 256, time/horizon t = 5
-        # b t ... -> t b ...
-        for key in batch:
-            if batch[key].ndim > 1:
-                batch[key] = batch[key].transpose(1, 0)
-
-        action = batch["action"]
-        reward = batch["next.reward"]
-        #  idxs = batch["index"]  # TODO(rcadene): use idxs to update sampling weights
-        done = torch.zeros_like(reward, dtype=torch.bool, device=reward.device)
-        mask = torch.ones_like(reward, dtype=torch.bool, device=reward.device)
-        weights = torch.ones(batch_size, dtype=torch.bool, device=reward.device)
-
-        obses = {
-            "rgb": batch["observation.image"],
-            "state": batch["observation.state"],
-        }
-
-        shapes = {}
-        for k in obses:
-            shapes[k] = obses[k].shape
-            obses[k] = einops.rearrange(obses[k], "t b ... -> (t b) ... ")
-
-        # Apply augmentations
-        aug_tf = h.aug(self.cfg)
-        obses = aug_tf(obses)
-
-        for k in obses:
-            t, b = shapes[k][:2]
-            obses[k] = einops.rearrange(obses[k], "(t b) ... -> t b ... ", b=b, t=t)
-
-        obs, next_obses = {}, {}
-        for k in obses:
-            obs[k] = obses[k][0]
-            next_obses[k] = obses[k][1:].clone()
-
-        horizon = next_obses["rgb"].shape[0]
-        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, 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.squeeze(1)
-
-        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")
-        qs = qs.squeeze(3)
-        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)
-        consistency_loss = (rho * torch.mean(h.mse(zs[1:], z_targets), dim=2) * 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).squeeze(2) * 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 * 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()
-
-        # TODO(rcadene): implement PrioritizedSampling by modifying sampler.weights with priorities computed by a criterion
-        # 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/utils.py

@@ -1,9 +1,28 @@
+#!/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:

lerobot/common/policies/vqbet/configuration_vqbet.py

@@ -0,0 +1,167 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Seungjae Lee and Yibin Wang and Haritheja Etukuru
+# and H. Jin Kim and Nur Muhammad Mahi Shafiullah and Lerrel Pinto
+# 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 VQBeTConfig:
+    """Configuration class for VQ-BeT.
+
+    Defaults are configured for training with PushT providing proprioceptive and single camera observations.
+
+    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
+    Those are: `input_shapes` and `output_shapes`.
+
+    Notes on the inputs and outputs:
+        - "observation.state" is required as an input key.
+        - At least one key starting with "observation.image is required as an input.
+        - If there are multiple keys beginning with "observation.image" they are treated as multiple camera
+          views. Right now we only support all images having the same shape.
+        - "action" is required as an output key.
+
+    Args:
+        n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
+            current step and additional steps going back).
+        n_action_pred_token: Total number of current token and future tokens that VQ-BeT predicts.
+        action_chunk_size: Action chunk size of each action prediction token.
+        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.
+        n_vqvae_training_steps: Number of optimization steps for training Residual VQ.
+        vqvae_n_embed: Number of embedding vectors in the RVQ dictionary (each layer).
+        vqvae_embedding_dim: Dimension of each embedding vector in the RVQ dictionary.
+        vqvae_enc_hidden_dim: Size of hidden dimensions of Encoder / Decoder part of Residaul VQ-VAE
+        gpt_block_size: Max block size of minGPT (should be larger than the number of input tokens)
+        gpt_input_dim: Size of output input of GPT. This is also used as the dimension of observation features.
+        gpt_output_dim: Size of output dimension of GPT. This is also used as a input dimension of offset / bin prediction headers.
+        gpt_n_layer: Number of layers of GPT
+        gpt_n_head: Number of headers of GPT
+        gpt_hidden_dim: Size of hidden dimensions of GPT
+        dropout: Dropout rate for GPT
+        mlp_hidden_dim: Size of hidden dimensions of offset header / bin prediction headers parts of VQ-BeT
+        offset_loss_weight:  A constant that is multiplied to the offset loss
+        primary_code_loss_weight: A constant that is multiplied to the primary code prediction loss
+        secondary_code_loss_weight: A constant that is multiplied to the secondary code prediction loss
+        bet_softmax_temperature: Sampling temperature of code for rollout with VQ-BeT
+        sequentially_select: Whether select code of primary / secondary as sequentially (pick primary code,
+            and then select secodnary code), or at the same time.
+    """
+
+    # Inputs / output structure.
+    n_obs_steps: int = 5
+    n_action_pred_token: int = 3
+    action_chunk_size: int = 5
+
+    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
+    # VQ-VAE
+    n_vqvae_training_steps: int = 20000
+    vqvae_n_embed: int = 16
+    vqvae_embedding_dim: int = 256
+    vqvae_enc_hidden_dim: int = 128
+    # VQ-BeT
+    gpt_block_size: int = 500
+    gpt_input_dim: int = 512
+    gpt_output_dim: int = 512
+    gpt_n_layer: int = 8
+    gpt_n_head: int = 8
+    gpt_hidden_dim: int = 512
+    dropout: float = 0.1
+    mlp_hidden_dim: int = 1024
+    offset_loss_weight: float = 10000.0
+    primary_code_loss_weight: float = 5.0
+    secondary_code_loss_weight: float = 0.5
+    bet_softmax_temperature: float = 0.1
+    sequentially_select: 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}."
+            )
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
+        if self.crop_shape is not None:
+            for image_key in 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}]`."
+                    )
+        # Check that all input images have the same shape.
+        first_image_key = next(iter(image_keys))
+        for image_key in image_keys:
+            if self.input_shapes[image_key] != self.input_shapes[first_image_key]:
+                raise ValueError(
+                    f"`input_shapes[{image_key}]` does not match `input_shapes[{first_image_key}]`, but we "
+                    "expect all image shapes to match."
+                )

lerobot/common/policies/vqbet/modeling_vqbet.py

@@ -0,0 +1,959 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Seungjae Lee and Yibin Wang and Haritheja Etukuru
+# and H. Jin Kim and Nur Muhammad Mahi Shafiullah and Lerrel Pinto
+# 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.
+
+import math
+import warnings
+from collections import deque
+from typing import Callable, List
+
+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 torch.optim.lr_scheduler import LambdaLR
+
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.utils import get_device_from_parameters, populate_queues
+from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
+from lerobot.common.policies.vqbet.vqbet_utils import GPT, ResidualVQ
+
+# ruff: noqa: N806
+
+
+class VQBeTPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "vqbet"],
+):
+    """
+    VQ-BeT Policy as per "Behavior Generation with Latent Actions"
+    """
+
+    name = "vqbet"
+
+    def __init__(
+        self,
+        config: VQBeTConfig | 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 = VQBeTConfig()
+        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
+        )
+
+        self.vqbet = VQBeTModel(config)
+
+        self.expected_image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+
+        self.reset()
+
+    def reset(self):
+        """
+        Clear observation and action queues. Should be called on `env.reset()`
+        queues are populated during rollout of the policy, they contain the n latest observations and actions
+        """
+        self._queues = {
+            "observation.images": deque(maxlen=self.config.n_obs_steps),
+            "observation.state": deque(maxlen=self.config.n_obs_steps),
+            "action": deque(maxlen=self.config.action_chunk_size),
+        }
+
+    @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.
+        """
+
+        batch = self.normalize_inputs(batch)
+        batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+        batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+        # Note: It's important that this happens after stacking the images into a single key.
+        self._queues = populate_queues(self._queues, batch)
+
+        if not self.vqbet.action_head.vqvae_model.discretized.item():
+            warnings.warn(
+                "To evaluate in the environment, your VQ-BeT model should contain a pretrained Residual VQ.",
+                stacklevel=1,
+            )
+
+        if len(self._queues["action"]) == 0:
+            batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
+            actions = self.vqbet(batch, rollout=True)[:, : self.config.action_chunk_size]
+
+            # the dimension of returned action is (batch_size, action_chunk_size, action_dim)
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+            # since the data in the action queue's dimension is (action_chunk_size, batch_size, action_dim), we transpose the action and fill the queue
+            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 = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+        batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+        batch = self.normalize_targets(batch)
+        # VQ-BeT discretizes action using VQ-VAE before training BeT (please refer to section 3.2 in the VQ-BeT paper https://arxiv.org/pdf/2403.03181)
+        if not self.vqbet.action_head.vqvae_model.discretized.item():
+            # loss: total loss of training RVQ
+            # n_different_codes: how many of the total possible VQ codes are being used in single batch (how many of them have at least one encoder embedding as a nearest neighbor). This can be at most `vqvae_n_embed * number of layers of RVQ (=2)`.
+            # n_different_combinations: how many different code combinations are being used out of all possible combinations in single batch. This can be at most `vqvae_n_embed ^ number of layers of RVQ (=2)` (hint consider the RVQ as a decision tree).
+            loss, n_different_codes, n_different_combinations, recon_l1_error = (
+                self.vqbet.action_head.discretize(self.config.n_vqvae_training_steps, batch["action"])
+            )
+            return {
+                "loss": loss,
+                "n_different_codes": n_different_codes,
+                "n_different_combinations": n_different_combinations,
+                "recon_l1_error": recon_l1_error,
+            }
+        # if Residual VQ is already trained, VQ-BeT trains its GPT and bin prediction head / offset prediction head parts.
+        _, loss_dict = self.vqbet(batch, rollout=False)
+
+        return loss_dict
+
+
+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 VQBeTModel(nn.Module):
+    """VQ-BeT: The underlying neural network for VQ-BeT
+
+    Note: In this code we use the terms `rgb_encoder`, 'policy', `action_head`. The meanings are as follows.
+        - The `rgb_encoder` process rgb-style image observations to one-dimensional embedding vectors
+        - A `policy` is a minGPT architecture, that takes observation sequences and action query tokens to generate `features`.
+        - These `features` pass through the action head, which passes through the code prediction, offset prediction head,
+        and finally generates a prediction for the action chunks.
+
+        -------------------------------** legend **-------------------------------
+        │   n = n_obs_steps, p = n_action_pred_token, c = action_chunk_size)   │
+        │   o_{t} : visual observation at timestep {t}                           │
+        │   s_{t} : state observation at timestep {t}                            │
+        │   a_{t} : action at timestep {t}                                       │
+        │   A_Q : action_query_token                                             │
+        --------------------------------------------------------------------------
+
+
+        Training Phase 1. Discretize action using Residual VQ (for config.n_vqvae_training_steps steps)
+
+
+        ┌─────────────────┐            ┌─────────────────┐            ┌─────────────────┐
+        │                 │            │                 │            │                 │
+        │   RVQ encoder   │    ─►      │     Residual    │    ─►      │   RVQ Decoder   │
+        │ (a_{t}~a_{t+p}) │            │  Code Quantizer │            │                 │
+        │                 │            │                 │            │                 │
+        └─────────────────┘            └─────────────────┘            └─────────────────┘
+
+        Training Phase 2.
+
+          timestep {t-n+1}   timestep {t-n+2}                timestep {t}
+            ┌─────┴─────┐     ┌─────┴─────┐                 ┌─────┴─────┐
+
+        o_{t-n+1}         o_{t-n+2}           ...         o_{t}
+            │                 │                             │
+            │ s_{t-n+1}       │ s_{t-n+2}         ...       │   s_{t}           p
+            │     │           │     │                       │     │     ┌───────┴───────┐
+            │     │    A_Q    │     │    A_Q          ...   │     │    A_Q     ...     A_Q
+            │     │     │     │     │     │                 │     │     │               │
+        ┌───▼─────▼─────▼─────▼─────▼─────▼─────────────────▼─────▼─────▼───────────────▼───┐
+        │                                                                                   │
+        │                                       GPT                                         │       =>    policy
+        │                                                                                   │
+        └───────────────▼─────────────────▼─────────────────────────────▼───────────────▼───┘
+                        │                 │                             │               │
+                    ┌───┴───┐         ┌───┴───┐                     ┌───┴───┐       ┌───┴───┐
+                  code    offset    code    offset                code    offset  code    offset
+                    ▼       │         ▼       │                     ▼       │       ▼       │       =>    action_head
+               RVQ Decoder  │    RVQ Decoder  │                RVQ Decoder  │  RVQ Decoder  │
+                    └── + ──┘         └── + ──┘                     └── + ──┘       └── + ──┘
+                        ▼                 ▼                             ▼               ▼
+                   action chunk      action chunk                  action chunk     action chunk
+                    a_{t-n+1} ~       a_{t-n+2} ~                   a_{t} ~     ...  a_{t+p-1} ~
+                     a_{t-n+c}         a_{t-n+c+1}                   a_{t+c-1}        a_{t+p+c-1}
+
+                                                                        ▼
+                                                      ONLY this chunk is used in rollout!
+    """
+
+    def __init__(self, config: VQBeTConfig):
+        super().__init__()
+        self.config = config
+
+        self.rgb_encoder = VQBeTRgbEncoder(config)
+        self.num_images = len([k for k in config.input_shapes if k.startswith("observation.image")])
+        # This action query token is used as a prompt for querying action chunks. Please refer to "A_Q" in the image above.
+        # Note: During the forward pass, this token is repeated as many times as needed. The authors also experimented with initializing the necessary number of tokens independently and observed inferior results.
+        self.action_token = nn.Parameter(torch.randn(1, 1, self.config.gpt_input_dim))
+
+        # To input state and observation features into GPT layers, we first project the features to fit the shape of input size of GPT.
+        self.state_projector = MLP(
+            config.input_shapes["observation.state"][0], hidden_channels=[self.config.gpt_input_dim]
+        )
+        self.rgb_feature_projector = MLP(
+            self.rgb_encoder.feature_dim, hidden_channels=[self.config.gpt_input_dim]
+        )
+
+        # GPT part of VQ-BeT
+        self.policy = GPT(config)
+        # bin prediction head / offset prediction head part of VQ-BeT
+        self.action_head = VQBeTHead(config)
+
+        # Action tokens for: each observation step, the current action token, and all future action tokens.
+        num_tokens = self.config.n_action_pred_token + self.config.n_obs_steps - 1
+        self.register_buffer(
+            "select_target_actions_indices",
+            torch.row_stack([torch.arange(i, i + self.config.action_chunk_size) for i in range(num_tokens)]),
+        )
+
+    def forward(self, batch: dict[str, Tensor], rollout: bool) -> Tensor:
+        # Input validation.
+        assert set(batch).issuperset({"observation.state", "observation.images"})
+        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.images"], "b s n ... -> (b s n) ...")
+        )
+        # Separate batch and sequence dims.
+        img_features = einops.rearrange(
+            img_features, "(b s n) ... -> b s n ...", b=batch_size, s=n_obs_steps, n=self.num_images
+        )
+
+        # Arrange prior and current observation step tokens as shown in the class docstring.
+        # First project features to token dimension.
+        rgb_tokens = self.rgb_feature_projector(
+            img_features
+        )  # (batch, obs_step, number of different cameras, projection dims)
+        input_tokens = [rgb_tokens[:, :, i] for i in range(rgb_tokens.size(2))]
+        input_tokens.append(
+            self.state_projector(batch["observation.state"])
+        )  # (batch, obs_step, projection dims)
+        input_tokens.append(einops.repeat(self.action_token, "1 1 d -> b n d", b=batch_size, n=n_obs_steps))
+        # Interleave tokens by stacking and rearranging.
+        input_tokens = torch.stack(input_tokens, dim=2)
+        input_tokens = einops.rearrange(input_tokens, "b n t d -> b (n t) d")
+
+        len_additional_action_token = self.config.n_action_pred_token - 1
+        future_action_tokens = self.action_token.repeat(batch_size, len_additional_action_token, 1)
+
+        # add additional action query tokens for predicting future action chunks
+        input_tokens = torch.cat([input_tokens, future_action_tokens], dim=1)
+
+        # get action features (pass through GPT)
+        features = self.policy(input_tokens)
+        # len(self.config.input_shapes) is the number of different observation modes. this line gets the index of action prompt tokens.
+        historical_act_pred_index = np.arange(0, n_obs_steps) * (len(self.config.input_shapes) + 1) + len(
+            self.config.input_shapes
+        )
+
+        # only extract the output tokens at the position of action query:
+        # Behavior Transformer (BeT), and VQ-BeT are both sequence-to-sequence prediction models, mapping sequential observation to sequential action (please refer to section 2.2 in BeT paper https://arxiv.org/pdf/2206.11251).
+        # Thus, it predict historical action sequence, in addition to current and future actions (predicting future actions : optional).
+        features = torch.cat(
+            [features[:, historical_act_pred_index], features[:, -len_additional_action_token:]], dim=1
+        )
+        # pass through action head
+        action_head_output = self.action_head(features)
+        # if rollout, VQ-BeT don't calculate loss
+        if rollout:
+            return action_head_output["predicted_action"][:, n_obs_steps - 1, :].reshape(
+                batch_size, self.config.action_chunk_size, -1
+            )
+        # else, it calculate overall loss (bin prediction loss, and offset loss)
+        else:
+            output = batch["action"][:, self.select_target_actions_indices]
+            loss = self.action_head.loss_fn(action_head_output, output, reduction="mean")
+            return action_head_output, loss
+
+
+class VQBeTHead(nn.Module):
+    def __init__(self, config: VQBeTConfig):
+        """
+        VQBeTHead takes output of GPT layers, and pass the feature through bin prediction head (`self.map_to_cbet_preds_bin`), and offset prediction head (`self.map_to_cbet_preds_offset`)
+
+        self.map_to_cbet_preds_bin: outputs probability of each code (for each layer).
+            The input dimension of `self.map_to_cbet_preds_bin` is same with the output of GPT,
+            and the output dimension of `self.map_to_cbet_preds_bin` is `self.vqvae_model.vqvae_num_layers (=fixed as 2) * self.config.vqvae_n_embed`.
+            if the agent select the code sequentially, we use self.map_to_cbet_preds_primary_bin and self.map_to_cbet_preds_secondary_bin instead of self._map_to_cbet_preds_bin.
+
+        self.map_to_cbet_preds_offset: output the predicted offsets for all the codes in all the layers.
+            The input dimension of ` self.map_to_cbet_preds_offset` is same with the output of GPT,
+            and the output dimension of ` self.map_to_cbet_preds_offset` is `self.vqvae_model.vqvae_num_layers (=fixed as 2) * self.config.vqvae_n_embed * config.action_chunk_size * config.output_shapes["action"][0]`.
+        """
+
+        super().__init__()
+        self.config = config
+        # init vqvae
+        self.vqvae_model = VqVae(config)
+        if config.sequentially_select:
+            self.map_to_cbet_preds_primary_bin = MLP(
+                in_channels=config.gpt_output_dim,
+                hidden_channels=[self.config.vqvae_n_embed],
+            )
+            self.map_to_cbet_preds_secondary_bin = MLP(
+                in_channels=config.gpt_output_dim + self.config.vqvae_n_embed,
+                hidden_channels=[self.config.vqvae_n_embed],
+            )
+        else:
+            self.map_to_cbet_preds_bin = MLP(
+                in_channels=config.gpt_output_dim,
+                hidden_channels=[self.vqvae_model.vqvae_num_layers * self.config.vqvae_n_embed],
+            )
+        self.map_to_cbet_preds_offset = MLP(
+            in_channels=config.gpt_output_dim,
+            hidden_channels=[
+                self.vqvae_model.vqvae_num_layers
+                * self.config.vqvae_n_embed
+                * config.action_chunk_size
+                * config.output_shapes["action"][0],
+            ],
+        )
+        # loss
+        self._focal_loss_fn = FocalLoss(gamma=2.0)
+
+    def discretize(self, n_vqvae_training_steps, actions):
+        # Resize the action sequence data to fit the action chunk size using a sliding window approach.
+        actions = torch.cat(
+            [
+                actions[:, j : j + self.config.action_chunk_size, :]
+                for j in range(actions.shape[1] + 1 - self.config.action_chunk_size)
+            ],
+            dim=0,
+        )
+        # `actions` is a tensor of shape (new_batch, action_chunk_size, action_dim) where new_batch is the number of possible chunks created from the original sequences using the sliding window.
+
+        loss, metric = self.vqvae_model.vqvae_forward(actions)
+        n_different_codes = sum(
+            [len(torch.unique(metric[2][:, i])) for i in range(self.vqvae_model.vqvae_num_layers)]
+        )
+        n_different_combinations = len(torch.unique(metric[2], dim=0))
+        recon_l1_error = metric[0].detach().cpu().item()
+        self.vqvae_model.optimized_steps += 1
+        # if we updated RVQ more than `n_vqvae_training_steps` steps, we freeze the RVQ part.
+        if self.vqvae_model.optimized_steps >= n_vqvae_training_steps:
+            self.vqvae_model.discretized = torch.tensor(True)
+            self.vqvae_model.vq_layer.freeze_codebook = torch.tensor(True)
+            print("Finished discretizing action data!")
+            self.vqvae_model.eval()
+            for param in self.vqvae_model.vq_layer.parameters():
+                param.requires_grad = False
+        return loss, n_different_codes, n_different_combinations, recon_l1_error
+
+    def forward(self, x, **kwargs):
+        # N is the batch size, and T is number of action query tokens, which are process through same GPT
+        N, T, _ = x.shape
+        # we calculate N and T side parallely. Thus, the dimensions would be
+        # (batch size * number of action query tokens, action chunk size, action dimension)
+        x = einops.rearrange(x, "N T WA -> (N T) WA")
+
+        # sample offsets
+        cbet_offsets = self.map_to_cbet_preds_offset(x)
+        cbet_offsets = einops.rearrange(
+            cbet_offsets,
+            "(NT) (G C WA) -> (NT) G C WA",
+            G=self.vqvae_model.vqvae_num_layers,
+            C=self.config.vqvae_n_embed,
+        )
+        # if self.config.sequentially_select is True, bin prediction head first sample the primary code, and then sample secondary code
+        if self.config.sequentially_select:
+            cbet_primary_logits = self.map_to_cbet_preds_primary_bin(x)
+
+            # select primary bin first
+            cbet_primary_probs = torch.softmax(
+                cbet_primary_logits / self.config.bet_softmax_temperature, dim=-1
+            )
+            NT, choices = cbet_primary_probs.shape
+            sampled_primary_centers = einops.rearrange(
+                torch.multinomial(cbet_primary_probs.view(-1, choices), num_samples=1),
+                "(NT) 1 -> NT",
+                NT=NT,
+            )
+
+            cbet_secondary_logits = self.map_to_cbet_preds_secondary_bin(
+                torch.cat(
+                    (x, F.one_hot(sampled_primary_centers, num_classes=self.config.vqvae_n_embed)),
+                    axis=1,
+                )
+            )
+            cbet_secondary_probs = torch.softmax(
+                cbet_secondary_logits / self.config.bet_softmax_temperature, dim=-1
+            )
+            sampled_secondary_centers = einops.rearrange(
+                torch.multinomial(cbet_secondary_probs.view(-1, choices), num_samples=1),
+                "(NT) 1 -> NT",
+                NT=NT,
+            )
+            sampled_centers = torch.stack((sampled_primary_centers, sampled_secondary_centers), axis=1)
+            cbet_logits = torch.stack([cbet_primary_logits, cbet_secondary_logits], dim=1)
+        # if self.config.sequentially_select is False, bin prediction head samples primary and secondary code at once.
+        else:
+            cbet_logits = self.map_to_cbet_preds_bin(x)
+            cbet_logits = einops.rearrange(
+                cbet_logits, "(NT) (G C) -> (NT) G C", G=self.vqvae_model.vqvae_num_layers
+            )
+            cbet_probs = torch.softmax(cbet_logits / self.config.bet_softmax_temperature, dim=-1)
+            NT, G, choices = cbet_probs.shape
+            sampled_centers = einops.rearrange(
+                torch.multinomial(cbet_probs.view(-1, choices), num_samples=1),
+                "(NT G) 1 -> NT G",
+                NT=NT,
+            )
+
+        device = get_device_from_parameters(self)
+        indices = (
+            torch.arange(NT, device=device).unsqueeze(1),
+            torch.arange(self.vqvae_model.vqvae_num_layers, device=device).unsqueeze(0),
+            sampled_centers,
+        )
+        # Use advanced indexing to sample the values (Extract the only offsets corresponding to the sampled codes.)
+        sampled_offsets = cbet_offsets[indices]
+        # Then, sum the offsets over the RVQ layers to get a net offset for the bin prediction
+        sampled_offsets = sampled_offsets.sum(dim=1)
+        with torch.no_grad():
+            # Get the centroids (= vectors corresponding to the codes) of each layer to pass it through RVQ decoder
+            return_decoder_input = self.vqvae_model.get_embeddings_from_code(sampled_centers).clone().detach()
+            # pass the centroids through decoder to get actions.
+            decoded_action = self.vqvae_model.get_action_from_latent(return_decoder_input).clone().detach()
+        # reshaped extracted offset to match with decoded centroids
+        sampled_offsets = einops.rearrange(
+            sampled_offsets, "NT (W A) -> NT W A", W=self.config.action_chunk_size
+        )
+        # add offset and decoded centroids
+        predicted_action = decoded_action + sampled_offsets
+        predicted_action = einops.rearrange(
+            predicted_action,
+            "(N T) W A -> N T (W A)",
+            N=N,
+            T=T,
+            W=self.config.action_chunk_size,
+        )
+
+        return {
+            "cbet_logits": cbet_logits,
+            "predicted_action": predicted_action,
+            "sampled_centers": sampled_centers,
+            "decoded_action": decoded_action,
+        }
+
+    def loss_fn(self, pred, target, **kwargs):
+        """
+        for given ground truth action values (target), and prediction (pred) this function calculates the overall loss.
+
+        predicted_action: predicted action chunk (offset + decoded centroids)
+        sampled_centers: sampled centroids (code of RVQ)
+        decoded_action: decoded action, which is produced by passing sampled_centers through RVQ decoder
+        NT: batch size * T
+        T: number of action query tokens, which are process through same GPT
+        cbet_logits: probability of all codes in each layer
+        """
+        action_seq = target
+        predicted_action = pred["predicted_action"]
+        sampled_centers = pred["sampled_centers"]
+        decoded_action = pred["decoded_action"]
+        NT = predicted_action.shape[0] * predicted_action.shape[1]
+
+        cbet_logits = pred["cbet_logits"]
+
+        predicted_action = einops.rearrange(
+            predicted_action, "N T (W A) -> (N T) W A", W=self.config.action_chunk_size
+        )
+
+        action_seq = einops.rearrange(action_seq, "N T W A -> (N T) W A")
+        # Figure out the loss for the actions.
+        # First, we need to find the closest cluster center for each ground truth action.
+        with torch.no_grad():
+            state_vq, action_bins = self.vqvae_model.get_code(action_seq)  # action_bins: NT, G
+
+        # Now we can compute the loss.
+
+        # offset loss is L1 distance between the predicted action and ground truth action
+        offset_loss = F.l1_loss(action_seq, predicted_action)
+
+        # calculate primary code prediction loss
+        cbet_loss1 = self._focal_loss_fn(
+            cbet_logits[:, 0, :],
+            action_bins[:, 0],
+        )
+        # calculate secondary code prediction loss
+        cbet_loss2 = self._focal_loss_fn(
+            cbet_logits[:, 1, :],
+            action_bins[:, 1],
+        )
+        # add all the prediction loss
+        cbet_loss = (
+            cbet_loss1 * self.config.primary_code_loss_weight
+            + cbet_loss2 * self.config.secondary_code_loss_weight
+        )
+
+        equal_primary_code_rate = torch.sum((action_bins[:, 0] == sampled_centers[:, 0]).int()) / (NT)
+        equal_secondary_code_rate = torch.sum((action_bins[:, 1] == sampled_centers[:, 1]).int()) / (NT)
+
+        action_mse_error = torch.mean((action_seq - predicted_action) ** 2)
+        vq_action_error = torch.mean(torch.abs(action_seq - decoded_action))
+        offset_action_error = torch.mean(torch.abs(action_seq - predicted_action))
+        action_error_max = torch.max(torch.abs(action_seq - predicted_action))
+
+        loss = cbet_loss + self.config.offset_loss_weight * offset_loss
+
+        loss_dict = {
+            "loss": loss,
+            "classification_loss": cbet_loss.detach().cpu().item(),
+            "offset_loss": offset_loss.detach().cpu().item(),
+            "equal_primary_code_rate": equal_primary_code_rate.detach().cpu().item(),
+            "equal_secondary_code_rate": equal_secondary_code_rate.detach().cpu().item(),
+            "vq_action_error": vq_action_error.detach().cpu().item(),
+            "offset_action_error": offset_action_error.detach().cpu().item(),
+            "action_error_max": action_error_max.detach().cpu().item(),
+            "action_mse_error": action_mse_error.detach().cpu().item(),
+        }
+        return loss_dict
+
+
+class VQBeTOptimizer(torch.optim.Adam):
+    def __init__(self, policy, cfg):
+        vqvae_params = (
+            list(policy.vqbet.action_head.vqvae_model.encoder.parameters())
+            + list(policy.vqbet.action_head.vqvae_model.decoder.parameters())
+            + list(policy.vqbet.action_head.vqvae_model.vq_layer.parameters())
+        )
+        decay_params, no_decay_params = policy.vqbet.policy.configure_parameters()
+        decay_params = (
+            decay_params
+            + list(policy.vqbet.rgb_encoder.parameters())
+            + list(policy.vqbet.state_projector.parameters())
+            + list(policy.vqbet.rgb_feature_projector.parameters())
+            + [policy.vqbet.action_token]
+            + list(policy.vqbet.action_head.map_to_cbet_preds_offset.parameters())
+        )
+
+        if cfg.policy.sequentially_select:
+            decay_params = (
+                decay_params
+                + list(policy.vqbet.action_head.map_to_cbet_preds_primary_bin.parameters())
+                + list(policy.vqbet.action_head.map_to_cbet_preds_secondary_bin.parameters())
+            )
+        else:
+            decay_params = decay_params + list(policy.vqbet.action_head.map_to_cbet_preds_bin.parameters())
+
+        optim_groups = [
+            {
+                "params": decay_params,
+                "weight_decay": cfg.training.adam_weight_decay,
+                "lr": cfg.training.lr,
+            },
+            {
+                "params": vqvae_params,
+                "weight_decay": 0.0001,
+                "lr": cfg.training.vqvae_lr,
+            },
+            {
+                "params": no_decay_params,
+                "weight_decay": 0.0,
+                "lr": cfg.training.lr,
+            },
+        ]
+        super().__init__(
+            optim_groups,
+            cfg.training.lr,
+            cfg.training.adam_betas,
+            cfg.training.adam_eps,
+        )
+
+
+class VQBeTScheduler(nn.Module):
+    def __init__(self, optimizer, cfg):
+        super().__init__()
+        n_vqvae_training_steps = cfg.training.n_vqvae_training_steps
+
+        num_warmup_steps = cfg.training.lr_warmup_steps
+        num_training_steps = cfg.training.offline_steps
+        num_cycles = 0.5
+
+        def lr_lambda(current_step):
+            if current_step < n_vqvae_training_steps:
+                return float(1)
+            else:
+                current_step = current_step - n_vqvae_training_steps
+                if current_step < num_warmup_steps:
+                    return float(current_step) / float(max(1, num_warmup_steps))
+                progress = float(current_step - num_warmup_steps) / float(
+                    max(1, num_training_steps - num_warmup_steps)
+                )
+                return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress)))
+
+        self.lr_scheduler = LambdaLR(optimizer, lr_lambda, -1)
+
+    def step(self):
+        self.lr_scheduler.step()
+
+
+class VQBeTRgbEncoder(nn.Module):
+    """Encode an RGB image into a 1D feature vector.
+
+    Includes the ability to normalize and crop the image first.
+
+    Same with DiffusionRgbEncoder from modeling_diffusion.py
+    """
+
+    def __init__(self, config: VQBeTConfig):
+        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` if it is provided, otherwise it should use the
+        # height and width from `config.input_shapes`.
+        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_h_w = (
+            config.crop_shape if config.crop_shape is not None else config.input_shapes[image_key][1:]
+        )
+        dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *dummy_input_h_w))
+        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 VqVae(nn.Module):
+    def __init__(
+        self,
+        config: VQBeTConfig,
+    ):
+        """
+        VQ-VAE is composed of three parts: encoder, vq_layer, and decoder.
+        Encoder and decoder are MLPs consisting of an input, output layer, and hidden layer, respectively.
+        The vq_layer uses residual VQs.
+
+        This class contains functions for training the encoder and decoder along with the residual VQ layer (for trainign phase 1),
+        as well as functions to help BeT training part in training phase 2.
+        """
+
+        super().__init__()
+        self.config = config
+        # 'discretized' indicates whether the Residual VQ part is trained or not. (After finishing the training, we set discretized=True)
+        self.register_buffer("discretized", torch.tensor(False))
+        self.optimized_steps = 0
+        # we use the fixed number of layers for Residual VQ across all environments.
+        self.vqvae_num_layers = 2
+
+        self.vq_layer = ResidualVQ(
+            dim=config.vqvae_embedding_dim,
+            num_quantizers=self.vqvae_num_layers,
+            codebook_size=config.vqvae_n_embed,
+        )
+
+        self.encoder = MLP(
+            in_channels=self.config.output_shapes["action"][0] * self.config.action_chunk_size,
+            hidden_channels=[
+                config.vqvae_enc_hidden_dim,
+                config.vqvae_enc_hidden_dim,
+                config.vqvae_embedding_dim,
+            ],
+        )
+        self.decoder = MLP(
+            in_channels=config.vqvae_embedding_dim,
+            hidden_channels=[
+                config.vqvae_enc_hidden_dim,
+                config.vqvae_enc_hidden_dim,
+                self.config.output_shapes["action"][0] * self.config.action_chunk_size,
+            ],
+        )
+
+    def get_embeddings_from_code(self, encoding_indices):
+        # This function gets code indices as inputs, and outputs embedding vectors corresponding to the code indices.
+        with torch.no_grad():
+            z_embed = self.vq_layer.get_codebook_vector_from_indices(encoding_indices)
+            # since the RVQ has multiple layers, it adds the vectors in the axis of layers to provide a vector for that code combination.
+            z_embed = z_embed.sum(dim=0)
+        return z_embed
+
+    def get_action_from_latent(self, latent):
+        # given latent vector, this function outputs the decoded action.
+        output = self.decoder(latent)
+        if self.config.action_chunk_size == 1:
+            return einops.rearrange(output, "N (T A) -> N T A", A=self.config.output_shapes["action"][0])
+        else:
+            return einops.rearrange(output, "N (T A) -> N T A", A=self.config.output_shapes["action"][0])
+
+    def get_code(self, state):
+        # in phase 2 of VQ-BeT training, we need a `ground truth labels of action data` to calculate the Focal loss for code prediction head. (please refer to section 3.3 in the paper https://arxiv.org/pdf/2403.03181)
+        # this function outputs the `GT code` of given action using frozen encoder and quantization layers. (please refer to Figure 2. in the paper https://arxiv.org/pdf/2403.03181)
+        state = einops.rearrange(state, "N T A -> N (T A)")
+        with torch.no_grad():
+            state_rep = self.encoder(state)
+            state_rep_shape = state_rep.shape[:-1]
+            state_rep_flat = state_rep.view(state_rep.size(0), -1, state_rep.size(1))
+            state_rep_flat, vq_code, vq_loss_state = self.vq_layer(state_rep_flat)
+            state_vq = state_rep_flat.view(*state_rep_shape, -1)
+            vq_code = vq_code.view(*state_rep_shape, -1)
+            vq_loss_state = torch.sum(vq_loss_state)
+            return state_vq, vq_code
+
+    def vqvae_forward(self, state):
+        # This function passes the given data through Residual VQ with Encoder and Decoder. Please refer to section 3.2 in the paper https://arxiv.org/pdf/2403.03181).
+        state = einops.rearrange(state, "N T A -> N (T A)")
+        # We start with passing action (or action chunk) at:t+n through the encoder ϕ.
+        state_rep = self.encoder(state)
+        state_rep_shape = state_rep.shape[:-1]
+        state_rep_flat = state_rep.view(state_rep.size(0), -1, state_rep.size(1))
+        # The resulting latent embedding vector x = ϕ(at:t+n) is then mapped to an embedding vector in the codebook of the RVQ layers by the nearest neighbor look-up.
+        state_rep_flat, vq_code, vq_loss_state = self.vq_layer(state_rep_flat)
+        state_vq = state_rep_flat.view(*state_rep_shape, -1)
+        vq_code = vq_code.view(*state_rep_shape, -1)
+        # since the RVQ has multiple layers, it adds the vectors in the axis of layers to provide a vector for that code combination.
+        vq_loss_state = torch.sum(vq_loss_state)
+        # Then, the discretized vector zq(x) is reconstructed as ψ(zq(x)) by passing through the decoder ψ.
+        dec_out = self.decoder(state_vq)
+        # Calculate L1 reconstruction loss
+        encoder_loss = (state - dec_out).abs().mean()
+        # add encoder reconstruction loss and commitment loss
+        rep_loss = encoder_loss + vq_loss_state * 5
+
+        metric = (
+            encoder_loss.clone().detach(),
+            vq_loss_state.clone().detach(),
+            vq_code,
+            rep_loss.item(),
+        )
+        return rep_loss, metric
+
+
+class FocalLoss(nn.Module):
+    """
+    From https://github.com/notmahi/miniBET/blob/main/behavior_transformer/bet.py
+    """
+
+    def __init__(self, gamma: float = 0, size_average: bool = True):
+        super().__init__()
+        self.gamma = gamma
+        self.size_average = size_average
+
+    def forward(self, input, target):
+        if len(input.shape) == 3:
+            N, T, _ = input.shape
+            logpt = F.log_softmax(input, dim=-1)
+            logpt = logpt.gather(-1, target.view(N, T, 1)).view(N, T)
+        elif len(input.shape) == 2:
+            logpt = F.log_softmax(input, dim=-1)
+            logpt = logpt.gather(-1, target.view(-1, 1)).view(-1)
+        pt = logpt.exp()
+
+        loss = -1 * (1 - pt) ** self.gamma * logpt
+        if self.size_average:
+            return loss.mean()
+        else:
+            return loss.sum()
+
+
+class MLP(torch.nn.Sequential):
+    def __init__(
+        self,
+        in_channels: int,
+        hidden_channels: List[int],
+    ):
+        layers = []
+        in_dim = in_channels
+        for hidden_dim in hidden_channels[:-1]:
+            layers.append(torch.nn.Linear(in_dim, hidden_dim))
+            layers.append(torch.nn.ReLU())
+            in_dim = hidden_dim
+
+        layers.append(torch.nn.Linear(in_dim, hidden_channels[-1]))
+
+        super().__init__(*layers)