post-merge fix

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

.dockerignore

@@ -1,3 +1,17 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 # Misc
 .git
 tmp

.gitattributes

@@ -1,3 +1,17 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 *.memmap filter=lfs diff=lfs merge=lfs -text
 *.stl filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -1,3 +1,17 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 name: "\U0001F41B Bug Report"
 description: Submit a bug report to help us improve LeRobot
 body:

.github/PULL_REQUEST_TEMPLATE.md

@@ -21,7 +21,7 @@ Provide a simple way for the reviewer to try out your changes.
 
 Examples:
 ```bash
-DATA_DIR=tests/data pytest -sx tests/test_stuff.py::test_something
+pytest -sx tests/test_stuff.py::test_something
 ```
 ```bash
 python lerobot/scripts/train.py --some.option=true

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

@@ -1,3 +1,17 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 # Inspired by
 # https://github.com/huggingface/peft/blob/main/.github/workflows/build_docker_images.yml
 name: Builds
@@ -8,6 +22,8 @@ on:
   schedule:
     - cron: "0 1 * * *"
 
+permissions: {}
+
 env:
   PYTHON_VERSION: "3.10"
 
@@ -25,11 +41,14 @@ jobs:
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
+        with:
+          cache-binary: false
 
       - name: Check out code
         uses: actions/checkout@v4
         with:
           lfs: true
+          persist-credentials: false
 
       - name: Login to DockerHub
         uses: docker/login-action@v3
@@ -60,11 +79,14 @@ jobs:
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
+        with:
+          cache-binary: false
 
       - name: Check out code
         uses: actions/checkout@v4
         with:
           lfs: true
+          persist-credentials: false
 
       - name: Login to DockerHub
         uses: docker/login-action@v3
@@ -89,9 +111,13 @@ jobs:
     steps:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
+        with:
+          cache-binary: false
 
       - name: Check out code
         uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       - name: Login to DockerHub
         uses: docker/login-action@v3

.github/workflows/nightly-tests.yml

@@ -1,3 +1,17 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 # Inspired by
 # https://github.com/huggingface/peft/blob/main/.github/workflows/nightly.yml
 name: Nightly
@@ -7,10 +21,10 @@ on:
   schedule:
     - cron: "0 2 * * *"
 
-env:
-  DATA_DIR: tests/data
-  # SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
+permissions: {}
 
+# env:
+  # SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
 jobs:
   run_all_tests_cpu:
     name: CPU
@@ -30,13 +44,9 @@ jobs:
         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
 
 

.github/workflows/quality.yml

@@ -1,15 +1,29 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 name: Quality
 
 on:
   workflow_dispatch:
   workflow_call:
   pull_request:
-    branches:
-      - main
   push:
     branches:
       - main
 
+permissions: {}
+
 env:
   PYTHON_VERSION: "3.10"
 
@@ -19,7 +33,9 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout Repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       - name: Set up Python
         uses: actions/setup-python@v4
@@ -30,55 +46,27 @@ jobs:
         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}" >> $GITHUB_ENV
+          echo "ruff_version=${RUFF_VERSION}" >> $GITHUB_OUTPUT
 
       - name: Install Ruff
-        run: python -m pip install "ruff==${{ env.RUFF_VERSION }}"
+        env:
+          RUFF_VERSION: ${{ steps.get-ruff-version.outputs.ruff_version }}
+        run: python -m pip install "ruff==${RUFF_VERSION}"
 
       - name: Ruff check
-        run: ruff check
+        run: ruff check --output-format=github
 
       - name: Ruff format
         run: ruff format --diff
 
-
-  poetry_check:
-    name: Poetry check
+  typos:
+    name: Typos
     runs-on: ubuntu-latest
     steps:
       - name: Checkout Repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
-      - 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
+      - name: typos-action
+        uses: crate-ci/typos@v1.29.10

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

@@ -1,15 +1,29 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 # Inspired by
 # https://github.com/huggingface/peft/blob/main/.github/workflows/test-docker-build.yml
 name: Test Dockerfiles
 
 on:
   pull_request:
-    branches:
-      - main
     paths:
       # Run only when DockerFile files are modified
       - "docker/**"
 
+permissions: {}
+
 env:
   PYTHON_VERSION: "3.10"
 
@@ -22,6 +36,8 @@ jobs:
     steps:
       - name: Check out code
         uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       - name: Get changed files
         id: changed-files
@@ -30,21 +46,18 @@ jobs:
           files: docker/**
           json: "true"
 
-      - name: Run step if only the files listed above change
+      - name: Run step if only the files listed above change  # zizmor: ignore[template-injection]
         if: steps.changed-files.outputs.any_changed == 'true'
         id: set-matrix
-        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 }} != ''
+    if: needs.get_changed_files.outputs.matrix != ''
     strategy:
       fail-fast: false
       matrix:
@@ -52,9 +65,13 @@ jobs:
     steps:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
+        with:
+          cache-binary: false
 
       - name: Check out code
         uses: actions/checkout@v4
+        with:
+          persist-credentials: false
 
       - name: Build Docker image
         uses: docker/build-push-action@v5

.github/workflows/test.yml

@@ -1,15 +1,28 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 name: Tests
 
 on:
   pull_request:
-    branches:
-      - main
     paths:
       - "lerobot/**"
       - "tests/**"
       - "examples/**"
       - ".github/**"
-      - "poetry.lock"
+      - "pyproject.toml"
+      - ".pre-commit-config.yaml"
       - "Makefile"
       - ".cache/**"
   push:
@@ -20,21 +33,27 @@ on:
       - "tests/**"
       - "examples/**"
       - ".github/**"
-      - "poetry.lock"
+      - "pyproject.toml"
+      - ".pre-commit-config.yaml"
       - "Makefile"
       - ".cache/**"
 
+permissions: {}
+
+env:
+  UV_VERSION: "0.6.0"
+
 jobs:
   pytest:
     name: Pytest
     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
+          persist-credentials: false
 
       - name: Install apt dependencies
       # portaudio19-dev is needed to install pyaudio
@@ -42,25 +61,19 @@ jobs:
           sudo apt-get update && \
           sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
 
-      - name: Install poetry
-        run: |
-          pipx install poetry && poetry config virtualenvs.in-project true
-          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
-
-      # TODO(rcadene, aliberts): python 3.12 seems to be used in the tests, not python 3.10
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
+      - name: Install uv and python
+        uses: astral-sh/setup-uv@v5
         with:
+          enable-cache: true
+          version: ${{ env.UV_VERSION }}
           python-version: "3.10"
-          cache: "poetry"
 
-      - name: Install poetry dependencies
-        run: |
-          poetry install --all-extras
+      - name: Install lerobot (all extras)
+        run: uv sync --all-extras
 
       - name: Test with pytest
         run: |
-          pytest tests -v --cov=./lerobot --durations=0 \
+          uv run pytest tests -v --cov=./lerobot --durations=0 \
             -W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
             -W ignore::UserWarning:torch.utils.data.dataloader:558 \
             -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
@@ -70,50 +83,44 @@ jobs:
     name: Pytest (minimal install)
     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
+          persist-credentials: false
 
       - 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
-
-      # TODO(rcadene, aliberts): python 3.12 seems to be used in the tests, not python 3.10
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
+      - name: Install uv and python
+        uses: astral-sh/setup-uv@v5
         with:
+          enable-cache: true
+          version: ${{ env.UV_VERSION }}
           python-version: "3.10"
 
-      - name: Install poetry dependencies
-        run: |
-          poetry install --extras "test"
+      - name: Install lerobot
+        run: uv sync --extra "test"
 
       - name: Test with pytest
         run: |
-          pytest tests -v --cov=./lerobot --durations=0 \
+          uv run pytest tests -v --cov=./lerobot --durations=0 \
             -W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
             -W ignore::UserWarning:torch.utils.data.dataloader:558 \
             -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
             && rm -rf tests/outputs outputs
 
-
   end-to-end:
     name: End-to-end
     runs-on: ubuntu-latest
     env:
-      DATA_DIR: tests/data
       MUJOCO_GL: egl
     steps:
       - uses: actions/checkout@v4
         with:
           lfs: true  # Ensure LFS files are pulled
+          persist-credentials: false
 
       - name: Install apt dependencies
       # portaudio19-dev is needed to install pyaudio
@@ -121,20 +128,21 @@ jobs:
           sudo apt-get update && \
           sudo apt-get install -y libegl1-mesa-dev portaudio19-dev
 
-      - name: Install poetry
-        run: |
-          pipx install poetry && poetry config virtualenvs.in-project true
-          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
-
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
+      - name: Install uv and python
+        uses: astral-sh/setup-uv@v5
         with:
+          enable-cache: true
+          version: ${{ env.UV_VERSION }}
           python-version: "3.10"
-          cache: "poetry"
 
-      - name: Install poetry dependencies
+      - name: Install lerobot (all extras)
         run: |
-          poetry install --all-extras
+          uv venv
+          uv sync --all-extras
+
+      - name: venv
+        run: |
+          echo "PYTHON_PATH=${{ github.workspace }}/.venv/bin/python" >> $GITHUB_ENV
 
       - name: Test end-to-end
         run: |

.github/workflows/trufflehog.yml

@@ -1,10 +1,23 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 on:
   push:
 
 name: Secret Leaks
 
-permissions:
-  contents: read
+permissions: {}
 
 jobs:
   trufflehog:
@@ -14,6 +27,8 @@ jobs:
       uses: actions/checkout@v4
       with:
         fetch-depth: 0
+        persist-credentials: false
+
     - name: Secret Scanning
       uses: trufflesecurity/trufflehog@main
       with:

.gitignore

@@ -1,3 +1,17 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 # Logging
 logs
 tmp
@@ -49,6 +63,10 @@ share/python-wheels/
 *.egg
 MANIFEST
 
+# uv/poetry lock files
+poetry.lock
+uv.lock
+
 # PyInstaller
 #  Usually these files are written by a python script from a template
 #  before PyInstaller builds the exe, so as to inject date/other infos into it.

.pre-commit-config.yaml

@@ -1,9 +1,24 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 exclude: ^(tests/data)
 default_language_version:
     python: python3.10
 repos:
+  ##### Style / Misc. #####
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.6.0
+    rev: v5.0.0
     hooks:
       - id: check-added-large-files
       - id: debug-statements
@@ -13,25 +28,34 @@ repos:
       - id: check-toml
       - id: end-of-file-fixer
       - id: trailing-whitespace
+  - repo: https://github.com/crate-ci/typos
+    rev: v1.30.0
+    hooks:
+      - id: typos
+        args: [--force-exclude]
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.16.0
+    rev: v3.19.1
     hooks:
     -   id: pyupgrade
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.5.2
+    rev: v0.9.9
     hooks:
       - id: ruff
         args: [--fix]
       - id: ruff-format
-  - repo: https://github.com/python-poetry/poetry
-    rev: 1.8.0
-    hooks:
-      - id: poetry-check
-      - id: poetry-lock
-        args:
-          - "--check"
-          - "--no-update"
+
+  ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.18.4
+    rev: v8.24.0
     hooks:
       - id: gitleaks
+  - repo: https://github.com/woodruffw/zizmor-pre-commit
+    rev: v1.4.1
+    hooks:
+      - id: zizmor
+  - repo: https://github.com/PyCQA/bandit
+    rev: 1.8.3
+    hooks:
+    - id: bandit
+      args: ["-c", "pyproject.toml"]
+      additional_dependencies: ["bandit[toml]"]

CONTRIBUTING.md

@@ -129,38 +129,71 @@ Follow these steps to start contributing:
 
    🚨 **Do not** work on the `main` branch.
 
-4. for development, we use `poetry` instead of just `pip` to easily track our dependencies.
-   If you don't have it already, follow the [instructions](https://python-poetry.org/docs/#installation) to install it.
+4. for development, we advise to use a tool like `poetry` or `uv` instead of just `pip` to easily track our dependencies.
+   Follow the instructions to [install poetry](https://python-poetry.org/docs/#installation) (use a version >=2.1.0) or to [install uv](https://docs.astral.sh/uv/getting-started/installation/#installation-methods) if you don't have one of them already.
 
    Set up a development environment with conda or miniconda:
    ```bash
    conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev
    ```
 
-   To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
+   If you're using `uv`, it can manage python versions so you can instead do:
    ```bash
-   poetry install --sync --extras "dev test"
+   uv venv --python 3.10 && source .venv/bin/activate
+   ```
+
+   To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
+
+   using `poetry`
+   ```bash
+   poetry sync --extras "dev test"
+   ```
+
+   using `uv`
+   ```bash
+   uv sync --extra dev --extra test
    ```
 
    You can also install the project with all its dependencies (including environments):
+
+   using `poetry`
    ```bash
-   poetry install --sync --all-extras
+   poetry sync --all-extras
+   ```
+
+   using `uv`
+   ```bash
+   uv sync --all-extras
    ```
 
    > **Note:** If you don't install simulation environments with `--all-extras`, the tests that require them will be skipped when running the pytest suite locally. However, they *will* be tested in the CI. In general, we advise you to install everything and test locally before pushing.
 
-   Whichever command you chose to install the project (e.g. `poetry 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.
+   Whichever command you chose to install the project (e.g. `poetry sync --all-extras`), you should run it again when pulling code with an updated version of `pyproject.toml` and `poetry.lock` in order to synchronize your virtual environment with the new dependencies.
 
    The equivalent of `pip install some-package`, would just be:
+
+   using `poetry`
    ```bash
    poetry add some-package
    ```
 
-   When making changes to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
+   using `uv`
    ```bash
-   poetry lock --no-update
+   uv add some-package
    ```
 
+   When making changes to the poetry sections of the `pyproject.toml`, you should run the following command to lock dependencies.
+   using `poetry`
+   ```bash
+   poetry lock
+   ```
+
+   using `uv`
+   ```bash
+   uv lock
+   ```
+
+
 5. Develop the features on your branch.
 
    As you work on the features, you should make sure that the test suite
@@ -195,7 +228,7 @@ Follow these steps to start contributing:
    git commit
    ```
 
-   Note, if you already commited some changes that have a wrong formatting, you can use:
+   Note, if you already committed some changes that have a wrong formatting, you can use:
    ```bash
    pre-commit run --all-files
    ```
@@ -267,7 +300,7 @@ We use `pytest` in order to run the tests. From the root of the
 repository, here's how to run tests with `pytest` for the library:
 
 ```bash
-DATA_DIR="tests/data" python -m pytest -sv ./tests
+python -m pytest -sv ./tests
 ```
 
 

Makefile

@@ -1,11 +1,25 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 .PHONY: tests
 
 PYTHON_PATH := $(shell which python)
 
-# If 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)
+# If uv is installed and a virtual environment exists, use it
+UV_CHECK := $(shell command -v uv)
+ifneq ($(UV_CHECK),)
+	PYTHON_PATH := $(shell .venv/bin/python)
 endif
 
 export PATH := $(dir $(PYTHON_PATH)):$(PATH)
@@ -20,171 +34,109 @@ build-gpu:
 
 test-end-to-end:
 	${MAKE} DEVICE=$(DEVICE) test-act-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train-resume
 	${MAKE} DEVICE=$(DEVICE) test-act-ete-eval
-	${MAKE} DEVICE=$(DEVICE) test-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=$(DEVICE) \
-		training.save_checkpoint=true \
-		training.save_freq=2 \
-		policy.n_action_steps=20 \
-		policy.chunk_size=20 \
-		training.batch_size=2 \
-		training.image_transforms.enable=true \
-		hydra.run.dir=tests/outputs/act/
+		--policy.type=act \
+		--policy.dim_model=64 \
+		--policy.n_action_steps=20 \
+		--policy.chunk_size=20 \
+		--policy.device=$(DEVICE) \
+		--env.type=aloha \
+		--env.episode_length=5 \
+		--dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
+		--dataset.image_transforms.enable=true \
+		--dataset.episodes="[0]" \
+		--batch_size=2 \
+		--steps=4 \
+		--eval_freq=2 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1 \
+		--save_freq=2 \
+		--save_checkpoint=true \
+		--log_freq=1 \
+		--wandb.enable=false \
+		--output_dir=tests/outputs/act/
+
+test-act-ete-train-resume:
+	python lerobot/scripts/train.py \
+		--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
+		--resume=true
 
 test-act-ete-eval:
 	python lerobot/scripts/eval.py \
-		-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=$(DEVICE) \
-		use_amp=true
+		--policy.path=tests/outputs/act/checkpoints/000004/pretrained_model \
+		--policy.device=$(DEVICE) \
+		--env.type=aloha \
+		--env.episode_length=5 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1
 
 test-diffusion-ete-train:
 	python lerobot/scripts/train.py \
-		policy=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=$(DEVICE) \
-		training.save_checkpoint=true \
-		training.save_freq=2 \
-		training.batch_size=2 \
-		training.image_transforms.enable=true \
-		hydra.run.dir=tests/outputs/diffusion/
+		--policy.type=diffusion \
+		--policy.down_dims='[64,128,256]' \
+		--policy.diffusion_step_embed_dim=32 \
+		--policy.num_inference_steps=10 \
+		--policy.device=$(DEVICE) \
+		--env.type=pusht \
+		--env.episode_length=5 \
+		--dataset.repo_id=lerobot/pusht \
+		--dataset.image_transforms.enable=true \
+		--dataset.episodes="[0]" \
+		--batch_size=2 \
+		--steps=2 \
+		--eval_freq=2 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1 \
+		--save_checkpoint=true \
+		--save_freq=2 \
+		--log_freq=1 \
+		--wandb.enable=false \
+		--output_dir=tests/outputs/diffusion/
 
 test-diffusion-ete-eval:
 	python lerobot/scripts/eval.py \
-		-p tests/outputs/diffusion/checkpoints/000002/pretrained_model \
-		eval.n_episodes=1 \
-		eval.batch_size=1 \
-		env.episode_length=8 \
-		device=$(DEVICE) \
+		--policy.path=tests/outputs/diffusion/checkpoints/000002/pretrained_model \
+		--policy.device=$(DEVICE) \
+		--env.type=pusht \
+		--env.episode_length=5 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1
 
 test-tdmpc-ete-train:
 	python lerobot/scripts/train.py \
-		policy=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/
-
+		--policy.type=tdmpc \
+		--policy.device=$(DEVICE) \
+		--env.type=xarm \
+		--env.task=XarmLift-v0 \
+		--env.episode_length=5 \
+		--dataset.repo_id=lerobot/xarm_lift_medium \
+		--dataset.image_transforms.enable=true \
+		--dataset.episodes="[0]" \
+		--batch_size=2 \
+		--steps=2 \
+		--eval_freq=2 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1 \
+		--save_checkpoint=true \
+		--save_freq=2 \
+		--log_freq=1 \
+		--wandb.enable=false \
+		--output_dir=tests/outputs/tdmpc/
 
 test-tdmpc-ete-eval:
 	python lerobot/scripts/eval.py \
-		-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=$(DEVICE) \
-		training.save_model=true \
-		training.save_freq=2 \
-		training.batch_size=2 \
-		training.image_transforms.enable=true \
-		hydra.run.dir=tests/outputs/act_pusht/
-	rm lerobot/configs/policy/created_by_Makefile.yaml
+		--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
+		--policy.device=$(DEVICE) \
+		--env.type=xarm \
+		--env.episode_length=5 \
+		--env.task=XarmLift-v0 \
+		--eval.n_episodes=1 \
+		--eval.batch_size=1

README.md

@@ -23,15 +23,24 @@
 </div>
 
 <h2 align="center">
-    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">New robot in town: SO-100</a></p>
+    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
+        Build Your Own SO-100 Robot!</a></p>
 </h2>
 
 <div align="center">
-    <img src="media/so100/leader_follower.webp?raw=true" alt="SO-100 leader and follower arms" title="SO-100 leader and follower arms" width="50%">
-    <p>We just added a new tutorial on how to build a more affordable robot, at the price of $110 per arm!</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>Follow the link to the <a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">full tutorial for SO-100</a>.</p>
+  <img src="media/so100/leader_follower.webp?raw=true" alt="SO-100 leader and follower arms" title="SO-100 leader and follower arms" width="50%">
+
+  <p><strong>Meet the SO-100 – Just $110 per arm!</strong></p>
+  <p>Train it in minutes with a few simple moves on your laptop.</p>
+  <p>Then sit back and watch your creation act autonomously! 🤯</p>
+
+  <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
+      Get the full SO-100 tutorial here.</a></p>
+
+  <p>Want to take it to the next level? Make your SO-100 mobile by building LeKiwi!</p>
+  <p>Check out the <a href="https://github.com/huggingface/lerobot/blob/main/examples/11_use_lekiwi.md">LeKiwi tutorial</a> and bring your robot to life on wheels.</p>
+
+  <img src="media/lekiwi/kiwi.webp?raw=true" alt="LeKiwi mobile robot" title="LeKiwi mobile robot" width="50%">
 </div>
 
 <br/>
@@ -68,7 +77,7 @@
 
 ### Acknowledgment
 
-- Thanks to Tony Zaho, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
+- Thanks to Tony Zhao, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
 - Thanks to Cheng Chi, Zhenjia Xu and colleagues for open sourcing Diffusion policy, Pusht environment and datasets, as well as UMI datasets. Ours are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu) and [UMI Gripper](https://umi-gripper.github.io).
 - Thanks to Nicklas Hansen, Yunhai Feng and colleagues for open sourcing TDMPC policy, Simxarm environments and datasets. Ours are adapted from [TDMPC](https://github.com/nicklashansen/tdmpc) and [FOWM](https://www.yunhaifeng.com/FOWM).
 - Thanks to Antonio Loquercio and Ashish Kumar for their early support.
@@ -122,10 +131,7 @@ wandb login
 ├── 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
-|   |   ├── env            # various sim environments and their datasets: aloha.yaml, pusht.yaml, xarm.yaml
-|   |   └── policy         # various policies: act.yaml, diffusion.yaml, tdmpc.yaml
+|   ├── configs          # contains config classes with all options that you can override in the command line
 |   ├── common           # contains classes and utilities
 |   |   ├── datasets       # various datasets of human demonstrations: aloha, pusht, xarm
 |   |   ├── envs           # various sim environments: aloha, pusht, xarm
@@ -153,10 +159,12 @@ python lerobot/scripts/visualize_dataset.py \
     --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`)
+or from a dataset in a local folder with the `root` option and the `--local-files-only` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
 ```bash
-DATA_DIR='./my_local_data_dir' python lerobot/scripts/visualize_dataset.py \
+python lerobot/scripts/visualize_dataset.py \
     --repo-id lerobot/pusht \
+    --root ./my_local_data_dir \
+    --local-files-only 1 \
     --episode-index 0
 ```
 
@@ -208,12 +216,10 @@ dataset attributes:
 
 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
+- videos are stored in mp4 format to save space
+- metadata are stored in plain json/jsonl files
 
-Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can set the `DATA_DIR` environment variable to your root dataset folder as illustrated in the above section on dataset visualization.
+Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
 
 ### Evaluate a pretrained policy
 
@@ -222,87 +228,48 @@ Check out [example 2](./examples/2_evaluate_pretrained_policy.py) that illustrat
 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 \
-    -p lerobot/diffusion_pusht \
-    eval.n_episodes=10 \
-    eval.batch_size=10
+    --policy.path=lerobot/diffusion_pusht \
+    --env.type=pusht \
+    --eval.batch_size=10 \
+    --eval.n_episodes=10 \
+    --use_amp=false \
+    --device=cuda
 ```
 
 Note: After training your own policy, you can re-evaluate the checkpoints with:
 
 ```bash
-python lerobot/scripts/eval.py -p {OUTPUT_DIR}/checkpoints/last/pretrained_model
+python lerobot/scripts/eval.py --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```
 
 See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-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.
+Check out [example 3](./examples/3_train_policy.py) that illustrate how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
 
-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:
+To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding `--wandb.enable=true`.
 
-```bash
-python lerobot/scripts/train.py \
-    policy=act \
-    env=aloha \
-    env.task=AlohaInsertion-v0 \
-    dataset_repo_id=lerobot/aloha_sim_insertion_human \
-```
-
-The experiment directory is automatically generated and will show up in yellow in your terminal. It looks like `outputs/train/2024-05-05/20-21-12_aloha_act_default`. You can manually specify an experiment directory by adding this argument to the `train.py` python command:
-```bash
-    hydra.run.dir=your/new/experiment/dir
-```
-
-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 explanation of some commonly used metrics in logs.
+A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser. Please also check [here](./examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explanation of some commonly used metrics in logs.
 
 ![](media/wandb.png)
 
-Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
+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:
-
+We provide some pretrained policies on our [hub page](https://huggingface.co/lerobot) that can achieve state-of-the-art performances.
+You can reproduce their training by loading the config from their run. Simply running:
 ```bash
-python lerobot/scripts/train.py policy=diffusion env=pusht
+python lerobot/scripts/train.py --config_path=lerobot/diffusion_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
 
 If you would like to contribute to 🤗 LeRobot, please check out our [contribution guide](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md).
 
-### Add a new dataset
+<!-- ### Add a new dataset
 
 To add a dataset to the hub, you need to login using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
 ```bash
@@ -320,7 +287,7 @@ python lerobot/scripts/push_dataset_to_hub.py \
 
 See `python lerobot/scripts/push_dataset_to_hub.py --help` for more instructions.
 
-If your dataset format is not supported, implement your own in `lerobot/common/datasets/push_dataset_to_hub/${raw_format}_format.py` by copying examples like [pusht_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py), [umi_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py), [aloha_hdf5](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py), or [xarm_pkl](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py).
+If your dataset format is not supported, implement your own in `lerobot/common/datasets/push_dataset_to_hub/${raw_format}_format.py` by copying examples like [pusht_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py), [umi_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py), [aloha_hdf5](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py), or [xarm_pkl](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py). -->
 
 
 ### Add a pretrained policy
@@ -330,7 +297,7 @@ Once you have trained a policy you may upload it to the Hugging Face hub using a
 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.
+- `train_config.json`: A consolidated configuration containing all parameter userd for training. The policy configuration should match `config.json` exactly. Thisis useful for anyone who wants to evaluate your policy or for reproducibility.
 
 To upload these to the hub, run the following:
 ```bash
@@ -417,3 +384,6 @@ Additionally, if you are using any of the particular policy architecture, pretra
   year={2024}
 }
 ```
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)

benchmarks/video/README.md

@@ -21,7 +21,7 @@ How to decode videos?
 
 ## 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).
+We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an apartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
 For these reasons, we run this benchmark on four representative datasets:
 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
 - `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
@@ -63,7 +63,7 @@ This of course is affected by the `-g` parameter during encoding, which specifie
 
 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:
+Additionally, because some policies might request single timestamps that are a few frames apart, we also have the following scenario:
 - `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
 
 However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
@@ -85,8 +85,8 @@ However, due to how video decoding is implemented with `pyav`, we don't have acc
 **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:
+One aspect that can't be measured here with those metrics is the compatibility of the encoding across platforms, in particular on web browser, for visualization purposes.
+h264, h265 and AV1 are all commonly used codecs and should not pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
 - `yuv420p` is more widely supported across various platforms, including web browsers.
 - `yuv444p` offers higher color fidelity but might not be supported as broadly.
 
@@ -114,9 +114,9 @@ We tried to measure the most impactful parameters for both encoding and decoding
 
 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.).
+- `-tune` which allows to optimize the encoding for certain 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.
+See the documentation mentioned above for more detailed info on these settings and for a more comprehensive list of other parameters.
 
 Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
 - `torchaudio`

benchmarks/video/run_video_benchmark.py

@@ -266,7 +266,7 @@ def benchmark_encoding_decoding(
         )
 
     ep_num_images = dataset.episode_data_index["to"][0].item()
-    width, height = tuple(dataset[0][dataset.camera_keys[0]].shape[-2:])
+    width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
     num_pixels = width * height
     video_size_bytes = video_path.stat().st_size
     images_size_bytes = get_directory_size(imgs_dir)

docker/lerobot-cpu/Dockerfile

@@ -1,32 +1,29 @@
 # Configure image
 ARG PYTHON_VERSION=3.10
-
 FROM python:${PYTHON_VERSION}-slim
+
+# Configure environment variables
 ARG PYTHON_VERSION
-ARG DEBIAN_FRONTEND=noninteractive
-
-# Install apt dependencies
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential cmake \
-    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
-    speech-dispatcher \
-    && apt-get clean && rm -rf /var/lib/apt/lists/*
-
-# Create virtual environment
-RUN ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python
-RUN python -m venv /opt/venv
+ENV DEBIAN_FRONTEND=noninteractive
+ENV MUJOCO_GL="egl"
 ENV PATH="/opt/venv/bin:$PATH"
-RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 
-# Install LeRobot
+# Install dependencies and set up Python in a single layer
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake git \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    speech-dispatcher libgeos-dev \
+    && ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python \
+    && python -m venv /opt/venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/* \
+    && echo "source /opt/venv/bin/activate" >> /root/.bashrc
+
+# Clone repository and install LeRobot in a single layer
 COPY . /lerobot
 WORKDIR /lerobot
-RUN pip install --upgrade --no-cache-dir pip
-RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]" \
-    --extra-index-url https://download.pytorch.org/whl/cpu
-
-# Set EGL as the rendering backend for MuJoCo
-ENV MUJOCO_GL="egl"
+RUN /opt/venv/bin/pip install --upgrade --no-cache-dir pip \
+    && /opt/venv/bin/pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]" \
+        --extra-index-url https://download.pytorch.org/whl/cpu
 
 # Execute in bash shell rather than python
 CMD ["/bin/bash"]

docker/lerobot-gpu-dev/Dockerfile

@@ -13,7 +13,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     sed gawk grep curl wget zip unzip \
     tcpdump sysstat screen tmux \
     libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
-    speech-dispatcher \
+    speech-dispatcher portaudio19-dev libgeos-dev \
     python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 

docker/lerobot-gpu/Dockerfile

@@ -1,30 +1,24 @@
 FROM nvidia/cuda:12.4.1-base-ubuntu22.04
 
-# Configure image
+# Configure environment variables
 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${PYTHON_VERSION} /usr/bin/python
-RUN python -m venv /opt/venv
+ENV DEBIAN_FRONTEND=noninteractive
+ENV MUJOCO_GL="egl"
 ENV PATH="/opt/venv/bin:$PATH"
-RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 
-# Install LeRobot
+# Install dependencies and set up Python in a single layer
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake git \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    speech-dispatcher libgeos-dev \
+    python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+    && ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python \
+    && python -m venv /opt/venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/* \
+    && echo "source /opt/venv/bin/activate" >> /root/.bashrc
+
+# Clone repository and install LeRobot in a single layer
 COPY . /lerobot
 WORKDIR /lerobot
-RUN pip install --upgrade --no-cache-dir pip
-RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]"
-
-# Set EGL as the rendering backend for MuJoCo
-ENV MUJOCO_GL="egl"
+RUN /opt/venv/bin/pip install --upgrade --no-cache-dir pip \
+    && /opt/venv/bin/pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]"

examples/10_use_so100.md

@@ -1,280 +0,0 @@
-This tutorial explains how to use [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot.
-
-## Source the parts
-
-Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with link to source the parts, as well as the instructions to 3D print the parts, and advices if it's your first time printing or if you don't own a 3D printer already.
-
-**Important**: Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
-
-## Install LeRobot
-
-On your computer:
-
-1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
-```bash
-mkdir -p ~/miniconda3
-wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
-bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
-rm ~/miniconda3/miniconda.sh
-~/miniconda3/bin/conda init bash
-```
-
-2. Restart shell or `source ~/.bashrc`
-
-3. Create and activate a fresh conda environment for lerobot
-```bash
-conda create -y -n lerobot python=3.10 && conda activate lerobot
-```
-
-4. Clone LeRobot:
-```bash
-git clone https://github.com/huggingface/lerobot.git ~/lerobot
-```
-
-5. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install -e ".[feetech]"
-```
-
-For Linux only (not Mac), install extra dependencies for recording datasets:
-```bash
-conda install -y -c conda-forge ffmpeg
-pip uninstall -y opencv-python
-conda install -y -c conda-forge "opencv>=4.10.0"
-```
-
-## Configure the motors
-
-Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the use of our scripts below.
-
-**Find USB ports associated to your arms**
-To find the correct ports for each arm, run the utility script twice:
-```bash
-python lerobot/scripts/find_motors_bus_port.py
-```
-
-Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
-
-[...Disconnect leader arm and press Enter...]
-
-The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
-Reconnect the usb cable.
-```
-
-Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
-
-[...Disconnect follower arm and press Enter...]
-
-The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
-Reconnect the usb cable.
-```
-
-Troubleshooting: On Linux, you might need to give access to the USB ports by running:
-```bash
-sudo chmod 666 /dev/ttyACM0
-sudo chmod 666 /dev/ttyACM1
-```
-
-**Configure your motors**
-Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 1
-```
-
-Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
-
-Then unplug your motor and plug the second motor and set its ID to 2.
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 2
-```
-
-Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
-
-**Remove the gears of the 6 leader motors**
-Follow step 2 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
-
-**Add motor horn to the motors**
-Follow step 3 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). For SO-100, you need to align the holes on the motor horn to the motor spline to be approximately 1:30, 4:30, 7:30 and 10:30.
-Try to avoid rotating the motor while doing so to keep position 2048 set during configuration. It is especially tricky for the leader motors as it is more sensible without the gears, but it's ok if it's a bit rotated.
-
-## Assemble the arms
-
-Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
-
-## Calibrate
-
-Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one SO-100 robot to work on another.
-
-**Manual calibration of follower arm**
-/!\ Contrarily to step 6 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
-
-You will need to move the follower arm to these positions sequentially:
-
-| 1. Zero position | 2. Rotated position | 3. Rest position |
-|---|---|---|
-| <img src="../media/so100/follower_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/so100/follower_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/so100/follower_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
-
-Make sure both arms are connected and run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py calibrate \
-    --robot-path lerobot/configs/robot/so100.yaml \
-    --robot-overrides '~cameras' --arms main_follower
-```
-
-**Manual calibration of leader arm**
-Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
-
-| 1. Zero position | 2. Rotated position | 3. Rest position |
-|---|---|---|
-| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
-
-Run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py calibrate \
-    --robot-path lerobot/configs/robot/so100.yaml \
-    --robot-overrides '~cameras' --arms main_leader
-```
-
-## Teleoperate
-
-**Simple teleop**
-Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
-```bash
-python lerobot/scripts/control_robot.py teleoperate \
-    --robot-path lerobot/configs/robot/so100.yaml \
-    --robot-overrides '~cameras' \
-    --display-cameras 0
-```
-
-
-**Teleop with displaying cameras**
-Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
-```bash
-python lerobot/scripts/control_robot.py teleoperate \
-    --robot-path lerobot/configs/robot/so100.yaml
-```
-
-## Record a dataset
-
-Once you're familiar with teleoperation, you can record your first dataset with SO-100.
-
-If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
-```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
-```
-
-Store your Hugging Face repository name in a variable to run these commands:
-```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
-echo $HF_USER
-```
-
-Record 2 episodes and upload your dataset to the hub:
-```bash
-python lerobot/scripts/control_robot.py record \
-    --robot-path lerobot/configs/robot/so100.yaml \
-    --fps 30 \
-    --root data \
-    --repo-id ${HF_USER}/so100_test \
-    --tags so100 tutorial \
-    --warmup-time-s 5 \
-    --episode-time-s 40 \
-    --reset-time-s 10 \
-    --num-episodes 2 \
-    --push-to-hub 1
-```
-
-## Visualize a dataset
-
-If you uploaded your dataset to the hub with `--push-to-hub 1`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
-```bash
-echo ${HF_USER}/so100_test
-```
-
-If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
-```bash
-python lerobot/scripts/visualize_dataset_html.py \
-  --root data \
-  --repo-id ${HF_USER}/so100_test
-```
-
-## Replay an episode
-
-Now try to replay the first episode on your robot:
-```bash
-DATA_DIR=data python lerobot/scripts/control_robot.py replay \
-    --robot-path lerobot/configs/robot/so100.yaml \
-    --fps 30 \
-    --root data \
-    --repo-id ${HF_USER}/so100_test \
-    --episode 0
-```
-
-## Train a policy
-
-To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-```bash
-DATA_DIR=data python lerobot/scripts/train.py \
-  dataset_repo_id=${HF_USER}/so100_test \
-  policy=act_so100_real \
-  env=so100_real \
-  hydra.run.dir=outputs/train/act_so100_test \
-  hydra.job.name=act_so100_test \
-  device=cuda \
-  wandb.enable=true
-```
-
-Let's explain it:
-1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/so100_test`.
-2. We provided the policy with `policy=act_so100_real`. This loads configurations from [`lerobot/configs/policy/act_so100_real.yaml`](../lerobot/configs/policy/act_so100_real.yaml). Importantly, this policy uses 2 cameras as input `laptop`, `phone`.
-3. We provided an environment as argument with `env=so100_real`. This loads configurations from [`lerobot/configs/env/so100_real.yaml`](../lerobot/configs/env/so100_real.yaml).
-4. We provided `device=cuda` since we are training on a Nvidia GPU, but you can also use `device=mps` if you are using a Mac with Apple silicon, or `device=cpu` otherwise.
-5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$HOME/.cache/hugginface`. In future versions of `lerobot`, both directories will be in sync.
-
-Training should take several hours. You will find checkpoints in `outputs/train/act_so100_test/checkpoints`.
-
-## Evaluate your policy
-
-You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
-```bash
-python lerobot/scripts/control_robot.py record \
-  --robot-path lerobot/configs/robot/so100.yaml \
-  --fps 30 \
-  --root data \
-  --repo-id ${HF_USER}/eval_act_so100_test \
-  --tags so100 tutorial eval \
-  --warmup-time-s 5 \
-  --episode-time-s 40 \
-  --reset-time-s 10 \
-  --num-episodes 10 \
-  -p outputs/train/act_so100_test/checkpoints/last/pretrained_model
-```
-
-As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_so100_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_so100_test`).
-
-## More
-
-Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
-
-If you have any question or need help, please reach out on Discord in the channel [`#so100-arm`](https://discord.com/channels/1216765309076115607/1237741463832363039).

examples/1_load_lerobot_dataset.py

@@ -1,80 +1,136 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 """
 This script demonstrates the use of `LeRobotDataset` class for handling and processing robotic datasets from Hugging Face.
 It illustrates how to load datasets, manipulate them, and apply transformations suitable for machine learning tasks in PyTorch.
 
 Features included in this script:
-- Loading a dataset and accessing its properties.
-- Filtering data by episode number.
-- Converting tensor data for visualization.
-- Saving video files from dataset frames.
+- Viewing a dataset's metadata and exploring its properties.
+- Loading an existing dataset from the hub or a subset of it.
+- Accessing frames by episode number.
 - Using advanced dataset features like timestamp-based frame selection.
 - Demonstrating compatibility with PyTorch DataLoader for batch processing.
 
 The script ends with examples of how to batch process data using PyTorch's DataLoader.
 """
 
-from pathlib import Path
 from pprint import pprint
 
-import imageio
 import torch
+from huggingface_hub import HfApi
 
 import lerobot
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 
+# We ported a number of existing datasets ourselves, use this to see the list:
 print("List of available datasets:")
 pprint(lerobot.available_datasets)
 
-# Let's take one for this example
-repo_id = "lerobot/pusht"
+# You can also browse through the datasets created/ported by the community on the hub using the hub api:
+hub_api = HfApi()
+repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
+pprint(repo_ids)
 
-# You can easily load a dataset from a Hugging Face repository
+# Or simply explore them in your web browser directly at:
+# https://huggingface.co/datasets?other=LeRobot
+
+# Let's take this one for this example
+repo_id = "lerobot/aloha_mobile_cabinet"
+# We can have a look and fetch its metadata to know more about it:
+ds_meta = LeRobotDatasetMetadata(repo_id)
+
+# By instantiating just this class, you can quickly access useful information about the content and the
+# structure of the dataset without downloading the actual data yet (only metadata files — which are
+# lightweight).
+print(f"Total number of episodes: {ds_meta.total_episodes}")
+print(f"Average number of frames per episode: {ds_meta.total_frames / ds_meta.total_episodes:.3f}")
+print(f"Frames per second used during data collection: {ds_meta.fps}")
+print(f"Robot type: {ds_meta.robot_type}")
+print(f"keys to access images from cameras: {ds_meta.camera_keys=}\n")
+
+print("Tasks:")
+print(ds_meta.tasks)
+print("Features:")
+pprint(ds_meta.features)
+
+# You can also get a short summary by simply printing the object:
+print(ds_meta)
+
+# You can then load the actual dataset from the hub.
+# Either load any subset of episodes:
+dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
+
+# And see how many frames you have:
+print(f"Selected episodes: {dataset.episodes}")
+print(f"Number of episodes selected: {dataset.num_episodes}")
+print(f"Number of frames selected: {dataset.num_frames}")
+
+# Or simply load the entire dataset:
 dataset = LeRobotDataset(repo_id)
+print(f"Number of episodes selected: {dataset.num_episodes}")
+print(f"Number of frames selected: {dataset.num_frames}")
 
-# LeRobotDataset is actually a thin wrapper around an underlying Hugging Face dataset
-# (see https://huggingface.co/docs/datasets/index for more information).
-print(dataset)
+# The previous metadata class is contained in the 'meta' attribute of the dataset:
+print(dataset.meta)
+
+# LeRobotDataset actually wraps an underlying Hugging Face dataset
+# (see https://huggingface.co/docs/datasets for more information).
 print(dataset.hf_dataset)
 
-# 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
+# LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
+# with the latter, like iterating through the dataset.
+# The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
+# episodes, you can access the frame indices of any episode using the episode_data_index. Here, we access
+# frame indices associated to the first episode:
 episode_index = 0
 from_idx = dataset.episode_data_index["from"][episode_index].item()
 to_idx = dataset.episode_data_index["to"][episode_index].item()
 
-# 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)]
+# Then we grab all the image frames from the first camera:
+camera_key = dataset.meta.camera_keys[0]
+frames = [dataset[idx][camera_key] for idx in range(from_idx, to_idx)]
 
-# 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]
+# The objects returned by the dataset are all torch.Tensors
+print(type(frames[0]))
+print(frames[0].shape)
 
-# 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)
+# Since we're using pytorch, the shape is in pytorch, channel-first convention (c, h, w).
+# We can compare this shape with the information available for that feature
+pprint(dataset.features[camera_key])
+# In particular:
+print(dataset.features[camera_key]["shape"])
+# The shape is in (h, w, c) which is a more universal format.
 
 # For many machine learning applications we need to load the history of past observations or trajectories of
 # future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
 # differences with the current loaded frame. For instance:
 delta_timestamps = {
     # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
-    "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],
+    camera_key: [-1, -0.5, -0.20, 0],
+    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
+    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
     # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
     "action": [t / dataset.fps for t in range(64)],
 }
+# Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
+# timestamp, you still get a valid timestamp.
+
 dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
-print(f"\n{dataset[0]['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)
+print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
+print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
+print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
 
 # Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
 # PyTorch datasets.
@@ -84,8 +140,9 @@ dataloader = torch.utils.data.DataLoader(
     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)
+    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
+    print(f"{batch['observation.state'].shape=}")  # (32, 5, c)
+    print(f"{batch['action'].shape=}")  # (32, 64, c)
     break

examples/2_evaluate_pretrained_policy.py

@@ -1,6 +1,25 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 """
 This scripts demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
 training outputs directory. In the latter case, you might want to run examples/3_train_policy.py first.
+
+It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
+```bash
+pip install -e ".[pusht]"`
+```
 """
 
 from pathlib import Path
@@ -10,7 +29,6 @@ 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
 
@@ -18,25 +36,15 @@ from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 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.
+# Select your device
+device = "cuda"
+
+# Provide the [hugging face repo id](https://huggingface.co/lerobot/diffusion_pusht):
+pretrained_policy_path = "lerobot/diffusion_pusht"
+# OR a path to a local outputs/train folder.
 # pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
 
 policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
-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
@@ -47,7 +55,17 @@ env = gym.make(
     max_episode_steps=300,
 )
 
-# Reset the policy and environmens to prepare for rollout
+# We can verify that the shapes of the features expected by the policy match the ones from the observations
+# produced by the environment
+print(policy.config.input_features)
+print(env.observation_space)
+
+# Similarly, we can check that the actions produced by the policy will match the actions expected by the
+# environment
+print(policy.config.output_features)
+print(env.action_space)
+
+# Reset the policy and environments to prepare for rollout
 policy.reset()
 numpy_observation, info = env.reset(seed=42)
 

examples/3_train_policy.py

@@ -1,3 +1,17 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 """This scripts demonstrates how to train Diffusion Policy on the PushT environment.
 
 Once you have trained a model with this script, you can try to evaluate it on
@@ -8,72 +22,99 @@ from pathlib import Path
 
 import torch
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.datasets.utils import dataset_to_policy_features
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+from lerobot.configs.types import FeatureType
 
-# Create a directory to store the training checkpoint.
-output_directory = Path("outputs/train/example_pusht_diffusion")
-output_directory.mkdir(parents=True, exist_ok=True)
 
-# Number of offline training steps (we'll only do offline training for this example.)
-# Adjust as you prefer. 5000 steps are needed to get something worth evaluating.
-training_steps = 5000
-device = torch.device("cuda")
-log_freq = 250
+def main():
+    # Create a directory to store the training checkpoint.
+    output_directory = Path("outputs/train/example_pusht_diffusion")
+    output_directory.mkdir(parents=True, exist_ok=True)
 
-# Set up the dataset.
-delta_timestamps = {
-    # Load the previous image and state at -0.1 seconds before current frame,
-    # then load current image and state corresponding to 0.0 second.
-    "observation.image": [-0.1, 0.0],
-    "observation.state": [-0.1, 0.0],
-    # Load the previous action (-0.1), the next action to be executed (0.0),
-    # and 14 future actions with a 0.1 seconds spacing. All these actions will be
-    # used to supervise the policy.
-    "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
-}
-dataset = LeRobotDataset("lerobot/pusht", delta_timestamps=delta_timestamps)
+    # # Select your device
+    device = torch.device("cuda")
 
-# Set up the the policy.
-# Policies are initialized with a configuration class, in this case `DiffusionConfig`.
-# For this example, no arguments need to be passed because the defaults are set up for PushT.
-# If you're doing something different, you will likely need to change at least some of the defaults.
-cfg = DiffusionConfig()
-policy = DiffusionPolicy(cfg, dataset_stats=dataset.stats)
-policy.train()
-policy.to(device)
+    # Number of offline training steps (we'll only do offline training for this example.)
+    # Adjust as you prefer. 5000 steps are needed to get something worth evaluating.
+    training_steps = 5000
+    log_freq = 1
 
-optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
+    # When starting from scratch (i.e. not from a pretrained policy), we need to specify 2 things before
+    # creating the policy:
+    #   - input/output shapes: to properly size the policy
+    #   - dataset stats: for normalization and denormalization of input/outputs
+    dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
+    features = dataset_to_policy_features(dataset_metadata.features)
+    output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    input_features = {key: ft for key, ft in features.items() if key not in output_features}
 
-# Create dataloader for offline training.
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    num_workers=4,
-    batch_size=64,
-    shuffle=True,
-    pin_memory=device != torch.device("cpu"),
-    drop_last=True,
-)
+    # Policies are initialized with a configuration class, in this case `DiffusionConfig`. For this example,
+    # we'll just use the defaults and so no arguments other than input/output features need to be passed.
+    cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
 
-# Run training loop.
-step = 0
-done = False
-while not done:
-    for batch in dataloader:
-        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
-        output_dict = policy.forward(batch)
-        loss = output_dict["loss"]
-        loss.backward()
-        optimizer.step()
-        optimizer.zero_grad()
+    # We can now instantiate our policy with this config and the dataset stats.
+    policy = DiffusionPolicy(cfg, dataset_stats=dataset_metadata.stats)
+    policy.train()
+    policy.to(device)
 
-        if step % log_freq == 0:
-            print(f"step: {step} loss: {loss.item():.3f}")
-        step += 1
-        if step >= training_steps:
-            done = True
-            break
+    # Another policy-dataset interaction is with the delta_timestamps. Each policy expects a given number frames
+    # which can differ for inputs, outputs and rewards (if there are some).
+    delta_timestamps = {
+        "observation.image": [i / dataset_metadata.fps for i in cfg.observation_delta_indices],
+        "observation.state": [i / dataset_metadata.fps for i in cfg.observation_delta_indices],
+        "action": [i / dataset_metadata.fps for i in cfg.action_delta_indices],
+    }
 
-# Save a policy checkpoint.
-policy.save_pretrained(output_directory)
+    # In this case with the standard configuration for Diffusion Policy, it is equivalent to this:
+    delta_timestamps = {
+        # Load the previous image and state at -0.1 seconds before current frame,
+        # then load current image and state corresponding to 0.0 second.
+        "observation.image": [-0.1, 0.0],
+        "observation.state": [-0.1, 0.0],
+        # Load the previous action (-0.1), the next action to be executed (0.0),
+        # and 14 future actions with a 0.1 seconds spacing. All these actions will be
+        # used to supervise the policy.
+        "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
+    }
+
+    # We can then instantiate the dataset with these delta_timestamps configuration.
+    dataset = LeRobotDataset("lerobot/pusht", delta_timestamps=delta_timestamps)
+
+    # Then we create our optimizer and dataloader for offline training.
+    optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=4,
+        batch_size=64,
+        shuffle=True,
+        pin_memory=device.type != "cpu",
+        drop_last=True,
+    )
+
+    # Run training loop.
+    step = 0
+    done = False
+    while not done:
+        for batch in dataloader:
+            batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
+            loss, _ = policy.forward(batch)
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+
+            if step % log_freq == 0:
+                print(f"step: {step} loss: {loss.item():.3f}")
+            step += 1
+            if step >= training_steps:
+                done = True
+                break
+
+    # Save a policy checkpoint.
+    policy.save_pretrained(output_directory)
+
+
+if __name__ == "__main__":
+    main()

examples/4_train_policy_with_script.md

@@ -1,193 +1,223 @@
-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.
+This tutorial will explain the training script, how to use it, and particularly how to configure everything needed for the training run.
+> **Note:** The following assume you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--device=cpu` (`--device=mps` respectively). However, be advised that the code executes much slower on cpu.
+
 
 ## The training script
 
 LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
 
-- 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.
+- Initialize/load a configuration for the following steps using.
+- Instantiates a dataset.
+- (Optional) Instantiates a simulation environment corresponding to that dataset.
+- Instantiates a policy.
 - Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
 
-## 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
+## Overview of the configuration system
 
+In the training script, the main function `train` expects a `TrainPipelineConfig` object:
 ```python
-python lerobot/scripts/train.py
+# train.py
+@parser.wrap()
+def train(cfg: TrainPipelineConfig):
 ```
 
-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:
+You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
 
-```yaml
-defaults:
-  - _self_
-  - env: pusht
-  - policy: diffusion
+When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated for this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
+
+Let's have a look at a simplified example. Amongst other attributes, the training config has the following attributes:
+```python
+@dataclass
+class TrainPipelineConfig:
+    dataset: DatasetConfig
+    env: envs.EnvConfig | None = None
+    policy: PreTrainedConfig | None = None
+```
+in which `DatasetConfig` for example is defined as such:
+```python
+@dataclass
+class DatasetConfig:
+    repo_id: str
+    episodes: list[int] | None = None
+    video_backend: str = "pyav"
 ```
 
-This 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`_.
+This creates a hierarchical relationship where, for example assuming we have a `cfg` instance of `TrainPipelineConfig`, we can access the `repo_id` value with `cfg.dataset.repo_id`.
+From the command line, we can specify this value with using a very similar syntax `--dataset.repo_id=repo/id`.
 
-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`.
+By default, every field takes its default value specified in the dataclass. If a field doesn't have a default value, it needs to be specified either from the command line or from a config file – which path is also given in the command line (more in this below). In the example above, the `dataset` field doesn't have a default value which means it must be specified.
 
-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:
+## Specifying values from the CLI
 
+Let's say that we want to train [Diffusion Policy](../../lerobot/common/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
 ```bash
 python lerobot/scripts/train.py \
-    policy=act \
-    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
-    env=aloha \
-    env.task=AlohaTransferCube-v0
+    --dataset.repo_id=lerobot/pusht \
+    --policy.type=diffusion \
+    --env.type=pusht
 ```
 
-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.
+Let's break this down:
+- To specify the dataset, we just need to specify its `repo_id` on the hub which is the only required argument in the `DatasetConfig`. The rest of the fields have default values and in this case we are fine with those so we can just add the option `--dataset.repo_id=lerobot/pusht`.
+- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../../lerobot/common/policies)
+- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../../lerobot/common/envs/configs.py)
 
+Let's see another example. Let's say you've been training [ACT](../../lerobot/common/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
 ```bash
 python lerobot/scripts/train.py \
-    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 \
+    --policy.type=act \
+    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
+    --env.type=aloha \
+    --output_dir=outputs/train/act_aloha_insertion
 ```
+> Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.
 
-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:
-
+We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
+Looking at the [`AlohaEnv`](../../lerobot/common/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
 ```bash
-python lerobot/scripts/train.py --config-dir PARENT/PATH --config-name FILE_NAME_WITHOUT_EXTENSION
+python lerobot/scripts/train.py \
+    --policy.type=act \
+    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
+    --env.type=aloha \
+    --env.task=AlohaTransferCube-v0 \
+    --output_dir=outputs/train/act_aloha_transfer
 ```
 
-Note: here we use regular syntax for providing CLI arguments to a Python script, not Hydra's `param_name=param_value` syntax.
+## Loading from a config file
 
-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:
+Now, let's assume that we want to reproduce the run just above. That run has produced a `train_config.json` file in its checkpoints, which serializes the `TrainPipelineConfig` instance it used:
+```json
+{
+    "dataset": {
+        "repo_id": "lerobot/aloha_sim_transfer_cube_human",
+        "episodes": null,
+        ...
+    },
+    "env": {
+        "type": "aloha",
+        "task": "AlohaTransferCube-v0",
+        "fps": 50,
+        ...
+    },
+    "policy": {
+        "type": "act",
+        "n_obs_steps": 1,
+        ...
+    },
+    ...
+}
+```
 
+We can then simply load the config values from this file using:
 ```bash
-python lerobot/scripts/train.py --config-dir outputs/train/my_experiment/checkpoints/last/pretrained_model --config-name config
+python lerobot/scripts/train.py \
+    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
+    --output_dir=outputs/train/act_aloha_transfer_2
+```
+`--config_path` is also a special argument which allows to initialize the config from a local config file. It can point to a directory that contains `train_config.json` or to the config file itself directly.
+
+Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
+```bash
+python lerobot/scripts/train.py \
+    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
+    --output_dir=outputs/train/act_aloha_transfer_2
+    --policy.n_action_steps=80
+```
+> Note: While `--output_dir` is not required in general, in this case we need to specify it since it will otherwise take the value from the `train_config.json` (which is `outputs/train/act_aloha_transfer`). In order to prevent accidental deletion of previous run checkpoints, we raise an error if you're trying to write in an existing directory. This is not the case when resuming a run, which is what you'll learn next.
+
+`--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:
+```bash
+python lerobot/scripts/train.py --config_path=lerobot/diffusion_pusht
+```
+will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
+
+
+## Resume training
+
+Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to that here.
+
+Let's reuse the command from the previous run and add a few more options:
+```bash
+python lerobot/scripts/train.py \
+    --policy.type=act \
+    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
+    --env.type=aloha \
+    --env.task=AlohaTransferCube-v0 \
+    --log_freq=25 \
+    --save_freq=100 \
+    --output_dir=outputs/train/run_resumption
 ```
 
-Note that you may still use the regular syntax for config parameter overrides (eg: by adding `training.offline_steps=200000`).
+Here we've taken care to set up the log frequency and checkpointing frequency to low numbers so we can showcase resumption. You should be able to see some logging and have a first checkpoint within 1 minute (depending on hardware). Wait for the first checkpoint to happen, you should see a line that looks like this in your terminal:
+```
+INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
+```
+Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:
+```bash
+python lerobot/scripts/train.py \
+    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
+    --resume=true
+```
+You should see from the logging that your training picks up from where it left off.
+
+Another reason for which you might want to resume a run is simply to extend training and add more training steps. The number of training steps is set by the option `--steps`, which is 100 000 by default.
+You could double the number of steps of the previous run with:
+```bash
+python lerobot/scripts/train.py \
+    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
+    --resume=true \
+    --steps=200000
+```
+
+## Outputs of a run
+In the output directory, there will be a folder called `checkpoints` with the following structure:
+```bash
+outputs/train/run_resumption/checkpoints
+├── 000100  # checkpoint_dir for training step 100
+│   ├── pretrained_model/
+│   │   ├── config.json  # policy config
+│   │   ├── model.safetensors  # policy weights
+│   │   └── train_config.json  # train config
+│   └── training_state/
+│       ├── optimizer_param_groups.json  #  optimizer param groups
+│       ├── optimizer_state.safetensors  # optimizer state
+│       ├── rng_state.safetensors  # rng states
+│       ├── scheduler_state.json  # scheduler state
+│       └── training_step.json  # training step
+├── 000200
+└── last -> 000200  # symlink to the last available checkpoint
+```
+
+## Fine-tuning a pre-trained policy
+
+In addition to the features currently in Draccus, we've added a special `.path` argument for the policy, which allows to load a policy as you would with `PreTrainedPolicy.from_pretrained()`. In that case, `path` can be a local directory that contains a checkpoint or a repo_id pointing to a pretrained policy on the hub.
+
+For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:
+```bash
+python lerobot/scripts/train.py \
+    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \
+    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
+    --env.type=aloha \
+    --env.task=AlohaInsertion-v0
+```
+
+When doing so, keep in mind that the features of the fine-tuning dataset would have to match the input/output features of the pretrained policy.
 
 ## Typical logs and metrics
 
-When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
+When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you configured your run correctly. The final configuration will also be saved with the checkpoint.
 
 After that, you will see training log like this one:
-
 ```
 INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
 ```
-
-or evaluation log like:
-
+or evaluation log:
 ```
 INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
 ```
 
 These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
-
 - `smpl`: number of samples seen during training.
 - `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
 - `epch`: number of time all unique samples are seen (epoch).
@@ -200,14 +230,45 @@ These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here
 
 Some metrics are useful for initial performance profiling. For example, if you find the current GPU utilization is low via the `nvidia-smi` command and `data_s` sometimes is too high, you may need to modify batch size or number of dataloading workers to accelerate dataloading. We also recommend [pytorch profiler](https://github.com/huggingface/lerobot?tab=readme-ov-file#improve-your-code-with-profiling) for detailed performance probing.
 
----
+## In short
 
-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):
+We'll summarize here the main use cases to remember from this tutorial.
 
+#### Train a policy from scratch – CLI
 ```bash
-python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
+python lerobot/scripts/train.py \
+    --policy.type=act \  # <- select 'act' policy
+    --env.type=pusht \  # <- select 'pusht' environment
+    --dataset.repo_id=lerobot/pusht  # <- train on this dataset
 ```
 
-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.
+#### Train a policy from scratch - config file + CLI
+```bash
+python lerobot/scripts/train.py \
+    --config_path=path/to/pretrained_model \  # <- can also be a repo_id
+    --policy.n_action_steps=80  # <- you may still override values
+```
 
-Or in the meantime, happy coding! 🤗
+#### Resume/continue a training run
+```bash
+python lerobot/scripts/train.py \
+    --config_path=checkpoint/pretrained_model/ \
+    --resume=true \
+    --steps=200000  # <- you can change some training parameters
+```
+
+#### Fine-tuning
+```bash
+python lerobot/scripts/train.py \
+    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \  # <- can also be a local path to a checkpoint
+    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
+    --env.type=aloha \
+    --env.task=AlohaInsertion-v0
+```
+
+---
+
+Now that you know the basics of how to train a policy, you might want to know how to apply this knowledge to actual robots, or how to record your own datasets and train policies on your specific task?
+If that's the case, head over to the next tutorial [`7_get_started_with_real_robot.md`](./7_get_started_with_real_robot.md).
+
+Or in the meantime, happy training! 🤗

examples/5_resume_training.md

@@ -1,37 +0,0 @@
-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/7_get_started_with_real_robot.md

@@ -29,16 +29,21 @@ For a visual walkthrough of the assembly process, you can refer to [this video t
 
 ## 2. Configure motors, calibrate arms, teleoperate your Koch v1.1
 
-First, install the additional dependencies required for robots built with dynamixel motors like Koch v1.1 by running one of the following commands.
+First, install the additional dependencies required for robots built with dynamixel motors like Koch v1.1 by running one of the following commands (make sure gcc is installed).
 
 Using `pip`:
 ```bash
 pip install -e ".[dynamixel]"
 ```
 
-Or using `poetry`:
+Using `poetry`:
 ```bash
-poetry install --sync --extras "dynamixel"
+poetry sync --extras "dynamixel"
+```
+
+Using `uv`:
+```bash
+uv sync --extra "dynamixel"
 ```
 
 /!\ For Linux only, ffmpeg and opencv requires conda install for now. Run this exact sequence of commands:
@@ -54,24 +59,53 @@ Then plug the 12V power supply to the motor bus of the follower arm. It has two
 
 Finally, connect both arms to your computer via USB. Note that the USB doesn't provide any power, and both arms need to be plugged in with their associated power supply to be detected by your computer.
 
-*Copy pasting python code*
+Now you are ready to configure your motors for the first time, as detailed in the sections below. In the upcoming sections, you'll learn about our classes and functions by running some python code in an interactive session, or by copy-pasting it in a python file.
 
-In the upcoming sections, you'll learn about our classes and functions by running some python code, in an interactive session, or by copy-pasting it in a python file. If this is your first time using the tutorial., we highly recommend going through these steps to get deeper intuition about how things work. Once you're more familiar, you can streamline the process by directly running the teleoperate script (which is detailed further in the tutorial):
+If you have already configured your motors the first time, you can streamline the process by directly running the teleoperate script (which is detailed further in the tutorial):
 ```bash
-python lerobot/scripts/control_robot.py teleoperate \
-  --robot-path lerobot/configs/robot/koch.yaml \
-  --robot-overrides '~cameras'  # do not instantiate the cameras
+python lerobot/scripts/control_robot.py \
+  --robot.type=koch \
+  --control.type=teleoperate
 ```
 
 It will automatically:
-1. Detect and help you correct any motor configuration issues.
-2. Identify any missing calibrations and initiate the calibration procedure.
-3. Connect the robot and start teleoperation.
+1. Identify any missing calibrations and initiate the calibration procedure.
+2. Connect the robot and start teleoperation.
 
 ### a. Control your motors with DynamixelMotorsBus
 
 You can use the [`DynamixelMotorsBus`](../lerobot/common/robot_devices/motors/dynamixel.py) to communicate with the motors connected as a chain to the corresponding USB bus. This class leverages the Python [Dynamixel SDK](https://emanual.robotis.com/docs/en/software/dynamixel/dynamixel_sdk/sample_code/python_read_write_protocol_2_0/#python-read-write-protocol-20) to facilitate reading from and writing to the motors.
 
+**First Configuration of your motors**
+
+You will need to unplug each motor in turn and run a command the identify the motor. The motor will save its own identification, so you only need to do this once. Start by unplugging all of the motors.
+
+Do the Leader arm first, as all of its motors are of the same type. Plug in your first motor on your leader arm and run this script to set its ID to 1.
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand dynamixel \
+  --model xl330-m288 \
+  --baudrate 1000000 \
+  --ID 1
+```
+
+Then unplug your first motor and plug the second motor and set its ID to 2.
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand dynamixel \
+  --model xl330-m288 \
+  --baudrate 1000000 \
+  --ID 2
+```
+
+Redo the process for all your motors until ID 6.
+
+The process for the follower arm is almost the same, but the follower arm has two types of motors. For the first two motors, make sure you set the model to `xl430-w250`. _Important: configuring follower motors requires plugging and unplugging power. Make sure you use the 5V power for the XL330s and the 12V power for the XL430s!_
+
+After all of your motors are configured properly, you're ready to plug them all together in a daisy-chain as shown in the original video.
+
 **Instantiate the DynamixelMotorsBus**
 
 To begin, create two instances of the  [`DynamixelMotorsBus`](../lerobot/common/robot_devices/motors/dynamixel.py), one for each arm, using their corresponding USB ports (e.g. `DynamixelMotorsBus(port="/dev/tty.usbmodem575E0031751"`).
@@ -105,10 +139,10 @@ The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
 Reconnect the usb cable.
 ```
 
-Troubleshooting: On Linux, you might need to give access to the USB ports by running:
+Troubleshooting: On Linux, you might need to give access to the USB ports by running this command with your ports:
 ```bash
-sudo chmod 666 /dev/ttyACM0
-sudo chmod 666 /dev/ttyACM1
+sudo chmod 666 /dev/tty.usbmodem575E0032081
+sudo chmod 666 /dev/tty.usbmodem575E0031751
 ```
 
 *Listing and Configuring Motors*
@@ -117,13 +151,11 @@ Next, you'll need to list the motors for each arm, including their name, index,
 
 To assign indices to the motors, run this code in an interactive Python session. Replace the `port` values with the ones you identified earlier:
 ```python
+from lerobot.common.robot_devices.motors.configs import DynamixelMotorsBusConfig
 from lerobot.common.robot_devices.motors.dynamixel import DynamixelMotorsBus
 
-leader_port = "/dev/tty.usbmodem575E0031751"
-follower_port = "/dev/tty.usbmodem575E0032081"
-
-leader_arm = DynamixelMotorsBus(
-    port=leader_port,
+leader_config = DynamixelMotorsBusConfig(
+    port="/dev/tty.usbmodem575E0031751",
     motors={
         # name: (index, model)
         "shoulder_pan": (1, "xl330-m077"),
@@ -135,8 +167,8 @@ leader_arm = DynamixelMotorsBus(
     },
 )
 
-follower_arm = DynamixelMotorsBus(
-    port=follower_port,
+follower_config = DynamixelMotorsBusConfig(
+    port="/dev/tty.usbmodem575E0032081",
     motors={
         # name: (index, model)
         "shoulder_pan": (1, "xl430-w250"),
@@ -147,45 +179,57 @@ follower_arm = DynamixelMotorsBus(
         "gripper": (6, "xl330-m288"),
     },
 )
+
+leader_arm = DynamixelMotorsBus(leader_config)
+follower_arm = DynamixelMotorsBus(follower_config)
 ```
 
-*Updating the YAML Configuration File*
+IMPORTANTLY: Now that you have your ports, update [`KochRobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
+```python
+@RobotConfig.register_subclass("koch")
+@dataclass
+class KochRobotConfig(ManipulatorRobotConfig):
+    calibration_dir: str = ".cache/calibration/koch"
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
 
-Next, update the port values in the YAML configuration file for the Koch robot at [`lerobot/configs/robot/koch.yaml`](../lerobot/configs/robot/koch.yaml) with the ports you've identified:
-```yaml
-[...]
-robot_type: koch
-leader_arms:
-  main:
-    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
-    port: /dev/tty.usbmodem575E0031751  # <- Update
-    motors:
-      # name: (index, model)
-      shoulder_pan: [1, "xl330-m077"]
-      shoulder_lift: [2, "xl330-m077"]
-      elbow_flex: [3, "xl330-m077"]
-      wrist_flex: [4, "xl330-m077"]
-      wrist_roll: [5, "xl330-m077"]
-      gripper: [6, "xl330-m077"]
-follower_arms:
-  main:
-    _target_: lerobot.common.robot_devices.motors.dynamixel.DynamixelMotorsBus
-    port: /dev/tty.usbmodem575E0032081  # <- Update
-    motors:
-      # name: (index, model)
-      shoulder_pan: [1, "xl430-w250"]
-      shoulder_lift: [2, "xl430-w250"]
-      elbow_flex: [3, "xl330-m288"]
-      wrist_flex: [4, "xl330-m288"]
-      wrist_roll: [5, "xl330-m288"]
-      gripper: [6, "xl330-m288"]
-[...]
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": DynamixelMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0085511", <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "xl330-m077"],
+                    "shoulder_lift": [2, "xl330-m077"],
+                    "elbow_flex": [3, "xl330-m077"],
+                    "wrist_flex": [4, "xl330-m077"],
+                    "wrist_roll": [5, "xl330-m077"],
+                    "gripper": [6, "xl330-m077"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": DynamixelMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0076891", <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "xl430-w250"],
+                    "shoulder_lift": [2, "xl430-w250"],
+                    "elbow_flex": [3, "xl330-m288"],
+                    "wrist_flex": [4, "xl330-m288"],
+                    "wrist_roll": [5, "xl330-m288"],
+                    "gripper": [6, "xl330-m288"],
+                },
+            ),
+        }
+    )
 ```
 
-Don't forget to set `robot_type: aloha` if you follow this tutorial with [Aloha bimanual robot](aloha-2.github.io) instead of Koch v1.1
-
-This configuration file is used to instantiate your robot across all scripts. We'll cover how this works later on.
-
 **Connect and Configure your Motors**
 
 Before you can start using your motors, you'll need to configure them to ensure proper communication. When you first connect the motors, the [`DynamixelMotorsBus`](../lerobot/common/robot_devices/motors/dynamixel.py) automatically detects any mismatch between the current motor indices (factory set to `1`) and the specified indices (e.g., `1, 2, 3, 4, 5, 6`). This triggers a configuration procedure that requires you to unplug the power cord and motors, then reconnect each motor sequentially, starting from the one closest to the bus.
@@ -312,27 +356,27 @@ Alternatively, you can unplug the power cord, which will automatically disable t
 
 **Instantiate the ManipulatorRobot**
 
-Before you can teleoperate your robot, you need to instantiate the  [`ManipulatorRobot`](../lerobot/common/robot_devices/robots/manipulator.py) using the previously defined `leader_arm` and `follower_arm`.
+Before you can teleoperate your robot, you need to instantiate the  [`ManipulatorRobot`](../lerobot/common/robot_devices/robots/manipulator.py) using the previously defined `leader_config` and `follower_config`.
 
-For the Koch v1.1 robot, we only have one leader, so we refer to it as `"main"` and define it as `leader_arms={"main": leader_arm}`. We do the same for the follower arm. For other robots (like the Aloha), which may have two pairs of leader and follower arms, you would define them like this: `leader_arms={"left": left_leader_arm, "right": right_leader_arm},`. Same thing for the follower arms.
+For the Koch v1.1 robot, we only have one leader, so we refer to it as `"main"` and define it as `leader_arms={"main": leader_config}`. We do the same for the follower arm. For other robots (like the Aloha), which may have two pairs of leader and follower arms, you would define them like this: `leader_arms={"left": left_leader_config, "right": right_leader_config},`. Same thing for the follower arms.
 
-You also need to provide a path to a calibration directory, such as  `calibration_dir=".cache/calibration/koch"`. More on this in the next section.
 
 Run the following code to instantiate your manipulator robot:
 ```python
+from lerobot.common.robot_devices.robots.configs import KochRobotConfig
 from lerobot.common.robot_devices.robots.manipulator import ManipulatorRobot
 
-robot = ManipulatorRobot(
-    robot_type="koch",
-    leader_arms={"main": leader_arm},
-    follower_arms={"main": follower_arm},
-    calibration_dir=".cache/calibration/koch",
+robot_config = KochRobotConfig(
+    leader_arms={"main": leader_config},
+    follower_arms={"main": follower_config},
+    cameras={},  # We don't use any camera for now
 )
+robot = ManipulatorRobot(robot_config)
 ```
 
-The `robot_type="koch"` is used to set the associated settings and calibration process. For instance, we activate the torque of the gripper of the leader Koch v1.1 arm and position it at a 40 degree angle to use it as a trigger.
+The `KochRobotConfig` is used to set the associated settings and calibration process. For instance, we activate the torque of the gripper of the leader Koch v1.1 arm and position it at a 40 degree angle to use it as a trigger.
 
-For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `robot_type="aloha"` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected. If you need to run manual calibration, simply update `calibration_dir` to `.cache/calibration/aloha`.
+For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `AlohaRobotConfig` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected for Aloha.
 
 **Calibrate and Connect the ManipulatorRobot**
 
@@ -354,7 +398,7 @@ And here are the corresponding positions for the leader arm:
 
 You can watch a [video tutorial of the calibration procedure](https://youtu.be/8drnU9uRY24) for more details.
 
-During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask yo to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to mesure if the values changed negatively or positively.
+During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask yo to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to measure if the values changed negatively or positively.
 
 Finally, the rest position ensures that the follower and leader arms are roughly aligned after calibration, preventing sudden movements that could damage the motors when starting teleoperation.
 
@@ -579,9 +623,11 @@ Note: Some cameras may take a few seconds to warm up, and the first frame might
 
 Finally, run this code to instantiate and connectyour camera:
 ```python
+from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
 from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
 
-camera = OpenCVCamera(camera_index=0)
+config = OpenCVCameraConfig(camera_index=0)
+camera = OpenCVCamera(config)
 camera.connect()
 color_image = camera.read()
 
@@ -603,7 +649,7 @@ uint8
 
 With certain camera, you can also specify additional parameters like frame rate, resolution, and color mode during instantiation. For instance:
 ```python
-camera = OpenCVCamera(camera_index=0, fps=30, width=640, height=480)
+config = OpenCVCameraConfig(camera_index=0, fps=30, width=640, height=480)
 ```
 
 If the provided arguments are not compatible with the camera, an exception will be raised.
@@ -617,18 +663,20 @@ camera.disconnect()
 
 **Instantiate your robot with cameras**
 
-Additionaly, you can set up your robot to work with your cameras.
+Additionally, you can set up your robot to work with your cameras.
 
 Modify the following Python code with the appropriate camera names and configurations:
 ```python
 robot = ManipulatorRobot(
-    leader_arms={"main": leader_arm},
-    follower_arms={"main": follower_arm},
-    calibration_dir=".cache/calibration/koch",
-    cameras={
-        "laptop": OpenCVCamera(0, fps=30, width=640, height=480),
-        "phone": OpenCVCamera(1, fps=30, width=640, height=480),
-    },
+    KochRobotConfig(
+        leader_arms={"main": leader_arm},
+        follower_arms={"main": follower_arm},
+        calibration_dir=".cache/calibration/koch",
+        cameras={
+            "laptop": OpenCVCameraConfig(0, fps=30, width=640, height=480),
+            "phone": OpenCVCameraConfig(1, fps=30, width=640, height=480),
+        },
+    )
 )
 robot.connect()
 ```
@@ -652,39 +700,20 @@ torch.Size([3, 480, 640])
 255
 ```
 
-Also, update the following lines of the yaml file for Koch robot [`lerobot/configs/robot/koch.yaml`](../lerobot/configs/robot/koch.yaml) with the names and configurations of your cameras:
-```yaml
-[...]
-cameras:
-  laptop:
-    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
-    camera_index: 0
-    fps: 30
-    width: 640
-    height: 480
-  phone:
-    _target_: lerobot.common.robot_devices.cameras.opencv.OpenCVCamera
-    camera_index: 1
-    fps: 30
-    width: 640
-    height: 480
-```
+### d. Use `control_robot.py` and our `teleoperate` function
 
-This file is used to instantiate your robot in all our scripts. We will explain how this works in the next section.
-
-### d. Use `koch.yaml` and our `teleoperate` function
-
-Instead of manually running the python code in a terminal window, you can use [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) to instantiate your robot by providing the path to the robot yaml file (e.g. [`lerobot/configs/robot/koch.yaml`](../lerobot/configs/robot/koch.yaml)) and control your robot with various modes as explained next.
+Instead of manually running the python code in a terminal window, you can use [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) to instantiate your robot by providing the robot configurations via command line and control your robot with various modes as explained next.
 
 Try running this code to teleoperate your robot (if you dont have a camera, keep reading):
 ```bash
-python lerobot/scripts/control_robot.py teleoperate \
-  --robot-path lerobot/configs/robot/koch.yaml
+python lerobot/scripts/control_robot.py \
+  --robot.type=koch \
+  --control.type=teleoperate
 ```
 
 You will see a lot of lines appearing like this one:
 ```
-INFO 2024-08-10 11:15:03 ol_robot.py:209 dt: 5.12 (195.1hz) dtRlead: 4.93 (203.0hz) dtRfoll: 0.19 (5239.0hz)
+INFO 2024-08-10 11:15:03 ol_robot.py:209 dt: 5.12 (195.1hz) dtRlead: 4.93 (203.0hz) dtWfoll: 0.19 (5239.0hz)
 ```
 
 It contains
@@ -694,21 +723,12 @@ It contains
 - `dtRlead: 4.93 (203.0hz)` which is the number of milliseconds it took to read the position of the leader arm using `leader_arm.read("Present_Position")`.
 - `dtWfoll: 0.22 (4446.9hz)` which is the number of milliseconds it took to set a new goal position for the follower arm using `follower_arm.write("Goal_position", leader_pos)` ; note that writing is done asynchronously so it takes less time than reading.
 
-Note: you can override any entry in the yaml file using `--robot-overrides` and the [hydra.cc](https://hydra.cc/docs/advanced/override_grammar/basic) syntax. If needed, you can override the ports like this:
+Importantly: If you don't have any camera, you can remove them dynamically with this [draccus](https://github.com/dlwh/draccus) syntax `--robot.cameras='{}'`:
 ```bash
-python lerobot/scripts/control_robot.py teleoperate \
-  --robot-path lerobot/configs/robot/koch.yaml \
-  --robot-overrides \
-    leader_arms.main.port=/dev/tty.usbmodem575E0031751 \
-    follower_arms.main.port=/dev/tty.usbmodem575E0032081
-```
-
-Importantly: If you don't have any camera, you can remove them dynamically with this [hydra.cc](https://hydra.cc/docs/advanced/override_grammar/basic) syntax `'~cameras'`:
-```bash
-python lerobot/scripts/control_robot.py teleoperate \
-  --robot-path lerobot/configs/robot/koch.yaml \
-  --robot-overrides \
-    '~cameras'
+python lerobot/scripts/control_robot.py \
+  --robot.type=koch \
+  --robot.cameras='{}' \
+  --control.type=teleoperate
 ```
 
 We advise to create a new yaml file when the command becomes too long.
@@ -744,23 +764,23 @@ for _ in range(record_time_s * fps):
 
 Importantly, many utilities are still missing. For instance, if you have cameras, you will need to save the images on disk to not go out of RAM, and to do so in threads to not slow down communication with your robot. Also, you will need to store your data in a format optimized for training and web sharing like [`LeRobotDataset`](../lerobot/common/datasets/lerobot_dataset.py). More on this in the next section.
 
-### a. Use `koch.yaml` and the `record` function
+### a. Use the `record` function
 
 You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) to achieve efficient data recording. It encompasses many recording utilities:
-1. Frames from cameras are saved on disk in threads, and encoded into videos at the end of recording.
+1. Frames from cameras are saved on disk in threads, and encoded into videos at the end of each episode recording.
 2. Video streams from cameras are displayed in window so that you can verify them.
-3. Data is stored with [`LeRobotDataset`](../lerobot/common/datasets/lerobot_dataset.py) format which is pushed to your Hugging Face page (unless `--push-to-hub 0` is provided).
-4. Checkpoints are done during recording, so if any issue occurs, you can resume recording by re-running the same command again. You can also use `--force-override 1` to start recording from scratch.
+3. Data is stored with [`LeRobotDataset`](../lerobot/common/datasets/lerobot_dataset.py) format which is pushed to your Hugging Face page (unless `--control.push_to_hub=false` is provided).
+4. Checkpoints are done during recording, so if any issue occurs, you can resume recording by re-running the same command again with `--control.resume=true`. You will need to manually delete the dataset directory if you want to start recording from scratch.
 5. Set the flow of data recording using command line arguments:
-   - `--warmup-time-s` defines the number of seconds before starting data collection. It allows the robot devices to warmup and synchronize (10 seconds by default).
-   - `--episode-time-s` defines the number of seconds for data recording for each episode (60 seconds by default).
-   - `--reset-time-s` defines the number of seconds for resetting the environment after each episode (60 seconds by default).
-   - `--num-episodes` defines the number of episodes to record (50 by default).
+   - `--control.warmup_time_s=10` defines the number of seconds before starting data collection. It allows the robot devices to warmup and synchronize (10 seconds by default).
+   - `--control.episode_time_s=60` defines the number of seconds for data recording for each episode (60 seconds by default).
+   - `--control.reset_time_s=60` defines the number of seconds for resetting the environment after each episode (60 seconds by default).
+   - `--control.num_episodes=50` defines the number of episodes to record (50 by default).
 6. Control the flow during data recording using keyboard keys:
    - Press right arrow `->` at any time during episode recording to early stop and go to resetting. Same during resetting, to early stop and to go to the next episode recording.
    - Press left arrow `<-` at any time during episode recording or resetting to early stop, cancel the current episode, and re-record it.
    - Press escape `ESC` at any time during episode recording to end the session early and go straight to video encoding and dataset uploading.
-7. Similarly to `teleoperate`, you can also use `--robot-path` and `--robot-overrides` to specify your robots.
+7. Similarly to `teleoperate`, you can also use the command line to override anything.
 
 Before trying `record`, if you want to push your dataset to the hub, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
 ```bash
@@ -771,28 +791,29 @@ Also, store your Hugging Face repository name in a variable (e.g. `cadene` or `l
 HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```
-If you don't want to push to hub, use `--push-to-hub 0`.
+If you don't want to push to hub, use `--control.push_to_hub=false`.
 
 Now run this to record 2 episodes:
 ```bash
-python lerobot/scripts/control_robot.py record \
-  --robot-path lerobot/configs/robot/koch.yaml \
-  --fps 30 \
-  --root data \
-  --repo-id ${HF_USER}/koch_test \
-  --tags tutorial \
-  --warmup-time-s 5 \
-  --episode-time-s 30 \
-  --reset-time-s 30 \
-  --num-episodes 2
+python lerobot/scripts/control_robot.py \
+  --robot.type=koch \
+  --control.type=record \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/koch_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
 ```
 
-This will write your dataset locally to `{root}/{repo-id}` (e.g. `data/cadene/koch_test`) and push it on the hub at `https://huggingface.co/datasets/{HF_USER}/{repo-id}`. Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
+
+This will write your dataset locally to `~/.cache/huggingface/lerobot/{repo-id}` (e.g. `data/cadene/koch_test`) and push it on the hub at `https://huggingface.co/datasets/{HF_USER}/{repo-id}`. Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
 
 You can look for other LeRobot datasets on the hub by searching for `LeRobot` tags: https://huggingface.co/datasets?other=LeRobot
 
-Remember to add `--robot-overrides '~cameras'` if you don't have any cameras and you still use the default `koch.yaml` configuration.
-
 You will see a lot of lines appearing like this one:
 ```
 INFO 2024-08-10 15:02:58 ol_robot.py:219 dt:33.34 (30.0hz) dtRlead: 5.06 (197.5hz) dtWfoll: 0.25 (3963.7hz) dtRfoll: 6.22 (160.7hz) dtRlaptop: 32.57 (30.7hz) dtRphone: 33.84 (29.5hz)
@@ -804,8 +825,8 @@ It contains:
 - `dtRlead: 5.06 (197.5hz)` which is the delta time of reading the present position of the leader arm.
 - `dtWfoll: 0.25 (3963.7hz)` which is the delta time of writing the goal position on the follower arm ; writing is asynchronous so it takes less time than reading.
 - `dtRfoll: 6.22 (160.7hz)` which is the delta time of reading the present position on the follower arm.
-- `dtRlaptop:32.57 (30.7hz) ` which is the delta time of capturing an image from the laptop camera in the thread running asynchrously.
-- `dtRphone:33.84 (29.5hz)` which is the delta time of capturing an image from the phone camera in the thread running asynchrously.
+- `dtRlaptop:32.57 (30.7hz) ` which is the delta time of capturing an image from the laptop camera in the thread running asynchronously.
+- `dtRphone:33.84 (29.5hz)` which is the delta time of capturing an image from the phone camera in the thread running asynchronously.
 
 Troubleshooting:
 - On Linux, if you encounter a hanging issue when using cameras, uninstall opencv and re-install it with conda:
@@ -825,7 +846,7 @@ At the end of data recording, your dataset will be uploaded on your Hugging Face
 echo https://huggingface.co/datasets/${HF_USER}/koch_test
 ```
 
-### b. Advices for recording dataset
+### b. Advice for recording dataset
 
 Once you're comfortable with data recording, it's time to create a larger dataset for training. A good starting task is grasping an object at different locations and placing it in a bin. We suggest recording at least 50 episodes, with 10 episodes per location. Keep the cameras fixed and maintain consistent grasping behavior throughout the recordings.
 
@@ -840,10 +861,11 @@ In the coming months, we plan to release a foundational model for robotics. We a
 You can visualize your dataset by running:
 ```bash
 python lerobot/scripts/visualize_dataset_html.py \
-  --root data \
   --repo-id ${HF_USER}/koch_test
 ```
 
+Note: You might need to add `--local-files-only 1` if your dataset was not uploaded to hugging face hub.
+
 This will launch a local web server that looks like this:
 <div style="text-align:center;">
   <img src="../media/tutorial/visualize_dataset_html.webp?raw=true" alt="Koch v1.1 leader and follower arms" title="Koch v1.1 leader and follower arms" width="100%">
@@ -855,12 +877,12 @@ A useful feature of [`lerobot/scripts/control_robot.py`](../lerobot/scripts/cont
 
 To replay the first episode of the dataset you just recorded, run the following command:
 ```bash
-python lerobot/scripts/control_robot.py replay \
-  --robot-path lerobot/configs/robot/koch.yaml \
-  --fps 30 \
-  --root data \
-  --repo-id ${HF_USER}/koch_test \
-  --episode 0
+python lerobot/scripts/control_robot.py \
+  --robot.type=koch \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/koch_test \
+  --control.episode=0
 ```
 
 Your robot should replicate movements similar to those you recorded. For example, check out [this video](https://x.com/RemiCadene/status/1793654950905680090) where we use `replay` on a Aloha robot from [Trossen Robotics](https://www.trossenrobotics.com).
@@ -871,54 +893,20 @@ Your robot should replicate movements similar to those you recorded. For example
 
 To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 ```bash
-DATA_DIR=data python lerobot/scripts/train.py \
-  dataset_repo_id=${HF_USER}/koch_test \
-  policy=act_koch_real \
-  env=koch_real \
-  hydra.run.dir=outputs/train/act_koch_test \
-  hydra.job.name=act_koch_test \
-  device=cuda \
-  wandb.enable=true
+python lerobot/scripts/train.py \
+  --dataset.repo_id=${HF_USER}/koch_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_koch_test \
+  --job_name=act_koch_test \
+  --device=cuda \
+  --wandb.enable=true
 ```
 
 Let's explain it:
-1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/koch_test`.
-2. We provided the policy with `policy=act_koch_real`. This loads configurations from [`lerobot/configs/policy/act_koch_real.yaml`](../lerobot/configs/policy/act_koch_real.yaml). Importantly, this policy uses 2 cameras as input `laptop` and `phone`. If your dataset has different cameras, update the yaml file to account for it in the following parts:
-```yaml
-...
-override_dataset_stats:
-  observation.images.laptop:
-    # 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)
-  observation.images.phone:
-    # 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)
-...
-  input_shapes:
-    observation.images.laptop: [3, 480, 640]
-    observation.images.phone: [3, 480, 640]
-...
-  input_normalization_modes:
-    observation.images.laptop: mean_std
-    observation.images.phone: mean_std
-...
-```
-3. We provided an environment as argument with `env=koch_real`. This loads configurations from [`lerobot/configs/env/koch_real.yaml`](../lerobot/configs/env/koch_real.yaml). It looks like
-```yaml
-fps: 30
-env:
-  name: real_world
-  task: null
-  state_dim: 6
-  action_dim: 6
-  fps: ${fps}
-```
-It should match your dataset (e.g. `fps: 30`) and your robot (e.g. `state_dim: 6` and `action_dim: 6`). We are still working on simplifying this in future versions of `lerobot`.
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/koch_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
 4. We provided `device=cuda` since we are training on a Nvidia GPU, but you could use `device=mps` to train on Apple silicon.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$HOME/.cache/hugginface`. In future versions of `lerobot`, both directories will be in sync.
 
 For more information on the `train` script see the previous tutorial: [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md)
 
@@ -982,36 +970,36 @@ for _ in range(inference_time_s * fps):
     busy_wait(1 / fps - dt_s)
 ```
 
-### a. Use `koch.yaml` and our `record` function
+### a. Use our `record` function
 
 Ideally, when controlling your robot with your neural network, you would want to record evaluation episodes and to be able to visualize them later on, or even train on them like in Reinforcement Learning. This pretty much corresponds to recording a new dataset but with a neural network providing the actions instead of teleoperation.
 
 To this end, you can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
 ```bash
-python lerobot/scripts/control_robot.py record \
-  --robot-path lerobot/configs/robot/koch.yaml \
-  --fps 30 \
-  --root data \
-  --repo-id ${HF_USER}/eval_koch_test \
-  --tags tutorial eval \
-  --warmup-time-s 5 \
-  --episode-time-s 30 \
-  --reset-time-s 30 \
-  --num-episodes 10 \
-  -p outputs/train/act_koch_test/checkpoints/last/pretrained_model
+python lerobot/scripts/control_robot.py \
+  --robot.type=koch \
+  --control.type=record \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/eval_act_koch_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=10 \
+  --control.push_to_hub=true \
+  --control.policy.path=outputs/train/act_koch_test/checkpoints/last/pretrained_model
 ```
 
 As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_koch_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_koch_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_koch_test`).
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_koch_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_koch_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_koch_test`).
 
 ### b. Visualize evaluation afterwards
 
 You can then visualize your evaluation dataset by running the same command as before but with the new inference dataset as argument:
 ```bash
 python lerobot/scripts/visualize_dataset.py \
-  --root data \
-  --repo-id ${HF_USER}/eval_koch_test
+  --repo-id ${HF_USER}/eval_act_koch_test
 ```
 
 ## 6. Next step

examples/advanced/1_add_image_transforms.py

@@ -1,7 +1,21 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 """
 This script demonstrates how to use torchvision's image transformation with LeRobotDataset for data
 augmentation purposes. The transformations are passed to the dataset as an argument upon creation, and
-transforms are applied to the observation images before they are returned in the dataset's __get_item__.
+transforms are applied to the observation images before they are returned in the dataset's __getitem__.
 """
 
 from pathlib import Path
@@ -10,17 +24,17 @@ from torchvision.transforms import ToPILImage, v2
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
-dataset_repo_id = "lerobot/aloha_static_tape"
+dataset_repo_id = "lerobot/aloha_static_screw_driver"
 
 # Create a LeRobotDataset with no transformations
-dataset = LeRobotDataset(dataset_repo_id)
+dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
 # This is equivalent to `dataset = LeRobotDataset(dataset_repo_id, image_transforms=None)`
 
 # Get the index of the first observation in the first episode
 first_idx = dataset.episode_data_index["from"][0].item()
 
 # Get the frame corresponding to the first camera
-frame = dataset[first_idx][dataset.camera_keys[0]]
+frame = dataset[first_idx][dataset.meta.camera_keys[0]]
 
 
 # Define the transformations
@@ -28,15 +42,16 @@ transforms = v2.Compose(
     [
         v2.ColorJitter(brightness=(0.5, 1.5)),
         v2.ColorJitter(contrast=(0.5, 1.5)),
+        v2.ColorJitter(hue=(-0.1, 0.1)),
         v2.RandomAdjustSharpness(sharpness_factor=2, p=1),
     ]
 )
 
 # Create another LeRobotDataset with the defined transformations
-transformed_dataset = LeRobotDataset(dataset_repo_id, image_transforms=transforms)
+transformed_dataset = LeRobotDataset(dataset_repo_id, episodes=[0], image_transforms=transforms)
 
 # Get a frame from the transformed dataset
-transformed_frame = transformed_dataset[first_idx][transformed_dataset.camera_keys[0]]
+transformed_frame = transformed_dataset[first_idx][transformed_dataset.meta.camera_keys[0]]
 
 # Create a directory to store output images
 output_dir = Path("outputs/image_transforms")

examples/advanced/1_train_act_pusht/act_pusht.yaml

@@ -1,87 +0,0 @@
-# @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

@@ -1,70 +0,0 @@
-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

@@ -1,3 +1,17 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 """This script demonstrates how to slice a dataset and calculate the loss on a subset of the data.
 
 This technique can be useful for debugging and testing purposes, as well as identifying whether a policy
@@ -9,82 +23,82 @@ 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.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 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")
+def main():
+    device = torch.device("cuda")
 
-policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
-policy.eval()
-policy.to(device)
+    # Download the diffusion policy for pusht environment
+    pretrained_policy_path = "lerobot/diffusion_pusht"
+    # OR uncomment the following to evaluate a policy from the local outputs/train folder.
+    # pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
 
-# 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],
-}
+    policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
+    policy.eval()
+    policy.to(device)
 
-# 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)}")
+    # 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],
+    }
 
-# 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,
-)
+    # Load the last 10% of episodes of the dataset as a validation set.
+    # - Load dataset metadata
+    dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
+    # - Calculate train and val episodes
+    total_episodes = dataset_metadata.total_episodes
+    episodes = list(range(dataset_metadata.total_episodes))
+    num_train_episodes = math.floor(total_episodes * 90 / 100)
+    train_episodes = episodes[:num_train_episodes]
+    val_episodes = episodes[num_train_episodes:]
+    print(f"Number of episodes in full dataset: {total_episodes}")
+    print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
+    print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
+    # - Load train an val datasets
+    train_dataset = LeRobotDataset(
+        "lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps
+    )
+    val_dataset = LeRobotDataset("lerobot/pusht", episodes=val_episodes, delta_timestamps=delta_timestamps)
+    print(f"Number of frames in training dataset (90% subset): {len(train_dataset)}")
+    print(f"Number of frames in validation dataset (10% subset): {len(val_dataset)}")
 
-# 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)
+    # 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,
+    )
 
-    loss_cumsum += output_dict["loss"].item()
-    n_examples_evaluated += batch["index"].shape[0]
+    # Run validation loop.
+    loss_cumsum = 0
+    n_examples_evaluated = 0
+    for batch in val_dataloader:
+        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+        loss, _ = policy.forward(batch)
 
-# Calculate the average loss over the validation set.
-average_loss = loss_cumsum / n_examples_evaluated
+        loss_cumsum += loss.item()
+        n_examples_evaluated += batch["index"].shape[0]
 
-print(f"Average loss on validation set: {average_loss:.4f}")
+    # Calculate the average loss over the validation set.
+    average_loss = loss_cumsum / n_examples_evaluated
+
+    print(f"Average loss on validation set: {average_loss:.4f}")
+
+
+if __name__ == "__main__":
+    main()

examples/port_datasets/pusht_zarr.py

@@ -0,0 +1,243 @@
+# 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 shutil
+from pathlib import Path
+
+import numpy as np
+from huggingface_hub import HfApi
+
+from lerobot.common.constants import HF_LEROBOT_HOME
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
+from lerobot.common.datasets.push_dataset_to_hub._download_raw import download_raw
+
+PUSHT_TASK = "Push the T-shaped blue block onto the T-shaped green target surface."
+PUSHT_FEATURES = {
+    "observation.state": {
+        "dtype": "float32",
+        "shape": (2,),
+        "names": {
+            "axes": ["x", "y"],
+        },
+    },
+    "action": {
+        "dtype": "float32",
+        "shape": (2,),
+        "names": {
+            "axes": ["x", "y"],
+        },
+    },
+    "next.reward": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": None,
+    },
+    "next.success": {
+        "dtype": "bool",
+        "shape": (1,),
+        "names": None,
+    },
+    "observation.environment_state": {
+        "dtype": "float32",
+        "shape": (16,),
+        "names": [
+            "keypoints",
+        ],
+    },
+    "observation.image": {
+        "dtype": None,
+        "shape": (3, 96, 96),
+        "names": [
+            "channels",
+            "height",
+            "width",
+        ],
+    },
+}
+
+
+def build_features(mode: str) -> dict:
+    features = PUSHT_FEATURES
+    if mode == "keypoints":
+        features.pop("observation.image")
+    else:
+        features.pop("observation.environment_state")
+        features["observation.image"]["dtype"] = mode
+
+    return features
+
+
+def load_raw_dataset(zarr_path: Path):
+    try:
+        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
+
+    zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
+    return zarr_data
+
+
+def calculate_coverage(zarr_data):
+    try:
+        import pymunk
+        from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+
+    block_pos = zarr_data["state"][:, 2:4]
+    block_angle = zarr_data["state"][:, 4]
+
+    num_frames = len(block_pos)
+
+    coverage = np.zeros((num_frames,), dtype=np.float32)
+    # 8 keypoints with 2 coords each
+    keypoints = np.zeros((num_frames, 16), dtype=np.float32)
+
+    # Set x, y, theta (in radians)
+    goal_pos_angle = np.array([256, 256, np.pi / 4])
+    goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
+
+    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[i] = intersection_area / goal_area
+        keypoints[i] = PushTEnv.get_keypoints(block_shapes).flatten()
+
+    return coverage, keypoints
+
+
+def calculate_success(coverage: float, success_threshold: float):
+    return coverage > success_threshold
+
+
+def calculate_reward(coverage: float, success_threshold: float):
+    return np.clip(coverage / success_threshold, 0, 1)
+
+
+def main(raw_dir: Path, repo_id: str, mode: str = "video", push_to_hub: bool = True):
+    if mode not in ["video", "image", "keypoints"]:
+        raise ValueError(mode)
+
+    if (HF_LEROBOT_HOME / repo_id).exists():
+        shutil.rmtree(HF_LEROBOT_HOME / repo_id)
+
+    if not raw_dir.exists():
+        download_raw(raw_dir, repo_id="lerobot-raw/pusht_raw")
+
+    zarr_data = load_raw_dataset(zarr_path=raw_dir / "pusht_cchi_v7_replay.zarr")
+
+    env_state = zarr_data["state"][:]
+    agent_pos = env_state[:, :2]
+
+    action = zarr_data["action"][:]
+    image = zarr_data["img"]  # (b, h, w, c)
+
+    if image.dtype == np.float32 and image.max() == np.float32(255):
+        # HACK: images are loaded as float32 but they actually encode uint8 data
+        image = image.astype(np.uint8)
+
+    episode_data_index = {
+        "from": np.concatenate(([0], zarr_data.meta["episode_ends"][:-1])),
+        "to": zarr_data.meta["episode_ends"],
+    }
+
+    # Calculate success and reward based on the overlapping area
+    # of the T-object and the T-area.
+    coverage, keypoints = calculate_coverage(zarr_data)
+    success = calculate_success(coverage, success_threshold=0.95)
+    reward = calculate_reward(coverage, success_threshold=0.95)
+
+    features = build_features(mode)
+    dataset = LeRobotDataset.create(
+        repo_id=repo_id,
+        fps=10,
+        robot_type="2d pointer",
+        features=features,
+        image_writer_threads=4,
+    )
+    episodes = range(len(episode_data_index["from"]))
+    for ep_idx in episodes:
+        from_idx = episode_data_index["from"][ep_idx]
+        to_idx = episode_data_index["to"][ep_idx]
+        num_frames = to_idx - from_idx
+
+        for frame_idx in range(num_frames):
+            i = from_idx + frame_idx
+            idx = i + (frame_idx < num_frames - 1)
+            frame = {
+                "action": action[i],
+                # Shift reward and success by +1 until the last item of the episode
+                "next.reward": reward[idx : idx + 1],
+                "next.success": success[idx : idx + 1],
+                "task": PUSHT_TASK,
+            }
+
+            frame["observation.state"] = agent_pos[i]
+
+            if mode == "keypoints":
+                frame["observation.environment_state"] = keypoints[i]
+            else:
+                frame["observation.image"] = image[i]
+
+            dataset.add_frame(frame)
+
+        dataset.save_episode()
+
+    if push_to_hub:
+        dataset.push_to_hub()
+        hub_api = HfApi()
+        hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
+
+
+if __name__ == "__main__":
+    # To try this script, modify the repo id with your own HuggingFace user (e.g cadene/pusht)
+    repo_id = "lerobot/pusht"
+
+    modes = ["video", "image", "keypoints"]
+    # Uncomment if you want to try with a specific mode
+    # modes = ["video"]
+    # modes = ["image"]
+    # modes = ["keypoints"]
+
+    raw_dir = Path("data/lerobot-raw/pusht_raw")
+    for mode in modes:
+        if mode in ["image", "keypoints"]:
+            repo_id += f"_{mode}"
+
+        # download and load raw dataset, create LeRobotDataset, populate it, push to hub
+        main(raw_dir, repo_id=repo_id, mode=mode)
+
+        # Uncomment if you want to load the local dataset and explore it
+        # dataset = LeRobotDataset(repo_id=repo_id)
+        # breakpoint()

lerobot/__init__.py

@@ -58,7 +58,6 @@ 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())
 
@@ -86,23 +85,6 @@ available_datasets_per_env = {
         "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_real_world_datasets = [
@@ -181,8 +163,8 @@ available_real_world_datasets = [
     "lerobot/usc_cloth_sim",
 ]
 
-available_datasets = list(
-    itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets)
+available_datasets = sorted(
+    set(itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets))
 )
 
 # lists all available policies from `lerobot/common/policies`
@@ -221,7 +203,6 @@ available_policies_per_env = {
     "xarm": ["tdmpc"],
     "koch_real": ["act_koch_real"],
     "aloha_real": ["act_aloha_real"],
-    "dora_aloha_real": ["act_aloha_real"],
 }
 
 env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]

lerobot/common/cameras/__init__.py

@@ -0,0 +1,4 @@
+from .camera import Camera
+from .configs import CameraConfig
+
+__all__ = ["Camera", "CameraConfig"]

lerobot/common/cameras/camera.py

@@ -0,0 +1,25 @@
+import abc
+
+import numpy as np
+
+
+class Camera(abc.ABC):
+    @abc.abstractmethod
+    def connect(self):
+        pass
+
+    @abc.abstractmethod
+    def read(self, temporary_color: str | None = None) -> np.ndarray:
+        pass
+
+    @abc.abstractmethod
+    def async_read(self) -> np.ndarray:
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self):
+        pass
+
+    def __del__(self):
+        if getattr(self, "is_connected", False):
+            self.disconnect()

lerobot/common/cameras/configs.py

@@ -0,0 +1,11 @@
+import abc
+from dataclasses import dataclass
+
+import draccus
+
+
+@dataclass
+class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)

lerobot/common/cameras/intel/__init__.py

@@ -0,0 +1,4 @@
+from .camera_realsense import RealSenseCamera
+from .configuration_realsense import RealSenseCameraConfig
+
+__all__ = ["RealSenseCamera", "RealSenseCameraConfig"]

lerobot/common/cameras/intel/camera_realsense.py

@@ -1,3 +1,17 @@
+# 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 utilities for recording frames from Intel Realsense cameras.
 """
@@ -11,20 +25,21 @@ import threading
 import time
 import traceback
 from collections import Counter
-from dataclasses import dataclass, replace
 from pathlib import Path
 from threading import Thread
 
 import numpy as np
 from PIL import Image
 
-from lerobot.common.robot_devices.utils import (
-    RobotDeviceAlreadyConnectedError,
-    RobotDeviceNotConnectedError,
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.common.utils.robot_utils import (
     busy_wait,
 )
 from lerobot.common.utils.utils import capture_timestamp_utc
 
+from ..camera import Camera
+from .configuration_realsense import RealSenseCameraConfig
+
 SERIAL_NUMBER_INDEX = 1
 
 
@@ -94,10 +109,11 @@ def save_images_from_cameras(
     cameras = []
     for cam_sn in serial_numbers:
         print(f"{cam_sn=}")
-        camera = IntelRealSenseCamera(cam_sn, fps=fps, width=width, height=height, mock=mock)
+        config = RealSenseCameraConfig(serial_number=cam_sn, fps=fps, width=width, height=height, mock=mock)
+        camera = RealSenseCamera(config)
         camera.connect()
         print(
-            f"IntelRealSenseCamera({camera.serial_number}, fps={camera.fps}, width={camera.width}, height={camera.height}, color_mode={camera.color_mode})"
+            f"RealSenseCamera({camera.serial_number}, fps={camera.fps}, width={camera.capture_width}, height={camera.capture_height}, color_mode={camera.color_mode})"
         )
         cameras.append(camera)
 
@@ -149,53 +165,11 @@ def save_images_from_cameras(
             camera.disconnect()
 
 
-@dataclass
-class IntelRealSenseCameraConfig:
+class RealSenseCamera(Camera):
     """
-    Example of tested options for Intel Real Sense D405:
-
-    ```python
-    IntelRealSenseCameraConfig(30, 640, 480)
-    IntelRealSenseCameraConfig(60, 640, 480)
-    IntelRealSenseCameraConfig(90, 640, 480)
-    IntelRealSenseCameraConfig(30, 1280, 720)
-    IntelRealSenseCameraConfig(30, 640, 480, use_depth=True)
-    IntelRealSenseCameraConfig(30, 640, 480, rotation=90)
-    ```
-    """
-
-    fps: int | None = None
-    width: int | None = None
-    height: int | None = None
-    color_mode: str = "rgb"
-    use_depth: bool = False
-    force_hardware_reset: bool = True
-    rotation: int | None = None
-    mock: bool = False
-
-    def __post_init__(self):
-        if self.color_mode not in ["rgb", "bgr"]:
-            raise ValueError(
-                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
-            )
-
-        at_least_one_is_not_none = self.fps is not None or self.width is not None or self.height is not None
-        at_least_one_is_none = self.fps is None or self.width is None or self.height is None
-        if at_least_one_is_not_none and at_least_one_is_none:
-            raise ValueError(
-                "For `fps`, `width` and `height`, either all of them need to be set, or none of them, "
-                f"but {self.fps=}, {self.width=}, {self.height=} were provided."
-            )
-
-        if self.rotation not in [-90, None, 90, 180]:
-            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")
-
-
-class IntelRealSenseCamera:
-    """
-    The IntelRealSenseCamera class is similar to OpenCVCamera class but adds additional features for Intel Real Sense cameras:
+    The RealSenseCamera class is similar to OpenCVCamera class but adds additional features for Intel Real Sense cameras:
     - is instantiated with the serial number of the camera - won't randomly change as it can be the case of OpenCVCamera for Linux,
-    - can also be instantiated with the camera's name — if it's unique — using IntelRealSenseCamera.init_from_name(),
+    - can also be instantiated with the camera's name — if it's unique — using RealSenseCamera.init_from_name(),
     - depth map can be returned.
 
     To find the camera indices of your cameras, you can run our utility script that will save a few frames for each camera:
@@ -203,36 +177,38 @@ class IntelRealSenseCamera:
     python lerobot/common/robot_devices/cameras/intelrealsense.py --images-dir outputs/images_from_intelrealsense_cameras
     ```
 
-    When an IntelRealSenseCamera is instantiated, if no specific config is provided, the default fps, width, height and color_mode
+    When an RealSenseCamera is instantiated, if no specific config is provided, the default fps, width, height and color_mode
     of the given camera will be used.
 
-    Example of usage:
+    Example of instantiating with a serial number:
     ```python
-    # Instantiate with its serial number
-    camera = IntelRealSenseCamera(128422271347)
-    # Or by its name if it's unique
-    camera = IntelRealSenseCamera.init_from_name("Intel RealSense D405")
+    from lerobot.common.robot_devices.cameras.configs import RealSenseCameraConfig
+
+    config = RealSenseCameraConfig(serial_number=128422271347)
+    camera = RealSenseCamera(config)
     camera.connect()
     color_image = camera.read()
     # when done using the camera, consider disconnecting
     camera.disconnect()
     ```
 
+    Example of instantiating with a name if it's unique:
+    ```
+    config = RealSenseCameraConfig(name="Intel RealSense D405")
+    ```
+
     Example of changing default fps, width, height and color_mode:
     ```python
-    camera = IntelRealSenseCamera(serial_number, fps=30, width=1280, height=720)
-    camera = connect()  # applies the settings, might error out if these settings are not compatible with the camera
-
-    camera = IntelRealSenseCamera(serial_number, fps=90, width=640, height=480)
-    camera = connect()
-
-    camera = IntelRealSenseCamera(serial_number, fps=90, width=640, height=480, color_mode="bgr")
-    camera = connect()
+    config = RealSenseCameraConfig(serial_number=128422271347, fps=30, width=1280, height=720)
+    config = RealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480)
+    config = RealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480, color_mode="bgr")
+    # Note: might error out upon `camera.connect()` if these settings are not compatible with the camera
     ```
 
     Example of returning depth:
     ```python
-    camera = IntelRealSenseCamera(serial_number, use_depth=True)
+    config = RealSenseCameraConfig(serial_number=128422271347, use_depth=True)
+    camera = RealSenseCamera(config)
     camera.connect()
     color_image, depth_map = camera.read()
     ```
@@ -240,20 +216,28 @@ class IntelRealSenseCamera:
 
     def __init__(
         self,
-        serial_number: int,
-        config: IntelRealSenseCameraConfig | None = None,
-        **kwargs,
+        config: RealSenseCameraConfig,
     ):
-        if config is None:
-            config = IntelRealSenseCameraConfig()
+        self.config = config
+        if config.name is not None:
+            self.serial_number = self.find_serial_number_from_name(config.name)
+        else:
+            self.serial_number = config.serial_number
 
-        # Overwrite the config arguments using kwargs
-        config = replace(config, **kwargs)
+        # Store the raw (capture) resolution from the config.
+        self.capture_width = config.width
+        self.capture_height = config.height
+
+        # If rotated by ±90, swap width and height.
+        if config.rotation in [-90, 90]:
+            self.width = config.height
+            self.height = config.width
+        else:
+            self.width = config.width
+            self.height = config.height
 
-        self.serial_number = serial_number
         self.fps = config.fps
-        self.width = config.width
-        self.height = config.height
+        self.channels = config.channels
         self.color_mode = config.color_mode
         self.use_depth = config.use_depth
         self.force_hardware_reset = config.force_hardware_reset
@@ -272,7 +256,6 @@ class IntelRealSenseCamera:
         else:
             import cv2
 
-        # TODO(alibets): Do we keep original width/height or do we define them after rotation?
         self.rotation = None
         if config.rotation == -90:
             self.rotation = cv2.ROTATE_90_COUNTERCLOCKWISE
@@ -281,8 +264,7 @@ class IntelRealSenseCamera:
         elif config.rotation == 180:
             self.rotation = cv2.ROTATE_180
 
-    @classmethod
-    def init_from_name(cls, name: str, config: IntelRealSenseCameraConfig | None = None, **kwargs):
+    def find_serial_number_from_name(self, name):
         camera_infos = find_cameras()
         camera_names = [cam["name"] for cam in camera_infos]
         this_name_count = Counter(camera_names)[name]
@@ -295,19 +277,11 @@ class IntelRealSenseCamera:
         name_to_serial_dict = {cam["name"]: cam["serial_number"] for cam in camera_infos}
         cam_sn = name_to_serial_dict[name]
 
-        if config is None:
-            config = IntelRealSenseCameraConfig()
-
-        # Overwrite the config arguments using kwargs
-        config = replace(config, **kwargs)
-
-        return cls(serial_number=cam_sn, config=config, **kwargs)
+        return cam_sn
 
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is already connected."
-            )
+            raise DeviceAlreadyConnectedError(f"RealSenseCamera({self.serial_number}) is already connected.")
 
         if self.mock:
             import tests.mock_pyrealsense2 as rs
@@ -317,15 +291,19 @@ class IntelRealSenseCamera:
         config = rs.config()
         config.enable_device(str(self.serial_number))
 
-        if self.fps and self.width and self.height:
+        if self.fps and self.capture_width and self.capture_height:
             # TODO(rcadene): can we set rgb8 directly?
-            config.enable_stream(rs.stream.color, self.width, self.height, rs.format.rgb8, self.fps)
+            config.enable_stream(
+                rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps
+            )
         else:
             config.enable_stream(rs.stream.color)
 
         if self.use_depth:
-            if self.fps and self.width and self.height:
-                config.enable_stream(rs.stream.depth, self.width, self.height, rs.format.z16, self.fps)
+            if self.fps and self.capture_width and self.capture_height:
+                config.enable_stream(
+                    rs.stream.depth, self.capture_width, self.capture_height, rs.format.z16, self.fps
+                )
             else:
                 config.enable_stream(rs.stream.depth)
 
@@ -349,7 +327,7 @@ class IntelRealSenseCamera:
                     "To find the serial number you should use, run `python lerobot/common/robot_devices/cameras/intelrealsense.py`."
                 )
 
-            raise OSError(f"Can't access IntelRealSenseCamera({self.serial_number}).")
+            raise OSError(f"Can't access RealSenseCamera({self.serial_number}).")
 
         color_stream = profile.get_stream(rs.stream.color)
         color_profile = color_stream.as_video_stream_profile()
@@ -361,20 +339,20 @@ class IntelRealSenseCamera:
         if self.fps is not None and not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
             # Using `OSError` since it's a broad that encompasses issues related to device communication
             raise OSError(
-                f"Can't set {self.fps=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_fps}."
+                f"Can't set {self.fps=} for RealSenseCamera({self.serial_number}). Actual value is {actual_fps}."
             )
-        if self.width is not None and self.width != actual_width:
+        if self.capture_width is not None and self.capture_width != actual_width:
             raise OSError(
-                f"Can't set {self.width=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_width}."
+                f"Can't set {self.capture_width=} for RealSenseCamera({self.serial_number}). Actual value is {actual_width}."
             )
-        if self.height is not None and self.height != actual_height:
+        if self.capture_height is not None and self.capture_height != actual_height:
             raise OSError(
-                f"Can't set {self.height=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_height}."
+                f"Can't set {self.capture_height=} for RealSenseCamera({self.serial_number}). Actual value is {actual_height}."
             )
 
         self.fps = round(actual_fps)
-        self.width = round(actual_width)
-        self.height = round(actual_height)
+        self.capture_width = round(actual_width)
+        self.capture_height = round(actual_height)
 
         self.is_connected = True
 
@@ -389,8 +367,8 @@ class IntelRealSenseCamera:
         If you are reading data from other sensors, we advise to use `camera.async_read()` which is non blocking version of `camera.read()`.
         """
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            raise DeviceNotConnectedError(
+                f"RealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
             )
 
         if self.mock:
@@ -405,7 +383,7 @@ class IntelRealSenseCamera:
         color_frame = frame.get_color_frame()
 
         if not color_frame:
-            raise OSError(f"Can't capture color image from IntelRealSenseCamera({self.serial_number}).")
+            raise OSError(f"Can't capture color image from RealSenseCamera({self.serial_number}).")
 
         color_image = np.asanyarray(color_frame.get_data())
 
@@ -420,7 +398,7 @@ class IntelRealSenseCamera:
             color_image = cv2.cvtColor(color_image, cv2.COLOR_RGB2BGR)
 
         h, w, _ = color_image.shape
-        if h != self.height or w != self.width:
+        if h != self.capture_height or w != self.capture_width:
             raise OSError(
                 f"Can't capture color image with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
             )
@@ -437,12 +415,12 @@ class IntelRealSenseCamera:
         if self.use_depth:
             depth_frame = frame.get_depth_frame()
             if not depth_frame:
-                raise OSError(f"Can't capture depth image from IntelRealSenseCamera({self.serial_number}).")
+                raise OSError(f"Can't capture depth image from RealSenseCamera({self.serial_number}).")
 
             depth_map = np.asanyarray(depth_frame.get_data())
 
             h, w = depth_map.shape
-            if h != self.height or w != self.width:
+            if h != self.capture_height or w != self.capture_width:
                 raise OSError(
                     f"Can't capture depth map with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
                 )
@@ -464,8 +442,8 @@ class IntelRealSenseCamera:
     def async_read(self):
         """Access the latest color image"""
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            raise DeviceNotConnectedError(
+                f"RealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
             )
 
         if self.thread is None:
@@ -491,8 +469,8 @@ class IntelRealSenseCamera:
 
     def disconnect(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
+            raise DeviceNotConnectedError(
+                f"RealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
             )
 
         if self.thread is not None and self.thread.is_alive():
@@ -514,14 +492,14 @@ class IntelRealSenseCamera:
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
-        description="Save a few frames using `IntelRealSenseCamera` for all cameras connected to the computer, or a selected subset."
+        description="Save a few frames using `RealSenseCamera` for all cameras connected to the computer, or a selected subset."
     )
     parser.add_argument(
         "--serial-numbers",
         type=int,
         nargs="*",
         default=None,
-        help="List of serial numbers used to instantiate the `IntelRealSenseCamera`. If not provided, find and use all available camera indices.",
+        help="List of serial numbers used to instantiate the `RealSenseCamera`. If not provided, find and use all available camera indices.",
     )
     parser.add_argument(
         "--fps",

lerobot/common/cameras/intel/configuration_realsense.py

@@ -0,0 +1,71 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from ..configs import CameraConfig
+
+
+@CameraConfig.register_subclass("intelrealsense")
+@dataclass
+class RealSenseCameraConfig(CameraConfig):
+    """
+    Example of tested options for Intel Real Sense D405:
+
+    ```python
+    RealSenseCameraConfig(128422271347, 30, 640, 480)
+    RealSenseCameraConfig(128422271347, 60, 640, 480)
+    RealSenseCameraConfig(128422271347, 90, 640, 480)
+    RealSenseCameraConfig(128422271347, 30, 1280, 720)
+    RealSenseCameraConfig(128422271347, 30, 640, 480, use_depth=True)
+    RealSenseCameraConfig(128422271347, 30, 640, 480, rotation=90)
+    ```
+    """
+
+    name: str | None = None
+    serial_number: int | None = None
+    fps: int | None = None
+    width: int | None = None
+    height: int | None = None
+    color_mode: str = "rgb"
+    channels: int | None = None
+    use_depth: bool = False
+    force_hardware_reset: bool = True
+    rotation: int | None = None
+    mock: bool = False
+
+    def __post_init__(self):
+        # bool is stronger than is None, since it works with empty strings
+        if bool(self.name) and bool(self.serial_number):
+            raise ValueError(
+                f"One of them must be set: name or serial_number, but {self.name=} and {self.serial_number=} provided."
+            )
+
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )
+
+        self.channels = 3
+
+        at_least_one_is_not_none = self.fps is not None or self.width is not None or self.height is not None
+        at_least_one_is_none = self.fps is None or self.width is None or self.height is None
+        if at_least_one_is_not_none and at_least_one_is_none:
+            raise ValueError(
+                "For `fps`, `width` and `height`, either all of them need to be set, or none of them, "
+                f"but {self.fps=}, {self.width=}, {self.height=} were provided."
+            )
+
+        if self.rotation not in [-90, None, 90, 180]:
+            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")

lerobot/common/cameras/opencv/__init__.py

@@ -0,0 +1,4 @@
+from .camera_opencv import OpenCVCamera
+from .configuration_opencv import OpenCVCameraConfig
+
+__all__ = ["OpenCVCamera", "OpenCVCameraConfig"]

lerobot/common/cameras/opencv/camera_opencv.py

@@ -1,3 +1,17 @@
+# 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 utilities for recording frames from cameras. For more info look at `OpenCVCamera` docstring.
 """
@@ -9,20 +23,21 @@ import platform
 import shutil
 import threading
 import time
-from dataclasses import dataclass, replace
 from pathlib import Path
 from threading import Thread
 
 import numpy as np
 from PIL import Image
 
-from lerobot.common.robot_devices.utils import (
-    RobotDeviceAlreadyConnectedError,
-    RobotDeviceNotConnectedError,
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.common.utils.robot_utils import (
     busy_wait,
 )
 from lerobot.common.utils.utils import capture_timestamp_utc
 
+from ..camera import Camera
+from .configuration_opencv import OpenCVCameraConfig
+
 # The maximum opencv device index depends on your operating system. For instance,
 # if you have 3 cameras, they should be associated to index 0, 1, and 2. This is the case
 # on MacOS. However, on Ubuntu, the indices are different like 6, 16, 23.
@@ -126,11 +141,12 @@ def save_images_from_cameras(
     print("Connecting cameras")
     cameras = []
     for cam_idx in camera_ids:
-        camera = OpenCVCamera(cam_idx, fps=fps, width=width, height=height, mock=mock)
+        config = OpenCVCameraConfig(camera_index=cam_idx, fps=fps, width=width, height=height, mock=mock)
+        camera = OpenCVCamera(config)
         camera.connect()
         print(
-            f"OpenCVCamera({camera.camera_index}, fps={camera.fps}, width={camera.width}, "
-            f"height={camera.height}, color_mode={camera.color_mode})"
+            f"OpenCVCamera({camera.camera_index}, fps={camera.fps}, width={camera.capture_width}, "
+            f"height={camera.capture_height}, color_mode={camera.color_mode})"
         )
         cameras.append(camera)
 
@@ -175,37 +191,7 @@ def save_images_from_cameras(
     print(f"Images have been saved to {images_dir}")
 
 
-@dataclass
-class OpenCVCameraConfig:
-    """
-    Example of tested options for Intel Real Sense D405:
-
-    ```python
-    OpenCVCameraConfig(30, 640, 480)
-    OpenCVCameraConfig(60, 640, 480)
-    OpenCVCameraConfig(90, 640, 480)
-    OpenCVCameraConfig(30, 1280, 720)
-    ```
-    """
-
-    fps: int | None = None
-    width: int | None = None
-    height: int | None = None
-    color_mode: str = "rgb"
-    rotation: int | None = None
-    mock: bool = False
-
-    def __post_init__(self):
-        if self.color_mode not in ["rgb", "bgr"]:
-            raise ValueError(
-                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
-            )
-
-        if self.rotation not in [-90, None, 90, 180]:
-            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")
-
-
-class OpenCVCamera:
+class OpenCVCamera(Camera):
     """
     The OpenCVCamera class allows to efficiently record images from cameras. It relies on opencv2 to communicate
     with the cameras. Most cameras are compatible. For more info, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
@@ -224,7 +210,10 @@ class OpenCVCamera:
 
     Example of usage:
     ```python
-    camera = OpenCVCamera(camera_index=0)
+    from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
+
+    config = OpenCVCameraConfig(camera_index=0)
+    camera = OpenCVCamera(config)
     camera.connect()
     color_image = camera.read()
     # when done using the camera, consider disconnecting
@@ -233,25 +222,16 @@ class OpenCVCamera:
 
     Example of changing default fps, width, height and color_mode:
     ```python
-    camera = OpenCVCamera(0, fps=30, width=1280, height=720)
-    camera = connect()  # applies the settings, might error out if these settings are not compatible with the camera
-
-    camera = OpenCVCamera(0, fps=90, width=640, height=480)
-    camera = connect()
-
-    camera = OpenCVCamera(0, fps=90, width=640, height=480, color_mode="bgr")
-    camera = connect()
+    config = OpenCVCameraConfig(camera_index=0, fps=30, width=1280, height=720)
+    config = OpenCVCameraConfig(camera_index=0, fps=90, width=640, height=480)
+    config = OpenCVCameraConfig(camera_index=0, fps=90, width=640, height=480, color_mode="bgr")
+    # Note: might error out open `camera.connect()` if these settings are not compatible with the camera
     ```
     """
 
-    def __init__(self, camera_index: int | str, config: OpenCVCameraConfig | None = None, **kwargs):
-        if config is None:
-            config = OpenCVCameraConfig()
-
-        # Overwrite config arguments using kwargs
-        config = replace(config, **kwargs)
-
-        self.camera_index = camera_index
+    def __init__(self, config: OpenCVCameraConfig):
+        self.config = config
+        self.camera_index = config.camera_index
         self.port = None
 
         # Linux uses ports for connecting to cameras
@@ -263,11 +243,22 @@ class OpenCVCamera:
                 # Retrieve the camera index from a potentially symlinked path
                 self.camera_index = get_camera_index_from_unix_port(self.port)
             else:
-                raise ValueError(f"Please check the provided camera_index: {camera_index}")
+                raise ValueError(f"Please check the provided camera_index: {self.camera_index}")
+
+        # Store the raw (capture) resolution from the config.
+        self.capture_width = config.width
+        self.capture_height = config.height
+
+        # If rotated by ±90, swap width and height.
+        if config.rotation in [-90, 90]:
+            self.width = config.height
+            self.height = config.width
+        else:
+            self.width = config.width
+            self.height = config.height
 
         self.fps = config.fps
-        self.width = config.width
-        self.height = config.height
+        self.channels = config.channels
         self.color_mode = config.color_mode
         self.mock = config.mock
 
@@ -283,7 +274,6 @@ class OpenCVCamera:
         else:
             import cv2
 
-        # TODO(aliberts): Do we keep original width/height or do we define them after rotation?
         self.rotation = None
         if config.rotation == -90:
             self.rotation = cv2.ROTATE_90_COUNTERCLOCKWISE
@@ -294,7 +284,7 @@ class OpenCVCamera:
 
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(f"OpenCVCamera({self.camera_index}) is already connected.")
+            raise DeviceAlreadyConnectedError(f"OpenCVCamera({self.camera_index}) is already connected.")
 
         if self.mock:
             import tests.mock_cv2 as cv2
@@ -305,10 +295,20 @@ class OpenCVCamera:
             # when other threads are used to save the images.
             cv2.setNumThreads(1)
 
+        backend = (
+            cv2.CAP_V4L2
+            if platform.system() == "Linux"
+            else cv2.CAP_DSHOW
+            if platform.system() == "Windows"
+            else cv2.CAP_AVFOUNDATION
+            if platform.system() == "Darwin"
+            else cv2.CAP_ANY
+        )
+
         camera_idx = f"/dev/video{self.camera_index}" if platform.system() == "Linux" else self.camera_index
         # First create a temporary camera trying to access `camera_index`,
         # and verify it is a valid camera by calling `isOpened`.
-        tmp_camera = cv2.VideoCapture(camera_idx)
+        tmp_camera = cv2.VideoCapture(camera_idx, backend)
         is_camera_open = tmp_camera.isOpened()
         # Release camera to make it accessible for `find_camera_indices`
         tmp_camera.release()
@@ -331,14 +331,14 @@ class OpenCVCamera:
         # Secondly, create the camera that will be used downstream.
         # Note: For some unknown reason, calling `isOpened` blocks the camera which then
         # needs to be re-created.
-        self.camera = cv2.VideoCapture(camera_idx)
+        self.camera = cv2.VideoCapture(camera_idx, backend)
 
         if self.fps is not None:
             self.camera.set(cv2.CAP_PROP_FPS, self.fps)
-        if self.width is not None:
-            self.camera.set(cv2.CAP_PROP_FRAME_WIDTH, self.width)
-        if self.height is not None:
-            self.camera.set(cv2.CAP_PROP_FRAME_HEIGHT, self.height)
+        if self.capture_width is not None:
+            self.camera.set(cv2.CAP_PROP_FRAME_WIDTH, self.capture_width)
+        if self.capture_height is not None:
+            self.camera.set(cv2.CAP_PROP_FRAME_HEIGHT, self.capture_height)
 
         actual_fps = self.camera.get(cv2.CAP_PROP_FPS)
         actual_width = self.camera.get(cv2.CAP_PROP_FRAME_WIDTH)
@@ -350,19 +350,22 @@ class OpenCVCamera:
             raise OSError(
                 f"Can't set {self.fps=} for OpenCVCamera({self.camera_index}). Actual value is {actual_fps}."
             )
-        if self.width is not None and not math.isclose(self.width, actual_width, rel_tol=1e-3):
+        if self.capture_width is not None and not math.isclose(
+            self.capture_width, actual_width, rel_tol=1e-3
+        ):
             raise OSError(
-                f"Can't set {self.width=} for OpenCVCamera({self.camera_index}). Actual value is {actual_width}."
+                f"Can't set {self.capture_width=} for OpenCVCamera({self.camera_index}). Actual value is {actual_width}."
             )
-        if self.height is not None and not math.isclose(self.height, actual_height, rel_tol=1e-3):
+        if self.capture_height is not None and not math.isclose(
+            self.capture_height, actual_height, rel_tol=1e-3
+        ):
             raise OSError(
-                f"Can't set {self.height=} for OpenCVCamera({self.camera_index}). Actual value is {actual_height}."
+                f"Can't set {self.capture_height=} for OpenCVCamera({self.camera_index}). Actual value is {actual_height}."
             )
 
         self.fps = round(actual_fps)
-        self.width = round(actual_width)
-        self.height = round(actual_height)
-
+        self.capture_width = round(actual_width)
+        self.capture_height = round(actual_height)
         self.is_connected = True
 
     def read(self, temporary_color_mode: str | None = None) -> np.ndarray:
@@ -373,7 +376,7 @@ class OpenCVCamera:
         If you are reading data from other sensors, we advise to use `camera.async_read()` which is non blocking version of `camera.read()`.
         """
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
             )
 
@@ -403,7 +406,7 @@ class OpenCVCamera:
             color_image = cv2.cvtColor(color_image, cv2.COLOR_BGR2RGB)
 
         h, w, _ = color_image.shape
-        if h != self.height or w != self.width:
+        if h != self.capture_height or w != self.capture_width:
             raise OSError(
                 f"Can't capture color image with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
             )
@@ -430,7 +433,7 @@ class OpenCVCamera:
 
     def async_read(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
             )
 
@@ -452,7 +455,7 @@ class OpenCVCamera:
 
     def disconnect(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
             )
 

lerobot/common/cameras/opencv/configuration_opencv.py

@@ -0,0 +1,38 @@
+from dataclasses import dataclass
+
+from ..configs import CameraConfig
+
+
+@CameraConfig.register_subclass("opencv")
+@dataclass
+class OpenCVCameraConfig(CameraConfig):
+    """
+    Example of tested options for Intel Real Sense D405:
+
+    ```python
+    OpenCVCameraConfig(0, 30, 640, 480)
+    OpenCVCameraConfig(0, 60, 640, 480)
+    OpenCVCameraConfig(0, 90, 640, 480)
+    OpenCVCameraConfig(0, 30, 1280, 720)
+    ```
+    """
+
+    camera_index: int
+    fps: int | None = None
+    width: int | None = None
+    height: int | None = None
+    color_mode: str = "rgb"
+    channels: int | None = None
+    rotation: int | None = None
+    mock: bool = False
+
+    def __post_init__(self):
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )
+
+        self.channels = 3
+
+        if self.rotation not in [-90, None, 90, 180]:
+            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")

lerobot/common/cameras/utils.py

@@ -0,0 +1,21 @@
+from .camera import Camera
+from .configs import CameraConfig
+
+
+def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[str, Camera]:
+    cameras = {}
+
+    for key, cfg in camera_configs.items():
+        if cfg.type == "opencv":
+            from .opencv import OpenCVCamera
+
+            cameras[key] = OpenCVCamera(cfg)
+
+        elif cfg.type == "intelrealsense":
+            from .intel.camera_realsense import RealSenseCamera
+
+            cameras[key] = RealSenseCamera(cfg)
+        else:
+            raise ValueError(f"The motor type '{cfg.type}' is not valid.")
+
+    return cameras

lerobot/common/constants.py

@@ -0,0 +1,52 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# keys
+import os
+from pathlib import Path
+
+from huggingface_hub.constants import HF_HOME
+
+OBS_ENV_STATE = "observation.environment_state"
+OBS_STATE = "observation.state"
+OBS_IMAGE = "observation.image"
+OBS_IMAGES = "observation.images"
+ACTION = "action"
+
+ROBOTS = "robots"
+TELEOPERATORS = "teleoperators"
+
+# files & directories
+CHECKPOINTS_DIR = "checkpoints"
+LAST_CHECKPOINT_LINK = "last"
+PRETRAINED_MODEL_DIR = "pretrained_model"
+TRAINING_STATE_DIR = "training_state"
+RNG_STATE = "rng_state.safetensors"
+TRAINING_STEP = "training_step.json"
+OPTIMIZER_STATE = "optimizer_state.safetensors"
+OPTIMIZER_PARAM_GROUPS = "optimizer_param_groups.json"
+SCHEDULER_STATE = "scheduler_state.json"
+
+if "LEROBOT_HOME" in os.environ:
+    raise ValueError(
+        f"You have a 'LEROBOT_HOME' environment variable set to '{os.getenv('LEROBOT_HOME')}'.\n"
+        "'LEROBOT_HOME' is deprecated, please use 'HF_LEROBOT_HOME' instead."
+    )
+
+# cache dir
+default_cache_path = Path(HF_HOME) / "lerobot"
+HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
+
+# calibration dir
+default_calibration_path = HF_LEROBOT_HOME / ".calibration"
+HF_LEROBOT_CALIBRATION = Path(os.getenv("HF_LEROBOT_CALIBRATION", default_calibration_path)).expanduser()

lerobot/common/datasets/backward_compatibility.py

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

lerobot/common/datasets/card_template.md

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

lerobot/common/datasets/compute_stats.py

@@ -13,201 +13,164 @@
 # 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 numpy as np
 
-import einops
-import torch
-import tqdm
-from datasets import Image
-
-from lerobot.common.datasets.video_utils import VideoFrame
+from lerobot.common.datasets.utils import load_image_as_numpy
 
 
-def get_stats_einops_patterns(dataset, num_workers=0):
-    """These einops patterns will be used to aggregate batches and compute statistics.
+def estimate_num_samples(
+    dataset_len: int, min_num_samples: int = 100, max_num_samples: int = 10_000, power: float = 0.75
+) -> int:
+    """Heuristic to estimate the number of samples based on dataset size.
+    The power controls the sample growth relative to dataset size.
+    Lower the power for less number of samples.
 
-    Note: We assume the images are in channel first format
+    For default arguments, we have:
+    - from 1 to ~500, num_samples=100
+    - at 1000, num_samples=177
+    - at 2000, num_samples=299
+    - at 5000, num_samples=594
+    - at 10000, num_samples=1000
+    - at 20000, num_samples=1681
     """
+    if dataset_len < min_num_samples:
+        min_num_samples = dataset_len
+    return max(min_num_samples, min(int(dataset_len**power), max_num_samples))
 
-    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():
-        # NOTE: skip language_instruction embedding in stats computation
-        if key == "language_instruction":
-            continue
+def sample_indices(data_len: int) -> list[int]:
+    num_samples = estimate_num_samples(data_len)
+    return np.round(np.linspace(0, data_len - 1, num_samples)).astype(int).tolist()
 
-        # 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}"
+def auto_downsample_height_width(img: np.ndarray, target_size: int = 150, max_size_threshold: int = 300):
+    _, height, width = img.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()=}"
+    if max(width, height) < max_size_threshold:
+        # no downsampling needed
+        return img
 
-            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"
+    downsample_factor = int(width / target_size) if width > height else int(height / target_size)
+    return img[:, ::downsample_factor, ::downsample_factor]
+
+
+def sample_images(image_paths: list[str]) -> np.ndarray:
+    sampled_indices = sample_indices(len(image_paths))
+
+    images = None
+    for i, idx in enumerate(sampled_indices):
+        path = image_paths[idx]
+        # we load as uint8 to reduce memory usage
+        img = load_image_as_numpy(path, dtype=np.uint8, channel_first=True)
+        img = auto_downsample_height_width(img)
+
+        if images is None:
+            images = np.empty((len(sampled_indices), *img.shape), dtype=np.uint8)
+
+        images[i] = img
+
+    return images
+
+
+def get_feature_stats(array: np.ndarray, axis: tuple, keepdims: bool) -> dict[str, np.ndarray]:
+    return {
+        "min": np.min(array, axis=axis, keepdims=keepdims),
+        "max": np.max(array, axis=axis, keepdims=keepdims),
+        "mean": np.mean(array, axis=axis, keepdims=keepdims),
+        "std": np.std(array, axis=axis, keepdims=keepdims),
+        "count": np.array([len(array)]),
+    }
+
+
+def compute_episode_stats(episode_data: dict[str, list[str] | np.ndarray], features: dict) -> dict:
+    ep_stats = {}
+    for key, data in episode_data.items():
+        if features[key]["dtype"] == "string":
+            continue  # HACK: we should receive np.arrays of strings
+        elif features[key]["dtype"] in ["image", "video"]:
+            ep_ft_array = sample_images(data)  # data is a list of image paths
+            axes_to_reduce = (0, 2, 3)  # keep channel dim
+            keepdims = True
         else:
-            raise ValueError(f"{key}, {feats_type}, {batch[key].shape}")
+            ep_ft_array = data  # data is already a np.ndarray
+            axes_to_reduce = 0  # compute stats over the first axis
+            keepdims = data.ndim == 1  # keep as np.array
 
-    return stats_patterns
+        ep_stats[key] = get_feature_stats(ep_ft_array, axis=axes_to_reduce, keepdims=keepdims)
+
+        # finally, we normalize and remove batch dim for images
+        if features[key]["dtype"] in ["image", "video"]:
+            ep_stats[key] = {
+                k: v if k == "count" else np.squeeze(v / 255.0, axis=0) for k, v in ep_stats[key].items()
+            }
+
+    return ep_stats
 
 
-def compute_stats(dataset, batch_size=8, num_workers=8, 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 _assert_type_and_shape(stats_list: list[dict[str, dict]]):
+    for i in range(len(stats_list)):
+        for fkey in stats_list[i]:
+            for k, v in stats_list[i][fkey].items():
+                if not isinstance(v, np.ndarray):
+                    raise ValueError(
+                        f"Stats must be composed of numpy array, but key '{k}' of feature '{fkey}' is of type '{type(v)}' instead."
+                    )
+                if v.ndim == 0:
+                    raise ValueError("Number of dimensions must be at least 1, and is 0 instead.")
+                if k == "count" and v.shape != (1,):
+                    raise ValueError(f"Shape of 'count' must be (1), but is {v.shape} instead.")
+                if "image" in fkey and k != "count" and v.shape != (3, 1, 1):
+                    raise ValueError(f"Shape of '{k}' must be (3,1,1), but is {v.shape} instead.")
 
 
-def aggregate_stats(ls_datasets) -> dict[str, torch.Tensor]:
-    """Aggregate stats of multiple LeRobot datasets into one set of stats without recomputing from scratch.
+def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
+    """Aggregates stats for a single feature."""
+    means = np.stack([s["mean"] for s in stats_ft_list])
+    variances = np.stack([s["std"] ** 2 for s in stats_ft_list])
+    counts = np.stack([s["count"] for s in stats_ft_list])
+    total_count = counts.sum(axis=0)
 
-    The final stats will have the union of all data keys from each of the datasets.
+    # Prepare weighted mean by matching number of dimensions
+    while counts.ndim < means.ndim:
+        counts = np.expand_dims(counts, axis=-1)
 
-    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, ...)
+    # Compute the weighted mean
+    weighted_means = means * counts
+    total_mean = weighted_means.sum(axis=0) / total_count
+
+    # Compute the variance using the parallel algorithm
+    delta_means = means - total_mean
+    weighted_variances = (variances + delta_means**2) * counts
+    total_variance = weighted_variances.sum(axis=0) / total_count
+
+    return {
+        "min": np.min(np.stack([s["min"] for s in stats_ft_list]), axis=0),
+        "max": np.max(np.stack([s["max"] for s in stats_ft_list]), axis=0),
+        "mean": total_mean,
+        "std": np.sqrt(total_variance),
+        "count": total_count,
+    }
+
+
+def aggregate_stats(stats_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
+    """Aggregate stats from multiple compute_stats outputs into a single set of stats.
+
+    The final stats will have the union of all data keys from each of the stats dicts.
+
+    For instance:
     - new_min = min(min_dataset_0, min_dataset_1, ...)
-    - new_mean = (mean of all data)
+    - new_max = max(max_dataset_0, max_dataset_1, ...)
+    - new_mean = (mean of all data, weighted by counts)
     - 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
+
+    _assert_type_and_shape(stats_list)
+
+    data_keys = {key for stats in stats_list for key in stats}
+    aggregated_stats = {key: {} for key in data_keys}
+
+    for key in data_keys:
+        stats_with_key = [stats[key] for stats in stats_list if key in stats]
+        aggregated_stats[key] = aggregate_feature_stats(stats_with_key)
+
+    return aggregated_stats

lerobot/common/datasets/factory.py

@@ -14,104 +14,105 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
+from pprint import pformat
 
 import torch
-from omegaconf import ListConfig, OmegaConf
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
-from lerobot.common.datasets.transforms import get_image_transforms
+from lerobot.common.datasets.lerobot_dataset import (
+    LeRobotDataset,
+    LeRobotDatasetMetadata,
+    MultiLeRobotDataset,
+)
+from lerobot.common.datasets.transforms import ImageTransforms
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.train import TrainPipelineConfig
+
+IMAGENET_STATS = {
+    "mean": [[[0.485]], [[0.456]], [[0.406]]],  # (c,1,1)
+    "std": [[[0.229]], [[0.224]], [[0.225]]],  # (c,1,1)
+}
 
 
-def resolve_delta_timestamps(cfg):
-    """Resolves delta_timestamps config key (in-place) by using `eval`.
+def resolve_delta_timestamps(
+    cfg: PreTrainedConfig, ds_meta: LeRobotDatasetMetadata
+) -> dict[str, list] | None:
+    """Resolves delta_timestamps by reading from the 'delta_indices' properties of the PreTrainedConfig.
 
-    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):
-                # TODO(rcadene, alexander-soare): remove `eval` to avoid exploit
-                cfg.training.delta_timestamps[key] = eval(delta_timestamps[key])
-
-
-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]"`.
+        cfg (PreTrainedConfig): The PreTrainedConfig to read delta_indices from.
+        ds_meta (LeRobotDatasetMetadata): The dataset from which features and fps are used to build
+            delta_timestamps against.
+
     Returns:
-        The LeRobotDataset.
+        dict[str, list] | None: A dictionary of delta_timestamps, e.g.:
+            {
+                "observation.state": [-0.04, -0.02, 0]
+                "observation.action": [-0.02, 0, 0.02]
+            }
+            returns `None` if the the resulting dict is empty.
     """
-    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."
+    delta_timestamps = {}
+    for key in ds_meta.features:
+        if key == "next.reward" and cfg.reward_delta_indices is not None:
+            delta_timestamps[key] = [i / ds_meta.fps for i in cfg.reward_delta_indices]
+        if key == "action" and cfg.action_delta_indices is not None:
+            delta_timestamps[key] = [i / ds_meta.fps for i in cfg.action_delta_indices]
+        if key.startswith("observation.") and cfg.observation_delta_indices is not None:
+            delta_timestamps[key] = [i / ds_meta.fps for i in cfg.observation_delta_indices]
+
+    if len(delta_timestamps) == 0:
+        delta_timestamps = None
+
+    return delta_timestamps
+
+
+def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDataset:
+    """Handles the logic of setting up delta timestamps and image transforms before creating a dataset.
+
+    Args:
+        cfg (TrainPipelineConfig): A TrainPipelineConfig config which contains a DatasetConfig and a PreTrainedConfig.
+
+    Raises:
+        NotImplementedError: The MultiLeRobotDataset is currently deactivated.
+
+    Returns:
+        LeRobotDataset | MultiLeRobotDataset
+    """
+    image_transforms = (
+        ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None
+    )
+
+    if isinstance(cfg.dataset.repo_id, str):
+        ds_meta = LeRobotDatasetMetadata(
+            cfg.dataset.repo_id, root=cfg.dataset.root, revision=cfg.dataset.revision
         )
-
-    # 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):
+        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
         dataset = LeRobotDataset(
-            cfg.dataset_repo_id,
-            split=split,
-            delta_timestamps=cfg.training.get("delta_timestamps"),
+            cfg.dataset.repo_id,
+            root=cfg.dataset.root,
+            episodes=cfg.dataset.episodes,
+            delta_timestamps=delta_timestamps,
             image_transforms=image_transforms,
-            video_backend=cfg.video_backend,
+            revision=cfg.dataset.revision,
+            video_backend=cfg.dataset.video_backend,
         )
     else:
+        raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")
         dataset = MultiLeRobotDataset(
-            cfg.dataset_repo_id,
-            split=split,
-            delta_timestamps=cfg.training.get("delta_timestamps"),
+            cfg.dataset.repo_id,
+            # TODO(aliberts): add proper support for multi dataset
+            # delta_timestamps=delta_timestamps,
             image_transforms=image_transforms,
-            video_backend=cfg.video_backend,
+            video_backend=cfg.dataset.video_backend,
+        )
+        logging.info(
+            "Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
+            f"{pformat(dataset.repo_id_to_index, indent=2)}"
         )
 
-    if cfg.get("override_dataset_stats"):
-        for key, stats_dict in cfg.override_dataset_stats.items():
-            for stats_type, listconfig in stats_dict.items():
-                # example of stats_type: min, max, mean, std
-                stats = OmegaConf.to_container(listconfig, resolve=True)
-                dataset.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
+    if cfg.dataset.use_imagenet_stats:
+        for key in dataset.meta.camera_keys:
+            for stats_type, stats in IMAGENET_STATS.items():
+                dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
     return dataset

lerobot/common/datasets/image_writer.py

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

lerobot/common/datasets/online_buffer.py

@@ -187,7 +187,7 @@ class OnlineBuffer(torch.utils.data.Dataset):
         assert data[OnlineBuffer.INDEX_KEY][0].item() == 0
 
         # Shift the incoming indices if necessary.
-        if self.num_samples > 0:
+        if self.num_frames > 0:
             last_episode_index = self._data[OnlineBuffer.EPISODE_INDEX_KEY][next_index - 1]
             last_data_index = self._data[OnlineBuffer.INDEX_KEY][next_index - 1]
             data[OnlineBuffer.EPISODE_INDEX_KEY] += last_episode_index + 1
@@ -227,11 +227,11 @@ class OnlineBuffer(torch.utils.data.Dataset):
         )
 
     @property
-    def num_samples(self) -> int:
+    def num_frames(self) -> int:
         return np.count_nonzero(self._data[OnlineBuffer.OCCUPANCY_MASK_KEY])
 
     def __len__(self):
-        return self.num_samples
+        return self.num_frames
 
     def _item_to_tensors(self, item: dict) -> dict:
         item_ = {}

lerobot/common/datasets/populate_dataset.py

@@ -1,468 +0,0 @@
-"""Functions to create an empty dataset, and populate it with frames."""
-# TODO(rcadene, aliberts): to adapt as class methods of next version of LeRobotDataset
-
-import concurrent
-import json
-import logging
-import multiprocessing
-import shutil
-from pathlib import Path
-
-import torch
-import tqdm
-from PIL import Image
-
-from lerobot.common.datasets.compute_stats import compute_stats
-from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import to_hf_dataset
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, get_default_encoding
-from lerobot.common.datasets.utils import calculate_episode_data_index, create_branch
-from lerobot.common.datasets.video_utils import encode_video_frames
-from lerobot.common.utils.utils import log_say
-from lerobot.scripts.push_dataset_to_hub import (
-    push_dataset_card_to_hub,
-    push_meta_data_to_hub,
-    push_videos_to_hub,
-    save_meta_data,
-)
-
-########################################################################################
-# Asynchrounous saving of images on disk
-########################################################################################
-
-
-def safe_stop_image_writer(func):
-    # TODO(aliberts): Allow to pass custom exceptions
-    # (e.g. ThreadServiceExit, KeyboardInterrupt, SystemExit, UnpluggedError, DynamixelCommError)
-    def wrapper(*args, **kwargs):
-        try:
-            return func(*args, **kwargs)
-        except Exception as e:
-            image_writer = kwargs.get("dataset", {}).get("image_writer")
-            if image_writer is not None:
-                print("Waiting for image writer to terminate...")
-                stop_image_writer(image_writer, timeout=20)
-            raise e
-
-    return wrapper
-
-
-def save_image(img_tensor, key, frame_index, episode_index, videos_dir: str):
-    img = Image.fromarray(img_tensor.numpy())
-    path = Path(videos_dir) / f"{key}_episode_{episode_index:06d}" / f"frame_{frame_index:06d}.png"
-    path.parent.mkdir(parents=True, exist_ok=True)
-    img.save(str(path), quality=100)
-
-
-def loop_to_save_images_in_threads(image_queue, num_threads):
-    if num_threads < 1:
-        raise NotImplementedError(f"Only `num_threads>=1` is supported for now, but {num_threads=} given.")
-
-    with concurrent.futures.ThreadPoolExecutor(max_workers=num_threads) as executor:
-        futures = []
-        while True:
-            # Blocks until a frame is available
-            frame_data = image_queue.get()
-
-            # As usually done, exit loop when receiving None to stop the worker
-            if frame_data is None:
-                break
-
-            image, key, frame_index, episode_index, videos_dir = frame_data
-            futures.append(executor.submit(save_image, image, key, frame_index, episode_index, videos_dir))
-
-        # Before exiting function, wait for all threads to complete
-        with tqdm.tqdm(total=len(futures), desc="Writing images") as progress_bar:
-            concurrent.futures.wait(futures)
-            progress_bar.update(len(futures))
-
-
-def start_image_writer_processes(image_queue, num_processes, num_threads_per_process):
-    if num_processes < 1:
-        raise ValueError(f"Only `num_processes>=1` is supported, but {num_processes=} given.")
-
-    if num_threads_per_process < 1:
-        raise NotImplementedError(
-            "Only `num_threads_per_process>=1` is supported for now, but {num_threads_per_process=} given."
-        )
-
-    processes = []
-    for _ in range(num_processes):
-        process = multiprocessing.Process(
-            target=loop_to_save_images_in_threads,
-            args=(image_queue, num_threads_per_process),
-        )
-        process.start()
-        processes.append(process)
-    return processes
-
-
-def stop_processes(processes, queue, timeout):
-    # Send None to each process to signal them to stop
-    for _ in processes:
-        queue.put(None)
-
-    # Wait maximum 20 seconds for all processes to terminate
-    for process in processes:
-        process.join(timeout=timeout)
-
-    # If not terminated after 20 seconds, force termination
-    if process.is_alive():
-        process.terminate()
-
-    # Close the queue, no more items can be put in the queue
-    queue.close()
-
-    # Ensure all background queue threads have finished
-    queue.join_thread()
-
-
-def start_image_writer(num_processes, num_threads):
-    """This function abstract away the initialisation of processes or/and threads to
-    save images on disk asynchrounously, which is critical to control a robot and record data
-    at a high frame rate.
-
-    When `num_processes=0`, it returns a dictionary containing a threads pool of size `num_threads`.
-    When `num_processes>0`, it returns a dictionary containing a processes pool of size `num_processes`,
-    where each subprocess starts their own threads pool of size `num_threads`.
-
-    The optimal number of processes and threads depends on your computer capabilities.
-    We advise to use 4 threads per camera with 0 processes. If the fps is not stable, try to increase or lower
-    the number of threads. If it is still not stable, try to use 1 subprocess, or more.
-    """
-    image_writer = {}
-
-    if num_processes == 0:
-        futures = []
-        threads_pool = concurrent.futures.ThreadPoolExecutor(max_workers=num_threads)
-        image_writer["threads_pool"], image_writer["futures"] = threads_pool, futures
-    else:
-        # TODO(rcadene): When using num_processes>1, `multiprocessing.Manager().Queue()`
-        # might be better than `multiprocessing.Queue()`. Source: https://www.geeksforgeeks.org/python-multiprocessing-queue-vs-multiprocessing-manager-queue
-        image_queue = multiprocessing.Queue()
-        processes_pool = start_image_writer_processes(
-            image_queue, num_processes=num_processes, num_threads_per_process=num_threads
-        )
-        image_writer["processes_pool"], image_writer["image_queue"] = processes_pool, image_queue
-
-    return image_writer
-
-
-def async_save_image(image_writer, image, key, frame_index, episode_index, videos_dir):
-    """This function abstract away the saving of an image on disk asynchrounously. It uses a dictionary
-    called image writer which contains either a pool of processes or a pool of threads.
-    """
-    if "threads_pool" in image_writer:
-        threads_pool, futures = image_writer["threads_pool"], image_writer["futures"]
-        futures.append(threads_pool.submit(save_image, image, key, frame_index, episode_index, videos_dir))
-    else:
-        image_queue = image_writer["image_queue"]
-        image_queue.put((image, key, frame_index, episode_index, videos_dir))
-
-
-def stop_image_writer(image_writer, timeout):
-    if "threads_pool" in image_writer:
-        futures = image_writer["futures"]
-        # Before exiting function, wait for all threads to complete
-        with tqdm.tqdm(total=len(futures), desc="Writing images") as progress_bar:
-            concurrent.futures.wait(futures, timeout=timeout)
-            progress_bar.update(len(futures))
-    else:
-        processes_pool, image_queue = image_writer["processes_pool"], image_writer["image_queue"]
-        stop_processes(processes_pool, image_queue, timeout=timeout)
-
-
-########################################################################################
-# Functions to initialize, resume and populate a dataset
-########################################################################################
-
-
-def init_dataset(
-    repo_id,
-    root,
-    force_override,
-    fps,
-    video,
-    write_images,
-    num_image_writer_processes,
-    num_image_writer_threads,
-):
-    local_dir = Path(root) / repo_id
-    if local_dir.exists() and force_override:
-        shutil.rmtree(local_dir)
-
-    episodes_dir = local_dir / "episodes"
-    episodes_dir.mkdir(parents=True, exist_ok=True)
-
-    videos_dir = local_dir / "videos"
-    videos_dir.mkdir(parents=True, exist_ok=True)
-
-    # Logic to resume data recording
-    rec_info_path = episodes_dir / "data_recording_info.json"
-    if rec_info_path.exists():
-        with open(rec_info_path) as f:
-            rec_info = json.load(f)
-        num_episodes = rec_info["last_episode_index"] + 1
-    else:
-        num_episodes = 0
-
-    dataset = {
-        "repo_id": repo_id,
-        "local_dir": local_dir,
-        "videos_dir": videos_dir,
-        "episodes_dir": episodes_dir,
-        "fps": fps,
-        "video": video,
-        "rec_info_path": rec_info_path,
-        "num_episodes": num_episodes,
-    }
-
-    if write_images:
-        # Initialize processes or/and threads dedicated to save images on disk asynchronously,
-        # which is critical to control a robot and record data at a high frame rate.
-        image_writer = start_image_writer(
-            num_processes=num_image_writer_processes,
-            num_threads=num_image_writer_threads,
-        )
-        dataset["image_writer"] = image_writer
-
-    return dataset
-
-
-def add_frame(dataset, observation, action):
-    if "current_episode" not in dataset:
-        # initialize episode dictionary
-        ep_dict = {}
-        for key in observation:
-            if key not in ep_dict:
-                ep_dict[key] = []
-        for key in action:
-            if key not in ep_dict:
-                ep_dict[key] = []
-
-        ep_dict["episode_index"] = []
-        ep_dict["frame_index"] = []
-        ep_dict["timestamp"] = []
-        ep_dict["next.done"] = []
-
-        dataset["current_episode"] = ep_dict
-        dataset["current_frame_index"] = 0
-
-    ep_dict = dataset["current_episode"]
-    episode_index = dataset["num_episodes"]
-    frame_index = dataset["current_frame_index"]
-    videos_dir = dataset["videos_dir"]
-    video = dataset["video"]
-    fps = dataset["fps"]
-
-    ep_dict["episode_index"].append(episode_index)
-    ep_dict["frame_index"].append(frame_index)
-    ep_dict["timestamp"].append(frame_index / fps)
-    ep_dict["next.done"].append(False)
-
-    img_keys = [key for key in observation if "image" in key]
-    non_img_keys = [key for key in observation if "image" not in key]
-
-    # Save all observed modalities except images
-    for key in non_img_keys:
-        ep_dict[key].append(observation[key])
-
-    # Save actions
-    for key in action:
-        ep_dict[key].append(action[key])
-
-    if "image_writer" not in dataset:
-        dataset["current_frame_index"] += 1
-        return
-
-    # Save images
-    image_writer = dataset["image_writer"]
-    for key in img_keys:
-        imgs_dir = videos_dir / f"{key}_episode_{episode_index:06d}"
-        async_save_image(
-            image_writer,
-            image=observation[key],
-            key=key,
-            frame_index=frame_index,
-            episode_index=episode_index,
-            videos_dir=str(videos_dir),
-        )
-
-        if video:
-            fname = f"{key}_episode_{episode_index:06d}.mp4"
-            frame_info = {"path": f"videos/{fname}", "timestamp": frame_index / fps}
-        else:
-            frame_info = str(imgs_dir / f"frame_{frame_index:06d}.png")
-
-        ep_dict[key].append(frame_info)
-
-    dataset["current_frame_index"] += 1
-
-
-def delete_current_episode(dataset):
-    del dataset["current_episode"]
-    del dataset["current_frame_index"]
-
-    # delete temporary images
-    episode_index = dataset["num_episodes"]
-    videos_dir = dataset["videos_dir"]
-    for tmp_imgs_dir in videos_dir.glob(f"*_episode_{episode_index:06d}"):
-        shutil.rmtree(tmp_imgs_dir)
-
-
-def save_current_episode(dataset):
-    episode_index = dataset["num_episodes"]
-    ep_dict = dataset["current_episode"]
-    episodes_dir = dataset["episodes_dir"]
-    rec_info_path = dataset["rec_info_path"]
-
-    ep_dict["next.done"][-1] = True
-
-    for key in ep_dict:
-        if "observation" in key and "image" not in key:
-            ep_dict[key] = torch.stack(ep_dict[key])
-
-    ep_dict["action"] = torch.stack(ep_dict["action"])
-    ep_dict["episode_index"] = torch.tensor(ep_dict["episode_index"])
-    ep_dict["frame_index"] = torch.tensor(ep_dict["frame_index"])
-    ep_dict["timestamp"] = torch.tensor(ep_dict["timestamp"])
-    ep_dict["next.done"] = torch.tensor(ep_dict["next.done"])
-
-    ep_path = episodes_dir / f"episode_{episode_index}.pth"
-    torch.save(ep_dict, ep_path)
-
-    rec_info = {
-        "last_episode_index": episode_index,
-    }
-    with open(rec_info_path, "w") as f:
-        json.dump(rec_info, f)
-
-    # force re-initialization of episode dictionnary during add_frame
-    del dataset["current_episode"]
-
-    dataset["num_episodes"] += 1
-
-
-def encode_videos(dataset, image_keys, play_sounds):
-    log_say("Encoding videos", play_sounds)
-
-    num_episodes = dataset["num_episodes"]
-    videos_dir = dataset["videos_dir"]
-    local_dir = dataset["local_dir"]
-    fps = dataset["fps"]
-
-    # Use ffmpeg to convert frames stored as png into mp4 videos
-    for episode_index in tqdm.tqdm(range(num_episodes)):
-        for key in image_keys:
-            # key = f"observation.images.{name}"
-            tmp_imgs_dir = videos_dir / f"{key}_episode_{episode_index:06d}"
-            fname = f"{key}_episode_{episode_index:06d}.mp4"
-            video_path = local_dir / "videos" / fname
-            if video_path.exists():
-                # Skip if video is already encoded. Could be the case when resuming data recording.
-                continue
-            # note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
-            # since video encoding with ffmpeg is already using multithreading.
-            encode_video_frames(tmp_imgs_dir, video_path, fps, overwrite=True)
-            shutil.rmtree(tmp_imgs_dir)
-
-
-def from_dataset_to_lerobot_dataset(dataset, play_sounds):
-    log_say("Consolidate episodes", play_sounds)
-
-    num_episodes = dataset["num_episodes"]
-    episodes_dir = dataset["episodes_dir"]
-    videos_dir = dataset["videos_dir"]
-    video = dataset["video"]
-    fps = dataset["fps"]
-    repo_id = dataset["repo_id"]
-
-    ep_dicts = []
-    for episode_index in tqdm.tqdm(range(num_episodes)):
-        ep_path = episodes_dir / f"episode_{episode_index}.pth"
-        ep_dict = torch.load(ep_path)
-        ep_dicts.append(ep_dict)
-    data_dict = concatenate_episodes(ep_dicts)
-
-    if video:
-        image_keys = [key for key in data_dict if "image" in key]
-        encode_videos(dataset, image_keys, play_sounds)
-
-    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()
-
-    lerobot_dataset = LeRobotDataset.from_preloaded(
-        repo_id=repo_id,
-        hf_dataset=hf_dataset,
-        episode_data_index=episode_data_index,
-        info=info,
-        videos_dir=videos_dir,
-    )
-
-    return lerobot_dataset
-
-
-def save_lerobot_dataset_on_disk(lerobot_dataset):
-    hf_dataset = lerobot_dataset.hf_dataset
-    info = lerobot_dataset.info
-    stats = lerobot_dataset.stats
-    episode_data_index = lerobot_dataset.episode_data_index
-    local_dir = lerobot_dataset.videos_dir.parent
-    meta_data_dir = local_dir / "meta_data"
-
-    hf_dataset = hf_dataset.with_format(None)  # to remove transforms that cant be saved
-    hf_dataset.save_to_disk(str(local_dir / "train"))
-
-    save_meta_data(info, stats, episode_data_index, meta_data_dir)
-
-
-def push_lerobot_dataset_to_hub(lerobot_dataset, tags):
-    hf_dataset = lerobot_dataset.hf_dataset
-    local_dir = lerobot_dataset.videos_dir.parent
-    videos_dir = lerobot_dataset.videos_dir
-    repo_id = lerobot_dataset.repo_id
-    video = lerobot_dataset.video
-    meta_data_dir = local_dir / "meta_data"
-
-    if not (local_dir / "train").exists():
-        raise ValueError(
-            "You need to run `save_lerobot_dataset_on_disk(lerobot_dataset)` before pushing to the hub."
-        )
-
-    hf_dataset.push_to_hub(repo_id, revision="main")
-    push_meta_data_to_hub(repo_id, meta_data_dir, revision="main")
-    push_dataset_card_to_hub(repo_id, revision="main", tags=tags)
-    if video:
-        push_videos_to_hub(repo_id, videos_dir, revision="main")
-    create_branch(repo_id, repo_type="dataset", branch=CODEBASE_VERSION)
-
-
-def create_lerobot_dataset(dataset, run_compute_stats, push_to_hub, tags, play_sounds):
-    if "image_writer" in dataset:
-        logging.info("Waiting for image writer to terminate...")
-        image_writer = dataset["image_writer"]
-        stop_image_writer(image_writer, timeout=20)
-
-    lerobot_dataset = from_dataset_to_lerobot_dataset(dataset, play_sounds)
-
-    if run_compute_stats:
-        log_say("Computing dataset statistics", play_sounds)
-        lerobot_dataset.stats = compute_stats(lerobot_dataset)
-    else:
-        logging.info("Skipping computation of the dataset statistics")
-        lerobot_dataset.stats = {}
-
-    save_lerobot_dataset_on_disk(lerobot_dataset)
-
-    if push_to_hub:
-        push_lerobot_dataset_to_hub(lerobot_dataset, tags)
-
-    return lerobot_dataset

lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md

@@ -1,56 +0,0 @@
-## 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/_download_raw.py

@@ -152,7 +152,7 @@ def download_raw(raw_dir: Path, repo_id: str):
             stacklevel=1,
         )
 
-    # Send warning if raw_dir isn't well formated
+    # Send warning if raw_dir isn't well formatted
     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

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -30,12 +30,12 @@ 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 (
+    calculate_episode_data_index,
     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

lerobot/common/datasets/push_dataset_to_hub/cam_png_format.py

@@ -24,8 +24,11 @@ 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.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+)
+from lerobot.common.datasets.utils import hf_transform_to_torch
 from lerobot.common.datasets.video_utils import VideoFrame
 
 

lerobot/common/datasets/push_dataset_to_hub/dora_parquet_format.py

@@ -26,8 +26,8 @@ import torch
 from datasets import Dataset, Features, Image, Sequence, Value
 
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import calculate_episode_data_index
 from lerobot.common.datasets.utils import (
-    calculate_episode_data_index,
     hf_transform_to_torch,
 )
 from lerobot.common.datasets.video_utils import VideoFrame
@@ -68,11 +68,11 @@ def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episod
             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".
+            # matching timestamps that are too far apart, 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.
+            # are too far apart.
             direction="nearest",
-            tolerance=pd.Timedelta(f"{1/fps} seconds"),
+            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]
@@ -126,7 +126,7 @@ def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episod
     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
+    # sanity check the video paths are well formatted
     for key in df:
         if "observation.images." not in key:
             continue
@@ -143,7 +143,7 @@ def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episod
             # 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
+            # sanity check the video path is well formatted
             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}")

lerobot/common/datasets/push_dataset_to_hub/openx/configs.yaml

@@ -1,639 +0,0 @@
-OPENX_DATASET_CONFIGS:
-  fractal20220817_data:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - base_pose_tool_reached
-      - gripper_closed
-    fps: 3
-
-  kuka:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - clip_function_input/base_pose_tool_reached
-      - gripper_closed
-    fps: 10
-
-  bridge_openx:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - EEF_state
-      - gripper_state
-    fps: 5
-
-  taco_play:
-    image_obs_keys:
-      - rgb_static
-      - rgb_gripper
-    depth_obs_keys:
-      - depth_static
-      - depth_gripper
-    state_obs_keys:
-      - state_eef
-      - state_gripper
-    fps: 15
-
-  jaco_play:
-    image_obs_keys:
-      - image
-      - image_wrist
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state_eef
-      - state_gripper
-    fps: 10
-
-  berkeley_cable_routing:
-    image_obs_keys:
-      - image
-      - top_image
-      - wrist45_image
-      - wrist225_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - robot_state
-    fps: 10
-
-  roboturk:
-    image_obs_keys:
-      - front_rgb
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 10
-
-  nyu_door_opening_surprising_effectiveness:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 3
-
-  viola:
-    image_obs_keys:
-      - agentview_rgb
-      - eye_in_hand_rgb
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_states
-      - gripper_states
-    fps: 20
-
-  berkeley_autolab_ur5:
-    image_obs_keys:
-      - image
-      - hand_image
-    depth_obs_keys:
-      - image_with_depth
-    state_obs_keys:
-      - state
-    fps: 5
-
-  toto:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 30
-
-  language_table:
-    image_obs_keys:
-      - rgb
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - effector_translation
-    fps: 10
-
-  columbia_cairlab_pusht_real:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - robot_state
-    fps: 10
-
-  stanford_kuka_multimodal_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - depth_image
-    state_obs_keys:
-      - ee_position
-      - ee_orientation
-    fps: 20
-
-  nyu_rot_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 3
-
-  io_ai_tech:
-    image_obs_keys:
-      - image
-      - image_fisheye
-      - image_left_side
-      - image_right_side
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 3
-
-  stanford_hydra_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  austin_buds_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  nyu_franka_play_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - image_additional_view
-    depth_obs_keys:
-      - depth
-      - depth_additional_view
-    state_obs_keys:
-      - eef_state
-    fps: 3
-
-  maniskill_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - depth
-      - wrist_depth
-    state_obs_keys:
-      - tcp_pose
-      - gripper_state
-    fps: 20
-
-  furniture_bench_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  cmu_franka_exploration_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - highres_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 10
-
-  ucsd_kitchen_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_state
-    fps: 2
-
-  ucsd_pick_and_place_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 3
-
-  spoc:
-    image_obs_keys:
-      - image
-      - image_manipulation
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 3
-
-  austin_sailor_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  austin_sirius_dataset_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  bc_z:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - present/xyz
-      - present/axis_angle
-      - present/sensed_close
-    fps: 10
-
-  utokyo_pr2_opening_fridge_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  utokyo_pr2_tabletop_manipulation_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  utokyo_xarm_pick_and_place_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - image2
-      - hand_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - end_effector_pose
-    fps: 10
-
-  utokyo_xarm_bimanual_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - pose_r
-    fps: 10
-
-  robo_net:
-    image_obs_keys:
-      - image
-      - image1
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 1
-
-  robo_set:
-    image_obs_keys:
-      - image_left
-      - image_right
-      - image_wrist
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-      - state_velocity
-    fps: 5
-
-  berkeley_mvp_converted_externally_to_rlds:
-    image_obs_keys:
-      - hand_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - gripper
-      - pose
-      - joint_pos
-    fps: 5
-
-  berkeley_rpt_converted_externally_to_rlds:
-    image_obs_keys:
-      - hand_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_pos
-      - gripper
-    fps: 30
-
-  kaist_nonprehensile_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  stanford_mask_vit_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-
-  tokyo_u_lsmo_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  dlr_sara_pour_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  dlr_sara_grid_clamp_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  dlr_edan_shared_control_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 5
-
-  asu_table_top_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 12.5
-
-  stanford_robocook_converted_externally_to_rlds:
-    image_obs_keys:
-      - image_1
-      - image_2
-    depth_obs_keys:
-      - depth_1
-      - depth_2
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 5
-
-  imperialcollege_sawyer_wrist_cam:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  iamlab_cmu_pickup_insert_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_state
-      - gripper_state
-    fps: 20
-
-  uiuc_d3field:
-    image_obs_keys:
-      - image_1
-      - image_2
-    depth_obs_keys:
-      - depth_1
-      - depth_2
-    state_obs_keys:
-      - null
-    fps: 1
-
-  utaustin_mutex:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  berkeley_fanuc_manipulation:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - joint_state
-      - gripper_state
-    fps: 10
-
-  cmu_playing_with_food:
-    image_obs_keys:
-      - image
-      - finger_vision_1
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 10
-
-  cmu_play_fusion:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 5
-
-  cmu_stretch:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - eef_state
-      - gripper_state
-    fps: 10
-
-  berkeley_gnm_recon:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-      - position
-      - yaw
-    fps: 3
-
-  berkeley_gnm_cory_hall:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-      - position
-      - yaw
-    fps: 5
-
-  berkeley_gnm_sac_son:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-      - position
-      - yaw
-    fps: 10
-
-  droid:
-    image_obs_keys:
-      - exterior_image_1_left
-      - exterior_image_2_left
-      - wrist_image_left
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - proprio
-    fps: 15
-
-  droid_100:
-    image_obs_keys:
-      - exterior_image_1_left
-      - exterior_image_2_left
-      - wrist_image_left
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - proprio
-    fps: 15
-
-  fmb:
-    image_obs_keys:
-      - image_side_1
-      - image_side_2
-      - image_wrist_1
-      - image_wrist_2
-    depth_obs_keys:
-      - image_side_1_depth
-      - image_side_2_depth
-      - image_wrist_1_depth
-      - image_wrist_2_depth
-    state_obs_keys:
-      - proprio
-    fps: 10
-
-  dobbe:
-    image_obs_keys:
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - proprio
-    fps: 3.75
-
-  usc_cloth_sim_converted_externally_to_rlds:
-    image_obs_keys:
-      - image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - null
-    fps: 10
-
-  plex_robosuite:
-    image_obs_keys:
-      - image
-      - wrist_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 20
-
-  conq_hose_manipulation:
-    image_obs_keys:
-      - frontleft_fisheye_image
-      - frontright_fisheye_image
-      - hand_color_image
-    depth_obs_keys:
-      - null
-    state_obs_keys:
-      - state
-    fps: 30

lerobot/common/datasets/push_dataset_to_hub/openx/data_utils.py

@@ -1,106 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the Licens    e.
-# 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.
-"""
-NOTE(YL): Adapted from:
-    Octo: https://github.com/octo-models/octo/blob/main/octo/data/utils/data_utils.py
-
-data_utils.py
-
-Additional utils for data processing.
-"""
-
-from typing import Any, Dict, List
-
-import tensorflow as tf
-
-
-def binarize_gripper_actions(actions: tf.Tensor) -> tf.Tensor:
-    """
-    Converts gripper actions from continuous to binary values (0 and 1).
-
-    We exploit that fact that most of the time, the gripper is fully open (near 1.0) or fully closed (near 0.0). As it
-    transitions between the two, it sometimes passes through a few intermediate values. We relabel those intermediate
-    values based on the state that is reached _after_ those intermediate values.
-
-    In the edge case that the trajectory ends with an intermediate value, we give up on binarizing and relabel that
-    chunk of intermediate values as the last action in the trajectory.
-
-    The `scan_fn` implements the following logic:
-        new_actions = np.empty_like(actions)
-        carry = actions[-1]
-        for i in reversed(range(actions.shape[0])):
-            if in_between_mask[i]:
-                carry = carry
-            else:
-                carry = float(open_mask[i])
-            new_actions[i] = carry
-    """
-    open_mask, closed_mask = actions > 0.95, actions < 0.05
-    in_between_mask = tf.logical_not(tf.logical_or(open_mask, closed_mask))
-    is_open_float = tf.cast(open_mask, tf.float32)
-
-    def scan_fn(carry, i):
-        return tf.cond(in_between_mask[i], lambda: tf.cast(carry, tf.float32), lambda: is_open_float[i])
-
-    return tf.scan(scan_fn, tf.range(tf.shape(actions)[0]), actions[-1], reverse=True)
-
-
-def invert_gripper_actions(actions: tf.Tensor) -> tf.Tensor:
-    return 1 - actions
-
-
-def rel2abs_gripper_actions(actions: tf.Tensor) -> tf.Tensor:
-    """
-    Converts relative gripper actions (+1 for closing, -1 for opening) to absolute actions (0 = closed; 1 = open).
-
-    Assumes that the first relative gripper is not redundant (i.e. close when already closed)!
-    """
-    # Note =>> -1 for closing, 1 for opening, 0 for no change
-    opening_mask, closing_mask = actions < -0.1, actions > 0.1
-    thresholded_actions = tf.where(opening_mask, 1, tf.where(closing_mask, -1, 0))
-
-    def scan_fn(carry, i):
-        return tf.cond(thresholded_actions[i] == 0, lambda: carry, lambda: thresholded_actions[i])
-
-    # If no relative grasp, assumes open for whole trajectory
-    start = -1 * thresholded_actions[tf.argmax(thresholded_actions != 0, axis=0)]
-    start = tf.cond(start == 0, lambda: 1, lambda: start)
-
-    # Note =>> -1 for closed, 1 for open
-    new_actions = tf.scan(scan_fn, tf.range(tf.shape(actions)[0]), start)
-    new_actions = tf.cast(new_actions, tf.float32) / 2 + 0.5
-
-    return new_actions
-
-
-# === Bridge-V2 =>> Dataset-Specific Transform ===
-def relabel_bridge_actions(traj: Dict[str, Any]) -> Dict[str, Any]:
-    """Relabels actions to use reached proprioceptive state; discards last timestep (no-action)."""
-    movement_actions = traj["observation"]["state"][1:, :6] - traj["observation"]["state"][:-1, :6]
-    traj_truncated = tf.nest.map_structure(lambda x: x[:-1], traj)
-    traj_truncated["action"] = tf.concat([movement_actions, traj["action"][:-1, -1:]], axis=1)
-
-    return traj_truncated
-
-
-# === RLDS Dataset Initialization Utilities ===
-def pprint_data_mixture(dataset_kwargs_list: List[Dict[str, Any]], dataset_weights: List[int]) -> None:
-    print("\n######################################################################################")
-    print(f"# Loading the following {len(dataset_kwargs_list)} datasets (incl. sampling weight):{'': >24} #")
-    for dataset_kwargs, weight in zip(dataset_kwargs_list, dataset_weights, strict=False):
-        pad = 80 - len(dataset_kwargs["name"])
-        print(f"# {dataset_kwargs['name']}: {weight:=>{pad}f} #")
-    print("######################################################################################\n")

lerobot/common/datasets/push_dataset_to_hub/openx/droid_utils.py

@@ -1,200 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""
-NOTE(YL): Adapted from:
-    OpenVLA: https://github.com/openvla/openvla
-
-Episode transforms for DROID dataset.
-"""
-
-from typing import Any, Dict
-
-import tensorflow as tf
-import tensorflow_graphics.geometry.transformation as tfg
-
-
-def rmat_to_euler(rot_mat):
-    return tfg.euler.from_rotation_matrix(rot_mat)
-
-
-def euler_to_rmat(euler):
-    return tfg.rotation_matrix_3d.from_euler(euler)
-
-
-def invert_rmat(rot_mat):
-    return tfg.rotation_matrix_3d.inverse(rot_mat)
-
-
-def rotmat_to_rot6d(mat):
-    """
-    Converts rotation matrix to R6 rotation representation (first two rows in rotation matrix).
-    Args:
-        mat: rotation matrix
-
-    Returns: 6d vector (first two rows of rotation matrix)
-
-    """
-    r6 = mat[..., :2, :]
-    r6_0, r6_1 = r6[..., 0, :], r6[..., 1, :]
-    r6_flat = tf.concat([r6_0, r6_1], axis=-1)
-    return r6_flat
-
-
-def velocity_act_to_wrist_frame(velocity, wrist_in_robot_frame):
-    """
-    Translates velocity actions (translation + rotation) from base frame of the robot to wrist frame.
-    Args:
-        velocity: 6d velocity action (3 x translation, 3 x rotation)
-        wrist_in_robot_frame: 6d pose of the end-effector in robot base frame
-
-    Returns: 9d velocity action in robot wrist frame (3 x translation, 6 x rotation as R6)
-
-    """
-    r_frame = euler_to_rmat(wrist_in_robot_frame[:, 3:6])
-    r_frame_inv = invert_rmat(r_frame)
-
-    # world to wrist: dT_pi = R^-1 dT_rbt
-    vel_t = (r_frame_inv @ velocity[:, :3][..., None])[..., 0]
-
-    # world to wrist: dR_pi = R^-1 dR_rbt R
-    dr_ = euler_to_rmat(velocity[:, 3:6])
-    dr_ = r_frame_inv @ (dr_ @ r_frame)
-    dr_r6 = rotmat_to_rot6d(dr_)
-    return tf.concat([vel_t, dr_r6], axis=-1)
-
-
-def rand_swap_exterior_images(img1, img2):
-    """
-    Randomly swaps the two exterior images (for training with single exterior input).
-    """
-    return tf.cond(tf.random.uniform(shape=[]) > 0.5, lambda: (img1, img2), lambda: (img2, img1))
-
-
-def droid_baseact_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    DROID dataset transformation for actions expressed in *base* frame of the robot.
-    """
-    dt = trajectory["action_dict"]["cartesian_velocity"][:, :3]
-    dr_ = trajectory["action_dict"]["cartesian_velocity"][:, 3:6]
-
-    trajectory["action"] = tf.concat(
-        (
-            dt,
-            dr_,
-            1 - trajectory["action_dict"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["exterior_image_1_left"], trajectory["observation"]["exterior_image_2_left"] = (
-        rand_swap_exterior_images(
-            trajectory["observation"]["exterior_image_1_left"],
-            trajectory["observation"]["exterior_image_2_left"],
-        )
-    )
-    trajectory["observation"]["proprio"] = tf.concat(
-        (
-            trajectory["observation"]["cartesian_position"],
-            trajectory["observation"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def droid_wristact_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    DROID dataset transformation for actions expressed in *wrist* frame of the robot.
-    """
-    wrist_act = velocity_act_to_wrist_frame(
-        trajectory["action_dict"]["cartesian_velocity"], trajectory["observation"]["cartesian_position"]
-    )
-    trajectory["action"] = tf.concat(
-        (
-            wrist_act,
-            trajectory["action_dict"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["exterior_image_1_left"], trajectory["observation"]["exterior_image_2_left"] = (
-        rand_swap_exterior_images(
-            trajectory["observation"]["exterior_image_1_left"],
-            trajectory["observation"]["exterior_image_2_left"],
-        )
-    )
-    trajectory["observation"]["proprio"] = tf.concat(
-        (
-            trajectory["observation"]["cartesian_position"],
-            trajectory["observation"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def droid_finetuning_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    DROID dataset transformation for actions expressed in *base* frame of the robot.
-    """
-    dt = trajectory["action_dict"]["cartesian_velocity"][:, :3]
-    dr_ = trajectory["action_dict"]["cartesian_velocity"][:, 3:6]
-    trajectory["action"] = tf.concat(
-        (
-            dt,
-            dr_,
-            1 - trajectory["action_dict"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["proprio"] = tf.concat(
-        (
-            trajectory["observation"]["cartesian_position"],
-            trajectory["observation"]["gripper_position"],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def zero_action_filter(traj: Dict) -> bool:
-    """
-    Filters transitions whose actions are all-0 (only relative actions, no gripper action).
-    Note: this filter is applied *after* action normalization, so need to compare to "normalized 0".
-    """
-    droid_q01 = tf.convert_to_tensor(
-        [
-            -0.7776297926902771,
-            -0.5803514122962952,
-            -0.5795090794563293,
-            -0.6464047729969025,
-            -0.7041108310222626,
-            -0.8895104378461838,
-        ]
-    )
-    droid_q99 = tf.convert_to_tensor(
-        [
-            0.7597932070493698,
-            0.5726242214441299,
-            0.7351000607013702,
-            0.6705610305070877,
-            0.6464948207139969,
-            0.8897542208433151,
-        ]
-    )
-    droid_norm_0_act = (
-        2 * (tf.zeros_like(traj["action"][:, :6]) - droid_q01) / (droid_q99 - droid_q01 + 1e-8) - 1
-    )
-
-    return tf.reduce_any(tf.math.abs(traj["action"][:, :6] - droid_norm_0_act) > 1e-5)

lerobot/common/datasets/push_dataset_to_hub/openx/transforms.py

@@ -1,859 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""
-NOTE(YL): Adapted from:
-    OpenVLA: https://github.com/openvla/openvla
-    Octo: https://github.com/octo-models/octo
-
-transforms.py
-
-Defines a registry of per-dataset standardization transforms for each dataset in Open-X Embodiment.
-
-Transforms adopt the following structure:
-    Input: Dictionary of *batched* features (i.e., has leading time dimension)
-    Output: Dictionary `step` =>> {
-        "observation": {
-            <image_keys, depth_image_keys>
-            State (in chosen state representation)
-        },
-        "action": Action (in chosen action representation),
-        "language_instruction": str
-    }
-"""
-
-from typing import Any, Dict
-
-import tensorflow as tf
-
-from lerobot.common.datasets.push_dataset_to_hub.openx.data_utils import (
-    binarize_gripper_actions,
-    invert_gripper_actions,
-    rel2abs_gripper_actions,
-    relabel_bridge_actions,
-)
-
-
-def droid_baseact_transform_fn():
-    from lerobot.common.datasets.push_dataset_to_hub.openx.droid_utils import droid_baseact_transform
-
-    return droid_baseact_transform
-
-
-def bridge_openx_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    Applies to version of Bridge V2 in Open X-Embodiment mixture.
-
-    Note =>> In original Bridge V2 dataset, the first timestep has an all-zero action, so we remove it!
-    """
-    for key in trajectory:
-        if key == "traj_metadata":
-            continue
-        elif key in ["observation", "action"]:
-            for key2 in trajectory[key]:
-                trajectory[key][key2] = trajectory[key][key2][1:]
-        else:
-            trajectory[key] = trajectory[key][1:]
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            tf.cast(trajectory["action"]["open_gripper"][:, None], tf.float32),
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    trajectory = relabel_bridge_actions(trajectory)
-    trajectory["observation"]["EEF_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def bridge_orig_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    Applies to original version of Bridge V2 from the official project website.
-
-    Note =>> In original Bridge V2 dataset, the first timestep has an all-zero action, so we remove it!
-    """
-    for key in trajectory:
-        if key == "traj_metadata":
-            continue
-        elif key == "observation":
-            for key2 in trajectory[key]:
-                trajectory[key][key2] = trajectory[key][key2][1:]
-        else:
-            trajectory[key] = trajectory[key][1:]
-
-    trajectory["action"] = tf.concat(
-        [
-            trajectory["action"][:, :6],
-            binarize_gripper_actions(trajectory["action"][:, -1])[:, None],
-        ],
-        axis=1,
-    )
-    trajectory = relabel_bridge_actions(trajectory)
-    trajectory["observation"]["EEF_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def ppgm_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        [
-            trajectory["action"][:, :6],
-            binarize_gripper_actions(trajectory["action"][:, -1])[:, None],
-        ],
-        axis=1,
-    )
-    trajectory["observation"]["EEF_state"] = trajectory["observation"]["cartesian_position"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["gripper_position"][:, -1:]
-    return trajectory
-
-
-def rt1_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, 0]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def kuka_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, 0]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    # decode compressed state
-    eef_value = tf.io.decode_compressed(
-        trajectory["observation"]["clip_function_input/base_pose_tool_reached"],
-        compression_type="ZLIB",
-    )
-    eef_value = tf.io.decode_raw(eef_value, tf.float32)
-    trajectory["observation"]["clip_function_input/base_pose_tool_reached"] = tf.reshape(eef_value, (-1, 7))
-    gripper_value = tf.io.decode_compressed(
-        trajectory["observation"]["gripper_closed"], compression_type="ZLIB"
-    )
-    gripper_value = tf.io.decode_raw(gripper_value, tf.float32)
-    trajectory["observation"]["gripper_closed"] = tf.reshape(gripper_value, (-1, 1))
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def taco_play_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state_eef"] = trajectory["observation"]["robot_obs"][:, :6]
-    trajectory["observation"]["state_gripper"] = trajectory["observation"]["robot_obs"][:, 7:8]
-    trajectory["action"] = trajectory["action"]["rel_actions_world"]
-
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            tf.clip_by_value(trajectory["action"][:, -1:], 0, 1),
-        ),
-        axis=-1,
-    )
-
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def jaco_play_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state_eef"] = trajectory["observation"]["end_effector_cartesian_pos"][:, :6]
-    trajectory["observation"]["state_gripper"] = trajectory["observation"]["end_effector_cartesian_pos"][
-        :, -1:
-    ]
-
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, 0]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            tf.zeros_like(trajectory["action"]["world_vector"]),
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def berkeley_cable_routing_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            tf.zeros_like(trajectory["action"]["world_vector"][:, :1]),
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def roboturk_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert absolute gripper action, +1 = open, 0 = close
-    gripper_action = invert_gripper_actions(
-        tf.clip_by_value(trajectory["action"]["gripper_closedness_action"], 0, 1)
-    )
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action,
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    trajectory["language_embedding"] = trajectory["observation"]["natural_language_embedding"]
-    return trajectory
-
-
-def nyu_door_opening_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, 0]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def viola_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # make gripper action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"][:, None]
-    gripper_action = tf.clip_by_value(gripper_action, 0, 1)
-    gripper_action = invert_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action,
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def berkeley_autolab_ur5_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = trajectory["observation"]["robot_state"][:, 6:14]
-
-    # make gripper action absolute action, +1 = open, 0 = close
-    gripper_action = trajectory["action"]["gripper_closedness_action"]
-    gripper_action = rel2abs_gripper_actions(gripper_action)
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            gripper_action[:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def toto_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            tf.cast(trajectory["action"]["open_gripper"][:, None], tf.float32),
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def language_table_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # default to "open" gripper
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"],
-            tf.zeros_like(trajectory["action"]),
-            tf.zeros_like(trajectory["action"]),
-            tf.ones_like(trajectory["action"][:, :1]),
-        ),
-        axis=-1,
-    )
-
-    # decode language instruction
-    instruction_bytes = trajectory["observation"]["instruction"]
-    instruction_encoded = tf.strings.unicode_encode(instruction_bytes, output_encoding="UTF-8")
-    # Remove trailing padding --> convert RaggedTensor to regular Tensor.
-    trajectory["language_instruction"] = tf.strings.split(instruction_encoded, "\x00")[:, :1].to_tensor()[
-        :, 0
-    ]
-    return trajectory
-
-
-def pusht_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["world_vector"],
-            trajectory["action"]["rotation_delta"],
-            trajectory["action"]["gripper_closedness_action"][:, None],
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def stanford_kuka_multimodal_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["depth_image"] = trajectory["observation"]["depth_image"][..., 0]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tf.zeros_like(trajectory["action"][:, :3]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def nyu_rot_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][..., :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][..., -1:]
-    trajectory["action"] = trajectory["action"][..., :7]
-    return trajectory
-
-
-def stanford_hydra_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(trajectory["action"][:, -1:]),
-        ),
-        axis=-1,
-    )
-
-    trajectory["observation"]["eef_state"] = tf.concat(
-        (
-            trajectory["observation"]["state"][:, :3],
-            trajectory["observation"]["state"][:, 7:10],
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -3:-2]
-    return trajectory
-
-
-def austin_buds_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-
-    trajectory["observation"]["state"] = trajectory["observation"]["state"][:, :8]
-    return trajectory
-
-
-def nyu_franka_play_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["depth"] = tf.cast(trajectory["observation"]["depth"][..., 0], tf.float32)
-    trajectory["observation"]["depth_additional_view"] = tf.cast(
-        trajectory["observation"]["depth_additional_view"][..., 0], tf.float32
-    )
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, -6:]
-
-    # clip gripper action, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, -8:-2],
-            tf.clip_by_value(trajectory["action"][:, -2:-1], 0, 1),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def maniskill_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][..., 7:8]
-    return trajectory
-
-
-def furniture_bench_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    import tensorflow_graphics.geometry.transformation as tft
-
-    trajectory["observation"]["state"] = tf.concat(
-        (
-            trajectory["observation"]["state"][:, :7],
-            trajectory["observation"]["state"][:, -1:],
-        ),
-        axis=-1,
-    )
-
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tft.euler.from_quaternion(trajectory["action"][:, 3:7]),
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def cmu_franka_exploration_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def ucsd_kitchen_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["joint_state"] = trajectory["observation"]["state"][:, :7]
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def ucsd_pick_place_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tf.zeros_like(trajectory["action"][:, :3]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def austin_sailor_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def austin_sirius_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def bc_z_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"]["future/xyz_residual"][:, :3],
-            trajectory["action"]["future/axis_angle_residual"][:, :3],
-            invert_gripper_actions(tf.cast(trajectory["action"]["future/target_close"][:, :1], tf.float32)),
-        ),
-        axis=-1,
-    )
-    trajectory["language_instruction"] = trajectory["observation"]["natural_language_instruction"]
-    return trajectory
-
-
-def tokyo_pr2_opening_fridge_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def tokyo_pr2_tabletop_manipulation_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def utokyo_xarm_bimanual_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = trajectory["action"][..., -7:]
-    return trajectory
-
-
-def robo_net_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = tf.concat(
-        (
-            trajectory["observation"]["state"][:, :4],
-            tf.zeros_like(trajectory["observation"]["state"][:, :2]),
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :4],
-            tf.zeros_like(trajectory["action"][:, :2]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def berkeley_mvp_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    """
-    trajectory["observation"]["state"] = tf.concat((
-        tf.cast(trajectory["observation"]["gripper"][:, None], tf.float32),
-                        trajectory["observation"]["pose"],
-                        trajectory["observation"]["joint_pos"],),
-                        axis=-1,)
-    """
-    trajectory["observation"]["gripper"] = tf.cast(trajectory["observation"]["gripper"][:, None], tf.float32)
-    return trajectory
-
-
-def berkeley_rpt_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["gripper"] = tf.cast(trajectory["observation"]["gripper"][:, None], tf.float32)
-    return trajectory
-
-
-def kaist_nonprehensible_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = trajectory["observation"]["state"][:, -7:]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            tf.zeros_like(trajectory["action"][:, :1]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def stanford_mask_vit_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = tf.concat(
-        (
-            trajectory["observation"]["end_effector_pose"][:, :4],
-            tf.zeros_like(trajectory["observation"]["end_effector_pose"][:, :2]),
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["end_effector_pose"][:, -1:]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :4],
-            tf.zeros_like(trajectory["action"][:, :2]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def tokyo_lsmo_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def dlr_sara_grid_clamp_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = trajectory["observation"]["state"][:, :6]
-    return trajectory
-
-
-def dlr_edan_shared_control_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # invert gripper action, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(trajectory["action"][:, -1:]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def asu_table_top_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["ground_truth_states"]["EE"]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def robocook_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    return trajectory
-
-
-def imperial_wristcam_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def iamlab_pick_insert_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    import tensorflow_graphics.geometry.transformation as tft
-
-    trajectory["observation"]["joint_state"] = trajectory["observation"]["state"][:, :7]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, 7:8]
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tft.euler.from_quaternion(trajectory["action"][:, 3:7]),
-            trajectory["action"][:, 7:8],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def uiuc_d3field_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"],
-            tf.zeros_like(trajectory["action"]),
-            tf.zeros_like(trajectory["action"][:, :1]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def utaustin_mutex_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = trajectory["observation"]["state"][:, :8]
-
-    # invert gripper action + clip, +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :6],
-            invert_gripper_actions(tf.clip_by_value(trajectory["action"][:, -1:], 0, 1)),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def berkeley_fanuc_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["joint_state"] = trajectory["observation"]["state"][:, :6]
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, 6:7]
-
-    # dataset does not store gripper actions, so use gripper state info, invert so +1 = open, 0 = close
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"],
-            invert_gripper_actions(trajectory["observation"]["gripper_state"]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def cmu_playing_with_food_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    import tensorflow_graphics.geometry.transformation as tft
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            tft.euler.from_quaternion(trajectory["action"][:, 3:7]),
-            trajectory["action"][:, -1:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def playfusion_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :3],
-            trajectory["action"][:, -4:],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def cmu_stretch_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["eef_state"] = tf.concat(
-        (
-            trajectory["observation"]["state"][:, :3],
-            tf.zeros_like(trajectory["observation"]["state"][:, :3]),
-        ),
-        axis=-1,
-    )
-    trajectory["observation"]["gripper_state"] = trajectory["observation"]["state"][:, -1:]
-    trajectory["action"] = trajectory["action"][..., :-1]
-    return trajectory
-
-
-def gnm_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    trajectory["observation"]["state"] = tf.concat(
-        (
-            trajectory["observation"]["position"],
-            tf.zeros_like(trajectory["observation"]["state"][:, :3]),
-            trajectory["observation"]["yaw"],
-        ),
-        axis=-1,
-    )
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"],
-            tf.zeros_like(trajectory["action"]),
-            tf.zeros_like(trajectory["action"]),
-            tf.zeros_like(trajectory["action"][:, :1]),
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def fmb_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # every input feature is batched, ie has leading batch dimension
-    trajectory["observation"]["proprio"] = tf.concat(
-        (
-            trajectory["observation"]["eef_pose"],
-            trajectory["observation"]["state_gripper_pose"][..., None],
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def dobbe_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # every input feature is batched, ie has leading batch dimension
-    trajectory["observation"]["proprio"] = trajectory["observation"]["state"]
-    return trajectory
-
-
-def robo_set_dataset_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    # gripper action is in -1...1 --> clip to 0...1, flip
-    gripper_action = trajectory["action"][:, -1:]
-    gripper_action = invert_gripper_actions(tf.clip_by_value(gripper_action, 0, 1))
-
-    trajectory["action"] = tf.concat(
-        (
-            trajectory["action"][:, :7],
-            gripper_action,
-        ),
-        axis=-1,
-    )
-    return trajectory
-
-
-def identity_transform(trajectory: Dict[str, Any]) -> Dict[str, Any]:
-    return trajectory
-
-
-# === Registry ===
-OPENX_STANDARDIZATION_TRANSFORMS = {
-    "bridge_openx": bridge_openx_dataset_transform,
-    "bridge_orig": bridge_orig_dataset_transform,
-    "bridge_dataset": bridge_orig_dataset_transform,
-    "ppgm": ppgm_dataset_transform,
-    "ppgm_static": ppgm_dataset_transform,
-    "ppgm_wrist": ppgm_dataset_transform,
-    "fractal20220817_data": rt1_dataset_transform,
-    "kuka": kuka_dataset_transform,
-    "taco_play": taco_play_dataset_transform,
-    "jaco_play": jaco_play_dataset_transform,
-    "berkeley_cable_routing": berkeley_cable_routing_dataset_transform,
-    "roboturk": roboturk_dataset_transform,
-    "nyu_door_opening_surprising_effectiveness": nyu_door_opening_dataset_transform,
-    "viola": viola_dataset_transform,
-    "berkeley_autolab_ur5": berkeley_autolab_ur5_dataset_transform,
-    "toto": toto_dataset_transform,
-    "language_table": language_table_dataset_transform,
-    "columbia_cairlab_pusht_real": pusht_dataset_transform,
-    "stanford_kuka_multimodal_dataset_converted_externally_to_rlds": stanford_kuka_multimodal_dataset_transform,
-    "nyu_rot_dataset_converted_externally_to_rlds": nyu_rot_dataset_transform,
-    "stanford_hydra_dataset_converted_externally_to_rlds": stanford_hydra_dataset_transform,
-    "austin_buds_dataset_converted_externally_to_rlds": austin_buds_dataset_transform,
-    "nyu_franka_play_dataset_converted_externally_to_rlds": nyu_franka_play_dataset_transform,
-    "maniskill_dataset_converted_externally_to_rlds": maniskill_dataset_transform,
-    "furniture_bench_dataset_converted_externally_to_rlds": furniture_bench_dataset_transform,
-    "cmu_franka_exploration_dataset_converted_externally_to_rlds": cmu_franka_exploration_dataset_transform,
-    "ucsd_kitchen_dataset_converted_externally_to_rlds": ucsd_kitchen_dataset_transform,
-    "ucsd_pick_and_place_dataset_converted_externally_to_rlds": ucsd_pick_place_dataset_transform,
-    "austin_sailor_dataset_converted_externally_to_rlds": austin_sailor_dataset_transform,
-    "austin_sirius_dataset_converted_externally_to_rlds": austin_sirius_dataset_transform,
-    "bc_z": bc_z_dataset_transform,
-    "utokyo_pr2_opening_fridge_converted_externally_to_rlds": tokyo_pr2_opening_fridge_dataset_transform,
-    "utokyo_pr2_tabletop_manipulation_converted_externally_to_rlds": tokyo_pr2_tabletop_manipulation_dataset_transform,
-    "utokyo_xarm_pick_and_place_converted_externally_to_rlds": identity_transform,
-    "utokyo_xarm_bimanual_converted_externally_to_rlds": utokyo_xarm_bimanual_dataset_transform,
-    "robo_net": robo_net_dataset_transform,
-    "berkeley_mvp_converted_externally_to_rlds": berkeley_mvp_dataset_transform,
-    "berkeley_rpt_converted_externally_to_rlds": berkeley_rpt_dataset_transform,
-    "kaist_nonprehensile_converted_externally_to_rlds": kaist_nonprehensible_dataset_transform,
-    "stanford_mask_vit_converted_externally_to_rlds": stanford_mask_vit_dataset_transform,
-    "tokyo_u_lsmo_converted_externally_to_rlds": tokyo_lsmo_dataset_transform,
-    "dlr_sara_pour_converted_externally_to_rlds": identity_transform,
-    "dlr_sara_grid_clamp_converted_externally_to_rlds": dlr_sara_grid_clamp_dataset_transform,
-    "dlr_edan_shared_control_converted_externally_to_rlds": dlr_edan_shared_control_dataset_transform,
-    "asu_table_top_converted_externally_to_rlds": asu_table_top_dataset_transform,
-    "stanford_robocook_converted_externally_to_rlds": robocook_dataset_transform,
-    "imperialcollege_sawyer_wrist_cam": imperial_wristcam_dataset_transform,
-    "iamlab_cmu_pickup_insert_converted_externally_to_rlds": iamlab_pick_insert_dataset_transform,
-    "uiuc_d3field": uiuc_d3field_dataset_transform,
-    "utaustin_mutex": utaustin_mutex_dataset_transform,
-    "berkeley_fanuc_manipulation": berkeley_fanuc_dataset_transform,
-    "cmu_playing_with_food": cmu_playing_with_food_dataset_transform,
-    "cmu_play_fusion": playfusion_dataset_transform,
-    "cmu_stretch": cmu_stretch_dataset_transform,
-    "berkeley_gnm_recon": gnm_dataset_transform,
-    "berkeley_gnm_cory_hall": gnm_dataset_transform,
-    "berkeley_gnm_sac_son": gnm_dataset_transform,
-    "droid": droid_baseact_transform_fn(),
-    "droid_100": droid_baseact_transform_fn(),  # first 100 episodes of droid
-    "fmb": fmb_transform,
-    "dobbe": dobbe_dataset_transform,
-    "robo_set": robo_set_dataset_transform,
-    "usc_cloth_sim_converted_externally_to_rlds": identity_transform,
-    "plex_robosuite": identity_transform,
-    "conq_hose_manipulation": identity_transform,
-    "io_ai_tech": identity_transform,
-    "spoc": identity_transform,
-}

lerobot/common/datasets/push_dataset_to_hub/openx_rlds_format.py

@@ -14,13 +14,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
+For all datasets in the RLDS format.
 For https://github.com/google-deepmind/open_x_embodiment (OPENX) datasets.
 
+NOTE: You need to install tensorflow and tensorflow_datasets before running this script.
+
 Example:
     python lerobot/scripts/push_dataset_to_hub.py \
-        --raw-dir /hdd/tensorflow_datasets/bridge_dataset/1.0.0/ \
-        --repo-id youliangtan/sampled_bridge_data_v2 \
-        --raw-format openx_rlds.bridge_orig \
+        --raw-dir /path/to/data/bridge_dataset/1.0.0/ \
+        --repo-id your_hub/sampled_bridge_data_v2 \
+        --raw-format rlds \
         --episodes 3 4 5 8 9
 
 Exact dataset fps defined in openx/config.py, obtained from:
@@ -35,28 +38,21 @@ import tensorflow as tf
 import tensorflow_datasets as tfds
 import torch
 import tqdm
-import yaml
 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.openx.transforms import OPENX_STANDARDIZATION_TRANSFORMS
 from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
     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
 
-with open("lerobot/common/datasets/push_dataset_to_hub/openx/configs.yaml") as f:
-    _openx_list = yaml.safe_load(f)
-
-OPENX_DATASET_CONFIGS = _openx_list["OPENX_DATASET_CONFIGS"]
-
 np.set_printoptions(precision=2)
 
 
@@ -108,7 +104,6 @@ def load_from_raw(
     video: bool,
     episodes: list[int] | None = None,
     encoding: dict | None = None,
-    openx_dataset_name: str | None = None,
 ):
     """
     Args:
@@ -136,16 +131,17 @@ def load_from_raw(
     # we will apply the standardization transform if the dataset_name is provided
     # if the dataset name is not provided and the goal is to convert any rlds formatted dataset
     # search for 'image' keys in the observations
-    if openx_dataset_name is not None:
-        print(" - applying standardization transform for dataset: ", openx_dataset_name)
-        assert openx_dataset_name in OPENX_STANDARDIZATION_TRANSFORMS
-        transform_fn = OPENX_STANDARDIZATION_TRANSFORMS[openx_dataset_name]
-        dataset = dataset.map(transform_fn)
-
-        image_keys = OPENX_DATASET_CONFIGS[openx_dataset_name]["image_obs_keys"]
-    else:
-        obs_keys = dataset_info.features["steps"]["observation"].keys()
-        image_keys = [key for key in obs_keys if "image" in key]
+    image_keys = []
+    state_keys = []
+    observation_info = dataset_info.features["steps"]["observation"]
+    for key in observation_info:
+        # check whether the key is for an image or a vector observation
+        if len(observation_info[key].shape) == 3:
+            # only adding uint8 images discards depth images
+            if observation_info[key].dtype == tf.uint8:
+                image_keys.append(key)
+        else:
+            state_keys.append(key)
 
     lang_key = "language_instruction" if "language_instruction" in dataset.element_spec else None
 
@@ -193,50 +189,31 @@ def load_from_raw(
 
         num_frames = episode["action"].shape[0]
 
-        ###########################################################
-        # Handle the episodic data
-
-        # last step of demonstration is considered done
-        done = torch.zeros(num_frames, dtype=torch.bool)
-        done[-1] = True
         ep_dict = {}
-        langs = []  # TODO: might be located in "observation"
+        for key in state_keys:
+            ep_dict[f"observation.{key}"] = tf_to_torch(episode["observation"][key])
 
-        image_array_dict = {key: [] for key in image_keys}
-
-        # We will create the state observation tensor by stacking the state
-        # obs keys defined in the openx/configs.py
-        if openx_dataset_name is not None:
-            state_obs_keys = OPENX_DATASET_CONFIGS[openx_dataset_name]["state_obs_keys"]
-            # stack the state observations, if is None, pad with zeros
-            states = []
-            for key in state_obs_keys:
-                if key in episode["observation"]:
-                    states.append(tf_to_torch(episode["observation"][key]))
-                else:
-                    states.append(torch.zeros(num_frames, 1))  # pad with zeros
-            states = torch.cat(states, dim=1)
-            # assert states.shape == (num_frames, 8), f"states shape: {states.shape}"
-        else:
-            states = tf_to_torch(episode["observation"]["state"])
-
-        actions = tf_to_torch(episode["action"])
-        rewards = tf_to_torch(episode["reward"]).float()
+        ep_dict["action"] = tf_to_torch(episode["action"])
+        ep_dict["next.reward"] = tf_to_torch(episode["reward"]).float()
+        ep_dict["next.done"] = tf_to_torch(episode["is_last"])
+        ep_dict["is_terminal"] = tf_to_torch(episode["is_terminal"])
+        ep_dict["is_first"] = tf_to_torch(episode["is_first"])
+        ep_dict["discount"] = tf_to_torch(episode["discount"])
 
         # If lang_key is present, convert the entire tensor at once
         if lang_key is not None:
-            langs = [str(x) for x in episode[lang_key]]
+            ep_dict["language_instruction"] = [x.numpy().decode("utf-8") for x in episode[lang_key]]
+
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+
+        image_array_dict = {key: [] for key in image_keys}
 
         for im_key in image_keys:
             imgs = episode["observation"][im_key]
             image_array_dict[im_key] = [tf_img_convert(img) for img in imgs]
 
-        # simple assertions
-        for item in [states, actions, rewards, done]:
-            assert len(item) == num_frames
-
-        ###########################################################
-
         # loop through all cameras
         for im_key in image_keys:
             img_key = f"observation.images.{im_key}"
@@ -262,17 +239,6 @@ def load_from_raw(
             else:
                 ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
 
-        if lang_key is not None:
-            ep_dict["language_instruction"] = langs
-
-        ep_dict["observation.state"] = states
-        ep_dict["action"] = actions
-        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
-        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
-        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
-        ep_dict["next.reward"] = rewards
-        ep_dict["next.done"] = done
-
         path_ep_dict = tmp_ep_dicts_dir.joinpath(
             "ep_dict_" + "0" * (10 - len(str(ep_idx))) + str(ep_idx) + ".pt"
         )
@@ -290,30 +256,28 @@ def load_from_raw(
 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()
+    for key in data_dict:
+        # check if vector state obs
+        if key.startswith("observation.") and "observation.images." not in key:
+            features[key] = Sequence(length=data_dict[key].shape[1], feature=Value(dtype="float32", id=None))
+        # check if image obs
+        elif "observation.images." in key:
+            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)
-        )
     if "language_instruction" in data_dict:
         features["language_instruction"] = Value(dtype="string", id=None)
 
     features["action"] = Sequence(
         length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
     )
+
+    features["is_terminal"] = Value(dtype="bool", id=None)
+    features["is_first"] = Value(dtype="bool", id=None)
+    features["discount"] = 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)
@@ -333,19 +297,8 @@ def from_raw_to_lerobot_format(
     video: bool = True,
     episodes: list[int] | None = None,
     encoding: dict | None = None,
-    openx_dataset_name: str | None = None,
 ):
-    """This is a test impl for rlds conversion"""
-    if openx_dataset_name is None:
-        # set a default rlds frame rate if the dataset is not from openx
-        fps = 30
-    elif "fps" not in OPENX_DATASET_CONFIGS[openx_dataset_name]:
-        raise ValueError(
-            "fps for this dataset is not specified in openx/configs.py yet," "means it is not yet tested"
-        )
-    fps = OPENX_DATASET_CONFIGS[openx_dataset_name]["fps"]
-
-    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding, openx_dataset_name)
+    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 = {

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -27,12 +27,12 @@ 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 (
+    calculate_episode_data_index,
     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

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -28,12 +28,12 @@ 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 (
+    calculate_episode_data_index,
     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

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -16,7 +16,9 @@
 import inspect
 from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
+from typing import Dict
 
+import datasets
 import numpy
 import PIL
 import torch
@@ -72,3 +74,58 @@ def check_repo_id(repo_id: str) -> None:
             f"""`repo_id` is expected to contain a community or user id `/` the name of the dataset
             (e.g. 'lerobot/pusht'), but contains '{repo_id}'."""
         )
+
+
+# TODO(aliberts): remove
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
+    """
+    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
+
+    Parameters:
+    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
+
+    Returns:
+    - episode_data_index: A dictionary containing the data index for each episode. The dictionary has two keys:
+        - "from": A tensor containing the starting index of each episode.
+        - "to": A tensor containing the ending index of each episode.
+    """
+    episode_data_index = {"from": [], "to": []}
+
+    current_episode = None
+    """
+    The episode_index is a list of integers, each representing the episode index of the corresponding example.
+    For instance, the following is a valid episode_index:
+      [0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2]
+
+    Below, we iterate through the episode_index and populate the episode_data_index dictionary with the starting and
+    ending index of each episode. For the episode_index above, the episode_data_index dictionary will look like this:
+        {
+            "from": [0, 3, 7],
+            "to": [3, 7, 12]
+        }
+    """
+    if len(hf_dataset) == 0:
+        episode_data_index = {
+            "from": torch.tensor([]),
+            "to": torch.tensor([]),
+        }
+        return episode_data_index
+    for idx, episode_idx in enumerate(hf_dataset["episode_index"]):
+        if episode_idx != current_episode:
+            # We encountered a new episode, so we append its starting location to the "from" list
+            episode_data_index["from"].append(idx)
+            # If this is not the first episode, we append the ending location of the previous episode to the "to" list
+            if current_episode is not None:
+                episode_data_index["to"].append(idx)
+            # Let's keep track of the current episode index
+            current_episode = episode_idx
+        else:
+            # We are still in the same episode, so there is nothing for us to do here
+            pass
+    # We have reached the end of the dataset, so we append the ending location of the last episode to the "to" list
+    episode_data_index["to"].append(idx + 1)
+
+    for k in ["from", "to"]:
+        episode_data_index[k] = torch.tensor(episode_data_index[k])
+
+    return episode_data_index

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py

@@ -27,12 +27,12 @@ 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 (
+    calculate_episode_data_index,
     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

lerobot/common/datasets/transforms.py

@@ -14,7 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import collections
-from typing import Any, Callable, Dict, Sequence
+from dataclasses import dataclass, field
+from typing import Any, Callable, Sequence
 
 import torch
 from torchvision.transforms import v2
@@ -65,6 +66,8 @@ class RandomSubsetApply(Transform):
         self.n_subset = n_subset
         self.random_order = random_order
 
+        self.selected_transforms = None
+
     def forward(self, *inputs: Any) -> Any:
         needs_unpacking = len(inputs) > 1
 
@@ -72,9 +75,9 @@ class RandomSubsetApply(Transform):
         if not self.random_order:
             selected_indices = selected_indices.sort().values
 
-        selected_transforms = [self.transforms[i] for i in selected_indices]
+        self.selected_transforms = [self.transforms[i] for i in selected_indices]
 
-        for transform in selected_transforms:
+        for transform in self.selected_transforms:
             outputs = transform(*inputs)
             inputs = outputs if needs_unpacking else (outputs,)
 
@@ -129,69 +132,118 @@ class SharpnessJitter(Transform):
 
         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 make_params(self, flat_inputs: list[Any]) -> dict[str, Any]:
+        sharpness_factor = torch.empty(1).uniform_(self.sharpness[0], self.sharpness[1]).item()
+        return {"sharpness_factor": sharpness_factor}
 
-    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
-        sharpness_factor = self._generate_value(self.sharpness[0], self.sharpness[1])
+    def transform(self, inpt: Any, params: dict[str, Any]) -> Any:
+        sharpness_factor = params["sharpness_factor"]
         return self._call_kernel(F.adjust_sharpness, inpt, sharpness_factor=sharpness_factor)
 
 
-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})."
-                )
+@dataclass
+class ImageTransformConfig:
+    """
+    For each transform, the following parameters are available:
+      weight: This represents the multinomial probability (with no replacement)
+            used for sampling the transform. If the sum of the weights is not 1,
+            they will be normalized.
+      type: The name of the class used. This is either a class available under torchvision.transforms.v2 or a
+            custom transform defined here.
+      kwargs: Lower & upper bound respectively used for sampling the transform's parameter
+            (following uniform distribution) when it's applied.
+    """
 
-    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)
+    weight: float = 1.0
+    type: str = "Identity"
+    kwargs: dict[str, Any] = field(default_factory=dict)
 
-    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)
+@dataclass
+class ImageTransformsConfig:
+    """
+    These transforms are all using standard torchvision.transforms.v2
+    You can find out how these transformations affect images here:
+    https://pytorch.org/vision/0.18/auto_examples/transforms/plot_transforms_illustrations.html
+    We use a custom RandomSubsetApply container to sample them.
+    """
 
-    if n_subset == 0:
-        return v2.Identity()
+    # Set this flag to `true` to enable transforms during training
+    enable: bool = False
+    # This is the maximum number of transforms (sampled from these below) that will be applied to each frame.
+    # It's an integer in the interval [1, number_of_available_transforms].
+    max_num_transforms: int = 3
+    # By default, transforms are applied in Torchvision's suggested order (shown below).
+    # Set this to True to apply them in a random order.
+    random_order: bool = False
+    tfs: dict[str, ImageTransformConfig] = field(
+        default_factory=lambda: {
+            "brightness": ImageTransformConfig(
+                weight=1.0,
+                type="ColorJitter",
+                kwargs={"brightness": (0.8, 1.2)},
+            ),
+            "contrast": ImageTransformConfig(
+                weight=1.0,
+                type="ColorJitter",
+                kwargs={"contrast": (0.8, 1.2)},
+            ),
+            "saturation": ImageTransformConfig(
+                weight=1.0,
+                type="ColorJitter",
+                kwargs={"saturation": (0.5, 1.5)},
+            ),
+            "hue": ImageTransformConfig(
+                weight=1.0,
+                type="ColorJitter",
+                kwargs={"hue": (-0.05, 0.05)},
+            ),
+            "sharpness": ImageTransformConfig(
+                weight=1.0,
+                type="SharpnessJitter",
+                kwargs={"sharpness": (0.5, 1.5)},
+            ),
+        }
+    )
+
+
+def make_transform_from_config(cfg: ImageTransformConfig):
+    if cfg.type == "Identity":
+        return v2.Identity(**cfg.kwargs)
+    elif cfg.type == "ColorJitter":
+        return v2.ColorJitter(**cfg.kwargs)
+    elif cfg.type == "SharpnessJitter":
+        return SharpnessJitter(**cfg.kwargs)
     else:
-        # TODO(rcadene, aliberts): add v2.ToDtype float16?
-        return RandomSubsetApply(transforms, p=weights, n_subset=n_subset, random_order=random_order)
+        raise ValueError(f"Transform '{cfg.type}' is not valid.")
+
+
+class ImageTransforms(Transform):
+    """A class to compose image transforms based on configuration."""
+
+    def __init__(self, cfg: ImageTransformsConfig) -> None:
+        super().__init__()
+        self._cfg = cfg
+
+        self.weights = []
+        self.transforms = {}
+        for tf_name, tf_cfg in cfg.tfs.items():
+            if tf_cfg.weight <= 0.0:
+                continue
+
+            self.transforms[tf_name] = make_transform_from_config(tf_cfg)
+            self.weights.append(tf_cfg.weight)
+
+        n_subset = min(len(self.transforms), cfg.max_num_transforms)
+        if n_subset == 0 or not cfg.enable:
+            self.tf = v2.Identity()
+        else:
+            self.tf = RandomSubsetApply(
+                transforms=list(self.transforms.values()),
+                p=self.weights,
+                n_subset=n_subset,
+                random_order=cfg.random_order,
+            )
+
+    def forward(self, *inputs: Any) -> Any:
+        return self.tf(*inputs)

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

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

lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py

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

lerobot/common/datasets/v21/_remove_language_instruction.py

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

lerobot/common/datasets/v21/batch_convert_dataset_v20_to_v21.py

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

lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py

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

lerobot/common/datasets/v21/convert_stats.py

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

lerobot/common/datasets/video_utils.py

@@ -13,6 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import json
 import logging
 import subprocess
 import warnings
@@ -25,47 +26,11 @@ 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
+from PIL import Image
 
 
 def decode_video_frames_torchvision(
-    video_path: str,
+    video_path: Path | str,
     timestamps: list[float],
     tolerance_s: float,
     backend: str = "pyav",
@@ -104,11 +69,11 @@ def decode_video_frames_torchvision(
 
     # 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]
+    first_ts = min(timestamps)
+    last_ts = max(timestamps)
 
     # 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)
+    # Note: closest key frame timestamp is usually smaller than `first_ts` (e.g. key frame can be the first frame of the video)
     # for details on what `seek` is doing see: https://pyav.basswood-io.com/docs/stable/api/container.html?highlight=inputcontainer#av.container.InputContainer.seek
     reader.seek(first_ts, keyframes_only=keyframes_only)
 
@@ -163,8 +128,8 @@ def decode_video_frames_torchvision(
 
 
 def encode_video_frames(
-    imgs_dir: Path,
-    video_path: Path,
+    imgs_dir: Path | str,
+    video_path: Path | str,
     fps: int,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
@@ -247,3 +212,104 @@ with warnings.catch_warnings():
     )
     # to make VideoFrame available in HuggingFace `datasets`
     register_feature(VideoFrame, "VideoFrame")
+
+
+def get_audio_info(video_path: Path | str) -> dict:
+    ffprobe_audio_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "a:0",
+        "-show_entries",
+        "stream=channels,codec_name,bit_rate,sample_rate,bit_depth,channel_layout,duration",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_audio_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+
+    info = json.loads(result.stdout)
+    audio_stream_info = info["streams"][0] if info.get("streams") else None
+    if audio_stream_info is None:
+        return {"has_audio": False}
+
+    # Return the information, defaulting to None if no audio stream is present
+    return {
+        "has_audio": True,
+        "audio.channels": audio_stream_info.get("channels", None),
+        "audio.codec": audio_stream_info.get("codec_name", None),
+        "audio.bit_rate": int(audio_stream_info["bit_rate"]) if audio_stream_info.get("bit_rate") else None,
+        "audio.sample_rate": int(audio_stream_info["sample_rate"])
+        if audio_stream_info.get("sample_rate")
+        else None,
+        "audio.bit_depth": audio_stream_info.get("bit_depth", None),
+        "audio.channel_layout": audio_stream_info.get("channel_layout", None),
+    }
+
+
+def get_video_info(video_path: Path | str) -> dict:
+    ffprobe_video_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "v:0",
+        "-show_entries",
+        "stream=r_frame_rate,width,height,codec_name,nb_frames,duration,pix_fmt",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_video_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+
+    info = json.loads(result.stdout)
+    video_stream_info = info["streams"][0]
+
+    # Calculate fps from r_frame_rate
+    r_frame_rate = video_stream_info["r_frame_rate"]
+    num, denom = map(int, r_frame_rate.split("/"))
+    fps = num / denom
+
+    pixel_channels = get_video_pixel_channels(video_stream_info["pix_fmt"])
+
+    video_info = {
+        "video.fps": fps,
+        "video.height": video_stream_info["height"],
+        "video.width": video_stream_info["width"],
+        "video.channels": pixel_channels,
+        "video.codec": video_stream_info["codec_name"],
+        "video.pix_fmt": video_stream_info["pix_fmt"],
+        "video.is_depth_map": False,
+        **get_audio_info(video_path),
+    }
+
+    return video_info
+
+
+def get_video_pixel_channels(pix_fmt: str) -> int:
+    if "gray" in pix_fmt or "depth" in pix_fmt or "monochrome" in pix_fmt:
+        return 1
+    elif "rgba" in pix_fmt or "yuva" in pix_fmt:
+        return 4
+    elif "rgb" in pix_fmt or "yuv" in pix_fmt:
+        return 3
+    else:
+        raise ValueError("Unknown format")
+
+
+def get_image_pixel_channels(image: Image):
+    if image.mode == "L":
+        return 1  # Grayscale
+    elif image.mode == "LA":
+        return 2  # Grayscale + Alpha
+    elif image.mode == "RGB":
+        return 3  # RGB
+    elif image.mode == "RGBA":
+        return 4  # RGBA
+    else:
+        raise ValueError("Unknown format")

lerobot/common/envs/__init__.py

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

lerobot/common/envs/configs.py

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

lerobot/common/envs/factory.py

@@ -16,43 +16,54 @@
 import importlib
 
 import gymnasium as gym
-from omegaconf import DictConfig
+
+from lerobot.common.envs.configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv
 
 
-def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv | None:
-    """Makes a gym vector environment according to the evaluation config.
+def make_env_config(env_type: str, **kwargs) -> EnvConfig:
+    if env_type == "aloha":
+        return AlohaEnv(**kwargs)
+    elif env_type == "pusht":
+        return PushtEnv(**kwargs)
+    elif env_type == "xarm":
+        return XarmEnv(**kwargs)
+    else:
+        raise ValueError(f"Policy type '{env_type}' is not available.")
 
-    n_envs can be used to override eval.batch_size in the configuration. Must be at least 1.
+
+def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> gym.vector.VectorEnv | None:
+    """Makes a gym vector environment according to the config.
+
+    Args:
+        cfg (EnvConfig): the config of the environment to instantiate.
+        n_envs (int, optional): The number of parallelized env to return. Defaults to 1.
+        use_async_envs (bool, optional): Whether to return an AsyncVectorEnv or a SyncVectorEnv. Defaults to
+            False.
+
+    Raises:
+        ValueError: if n_envs < 1
+        ModuleNotFoundError: If the requested env package is not installed
+
+    Returns:
+        gym.vector.VectorEnv: The parallelized gym.env instance.
     """
-    if n_envs is not None and n_envs < 1:
+    if 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}"
+    package_name = f"gym_{cfg.type}"
 
     try:
         importlib.import_module(package_name)
     except ModuleNotFoundError as e:
-        print(
-            f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.env.name}]'`"
-        )
+        print(f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.type}]'`")
         raise e
 
-    gym_handle = f"{package_name}/{cfg.env.task}"
-    gym_kwgs = dict(cfg.env.get("gym", {}))
-
-    if cfg.env.get("episode_length"):
-        gym_kwgs["max_episode_steps"] = cfg.env.episode_length
+    gym_handle = f"{package_name}/{cfg.task}"
 
     # batched version of the env that returns an observation of shape (b, c)
-    env_cls = gym.vector.AsyncVectorEnv if cfg.eval.use_async_envs else gym.vector.SyncVectorEnv
+    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
     env = env_cls(
-        [
-            lambda: gym.make(gym_handle, disable_env_checker=True, **gym_kwgs)
-            for _ in range(n_envs if n_envs is not None else cfg.eval.batch_size)
-        ]
+        [lambda: gym.make(gym_handle, disable_env_checker=True, **cfg.gym_kwargs) for _ in range(n_envs)]
     )
 
     return env

lerobot/common/envs/utils.py

@@ -18,8 +18,13 @@ import numpy as np
 import torch
 from torch import Tensor
 
+from lerobot.common.envs.configs import EnvConfig
+from lerobot.common.utils.utils import get_channel_first_image_shape
+from lerobot.configs.types import FeatureType, PolicyFeature
+
 
 def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
+    # TODO(aliberts, rcadene): refactor this to use features from the environment (no hardcoding)
     """Convert environment observation to LeRobot format observation.
     Args:
         observation: Dictionary of observation batches from a Gym vector environment.
@@ -35,6 +40,7 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
             imgs = {"observation.image": observations["pixels"]}
 
         for imgkey, img in imgs.items():
+            # TODO(aliberts, rcadene): use transforms.ToTensor()?
             img = torch.from_numpy(img)
 
             # sanity check that images are channel last
@@ -60,3 +66,23 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
     # requirement for "agent_pos"
     return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
     return return_observations
+
+
+def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
+    # TODO(aliberts, rcadene): remove this hardcoding of keys and just use the nested keys as is
+    # (need to also refactor preprocess_observation and externalize normalization from policies)
+    policy_features = {}
+    for key, ft in env_cfg.features.items():
+        if ft.type is FeatureType.VISUAL:
+            if len(ft.shape) != 3:
+                raise ValueError(f"Number of dimensions of {key} != 3 (shape={ft.shape})")
+
+            shape = get_channel_first_image_shape(ft.shape)
+            feature = PolicyFeature(type=ft.type, shape=shape)
+        else:
+            feature = ft
+
+        policy_key = env_cfg.features_map[key]
+        policy_features[policy_key] = feature
+
+    return policy_features

lerobot/common/errors.py

@@ -0,0 +1,17 @@
+class DeviceNotConnectedError(ConnectionError):
+    """Exception raised when the device is not connected."""
+
+    def __init__(self, message="This device is not connected. Try calling `connect()` first."):
+        self.message = message
+        super().__init__(self.message)
+
+
+class DeviceAlreadyConnectedError(ConnectionError):
+    """Exception raised when the device is already connected."""
+
+    def __init__(
+        self,
+        message="This device is already connected. Try not calling `connect()` twice.",
+    ):
+        self.message = message
+        super().__init__(self.message)

lerobot/common/logger.py

@@ -1,246 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""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
-
-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: 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)
-
-
-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
-
-
-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.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)
-        run_offline = not enable_wandb or not project
-        if run_offline:
-            logging.info(colored("Logs will be saved locally.", "yellow", attrs=["bold"]))
-            self._wandb = None
-        else:
-            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=wandb_job_name,
-                notes=cfg.get("wandb", {}).get("notes"),
-                tags=cfg_to_group(cfg, return_list=True),
-                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",
-                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
-
-    @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"
-
-    @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 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_checkpoint(
-        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_path: str, step: int, mode: str = "train"):
-        assert mode in {"train", "eval"}
-        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/motors/__init__.py

@@ -0,0 +1,3 @@
+from .motors_bus import MotorsBus
+
+__all__ = ["MotorsBus"]

lerobot/common/motors/configs.py

@@ -0,0 +1,41 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import abc
+from dataclasses import dataclass
+
+import draccus
+
+
+@dataclass
+class MotorsBusConfig(draccus.ChoiceRegistry, abc.ABC):
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)
+
+
+@MotorsBusConfig.register_subclass("dynamixel")
+@dataclass
+class DynamixelMotorsBusConfig(MotorsBusConfig):
+    port: str
+    motors: dict[str, tuple[int, str]]
+    mock: bool = False
+
+
+@MotorsBusConfig.register_subclass("feetech")
+@dataclass
+class FeetechMotorsBusConfig(MotorsBusConfig):
+    port: str
+    motors: dict[str, tuple[int, str]]
+    mock: bool = False

lerobot/common/motors/dynamixel/__init__.py

@@ -0,0 +1,4 @@
+from .dynamixel import DynamixelMotorsBus, TorqueMode, set_operating_mode
+from .dynamixel_calibration import run_arm_calibration
+
+__all__ = ["DynamixelMotorsBus", "TorqueMode", "set_operating_mode", "run_arm_calibration"]

lerobot/common/motors/dynamixel/dynamixel.py

@@ -1,3 +1,17 @@
+# 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 enum
 import logging
 import math
@@ -8,7 +22,7 @@ from copy import deepcopy
 import numpy as np
 import tqdm
 
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.common.utils.utils import capture_timestamp_utc
 
 PROTOCOL_VERSION = 2.0
@@ -241,7 +255,7 @@ class DriveMode(enum.Enum):
 class CalibrationMode(enum.Enum):
     # Joints with rotational motions are expressed in degrees in nominal range of [-180, 180]
     DEGREE = 0
-    # Joints with linear motions (like gripper of Aloha) are experessed in nominal range of [0, 100]
+    # Joints with linear motions (like gripper of Aloha) are expressed in nominal range of [0, 100]
     LINEAR = 1
 
 
@@ -252,7 +266,6 @@ class JointOutOfRangeError(Exception):
 
 
 class DynamixelMotorsBus:
-    # TODO(rcadene): Add a script to find the motor indices without DynamixelWizzard2
     """
     The DynamixelMotorsBus class allows to efficiently read and write to the attached motors. It relies on
     the python dynamixel sdk to communicate with the motors. For more info, see the [Dynamixel SDK Documentation](https://emanual.robotis.com/docs/en/software/dynamixel/dynamixel_sdk/sample_code/python_read_write_protocol_2_0/#python-read-write-protocol-20).
@@ -295,21 +308,14 @@ class DynamixelMotorsBus:
         self,
         port: str,
         motors: dict[str, tuple[int, str]],
-        extra_model_control_table: dict[str, list[tuple]] | None = None,
-        extra_model_resolution: dict[str, int] | None = None,
-        mock=False,
+        mock: bool = False,
     ):
         self.port = port
         self.motors = motors
         self.mock = mock
 
         self.model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
-        if extra_model_control_table:
-            self.model_ctrl_table.update(extra_model_control_table)
-
         self.model_resolution = deepcopy(MODEL_RESOLUTION)
-        if extra_model_resolution:
-            self.model_resolution.update(extra_model_resolution)
 
         self.port_handler = None
         self.packet_handler = None
@@ -321,7 +327,7 @@ class DynamixelMotorsBus:
 
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
+            raise DeviceAlreadyConnectedError(
                 f"DynamixelMotorsBus({self.port}) is already connected. Do not call `motors_bus.connect()` twice."
             )
 
@@ -618,7 +624,7 @@ class DynamixelMotorsBus:
                 # 0-centered resolution range (e.g. [-2048, 2048] for resolution=4096)
                 values[i] = values[i] / HALF_TURN_DEGREE * (resolution // 2)
 
-                # Substract the homing offsets to come back to actual motor range of values
+                # Subtract the homing offsets to come back to actual motor range of values
                 # which can be arbitrary.
                 values[i] -= homing_offset
 
@@ -678,7 +684,7 @@ class DynamixelMotorsBus:
 
     def read(self, data_name, motor_names: str | list[str] | None = None):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"DynamixelMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
             )
 
@@ -780,7 +786,7 @@ class DynamixelMotorsBus:
 
     def write(self, data_name, values: int | float | np.ndarray, motor_names: str | list[str] | None = None):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"DynamixelMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
             )
 
@@ -849,7 +855,7 @@ class DynamixelMotorsBus:
 
     def disconnect(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"DynamixelMotorsBus({self.port}) is not connected. Try running `motors_bus.connect()` first."
             )
 
@@ -865,3 +871,25 @@ class DynamixelMotorsBus:
     def __del__(self):
         if getattr(self, "is_connected", False):
             self.disconnect()
+
+
+def set_operating_mode(arm: DynamixelMotorsBus):
+    if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+        raise ValueError("To run set robot preset, the torque must be disabled on all motors.")
+
+    # Use 'extended position mode' for all motors except gripper, because in joint mode the servos can't
+    # rotate more than 360 degrees (from 0 to 4095) And some mistake can happen while assembling the arm,
+    # you could end up with a servo with a position 0 or 4095 at a crucial point See [
+    # https://emanual.robotis.com/docs/en/dxl/x/x_series/#operating-mode11]
+    all_motors_except_gripper = [name for name in arm.motor_names if name != "gripper"]
+    if len(all_motors_except_gripper) > 0:
+        # 4 corresponds to Extended Position on Koch motors
+        arm.write("Operating_Mode", 4, all_motors_except_gripper)
+
+    # Use 'position control current based' for gripper to be limited by the limit of the current.
+    # For the follower gripper, it means it can grasp an object without forcing too much even tho,
+    # it's goal position is a complete grasp (both gripper fingers are ordered to join and reach a touch).
+    # For the leader gripper, it means we can use it as a physical trigger, since we can force with our finger
+    # to make it move, and it will move back to its original target position when we release the force.
+    # 5 corresponds to Current Controlled Position on Koch gripper motors "xl330-m077, xl330-m288"
+    arm.write("Operating_Mode", 5, "gripper")

lerobot/common/motors/dynamixel/dynamixel_calibration.py

@@ -1,14 +1,28 @@
+# 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.
+
 """Logic to calibrate a robot arm built with dynamixel motors"""
 # TODO(rcadene, aliberts): move this logic into the robot code when refactoring
 
 import numpy as np
 
-from lerobot.common.robot_devices.motors.dynamixel import (
+from ..motors_bus import MotorsBus
+from .dynamixel import (
     CalibrationMode,
     TorqueMode,
     convert_degrees_to_steps,
 )
-from lerobot.common.robot_devices.motors.utils import MotorsBus
 
 URL_TEMPLATE = (
     "https://raw.githubusercontent.com/huggingface/lerobot/main/media/{robot}/{arm}_{position}.webp"
@@ -87,7 +101,7 @@ def run_arm_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type
     # For instance, if the motor rotates 90 degree, and its value is -90 after applying the homing offset, then we know its rotation direction
     # is inverted. However, for the calibration being successful, we need everyone to follow the same target position.
     # Sometimes, there is only one possible rotation direction. For instance, if the gripper is closed, there is only one direction which
-    # corresponds to opening the gripper. When the rotation direction is ambiguous, we arbitrarely rotate clockwise from the point of view
+    # corresponds to opening the gripper. When the rotation direction is ambiguous, we arbitrarily rotate clockwise from the point of view
     # of the previous motor in the kinetic chain.
     print("\nMove arm to rotated target position")
     print("See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="rotated"))
@@ -115,7 +129,7 @@ def run_arm_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type
 
     # TODO(rcadene): make type of joints (DEGREE or LINEAR) configurable from yaml?
     if robot_type in ["aloha"] and "gripper" in arm.motor_names:
-        # Joints with linear motions (like gripper of Aloha) are experessed in nominal range of [0, 100]
+        # Joints with linear motions (like gripper of Aloha) are expressed in nominal range of [0, 100]
         calib_idx = arm.motor_names.index("gripper")
         calib_mode[calib_idx] = CalibrationMode.LINEAR.name
 

lerobot/common/motors/feetech/__init__.py

@@ -0,0 +1,4 @@
+from .feetech import FeetechMotorsBus, TorqueMode
+from .feetech_calibration import run_full_arm_calibration
+
+__all__ = ["FeetechMotorsBus", "TorqueMode", "run_full_arm_calibration"]

lerobot/common/motors/feetech/feetech.py

@@ -1,6 +1,18 @@
+# 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 enum
-import logging
-import math
 import time
 import traceback
 from copy import deepcopy
@@ -8,7 +20,7 @@ from copy import deepcopy
 import numpy as np
 import tqdm
 
-from lerobot.common.robot_devices.utils import RobotDeviceAlreadyConnectedError, RobotDeviceNotConnectedError
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.common.utils.utils import capture_timestamp_utc
 
 PROTOCOL_VERSION = 0
@@ -17,13 +29,6 @@ TIMEOUT_MS = 1000
 
 MAX_ID_RANGE = 252
 
-# The following bounds define the lower and upper joints range (after calibration).
-# For joints in degree (i.e. revolute joints), their nominal range is [-180, 180] degrees
-# which corresponds to a half rotation on the left and half rotation on the right.
-# Some joints might require higher range, so we allow up to [-270, 270] degrees until
-# an error is raised.
-LOWER_BOUND_DEGREE = -270
-UPPER_BOUND_DEGREE = 270
 # For joints in percentage (i.e. joints that move linearly like the prismatic joint of a gripper),
 # their nominal range is [0, 100] %. For instance, for Aloha gripper, 0% is fully
 # closed, and 100% is fully open. To account for slight calibration issue, we allow up to
@@ -33,7 +38,6 @@ UPPER_BOUND_LINEAR = 110
 
 HALF_TURN_DEGREE = 180
 
-
 # See this link for STS3215 Memory Table:
 # https://docs.google.com/spreadsheets/d/1GVs7W1VS1PqdhA1nW-abeyAHhTUxKUdR/edit?usp=sharing&ouid=116566590112741600240&rtpof=true&sd=true
 # data_name: (address, size_byte)
@@ -99,8 +103,6 @@ SCS_SERIES_BAUDRATE_TABLE = {
 }
 
 CALIBRATION_REQUIRED = ["Goal_Position", "Present_Position"]
-CONVERT_UINT32_TO_INT32_REQUIRED = ["Goal_Position", "Present_Position"]
-
 
 MODEL_CONTROL_TABLE = {
     "scs_series": SCS_SERIES_CONTROL_TABLE,
@@ -122,15 +124,68 @@ NUM_READ_RETRY = 20
 NUM_WRITE_RETRY = 20
 
 
-def convert_degrees_to_steps(degrees: float | np.ndarray, models: str | list[str]) -> np.ndarray:
-    """This function converts the degree range to the step range for indicating motors rotation.
-    It assumes a motor achieves a full rotation by going from -180 degree position to +180.
-    The motor resolution (e.g. 4096) corresponds to the number of steps needed to achieve a full rotation.
+def convert_ticks_to_degrees(ticks, model):
+    resolutions = MODEL_RESOLUTION[model]
+    # Convert the ticks to degrees
+    return ticks * (360.0 / resolutions)
+
+
+def convert_degrees_to_ticks(degrees, model):
+    resolutions = MODEL_RESOLUTION[model]
+    # Convert degrees to motor ticks
+    return int(degrees * (resolutions / 360.0))
+
+
+def adjusted_to_homing_ticks(
+    raw_motor_ticks: int, encoder_offset: int, model: str, motorbus, motor_id: int
+) -> int:
     """
-    resolutions = [MODEL_RESOLUTION[model] for model in models]
-    steps = degrees / 180 * np.array(resolutions) / 2
-    steps = steps.astype(int)
-    return steps
+    Shifts raw [0..4095] ticks by an encoder offset, modulo a single turn [0..4095].
+    """
+    resolutions = MODEL_RESOLUTION[model]
+
+    # Add offset and wrap within resolution
+    ticks = (raw_motor_ticks + encoder_offset) % resolutions
+
+    # # Re-center into a symmetric range (e.g., [-2048, 2047] if resolutions==4096) Thus the middle homing position will be virtual 0.
+    if ticks > resolutions // 2:
+        ticks -= resolutions
+
+    # Update direction of rotation of the motor to match between leader and follower.
+    # In fact, the motor of the leader for a given joint can be assembled in an
+    # opposite direction in term of rotation than the motor of the follower on the same joint.
+    drive_mode = 0
+    if motorbus.calibration is not None:
+        drive_mode = motorbus.calibration["drive_mode"][motor_id - 1]
+
+    if drive_mode:
+        ticks *= -1
+
+    return ticks
+
+
+def adjusted_to_motor_ticks(
+    adjusted_pos: int, encoder_offset: int, model: str, motorbus, motor_id: int
+) -> int:
+    """
+    Inverse of adjusted_to_homing_ticks().
+    """
+    # Update direction of rotation of the motor to match between leader and follower.
+    # In fact, the motor of the leader for a given joint can be assembled in an
+    # opposite direction in term of rotation than the motor of the follower on the same joint.
+    drive_mode = 0
+    if motorbus.calibration is not None:
+        drive_mode = motorbus.calibration["drive_mode"][motor_id - 1]
+
+    if drive_mode:
+        adjusted_pos *= -1
+
+    resolutions = MODEL_RESOLUTION[model]
+
+    # Remove offset and wrap within resolution
+    ticks = (adjusted_pos - encoder_offset) % resolutions
+
+    return ticks
 
 
 def convert_to_bytes(value, bytes, mock=False):
@@ -220,7 +275,7 @@ class DriveMode(enum.Enum):
 class CalibrationMode(enum.Enum):
     # Joints with rotational motions are expressed in degrees in nominal range of [-180, 180]
     DEGREE = 0
-    # Joints with linear motions (like gripper of Aloha) are experessed in nominal range of [0, 100]
+    # Joints with linear motions (like gripper of Aloha) are expressed in nominal range of [0, 100]
     LINEAR = 1
 
 
@@ -273,21 +328,14 @@ class FeetechMotorsBus:
         self,
         port: str,
         motors: dict[str, tuple[int, str]],
-        extra_model_control_table: dict[str, list[tuple]] | None = None,
-        extra_model_resolution: dict[str, int] | None = None,
-        mock=False,
+        mock: bool = False,
     ):
         self.port = port
         self.motors = motors
         self.mock = mock
 
         self.model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
-        if extra_model_control_table:
-            self.model_ctrl_table.update(extra_model_control_table)
-
         self.model_resolution = deepcopy(MODEL_RESOLUTION)
-        if extra_model_resolution:
-            self.model_resolution.update(extra_model_resolution)
 
         self.port_handler = None
         self.packet_handler = None
@@ -297,11 +345,9 @@ class FeetechMotorsBus:
         self.group_writers = {}
         self.logs = {}
 
-        self.track_positions = {}
-
     def connect(self):
         if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
+            raise DeviceAlreadyConnectedError(
                 f"FeetechMotorsBus({self.port}) is already connected. Do not call `motors_bus.connect()` twice."
             )
 
@@ -395,33 +441,7 @@ class FeetechMotorsBus:
     def set_calibration(self, calibration: dict[str, list]):
         self.calibration = calibration
 
-    def apply_calibration_autocorrect(self, values: np.ndarray | list, motor_names: list[str] | None):
-        """This function apply the calibration, automatically detects out of range errors for motors values and attempt to correct.
-
-        For more info, see docstring of `apply_calibration` and `autocorrect_calibration`.
-        """
-        try:
-            values = self.apply_calibration(values, motor_names)
-        except JointOutOfRangeError as e:
-            print(e)
-            self.autocorrect_calibration(values, motor_names)
-            values = self.apply_calibration(values, motor_names)
-        return values
-
     def apply_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
-        """Convert from unsigned int32 joint position range [0, 2**32[ to the universal float32 nominal degree range ]-180.0, 180.0[ with
-        a "zero position" at 0 degree.
-
-        Note: We say "nominal degree range" since the motors can take values outside this range. For instance, 190 degrees, if the motor
-        rotate more than a half a turn from the zero position. However, most motors can't rotate more than 180 degrees and will stay in this range.
-
-        Joints values are original in [0, 2**32[ (unsigned int32). Each motor are expected to complete a full rotation
-        when given a goal position that is + or - their resolution. For instance, feetech xl330-m077 have a resolution of 4096, and
-        at any position in their original range, let's say the position 56734, they complete a full rotation clockwise by moving to 60830,
-        or anticlockwise by moving to 52638. The position in the original range is arbitrary and might change a lot between each motor.
-        To harmonize between motors of the same model, different robots, or even models of different brands, we propose to work
-        in the centered nominal degree range ]-180, 180[.
-        """
         if motor_names is None:
             motor_names = self.motor_names
 
@@ -433,34 +453,12 @@ class FeetechMotorsBus:
             calib_mode = self.calibration["calib_mode"][calib_idx]
 
             if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
-                drive_mode = self.calibration["drive_mode"][calib_idx]
                 homing_offset = self.calibration["homing_offset"][calib_idx]
-                _, model = self.motors[name]
-                resolution = self.model_resolution[model]
+                motor_idx, model = self.motors[name]
 
-                # Update direction of rotation of the motor to match between leader and follower.
-                # In fact, the motor of the leader for a given joint can be assembled in an
-                # opposite direction in term of rotation than the motor of the follower on the same joint.
-                if drive_mode:
-                    values[i] *= -1
-
-                # Convert from range [-2**31, 2**31[ to
-                # nominal range ]-resolution, resolution[ (e.g. ]-2048, 2048[)
-                values[i] += homing_offset
-
-                # Convert from range ]-resolution, resolution[ to
-                # universal float32 centered degree range ]-180, 180[
-                values[i] = values[i] / (resolution // 2) * HALF_TURN_DEGREE
-
-                if (values[i] < LOWER_BOUND_DEGREE) or (values[i] > UPPER_BOUND_DEGREE):
-                    raise JointOutOfRangeError(
-                        f"Wrong motor position range detected for {name}. "
-                        f"Expected to be in nominal range of [-{HALF_TURN_DEGREE}, {HALF_TURN_DEGREE}] degrees (a full rotation), "
-                        f"with a maximum range of [{LOWER_BOUND_DEGREE}, {UPPER_BOUND_DEGREE}] degrees to account for joints that can rotate a bit more, "
-                        f"but present value is {values[i]} degree. "
-                        "This might be due to a cable connection issue creating an artificial 360 degrees jump in motor values. "
-                        "You need to recalibrate by running: `python lerobot/scripts/control_robot.py calibrate`"
-                    )
+                # Convert raw motor ticks to homed ticks, then convert the homed ticks to degrees
+                values[i] = adjusted_to_homing_ticks(values[i], homing_offset, model, self, motor_idx)
+                values[i] = convert_ticks_to_degrees(values[i], model)
 
             elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
                 start_pos = self.calibration["start_pos"][calib_idx]
@@ -482,103 +480,6 @@ class FeetechMotorsBus:
 
         return values
 
-    def autocorrect_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
-        """This function automatically detects issues with values of motors after calibration, and correct for these issues.
-
-        Some motors might have values outside of expected maximum bounds after calibration.
-        For instance, for a joint in degree, its value can be outside [-270, 270] degrees, which is totally unexpected given
-        a nominal range of [-180, 180] degrees, which represents half a turn to the left or right starting from zero position.
-
-        Known issues:
-        #1: Motor value randomly shifts of a full turn, caused by hardware/connection errors.
-        #2: Motor internal homing offset is shifted of a full turn, caused by using default calibration (e.g Aloha).
-        #3: motor internal homing offset is shifted of less or more than a full turn, caused by using default calibration
-            or by human error during manual calibration.
-
-        Issues #1 and #2 can be solved by shifting the calibration homing offset by a full turn.
-        Issue #3 will be visually detected by user and potentially captured by the safety feature `max_relative_target`,
-        that will slow down the motor, raise an error asking to recalibrate. Manual recalibrating will solve the issue.
-
-        Note: A full turn corresponds to 360 degrees but also to 4096 steps for a motor resolution of 4096.
-        """
-        if motor_names is None:
-            motor_names = self.motor_names
-
-        # Convert from unsigned int32 original range [0, 2**32] to signed float32 range
-        values = values.astype(np.float32)
-
-        for i, name in enumerate(motor_names):
-            calib_idx = self.calibration["motor_names"].index(name)
-            calib_mode = self.calibration["calib_mode"][calib_idx]
-
-            if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
-                drive_mode = self.calibration["drive_mode"][calib_idx]
-                homing_offset = self.calibration["homing_offset"][calib_idx]
-                _, model = self.motors[name]
-                resolution = self.model_resolution[model]
-
-                if drive_mode:
-                    values[i] *= -1
-
-                # Convert from initial range to range [-180, 180] degrees
-                calib_val = (values[i] + homing_offset) / (resolution // 2) * HALF_TURN_DEGREE
-                in_range = (calib_val > LOWER_BOUND_DEGREE) and (calib_val < UPPER_BOUND_DEGREE)
-
-                # Solve this inequality to find the factor to shift the range into [-180, 180] degrees
-                # values[i] = (values[i] + homing_offset + resolution * factor) / (resolution // 2) * HALF_TURN_DEGREE
-                # - HALF_TURN_DEGREE <= (values[i] + homing_offset + resolution * factor) / (resolution // 2) * HALF_TURN_DEGREE <= HALF_TURN_DEGREE
-                # (- HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset) / resolution <= factor <= (HALF_TURN_DEGREE / 180 * (resolution // 2) - values[i] - homing_offset) / resolution
-                low_factor = (
-                    -HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset
-                ) / resolution
-                upp_factor = (
-                    HALF_TURN_DEGREE / HALF_TURN_DEGREE * (resolution // 2) - values[i] - homing_offset
-                ) / resolution
-
-            elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
-                start_pos = self.calibration["start_pos"][calib_idx]
-                end_pos = self.calibration["end_pos"][calib_idx]
-
-                # Convert from initial range to range [0, 100] in %
-                calib_val = (values[i] - start_pos) / (end_pos - start_pos) * 100
-                in_range = (calib_val > LOWER_BOUND_LINEAR) and (calib_val < UPPER_BOUND_LINEAR)
-
-                # Solve this inequality to find the factor to shift the range into [0, 100] %
-                # values[i] = (values[i] - start_pos + resolution * factor) / (end_pos + resolution * factor - start_pos - resolution * factor) * 100
-                # values[i] = (values[i] - start_pos + resolution * factor) / (end_pos - start_pos) * 100
-                # 0 <= (values[i] - start_pos + resolution * factor) / (end_pos - start_pos) * 100 <= 100
-                # (start_pos - values[i]) / resolution <= factor <= (end_pos - values[i]) / resolution
-                low_factor = (start_pos - values[i]) / resolution
-                upp_factor = (end_pos - values[i]) / resolution
-
-            if not in_range:
-                # Get first integer between the two bounds
-                if low_factor < upp_factor:
-                    factor = math.ceil(low_factor)
-
-                    if factor > upp_factor:
-                        raise ValueError(f"No integer found between bounds [{low_factor=}, {upp_factor=}]")
-                else:
-                    factor = math.ceil(upp_factor)
-
-                    if factor > low_factor:
-                        raise ValueError(f"No integer found between bounds [{low_factor=}, {upp_factor=}]")
-
-                if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
-                    out_of_range_str = f"{LOWER_BOUND_DEGREE} < {calib_val} < {UPPER_BOUND_DEGREE} degrees"
-                    in_range_str = f"{LOWER_BOUND_DEGREE} < {calib_val} < {UPPER_BOUND_DEGREE} degrees"
-                elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
-                    out_of_range_str = f"{LOWER_BOUND_LINEAR} < {calib_val} < {UPPER_BOUND_LINEAR} %"
-                    in_range_str = f"{LOWER_BOUND_LINEAR} < {calib_val} < {UPPER_BOUND_LINEAR} %"
-
-                logging.warning(
-                    f"Auto-correct calibration of motor '{name}' by shifting value by {abs(factor)} full turns, "
-                    f"from '{out_of_range_str}' to '{in_range_str}'."
-                )
-
-                # A full turn corresponds to 360 degrees but also to 4096 steps for a motor resolution of 4096.
-                self.calibration["homing_offset"][calib_idx] += resolution * factor
-
     def revert_calibration(self, values: np.ndarray | list, motor_names: list[str] | None):
         """Inverse of `apply_calibration`."""
         if motor_names is None:
@@ -589,23 +490,12 @@ class FeetechMotorsBus:
             calib_mode = self.calibration["calib_mode"][calib_idx]
 
             if CalibrationMode[calib_mode] == CalibrationMode.DEGREE:
-                drive_mode = self.calibration["drive_mode"][calib_idx]
                 homing_offset = self.calibration["homing_offset"][calib_idx]
-                _, model = self.motors[name]
-                resolution = self.model_resolution[model]
+                motor_idx, model = self.motors[name]
 
-                # Convert from nominal 0-centered degree range [-180, 180] to
-                # 0-centered resolution range (e.g. [-2048, 2048] for resolution=4096)
-                values[i] = values[i] / HALF_TURN_DEGREE * (resolution // 2)
-
-                # Substract the homing offsets to come back to actual motor range of values
-                # which can be arbitrary.
-                values[i] -= homing_offset
-
-                # Remove drive mode, which is the rotation direction of the motor, to come back to
-                # actual motor rotation direction which can be arbitrary.
-                if drive_mode:
-                    values[i] *= -1
+                # Convert degrees to homed ticks, then convert the homed ticks to raw ticks
+                values[i] = convert_degrees_to_ticks(values[i], model)
+                values[i] = adjusted_to_motor_ticks(values[i], homing_offset, model, self, motor_idx)
 
             elif CalibrationMode[calib_mode] == CalibrationMode.LINEAR:
                 start_pos = self.calibration["start_pos"][calib_idx]
@@ -618,43 +508,6 @@ class FeetechMotorsBus:
         values = np.round(values).astype(np.int32)
         return values
 
-    def avoid_rotation_reset(self, values, motor_names, data_name):
-        if data_name not in self.track_positions:
-            self.track_positions[data_name] = {
-                "prev": [None] * len(self.motor_names),
-                # Assume False at initialization
-                "below_zero": [False] * len(self.motor_names),
-                "above_max": [False] * len(self.motor_names),
-            }
-
-        track = self.track_positions[data_name]
-
-        if motor_names is None:
-            motor_names = self.motor_names
-
-        for i, name in enumerate(motor_names):
-            idx = self.motor_names.index(name)
-
-            if track["prev"][idx] is None:
-                track["prev"][idx] = values[i]
-                continue
-
-            # Detect a full rotation occured
-            if abs(track["prev"][idx] - values[i]) > 2048:
-                # Position went below 0 and got reset to 4095
-                if track["prev"][idx] < values[i]:
-                    # So we set negative value by adding a full rotation
-                    values[i] -= 4096
-
-                # Position went above 4095 and got reset to 0
-                elif track["prev"][idx] > values[i]:
-                    # So we add a full rotation
-                    values[i] += 4096
-
-            track["prev"][idx] = values[i]
-
-        return values
-
     def read_with_motor_ids(self, motor_models, motor_ids, data_name, num_retry=NUM_READ_RETRY):
         if self.mock:
             import tests.mock_scservo_sdk as scs
@@ -700,7 +553,7 @@ class FeetechMotorsBus:
             import scservo_sdk as scs
 
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"FeetechMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
             )
 
@@ -724,7 +577,11 @@ class FeetechMotorsBus:
         group_key = get_group_sync_key(data_name, motor_names)
 
         if data_name not in self.group_readers:
-            # create new group reader
+            # Very Important to flush the buffer!
+            self.port_handler.ser.reset_output_buffer()
+            self.port_handler.ser.reset_input_buffer()
+
+            # Create new group reader
             self.group_readers[group_key] = scs.GroupSyncRead(
                 self.port_handler, self.packet_handler, addr, bytes
             )
@@ -749,15 +606,8 @@ class FeetechMotorsBus:
 
         values = np.array(values)
 
-        # Convert to signed int to use range [-2048, 2048] for our motor positions.
-        if data_name in CONVERT_UINT32_TO_INT32_REQUIRED:
-            values = values.astype(np.int32)
-
-        if data_name in CALIBRATION_REQUIRED:
-            values = self.avoid_rotation_reset(values, motor_names, data_name)
-
         if data_name in CALIBRATION_REQUIRED and self.calibration is not None:
-            values = self.apply_calibration_autocorrect(values, motor_names)
+            values = self.apply_calibration(values, motor_names)
 
         # log the number of seconds it took to read the data from the motors
         delta_ts_name = get_log_name("delta_timestamp_s", "read", data_name, motor_names)
@@ -800,7 +650,7 @@ class FeetechMotorsBus:
 
     def write(self, data_name, values: int | float | np.ndarray, motor_names: str | list[str] | None = None):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"FeetechMotorsBus({self.port}) is not connected. You need to run `motors_bus.connect()`."
             )
 
@@ -869,7 +719,7 @@ class FeetechMotorsBus:
 
     def disconnect(self):
         if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
+            raise DeviceNotConnectedError(
                 f"FeetechMotorsBus({self.port}) is not connected. Try running `motors_bus.connect()` first."
             )
 

lerobot/common/motors/feetech/feetech_calibration.py

@@ -0,0 +1,184 @@
+# 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.
+
+"""Logic to calibrate a robot arm built with feetech motors"""
+# TODO(rcadene, aliberts): move this logic into the robot code when refactoring
+
+import numpy as np
+
+from ..motors_bus import MotorsBus
+from .feetech import (
+    CalibrationMode,
+    TorqueMode,
+)
+
+URL_TEMPLATE = (
+    "https://raw.githubusercontent.com/huggingface/lerobot/main/media/{robot}/{arm}_{position}.webp"
+)
+
+
+def disable_torque(arm: MotorsBus):
+    if (arm.read("Torque_Enable") != TorqueMode.DISABLED.value).any():
+        raise ValueError("To run calibration, the torque must be disabled on all motors.")
+
+
+def get_calibration_modes(arm: MotorsBus):
+    """Returns calibration modes for each motor (DEGREE for rotational, LINEAR for gripper)."""
+    return [
+        CalibrationMode.LINEAR.name if name == "gripper" else CalibrationMode.DEGREE.name
+        for name in arm.motor_names
+    ]
+
+
+def calibrate_homing_motor(motor_id, motor_bus):
+    """
+    1) Reads servo ticks.
+    2) Calculates the offset so that 'home_ticks' becomes 0.
+    3) Returns the offset
+    """
+
+    home_ticks = motor_bus.read("Present_Position")[motor_id - 1]  # Read index starts at 0
+
+    # Calculate how many ticks to shift so that 'home_ticks' becomes 0
+    raw_offset = -home_ticks  # negative of home_ticks
+    encoder_offset = raw_offset % 4096  # wrap to [0..4095]
+
+    # Convert to a signed range [-2048..2047]
+    if encoder_offset > 2047:
+        encoder_offset -= 4096
+
+    print(f"Encoder offset: {encoder_offset}")
+
+    return encoder_offset
+
+
+def calibrate_linear_motor(motor_id, motor_bus):
+    motor_names = motor_bus.motor_names
+    motor_name = motor_names[motor_id - 1]  # TODO(pepijn): replace motor_id with motor index when (id-1)
+
+    input(f"Close the {motor_name}, then press Enter...")
+    start_pos = motor_bus.read("Present_Position")[motor_id - 1]  # Read index starts ar 0
+    print(f"  [Motor {motor_id}] start position recorded: {start_pos}")
+
+    input(f"Open the {motor_name} fully, then press Enter...")
+    end_pos = motor_bus.read("Present_Position")[motor_id - 1]  # Read index starts ar 0
+    print(f"  [Motor {motor_id}] end position recorded: {end_pos}")
+
+    return start_pos, end_pos
+
+
+def single_motor_calibration(arm: MotorsBus, motor_id: int):
+    """Calibrates a single motor and returns its calibration data for updating the calibration file."""
+
+    disable_torque(arm)
+    print(f"\n--- Calibrating Motor {motor_id} ---")
+
+    start_pos = 0
+    end_pos = 0
+    encoder_offset = 0
+
+    if motor_id == 6:
+        start_pos, end_pos = calibrate_linear_motor(motor_id, arm)
+    else:
+        input("Move the motor to (zero) position, then press Enter...")
+        encoder_offset = calibrate_homing_motor(motor_id, arm)
+
+    print(f"Calibration for motor ID:{motor_id} done.")
+
+    # Create a calibration dictionary for the single motor
+    calib_dict = {
+        "homing_offset": int(encoder_offset),
+        "drive_mode": 0,
+        "start_pos": int(start_pos),
+        "end_pos": int(end_pos),
+        "calib_mode": get_calibration_modes(arm)[motor_id - 1],
+        "motor_name": arm.motor_names[motor_id - 1],
+    }
+
+    return calib_dict
+
+
+def run_full_arm_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    """
+    Runs a full calibration process for all motors in a robotic arm.
+
+    This function calibrates each motor in the arm, determining encoder offsets and
+    start/end positions for linear and rotational motors. The calibration data is then
+    stored in a dictionary for later use.
+
+    **Calibration Process:**
+    - The user is prompted to move the arm to its homing position before starting.
+    - Motors with rotational motion are calibrated using a homing method.
+    - Linear actuators (e.g., grippers) are calibrated separately.
+    - Encoder offsets, start positions, and end positions are recorded.
+
+    **Example Usage:**
+    ```python
+    run_full_arm_calibration(arm, "so100", "left", "follower")
+    ```
+    """
+    disable_torque(arm)
+
+    print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")
+
+    print("\nMove arm to homing position (middle)")
+    print(
+        "See: " + URL_TEMPLATE.format(robot=robot_type, arm=arm_type, position="zero")
+    )  # TODO(pepijn): replace with new instruction homing pos (all motors in middle) in tutorial
+    input("Press Enter to continue...")
+
+    start_positions = np.zeros(len(arm.motor_indices))
+    end_positions = np.zeros(len(arm.motor_indices))
+    encoder_offsets = np.zeros(len(arm.motor_indices))
+
+    modes = get_calibration_modes(arm)
+
+    for i, motor_id in enumerate(arm.motor_indices):
+        if modes[i] == CalibrationMode.DEGREE.name:
+            encoder_offsets[i] = calibrate_homing_motor(motor_id, arm)
+            start_positions[i] = 0
+            end_positions[i] = 0
+
+    for i, motor_id in enumerate(arm.motor_indices):
+        if modes[i] == CalibrationMode.LINEAR.name:
+            start_positions[i], end_positions[i] = calibrate_linear_motor(motor_id, arm)
+            encoder_offsets[i] = 0
+
+    print("\nMove arm to rest position")
+    input("Press Enter to continue...")
+
+    print(f"\n calibration of {robot_type} {arm_name} {arm_type} done!")
+
+    # Force drive_mode values: motors 2 and 5 -> drive_mode 1; all others -> 0. 1 = clockwise = positive range (0..180), 0 = clockwise = negative range (0..-180)
+    drive_modes = [0, 1, 0, 0, 1, 0]
+
+    calib_dict = {
+        "homing_offset": encoder_offsets.astype(int).tolist(),
+        "drive_mode": drive_modes,
+        "start_pos": start_positions.astype(int).tolist(),
+        "end_pos": end_positions.astype(int).tolist(),
+        "calib_mode": get_calibration_modes(arm),
+        "motor_names": arm.motor_names,
+    }
+    return calib_dict
+
+
+def run_full_auto_arm_calibration(arm: MotorsBus, robot_type: str, arm_name: str, arm_type: str):
+    """TODO(pepijn): Add this method later as extra
+    Example of usage:
+    ```python
+    run_full_auto_arm_calibration(arm, "so100", "left", "follower")
+    ```
+    """
+    print(f"\nRunning calibration of {robot_type} {arm_name} {arm_type}...")

lerobot/common/motors/motors_bus.py

@@ -0,0 +1,46 @@
+import abc
+
+
+class MotorsBus(abc.ABC):
+    """The main LeRobot class for implementing motors buses."""
+
+    def __init__(
+        self,
+        motors: dict[str, tuple[int, str]],
+    ):
+        self.motors = motors
+
+    def __len__(self):
+        return len(self.motors)
+
+    @abc.abstractmethod
+    def connect(self):
+        pass
+
+    @abc.abstractmethod
+    def reconnect(self):
+        pass
+
+    @abc.abstractmethod
+    def set_calibration(self, calibration: dict[str, list]):
+        pass
+
+    @abc.abstractmethod
+    def apply_calibration(self):
+        pass
+
+    @abc.abstractmethod
+    def revert_calibration(self):
+        pass
+
+    @abc.abstractmethod
+    def read(self):
+        pass
+
+    @abc.abstractmethod
+    def write(self):
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self):
+        pass

lerobot/common/motors/utils.py

@@ -0,0 +1,56 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configs import MotorsBusConfig
+from .motors_bus import MotorsBus
+
+
+def make_motors_buses_from_configs(motors_bus_configs: dict[str, MotorsBusConfig]) -> list[MotorsBus]:
+    motors_buses = {}
+
+    for key, cfg in motors_bus_configs.items():
+        if cfg.type == "dynamixel":
+            from .dynamixel import DynamixelMotorsBus
+
+            motors_buses[key] = DynamixelMotorsBus(cfg)
+
+        elif cfg.type == "feetech":
+            from lerobot.common.motors.feetech.feetech import FeetechMotorsBus
+
+            motors_buses[key] = FeetechMotorsBus(cfg)
+
+        else:
+            raise ValueError(f"The motor type '{cfg.type}' is not valid.")
+
+    return motors_buses
+
+
+def make_motors_bus(motor_type: str, **kwargs) -> MotorsBus:
+    if motor_type == "dynamixel":
+        from .configs import DynamixelMotorsBusConfig
+        from .dynamixel import DynamixelMotorsBus
+
+        config = DynamixelMotorsBusConfig(**kwargs)
+        return DynamixelMotorsBus(config)
+
+    elif motor_type == "feetech":
+        from feetech import FeetechMotorsBus
+
+        from .configs import FeetechMotorsBusConfig
+
+        config = FeetechMotorsBusConfig(**kwargs)
+        return FeetechMotorsBus(config)
+
+    else:
+        raise ValueError(f"The motor type '{motor_type}' is not valid.")

lerobot/common/optim/__init__.py

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

lerobot/common/optim/factory.py

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

lerobot/common/optim/optimizers.py

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

lerobot/common/optim/schedulers.py

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

lerobot/common/policies/__init__.py

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

lerobot/common/policies/act/configuration_act.py

@@ -15,9 +15,14 @@
 # limitations under the License.
 from dataclasses import dataclass, field
 
+from lerobot.common.optim.optimizers import AdamWConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
 
+
+@PreTrainedConfig.register_subclass("act")
 @dataclass
-class ACTConfig:
+class ACTConfig(PreTrainedConfig):
     """Configuration class for the Action Chunking Transformers policy.
 
     Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
@@ -59,7 +64,7 @@ class ACTConfig:
         output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
             original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
-        pretrained_backbone_weights: Pretrained weights from torchvision to initalize the backbone.
+        pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.
         replace_final_stride_with_dilation: Whether to replace the ResNet's final 2x2 stride with a dilated
             convolution.
@@ -90,28 +95,11 @@ class ACTConfig:
     chunk_size: int = 100
     n_action_steps: int = 100
 
-    input_shapes: dict[str, list[int]] = field(
+    normalization_mapping: dict[str, NormalizationMode] = field(
         default_factory=lambda: {
-            "observation.images.top": [3, 480, 640],
-            "observation.state": [14],
-        }
-    )
-    output_shapes: dict[str, list[int]] = field(
-        default_factory=lambda: {
-            "action": [14],
-        }
-    )
-
-    # Normalization / Unnormalization
-    input_normalization_modes: dict[str, str] = field(
-        default_factory=lambda: {
-            "observation.images.top": "mean_std",
-            "observation.state": "mean_std",
-        }
-    )
-    output_normalization_modes: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": "mean_std",
+            "VISUAL": NormalizationMode.MEAN_STD,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MEAN_STD,
         }
     )
 
@@ -144,7 +132,14 @@ class ACTConfig:
     dropout: float = 0.1
     kl_weight: float = 10.0
 
+    # Training preset
+    optimizer_lr: float = 1e-5
+    optimizer_weight_decay: float = 1e-4
+    optimizer_lr_backbone: float = 1e-5
+
     def __post_init__(self):
+        super().__post_init__()
+
         """Input validation (not exhaustive)."""
         if not self.vision_backbone.startswith("resnet"):
             raise ValueError(
@@ -164,8 +159,28 @@ class ACTConfig:
             raise ValueError(
                 f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
             )
-        if (
-            not any(k.startswith("observation.image") for k in self.input_shapes)
-            and "observation.environment_state" not in self.input_shapes
-        ):
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            weight_decay=self.optimizer_weight_decay,
+        )
+
+    def get_scheduler_preset(self) -> None:
+        return None
+
+    def validate_features(self) -> None:
+        if not self.image_features and not self.env_state_feature:
             raise ValueError("You must provide at least one image or the environment state among the inputs.")
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

lerobot/common/policies/act/modeling_act.py

@@ -29,32 +29,27 @@ import numpy as np
 import torch
 import torch.nn.functional as F  # noqa: N812
 import torchvision
-from huggingface_hub import PyTorchModelHubMixin
 from torch import Tensor, nn
 from torchvision.models._utils import IntermediateLayerGetter
 from torchvision.ops.misc import FrozenBatchNorm2d
 
 from lerobot.common.policies.act.configuration_act import ACTConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pretrained import PreTrainedPolicy
 
 
-class ACTPolicy(
-    nn.Module,
-    PyTorchModelHubMixin,
-    library_name="lerobot",
-    repo_url="https://github.com/huggingface/lerobot",
-    tags=["robotics", "act"],
-):
+class ACTPolicy(PreTrainedPolicy):
     """
     Action Chunking Transformer Policy as per Learning Fine-Grained Bimanual Manipulation with Low-Cost
     Hardware (paper: https://arxiv.org/abs/2304.13705, code: https://github.com/tonyzhaozh/act)
     """
 
+    config_class = ACTConfig
     name = "act"
 
     def __init__(
         self,
-        config: ACTConfig | None = None,
+        config: ACTConfig,
         dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
@@ -64,30 +59,46 @@ class ACTPolicy(
             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
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
 
-        self.normalize_inputs = Normalize(
-            config.input_shapes, config.input_normalization_modes, dataset_stats
-        )
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
         self.normalize_targets = Normalize(
-            config.output_shapes, config.output_normalization_modes, dataset_stats
+            config.output_features, config.normalization_mapping, dataset_stats
         )
         self.unnormalize_outputs = Unnormalize(
-            config.output_shapes, config.output_normalization_modes, dataset_stats
+            config.output_features, config.normalization_mapping, 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 get_optim_params(self) -> dict:
+        # TODO(aliberts, rcadene): As of now, lr_backbone == lr
+        # Should we remove this and just `return self.parameters()`?
+        return [
+            {
+                "params": [
+                    p
+                    for n, p in self.named_parameters()
+                    if not n.startswith("model.backbone") and p.requires_grad
+                ]
+            },
+            {
+                "params": [
+                    p
+                    for n, p in self.named_parameters()
+                    if n.startswith("model.backbone") and p.requires_grad
+                ],
+                "lr": self.config.optimizer_lr_backbone,
+            },
+        ]
+
     def reset(self):
         """This should be called whenever the environment is reset."""
         if self.config.temporal_ensemble_coeff is not None:
@@ -106,9 +117,11 @@ class ACTPolicy(
         self.eval()
 
         batch = self.normalize_inputs(batch)
-        if len(self.expected_image_keys) > 0:
+        if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+            batch["observation.images"] = torch.stack(
+                [batch[key] for key in self.config.image_features], dim=-4
+            )
 
         # If we are doing temporal ensembling, do online updates where we keep track of the number of actions
         # we are ensembling over.
@@ -131,12 +144,14 @@ class ACTPolicy(
             self._action_queue.extend(actions.transpose(0, 1))
         return self._action_queue.popleft()
 
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
-        if len(self.expected_image_keys) > 0:
+        if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+            batch["observation.images"] = torch.stack(
+                [batch[key] for key in self.config.image_features], dim=-4
+            )
         batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
@@ -154,11 +169,11 @@ class ACTPolicy(
                 (-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
+            loss = l1_loss + mean_kld * self.config.kl_weight
         else:
-            loss_dict["loss"] = l1_loss
+            loss = l1_loss
 
-        return loss_dict
+        return loss, loss_dict
 
 
 class ACTTemporalEnsembler:
@@ -288,31 +303,30 @@ class ACT(nn.Module):
     """
 
     def __init__(self, config: ACTConfig):
-        super().__init__()
-        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]).
-        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
+        super().__init__()
+        self.config = config
+
         if self.config.use_vae:
             self.vae_encoder = ACTEncoder(config, is_vae_encoder=True)
             self.vae_encoder_cls_embed = nn.Embedding(1, config.dim_model)
             # Projection layer for joint-space configuration to hidden dimension.
-            if self.use_robot_state:
+            if self.config.robot_state_feature:
                 self.vae_encoder_robot_state_input_proj = nn.Linear(
-                    config.input_shapes["observation.state"][0], config.dim_model
+                    self.config.robot_state_feature.shape[0], config.dim_model
                 )
             # Projection layer for action (joint-space target) to hidden dimension.
             self.vae_encoder_action_input_proj = nn.Linear(
-                config.output_shapes["action"][0], config.dim_model
+                self.config.action_feature.shape[0],
+                config.dim_model,
             )
             # Projection layer from the VAE encoder's output to the latent distribution's parameter space.
             self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, config.latent_dim * 2)
             # Fixed sinusoidal positional embedding for the input to the VAE encoder. Unsqueeze for batch
             # dimension.
             num_input_token_encoder = 1 + config.chunk_size
-            if self.use_robot_state:
+            if self.config.robot_state_feature:
                 num_input_token_encoder += 1
             self.register_buffer(
                 "vae_encoder_pos_enc",
@@ -320,7 +334,7 @@ class ACT(nn.Module):
             )
 
         # Backbone for image feature extraction.
-        if self.use_images:
+        if self.config.image_features:
             backbone_model = getattr(torchvision.models, config.vision_backbone)(
                 replace_stride_with_dilation=[False, False, config.replace_final_stride_with_dilation],
                 weights=config.pretrained_backbone_weights,
@@ -337,27 +351,27 @@ class ACT(nn.Module):
 
         # Transformer encoder input projections. The tokens will be structured like
         # [latent, (robot_state), (env_state), (image_feature_map_pixels)].
-        if self.use_robot_state:
+        if self.config.robot_state_feature:
             self.encoder_robot_state_input_proj = nn.Linear(
-                config.input_shapes["observation.state"][0], config.dim_model
+                self.config.robot_state_feature.shape[0], config.dim_model
             )
-        if self.use_env_state:
+        if self.config.env_state_feature:
             self.encoder_env_state_input_proj = nn.Linear(
-                config.input_shapes["observation.environment_state"][0], config.dim_model
+                self.config.env_state_feature.shape[0], config.dim_model
             )
         self.encoder_latent_input_proj = nn.Linear(config.latent_dim, config.dim_model)
-        if self.use_images:
+        if self.config.image_features:
             self.encoder_img_feat_input_proj = nn.Conv2d(
                 backbone_model.fc.in_features, config.dim_model, kernel_size=1
             )
         # Transformer encoder positional embeddings.
         n_1d_tokens = 1  # for the latent
-        if self.use_robot_state:
+        if self.config.robot_state_feature:
             n_1d_tokens += 1
-        if self.use_env_state:
+        if self.config.env_state_feature:
             n_1d_tokens += 1
         self.encoder_1d_feature_pos_embed = nn.Embedding(n_1d_tokens, config.dim_model)
-        if self.use_images:
+        if self.config.image_features:
             self.encoder_cam_feat_pos_embed = ACTSinusoidalPositionEmbedding2d(config.dim_model // 2)
 
         # Transformer decoder.
@@ -365,7 +379,7 @@ class ACT(nn.Module):
         self.decoder_pos_embed = nn.Embedding(config.chunk_size, config.dim_model)
 
         # Final action regression head on the output of the transformer's decoder.
-        self.action_head = nn.Linear(config.dim_model, config.output_shapes["action"][0])
+        self.action_head = nn.Linear(config.dim_model, self.config.action_feature.shape[0])
 
         self._reset_parameters()
 
@@ -380,13 +394,13 @@ class ACT(nn.Module):
 
         `batch` should have the following structure:
         {
-            "observation.state" (optional): (B, state_dim) batch of robot states.
+            [robot_state_feature] (optional): (B, state_dim) batch of robot states.
 
-            "observation.images": (B, n_cameras, C, H, W) batch of images.
+            [image_features]: (B, n_cameras, C, H, W) batch of images.
                 AND/OR
-            "observation.environment_state": (B, env_dim) batch of environment states.
+            [env_state_feature]: (B, env_dim) batch of environment states.
 
-            "action" (optional, only if training with VAE): (B, chunk_size, action dim) batch of actions.
+            [action_feature] (optional, only if training with VAE): (B, chunk_size, action dim) batch of actions.
         }
 
         Returns:
@@ -395,9 +409,9 @@ class ACT(nn.Module):
             latent dimension.
         """
         if self.config.use_vae and self.training:
-            assert (
-                "action" in batch
-            ), "actions must be provided when using the variational objective in training mode."
+            assert "action" in batch, (
+                "actions must be provided when using the variational objective in training mode."
+            )
 
         batch_size = (
             batch["observation.images"]
@@ -411,12 +425,12 @@ class ACT(nn.Module):
             cls_embed = einops.repeat(
                 self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
             )  # (B, 1, D)
-            if self.use_robot_state:
+            if self.config.robot_state_feature:
                 robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
                 robot_state_embed = robot_state_embed.unsqueeze(1)  # (B, 1, D)
             action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
 
-            if self.use_robot_state:
+            if self.config.robot_state_feature:
                 vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
             else:
                 vae_encoder_input = [cls_embed, action_embed]
@@ -430,7 +444,7 @@ class ACT(nn.Module):
             # 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),
+                (batch_size, 2 if self.config.robot_state_feature else 1),
                 False,
                 device=batch["observation.state"].device,
             )
@@ -463,16 +477,16 @@ class ACT(nn.Module):
         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:
+        if self.config.robot_state_feature:
             encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch["observation.state"]))
         # Environment state token.
-        if self.use_env_state:
+        if self.config.env_state_feature:
             encoder_in_tokens.append(
                 self.encoder_env_state_input_proj(batch["observation.environment_state"])
             )
 
         # Camera observation features and positional embeddings.
-        if self.use_images:
+        if self.config.image_features:
             all_cam_features = []
             all_cam_pos_embeds = []
 

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -16,9 +16,15 @@
 # limitations under the License.
 from dataclasses import dataclass, field
 
+from lerobot.common.optim.optimizers import AdamConfig
+from lerobot.common.optim.schedulers import DiffuserSchedulerConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
 
+
+@PreTrainedConfig.register_subclass("diffusion")
 @dataclass
-class DiffusionConfig:
+class DiffusionConfig(PreTrainedConfig):
     """Configuration class for DiffusionPolicy.
 
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
@@ -62,7 +68,7 @@ class DiffusionConfig:
             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.
+        pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
             The group sizes are set to be about 16 (to be precise, feature_dim // 16).
@@ -93,7 +99,7 @@ class DiffusionConfig:
         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
+            `LeRobotDataset` and `load_previous_and_future_frames` for more information. Note, this defaults
             to False as the original Diffusion Policy implementation does the same.
     """
 
@@ -102,26 +108,17 @@ class DiffusionConfig:
     horizon: int = 16
     n_action_steps: int = 8
 
-    input_shapes: dict[str, list[int]] = field(
+    normalization_mapping: dict[str, NormalizationMode] = field(
         default_factory=lambda: {
-            "observation.image": [3, 96, 96],
-            "observation.state": [2],
-        }
-    )
-    output_shapes: dict[str, list[int]] = field(
-        default_factory=lambda: {
-            "action": [2],
+            "VISUAL": NormalizationMode.MEAN_STD,
+            "STATE": NormalizationMode.MIN_MAX,
+            "ACTION": NormalizationMode.MIN_MAX,
         }
     )
 
-    # 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"})
+    # The original implementation doesn't sample frames for the last 7 steps,
+    # which avoids excessive padding and leads to improved training results.
+    drop_n_last_frames: int = 7  # horizon - n_action_steps - n_obs_steps + 1
 
     # Architecture / modeling.
     # Vision backbone.
@@ -154,39 +151,23 @@ class DiffusionConfig:
     # Loss computation
     do_mask_loss_for_padding: bool = False
 
+    # Training presets
+    optimizer_lr: float = 1e-4
+    optimizer_betas: tuple = (0.95, 0.999)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-6
+    scheduler_name: str = "cosine"
+    scheduler_warmup_steps: int = 500
+
     def __post_init__(self):
+        super().__post_init__()
+
         """Input validation (not exhaustive)."""
         if not self.vision_backbone.startswith("resnet"):
             raise ValueError(
                 f"`vision_backbone` must be one of the ResNet variants. Got {self.vision_backbone}."
             )
 
-        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(
@@ -207,3 +188,50 @@ class DiffusionConfig:
                 "The horizon should be an integer multiple of the downsampling factor (which is determined "
                 f"by `len(down_dims)`). Got {self.horizon=} and {self.down_dims=}"
             )
+
+    def get_optimizer_preset(self) -> AdamConfig:
+        return AdamConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+        )
+
+    def get_scheduler_preset(self) -> DiffuserSchedulerConfig:
+        return DiffuserSchedulerConfig(
+            name=self.scheduler_name,
+            num_warmup_steps=self.scheduler_warmup_steps,
+        )
+
+    def validate_features(self) -> None:
+        if len(self.image_features) == 0 and self.env_state_feature is None:
+            raise ValueError("You must provide at least one image or the environment state among the inputs.")
+
+        if self.crop_shape is not None:
+            for key, image_ft in self.image_features.items():
+                if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
+                    raise ValueError(
+                        f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
+                        f"for `crop_shape` and {image_ft.shape} for "
+                        f"`{key}`."
+                    )
+
+        # Check that all input images have the same shape.
+        first_image_key, first_image_ft = next(iter(self.image_features.items()))
+        for key, image_ft in self.image_features.items():
+            if image_ft.shape != first_image_ft.shape:
+                raise ValueError(
+                    f"`{key}` does not match `{first_image_key}`, but we expect all image shapes to match."
+                )
+
+    @property
+    def observation_delta_indices(self) -> list:
+        return list(range(1 - self.n_obs_steps, 1))
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(1 - self.n_obs_steps, 1 - self.n_obs_steps + self.horizon))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -31,35 +31,32 @@ import torch.nn.functional as F  # noqa: N812
 import torchvision
 from diffusers.schedulers.scheduling_ddim import DDIMScheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
-from huggingface_hub import PyTorchModelHubMixin
 from torch import Tensor, nn
 
+from lerobot.common.constants import OBS_ENV_STATE, OBS_STATE
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.utils import (
     get_device_from_parameters,
     get_dtype_from_parameters,
+    get_output_shape,
     populate_queues,
 )
 
 
-class DiffusionPolicy(
-    nn.Module,
-    PyTorchModelHubMixin,
-    library_name="lerobot",
-    repo_url="https://github.com/huggingface/lerobot",
-    tags=["robotics", "diffusion-policy"],
-):
+class DiffusionPolicy(PreTrainedPolicy):
     """
     Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
     (paper: https://arxiv.org/abs/2303.04137, code: https://github.com/real-stanford/diffusion_policy).
     """
 
+    config_class = DiffusionConfig
     name = "diffusion"
 
     def __init__(
         self,
-        config: DiffusionConfig | None = None,
+        config: DiffusionConfig,
         dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
@@ -69,18 +66,16 @@ class DiffusionPolicy(
             dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
-        super().__init__()
-        if config is None:
-            config = DiffusionConfig()
+        super().__init__(config)
+        config.validate_features()
         self.config = config
-        self.normalize_inputs = Normalize(
-            config.input_shapes, config.input_normalization_modes, dataset_stats
-        )
+
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
         self.normalize_targets = Normalize(
-            config.output_shapes, config.output_normalization_modes, dataset_stats
+            config.output_features, config.normalization_mapping, dataset_stats
         )
         self.unnormalize_outputs = Unnormalize(
-            config.output_shapes, config.output_normalization_modes, dataset_stats
+            config.output_features, config.normalization_mapping, dataset_stats
         )
 
         # queues are populated during rollout of the policy, they contain the n latest observations and actions
@@ -88,20 +83,20 @@ class DiffusionPolicy(
 
         self.diffusion = DiffusionModel(config)
 
-        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
-
         self.reset()
 
+    def get_optim_params(self) -> dict:
+        return self.diffusion.parameters()
+
     def reset(self):
         """Clear observation and action queues. Should be called on `env.reset()`"""
         self._queues = {
             "observation.state": deque(maxlen=self.config.n_obs_steps),
             "action": deque(maxlen=self.config.n_action_steps),
         }
-        if len(self.expected_image_keys) > 0:
+        if self.config.image_features:
             self._queues["observation.images"] = deque(maxlen=self.config.n_obs_steps)
-        if self.use_env_state:
+        if self.config.env_state_feature:
             self._queues["observation.environment_state"] = deque(maxlen=self.config.n_obs_steps)
 
     @torch.no_grad
@@ -127,9 +122,11 @@ class DiffusionPolicy(
         actually measured from the first observation which (if `n_obs_steps` > 1) happened in the past.
         """
         batch = self.normalize_inputs(batch)
-        if len(self.expected_image_keys) > 0:
+        if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+            batch["observation.images"] = torch.stack(
+                [batch[key] for key in self.config.image_features], dim=-4
+            )
         # Note: It's important that this happens after stacking the images into a single key.
         self._queues = populate_queues(self._queues, batch)
 
@@ -146,15 +143,18 @@ class DiffusionPolicy(
         action = self._queues["action"].popleft()
         return action
 
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, None]:
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
-        if len(self.expected_image_keys) > 0:
+        if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
+            batch["observation.images"] = torch.stack(
+                [batch[key] for key in self.config.image_features], dim=-4
+            )
         batch = self.normalize_targets(batch)
         loss = self.diffusion.compute_loss(batch)
-        return {"loss": loss}
+        # no output_dict so returning None
+        return loss, None
 
 
 def _make_noise_scheduler(name: str, **kwargs: dict) -> DDPMScheduler | DDIMScheduler:
@@ -176,12 +176,9 @@ class DiffusionModel(nn.Module):
         self.config = config
 
         # 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
+        global_cond_dim = self.config.robot_state_feature.shape[0]
+        if self.config.image_features:
+            num_images = len(self.config.image_features)
             if self.config.use_separate_rgb_encoder_per_camera:
                 encoders = [DiffusionRgbEncoder(config) for _ in range(num_images)]
                 self.rgb_encoder = nn.ModuleList(encoders)
@@ -189,9 +186,8 @@ class DiffusionModel(nn.Module):
             else:
                 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]
+        if self.config.env_state_feature:
+            global_cond_dim += self.config.env_state_feature.shape[0]
 
         self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
 
@@ -220,7 +216,7 @@ class DiffusionModel(nn.Module):
 
         # Sample prior.
         sample = torch.randn(
-            size=(batch_size, self.config.horizon, self.config.output_shapes["action"][0]),
+            size=(batch_size, self.config.horizon, self.config.action_feature.shape[0]),
             dtype=dtype,
             device=device,
             generator=generator,
@@ -242,10 +238,10 @@ class DiffusionModel(nn.Module):
 
     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"]]
+        batch_size, n_obs_steps = batch[OBS_STATE].shape[:2]
+        global_cond_feats = [batch[OBS_STATE]]
         # Extract image features.
-        if self._use_images:
+        if self.config.image_features:
             if self.config.use_separate_rgb_encoder_per_camera:
                 # Combine batch and sequence dims while rearranging to make the camera index dimension first.
                 images_per_camera = einops.rearrange(batch["observation.images"], "b s n ... -> n (b s) ...")
@@ -272,8 +268,8 @@ class DiffusionModel(nn.Module):
                 )
             global_cond_feats.append(img_features)
 
-        if self._use_env_state:
-            global_cond_feats.append(batch["observation.environment_state"])
+        if self.config.env_state_feature:
+            global_cond_feats.append(batch[OBS_ENV_STATE])
 
         # Concatenate features then flatten to (B, global_cond_dim).
         return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)
@@ -443,7 +439,7 @@ class SpatialSoftmax(nn.Module):
 
 
 class DiffusionRgbEncoder(nn.Module):
-    """Encoder an RGB image into a 1D feature vector.
+    """Encodes an RGB image into a 1D feature vector.
 
     Includes the ability to normalize and crop the image first.
     """
@@ -482,19 +478,16 @@ class DiffusionRgbEncoder(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
+        # The dummy input should take the number of image channels from `config.image_features` and it should
         # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
-        # height and width from `config.input_shapes`.
-        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        # height and width from `config.image_features`.
+
         # Note: we have a check in the config class to make sure all images have the same shape.
-        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:])
+        images_shape = next(iter(config.image_features.values())).shape
+        dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
+        dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
+        feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
+
         self.pool = SpatialSoftmax(feature_map_shape, num_kp=config.spatial_softmax_num_keypoints)
         self.feature_dim = config.spatial_softmax_num_keypoints * 2
         self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
@@ -611,7 +604,7 @@ class DiffusionConditionalUnet1d(nn.Module):
 
         # In channels / out channels for each downsampling block in the Unet's encoder. For the decoder, we
         # just reverse these.
-        in_out = [(config.output_shapes["action"][0], config.down_dims[0])] + list(
+        in_out = [(config.action_feature.shape[0], config.down_dims[0])] + list(
             zip(config.down_dims[:-1], config.down_dims[1:], strict=True)
         )
 
@@ -666,7 +659,7 @@ class DiffusionConditionalUnet1d(nn.Module):
 
         self.final_conv = nn.Sequential(
             DiffusionConv1dBlock(config.down_dims[0], config.down_dims[0], kernel_size=config.kernel_size),
-            nn.Conv1d(config.down_dims[0], config.output_shapes["action"][0], 1),
+            nn.Conv1d(config.down_dims[0], config.action_feature.shape[0], 1),
         )
 
     def forward(self, x: Tensor, timestep: Tensor | int, global_cond=None) -> Tensor: