added tdmpc2 to policy factory; shape fixes in tdmpc2

entire policy

.github/PULL_REQUEST_TEMPLATE.md

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

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

@@ -8,8 +8,6 @@ on:
   schedule:
     - cron: "0 1 * * *"
 
-permissions: {}
-
 env:
   PYTHON_VERSION: "3.10"
 
@@ -27,14 +25,11 @@ 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
@@ -65,14 +60,11 @@ 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
@@ -97,13 +89,9 @@ 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

@@ -7,10 +7,10 @@ on:
   schedule:
     - cron: "0 2 * * *"
 
-permissions: {}
-
-# env:
+env:
+  DATA_DIR: tests/data
   # SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
+
 jobs:
   run_all_tests_cpu:
     name: CPU
@@ -30,9 +30,13 @@ 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

@@ -4,12 +4,12 @@ on:
   workflow_dispatch:
   workflow_call:
   pull_request:
+    branches:
+      - main
   push:
     branches:
       - main
 
-permissions: {}
-
 env:
   PYTHON_VERSION: "3.10"
 
@@ -19,9 +19,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout Repository
-        uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+        uses: actions/checkout@v3
 
       - name: Set up Python
         uses: actions/setup-python@v4
@@ -38,7 +36,49 @@ jobs:
         run: python -m pip install "ruff==${{ env.RUFF_VERSION }}"
 
       - name: Ruff check
-        run: ruff check --output-format=github
+        run: ruff check
 
       - name: Ruff format
         run: ruff format --diff
+
+
+  poetry_check:
+    name: Poetry check
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Install poetry
+        run: pipx install poetry
+
+      - name: Poetry check
+        run: poetry check
+
+
+  poetry_relax:
+    name: Poetry relax
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Install poetry
+        run: pipx install poetry
+
+      - name: Install poetry-relax
+        run: poetry self add poetry-relax
+
+      - name: Poetry relax
+        id: poetry_relax
+        run: |
+          output=$(poetry relax --check 2>&1)
+          if echo "$output" | grep -q "Proposing updates"; then
+            echo "$output"
+            echo ""
+            echo "Some dependencies have caret '^' version requirement added by poetry by default."
+            echo "Please replace them with '>='. You can do this by hand or use poetry-relax to do this."
+            exit 1
+          else
+            echo "$output"
+          fi

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

@@ -4,12 +4,12 @@ name: Test Dockerfiles
 
 on:
   pull_request:
+    branches:
+      - main
     paths:
       # Run only when DockerFile files are modified
       - "docker/**"
 
-permissions: {}
-
 env:
   PYTHON_VERSION: "3.10"
 
@@ -22,8 +22,6 @@ jobs:
     steps:
       - name: Check out code
         uses: actions/checkout@v4
-        with:
-          persist-credentials: false
 
       - name: Get changed files
         id: changed-files
@@ -32,12 +30,15 @@ jobs:
           files: docker/**
           json: "true"
 
-      - name: Run step if only the files listed above change  # zizmor: ignore[template-injection]
+      - name: Run step if only the files listed above change
         if: steps.changed-files.outputs.any_changed == 'true'
         id: set-matrix
+        env:
+          ALL_CHANGED_FILES: ${{ steps.changed-files.outputs.all_changed_files }}
         run: |
           echo "matrix=${{ steps.changed-files.outputs.all_changed_files}}" >> $GITHUB_OUTPUT
 
+
   build_modified_dockerfiles:
     name: Build modified Docker images
     needs: get_changed_files
@@ -51,13 +52,9 @@ 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

@@ -2,13 +2,14 @@ name: Tests
 
 on:
   pull_request:
+    branches:
+      - main
     paths:
       - "lerobot/**"
       - "tests/**"
       - "examples/**"
       - ".github/**"
-      - "pyproject.toml"
-      - ".pre-commit-config.yaml"
+      - "poetry.lock"
       - "Makefile"
       - ".cache/**"
   push:
@@ -19,27 +20,21 @@ on:
       - "tests/**"
       - "examples/**"
       - ".github/**"
-      - "pyproject.toml"
-      - ".pre-commit-config.yaml"
+      - "poetry.lock"
       - "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
@@ -47,19 +42,25 @@ jobs:
           sudo apt-get update && \
           sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
 
-      - name: Install uv and python
-        uses: astral-sh/setup-uv@v5
-        with:
-          enable-cache: true
-          version: ${{ env.UV_VERSION }}
-          python-version: "3.10"
+      - name: Install poetry
+        run: |
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
 
-      - name: Install lerobot (all extras)
-        run: uv sync --all-extras
+      # 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
+        with:
+          python-version: "3.10"
+          cache: "poetry"
+
+      - name: Install poetry dependencies
+        run: |
+          poetry install --all-extras
 
       - name: Test with pytest
         run: |
-          uv run pytest tests -v --cov=./lerobot --durations=0 \
+          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 \
@@ -69,44 +70,50 @@ 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 uv and python
-        uses: astral-sh/setup-uv@v5
+      - 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
         with:
-          enable-cache: true
-          version: ${{ env.UV_VERSION }}
           python-version: "3.10"
 
-      - name: Install lerobot
-        run: uv sync --extra "test"
+      - name: Install poetry dependencies
+        run: |
+          poetry install --extras "test"
 
       - name: Test with pytest
         run: |
-          uv run pytest tests -v --cov=./lerobot --durations=0 \
+          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
@@ -114,21 +121,20 @@ jobs:
           sudo apt-get update && \
           sudo apt-get install -y libegl1-mesa-dev portaudio19-dev
 
-      - name: Install uv and python
-        uses: astral-sh/setup-uv@v5
+      - name: Install poetry
+        run: |
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+
+      - name: Set up Python 3.10
+        uses: actions/setup-python@v5
         with:
-          enable-cache: true
-          version: ${{ env.UV_VERSION }}
           python-version: "3.10"
+          cache: "poetry"
 
-      - name: Install lerobot (all extras)
+      - name: Install poetry dependencies
         run: |
-          uv venv
-          uv sync --all-extras
-
-      - name: venv
-        run: |
-          echo "PYTHON_PATH=${{ github.workspace }}/.venv/bin/python" >> $GITHUB_ENV
+          poetry install --all-extras
 
       - name: Test end-to-end
         run: |

.github/workflows/trufflehog.yml

@@ -3,7 +3,8 @@ on:
 
 name: Secret Leaks
 
-permissions: {}
+permissions:
+  contents: read
 
 jobs:
   trufflehog:
@@ -13,8 +14,6 @@ jobs:
       uses: actions/checkout@v4
       with:
         fetch-depth: 0
-        persist-credentials: false
-
     - name: Secret Scanning
       uses: trufflesecurity/trufflehog@main
       with:

.gitignore

@@ -49,10 +49,6 @@ 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

@@ -3,7 +3,7 @@ default_language_version:
     python: python3.10
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+    rev: v4.6.0
     hooks:
       - id: check-added-large-files
       - id: debug-statements
@@ -14,20 +14,24 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.19.1
+    rev: v3.16.0
     hooks:
     -   id: pyupgrade
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.9.6
+    rev: v0.5.2
     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"
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.23.3
+    rev: v8.18.4
     hooks:
       - id: gitleaks
-  - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.3.1
-    hooks:
-      - id: zizmor

CONTRIBUTING.md

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

Makefile

@@ -2,10 +2,10 @@
 
 PYTHON_PATH := $(shell which python)
 
-# If uv is installed and a virtual environment exists, use it
-UV_CHECK := $(shell command -v uv)
-ifneq ($(UV_CHECK),)
-	PYTHON_PATH := $(shell .venv/bin/python)
+# If Poetry is installed, redefine PYTHON_PATH to use the Poetry-managed Python
+POETRY_CHECK := $(shell command -v poetry)
+ifneq ($(POETRY_CHECK),)
+	PYTHON_PATH := $(shell poetry run which python)
 endif
 
 export PATH := $(dir $(PYTHON_PATH)):$(PATH)
@@ -20,109 +20,171 @@ 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.type=act \
-		--policy.dim_model=64 \
-		--policy.n_action_steps=20 \
-		--policy.chunk_size=20 \
-		--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 \
-		--device=$(DEVICE) \
-		--output_dir=tests/outputs/act/
-
-test-act-ete-train-resume:
-	python lerobot/scripts/train.py \
-		--config_path=tests/outputs/act/checkpoints/000002/pretrained_model/train_config.json \
-		--resume=true
+		policy=act \
+		policy.dim_model=64 \
+		env=aloha \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		policy.n_action_steps=20 \
+		policy.chunk_size=20 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/act/
 
 test-act-ete-eval:
 	python lerobot/scripts/eval.py \
-		--policy.path=tests/outputs/act/checkpoints/000004/pretrained_model \
-		--env.type=aloha \
-		--env.episode_length=5 \
-		--eval.n_episodes=1 \
-		--eval.batch_size=1 \
-		--device=$(DEVICE)
+		-p tests/outputs/act/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+
+test-act-ete-train-amp:
+	python lerobot/scripts/train.py \
+		policy=act \
+		policy.dim_model=64 \
+		env=aloha \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		policy.n_action_steps=20 \
+		policy.chunk_size=20 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/act_amp/ \
+		training.image_transforms.enable=true \
+		use_amp=true
+
+test-act-ete-eval-amp:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/act_amp/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+		use_amp=true
 
 test-diffusion-ete-train:
 	python lerobot/scripts/train.py \
-		--policy.type=diffusion \
-		--policy.down_dims='[64,128,256]' \
-		--policy.diffusion_step_embed_dim=32 \
-		--policy.num_inference_steps=10 \
-		--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 \
-		--device=$(DEVICE) \
-		--output_dir=tests/outputs/diffusion/
+		policy=diffusion \
+		policy.down_dims=\[64,128,256\] \
+		policy.diffusion_step_embed_dim=32 \
+		policy.num_inference_steps=10 \
+		env=pusht \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/diffusion/
 
 test-diffusion-ete-eval:
 	python lerobot/scripts/eval.py \
-		--policy.path=tests/outputs/diffusion/checkpoints/000002/pretrained_model \
-		--env.type=pusht \
-		--env.episode_length=5 \
-		--eval.n_episodes=1 \
-		--eval.batch_size=1 \
-		--device=$(DEVICE)
+		-p tests/outputs/diffusion/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
 
 test-tdmpc-ete-train:
 	python lerobot/scripts/train.py \
-		--policy.type=tdmpc \
-		--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 \
-		--device=$(DEVICE) \
-		--output_dir=tests/outputs/tdmpc/
+		policy=tdmpc \
+		env=xarm \
+		env.task=XarmLift-v0 \
+		dataset_repo_id=lerobot/xarm_lift_medium \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=2 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/tdmpc/
+
+test-tdmpc-ete-train-with-online:
+	python lerobot/scripts/train.py \
+		env=pusht \
+		env.gym.obs_type=environment_state_agent_pos \
+		policy=tdmpc_pusht_keypoints \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=10 \
+		device=$(DEVICE) \
+		training.offline_steps=2 \
+		training.online_steps=20 \
+		training.save_checkpoint=false \
+		training.save_freq=10 \
+		training.batch_size=2 \
+		training.online_rollout_n_episodes=2 \
+		training.online_rollout_batch_size=2 \
+		training.online_steps_between_rollouts=10 \
+		training.online_buffer_capacity=15 \
+		eval.use_async_envs=true \
+		hydra.run.dir=tests/outputs/tdmpc_online/
+
 
 test-tdmpc-ete-eval:
 	python lerobot/scripts/eval.py \
-		--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
-		--env.type=xarm \
-		--env.episode_length=5 \
-		--env.task=XarmLift-v0 \
-		--eval.n_episodes=1 \
-		--eval.batch_size=1 \
-		--device=$(DEVICE)
+		-p tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+
+test-default-ete-eval:
+	python lerobot/scripts/eval.py \
+		--config lerobot/configs/default.yaml \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+
+test-act-pusht-tutorial:
+	cp examples/advanced/1_train_act_pusht/act_pusht.yaml lerobot/configs/policy/created_by_Makefile.yaml
+	python lerobot/scripts/train.py \
+		policy=created_by_Makefile.yaml \
+		env=pusht \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=2 \
+		device=$(DEVICE) \
+		training.save_model=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/act_pusht/
+	rm lerobot/configs/policy/created_by_Makefile.yaml

README.md

@@ -68,7 +68,7 @@
 
 ### Acknowledgment
 
-- Thanks to Tony Zhao, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
+- Thanks to Tony Zaho, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
 - Thanks to Cheng Chi, Zhenjia Xu and colleagues for open sourcing Diffusion policy, Pusht environment and datasets, as well as UMI datasets. Ours are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu) and [UMI Gripper](https://umi-gripper.github.io).
 - Thanks to Nicklas Hansen, Yunhai Feng and colleagues for open sourcing TDMPC policy, Simxarm environments and datasets. Ours are adapted from [TDMPC](https://github.com/nicklashansen/tdmpc) and [FOWM](https://www.yunhaifeng.com/FOWM).
 - Thanks to Antonio Loquercio and Ashish Kumar for their early support.
@@ -122,7 +122,10 @@ 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 config classes with all options that you can override in the command line
+|   ├── configs          # contains hydra yaml files with all options that you can override in the command line
+|   |   ├── default.yaml   # selected by default, it loads pusht environment and diffusion policy
+|   |   ├── env            # various sim environments and their datasets: aloha.yaml, pusht.yaml, xarm.yaml
+|   |   └── policy         # various policies: act.yaml, diffusion.yaml, tdmpc.yaml
 |   ├── common           # contains classes and utilities
 |   |   ├── datasets       # various datasets of human demonstrations: aloha, pusht, xarm
 |   |   ├── envs           # various sim environments: aloha, pusht, xarm
@@ -150,12 +153,10 @@ python lerobot/scripts/visualize_dataset.py \
     --episode-index 0
 ```
 
-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`)
+or from a dataset in a local folder with the root `DATA_DIR` environment variable (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
 ```bash
-python lerobot/scripts/visualize_dataset.py \
+DATA_DIR='./my_local_data_dir' python lerobot/scripts/visualize_dataset.py \
     --repo-id lerobot/pusht \
-    --root ./my_local_data_dir \
-    --local-files-only 1 \
     --episode-index 0
 ```
 
@@ -207,10 +208,12 @@ 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
-- metadata are stored in plain json/jsonl files
+- videos are stored in mp4 format to save space or png files
+- episode_data_index saved using `safetensor` tensor serialization format
+- stats saved using `safetensor` tensor serialization format
+- info are saved using JSON
 
-Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can use the `local_files_only` argument and specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
+Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can set the `DATA_DIR` environment variable to your root dataset folder as illustrated in the above section on dataset visualization.
 
 ### Evaluate a pretrained policy
 
@@ -219,48 +222,87 @@ 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 \
-    --policy.path=lerobot/diffusion_pusht \
-    --env.type=pusht \
-    --eval.batch_size=10 \
-    --eval.n_episodes=10 \
-    --use_amp=false \
-    --device=cuda
+    -p lerobot/diffusion_pusht \
+    eval.n_episodes=10 \
+    eval.batch_size=10
 ```
 
 Note: After training your own policy, you can re-evaluate the checkpoints with:
 
 ```bash
-python lerobot/scripts/eval.py --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
+python lerobot/scripts/eval.py -p {OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```
 
 See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-Check out [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.
+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.
 
-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`.
+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:
 
-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.
+```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.
 
 ![](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 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:
+We have organized our configuration files (found under [`lerobot/configs`](./lerobot/configs)) such that they reproduce SOTA results from a given model variant in their respective original works. Simply running:
+
 ```bash
-python lerobot/scripts/train.py --config_path=lerobot/diffusion_pusht
+python lerobot/scripts/train.py policy=diffusion env=pusht
 ```
+
 reproduces SOTA results for Diffusion Policy on the PushT task.
 
+Pretrained policies, along with reproduction details, can be found under the "Models" section of https://huggingface.co/lerobot.
+
 ## Contribute
 
 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
@@ -278,7 +320,7 @@ 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
@@ -288,7 +330,7 @@ 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.
-- `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.
+- `config.yaml`: A consolidated Hydra training configuration containing the policy, environment, and dataset configs. The policy configuration should match `config.json` exactly. The environment config is useful for anyone who wants to evaluate your policy. The dataset config just serves as a paper trail for reproducibility.
 
 To upload these to the hub, run the following:
 ```bash

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 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).
+We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an appartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
 For these reasons, we run this benchmark on four representative datasets:
 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
 - `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
@@ -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 apart, we also have the following scenario:
+Additionally, because some policies might request single timestamps that are a few frames appart, we also have the following scenario:
 - `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
 
 However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
@@ -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 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:
+One aspect that can't be measured here with those metrics is the compatibility of the encoding accross platforms, in particular on web browser, for visualization purposes.
+h264, h265 and AV1 are all commonly used codecs and should not be pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
 - `yuv420p` is more widely supported across various platforms, including web browsers.
 - `yuv444p` offers higher color fidelity but might not be supported as broadly.
 
@@ -116,7 +116,7 @@ Additional encoding parameters exist that are not included in this benchmark. In
 - `-preset` which allows for selecting encoding presets. This represents a collection of options that will provide a certain encoding speed to compression ratio. By leaving this parameter unspecified, it is considered to be `medium` for libx264 and libx265 and `8` for libsvtav1.
 - `-tune` which allows to optimize the encoding for certains aspects (e.g. film quality, fast decoding, etc.).
 
-See the documentation mentioned above for more detailed info on these settings and for a more comprehensive list of other parameters.
+See the documentation mentioned above for more detailled info on these settings and for a more comprehensive list of other parameters.
 
 Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
 - `torchaudio`

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.meta.camera_keys[0]].shape[-2:])
+    width, height = tuple(dataset[0][dataset.camera_keys[0]].shape[-2:])
     num_pixels = width * height
     video_size_bytes = video_path.stat().st_size
     images_size_bytes = get_directory_size(imgs_dir)

docker/lerobot-cpu/Dockerfile

@@ -7,9 +7,9 @@ ARG DEBIAN_FRONTEND=noninteractive
 
 # Install apt dependencies
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential cmake git git-lfs \
+    build-essential cmake \
     libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
-    speech-dispatcher libgeos-dev \
+    speech-dispatcher \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
 # Create virtual environment
@@ -19,8 +19,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 
 # Install LeRobot
-RUN git lfs install
-RUN git clone https://github.com/huggingface/lerobot.git /lerobot
+COPY . /lerobot
 WORKDIR /lerobot
 RUN pip install --upgrade --no-cache-dir pip
 RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]" \

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 portaudio19-dev libgeos-dev \
+    speech-dispatcher \
     python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 

docker/lerobot-gpu/Dockerfile

@@ -7,9 +7,9 @@ ARG DEBIAN_FRONTEND=noninteractive
 
 # Install apt dependencies
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential cmake git git-lfs \
+    build-essential cmake \
     libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
-    speech-dispatcher libgeos-dev \
+    speech-dispatcher \
     python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
@@ -21,8 +21,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 
 # Install LeRobot
-RUN git lfs install
-RUN git clone https://github.com/huggingface/lerobot.git /lerobot
+COPY . /lerobot
 WORKDIR /lerobot
 RUN pip install --upgrade --no-cache-dir pip
 RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]"

examples/10_use_so100.md

@@ -1,99 +1,58 @@
-# Using the [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot
+This tutorial explains how to use [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot.
 
-## Table of Contents
+## Source the parts
 
-  - [A. Source the parts](#a-source-the-parts)
-  - [B. Install LeRobot](#b-install-lerobot)
-  - [C. Configure the motors](#c-configure-the-motors)
-  - [D. Assemble the arms](#d-assemble-the-arms)
-  - [E. Calibrate](#e-calibrate)
-  - [F. Teleoperate](#f-teleoperate)
-  - [G. Record a dataset](#g-record-a-dataset)
-  - [H. Visualize a dataset](#h-visualize-a-dataset)
-  - [I. Replay an episode](#i-replay-an-episode)
-  - [J. Train a policy](#j-train-a-policy)
-  - [K. Evaluate your policy](#k-evaluate-your-policy)
-  - [L. More Information](#l-more-information)
+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.
 
-## A. Source the parts
+**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.
 
-Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts,
-and advice if it's your first time printing or if you don't own a 3D printer.
-
-Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
-
-## B. Install LeRobot
-
-> [!TIP]
-> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
+## Install LeRobot
 
 On your computer:
 
-#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
-
-#### 2. Restart shell
-Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
-
-#### 3. Create and activate a fresh conda environment for lerobot
-
-<details>
-<summary><strong>Video install instructions</strong></summary>
-
-<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
-
-</details>
-
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
 ```bash
-conda create -y -n lerobot python=3.10
+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
 ```
 
-Then activate your conda environment (do this each time you open a shell to use lerobot!):
+2. Restart shell or `source ~/.bashrc`
+
+3. Create and activate a fresh conda environment for lerobot
 ```bash
-conda activate lerobot
+conda create -y -n lerobot python=3.10 && conda activate lerobot
 ```
 
-#### 4. Clone LeRobot:
+4. Clone LeRobot:
 ```bash
 git clone https://github.com/huggingface/lerobot.git ~/lerobot
 ```
 
-#### 5. Install LeRobot with dependencies for the feetech motors:
+5. Install LeRobot with dependencies for the feetech motors:
 ```bash
 cd ~/lerobot && pip install -e ".[feetech]"
 ```
 
-*EXTRA: For Linux only (not Mac)*: install extra dependencies for recording datasets:
+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"
 ```
-Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms :robot:.
-Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
-## C. Configure the motors
 
-> [!NOTE]
-> Throughout this tutorial you will find videos on how to do the steps, the full video tutorial can be found here: [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I).
+## Configure the motors
 
-### 1. Find the USB ports associated to each arm
+Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the use of our scripts below.
 
-Designate one bus servo adapter and 6 motors for your leader arm, and similarly the other bus servo adapter and 6 motors for the follower arm. It's convenient to label them and write on each motor if it's for the follower `F` or for the leader `L` and it's ID from 1 to 6 (F1...F6 and L1...L6).
-
-#### a. Run the script to find port
-
-<details>
-<summary><strong>Video finding port</strong></summary>
-  <video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
-  <video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
-</details>
-
-To find the port for each bus servo adapter, run the utility script:
+**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
 ```
 
-#### b. Example outputs
-
 Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
 ```
 Finding all available ports for the MotorBus.
@@ -105,6 +64,7 @@ Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
 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.
@@ -117,73 +77,14 @@ The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
 Reconnect the usb cable.
 ```
 
-#### c. Troubleshooting
-On Linux, you might need to give access to the USB ports by running:
+Troubleshooting: On Linux, you might need to give access to the USB ports by running:
 ```bash
 sudo chmod 666 /dev/ttyACM0
 sudo chmod 666 /dev/ttyACM1
 ```
 
-#### d. Update config file
-
-IMPORTANTLY: Now that you have your ports, update the **port** default values of [`SO100RobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
-```python
-@RobotConfig.register_subclass("so100")
-@dataclass
-class So100RobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/so100"
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-```
-
-### 2. Assembling the Base
-Let's begin with assembling the follower arm base
-
-#### a. Set IDs for all 12 motors
-
-<details>
-<summary><strong>Video configuring motor</strong></summary>
-  <video src="https://github.com/user-attachments/assets/ef9b3317-2e11-4858-b9d3-f0a02fb48ecf"></video>
-  <video src="https://github.com/user-attachments/assets/f36b5ed5-c803-4ebe-8947-b39278776a0d"></video>
-</details>
-
-Plug your first motor F1 and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate. Replace the text after --port to the corresponding follower control board port and run this command in cmd:
+**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 \
@@ -193,8 +94,7 @@ python lerobot/scripts/configure_motor.py \
   --ID 1
 ```
 
-> [!NOTE]
-> These motors are currently limited. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
+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
@@ -208,50 +108,23 @@ python lerobot/scripts/configure_motor.py \
 
 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.
 
-#### b. Remove the gears of the 6 leader motors
-
-<details>
-<summary><strong>Video removing gears</strong></summary>
-
-<video src="https://github.com/user-attachments/assets/0c95b88c-5b85-413d-ba19-aee2f864f2a7"></video>
-
-</details>
-
-
-Follow the video for removing gears. You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
-
-#### c. Add motor horn to all 12 motors
-
-<details>
-<summary><strong>Video adding motor horn</strong></summary>
-
-<video src="https://github.com/user-attachments/assets/ef3391a4-ad05-4100-b2bd-1699bf86c969"></video>
-
-</details>
-
-Follow the video for adding the motor horn. 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.
+**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.
 
-## D. Assemble the arms
+## Assemble the arms
 
-<details>
-<summary><strong>Video assembling arms</strong></summary>
+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.
 
-<video src="https://github.com/user-attachments/assets/488a39de-0189-4461-9de3-05b015f90cca"></video>
-
-</details>
-
-Follow the video for assembling the arms. It is important to insert the cables into the motor that is being assembled before you assemble the motor into the arm! Inserting the cables beforehand is much easier than doing this afterward. The first arm should take a bit more than 1 hour to assemble, but once you get used to it, you can do it under 1 hour for the second arm.
-
-## E. Calibrate
+## Calibrate
 
 Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one SO-100 robot to work on another.
 
-#### a. Manual calibration of follower arm
-
-> [!IMPORTANT]
-> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
+**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:
 
@@ -261,15 +134,13 @@ You will need to move the follower arm to these positions sequentially:
 
 Make sure both arms are connected and run this script to launch manual calibration:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_follower"]'
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --robot-overrides '~cameras' --arms main_follower
 ```
 
-#### b. Manual calibration of leader arm
-Follow step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
+**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 |
 |---|---|---|
@@ -277,34 +148,31 @@ Follow step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which
 
 Run this script to launch manual calibration:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_leader"]'
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --robot-overrides '~cameras' --arms main_leader
 ```
 
-## F. Teleoperate
+## Teleoperate
 
 **Simple teleop**
 Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --robot.cameras='{}' \
-  --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --robot-overrides '~cameras' \
+    --display-cameras 0
 ```
 
 
-#### a. Teleop with displaying cameras
+**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 \
-  --robot.type=so100 \
-  --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/so100.yaml
 ```
 
-## G. Record a dataset
+## Record a dataset
 
 Once you're familiar with teleoperation, you can record your first dataset with SO-100.
 
@@ -321,99 +189,92 @@ echo $HF_USER
 
 Record 2 episodes and upload your dataset to the hub:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/so100_test \
-  --control.tags='["so100","tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.push_to_hub=true
+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
 ```
 
-Note: You can resume recording by adding `--control.resume=true`. Also if you didn't push your dataset yet, add `--control.local_files_only=true`.
+## Visualize a dataset
 
-## H. Visualize a dataset
-
-If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+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 `--control.push_to_hub=false`, you can also visualize it locally with:
+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 \
-  --repo-id ${HF_USER}/so100_test \
-  --local-files-only 1
+  --root data \
+  --repo-id ${HF_USER}/so100_test
 ```
 
-## I. Replay an episode
+## Replay an episode
 
 Now try to replay the first episode on your robot:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/so100_test \
-  --control.episode=0
+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
 ```
 
-Note: If you didn't push your dataset yet, add `--control.local_files_only=true`.
-
-## J. Train a policy
+## Train a policy
 
 To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 ```bash
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/so100_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_so100_test \
-  --job_name=act_so100_test \
-  --device=cuda \
-  --wandb.enable=true
+DATA_DIR=data python lerobot/scripts/train.py \
+  dataset_repo_id=${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
 ```
 
-Note: If you didn't push your dataset yet, add `--control.local_files_only=true`.
-
 Let's explain it:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so100_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `device=cuda` since we are training on a Nvidia GPU, but you could use `device=mps` to train on Apple silicon.
+1. We provided the dataset as argument with `dataset_repo_id=${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`.
 
-## K. Evaluate your policy
+## Evaluate your policy
 
 You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/eval_act_so100_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=10 \
-  --control.push_to_hub=true \
-  --control.policy.path=outputs/train/act_so100_test/checkpoints/last/pretrained_model
+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 `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so100_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so100_test`).
+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`).
 
-## L. More Information
+## 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.
 
-> [!TIP]
->  If you have any questions or need help, please reach out on Discord in the channel [`#so100-arm`](https://discord.com/channels/1216765309076115607/1237741463832363039).
+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/11_use_moss.md

@@ -83,54 +83,6 @@ sudo chmod 666 /dev/ttyACM0
 sudo chmod 666 /dev/ttyACM1
 ```
 
-#### Update config file
-
-IMPORTANTLY: Now that you have your ports, update the **port** default values of [`MossRobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
-```python
-@RobotConfig.register_subclass("moss")
-@dataclass
-class MossRobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/moss"
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-```
-
 **Configure your motors**
 Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
 ```bash
@@ -182,11 +134,9 @@ You will need to move the follower arm to these positions sequentially:
 
 Make sure both arms are connected and run this script to launch manual calibration:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_follower"]'
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --robot-overrides '~cameras' --arms main_follower
 ```
 
 **Manual calibration of leader arm**
@@ -198,11 +148,9 @@ Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMi
 
 Run this script to launch manual calibration:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_leader"]'
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --robot-overrides '~cameras' --arms main_leader
 ```
 
 ## Teleoperate
@@ -210,19 +158,18 @@ python lerobot/scripts/control_robot.py \
 **Simple teleop**
 Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --robot.cameras='{}' \
-  --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/moss.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 \
-  --robot.type=moss \
-  --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/moss.yaml
 ```
 
 ## Record a dataset
@@ -242,70 +189,66 @@ echo $HF_USER
 
 Record 2 episodes and upload your dataset to the hub:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/moss_test \
-  --control.tags='["moss","tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.push_to_hub=true
+python lerobot/scripts/control_robot.py record \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --fps 30 \
+    --root data \
+    --repo-id ${HF_USER}/moss_test \
+    --tags moss tutorial \
+    --warmup-time-s 5 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 2 \
+    --push-to-hub 1
 ```
 
-Note: You can resume recording by adding `--control.resume=true`. Also if you didn't push your dataset yet, add `--control.local_files_only=true`.
-
 ## Visualize a dataset
 
-If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+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}/moss_test
 ```
 
-If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with:
+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 \
-  --repo-id ${HF_USER}/moss_test \
-  --local-files-only 1
+  --root data \
+  --repo-id ${HF_USER}/moss_test
 ```
 
 ## Replay an episode
 
 Now try to replay the first episode on your robot:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/moss_test \
-  --control.episode=0
+DATA_DIR=data python lerobot/scripts/control_robot.py replay \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --fps 30 \
+    --root data \
+    --repo-id ${HF_USER}/moss_test \
+    --episode 0
 ```
 
-Note: If you didn't push your dataset yet, add `--control.local_files_only=true`.
-
 ## Train a policy
 
 To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
 ```bash
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/moss_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_moss_test \
-  --job_name=act_moss_test \
-  --device=cuda \
-  --wandb.enable=true
+DATA_DIR=data python lerobot/scripts/train.py \
+  dataset_repo_id=${HF_USER}/moss_test \
+  policy=act_moss_real \
+  env=moss_real \
+  hydra.run.dir=outputs/train/act_moss_test \
+  hydra.job.name=act_moss_test \
+  device=cuda \
+  wandb.enable=true
 ```
 
-Note: If you didn't push your dataset yet, add `--control.local_files_only=true`.
-
 Let's explain it:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/moss_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `device=cuda` since we are training on a Nvidia GPU, but you could use `device=mps` to train on Apple silicon.
+1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/moss_test`.
+2. We provided the policy with `policy=act_moss_real`. This loads configurations from [`lerobot/configs/policy/act_moss_real.yaml`](../lerobot/configs/policy/act_moss_real.yaml). Importantly, this policy uses 2 cameras as input `laptop`, `phone`.
+3. We provided an environment as argument with `env=moss_real`. This loads configurations from [`lerobot/configs/env/moss_real.yaml`](../lerobot/configs/env/moss_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_moss_test/checkpoints`.
 
@@ -313,24 +256,22 @@ Training should take several hours. You will find checkpoints in `outputs/train/
 
 You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/eval_act_moss_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=10 \
-  --control.push_to_hub=true \
-  --control.policy.path=outputs/train/act_moss_test/checkpoints/last/pretrained_model
+python lerobot/scripts/control_robot.py record \
+  --robot-path lerobot/configs/robot/moss.yaml \
+  --fps 30 \
+  --root data \
+  --repo-id ${HF_USER}/eval_act_moss_test \
+  --tags moss tutorial eval \
+  --warmup-time-s 5 \
+  --episode-time-s 40 \
+  --reset-time-s 10 \
+  --num-episodes 10 \
+  -p outputs/train/act_moss_test/checkpoints/last/pretrained_model
 ```
 
 As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_moss_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_moss_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_moss_test`).
+1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_moss_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_moss_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_moss_test`).
 
 ## More
 

examples/1_load_lerobot_dataset.py

@@ -3,120 +3,78 @@ This script demonstrates the use of `LeRobotDataset` class for handling and proc
 It illustrates how to load datasets, manipulate them, and apply transformations suitable for machine learning tasks in PyTorch.
 
 Features included in this script:
-- Viewing a dataset's metadata and exploring its properties.
-- Loading an existing dataset from the hub or a subset of it.
-- Accessing frames by episode number.
+- Loading a dataset and accessing its properties.
+- Filtering data by episode number.
+- Converting tensor data for visualization.
+- Saving video files from dataset frames.
 - Using advanced dataset features like timestamp-based frame selection.
 - Demonstrating compatibility with PyTorch DataLoader for batch processing.
 
 The script ends with examples of how to batch process data using PyTorch's DataLoader.
 """
 
+from pathlib import Path
 from pprint import pprint
 
+import imageio
 import torch
-from huggingface_hub import HfApi
 
 import lerobot
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
-# We ported a number of existing datasets ourselves, use this to see the list:
 print("List of available datasets:")
 pprint(lerobot.available_datasets)
 
-# You can also browse through the datasets created/ported by the community on the hub using the hub api:
-hub_api = HfApi()
-repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
-pprint(repo_ids)
+# Let's take one for this example
+repo_id = "lerobot/pusht"
 
-# 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:
+# You can easily load a dataset from a Hugging Face repository
 dataset = LeRobotDataset(repo_id)
-print(f"Number of episodes selected: {dataset.num_episodes}")
-print(f"Number of frames selected: {dataset.num_frames}")
 
-# The previous metadata class is contained in the 'meta' attribute of the dataset:
-print(dataset.meta)
-
-# LeRobotDataset actually wraps an underlying Hugging Face dataset
-# (see https://huggingface.co/docs/datasets for more information).
+# LeRobotDataset is actually a thin wrapper around an underlying Hugging Face dataset
+# (see https://huggingface.co/docs/datasets/index for more information).
+print(dataset)
 print(dataset.hf_dataset)
 
-# 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:
+# And provides additional utilities for robotics and compatibility with Pytorch
+print(f"\naverage number of frames per episode: {dataset.num_samples / dataset.num_episodes:.3f}")
+print(f"frames per second used during data collection: {dataset.fps=}")
+print(f"keys to access images from cameras: {dataset.camera_keys=}\n")
+
+# Access frame indexes associated to first episode
 episode_index = 0
 from_idx = dataset.episode_data_index["from"][episode_index].item()
 to_idx = dataset.episode_data_index["to"][episode_index].item()
 
-# 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)]
+# 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)]
 
-# The objects returned by the dataset are all torch.Tensors
-print(type(frames[0]))
-print(frames[0].shape)
+# 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]
 
-# 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.
+# Finally, we save the frames to a mp4 video for visualization.
+Path("outputs/examples/1_load_lerobot_dataset").mkdir(parents=True, exist_ok=True)
+imageio.mimsave("outputs/examples/1_load_lerobot_dataset/episode_0.mp4", frames, fps=dataset.fps)
 
 # For many machine learning applications we need to load the history of past observations or trajectories of
 # future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
 # differences with the current loaded frame. For instance:
 delta_timestamps = {
     # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
-    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],
+    "observation.image": [-1, -0.5, -0.20, 0],
+    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 20 ms, 10 ms, and current frame
+    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, -0.02, -0.01, 0],
     # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
     "action": [t / dataset.fps for t in range(64)],
 }
-# Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
-# timestamp, you still get a valid timestamp.
-
 dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
-print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
-print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
-print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
+print(f"\n{dataset[0]['observation.image'].shape=}")  # (4,c,h,w)
+print(f"{dataset[0]['observation.state'].shape=}")  # (8,c)
+print(f"{dataset[0]['action'].shape=}\n")  # (64,c)
 
 # Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
 # PyTorch datasets.
@@ -126,9 +84,8 @@ dataloader = torch.utils.data.DataLoader(
     batch_size=32,
     shuffle=True,
 )
-
 for batch in dataloader:
-    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
-    print(f"{batch['observation.state'].shape=}")  # (32, 5, c)
-    print(f"{batch['action'].shape=}")  # (32, 64, c)
+    print(f"{batch['observation.image'].shape=}")  # (32,4,c,h,w)
+    print(f"{batch['observation.state'].shape=}")  # (32,8,c)
+    print(f"{batch['action'].shape=}")  # (32,64,c)
     break

examples/2_evaluate_pretrained_policy.py

@@ -1,11 +1,6 @@
 """
 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
@@ -15,6 +10,7 @@ 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
 
@@ -22,15 +18,25 @@ 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)
 
-# 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.
+# Download the diffusion policy for pusht environment
+pretrained_policy_path = Path(snapshot_download("lerobot/diffusion_pusht"))
+# OR uncomment the following to evaluate a policy from the local outputs/train folder.
 # pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
 
-policy = DiffusionPolicy.from_pretrained(pretrained_policy_path, map_location=device)
+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
@@ -41,17 +47,7 @@ env = gym.make(
     max_episode_steps=300,
 )
 
-# We can verify that the shapes of the features expected by the policy match the ones from the observations
-# produced by the environment
-print(policy.config.input_features)
-print(env.observation_space)
-
-# Similarly, we can check that the actions produced by the policy will match the actions expected by the
-# environment
-print(policy.config.output_features)
-print(env.action_space)
-
-# Reset the policy and environments to prepare for rollout
+# Reset the policy and environmens to prepare for rollout
 policy.reset()
 numpy_observation, info = env.reset(seed=42)
 

examples/3_train_policy.py

@@ -8,99 +8,72 @@ from pathlib import Path
 
 import torch
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.common.datasets.utils import dataset_to_policy_features
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.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)
 
-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)
+# 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
 
-    # # Select your device
-    device = torch.device("cuda")
+# 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)
 
-    # 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
+# 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)
 
-    # 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}
+optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
 
-    # 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)
+# 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,
+)
 
-    # 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)
+# 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()
 
-    # 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],
-    }
+        if step % log_freq == 0:
+            print(f"step: {step} loss: {loss.item():.3f}")
+        step += 1
+        if step >= training_steps:
+            done = True
+            break
 
-    # 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, non_blocking=True) 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()
+# Save a policy checkpoint.
+policy.save_pretrained(output_directory)

examples/4_train_policy_with_script.md

@@ -1,223 +1,193 @@
-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.
-
+This tutorial will explain the training script, how to use it, and particularly the use of Hydra to configure everything needed for the training run.
 
 ## The training script
 
 LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
 
-- Initialize/load a configuration for the following steps using.
-- Instantiates a dataset.
-- (Optional) Instantiates a simulation environment corresponding to that dataset.
-- Instantiates a policy.
+- Loads a Hydra configuration file for the following steps (more on Hydra in a moment).
+- Makes a simulation environment.
+- Makes a dataset corresponding to that simulation environment.
+- Makes a policy.
 - Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
 
-## Overview of the configuration system
+## 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
 
-In the training script, the main function `train` expects a `TrainPipelineConfig` object:
 ```python
-# train.py
-@parser.wrap()
-def train(cfg: TrainPipelineConfig):
+python lerobot/scripts/train.py
 ```
 
-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)
+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:
 
-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"
+```yaml
+defaults:
+  - _self_
+  - env: pusht
+  - policy: diffusion
 ```
 
-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`.
+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`_.
 
-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.
+Then, `default.yaml` also contains common configuration parameters such as `device: cuda` or `use_amp: false` (for enabling fp16 training). Some other parameters are set to `???` which indicates that they are expected to be set in additional yaml files. For instance, `training.offline_steps: ???` in `default.yaml` is set to `200000` in `diffusion.yaml`.
 
+Thanks to this `defaults` section in `default.yaml`, if you want to train Diffusion Policy with PushT, you really only need to run:
 
-## Specifying values from the CLI
+```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:
 
-Let's say that we want to train [Diffusion Policy](../../lerobot/common/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
 ```bash
 python lerobot/scripts/train.py \
-    --dataset.repo_id=lerobot/pusht \
-    --policy.type=diffusion \
-    --env.type=pusht
+    policy=act \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    env=aloha \
+    env.task=AlohaTransferCube-v0
 ```
 
-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)
+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 see another example. Let's say you've been training [ACT](../../lerobot/common/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
 ```bash
 python lerobot/scripts/train.py \
-    --policy.type=act \
-    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
-    --env.type=aloha \
-    --output_dir=outputs/train/act_aloha_insertion
+    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 \
 ```
-> Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.
 
-We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
-Looking at the [`AlohaEnv`](../../lerobot/common/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
+There's one new thing here: `hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human`, which specifies where to save the training output.
+
+## Using a configuration file not in `lerobot/configs`
+
+Above we discusses the our training script is set up such that Hydra looks for `default.yaml` in `lerobot/configs`. But, if you have a configuration file elsewhere in your filesystem you may use:
+
 ```bash
-python lerobot/scripts/train.py \
-    --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
+python lerobot/scripts/train.py --config-dir PARENT/PATH --config-name FILE_NAME_WITHOUT_EXTENSION
 ```
 
-## Loading from a config file
+Note: here we use regular syntax for providing CLI arguments to a Python script, not Hydra's `param_name=param_value` syntax.
 
-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,
-        ...
-    },
-    ...
-}
-```
+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:
 
-We can then simply load the config values from this file using:
 ```bash
-python lerobot/scripts/train.py \
-    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
-    --output_dir=outputs/train/act_aloha_transfer_2
-```
-`--config_path` is also a special argument which allows to initialize the config from a local config file. It can point to a directory that contains `train_config.json` or to the config file itself directly.
-
-Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
-```bash
-python lerobot/scripts/train.py \
-    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
-    --output_dir=outputs/train/act_aloha_transfer_2
-    --policy.n_action_steps=80
-```
-> Note: While `--output_dir` is not required in general, in this case we need to specify it since it will otherwise take the value from the `train_config.json` (which is `outputs/train/act_aloha_transfer`). In order to prevent accidental deletion of previous run checkpoints, we raise an error if you're trying to write in an existing directory. This is not the case when resuming a run, which is what you'll learn next.
-
-`--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:
-```bash
-python lerobot/scripts/train.py --config_path=lerobot/diffusion_pusht
-```
-will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
-
-
-## Resume training
-
-Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to that here.
-
-Let's reuse the command from the previous run and add a few more options:
-```bash
-python lerobot/scripts/train.py \
-    --policy.type=act \
-    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
-    --env.type=aloha \
-    --env.task=AlohaTransferCube-v0 \
-    --log_freq=25 \
-    --save_freq=100 \
-    --output_dir=outputs/train/run_resumption
+python lerobot/scripts/train.py --config-dir outputs/train/my_experiment/checkpoints/last/pretrained_model --config-name config
 ```
 
-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.
+Note that you may still use the regular syntax for config parameter overrides (eg: by adding `training.offline_steps=200000`).
 
 ## Typical logs and metrics
 
-When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you configured your run correctly. 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 config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
 
 After that, you will see training log like this one:
+
 ```
 INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
 ```
-or evaluation log:
+
+or evaluation log like:
+
 ```
 INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
 ```
 
 These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
+
 - `smpl`: number of samples seen during training.
 - `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
 - `epch`: number of time all unique samples are seen (epoch).
@@ -230,45 +200,14 @@ 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
-
-We'll summarize here the main use cases to remember from this tutorial.
-
-#### Train a policy from scratch – CLI
-```bash
-python lerobot/scripts/train.py \
-    --policy.type=act \  # <- select 'act' policy
-    --env.type=pusht \  # <- select 'pusht' environment
-    --dataset.repo_id=lerobot/pusht  # <- train on this dataset
-```
-
-#### Train a policy from scratch - config file + CLI
-```bash
-python lerobot/scripts/train.py \
-    --config_path=path/to/pretrained_model \  # <- can also be a repo_id
-    --policy.n_action_steps=80  # <- you may still override values
-```
-
-#### Resume/continue a training run
-```bash
-python lerobot/scripts/train.py \
-    --config_path=checkpoint/pretrained_model/ \
-    --resume=true \
-    --steps=200000  # <- you can change some training parameters
-```
-
-#### Fine-tuning
-```bash
-python lerobot/scripts/train.py \
-    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \  # <- can also be a local path to a checkpoint
-    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
-    --env.type=aloha \
-    --env.task=AlohaInsertion-v0
-```
-
 ---
 
-Now that you know the basics of how to train a policy, you might want to know how to apply this knowledge to actual robots, or how to record your own datasets and train policies on your specific task?
-If that's the case, head over to the next tutorial [`7_get_started_with_real_robot.md`](./7_get_started_with_real_robot.md).
+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):
 
-Or in the meantime, happy training! 🤗
+```bash
+python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
+```
+
+Please, head on over to our [advanced tutorial on adapting policy configuration to various environments](./advanced/train_act_pusht/train_act_pusht.md) to learn more.
+
+Or in the meantime, happy coding! 🤗

examples/5_resume_training.md

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

examples/6_add_image_transforms.py

@@ -1,7 +1,7 @@
 """
 This script demonstrates how to use torchvision's image transformation with LeRobotDataset for data
 augmentation purposes. The transformations are passed to the dataset as an argument upon creation, and
-transforms are applied to the observation images before they are returned in the dataset's __getitem__.
+transforms are applied to the observation images before they are returned in the dataset's __get_item__.
 """
 
 from pathlib import Path
@@ -10,17 +10,17 @@ from torchvision.transforms import ToPILImage, v2
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
-dataset_repo_id = "lerobot/aloha_static_screw_driver"
+dataset_repo_id = "lerobot/aloha_static_tape"
 
 # Create a LeRobotDataset with no transformations
-dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
+dataset = LeRobotDataset(dataset_repo_id)
 # This is equivalent to `dataset = LeRobotDataset(dataset_repo_id, image_transforms=None)`
 
 # Get the index of the first observation in the first episode
 first_idx = dataset.episode_data_index["from"][0].item()
 
 # Get the frame corresponding to the first camera
-frame = dataset[first_idx][dataset.meta.camera_keys[0]]
+frame = dataset[first_idx][dataset.camera_keys[0]]
 
 
 # Define the transformations
@@ -28,16 +28,15 @@ transforms = v2.Compose(
     [
         v2.ColorJitter(brightness=(0.5, 1.5)),
         v2.ColorJitter(contrast=(0.5, 1.5)),
-        v2.ColorJitter(hue=(-0.1, 0.1)),
         v2.RandomAdjustSharpness(sharpness_factor=2, p=1),
     ]
 )
 
 # Create another LeRobotDataset with the defined transformations
-transformed_dataset = LeRobotDataset(dataset_repo_id, episodes=[0], image_transforms=transforms)
+transformed_dataset = LeRobotDataset(dataset_repo_id, image_transforms=transforms)
 
 # Get a frame from the transformed dataset
-transformed_frame = transformed_dataset[first_idx][transformed_dataset.meta.camera_keys[0]]
+transformed_frame = transformed_dataset[first_idx][transformed_dataset.camera_keys[0]]
 
 # Create a directory to store output images
 output_dir = Path("outputs/image_transforms")

examples/7_get_started_with_real_robot.md

@@ -29,21 +29,16 @@ 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 (make sure gcc is installed).
+First, install the additional dependencies required for robots built with dynamixel motors like Koch v1.1 by running one of the following commands.
 
 Using `pip`:
 ```bash
 pip install -e ".[dynamixel]"
 ```
 
-Using `poetry`:
+Or using `poetry`:
 ```bash
-poetry sync --extras "dynamixel"
-```
-
-Using `uv`:
-```bash
-uv sync --extra "dynamixel"
+poetry install --sync --extras "dynamixel"
 ```
 
 /!\ For Linux only, ffmpeg and opencv requires conda install for now. Run this exact sequence of commands:
@@ -59,53 +54,24 @@ 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.
 
-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.
+*Copy pasting python code*
 
-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):
+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):
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=koch \
-  --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+  --robot-path lerobot/configs/robot/koch.yaml \
+  --robot-overrides '~cameras'  # do not instantiate the cameras
 ```
 
 It will automatically:
-1. Identify any missing calibrations and initiate the calibration procedure.
-2. Connect the robot and start teleoperation.
+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.
 
 ### 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"`).
@@ -139,10 +105,10 @@ 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 this command with your ports:
+Troubleshooting: On Linux, you might need to give access to the USB ports by running:
 ```bash
-sudo chmod 666 /dev/tty.usbmodem575E0032081
-sudo chmod 666 /dev/tty.usbmodem575E0031751
+sudo chmod 666 /dev/ttyACM0
+sudo chmod 666 /dev/ttyACM1
 ```
 
 *Listing and Configuring Motors*
@@ -151,11 +117,13 @@ 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_config = DynamixelMotorsBusConfig(
-    port="/dev/tty.usbmodem575E0031751",
+leader_port = "/dev/tty.usbmodem575E0031751"
+follower_port = "/dev/tty.usbmodem575E0032081"
+
+leader_arm = DynamixelMotorsBus(
+    port=leader_port,
     motors={
         # name: (index, model)
         "shoulder_pan": (1, "xl330-m077"),
@@ -167,8 +135,8 @@ leader_config = DynamixelMotorsBusConfig(
     },
 )
 
-follower_config = DynamixelMotorsBusConfig(
-    port="/dev/tty.usbmodem575E0032081",
+follower_arm = DynamixelMotorsBus(
+    port=follower_port,
     motors={
         # name: (index, model)
         "shoulder_pan": (1, "xl430-w250"),
@@ -179,57 +147,45 @@ follower_config = DynamixelMotorsBusConfig(
         "gripper": (6, "xl330-m288"),
     },
 )
-
-leader_arm = DynamixelMotorsBus(leader_config)
-follower_arm = DynamixelMotorsBus(follower_config)
 ```
 
-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
+*Updating the YAML Configuration File*
 
-    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"],
-                },
-            ),
-        }
-    )
+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"]
+[...]
 ```
 
+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.
@@ -356,27 +312,27 @@ 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_config` and `follower_config`.
+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`.
 
-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.
+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.
 
+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_config = KochRobotConfig(
-    leader_arms={"main": leader_config},
-    follower_arms={"main": follower_config},
-    cameras={},  # We don't use any camera for now
+robot = ManipulatorRobot(
+    robot_type="koch",
+    leader_arms={"main": leader_arm},
+    follower_arms={"main": follower_arm},
+    calibration_dir=".cache/calibration/koch",
 )
-robot = ManipulatorRobot(robot_config)
 ```
 
-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.
+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.
 
-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.
+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`.
 
 **Calibrate and Connect the ManipulatorRobot**
 
@@ -623,11 +579,9 @@ 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_config = OpenCVCameraConfig(camera_index=0)
-camera = OpenCVCamera(config)
+camera = OpenCVCamera(camera_index=0)
 camera.connect()
 color_image = camera.read()
 
@@ -649,7 +603,7 @@ uint8
 
 With certain camera, you can also specify additional parameters like frame rate, resolution, and color mode during instantiation. For instance:
 ```python
-config = OpenCVCameraConfig(camera_index=0, fps=30, width=640, height=480)
+camera = OpenCVCamera(camera_index=0, fps=30, width=640, height=480)
 ```
 
 If the provided arguments are not compatible with the camera, an exception will be raised.
@@ -672,8 +626,8 @@ robot = ManipulatorRobot(
     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),
+        "laptop": OpenCVCamera(0, fps=30, width=640, height=480),
+        "phone": OpenCVCamera(1, fps=30, width=640, height=480),
     },
 )
 robot.connect()
@@ -698,15 +652,34 @@ torch.Size([3, 480, 640])
 255
 ```
 
-### d. Use `control_robot.py` and our `teleoperate` function
+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
+```
 
-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.
+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.
 
 Try running this code to teleoperate your robot (if you dont have a camera, keep reading):
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=koch \
-  --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+  --robot-path lerobot/configs/robot/koch.yaml
 ```
 
 You will see a lot of lines appearing like this one:
@@ -721,12 +694,21 @@ 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.
 
-Importantly: If you don't have any camera, you can remove them dynamically with this [draccus](https://github.com/dlwh/draccus) syntax `--robot.cameras='{}'`:
+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:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=koch \
-  --robot.cameras='{}' \
-  --control.type=teleoperate
+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'
 ```
 
 We advise to create a new yaml file when the command becomes too long.
@@ -762,23 +744,23 @@ 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 the `record` function
+### a. Use `koch.yaml` and 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 each episode recording.
+1. Frames from cameras are saved on disk in threads, and encoded into videos at the end of 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 `--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 might need to add `--control.local_files_only=true` if your dataset was not uploaded to hugging face hub. Also you will need to manually delete the dataset directory 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 `--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.
 5. Set the flow of data recording using command line arguments:
-   - `--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).
+   - `--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).
 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 the command line to override anything.
+7. Similarly to `teleoperate`, you can also use `--robot-path` and `--robot-overrides` to specify your robots.
 
 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
@@ -789,29 +771,28 @@ 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 `--control.push_to_hub=false`.
+If you don't want to push to hub, use `--push-to-hub 0`.
 
 Now run this to record 2 episodes:
 ```bash
-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
+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
 ```
 
-
-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).
+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).
 
 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)
@@ -859,11 +840,10 @@ 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%">
@@ -875,16 +855,14 @@ 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 \
-  --robot.type=koch \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/koch_test \
-  --control.episode=0
+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
 ```
 
-Note: You might need to add `--control.local_files_only=true` if your dataset was not uploaded to hugging face hub.
-
 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).
 
 ## 4. Train a policy on your data
@@ -893,22 +871,54 @@ 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
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/koch_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_koch_test \
-  --job_name=act_koch_test \
-  --device=cuda \
-  --wandb.enable=true
+DATA_DIR=data python lerobot/scripts/train.py \
+  dataset_repo_id=${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
 ```
 
-Note: You might need to add `--dataset.local_files_only=true` if your dataset was not uploaded to hugging face hub.
-
 Let's explain it:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/koch_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+1. We provided the dataset as argument with `dataset_repo_id=${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: ${fps}
+```
+It should match your dataset (e.g. `fps: 30`) and your robot (e.g. `state_dim: 6` and `action_dim: 6`). We are still working on simplifying this in future versions of `lerobot`.
 4. We provided `device=cuda` since we are training on a Nvidia GPU, but you could use `device=mps` to train on Apple silicon.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+6. We added `DATA_DIR=data` to access your dataset stored in your local `data` directory. If you dont provide `DATA_DIR`, your dataset will be downloaded from Hugging Face hub to your cache folder `$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)
 
@@ -972,36 +982,36 @@ for _ in range(inference_time_s * fps):
     busy_wait(1 / fps - dt_s)
 ```
 
-### a. Use our `record` function
+### a. Use `koch.yaml` and 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 \
-  --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
+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
 ```
 
 As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_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`).
+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`).
 
 ### 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 \
-  --repo-id ${HF_USER}/eval_act_koch_test
+  --root data \
+  --repo-id ${HF_USER}/eval_koch_test
 ```
 
 ## 6. Next step

examples/8_use_stretch.md

@@ -92,9 +92,8 @@ Serial Number = stretch-se3-3054
 **Calibrate (Optional)**
 Before operating Stretch, you need to [home](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#homing) it first. Be mindful about giving Stretch some space as this procedure will move the robot's arm and gripper. Now run this command:
 ```bash
-python lerobot/scripts/control_robot.py \
-    --robot.type=stretch \
-    --control.type=calibrate
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/stretch.yaml
 ```
 This is equivalent to running `stretch_robot_home.py`
 
@@ -105,9 +104,8 @@ Before trying teleoperation, you need activate the gamepad controller by pressin
 
 Now try out teleoperation (see above documentation to learn about the gamepad controls):
 ```bash
-python lerobot/scripts/control_robot.py \
-    --robot.type=stretch \
-    --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/stretch.yaml
 ```
 This is essentially the same as running `stretch_gamepad_teleop.py`
 
@@ -127,18 +125,17 @@ echo $HF_USER
 
 Record one episode:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=stretch \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/stretch_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.push_to_hub=true
+python lerobot/scripts/control_robot.py record \
+    --robot-path lerobot/configs/robot/stretch.yaml \
+    --fps 20 \
+    --root data \
+    --repo-id ${HF_USER}/stretch_test \
+    --tags stretch tutorial \
+    --warmup-time-s 3 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 1 \
+    --push-to-hub 0
 ```
 
 > **Note:** If you're using ssh to connect to Stretch and run this script, you won't be able to visualize its cameras feed (though they will still be recording). To see the cameras stream, use [tethered](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#tethered-setup) or [untethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#untethered-setup).
@@ -146,12 +143,12 @@ python lerobot/scripts/control_robot.py \
 **Replay an episode**
 Now try to replay this episode (make sure the robot's initial position is the same):
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=stretch \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/stretch_test \
-  --control.episode=0
+python lerobot/scripts/control_robot.py replay \
+    --robot-path lerobot/configs/robot/stretch.yaml \
+    --fps 20 \
+    --root data \
+    --repo-id ${HF_USER}/stretch_test \
+    --episode 0
 ```
 
 Follow [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) to train a policy on your data and run inference on your robot. You will need to adapt the code for Stretch.

examples/9_use_aloha.md

@@ -51,18 +51,16 @@ Teleoperation consists in manually operating the leader arms to move the followe
 
 By running the following code, you can start your first **SAFE** teleoperation:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=aloha \
-  --robot.max_relative_target=5 \
-  --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=5
 ```
 
-By adding `--robot.max_relative_target=5`, we override the default value for `max_relative_target` defined in [`AlohaRobotConfig`](lerobot/common/robot_devices/robots/configs.py). It is expected to be `5` to limit the magnitude of the movement for more safety, but the teleoperation won't be smooth. When you feel confident, you can disable this limit by adding `--robot.max_relative_target=null` to the command line:
+By adding `--robot-overrides max_relative_target=5`, we override the default value for `max_relative_target` defined in `lerobot/configs/robot/aloha.yaml`. It is expected to be `5` to limit the magnitude of the movement for more safety, but the teloperation won't be smooth. When you feel confident, you can disable this limit by adding `--robot-overrides max_relative_target=null` to the command line:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=aloha \
-  --robot.max_relative_target=null \
-  --control.type=teleoperate
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=null
 ```
 
 ## Record a dataset
@@ -82,70 +80,71 @@ echo $HF_USER
 
 Record 2 episodes and upload your dataset to the hub:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=aloha \
-  --robot.max_relative_target=null \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/aloha_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.push_to_hub=true
+python lerobot/scripts/control_robot.py record \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=null \
+    --fps 30 \
+    --root data \
+    --repo-id ${HF_USER}/aloha_test \
+    --tags aloha 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 `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+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}/aloha_test
 ```
 
-If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with:
+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}/aloha_test
 ```
 
 ## Replay an episode
 
 **/!\ FOR SAFETY, READ THIS /!\**
-Replay consists in automatically replaying the sequence of actions (i.e. goal positions for your motors) recorded in a given dataset episode. Make sure the current initial position of your robot is similar to the one in your episode, so that your follower arms don't move too fast to go to the first goal positions. For safety, you might want to add `--robot.max_relative_target=5` to your command line as explained above.
+Replay consists in automatically replaying the sequence of actions (i.e. goal positions for your motors) recorded in a given dataset episode. Make sure the current initial position of your robot is similar to the one in your episode, so that your follower arms don't move too fast to go to the first goal positions. For safety, you might want to add `--robot-overrides max_relative_target=5` to your command line as explained above.
 
 Now try to replay the first episode on your robot:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=aloha \
-  --robot.max_relative_target=null \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/aloha_test \
-  --control.episode=0
+python lerobot/scripts/control_robot.py replay \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=null \
+    --fps 30 \
+    --root data \
+    --repo-id ${HF_USER}/aloha_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
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/aloha_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_aloha_test \
-  --job_name=act_aloha_test \
-  --device=cuda \
-  --wandb.enable=true
+DATA_DIR=data python lerobot/scripts/train.py \
+  dataset_repo_id=${HF_USER}/aloha_test \
+  policy=act_aloha_real \
+  env=aloha_real \
+  hydra.run.dir=outputs/train/act_aloha_test \
+  hydra.job.name=act_aloha_test \
+  device=cuda \
+  wandb.enable=true
 ```
 
 Let's explain it:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/aloha_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `device=cuda` since we are training on a Nvidia GPU, but you could use `device=mps` to train on Apple silicon.
+1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/aloha_test`.
+2. We provided the policy with `policy=act_aloha_real`. This loads configurations from [`lerobot/configs/policy/act_aloha_real.yaml`](../lerobot/configs/policy/act_aloha_real.yaml). Importantly, this policy uses 4 cameras as input `cam_right_wrist`, `cam_left_wrist`, `cam_high`, and `cam_low`.
+3. We provided an environment as argument with `env=aloha_real`. This loads configurations from [`lerobot/configs/env/aloha_real.yaml`](../lerobot/configs/env/aloha_real.yaml). Note: this yaml defines 18 dimensions for the `state_dim` and `action_dim`, corresponding to 18 motors, not 14 motors as used in previous Aloha work. This is because, we include the `shoulder_shadow` and `elbow_shadow` motors for simplicity.
+4. We provided `device=cuda` since we are training on a Nvidia GPU.
 5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-
-For more information on the `train` script see the previous tutorial: [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md)
+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_aloha_test/checkpoints`.
 
@@ -153,26 +152,25 @@ Training should take several hours. You will find checkpoints in `outputs/train/
 
 You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
 ```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=aloha \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/eval_act_aloha_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=10 \
-  --control.push_to_hub=true \
-  --control.policy.path=outputs/train/act_aloha_test/checkpoints/last/pretrained_model \
-  --control.num_image_writer_processes=1
+python lerobot/scripts/control_robot.py record \
+  --robot-path lerobot/configs/robot/aloha.yaml \
+  --robot-overrides max_relative_target=null \
+  --fps 30 \
+  --root data \
+  --repo-id ${HF_USER}/eval_act_aloha_test \
+  --tags aloha tutorial eval \
+  --warmup-time-s 5 \
+  --episode-time-s 40 \
+  --reset-time-s 10 \
+  --num-episodes 10 \
+  --num-image-writer-processes 1 \
+  -p outputs/train/act_aloha_test/checkpoints/last/pretrained_model
 ```
 
 As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_aloha_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_aloha_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_aloha_test`).
-3. We use `--control.num_image_writer_processes=1` instead of the default value (`0`). On our computer, using a dedicated process to write images from the 4 cameras on disk allows to reach constent 30 fps during inference. Feel free to explore different values for `--control.num_image_writer_processes`.
+1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_aloha_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_aloha_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_aloha_test`).
+3. We use `--num-image-writer-processes 1` instead of the default value (`0`). On our computer, using a dedicated process to write images from the 4 cameras on disk allows to reach constent 30 fps during inference. Feel free to explore different values for `--num-image-writer-processes`.
 
 ## More
 

examples/advanced/1_train_act_pusht/act_pusht.yaml

@@ -0,0 +1,87 @@
+# @package _global_
+
+# Change the seed to match what PushT eval uses
+# (to avoid evaluating on seeds used for generating the training data).
+seed: 100000
+# Change the dataset repository to the PushT one.
+dataset_repo_id: lerobot/pusht
+
+override_dataset_stats:
+  observation.image:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+training:
+  offline_steps: 80000
+  online_steps: 0
+  eval_freq: 10000
+  save_freq: 100000
+  log_freq: 250
+  save_model: true
+
+  batch_size: 8
+  lr: 1e-5
+  lr_backbone: 1e-5
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
+
+  delta_timestamps:
+    action: "[i / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    observation.image: [3, 96, 96]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.image: mean_std
+    # Use min_max normalization just because it's more standard.
+    observation.state: min_max
+  output_normalization_modes:
+    # Use min_max normalization just because it's more standard.
+    action: min_max
+
+  # Architecture.
+  # Vision backbone.
+  vision_backbone: resnet18
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
+  replace_final_stride_with_dilation: false
+  # Transformer layers.
+  pre_norm: false
+  dim_model: 512
+  n_heads: 8
+  dim_feedforward: 3200
+  feedforward_activation: relu
+  n_encoder_layers: 4
+    # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
+  # that means only the first layer is used. Here we match the original implementation by setting this to 1.
+  # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
+  n_decoder_layers: 1
+  # VAE.
+  use_vae: true
+  latent_dim: 32
+  n_vae_encoder_layers: 4
+
+  # Inference.
+  temporal_ensemble_coeff: null
+
+  # Training and loss computation.
+  dropout: 0.1
+  kl_weight: 10.0

examples/advanced/1_train_act_pusht/train_act_pusht.md

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

examples/advanced/2_calculate_validation_loss.py

@@ -9,82 +9,82 @@ 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, LeRobotDatasetMetadata
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 
+device = torch.device("cuda")
 
-def main():
-    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")
 
-    # Download the diffusion policy for pusht environment
-    pretrained_policy_path = "lerobot/diffusion_pusht"
-    # OR uncomment the following to evaluate a policy from the local outputs/train folder.
-    # pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
+policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
+policy.eval()
+policy.to(device)
 
-    policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
-    policy.eval()
-    policy.to(device)
+# Set up the dataset.
+delta_timestamps = {
+    # Load the previous image and state at -0.1 seconds before current frame,
+    # then load current image and state corresponding to 0.0 second.
+    "observation.image": [-0.1, 0.0],
+    "observation.state": [-0.1, 0.0],
+    # Load the previous action (-0.1), the next action to be executed (0.0),
+    # and 14 future actions with a 0.1 seconds spacing. All these actions will be
+    # used to calculate the loss.
+    "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
+}
 
-    # Set up the dataset.
-    delta_timestamps = {
-        # Load the previous image and state at -0.1 seconds before current frame,
-        # then load current image and state corresponding to 0.0 second.
-        "observation.image": [-0.1, 0.0],
-        "observation.state": [-0.1, 0.0],
-        # Load the previous action (-0.1), the next action to be executed (0.0),
-        # and 14 future actions with a 0.1 seconds spacing. All these actions will be
-        # used to calculate the loss.
-        "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
-    }
+# Load the last 10% of episodes of the dataset as a validation set.
+# - Load full dataset
+full_dataset = LeRobotDataset("lerobot/pusht", split="train")
+# - Calculate train and val subsets
+num_train_episodes = math.floor(full_dataset.num_episodes * 90 / 100)
+num_val_episodes = full_dataset.num_episodes - num_train_episodes
+print(f"Number of episodes in full dataset: {full_dataset.num_episodes}")
+print(f"Number of episodes in training dataset (90% subset): {num_train_episodes}")
+print(f"Number of episodes in validation dataset (10% subset): {num_val_episodes}")
+# - Get first frame index of the validation set
+first_val_frame_index = full_dataset.episode_data_index["from"][num_train_episodes].item()
+# - Load frames subset belonging to validation set using the `split` argument.
+#   It utilizes the `datasets` library's syntax for slicing datasets.
+#   For more information on the Slice API, please see:
+#   https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
+train_dataset = LeRobotDataset(
+    "lerobot/pusht", split=f"train[:{first_val_frame_index}]", delta_timestamps=delta_timestamps
+)
+val_dataset = LeRobotDataset(
+    "lerobot/pusht", split=f"train[{first_val_frame_index}:]", delta_timestamps=delta_timestamps
+)
+print(f"Number of frames in training dataset (90% subset): {len(train_dataset)}")
+print(f"Number of frames in validation dataset (10% subset): {len(val_dataset)}")
 
-    # Load the last 10% of episodes of the dataset as a validation set.
-    # - Load dataset metadata
-    dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
-    # - Calculate train and val episodes
-    total_episodes = dataset_metadata.total_episodes
-    episodes = list(range(dataset_metadata.total_episodes))
-    num_train_episodes = math.floor(total_episodes * 90 / 100)
-    train_episodes = episodes[:num_train_episodes]
-    val_episodes = episodes[num_train_episodes:]
-    print(f"Number of episodes in full dataset: {total_episodes}")
-    print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
-    print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
-    # - Load train an val datasets
-    train_dataset = LeRobotDataset(
-        "lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps
-    )
-    val_dataset = LeRobotDataset("lerobot/pusht", episodes=val_episodes, delta_timestamps=delta_timestamps)
-    print(f"Number of frames in training dataset (90% subset): {len(train_dataset)}")
-    print(f"Number of frames in validation dataset (10% subset): {len(val_dataset)}")
+# Create dataloader for evaluation.
+val_dataloader = torch.utils.data.DataLoader(
+    val_dataset,
+    num_workers=4,
+    batch_size=64,
+    shuffle=False,
+    pin_memory=device != torch.device("cpu"),
+    drop_last=False,
+)
 
-    # Create dataloader for evaluation.
-    val_dataloader = torch.utils.data.DataLoader(
-        val_dataset,
-        num_workers=4,
-        batch_size=64,
-        shuffle=False,
-        pin_memory=device != torch.device("cpu"),
-        drop_last=False,
-    )
+# Run validation loop.
+loss_cumsum = 0
+n_examples_evaluated = 0
+for batch in val_dataloader:
+    batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+    output_dict = policy.forward(batch)
 
-    # Run validation loop.
-    loss_cumsum = 0
-    n_examples_evaluated = 0
-    for batch in val_dataloader:
-        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
-        loss, _ = policy.forward(batch)
+    loss_cumsum += output_dict["loss"].item()
+    n_examples_evaluated += batch["index"].shape[0]
 
-        loss_cumsum += loss.item()
-        n_examples_evaluated += batch["index"].shape[0]
+# Calculate the average loss over the validation set.
+average_loss = loss_cumsum / n_examples_evaluated
 
-    # 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()
+print(f"Average loss on validation set: {average_loss:.4f}")

examples/port_datasets/pusht_zarr.py

@@ -1,222 +0,0 @@
-import shutil
-from pathlib import Path
-
-import numpy as np
-import torch
-
-from lerobot.common.datasets.lerobot_dataset import LEROBOT_HOME, 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,))
-    # 8 keypoints with 2 coords each
-    keypoints = np.zeros((num_frames, 16))
-
-    # 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] = torch.from_numpy(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 (LEROBOT_HOME / repo_id).exists():
-        shutil.rmtree(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)
-
-    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
-            frame = {
-                "action": torch.from_numpy(action[i]),
-                # Shift reward and success by +1 until the last item of the episode
-                "next.reward": reward[i + (frame_idx < num_frames - 1)],
-                "next.success": success[i + (frame_idx < num_frames - 1)],
-            }
-
-            frame["observation.state"] = torch.from_numpy(agent_pos[i])
-
-            if mode == "keypoints":
-                frame["observation.environment_state"] = torch.from_numpy(keypoints[i])
-            else:
-                frame["observation.image"] = torch.from_numpy(image[i])
-
-            dataset.add_frame(frame)
-
-        dataset.save_episode(task=PUSHT_TASK)
-
-    dataset.consolidate()
-
-    if push_to_hub:
-        dataset.push_to_hub()
-
-
-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, local_files_only=True)
-        # breakpoint()

lerobot/__init__.py

@@ -58,6 +58,7 @@ available_tasks_per_env = {
     ],
     "pusht": ["PushT-v0"],
     "xarm": ["XarmLift-v0"],
+    "dora_aloha_real": ["DoraAloha-v0", "DoraKoch-v0", "DoraReachy2-v0"],
 }
 available_envs = list(available_tasks_per_env.keys())
 
@@ -85,6 +86,23 @@ 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 = [
@@ -163,8 +181,8 @@ available_real_world_datasets = [
     "lerobot/usc_cloth_sim",
 ]
 
-available_datasets = sorted(
-    set(itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets))
+available_datasets = list(
+    itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets)
 )
 
 # lists all available policies from `lerobot/common/policies`
@@ -203,6 +221,7 @@ 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/constants.py

@@ -1,17 +0,0 @@
-# keys
-OBS_ENV = "observation.environment_state"
-OBS_ROBOT = "observation.state"
-OBS_IMAGE = "observation.image"
-OBS_IMAGES = "observation.images"
-ACTION = "action"
-
-# files & directories
-CHECKPOINTS_DIR = "checkpoints"
-LAST_CHECKPOINT_LINK = "last"
-PRETRAINED_MODEL_DIR = "pretrained_model"
-TRAINING_STATE_DIR = "training_state"
-RNG_STATE = "rng_state.safetensors"
-TRAINING_STEP = "training_step.json"
-OPTIMIZER_STATE = "optimizer_state.safetensors"
-OPTIMIZER_PARAM_GROUPS = "optimizer_param_groups.json"
-SCHEDULER_STATE = "scheduler_state.json"

lerobot/common/datasets/card_template.md

@@ -1,27 +0,0 @@
----
-# 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

@@ -19,6 +19,9 @@ from math import ceil
 import einops
 import torch
 import tqdm
+from datasets import Image
+
+from lerobot.common.datasets.video_utils import VideoFrame
 
 
 def get_stats_einops_patterns(dataset, num_workers=0):
@@ -36,13 +39,15 @@ def get_stats_einops_patterns(dataset, num_workers=0):
     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
 
-    for key in dataset.features:
         # sanity check that tensors are not float64
         assert batch[key].dtype != torch.float64
 
-        # if isinstance(feats_type, (VideoFrame, Image)):
-        if key in dataset.meta.camera_keys:
+        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}"
@@ -58,7 +63,7 @@ def get_stats_einops_patterns(dataset, num_workers=0):
         elif batch[key].ndim == 1:
             stats_patterns[key] = "b -> 1"
         else:
-            raise ValueError(f"{key}, {batch[key].shape}")
+            raise ValueError(f"{key}, {feats_type}, {batch[key].shape}")
 
     return stats_patterns
 
@@ -170,45 +175,39 @@ def aggregate_stats(ls_datasets) -> dict[str, torch.Tensor]:
     """
     data_keys = set()
     for dataset in ls_datasets:
-        data_keys.update(dataset.meta.stats.keys())
+        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(
-                    [ds.meta.stats[data_key][stat_key] for ds in ls_datasets if data_key in ds.meta.stats],
-                    dim=0,
-                ),
+                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_frames for d in ls_datasets if data_key in d.meta.stats)
+        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_frames / total_samples) are needed tor minimize the risk of
+        # NOTE: the brackets around (d.num_samples / total_samples) are needed tor minimize the risk of
         # numerical overflow!
         stats[data_key]["mean"] = sum(
-            d.meta.stats[data_key]["mean"] * (d.num_frames / total_samples)
+            d.stats[data_key]["mean"] * (d.num_samples / total_samples)
             for d in ls_datasets
-            if data_key in d.meta.stats
+            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_frames / total_samples) are needed tor minimize the risk of
+        # 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.meta.stats[data_key]["std"] ** 2
-                    + (d.meta.stats[data_key]["mean"] - stats[data_key]["mean"]) ** 2
-                )
-                * (d.num_frames / total_samples)
+                (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.meta.stats
+                if data_key in d.stats
             )
         )
     return stats

lerobot/common/datasets/factory.py

@@ -14,102 +14,104 @@
 # 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,
-    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)
-}
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
+from lerobot.common.datasets.transforms import get_image_transforms
 
 
-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.
+def resolve_delta_timestamps(cfg):
+    """Resolves delta_timestamps config key (in-place) by using `eval`.
 
-    Args:
-        cfg (PreTrainedConfig): The PreTrainedConfig to read delta_indices from.
-        ds_meta (LeRobotDatasetMetadata): The dataset from which features and fps are used to build
-            delta_timestamps against.
-
-    Returns:
-        dict[str, list] | None: A dictionary of delta_timestamps, e.g.:
-            {
-                "observation.state": [-0.04, -0.02, 0]
-                "observation.action": [-0.02, 0, 0.02]
-            }
-            returns `None` if the the resulting dict is empty.
+    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 = {}
-    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
+    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: 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
+def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotDataset:
     """
-    image_transforms = (
-        ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None
-    )
+    Args:
+        cfg: A Hydra config as per the LeRobot config scheme.
+        split: Select the data subset used to create an instance of LeRobotDataset.
+            All datasets hosted on [lerobot](https://huggingface.co/lerobot) contain only one subset: "train".
+            Thus, by default, `split="train"` selects all the available data. `split` aims to work like the
+            slicer in the hugging face datasets:
+            https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
+            As of now, it only supports `split="train[:n]"` to load the first n frames of the dataset or
+            `split="train[n:]"` to load the last n frames. For instance `split="train[:1000]"`.
+    Returns:
+        The LeRobotDataset.
+    """
+    if not isinstance(cfg.dataset_repo_id, (str, ListConfig)):
+        raise ValueError(
+            "Expected cfg.dataset_repo_id to be either a single string to load one dataset or a list of "
+            "strings to load multiple datasets."
+        )
 
-    if isinstance(cfg.dataset.repo_id, str):
-        ds_meta = LeRobotDatasetMetadata(cfg.dataset.repo_id, local_files_only=cfg.dataset.local_files_only)
-        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
+    # A soft check to warn if the environment matches the dataset. Don't check if we are using a real world env (dora).
+    if cfg.env.name != "dora":
+        if isinstance(cfg.dataset_repo_id, str):
+            dataset_repo_ids = [cfg.dataset_repo_id]  # single dataset
+        else:
+            dataset_repo_ids = cfg.dataset_repo_id  # multiple datasets
+
+        for dataset_repo_id in dataset_repo_ids:
+            if cfg.env.name not in dataset_repo_id:
+                logging.warning(
+                    f"There might be a mismatch between your training dataset ({dataset_repo_id=}) and your "
+                    f"environment ({cfg.env.name=})."
+                )
+
+    resolve_delta_timestamps(cfg)
+
+    image_transforms = None
+    if cfg.training.image_transforms.enable:
+        cfg_tf = cfg.training.image_transforms
+        image_transforms = get_image_transforms(
+            brightness_weight=cfg_tf.brightness.weight,
+            brightness_min_max=cfg_tf.brightness.min_max,
+            contrast_weight=cfg_tf.contrast.weight,
+            contrast_min_max=cfg_tf.contrast.min_max,
+            saturation_weight=cfg_tf.saturation.weight,
+            saturation_min_max=cfg_tf.saturation.min_max,
+            hue_weight=cfg_tf.hue.weight,
+            hue_min_max=cfg_tf.hue.min_max,
+            sharpness_weight=cfg_tf.sharpness.weight,
+            sharpness_min_max=cfg_tf.sharpness.min_max,
+            max_num_transforms=cfg_tf.max_num_transforms,
+            random_order=cfg_tf.random_order,
+        )
+
+    if isinstance(cfg.dataset_repo_id, str):
         dataset = LeRobotDataset(
-            cfg.dataset.repo_id,
-            episodes=cfg.dataset.episodes,
-            delta_timestamps=delta_timestamps,
+            cfg.dataset_repo_id,
+            split=split,
+            delta_timestamps=cfg.training.get("delta_timestamps"),
             image_transforms=image_transforms,
-            video_backend=cfg.dataset.video_backend,
-            local_files_only=cfg.dataset.local_files_only,
+            video_backend=cfg.video_backend,
         )
     else:
-        raise NotImplementedError("The MultiLeRobotDataset isn't supported for now.")
         dataset = MultiLeRobotDataset(
-            cfg.dataset.repo_id,
-            # TODO(aliberts): add proper support for multi dataset
-            # delta_timestamps=delta_timestamps,
+            cfg.dataset_repo_id,
+            split=split,
+            delta_timestamps=cfg.training.get("delta_timestamps"),
             image_transforms=image_transforms,
-            video_backend=cfg.dataset.video_backend,
-        )
-        logging.info(
-            "Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
-            f"{pformat(dataset.repo_id_to_index, indent=2)}"
+            video_backend=cfg.video_backend,
         )
 
-    if cfg.dataset.use_imagenet_stats:
-        for key in dataset.meta.camera_keys:
-            for stats_type, stats in IMAGENET_STATS.items():
-                dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
+    if cfg.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)
 
     return dataset

lerobot/common/datasets/image_writer.py

@@ -1,160 +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.
-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_image(image_array: np.ndarray) -> PIL.Image.Image:
-    # TODO(aliberts): handle 1 channel and 4 for depth images
-    if image_array.ndim == 3 and image_array.shape[0] in [1, 3]:
-        # Transpose from pytorch convention (C, H, W) to (H, W, C)
-        image_array = image_array.transpose(1, 2, 0)
-    if image_array.dtype != np.uint8:
-        # Assume the image is in [0, 1] range for floating-point data
-        image_array = np.clip(image_array, 0, 1)
-        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_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_frames > 0:
+        if self.num_samples > 0:
             last_episode_index = self._data[OnlineBuffer.EPISODE_INDEX_KEY][next_index - 1]
             last_data_index = self._data[OnlineBuffer.INDEX_KEY][next_index - 1]
             data[OnlineBuffer.EPISODE_INDEX_KEY] += last_episode_index + 1
@@ -227,11 +227,11 @@ class OnlineBuffer(torch.utils.data.Dataset):
         )
 
     @property
-    def num_frames(self) -> int:
+    def num_samples(self) -> int:
         return np.count_nonzero(self._data[OnlineBuffer.OCCUPANCY_MASK_KEY])
 
     def __len__(self):
-        return self.num_frames
+        return self.num_samples
 
     def _item_to_tensors(self, item: dict) -> dict:
         item_ = {}

lerobot/common/datasets/populate_dataset.py

@@ -0,0 +1,468 @@
+"""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/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

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

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
@@ -72,7 +72,7 @@ def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episod
             # However, note that "nearest" might synchronize the reference camera with other cameras on slightly future timestamps.
             # are too far appart.
             direction="nearest",
-            tolerance=pd.Timedelta(f"{1 / fps} seconds"),
+            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]

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

@@ -0,0 +1,639 @@
+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

@@ -0,0 +1,106 @@
+#!/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

@@ -0,0 +1,200 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+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

@@ -0,0 +1,859 @@
+#!/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

@@ -14,16 +14,13 @@
 # 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_datsets before running this script.
-
 Example:
     python lerobot/scripts/push_dataset_to_hub.py \
-        --raw-dir /path/to/data/bridge_dataset/1.0.0/ \
-        --repo-id your_hub/sampled_bridge_data_v2 \
-        --raw-format rlds \
+        --raw-dir /hdd/tensorflow_datasets/bridge_dataset/1.0.0/ \
+        --repo-id youliangtan/sampled_bridge_data_v2 \
+        --raw-format openx_rlds.bridge_orig \
         --episodes 3 4 5 8 9
 
 Exact dataset fps defined in openx/config.py, obtained from:
@@ -38,21 +35,28 @@ 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)
 
 
@@ -104,6 +108,7 @@ def load_from_raw(
     video: bool,
     episodes: list[int] | None = None,
     encoding: dict | None = None,
+    openx_dataset_name: str | None = None,
 ):
     """
     Args:
@@ -131,17 +136,16 @@ 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
-    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)
+    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]
 
     lang_key = "language_instruction" if "language_instruction" in dataset.element_spec else None
 
@@ -189,31 +193,50 @@ def load_from_raw(
 
         num_frames = episode["action"].shape[0]
 
-        ep_dict = {}
-        for key in state_keys:
-            ep_dict[f"observation.{key}"] = tf_to_torch(episode["observation"][key])
+        ###########################################################
+        # Handle the episodic data
 
-        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"])
+        # 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"
+
+        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()
 
         # If lang_key is present, convert the entire tensor at once
         if lang_key is not None:
-            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}
+            langs = [str(x) for x in episode[lang_key]]
 
         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}"
@@ -239,6 +262,17 @@ 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"
         )
@@ -256,28 +290,30 @@ def load_from_raw(
 def to_hf_dataset(data_dict, video) -> Dataset:
     features = {}
 
-    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()
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        if video:
+            features[key] = VideoFrame()
+        else:
+            features[key] = Image()
 
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if "observation.velocity" in data_dict:
+        features["observation.velocity"] = Sequence(
+            length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    if "observation.effort" in data_dict:
+        features["observation.effort"] = Sequence(
+            length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
+        )
     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)
@@ -297,8 +333,19 @@ def from_raw_to_lerobot_format(
     video: bool = True,
     episodes: list[int] | None = None,
     encoding: dict | None = None,
+    openx_dataset_name: str | None = None,
 ):
-    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    """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)
     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

@@ -16,9 +16,7 @@
 import inspect
 from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
-from typing import Dict
 
-import datasets
 import numpy
 import PIL
 import torch
@@ -74,58 +72,3 @@ 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

@@ -14,8 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import collections
-from dataclasses import dataclass, field
-from typing import Any, Callable, Sequence
+from typing import Any, Callable, Dict, Sequence
 
 import torch
 from torchvision.transforms import v2
@@ -66,8 +65,6 @@ 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
 
@@ -75,9 +72,9 @@ class RandomSubsetApply(Transform):
         if not self.random_order:
             selected_indices = selected_indices.sort().values
 
-        self.selected_transforms = [self.transforms[i] for i in selected_indices]
+        selected_transforms = [self.transforms[i] for i in selected_indices]
 
-        for transform in self.selected_transforms:
+        for transform in selected_transforms:
             outputs = transform(*inputs)
             inputs = outputs if needs_unpacking else (outputs,)
 
@@ -132,118 +129,69 @@ class SharpnessJitter(Transform):
 
         return float(sharpness[0]), float(sharpness[1])
 
-    def make_params(self, flat_inputs: list[Any]) -> dict[str, Any]:
-        sharpness_factor = torch.empty(1).uniform_(self.sharpness[0], self.sharpness[1]).item()
-        return {"sharpness_factor": sharpness_factor}
+    def _generate_value(self, left: float, right: float) -> float:
+        return torch.empty(1).uniform_(left, right).item()
 
-    def transform(self, inpt: Any, params: dict[str, Any]) -> Any:
-        sharpness_factor = params["sharpness_factor"]
+    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+        sharpness_factor = self._generate_value(self.sharpness[0], self.sharpness[1])
         return self._call_kernel(F.adjust_sharpness, inpt, sharpness_factor=sharpness_factor)
 
 
-@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.
-    """
+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})."
+                )
 
-    weight: float = 1.0
-    type: str = "Identity"
-    kwargs: dict[str, Any] = field(default_factory=dict)
+    check_value("brightness", brightness_weight, brightness_min_max)
+    check_value("contrast", contrast_weight, contrast_min_max)
+    check_value("saturation", saturation_weight, saturation_min_max)
+    check_value("hue", hue_weight, hue_min_max)
+    check_value("sharpness", sharpness_weight, sharpness_min_max)
 
+    weights = []
+    transforms = []
+    if brightness_min_max is not None and brightness_weight > 0.0:
+        weights.append(brightness_weight)
+        transforms.append(v2.ColorJitter(brightness=brightness_min_max))
+    if contrast_min_max is not None and contrast_weight > 0.0:
+        weights.append(contrast_weight)
+        transforms.append(v2.ColorJitter(contrast=contrast_min_max))
+    if saturation_min_max is not None and saturation_weight > 0.0:
+        weights.append(saturation_weight)
+        transforms.append(v2.ColorJitter(saturation=saturation_min_max))
+    if hue_min_max is not None and hue_weight > 0.0:
+        weights.append(hue_weight)
+        transforms.append(v2.ColorJitter(hue=hue_min_max))
+    if sharpness_min_max is not None and sharpness_weight > 0.0:
+        weights.append(sharpness_weight)
+        transforms.append(SharpnessJitter(sharpness=sharpness_min_max))
 
-@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.
-    """
+    n_subset = len(transforms)
+    if max_num_transforms is not None:
+        n_subset = min(n_subset, max_num_transforms)
 
-    # 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)
+    if n_subset == 0:
+        return v2.Identity()
     else:
-        raise ValueError(f"Transform '{cfg.type}' is not valid.")
-
-
-class ImageTransforms(Transform):
-    """A class to compose image transforms based on configuration."""
-
-    def __init__(self, cfg: ImageTransformsConfig) -> None:
-        super().__init__()
-        self._cfg = cfg
-
-        self.weights = []
-        self.transforms = {}
-        for tf_name, tf_cfg in cfg.tfs.items():
-            if tf_cfg.weight <= 0.0:
-                continue
-
-            self.transforms[tf_name] = make_transform_from_config(tf_cfg)
-            self.weights.append(tf_cfg.weight)
-
-        n_subset = min(len(self.transforms), cfg.max_num_transforms)
-        if n_subset == 0 or not cfg.enable:
-            self.tf = v2.Identity()
-        else:
-            self.tf = RandomSubsetApply(
-                transforms=list(self.transforms.values()),
-                p=self.weights,
-                n_subset=n_subset,
-                random_order=cfg.random_order,
-            )
-
-    def forward(self, *inputs: Any) -> Any:
-        return self.tf(*inputs)
+        # TODO(rcadene, aliberts): add v2.ToDtype float16?
+        return RandomSubsetApply(transforms, p=weights, n_subset=n_subset, random_order=random_order)

lerobot/common/datasets/utils.py

@@ -13,61 +13,31 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import importlib.resources
 import json
-import logging
-import textwrap
-from collections.abc import Iterator
-from itertools import accumulate
+import re
+import warnings
+from functools import cache
 from pathlib import Path
-from pprint import pformat
-from types import SimpleNamespace
-from typing import Any
+from typing import Dict
 
 import datasets
-import jsonlines
-import numpy as np
-import pyarrow.compute as pc
 import torch
-from datasets.table import embed_table_storage
-from huggingface_hub import DatasetCard, DatasetCardData, HfApi
+from datasets import load_dataset, load_from_disk
+from huggingface_hub import DatasetCard, HfApi, hf_hub_download, snapshot_download
 from PIL import Image as PILImage
+from safetensors.torch import load_file
 from torchvision import transforms
 
-from lerobot.common.robot_devices.robots.utils import Robot
-from lerobot.configs.types import DictLike, FeatureType, PolicyFeature
-
-DEFAULT_CHUNK_SIZE = 1000  # Max number of episodes per chunk
-
-INFO_PATH = "meta/info.json"
-EPISODES_PATH = "meta/episodes.jsonl"
-STATS_PATH = "meta/stats.json"
-TASKS_PATH = "meta/tasks.jsonl"
-
-DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
-DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
-DEFAULT_IMAGE_PATH = "images/{image_key}/episode_{episode_index:06d}/frame_{frame_index:06d}.png"
-
 DATASET_CARD_TEMPLATE = """
 ---
 # Metadata will go there
 ---
 This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
 
-## {}
-
 """
 
-DEFAULT_FEATURES = {
-    "timestamp": {"dtype": "float32", "shape": (1,), "names": None},
-    "frame_index": {"dtype": "int64", "shape": (1,), "names": None},
-    "episode_index": {"dtype": "int64", "shape": (1,), "names": None},
-    "index": {"dtype": "int64", "shape": (1,), "names": None},
-    "task_index": {"dtype": "int64", "shape": (1,), "names": None},
-}
 
-
-def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
+def flatten_dict(d, parent_key="", sep="/"):
     """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
 
     For example:
@@ -86,7 +56,7 @@ def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
     return dict(items)
 
 
-def unflatten_dict(d: dict, sep: str = "/") -> dict:
+def unflatten_dict(d, sep="/"):
     outdict = {}
     for key, value in d.items():
         parts = key.split(sep)
@@ -99,94 +69,6 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
     return outdict
 
 
-def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
-    split_keys = flattened_key.split(sep)
-    getter = obj[split_keys[0]]
-    if len(split_keys) == 1:
-        return getter
-
-    for key in split_keys[1:]:
-        getter = getter[key]
-
-    return getter
-
-
-def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
-    serialized_dict = {key: value.tolist() for key, value in flatten_dict(stats).items()}
-    return unflatten_dict(serialized_dict)
-
-
-def write_parquet(dataset: datasets.Dataset, fpath: Path) -> None:
-    # Embed image bytes into the table before saving to parquet
-    format = dataset.format
-    dataset = dataset.with_format("arrow")
-    dataset = dataset.map(embed_table_storage, batched=False)
-    dataset = dataset.with_format(**format)
-    dataset.to_parquet(fpath)
-
-
-def load_json(fpath: Path) -> Any:
-    with open(fpath) as f:
-        return json.load(f)
-
-
-def write_json(data: dict, fpath: Path) -> None:
-    fpath.parent.mkdir(exist_ok=True, parents=True)
-    with open(fpath, "w") as f:
-        json.dump(data, f, indent=4, ensure_ascii=False)
-
-
-def load_jsonlines(fpath: Path) -> list[Any]:
-    with jsonlines.open(fpath, "r") as reader:
-        return list(reader)
-
-
-def write_jsonlines(data: dict, fpath: Path) -> None:
-    fpath.parent.mkdir(exist_ok=True, parents=True)
-    with jsonlines.open(fpath, "w") as writer:
-        writer.write_all(data)
-
-
-def append_jsonlines(data: dict, fpath: Path) -> None:
-    fpath.parent.mkdir(exist_ok=True, parents=True)
-    with jsonlines.open(fpath, "a") as writer:
-        writer.write(data)
-
-
-def load_info(local_dir: Path) -> dict:
-    info = load_json(local_dir / INFO_PATH)
-    for ft in info["features"].values():
-        ft["shape"] = tuple(ft["shape"])
-    return info
-
-
-def load_stats(local_dir: Path) -> dict:
-    if not (local_dir / STATS_PATH).exists():
-        return None
-    stats = load_json(local_dir / STATS_PATH)
-    stats = {key: torch.tensor(value) for key, value in flatten_dict(stats).items()}
-    return unflatten_dict(stats)
-
-
-def load_tasks(local_dir: Path) -> dict:
-    tasks = load_jsonlines(local_dir / TASKS_PATH)
-    return {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
-
-
-def load_episodes(local_dir: Path) -> dict:
-    return load_jsonlines(local_dir / EPISODES_PATH)
-
-
-def load_image_as_numpy(fpath: str | Path, dtype="float32", channel_first: bool = True) -> np.ndarray:
-    img = PILImage.open(fpath).convert("RGB")
-    img_array = np.array(img, dtype=dtype)
-    if channel_first:  # (H, W, C) -> (C, H, W)
-        img_array = np.transpose(img_array, (2, 0, 1))
-    if "float" in dtype:
-        img_array /= 255.0
-    return img_array
-
-
 def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     """Get a transform function that convert items from Hugging Face dataset (pyarrow)
     to torch tensors. Importantly, images are converted from PIL, which corresponds to
@@ -198,6 +80,14 @@ def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
         if isinstance(first_item, PILImage.Image):
             to_tensor = transforms.ToTensor()
             items_dict[key] = [to_tensor(img) for img in items_dict[key]]
+        elif isinstance(first_item, str):
+            # TODO (michel-aractingi): add str2embedding via language tokenizer
+            # For now we leave this part up to the user to choose how to address
+            # language conditioned tasks
+            pass
+        elif isinstance(first_item, dict) and "path" in first_item and "timestamp" in first_item:
+            # video frame will be processed downstream
+            pass
         elif first_item is None:
             pass
         else:
@@ -205,67 +95,19 @@ def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     return items_dict
 
 
-def _get_major_minor(version: str) -> tuple[int]:
-    split = version.strip("v").split(".")
-    return int(split[0]), int(split[1])
-
-
-class BackwardCompatibilityError(Exception):
-    def __init__(self, repo_id, version):
-        message = textwrap.dedent(f"""
-            BackwardCompatibilityError: 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).
-        """)
-        super().__init__(message)
-
-
-def check_version_compatibility(
-    repo_id: str, version_to_check: str, current_version: str, enforce_breaking_major: bool = True
-) -> None:
-    current_major, _ = _get_major_minor(current_version)
-    major_to_check, _ = _get_major_minor(version_to_check)
-    if major_to_check < current_major and enforce_breaking_major:
-        raise BackwardCompatibilityError(repo_id, version_to_check)
-    elif float(version_to_check.strip("v")) < float(current_version.strip("v")):
-        logging.warning(
-            f"""The dataset you requested ({repo_id}) was created with a previous version ({version_to_check}) of the
-            codebase. The current codebase version is {current_version}. You should be fine since
-            backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
-            Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
-        )
-
-
-def get_hub_safe_version(repo_id: str, version: str) -> str:
+@cache
+def get_hf_dataset_safe_version(repo_id: str, version: str) -> str:
     api = HfApi()
     dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
     branches = [b.name for b in dataset_info.branches]
     if version not in branches:
-        num_version = float(version.strip("v"))
-        hub_num_versions = [float(v.strip("v")) for v in branches if v.startswith("v")]
-        if num_version >= 2.0 and all(v < 2.0 for v in hub_num_versions):
-            raise BackwardCompatibilityError(repo_id, version)
-
-        logging.warning(
+        warnings.warn(
             f"""You are trying to load a dataset from {repo_id} created with a previous version of the
             codebase. The following versions are available: {branches}.
             The requested version ('{version}') is not found. You should be fine since
             backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
             Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
+            stacklevel=1,
         )
         if "main" not in branches:
             raise ValueError(f"Version 'main' not found on {repo_id}")
@@ -274,215 +116,275 @@ def get_hub_safe_version(repo_id: str, version: str) -> str:
         return version
 
 
-def get_hf_features_from_features(features: dict) -> datasets.Features:
-    hf_features = {}
-    for key, ft in features.items():
-        if ft["dtype"] == "video":
-            continue
-        elif ft["dtype"] == "image":
-            hf_features[key] = datasets.Image()
-        elif ft["shape"] == (1,):
-            hf_features[key] = datasets.Value(dtype=ft["dtype"])
-        else:
-            assert len(ft["shape"]) == 1
-            hf_features[key] = datasets.Sequence(
-                length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
-            )
+def load_hf_dataset(repo_id: str, version: str, root: Path, split: str) -> datasets.Dataset:
+    """hf_dataset contains all the observations, states, actions, rewards, etc."""
+    if root is not None:
+        hf_dataset = load_from_disk(str(Path(root) / repo_id / "train"))
+        # TODO(rcadene): clean this which enables getting a subset of dataset
+        if split != "train":
+            if "%" in split:
+                raise NotImplementedError(f"We dont support splitting based on percentage for now ({split}).")
+            match_from = re.search(r"train\[(\d+):\]", split)
+            match_to = re.search(r"train\[:(\d+)\]", split)
+            if match_from:
+                from_frame_index = int(match_from.group(1))
+                hf_dataset = hf_dataset.select(range(from_frame_index, len(hf_dataset)))
+            elif match_to:
+                to_frame_index = int(match_to.group(1))
+                hf_dataset = hf_dataset.select(range(to_frame_index))
+            else:
+                raise ValueError(
+                    f'`split` ({split}) should either be "train", "train[INT:]", or "train[:INT]"'
+                )
+    else:
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        hf_dataset = load_dataset(repo_id, revision=safe_version, split=split)
 
-    return datasets.Features(hf_features)
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
 
 
-def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
-    camera_ft = {}
-    if robot.cameras:
-        camera_ft = {
-            key: {"dtype": "video" if use_videos else "image", **ft}
-            for key, ft in robot.camera_features.items()
-        }
-    return {**robot.motor_features, **camera_ft, **DEFAULT_FEATURES}
+def load_episode_data_index(repo_id, version, root) -> dict[str, torch.Tensor]:
+    """episode_data_index contains the range of indices for each episode
 
-
-def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFeature]:
-    # TODO(aliberts): Implement "type" in dataset features and simplify this
-    policy_features = {}
-    for key, ft in features.items():
-        shape = ft["shape"]
-        if ft["dtype"] in ["image", "video"]:
-            type = FeatureType.VISUAL
-            if len(shape) != 3:
-                raise ValueError(f"Number of dimensions of {key} != 3 (shape={shape})")
-
-            names = ft["names"]
-            # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
-            if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
-                shape = (shape[2], shape[0], shape[1])
-        elif key == "observation.environment_state":
-            type = FeatureType.ENV
-        elif key.startswith("observation"):
-            type = FeatureType.STATE
-        elif key == "action":
-            type = FeatureType.ACTION
-        else:
-            continue
-
-        policy_features[key] = PolicyFeature(
-            type=type,
-            shape=shape,
+    Example:
+    ```python
+    from_id = episode_data_index["from"][episode_id].item()
+    to_id = episode_data_index["to"][episode_id].item()
+    episode_frames = [dataset[i] for i in range(from_id, to_id)]
+    ```
+    """
+    if root is not None:
+        path = Path(root) / repo_id / "meta_data" / "episode_data_index.safetensors"
+    else:
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        path = hf_hub_download(
+            repo_id, "meta_data/episode_data_index.safetensors", repo_type="dataset", revision=safe_version
         )
 
-    return policy_features
+    return load_file(path)
 
 
-def create_empty_dataset_info(
-    codebase_version: str,
-    fps: int,
-    robot_type: str,
-    features: dict,
-    use_videos: bool,
-) -> dict:
-    return {
-        "codebase_version": codebase_version,
-        "robot_type": robot_type,
-        "total_episodes": 0,
-        "total_frames": 0,
-        "total_tasks": 0,
-        "total_videos": 0,
-        "total_chunks": 0,
-        "chunks_size": DEFAULT_CHUNK_SIZE,
-        "fps": fps,
-        "splits": {},
-        "data_path": DEFAULT_PARQUET_PATH,
-        "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
-        "features": features,
-    }
+def load_stats(repo_id, version, root) -> dict[str, dict[str, torch.Tensor]]:
+    """stats contains the statistics per modality computed over the full dataset, such as max, min, mean, std
+
+    Example:
+    ```python
+    normalized_action = (action - stats["action"]["mean"]) / stats["action"]["std"]
+    ```
+    """
+    if root is not None:
+        path = Path(root) / repo_id / "meta_data" / "stats.safetensors"
+    else:
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        path = hf_hub_download(
+            repo_id, "meta_data/stats.safetensors", repo_type="dataset", revision=safe_version
+        )
+
+    stats = load_file(path)
+    return unflatten_dict(stats)
 
 
-def get_episode_data_index(
-    episode_dicts: list[dict], episodes: list[int] | None = None
-) -> dict[str, torch.Tensor]:
-    episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in enumerate(episode_dicts)}
-    if episodes is not None:
-        episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
+def load_info(repo_id, version, root) -> dict:
+    """info contains useful information regarding the dataset that are not stored elsewhere
 
-    cumulative_lenghts = list(accumulate(episode_lengths.values()))
-    return {
-        "from": torch.LongTensor([0] + cumulative_lenghts[:-1]),
-        "to": torch.LongTensor(cumulative_lenghts),
-    }
+    Example:
+    ```python
+    print("frame per second used to collect the video", info["fps"])
+    ```
+    """
+    if root is not None:
+        path = Path(root) / repo_id / "meta_data" / "info.json"
+    else:
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        path = hf_hub_download(repo_id, "meta_data/info.json", repo_type="dataset", revision=safe_version)
+
+    with open(path) as f:
+        info = json.load(f)
+    return info
 
 
-def calculate_total_episode(
-    hf_dataset: datasets.Dataset, raise_if_not_contiguous: bool = True
-) -> dict[str, torch.Tensor]:
-    episode_indices = sorted(hf_dataset.unique("episode_index"))
-    total_episodes = len(episode_indices)
-    if raise_if_not_contiguous and episode_indices != list(range(total_episodes)):
-        raise ValueError("episode_index values are not sorted and contiguous.")
-    return total_episodes
+def load_videos(repo_id, version, root) -> Path:
+    if root is not None:
+        path = Path(root) / repo_id / "videos"
+    else:
+        # TODO(rcadene): we download the whole repo here. see if we can avoid this
+        safe_version = get_hf_dataset_safe_version(repo_id, version)
+        repo_dir = snapshot_download(repo_id, repo_type="dataset", revision=safe_version)
+        path = Path(repo_dir) / "videos"
+
+    return path
 
 
-def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> dict[str, torch.Tensor]:
-    episode_lengths = []
-    table = hf_dataset.data.table
-    total_episodes = calculate_total_episode(hf_dataset)
-    for ep_idx in range(total_episodes):
-        ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
-        episode_lengths.insert(ep_idx, len(ep_table))
-
-    cumulative_lenghts = list(accumulate(episode_lengths))
-    return {
-        "from": torch.LongTensor([0] + cumulative_lenghts[:-1]),
-        "to": torch.LongTensor(cumulative_lenghts),
-    }
-
-
-def check_timestamps_sync(
+def load_previous_and_future_frames(
+    item: dict[str, torch.Tensor],
     hf_dataset: datasets.Dataset,
     episode_data_index: dict[str, torch.Tensor],
-    fps: int,
+    delta_timestamps: dict[str, list[float]],
     tolerance_s: float,
-    raise_value_error: bool = True,
-) -> bool:
+) -> dict[torch.Tensor]:
     """
-    This check is to make sure that each timestamps is separated to the next by 1/fps +/- tolerance to
-    account for possible numerical error.
+    Given a current item in the dataset containing a timestamp (e.g. 0.6 seconds), and a list of time differences of
+    some modalities (e.g. delta_timestamps={"observation.image": [-0.8, -0.2, 0, 0.2]}), this function computes for each
+    given modality (e.g. "observation.image") a list of query timestamps (e.g. [-0.2, 0.4, 0.6, 0.8]) and loads the closest
+    frames in the dataset.
+
+    Importantly, when no frame can be found around a query timestamp within a specified tolerance window, this function
+    raises an AssertionError. When a timestamp is queried before the first available timestamp of the episode or after
+    the last available timestamp, the violation of the tolerance doesnt raise an AssertionError, and the function
+    populates a boolean array indicating which frames are outside of the episode range. For instance, this boolean array
+    is useful during batched training to not supervise actions associated to timestamps coming after the end of the
+    episode, or to pad the observations in a specific way. Note that by default the observation frames before the start
+    of the episode are the same as the first frame of the episode.
+
+    Parameters:
+    - item (dict): A dictionary containing all the data related to a frame. It is the result of `dataset[idx]`. Each key
+      corresponds to a different modality (e.g., "timestamp", "observation.image", "action").
+    - hf_dataset (datasets.Dataset): A dictionary containing the full dataset. Each key corresponds to a different
+      modality (e.g., "timestamp", "observation.image", "action").
+    - episode_data_index (dict): A dictionary containing two keys ("from" and "to") associated to dataset indices.
+      They indicate the start index and end index of each episode in the dataset.
+    - delta_timestamps (dict): A dictionary containing lists of delta timestamps for each possible modality to be
+      retrieved. These deltas are added to the item timestamp to form the query timestamps.
+    - tolerance_s (float, optional): The tolerance level (in seconds) used to determine if a data point is close enough to the query
+      timestamp by asserting `tol > difference`. It is suggested to set `tol` to a smaller value than the
+      smallest expected inter-frame period, but large enough to account for jitter.
+
+    Returns:
+    - The same item with the queried frames for each modality specified in delta_timestamps, with an additional key for
+      each modality (e.g. "observation.image_is_pad").
+
+    Raises:
+    - AssertionError: If any of the frames unexpectedly violate the tolerance level. This could indicate synchronization
+      issues with timestamps during data collection.
     """
-    timestamps = torch.stack(hf_dataset["timestamp"])
-    diffs = torch.diff(timestamps)
-    within_tolerance = torch.abs(diffs - 1 / fps) <= tolerance_s
+    # get indices of the frames associated to the episode, and their timestamps
+    ep_id = item["episode_index"].item()
+    ep_data_id_from = episode_data_index["from"][ep_id].item()
+    ep_data_id_to = episode_data_index["to"][ep_id].item()
+    ep_data_ids = torch.arange(ep_data_id_from, ep_data_id_to, 1)
 
-    # We mask differences between the timestamp at the end of an episode
-    # and the one at the start of the next episode since these are expected
-    # to be outside tolerance.
-    mask = torch.ones(len(diffs), dtype=torch.bool)
-    ignored_diffs = episode_data_index["to"][:-1] - 1
-    mask[ignored_diffs] = False
-    filtered_within_tolerance = within_tolerance[mask]
+    # load timestamps
+    ep_timestamps = hf_dataset.select_columns("timestamp")[ep_data_id_from:ep_data_id_to]["timestamp"]
+    ep_timestamps = torch.stack(ep_timestamps)
 
-    if not torch.all(filtered_within_tolerance):
-        # Track original indices before masking
-        original_indices = torch.arange(len(diffs))
-        filtered_indices = original_indices[mask]
-        outside_tolerance_filtered_indices = torch.nonzero(~filtered_within_tolerance)  # .squeeze()
-        outside_tolerance_indices = filtered_indices[outside_tolerance_filtered_indices]
-        episode_indices = torch.stack(hf_dataset["episode_index"])
+    # we make the assumption that the timestamps are sorted
+    ep_first_ts = ep_timestamps[0]
+    ep_last_ts = ep_timestamps[-1]
+    current_ts = item["timestamp"].item()
 
-        outside_tolerances = []
-        for idx in outside_tolerance_indices:
-            entry = {
-                "timestamps": [timestamps[idx], timestamps[idx + 1]],
-                "diff": diffs[idx],
-                "episode_index": episode_indices[idx].item(),
-            }
-            outside_tolerances.append(entry)
+    for key in delta_timestamps:
+        # get timestamps used as query to retrieve data of previous/future frames
+        delta_ts = delta_timestamps[key]
+        query_ts = current_ts + torch.tensor(delta_ts)
 
-        if raise_value_error:
-            raise ValueError(
-                f"""One or several timestamps unexpectedly violate the tolerance inside episode range.
-                This might be due to synchronization issues with timestamps during data collection.
-                \n{pformat(outside_tolerances)}"""
-            )
-        return False
+        # compute distances between each query timestamp and all timestamps of all the frames belonging to the episode
+        dist = torch.cdist(query_ts[:, None], ep_timestamps[:, None], p=1)
+        min_, argmin_ = dist.min(1)
 
-    return True
+        # TODO(rcadene): synchronize timestamps + interpolation if needed
+
+        is_pad = min_ > tolerance_s
+
+        # check violated query timestamps are all outside the episode range
+        assert ((query_ts[is_pad] < ep_first_ts) | (ep_last_ts < query_ts[is_pad])).all(), (
+            f"One or several timestamps unexpectedly violate the tolerance ({min_} > {tolerance_s=}) inside episode range."
+            "This might be due to synchronization issues with timestamps during data collection."
+        )
+
+        # get dataset indices corresponding to frames to be loaded
+        data_ids = ep_data_ids[argmin_]
+
+        # load frames modality
+        item[key] = hf_dataset.select_columns(key)[data_ids][key]
+
+        if isinstance(item[key][0], dict) and "path" in item[key][0]:
+            # video mode where frame are expressed as dict of path and timestamp
+            item[key] = item[key]
+        else:
+            item[key] = torch.stack(item[key])
+
+        item[f"{key}_is_pad"] = is_pad
+
+    return item
 
 
-def check_delta_timestamps(
-    delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
-) -> bool:
-    """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
-    This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be
-    actual timestamps from the dataset.
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
     """
-    outside_tolerance = {}
-    for key, delta_ts in delta_timestamps.items():
-        within_tolerance = [abs(ts * fps - round(ts * fps)) / fps <= tolerance_s for ts in delta_ts]
-        if not all(within_tolerance):
-            outside_tolerance[key] = [
-                ts for ts, is_within in zip(delta_ts, within_tolerance, strict=True) if not is_within
-            ]
+    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
 
-    if len(outside_tolerance) > 0:
-        if raise_value_error:
-            raise ValueError(
-                f"""
-                The following delta_timestamps are found outside of tolerance range.
-                Please make sure they are multiples of 1/{fps} +/- tolerance and adjust
-                their values accordingly.
-                \n{pformat(outside_tolerance)}
-                """
-            )
-        return False
+    Parameters:
+    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
 
-    return True
+    Returns:
+    - episode_data_index: A dictionary containing the data index for each episode. The dictionary has two keys:
+        - "from": A tensor containing the starting index of each episode.
+        - "to": A tensor containing the ending index of each episode.
+    """
+    episode_data_index = {"from": [], "to": []}
+
+    current_episode = None
+    """
+    The episode_index is a list of integers, each representing the episode index of the corresponding example.
+    For instance, the following is a valid episode_index:
+      [0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2]
+
+    Below, we iterate through the episode_index and populate the episode_data_index dictionary with the starting and
+    ending index of each episode. For the episode_index above, the episode_data_index dictionary will look like this:
+        {
+            "from": [0, 3, 7],
+            "to": [3, 7, 12]
+        }
+    """
+    if len(hf_dataset) == 0:
+        episode_data_index = {
+            "from": torch.tensor([]),
+            "to": torch.tensor([]),
+        }
+        return episode_data_index
+    for idx, episode_idx in enumerate(hf_dataset["episode_index"]):
+        if episode_idx != current_episode:
+            # We encountered a new episode, so we append its starting location to the "from" list
+            episode_data_index["from"].append(idx)
+            # If this is not the first episode, we append the ending location of the previous episode to the "to" list
+            if current_episode is not None:
+                episode_data_index["to"].append(idx)
+            # Let's keep track of the current episode index
+            current_episode = episode_idx
+        else:
+            # We are still in the same episode, so there is nothing for us to do here
+            pass
+    # We have reached the end of the dataset, so we append the ending location of the last episode to the "to" list
+    episode_data_index["to"].append(idx + 1)
+
+    for k in ["from", "to"]:
+        episode_data_index[k] = torch.tensor(episode_data_index[k])
+
+    return episode_data_index
 
 
-def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
-    delta_indices = {}
-    for key, delta_ts in delta_timestamps.items():
-        delta_indices[key] = [round(d * fps) for d in delta_ts]
+def reset_episode_index(hf_dataset: datasets.Dataset) -> datasets.Dataset:
+    """Reset the `episode_index` of the provided HuggingFace Dataset.
 
-    return delta_indices
+    `episode_data_index` (and related functionality such as `load_previous_and_future_frames`) requires the
+    `episode_index` to be sorted, continuous (1,1,1 and not 1,2,1) and start at 0.
+
+    This brings the `episode_index` to the required format.
+    """
+    if len(hf_dataset) == 0:
+        return hf_dataset
+    unique_episode_idxs = torch.stack(hf_dataset["episode_index"]).unique().tolist()
+    episode_idx_to_reset_idx_mapping = {
+        ep_id: reset_ep_id for reset_ep_id, ep_id in enumerate(unique_episode_idxs)
+    }
+
+    def modify_ep_idx_func(example):
+        example["episode_index"] = episode_idx_to_reset_idx_mapping[example["episode_index"].item()]
+        return example
+
+    hf_dataset = hf_dataset.map(modify_ep_idx_func)
+
+    return hf_dataset
 
 
 def cycle(iterable):
@@ -498,7 +400,7 @@ def cycle(iterable):
             iterator = iter(iterable)
 
 
-def create_branch(repo_id, *, branch: str, repo_type: str | None = None) -> None:
+def create_branch(repo_id, *, branch: str, repo_type: str | None = None):
     """Create a branch on a existing Hugging Face repo. Delete the branch if it already
     exists before creating it.
     """
@@ -513,94 +415,12 @@ def create_branch(repo_id, *, branch: str, repo_type: str | None = None) -> None
     api.create_branch(repo_id, repo_type=repo_type, branch=branch)
 
 
-def create_lerobot_dataset_card(
-    tags: list | None = None,
-    dataset_info: dict | None = None,
-    **kwargs,
-) -> DatasetCard:
-    """
-    Keyword arguments will be used to replace values in ./lerobot/common/datasets/card_template.md.
-    Note: If specified, license must be one of https://huggingface.co/docs/hub/repositories-licenses.
-    """
-    card_tags = ["LeRobot"]
-
-    if tags:
-        card_tags += tags
-    if dataset_info:
-        dataset_structure = "[meta/info.json](meta/info.json):\n"
-        dataset_structure += f"```json\n{json.dumps(dataset_info, indent=4)}\n```\n"
-        kwargs = {**kwargs, "dataset_structure": dataset_structure}
-    card_data = DatasetCardData(
-        license=kwargs.get("license"),
-        tags=card_tags,
-        task_categories=["robotics"],
-        configs=[
-            {
-                "config_name": "default",
-                "data_files": "data/*/*.parquet",
-            }
-        ],
-    )
-
-    card_template = (importlib.resources.files("lerobot.common.datasets") / "card_template.md").read_text()
-
-    return DatasetCard.from_template(
-        card_data=card_data,
-        template_str=card_template,
-        **kwargs,
-    )
-
-
-class IterableNamespace(SimpleNamespace):
-    """
-    A namespace object that supports both dictionary-like iteration and dot notation access.
-    Automatically converts nested dictionaries into IterableNamespaces.
-
-    This class extends SimpleNamespace to provide:
-    - Dictionary-style iteration over keys
-    - Access to items via both dot notation (obj.key) and brackets (obj["key"])
-    - Dictionary-like methods: items(), keys(), values()
-    - Recursive conversion of nested dictionaries
-
-    Args:
-        dictionary: Optional dictionary to initialize the namespace
-        **kwargs: Additional keyword arguments passed to SimpleNamespace
-
-    Examples:
-        >>> data = {"name": "Alice", "details": {"age": 25}}
-        >>> ns = IterableNamespace(data)
-        >>> ns.name
-        'Alice'
-        >>> ns.details.age
-        25
-        >>> list(ns.keys())
-        ['name', 'details']
-        >>> for key, value in ns.items():
-        ...     print(f"{key}: {value}")
-        name: Alice
-        details: IterableNamespace(age=25)
-    """
-
-    def __init__(self, dictionary: dict[str, Any] = None, **kwargs):
-        super().__init__(**kwargs)
-        if dictionary is not None:
-            for key, value in dictionary.items():
-                if isinstance(value, dict):
-                    setattr(self, key, IterableNamespace(value))
-                else:
-                    setattr(self, key, value)
-
-    def __iter__(self) -> Iterator[str]:
-        return iter(vars(self))
-
-    def __getitem__(self, key: str) -> Any:
-        return vars(self)[key]
-
-    def items(self):
-        return vars(self).items()
-
-    def values(self):
-        return vars(self).values()
-
-    def keys(self):
-        return vars(self).keys()
+def create_lerobot_dataset_card(tags: list | None = None, text: str | None = None) -> DatasetCard:
+    card = DatasetCard(DATASET_CARD_TEMPLATE)
+    card.data.task_categories = ["robotics"]
+    card.data.tags = ["LeRobot"]
+    if tags is not None:
+        card.data.tags += tags
+    if text is not None:
+        card.text += text
+    return card

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -1,882 +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.
-
-"""
-This script is for internal use to convert all datasets under the 'lerobot' hub user account to v2.
-
-Note: Since the original Aloha datasets don't use shadow motors, you need to comment those out in
-lerobot/configs/robot/aloha.yaml before running this script.
-"""
-
-import traceback
-from pathlib import Path
-from textwrap import dedent
-
-from lerobot import available_datasets
-from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset
-from lerobot.common.robot_devices.robots.configs import AlohaRobotConfig
-
-LOCAL_DIR = Path("data/")
-
-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(),
-    },
-}
-
-
-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

@@ -1,664 +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.
-
-"""
-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-conditionning.
-
-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_hub_safe_version,
-    load_json,
-    unflatten_dict,
-    write_json,
-    write_jsonlines,
-)
-from lerobot.common.datasets.video_utils import (
-    VideoFrame,  # noqa: F401
-    get_image_pixel_channels,
-    get_video_info,
-)
-from lerobot.common.robot_devices.robots.configs import RobotConfig
-from lerobot.common.robot_devices.robots.utils import make_robot_config
-
-V16 = "v1.6"
-V20 = "v2.0"
-
-GITATTRIBUTES_REF = "aliberts/gitattributes_reference"
-V1_VIDEO_FILE = "{video_key}_episode_{episode_index:06d}.mp4"
-V1_INFO_PATH = "meta_data/info.json"
-V1_STATS_PATH = "meta_data/stats.safetensors"
-
-
-def parse_robot_config(robot_cfg: RobotConfig) -> tuple[str, dict]:
-    if robot_cfg.type in ["aloha", "koch"]:
-        state_names = [
-            f"{arm}_{motor}" if len(robot_cfg.follower_arms) > 1 else motor
-            for arm in robot_cfg.follower_arms
-            for motor in robot_cfg.follower_arms[arm].motors
-        ]
-        action_names = [
-            # f"{arm}_{motor}" for arm in ["left", "right"] for motor in robot_cfg["leader_arms"][arm]["motors"]
-            f"{arm}_{motor}" if len(robot_cfg.leader_arms) > 1 else motor
-            for arm in robot_cfg.leader_arms
-            for motor in robot_cfg.leader_arms[arm].motors
-        ]
-    # elif robot_cfg["robot_type"] == "stretch3": TODO
-    else:
-        raise NotImplementedError(
-            "Please provide robot_config={'robot_type': ..., 'names': ...} directly to convert_dataset()."
-        )
-
-    return {
-        "robot_type": robot_cfg.type,
-        "names": {
-            "observation.state": state_names,
-            "observation.effort": state_names,
-            "action": action_names,
-        },
-    }
-
-
-def convert_stats_to_json(v1_dir: Path, v2_dir: Path) -> None:
-    safetensor_path = v1_dir / V1_STATS_PATH
-    stats = load_file(safetensor_path)
-    serialized_stats = {key: value.tolist() for key, value in stats.items()}
-    serialized_stats = unflatten_dict(serialized_stats)
-
-    json_path = v2_dir / STATS_PATH
-    json_path.parent.mkdir(exist_ok=True, parents=True)
-    with open(json_path, "w") as f:
-        json.dump(serialized_stats, f, indent=4)
-
-    # Sanity check
-    with open(json_path) as f:
-        stats_json = json.load(f)
-
-    stats_json = flatten_dict(stats_json)
-    stats_json = {key: torch.tensor(value) for key, value in stats_json.items()}
-    for key in stats:
-        torch.testing.assert_close(stats_json[key], stats[key])
-
-
-def get_features_from_hf_dataset(
-    dataset: Dataset, robot_config: RobotConfig | None = None
-) -> dict[str, list]:
-    robot_config = parse_robot_config(robot_config)
-    features = {}
-    for key, ft in dataset.features.items():
-        if isinstance(ft, datasets.Value):
-            dtype = ft.dtype
-            shape = (1,)
-            names = None
-        if isinstance(ft, datasets.Sequence):
-            assert isinstance(ft.feature, datasets.Value)
-            dtype = ft.feature.dtype
-            shape = (ft.length,)
-            motor_names = (
-                robot_config["names"][key] if robot_config else [f"motor_{i}" for i in range(ft.length)]
-            )
-            assert len(motor_names) == shape[0]
-            names = {"motors": motor_names}
-        elif isinstance(ft, datasets.Image):
-            dtype = "image"
-            image = dataset[0][key]  # Assuming first row
-            channels = get_image_pixel_channels(image)
-            shape = (image.height, image.width, channels)
-            names = ["height", "width", "channels"]
-        elif ft._type == "VideoFrame":
-            dtype = "video"
-            shape = None  # Add shape later
-            names = ["height", "width", "channels"]
-
-        features[key] = {
-            "dtype": dtype,
-            "shape": shape,
-            "names": names,
-        }
-
-    return features
-
-
-def add_task_index_by_episodes(dataset: Dataset, tasks_by_episodes: dict) -> tuple[Dataset, list[str]]:
-    df = dataset.to_pandas()
-    tasks = list(set(tasks_by_episodes.values()))
-    tasks_to_task_index = {task: task_idx for task_idx, task in enumerate(tasks)}
-    episodes_to_task_index = {ep_idx: tasks_to_task_index[task] for ep_idx, task in tasks_by_episodes.items()}
-    df["task_index"] = df["episode_index"].map(episodes_to_task_index).astype(int)
-
-    features = dataset.features
-    features["task_index"] = datasets.Value(dtype="int64")
-    dataset = Dataset.from_pandas(df, features=features, split="train")
-    return dataset, tasks
-
-
-def add_task_index_from_tasks_col(
-    dataset: Dataset, tasks_col: str
-) -> tuple[Dataset, dict[str, list[str]], list[str]]:
-    df = dataset.to_pandas()
-
-    # HACK: This is to clean some of the instructions in our version of Open X datasets
-    prefix_to_clean = "tf.Tensor(b'"
-    suffix_to_clean = "', shape=(), dtype=string)"
-    df[tasks_col] = df[tasks_col].str.removeprefix(prefix_to_clean).str.removesuffix(suffix_to_clean)
-
-    # Create task_index col
-    tasks_by_episode = df.groupby("episode_index")[tasks_col].unique().apply(lambda x: x.tolist()).to_dict()
-    tasks = df[tasks_col].unique().tolist()
-    tasks_to_task_index = {task: idx for idx, task in enumerate(tasks)}
-    df["task_index"] = df[tasks_col].map(tasks_to_task_index).astype(int)
-
-    # Build the dataset back from df
-    features = dataset.features
-    features["task_index"] = datasets.Value(dtype="int64")
-    dataset = Dataset.from_pandas(df, features=features, split="train")
-    dataset = dataset.remove_columns(tasks_col)
-
-    return dataset, tasks, tasks_by_episode
-
-
-def split_parquet_by_episodes(
-    dataset: Dataset,
-    total_episodes: int,
-    total_chunks: int,
-    output_dir: Path,
-) -> list:
-    table = dataset.data.table
-    episode_lengths = []
-    for ep_chunk in range(total_chunks):
-        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
-        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
-        chunk_dir = "/".join(DEFAULT_PARQUET_PATH.split("/")[:-1]).format(episode_chunk=ep_chunk)
-        (output_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
-        for ep_idx in range(ep_chunk_start, ep_chunk_end):
-            ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
-            episode_lengths.insert(ep_idx, len(ep_table))
-            output_file = output_dir / DEFAULT_PARQUET_PATH.format(
-                episode_chunk=ep_chunk, episode_index=ep_idx
-            )
-            pq.write_table(ep_table, output_file)
-
-    return episode_lengths
-
-
-def move_videos(
-    repo_id: str,
-    video_keys: list[str],
-    total_episodes: int,
-    total_chunks: int,
-    work_dir: Path,
-    clean_gittatributes: Path,
-    branch: str = "main",
-) -> None:
-    """
-    HACK: Since HfApi() doesn't provide a way to move files directly in a repo, this function will run git
-    commands to fetch git lfs video files references to move them into subdirectories without having to
-    actually download them.
-    """
-    _lfs_clone(repo_id, work_dir, branch)
-
-    videos_moved = False
-    video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*.mp4")]
-    if len(video_files) == 0:
-        video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*/*/*.mp4")]
-        videos_moved = True  # Videos have already been moved
-
-    assert len(video_files) == total_episodes * len(video_keys)
-
-    lfs_untracked_videos = _get_lfs_untracked_videos(work_dir, video_files)
-
-    current_gittatributes = work_dir / ".gitattributes"
-    if not filecmp.cmp(current_gittatributes, clean_gittatributes, shallow=False):
-        fix_gitattributes(work_dir, current_gittatributes, clean_gittatributes)
-
-    if lfs_untracked_videos:
-        fix_lfs_video_files_tracking(work_dir, video_files)
-
-    if videos_moved:
-        return
-
-    video_dirs = sorted(work_dir.glob("videos*/"))
-    for ep_chunk in range(total_chunks):
-        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
-        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
-        for vid_key in video_keys:
-            chunk_dir = "/".join(DEFAULT_VIDEO_PATH.split("/")[:-1]).format(
-                episode_chunk=ep_chunk, video_key=vid_key
-            )
-            (work_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
-
-            for ep_idx in range(ep_chunk_start, ep_chunk_end):
-                target_path = DEFAULT_VIDEO_PATH.format(
-                    episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_idx
-                )
-                video_file = V1_VIDEO_FILE.format(video_key=vid_key, episode_index=ep_idx)
-                if len(video_dirs) == 1:
-                    video_path = video_dirs[0] / video_file
-                else:
-                    for dir in video_dirs:
-                        if (dir / video_file).is_file():
-                            video_path = dir / video_file
-                            break
-
-                video_path.rename(work_dir / target_path)
-
-    commit_message = "Move video files into chunk subdirectories"
-    subprocess.run(["git", "add", "."], cwd=work_dir, check=True)
-    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
-    subprocess.run(["git", "push"], cwd=work_dir, check=True)
-
-
-def fix_lfs_video_files_tracking(work_dir: Path, lfs_untracked_videos: list[str]) -> None:
-    """
-    HACK: This function fixes the tracking by git lfs which was not properly set on some repos. In that case,
-    there's no other option than to download the actual files and reupload them with lfs tracking.
-    """
-    for i in range(0, len(lfs_untracked_videos), 100):
-        files = lfs_untracked_videos[i : i + 100]
-        try:
-            subprocess.run(["git", "rm", "--cached", *files], cwd=work_dir, capture_output=True, check=True)
-        except subprocess.CalledProcessError as e:
-            print("git rm --cached ERROR:")
-            print(e.stderr)
-        subprocess.run(["git", "add", *files], cwd=work_dir, check=True)
-
-    commit_message = "Track video files with git lfs"
-    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
-    subprocess.run(["git", "push"], cwd=work_dir, check=True)
-
-
-def fix_gitattributes(work_dir: Path, current_gittatributes: Path, clean_gittatributes: Path) -> None:
-    shutil.copyfile(clean_gittatributes, current_gittatributes)
-    subprocess.run(["git", "add", ".gitattributes"], cwd=work_dir, check=True)
-    subprocess.run(["git", "commit", "-m", "Fix .gitattributes"], cwd=work_dir, check=True)
-    subprocess.run(["git", "push"], cwd=work_dir, check=True)
-
-
-def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:
-    subprocess.run(["git", "lfs", "install"], cwd=work_dir, check=True)
-    repo_url = f"https://huggingface.co/datasets/{repo_id}"
-    env = {"GIT_LFS_SKIP_SMUDGE": "1"}  # Prevent downloading LFS files
-    subprocess.run(
-        ["git", "clone", "--branch", branch, "--single-branch", "--depth", "1", repo_url, str(work_dir)],
-        check=True,
-        env=env,
-    )
-
-
-def _get_lfs_untracked_videos(work_dir: Path, video_files: list[str]) -> list[str]:
-    lfs_tracked_files = subprocess.run(
-        ["git", "lfs", "ls-files", "-n"], cwd=work_dir, capture_output=True, text=True, check=True
-    )
-    lfs_tracked_files = set(lfs_tracked_files.stdout.splitlines())
-    return [f for f in video_files if f not in lfs_tracked_files]
-
-
-def get_videos_info(repo_id: str, local_dir: Path, video_keys: list[str], branch: str) -> dict:
-    # Assumes first episode
-    video_files = [
-        DEFAULT_VIDEO_PATH.format(episode_chunk=0, video_key=vid_key, episode_index=0)
-        for vid_key in video_keys
-    ]
-    hub_api = HfApi()
-    hub_api.snapshot_download(
-        repo_id=repo_id, repo_type="dataset", local_dir=local_dir, revision=branch, allow_patterns=video_files
-    )
-    videos_info_dict = {}
-    for vid_key, vid_path in zip(video_keys, video_files, strict=True):
-        videos_info_dict[vid_key] = get_video_info(local_dir / vid_path)
-
-    return videos_info_dict
-
-
-def convert_dataset(
-    repo_id: str,
-    local_dir: Path,
-    single_task: str | None = None,
-    tasks_path: Path | None = None,
-    tasks_col: Path | None = None,
-    robot_config: RobotConfig | None = None,
-    test_branch: str | None = None,
-    **card_kwargs,
-):
-    v1 = get_hub_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/video_utils.py

@@ -13,7 +13,6 @@
 # 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
@@ -26,11 +25,47 @@ import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
-from PIL import Image
+
+
+def load_from_videos(
+    item: dict[str, torch.Tensor],
+    video_frame_keys: list[str],
+    videos_dir: Path,
+    tolerance_s: float,
+    backend: str = "pyav",
+):
+    """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
+    in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a Segmentation Fault.
+    This probably happens because a memory reference to the video loader is created in the main process and a
+    subprocess fails to access it.
+    """
+    # since video path already contains "videos" (e.g. videos_dir="data/videos", path="videos/episode_0.mp4")
+    data_dir = videos_dir.parent
+
+    for key in video_frame_keys:
+        if isinstance(item[key], list):
+            # load multiple frames at once (expected when delta_timestamps is not None)
+            timestamps = [frame["timestamp"] for frame in item[key]]
+            paths = [frame["path"] for frame in item[key]]
+            if len(set(paths)) > 1:
+                raise NotImplementedError("All video paths are expected to be the same for now.")
+            video_path = data_dir / paths[0]
+
+            frames = decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
+            item[key] = frames
+        else:
+            # load one frame
+            timestamps = [item[key]["timestamp"]]
+            video_path = data_dir / item[key]["path"]
+
+            frames = decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
+            item[key] = frames[0]
+
+    return item
 
 
 def decode_video_frames_torchvision(
-    video_path: Path | str,
+    video_path: str,
     timestamps: list[float],
     tolerance_s: float,
     backend: str = "pyav",
@@ -128,8 +163,8 @@ def decode_video_frames_torchvision(
 
 
 def encode_video_frames(
-    imgs_dir: Path | str,
-    video_path: Path | str,
+    imgs_dir: Path,
+    video_path: Path,
     fps: int,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
@@ -212,104 +247,3 @@ 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

@@ -1 +0,0 @@
-from .configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv  # noqa: F401

lerobot/common/envs/configs.py

@@ -1,142 +0,0 @@
-import abc
-from dataclasses import dataclass, field
-
-import draccus
-
-from lerobot.common.constants import ACTION, OBS_ENV, OBS_IMAGE, OBS_IMAGES, OBS_ROBOT
-from lerobot.configs.types import FeatureType, PolicyFeature
-
-
-@dataclass
-class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
-    task: str | None = None
-    fps: int = 30
-    features: dict[str, PolicyFeature] = field(default_factory=dict)
-    features_map: dict[str, str] = field(default_factory=dict)
-
-    @property
-    def type(self) -> str:
-        return self.get_choice_name(self.__class__)
-
-    @abc.abstractproperty
-    def gym_kwargs(self) -> dict:
-        raise NotImplementedError()
-
-
-@EnvConfig.register_subclass("aloha")
-@dataclass
-class AlohaEnv(EnvConfig):
-    task: str = "AlohaInsertion-v0"
-    fps: int = 50
-    episode_length: int = 400
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": ACTION,
-            "agent_pos": OBS_ROBOT,
-            "top": f"{OBS_IMAGE}.top",
-            "pixels/top": f"{OBS_IMAGES}.top",
-        }
-    )
-
-    def __post_init__(self):
-        if self.obs_type == "pixels":
-            self.features["top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
-        elif self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(14,))
-            self.features["pixels/top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "max_episode_steps": self.episode_length,
-        }
-
-
-@EnvConfig.register_subclass("pusht")
-@dataclass
-class PushtEnv(EnvConfig):
-    task: str = "PushT-v0"
-    fps: int = 10
-    episode_length: int = 300
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    visualization_width: int = 384
-    visualization_height: int = 384
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(2,)),
-            "agent_pos": PolicyFeature(type=FeatureType.STATE, shape=(2,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": ACTION,
-            "agent_pos": OBS_ROBOT,
-            "environment_state": OBS_ENV,
-            "pixels": OBS_IMAGE,
-        }
-    )
-
-    def __post_init__(self):
-        if self.obs_type == "pixels_agent_pos":
-            self.features["pixels"] = PolicyFeature(type=FeatureType.VISUAL, shape=(384, 384, 3))
-        elif self.obs_type == "environment_state_agent_pos":
-            self.features["environment_state"] = PolicyFeature(type=FeatureType.ENV, shape=(16,))
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "visualization_width": self.visualization_width,
-            "visualization_height": self.visualization_height,
-            "max_episode_steps": self.episode_length,
-        }
-
-
-@EnvConfig.register_subclass("xarm")
-@dataclass
-class XarmEnv(EnvConfig):
-    task: str = "XarmLift-v0"
-    fps: int = 15
-    episode_length: int = 200
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    visualization_width: int = 384
-    visualization_height: int = 384
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
-            "pixels": PolicyFeature(type=FeatureType.VISUAL, shape=(84, 84, 3)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": ACTION,
-            "agent_pos": OBS_ROBOT,
-            "pixels": OBS_IMAGE,
-        }
-    )
-
-    def __post_init__(self):
-        if self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(4,))
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "visualization_width": self.visualization_width,
-            "visualization_height": self.visualization_height,
-            "max_episode_steps": self.episode_length,
-        }

lerobot/common/envs/factory.py

@@ -16,54 +16,43 @@
 import importlib
 
 import gymnasium as gym
-
-from lerobot.common.envs.configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv
+from omegaconf import DictConfig
 
 
-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.")
+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(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): Wether to return an AsyncVectorEnv or a SyncVectorEnv. Defaults to
-            False.
-
-    Raises:
-        ValueError: if n_envs < 1
-        ModuleNotFoundError: If the requested env package is not intalled
-
-    Returns:
-        gym.vector.VectorEnv: The parallelized gym.env instance.
+    n_envs can be used to override eval.batch_size in the configuration. Must be at least 1.
     """
-    if n_envs < 1:
+    if n_envs is not None and n_envs < 1:
         raise ValueError("`n_envs must be at least 1")
 
-    package_name = f"gym_{cfg.type}"
+    if cfg.env.name == "real_world":
+        return
+
+    package_name = f"gym_{cfg.env.name}"
 
     try:
         importlib.import_module(package_name)
     except ModuleNotFoundError as e:
-        print(f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.type}]'`")
+        print(
+            f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.env.name}]'`"
+        )
         raise e
 
-    gym_handle = f"{package_name}/{cfg.task}"
+    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
 
     # batched version of the env that returns an observation of shape (b, c)
-    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+    env_cls = gym.vector.AsyncVectorEnv if cfg.eval.use_async_envs else gym.vector.SyncVectorEnv
     env = env_cls(
-        [lambda: gym.make(gym_handle, disable_env_checker=True, **cfg.gym_kwargs) for _ in range(n_envs)]
+        [
+            lambda: gym.make(gym_handle, disable_env_checker=True, **gym_kwgs)
+            for _ in range(n_envs if n_envs is not None else cfg.eval.batch_size)
+        ]
     )
 
     return env

lerobot/common/envs/utils.py

@@ -18,13 +18,8 @@ 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.
@@ -40,7 +35,6 @@ 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
@@ -66,23 +60,3 @@ 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/logger.py

@@ -0,0 +1,246 @@
+#!/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/optim/__init__.py

@@ -1 +0,0 @@
-from .optimizers import OptimizerConfig as OptimizerConfig

lerobot/common/optim/factory.py

@@ -1,40 +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.
-
-
-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

@@ -1,118 +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.
-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

@@ -1,122 +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.
-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

@@ -1,5 +0,0 @@
-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

@@ -15,14 +15,9 @@
 # limitations under the License.
 from dataclasses import dataclass, field
 
-from lerobot.common.optim.optimizers import AdamWConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
 
-
-@PreTrainedConfig.register_subclass("act")
 @dataclass
-class ACTConfig(PreTrainedConfig):
+class ACTConfig:
     """Configuration class for the Action Chunking Transformers policy.
 
     Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
@@ -95,11 +90,28 @@ class ACTConfig(PreTrainedConfig):
     chunk_size: int = 100
     n_action_steps: int = 100
 
-    normalization_mapping: dict[str, NormalizationMode] = field(
+    input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
-            "VISUAL": NormalizationMode.MEAN_STD,
-            "STATE": NormalizationMode.MEAN_STD,
-            "ACTION": NormalizationMode.MEAN_STD,
+            "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",
         }
     )
 
@@ -132,14 +144,7 @@ class ACTConfig(PreTrainedConfig):
     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(
@@ -159,28 +164,8 @@ class ACTConfig(PreTrainedConfig):
             raise ValueError(
                 f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
             )
-
-    def get_optimizer_preset(self) -> AdamWConfig:
-        return AdamWConfig(
-            lr=self.optimizer_lr,
-            weight_decay=self.optimizer_weight_decay,
-        )
-
-    def get_scheduler_preset(self) -> None:
-        return None
-
-    def validate_features(self) -> None:
-        if not self.image_features and not self.env_state_feature:
+        if (
+            not any(k.startswith("observation.image") for k in self.input_shapes)
+            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.")
-
-    @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

@@ -29,27 +29,32 @@ 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(PreTrainedPolicy):
+class ACTPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "act"],
+):
     """
     Action Chunking Transformer Policy as per Learning Fine-Grained Bimanual Manipulation with Low-Cost
     Hardware (paper: https://arxiv.org/abs/2304.13705, code: https://github.com/tonyzhaozh/act)
     """
 
-    config_class = ACTConfig
     name = "act"
 
     def __init__(
         self,
-        config: ACTConfig,
+        config: ACTConfig | None = None,
         dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
@@ -59,46 +64,30 @@ class ACTPolicy(PreTrainedPolicy):
             dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
-        super().__init__(config)
-        config.validate_features()
-        self.config = config
+        super().__init__()
+        if config is None:
+            config = ACTConfig()
+        self.config: ACTConfig = config
 
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_inputs = Normalize(
+            config.input_shapes, config.input_normalization_modes, dataset_stats
+        )
         self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
+            config.output_shapes, config.output_normalization_modes, dataset_stats
         )
         self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
+            config.output_shapes, config.output_normalization_modes, dataset_stats
         )
 
         self.model = ACT(config)
 
+        self.expected_image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+
         if config.temporal_ensemble_coeff is not None:
             self.temporal_ensembler = ACTTemporalEnsembler(config.temporal_ensemble_coeff, config.chunk_size)
 
         self.reset()
 
-    def 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:
@@ -117,11 +106,9 @@ class ACTPolicy(PreTrainedPolicy):
         self.eval()
 
         batch = self.normalize_inputs(batch)
-        if self.config.image_features:
+        if len(self.expected_image_keys) > 0:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = torch.stack(
-                [batch[key] for key in self.config.image_features], dim=-4
-            )
+            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
 
         # If we are doing temporal ensembling, do online updates where we keep track of the number of actions
         # we are ensembling over.
@@ -144,45 +131,34 @@ class ACTPolicy(PreTrainedPolicy):
             self._action_queue.extend(actions.transpose(0, 1))
         return self._action_queue.popleft()
 
-    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
-        if self.config.image_features:
+        if len(self.expected_image_keys) > 0:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = torch.stack(
-                [batch[key] for key in self.config.image_features], dim=-4
-            )
+            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         batch = self.normalize_targets(batch)
         actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
 
-        l1_loss = F.l1_loss(batch["action"], actions_hat, reduction="none")
-        l1_loss *= ~batch["action_is_pad"].unsqueeze(-1)
+        l1_loss = (
+            F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
+        ).mean()
 
-        bsize, seqlen, num_motors = l1_loss.shape
-        output_dict = {
-            "l1_loss": l1_loss.mean().item(),
-            "l1_loss_per_item": l1_loss.view(bsize, seqlen * num_motors).mean(dim=1),
-            "action": self.unnormalize_outputs({"action": actions_hat})["action"],
-        }
+        loss_dict = {"l1_loss": l1_loss.item()}
         if self.config.use_vae:
             # Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for
             # each dimension independently, we sum over the latent dimension to get the total
             # KL-divergence per batch element, then take the mean over the batch.
             # (See App. B of https://arxiv.org/abs/1312.6114 for more details).
-            mean_kld = (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1)
-            output_dict["kld_loss_per_item"] = mean_kld
-
-            mean_kld = mean_kld.mean()
-            output_dict["kld_loss"] = mean_kld.item()
-
-            loss = l1_loss + mean_kld * self.config.kl_weight
-            output_dict["loss_per_item"] = (
-                output_dict["l1_loss_per_item"] + output_dict["kld_loss_per_item"] * self.config.kl_weight
+            mean_kld = (
+                (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
             )
+            loss_dict["kld_loss"] = mean_kld.item()
+            loss_dict["loss"] = l1_loss + mean_kld * self.config.kl_weight
         else:
-            loss = l1_loss
+            loss_dict["loss"] = l1_loss
 
-        return loss, output_dict
+        return loss_dict
 
 
 class ACTTemporalEnsembler:
@@ -312,30 +288,31 @@ class ACT(nn.Module):
     """
 
     def __init__(self, config: ACTConfig):
-        # BERT style VAE encoder with input tokens [cls, robot_state, *action_sequence].
-        # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
         super().__init__()
         self.config = config
-
+        # 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
         if self.config.use_vae:
             self.vae_encoder = ACTEncoder(config, is_vae_encoder=True)
             self.vae_encoder_cls_embed = nn.Embedding(1, config.dim_model)
             # Projection layer for joint-space configuration to hidden dimension.
-            if self.config.robot_state_feature:
+            if self.use_robot_state:
                 self.vae_encoder_robot_state_input_proj = nn.Linear(
-                    self.config.robot_state_feature.shape[0], config.dim_model
+                    config.input_shapes["observation.state"][0], config.dim_model
                 )
             # Projection layer for action (joint-space target) to hidden dimension.
             self.vae_encoder_action_input_proj = nn.Linear(
-                self.config.action_feature.shape[0],
-                config.dim_model,
+                config.output_shapes["action"][0], config.dim_model
             )
             # Projection layer from the VAE encoder's output to the latent distribution's parameter space.
             self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, config.latent_dim * 2)
             # Fixed sinusoidal positional embedding for the input to the VAE encoder. Unsqueeze for batch
             # dimension.
             num_input_token_encoder = 1 + config.chunk_size
-            if self.config.robot_state_feature:
+            if self.use_robot_state:
                 num_input_token_encoder += 1
             self.register_buffer(
                 "vae_encoder_pos_enc",
@@ -343,7 +320,7 @@ class ACT(nn.Module):
             )
 
         # Backbone for image feature extraction.
-        if self.config.image_features:
+        if self.use_images:
             backbone_model = getattr(torchvision.models, config.vision_backbone)(
                 replace_stride_with_dilation=[False, False, config.replace_final_stride_with_dilation],
                 weights=config.pretrained_backbone_weights,
@@ -360,27 +337,27 @@ 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.config.robot_state_feature:
+        if self.use_robot_state:
             self.encoder_robot_state_input_proj = nn.Linear(
-                self.config.robot_state_feature.shape[0], config.dim_model
+                config.input_shapes["observation.state"][0], config.dim_model
             )
-        if self.config.env_state_feature:
+        if self.use_env_state:
             self.encoder_env_state_input_proj = nn.Linear(
-                self.config.env_state_feature.shape[0], config.dim_model
+                config.input_shapes["observation.environment_state"][0], config.dim_model
             )
         self.encoder_latent_input_proj = nn.Linear(config.latent_dim, config.dim_model)
-        if self.config.image_features:
+        if self.use_images:
             self.encoder_img_feat_input_proj = nn.Conv2d(
                 backbone_model.fc.in_features, config.dim_model, kernel_size=1
             )
         # Transformer encoder positional embeddings.
         n_1d_tokens = 1  # for the latent
-        if self.config.robot_state_feature:
+        if self.use_robot_state:
             n_1d_tokens += 1
-        if self.config.env_state_feature:
+        if self.use_env_state:
             n_1d_tokens += 1
         self.encoder_1d_feature_pos_embed = nn.Embedding(n_1d_tokens, config.dim_model)
-        if self.config.image_features:
+        if self.use_images:
             self.encoder_cam_feat_pos_embed = ACTSinusoidalPositionEmbedding2d(config.dim_model // 2)
 
         # Transformer decoder.
@@ -388,7 +365,7 @@ 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, self.config.action_feature.shape[0])
+        self.action_head = nn.Linear(config.dim_model, config.output_shapes["action"][0])
 
         self._reset_parameters()
 
@@ -403,13 +380,13 @@ class ACT(nn.Module):
 
         `batch` should have the following structure:
         {
-            [robot_state_feature] (optional): (B, state_dim) batch of robot states.
+            "observation.state" (optional): (B, state_dim) batch of robot states.
 
-            [image_features]: (B, n_cameras, C, H, W) batch of images.
+            "observation.images": (B, n_cameras, C, H, W) batch of images.
                 AND/OR
-            [env_state_feature]: (B, env_dim) batch of environment states.
+            "observation.environment_state": (B, env_dim) batch of environment states.
 
-            [action_feature] (optional, only if training with VAE): (B, chunk_size, action dim) batch of actions.
+            "action" (optional, only if training with VAE): (B, chunk_size, action dim) batch of actions.
         }
 
         Returns:
@@ -418,9 +395,9 @@ 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"]
@@ -434,12 +411,12 @@ 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.config.robot_state_feature:
+            if self.use_robot_state:
                 robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
                 robot_state_embed = robot_state_embed.unsqueeze(1)  # (B, 1, D)
             action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
 
-            if self.config.robot_state_feature:
+            if self.use_robot_state:
                 vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
             else:
                 vae_encoder_input = [cls_embed, action_embed]
@@ -453,7 +430,7 @@ 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.config.robot_state_feature else 1),
+                (batch_size, 2 if self.use_robot_state else 1),
                 False,
                 device=batch["observation.state"].device,
             )
@@ -486,16 +463,16 @@ 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.config.robot_state_feature:
+        if self.use_robot_state:
             encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch["observation.state"]))
         # Environment state token.
-        if self.config.env_state_feature:
+        if self.use_env_state:
             encoder_in_tokens.append(
                 self.encoder_env_state_input_proj(batch["observation.environment_state"])
             )
 
         # Camera observation features and positional embeddings.
-        if self.config.image_features:
+        if self.use_images:
             all_cam_features = []
             all_cam_pos_embeds = []
 

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -16,15 +16,9 @@
 # limitations under the License.
 from dataclasses import dataclass, field
 
-from lerobot.common.optim.optimizers import AdamConfig
-from lerobot.common.optim.schedulers import DiffuserSchedulerConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
 
-
-@PreTrainedConfig.register_subclass("diffusion")
 @dataclass
-class DiffusionConfig(PreTrainedConfig):
+class DiffusionConfig:
     """Configuration class for DiffusionPolicy.
 
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
@@ -108,17 +102,26 @@ class DiffusionConfig(PreTrainedConfig):
     horizon: int = 16
     n_action_steps: int = 8
 
-    normalization_mapping: dict[str, NormalizationMode] = field(
+    input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
-            "VISUAL": NormalizationMode.MEAN_STD,
-            "STATE": NormalizationMode.MIN_MAX,
-            "ACTION": NormalizationMode.MIN_MAX,
+            "observation.image": [3, 96, 96],
+            "observation.state": [2],
+        }
+    )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [2],
         }
     )
 
-    # 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
+    # Normalization / Unnormalization
+    input_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {
+            "observation.image": "mean_std",
+            "observation.state": "min_max",
+        }
+    )
+    output_normalization_modes: dict[str, str] = field(default_factory=lambda: {"action": "min_max"})
 
     # Architecture / modeling.
     # Vision backbone.
@@ -151,23 +154,39 @@ class DiffusionConfig(PreTrainedConfig):
     # 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(
@@ -188,50 +207,3 @@ class DiffusionConfig(PreTrainedConfig):
                 "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

@@ -31,32 +31,35 @@ 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, OBS_ROBOT
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
-from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.utils import (
     get_device_from_parameters,
     get_dtype_from_parameters,
-    get_output_shape,
     populate_queues,
 )
 
 
-class DiffusionPolicy(PreTrainedPolicy):
+class DiffusionPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "diffusion-policy"],
+):
     """
     Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
     (paper: https://arxiv.org/abs/2303.04137, code: https://github.com/real-stanford/diffusion_policy).
     """
 
-    config_class = DiffusionConfig
     name = "diffusion"
 
     def __init__(
         self,
-        config: DiffusionConfig,
+        config: DiffusionConfig | None = None,
         dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
@@ -66,16 +69,18 @@ class DiffusionPolicy(PreTrainedPolicy):
             dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
-        super().__init__(config)
-        config.validate_features()
+        super().__init__()
+        if config is None:
+            config = DiffusionConfig()
         self.config = config
-
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_inputs = Normalize(
+            config.input_shapes, config.input_normalization_modes, dataset_stats
+        )
         self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
+            config.output_shapes, config.output_normalization_modes, dataset_stats
         )
         self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
+            config.output_shapes, config.output_normalization_modes, dataset_stats
         )
 
         # queues are populated during rollout of the policy, they contain the n latest observations and actions
@@ -83,10 +88,10 @@ class DiffusionPolicy(PreTrainedPolicy):
 
         self.diffusion = DiffusionModel(config)
 
-        self.reset()
+        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
 
-    def get_optim_params(self) -> dict:
-        return self.diffusion.parameters()
+        self.reset()
 
     def reset(self):
         """Clear observation and action queues. Should be called on `env.reset()`"""
@@ -94,9 +99,9 @@ class DiffusionPolicy(PreTrainedPolicy):
             "observation.state": deque(maxlen=self.config.n_obs_steps),
             "action": deque(maxlen=self.config.n_action_steps),
         }
-        if self.config.image_features:
+        if len(self.expected_image_keys) > 0:
             self._queues["observation.images"] = deque(maxlen=self.config.n_obs_steps)
-        if self.config.env_state_feature:
+        if self.use_env_state:
             self._queues["observation.environment_state"] = deque(maxlen=self.config.n_obs_steps)
 
     @torch.no_grad
@@ -122,11 +127,9 @@ class DiffusionPolicy(PreTrainedPolicy):
         actually measured from the first observation which (if `n_obs_steps` > 1) happened in the past.
         """
         batch = self.normalize_inputs(batch)
-        if self.config.image_features:
+        if len(self.expected_image_keys) > 0:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = torch.stack(
-                [batch[key] for key in self.config.image_features], dim=-4
-            )
+            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         # Note: It's important that this happens after stacking the images into a single key.
         self._queues = populate_queues(self._queues, batch)
 
@@ -143,18 +146,15 @@ class DiffusionPolicy(PreTrainedPolicy):
         action = self._queues["action"].popleft()
         return action
 
-    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, None]:
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
-        if self.config.image_features:
+        if len(self.expected_image_keys) > 0:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.images"] = torch.stack(
-                [batch[key] for key in self.config.image_features], dim=-4
-            )
+            batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         batch = self.normalize_targets(batch)
         loss = self.diffusion.compute_loss(batch)
-        # no output_dict so returning None
-        return loss, None
+        return {"loss": loss}
 
 
 def _make_noise_scheduler(name: str, **kwargs: dict) -> DDPMScheduler | DDIMScheduler:
@@ -176,9 +176,12 @@ class DiffusionModel(nn.Module):
         self.config = config
 
         # Build observation encoders (depending on which observations are provided).
-        global_cond_dim = self.config.robot_state_feature.shape[0]
-        if self.config.image_features:
-            num_images = len(self.config.image_features)
+        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
             if self.config.use_separate_rgb_encoder_per_camera:
                 encoders = [DiffusionRgbEncoder(config) for _ in range(num_images)]
                 self.rgb_encoder = nn.ModuleList(encoders)
@@ -186,8 +189,9 @@ class DiffusionModel(nn.Module):
             else:
                 self.rgb_encoder = DiffusionRgbEncoder(config)
                 global_cond_dim += self.rgb_encoder.feature_dim * num_images
-        if self.config.env_state_feature:
-            global_cond_dim += self.config.env_state_feature.shape[0]
+        if "observation.environment_state" in config.input_shapes:
+            self._use_env_state = True
+            global_cond_dim += config.input_shapes["observation.environment_state"][0]
 
         self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
 
@@ -216,7 +220,7 @@ class DiffusionModel(nn.Module):
 
         # Sample prior.
         sample = torch.randn(
-            size=(batch_size, self.config.horizon, self.config.action_feature.shape[0]),
+            size=(batch_size, self.config.horizon, self.config.output_shapes["action"][0]),
             dtype=dtype,
             device=device,
             generator=generator,
@@ -238,10 +242,10 @@ 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[OBS_ROBOT].shape[:2]
-        global_cond_feats = [batch[OBS_ROBOT]]
+        batch_size, n_obs_steps = batch["observation.state"].shape[:2]
+        global_cond_feats = [batch["observation.state"]]
         # Extract image features.
-        if self.config.image_features:
+        if self._use_images:
             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) ...")
@@ -268,8 +272,8 @@ class DiffusionModel(nn.Module):
                 )
             global_cond_feats.append(img_features)
 
-        if self.config.env_state_feature:
-            global_cond_feats.append(batch[OBS_ENV])
+        if self._use_env_state:
+            global_cond_feats.append(batch["observation.environment_state"])
 
         # Concatenate features then flatten to (B, global_cond_dim).
         return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)
@@ -439,7 +443,7 @@ class SpatialSoftmax(nn.Module):
 
 
 class DiffusionRgbEncoder(nn.Module):
-    """Encodes an RGB image into a 1D feature vector.
+    """Encoder an RGB image into a 1D feature vector.
 
     Includes the ability to normalize and crop the image first.
     """
@@ -478,16 +482,19 @@ 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.image_features` and it should
+        # The dummy input should take the number of image channels from `config.input_shapes` and it should
         # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
-        # height and width from `config.image_features`.
-
+        # height and width from `config.input_shapes`.
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
         # Note: we have a check in the config class to make sure all images have the same shape.
-        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:]
-
+        image_key = image_keys[0]
+        dummy_input_h_w = (
+            config.crop_shape if config.crop_shape is not None else config.input_shapes[image_key][1:]
+        )
+        dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *dummy_input_h_w))
+        with torch.inference_mode():
+            dummy_feature_map = self.backbone(dummy_input)
+        feature_map_shape = tuple(dummy_feature_map.shape[1:])
         self.pool = SpatialSoftmax(feature_map_shape, num_kp=config.spatial_softmax_num_keypoints)
         self.feature_dim = config.spatial_softmax_num_keypoints * 2
         self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
@@ -604,7 +611,7 @@ 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.action_feature.shape[0], config.down_dims[0])] + list(
+        in_out = [(config.output_shapes["action"][0], config.down_dims[0])] + list(
             zip(config.down_dims[:-1], config.down_dims[1:], strict=True)
         )
 
@@ -659,7 +666,7 @@ 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.action_feature.shape[0], 1),
+            nn.Conv1d(config.down_dims[0], config.output_shapes["action"][0], 1),
         )
 
     def forward(self, x: Tensor, timestep: Tensor | int, global_cond=None) -> Tensor:

lerobot/common/policies/factory.py

@@ -13,141 +13,106 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
+import inspect
 import logging
 
-import torch
-from torch import nn
+from omegaconf import DictConfig, OmegaConf
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.common.datasets.utils import dataset_to_policy_features
-from lerobot.common.envs.configs import EnvConfig
-from lerobot.common.envs.utils import env_to_policy_features
-from lerobot.common.policies.act.configuration_act import ACTConfig
-from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
-from lerobot.common.policies.pretrained import PreTrainedPolicy
-from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
-from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import FeatureType
+from lerobot.common.policies.policy_protocol import Policy
+from lerobot.common.utils.utils import get_safe_torch_device
 
 
-def get_policy_class(name: str) -> PreTrainedPolicy:
+def _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg):
+    expected_kwargs = set(inspect.signature(policy_cfg_class).parameters)
+    if not set(hydra_cfg.policy).issuperset(expected_kwargs):
+        logging.warning(
+            f"Hydra config is missing arguments: {set(expected_kwargs).difference(hydra_cfg.policy)}"
+        )
+
+    # OmegaConf.to_container returns lists where sequences are found, but our dataclasses use tuples to avoid
+    # issues with mutable defaults. This filter changes all lists to tuples.
+    def list_to_tuple(item):
+        return tuple(item) if isinstance(item, list) else item
+
+    policy_cfg = policy_cfg_class(
+        **{
+            k: list_to_tuple(v)
+            for k, v in OmegaConf.to_container(hydra_cfg.policy, resolve=True).items()
+            if k in expected_kwargs
+        }
+    )
+    return policy_cfg
+
+
+def get_policy_and_config_classes(name: str) -> tuple[Policy, object]:
     """Get the policy's class and config class given a name (matching the policy class' `name` attribute)."""
     if name == "tdmpc":
+        from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
         from lerobot.common.policies.tdmpc.modeling_tdmpc import TDMPCPolicy
 
-        return TDMPCPolicy
+        return TDMPCPolicy, TDMPCConfig
+
+    elif name == "tdmpc2":
+        from lerobot.common.policies.tdmpc2.configuration_tdmpc2 import TDMPC2Config
+        from lerobot.common.policies.tdmpc2.modeling_tdmpc2 import TDMPC2Policy
+
+        return TDMPC2Policy, TDMPC2Config
+
     elif name == "diffusion":
+        from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
         from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 
-        return DiffusionPolicy
+        return DiffusionPolicy, DiffusionConfig
     elif name == "act":
+        from lerobot.common.policies.act.configuration_act import ACTConfig
         from lerobot.common.policies.act.modeling_act import ACTPolicy
 
-        return ACTPolicy
+        return ACTPolicy, ACTConfig
     elif name == "vqbet":
+        from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
         from lerobot.common.policies.vqbet.modeling_vqbet import VQBeTPolicy
 
-        return VQBeTPolicy
-    elif name == "pi0":
-        from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
-
-        return PI0Policy
+        return VQBeTPolicy, VQBeTConfig
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
 
-def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
-    if policy_type == "tdmpc":
-        return TDMPCConfig(**kwargs)
-    elif policy_type == "diffusion":
-        return DiffusionConfig(**kwargs)
-    elif policy_type == "act":
-        return ACTConfig(**kwargs)
-    elif policy_type == "vqbet":
-        return VQBeTConfig(**kwargs)
-    elif policy_type == "pi0":
-        return PI0Config(**kwargs)
-    else:
-        raise ValueError(f"Policy type '{policy_type}' is not available.")
-
-
 def make_policy(
-    cfg: PreTrainedConfig,
-    device: str | torch.device,
-    ds_meta: LeRobotDatasetMetadata | None = None,
-    env_cfg: EnvConfig | None = None,
-) -> PreTrainedPolicy:
+    hydra_cfg: DictConfig, pretrained_policy_name_or_path: str | None = None, dataset_stats=None
+) -> Policy:
     """Make an instance of a policy class.
 
-    This function exists because (for now) we need to parse features from either a dataset or an environment
-    in order to properly dimension and instantiate a policy for that dataset or environment.
-
     Args:
-        cfg (PreTrainedConfig): The config of the policy to make. If `pretrained_path` is set, the policy will
-            be loaded with the weights from that path.
-        device (str): the device to load the policy onto.
-        ds_meta (LeRobotDatasetMetadata | None, optional): Dataset metadata to take input/output shapes and
-            statistics to use for (un)normalization of inputs/outputs in the policy. Defaults to None.
-        env_cfg (EnvConfig | None, optional): The config of a gym environment to parse features from. Must be
-            provided if ds_meta is not. Defaults to None.
-
-    Raises:
-        ValueError: Either ds_meta or env and env_cfg must be provided.
-        NotImplementedError: if the policy.type is 'vqbet' and the device 'mps' (due to an incompatibility)
-
-    Returns:
-        PreTrainedPolicy: _description_
+        hydra_cfg: A parsed Hydra configuration (see scripts). If `pretrained_policy_name_or_path` is
+            provided, only `hydra_cfg.policy.name` is used while everything else is ignored.
+        pretrained_policy_name_or_path: Either the repo ID of a model hosted on the Hub or a path to a
+            directory containing weights saved using `Policy.save_pretrained`. Note that providing this
+            argument overrides everything in `hydra_cfg.policy` apart from `hydra_cfg.policy.name`.
+        dataset_stats: Dataset statistics to use for (un)normalization of inputs/outputs in the policy. Must
+            be provided when initializing a new policy, and must not be provided when loading a pretrained
+            policy. Therefore, this argument is mutually exclusive with `pretrained_policy_name_or_path`.
     """
-    if bool(ds_meta) == bool(env_cfg):
-        raise ValueError("Either one of a dataset metadata or a sim env must be provided.")
-
-    # NOTE: Currently, if you try to run vqbet with mps backend, you'll get this error.
-    # TODO(aliberts, rcadene): Implement a check_backend_compatibility in policies?
-    # NotImplementedError: The operator 'aten::unique_dim' is not currently implemented for the MPS device. If
-    # you want this op to be added in priority during the prototype phase of this feature, please comment on
-    # https://github.com/pytorch/pytorch/issues/77764. As a temporary fix, you can set the environment
-    # variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be
-    # slower than running natively on MPS.
-    if cfg.type == "vqbet" and str(device) == "mps":
-        raise NotImplementedError(
-            "Current implementation of VQBeT does not support `mps` backend. "
-            "Please use `cpu` or `cuda` backend."
+    if not (pretrained_policy_name_or_path is None) ^ (dataset_stats is None):
+        raise ValueError(
+            "Exactly one of `pretrained_policy_name_or_path` and `dataset_stats` must be provided."
         )
 
-    policy_cls = get_policy_class(cfg.type)
+    policy_cls, policy_cfg_class = get_policy_and_config_classes(hydra_cfg.policy.name)
 
-    kwargs = {}
-    if ds_meta is not None:
-        features = dataset_to_policy_features(ds_meta.features)
-        kwargs["dataset_stats"] = ds_meta.stats
+    policy_cfg = _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg)
+    if pretrained_policy_name_or_path is None:
+        # Make a fresh policy.
+        policy = policy_cls(policy_cfg, dataset_stats)
     else:
-        if not cfg.pretrained_path:
-            logging.warning(
-                "You are instantiating a policy from scratch and its features are parsed from an environment "
-                "rather than a dataset. Normalization modules inside the policy will have infinite values "
-                "by default without stats from a dataset."
-            )
-        features = env_to_policy_features(env_cfg)
-
-    cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
-    cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
-    kwargs["config"] = cfg
-
-    if cfg.pretrained_path:
         # Load a pretrained policy and override the config if needed (for example, if there are inference-time
         # hyperparameters that we want to vary).
-        kwargs["pretrained_name_or_path"] = cfg.pretrained_path
-        policy = policy_cls.from_pretrained(**kwargs)
-    else:
-        # Make a fresh policy.
-        policy = policy_cls(**kwargs)
+        # TODO(alexander-soare): This hack makes use of huggingface_hub's tooling to load the policy with,
+        # pretrained weights which are then loaded into a fresh policy with the desired config. This PR in
+        # huggingface_hub should make it possible to avoid the hack:
+        # https://github.com/huggingface/huggingface_hub/pull/2274.
+        policy = policy_cls(policy_cfg)
+        policy.load_state_dict(policy_cls.from_pretrained(pretrained_policy_name_or_path).state_dict())
 
-    policy.to(device)
-    assert isinstance(policy, nn.Module)
-
-    # policy = torch.compile(policy, mode="reduce-overhead")
+    policy.to(get_safe_torch_device(hydra_cfg.device))
 
     return policy

lerobot/common/policies/normalize.py

@@ -16,12 +16,10 @@
 import torch
 from torch import Tensor, nn
 
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-
 
 def create_stats_buffers(
-    features: dict[str, PolicyFeature],
-    norm_map: dict[str, NormalizationMode],
+    shapes: dict[str, list[int]],
+    modes: dict[str, str],
     stats: dict[str, dict[str, Tensor]] | None = None,
 ) -> dict[str, dict[str, nn.ParameterDict]]:
     """
@@ -36,16 +34,12 @@ def create_stats_buffers(
     """
     stats_buffers = {}
 
-    for key, ft in features.items():
-        norm_mode = norm_map.get(ft.type, NormalizationMode.IDENTITY)
-        if norm_mode is NormalizationMode.IDENTITY:
-            continue
+    for key, mode in modes.items():
+        assert mode in ["mean_std", "min_max"]
 
-        assert isinstance(norm_mode, NormalizationMode)
+        shape = tuple(shapes[key])
 
-        shape = tuple(ft.shape)
-
-        if ft.type is FeatureType.VISUAL:
+        if "image" in key:
             # sanity checks
             assert len(shape) == 3, f"number of dimensions of {key} != 3 ({shape=}"
             c, h, w = shape
@@ -58,7 +52,7 @@ def create_stats_buffers(
         # we assert they are not infinity anymore.
 
         buffer = {}
-        if norm_mode is NormalizationMode.MEAN_STD:
+        if mode == "mean_std":
             mean = torch.ones(shape, dtype=torch.float32) * torch.inf
             std = torch.ones(shape, dtype=torch.float32) * torch.inf
             buffer = nn.ParameterDict(
@@ -67,7 +61,7 @@ def create_stats_buffers(
                     "std": nn.Parameter(std, requires_grad=False),
                 }
             )
-        elif norm_mode is NormalizationMode.MIN_MAX:
+        elif mode == "min_max":
             min = torch.ones(shape, dtype=torch.float32) * torch.inf
             max = torch.ones(shape, dtype=torch.float32) * torch.inf
             buffer = nn.ParameterDict(
@@ -77,15 +71,15 @@ def create_stats_buffers(
                 }
             )
 
-        if stats:
+        if stats is not None:
             # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
             # tensors anywhere (for example, when we use the same stats for normalization and
             # unnormalization). See the logic here
             # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
-            if norm_mode is NormalizationMode.MEAN_STD:
+            if mode == "mean_std":
                 buffer["mean"].data = stats[key]["mean"].clone()
                 buffer["std"].data = stats[key]["std"].clone()
-            elif norm_mode is NormalizationMode.MIN_MAX:
+            elif mode == "min_max":
                 buffer["min"].data = stats[key]["min"].clone()
                 buffer["max"].data = stats[key]["max"].clone()
 
@@ -105,8 +99,8 @@ class Normalize(nn.Module):
 
     def __init__(
         self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
+        shapes: dict[str, list[int]],
+        modes: dict[str, str],
         stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
@@ -128,10 +122,10 @@ class Normalize(nn.Module):
                 dataset is not needed to get the stats, since they are already in the policy state_dict.
         """
         super().__init__()
-        self.features = features
-        self.norm_map = norm_map
+        self.shapes = shapes
+        self.modes = modes
         self.stats = stats
-        stats_buffers = create_stats_buffers(features, norm_map, stats)
+        stats_buffers = create_stats_buffers(shapes, modes, stats)
         for key, buffer in stats_buffers.items():
             setattr(self, "buffer_" + key.replace(".", "_"), buffer)
 
@@ -139,23 +133,16 @@ class Normalize(nn.Module):
     @torch.no_grad
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         batch = dict(batch)  # shallow copy avoids mutating the input batch
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
+        for key, mode in self.modes.items():
             buffer = getattr(self, "buffer_" + key.replace(".", "_"))
 
-            if norm_mode is NormalizationMode.MEAN_STD:
+            if mode == "mean_std":
                 mean = buffer["mean"]
                 std = buffer["std"]
                 assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
                 assert not torch.isinf(std).any(), _no_stats_error_str("std")
                 batch[key] = (batch[key] - mean) / (std + 1e-8)
-            elif norm_mode is NormalizationMode.MIN_MAX:
+            elif mode == "min_max":
                 min = buffer["min"]
                 max = buffer["max"]
                 assert not torch.isinf(min).any(), _no_stats_error_str("min")
@@ -165,7 +152,7 @@ class Normalize(nn.Module):
                 # normalize to [-1, 1]
                 batch[key] = batch[key] * 2 - 1
             else:
-                raise ValueError(norm_mode)
+                raise ValueError(mode)
         return batch
 
 
@@ -177,8 +164,8 @@ class Unnormalize(nn.Module):
 
     def __init__(
         self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
+        shapes: dict[str, list[int]],
+        modes: dict[str, str],
         stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
@@ -200,11 +187,11 @@ class Unnormalize(nn.Module):
                 dataset is not needed to get the stats, since they are already in the policy state_dict.
         """
         super().__init__()
-        self.features = features
-        self.norm_map = norm_map
+        self.shapes = shapes
+        self.modes = modes
         self.stats = stats
         # `self.buffer_observation_state["mean"]` contains `torch.tensor(state_dim)`
-        stats_buffers = create_stats_buffers(features, norm_map, stats)
+        stats_buffers = create_stats_buffers(shapes, modes, stats)
         for key, buffer in stats_buffers.items():
             setattr(self, "buffer_" + key.replace(".", "_"), buffer)
 
@@ -212,23 +199,16 @@ class Unnormalize(nn.Module):
     @torch.no_grad
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         batch = dict(batch)  # shallow copy avoids mutating the input batch
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
+        for key, mode in self.modes.items():
             buffer = getattr(self, "buffer_" + key.replace(".", "_"))
 
-            if norm_mode is NormalizationMode.MEAN_STD:
+            if mode == "mean_std":
                 mean = buffer["mean"]
                 std = buffer["std"]
                 assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
                 assert not torch.isinf(std).any(), _no_stats_error_str("std")
                 batch[key] = batch[key] * std + mean
-            elif norm_mode is NormalizationMode.MIN_MAX:
+            elif mode == "min_max":
                 min = buffer["min"]
                 max = buffer["max"]
                 assert not torch.isinf(min).any(), _no_stats_error_str("min")
@@ -236,5 +216,5 @@ class Unnormalize(nn.Module):
                 batch[key] = (batch[key] + 1) / 2
                 batch[key] = batch[key] * (max - min) + min
             else:
-                raise ValueError(norm_mode)
+                raise ValueError(mode)
         return batch

lerobot/common/policies/pi0/configuration_pi0.py

@@ -1,134 +0,0 @@
-from dataclasses import dataclass, field
-
-from lerobot.common.optim.optimizers import AdamWConfig
-from lerobot.common.optim.schedulers import (
-    CosineDecayWithWarmupSchedulerConfig,
-)
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-
-
-@PreTrainedConfig.register_subclass("pi0")
-@dataclass
-class PI0Config(PreTrainedConfig):
-    # Input / output structure.
-    n_obs_steps: int = 1
-    chunk_size: int = 50
-    n_action_steps: int = 50
-
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.MEAN_STD,
-            "ACTION": NormalizationMode.MEAN_STD,
-        }
-    )
-
-    # Shorter state and action vectors will be padded
-    max_state_dim: int = 32
-    max_action_dim: int = 32
-
-    # Image preprocessing
-    resize_imgs_with_padding: tuple[int, int] = (224, 224)
-
-    # Add empty images. Used by pi0_aloha_sim which adds the empty
-    # left and right wrist cameras in addition to the top camera.
-    empty_cameras: int = 0
-
-    # Converts the joint and gripper values from the standard Aloha space to
-    # the space used by the pi internal runtime which was used to train the base model.
-    adapt_to_pi_aloha: bool = False
-
-    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
-    # Gripper dimensions will remain in absolute values.
-    use_delta_joint_actions_aloha: bool = False
-
-    # Tokenizer
-    tokenizer_max_length: int = 48
-
-    # Projector
-    proj_width: int = 1024
-
-    # Decoding
-    num_steps: int = 10
-
-    # Attention utils
-    use_cache: bool = True
-    attention_implementation: str = "eager"  # or fa2, flex
-
-    # Finetuning settings
-    freeze_vision_encoder: bool = True
-    train_expert_only: bool = False
-    train_state_proj: bool = True
-
-    # Training presets
-    optimizer_lr: float = 2.5e-5
-    optimizer_betas: tuple[float, float] = (0.9, 0.95)
-    optimizer_eps: float = 1e-8
-    optimizer_weight_decay: float = 1e-10
-
-    scheduler_warmup_steps: int = 1_000
-    scheduler_decay_steps: int = 30_000
-    scheduler_decay_lr: float = 2.5e-6
-
-    # TODO: Add EMA
-
-    def __post_init__(self):
-        super().__post_init__()
-
-        """Input validation (not exhaustive)."""
-        if self.n_action_steps > self.chunk_size:
-            raise ValueError(
-                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
-                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
-            )
-        if self.n_obs_steps != 1:
-            raise ValueError(
-                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
-            )
-
-        if self.use_delta_joint_actions_aloha:
-            raise NotImplementedError(
-                "`use_delta_joint_actions_aloha` is used by pi0 for aloha real models. It is not ported yet in LeRobot."
-            )
-
-    def validate_features(self) -> None:
-        # TODO: implement value error
-        # if not self.image_features and not self.env_state_feature:
-        #     raise ValueError("You must provide at least one image or the environment state among the inputs.")
-
-        for i in range(self.empty_cameras):
-            key = f"observation.images.empty_camera_{i}"
-            empty_camera = PolicyFeature(
-                type=FeatureType.VISUAL,
-                shape=(3, 480, 640),
-            )
-            self.input_features[key] = empty_camera
-
-    def get_optimizer_preset(self) -> AdamWConfig:
-        return AdamWConfig(
-            lr=self.optimizer_lr,
-            betas=self.optimizer_betas,
-            eps=self.optimizer_eps,
-            weight_decay=self.optimizer_weight_decay,
-        )
-
-    def get_scheduler_preset(self):
-        return CosineDecayWithWarmupSchedulerConfig(
-            peak_lr=self.optimizer_lr,
-            decay_lr=self.scheduler_decay_lr,
-            num_warmup_steps=self.scheduler_warmup_steps,
-            num_decay_steps=self.scheduler_decay_steps,
-        )
-
-    @property
-    def observation_delta_indices(self) -> None:
-        return None
-
-    @property
-    def action_delta_indices(self) -> list:
-        return list(range(self.chunk_size))
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None

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

@@ -1,68 +0,0 @@
-import torch
-
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.common.policies.factory import make_policy
-from lerobot.configs.policies import PreTrainedConfig
-
-torch.backends.cudnn.benchmark = True
-
-
-def main():
-    device = "cuda"
-    dataset_repo_id = "danaaubakirova/koch_test"
-    # model_name = "pi0_base"
-    # ckpt_torch_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}_pytorch"
-    ckpt_torch_dir = "lerobot/pi0"
-
-    dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
-
-    dataloader = torch.utils.data.DataLoader(
-        dataset,
-        num_workers=0,
-        batch_size=1,
-    )
-
-    batch = next(iter(dataloader))
-
-    # To device
-    for k in batch:
-        if isinstance(batch[k], torch.Tensor):
-            batch[k] = batch[k].to(device=device, dtype=torch.float32)
-
-    cfg = PreTrainedConfig.from_pretrained(ckpt_torch_dir)
-    cfg.pretrained_path = ckpt_torch_dir
-    policy = make_policy(cfg, device, ds_meta=dataset.meta)
-
-    # policy = torch.compile(policy, mode="reduce-overhead")
-
-    warmup_iters = 10
-    benchmark_iters = 30
-
-    # Warmup
-    for _ in range(warmup_iters):
-        torch.cuda.synchronize()
-        policy.select_action(batch)
-        policy.reset()
-        torch.cuda.synchronize()
-
-    # Benchmark
-    start_event = torch.cuda.Event(enable_timing=True)
-    end_event = torch.cuda.Event(enable_timing=True)
-
-    start_event.record()
-    for _ in range(benchmark_iters):
-        policy.select_action(batch)
-        policy.reset()
-    end_event.record()
-
-    # Synchronize and measure time
-    torch.cuda.synchronize()
-    elapsed_time_ms = start_event.elapsed_time(end_event)
-
-    avg_time_per_iter = elapsed_time_ms / benchmark_iters
-    print(f"Average execution time per iteration: {avg_time_per_iter:.3f} ms")
-
-
-if __name__ == "__main__":
-    with torch.inference_mode():
-        main()

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

@@ -1,117 +0,0 @@
-import json
-import pickle
-from pathlib import Path
-
-import torch
-
-from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.common.policies.factory import make_policy
-from lerobot.configs.policies import PreTrainedConfig
-
-
-def display(tensor: torch.Tensor):
-    if tensor.dtype == torch.bool:
-        tensor = tensor.float()
-    print(f"Shape: {tensor.shape}")
-    print(f"Mean: {tensor.mean().item()}")
-    print(f"Std: {tensor.std().item()}")
-    print(f"Min: {tensor.min().item()}")
-    print(f"Max: {tensor.max().item()}")
-
-
-def main():
-    num_motors = 14
-    device = "cuda"
-    # model_name = "pi0_aloha_towel"
-    model_name = "pi0_aloha_sim"
-
-    if model_name == "pi0_aloha_towel":
-        dataset_repo_id = "lerobot/aloha_static_towel"
-    else:
-        dataset_repo_id = "lerobot/aloha_sim_transfer_cube_human"
-
-    ckpt_torch_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}_pytorch"
-    ckpt_jax_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}"
-    save_dir = Path(f"../openpi/data/{model_name}/save")
-
-    with open(save_dir / "example.pkl", "rb") as f:
-        example = pickle.load(f)
-    with open(save_dir / "outputs.pkl", "rb") as f:
-        outputs = pickle.load(f)
-    with open(save_dir / "noise.pkl", "rb") as f:
-        noise = pickle.load(f)
-
-    with open(ckpt_jax_dir / "assets/norm_stats.json") as f:
-        norm_stats = json.load(f)
-
-    # Override stats
-    dataset_meta = LeRobotDatasetMetadata(dataset_repo_id)
-    dataset_meta.stats["observation.state"]["mean"] = torch.tensor(
-        norm_stats["norm_stats"]["state"]["mean"][:num_motors], dtype=torch.float32
-    )
-    dataset_meta.stats["observation.state"]["std"] = torch.tensor(
-        norm_stats["norm_stats"]["state"]["std"][:num_motors], dtype=torch.float32
-    )
-
-    # Create LeRobot batch from Jax
-    batch = {}
-    for cam_key, uint_chw_array in example["images"].items():
-        batch[f"observation.images.{cam_key}"] = torch.from_numpy(uint_chw_array) / 255.0
-    batch["observation.state"] = torch.from_numpy(example["state"])
-    batch["action"] = torch.from_numpy(outputs["actions"])
-    batch["task"] = example["prompt"]
-
-    if model_name == "pi0_aloha_towel":
-        del batch["observation.images.cam_low"]
-    elif model_name == "pi0_aloha_sim":
-        batch["observation.images.top"] = batch["observation.images.cam_high"]
-        del batch["observation.images.cam_high"]
-
-    # Batchify
-    for key in batch:
-        if isinstance(batch[key], torch.Tensor):
-            batch[key] = batch[key].unsqueeze(0)
-        elif isinstance(batch[key], str):
-            batch[key] = [batch[key]]
-        else:
-            raise ValueError(f"{key}, {batch[key]}")
-
-    # To device
-    for k in batch:
-        if isinstance(batch[k], torch.Tensor):
-            batch[k] = batch[k].to(device=device, dtype=torch.float32)
-
-    noise = torch.from_numpy(noise).to(device=device, dtype=torch.float32)
-
-    from lerobot.common import policies  # noqa
-
-    cfg = PreTrainedConfig.from_pretrained(ckpt_torch_dir)
-    cfg.pretrained_path = ckpt_torch_dir
-    policy = make_policy(cfg, device, dataset_meta)
-
-    # loss_dict = policy.forward(batch, noise=noise, time=time_beta)
-    # loss_dict["loss"].backward()
-    # print("losses")
-    # display(loss_dict["losses_after_forward"])
-    # print("pi_losses")
-    # display(pi_losses)
-
-    actions = []
-    for _ in range(50):
-        action = policy.select_action(batch, noise=noise)
-        actions.append(action)
-
-    actions = torch.stack(actions, dim=1)
-    pi_actions = batch["action"]
-    print("actions")
-    display(actions)
-    print()
-    print("pi_actions")
-    display(pi_actions)
-    print("atol=3e-2", torch.allclose(actions, pi_actions, atol=3e-2))
-    print("atol=2e-2", torch.allclose(actions, pi_actions, atol=2e-2))
-    print("atol=1e-2", torch.allclose(actions, pi_actions, atol=1e-2))
-
-
-if __name__ == "__main__":
-    main()

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

@@ -1,70 +0,0 @@
-from transformers import GemmaConfig, PaliGemmaConfig
-
-
-def get_paligemma_config(precision: str):
-    config = {
-        "image_token_index": None,
-        "pad_token_id": 0,
-        "bos_token_id": 2,
-        "eos_token_id": 1,
-    }
-
-    # image_sizes = {"2b-test": 224, "3b-224px": 224, "3b-448px": 448, "3b-896px": 896}
-
-    image_size = 224  # image_sizes[variant]
-    patch_size = 14
-    num_image_tokens = (image_size**2) // (patch_size**2)
-
-    config["image_token_index"] = 257152
-    text_config = {
-        "vocab_size": 257152,
-        "num_hidden_layers": 18,
-        "num_key_value_heads": 1,
-        "head_dim": 256,
-        "torch_dtype": precision,
-        "hidden_size": 2048,
-        "hidden_activation": "gelu_pytorch_tanh",
-        "num_attention_heads": 8,
-        "intermediate_size": 16384,
-        "is_encoder_decoder": False,
-    }
-    vision_config = {
-        "torch_dtype": precision,
-        "image_size": image_size,
-        "patch_size": patch_size,
-        "num_image_tokens": num_image_tokens,
-        "hidden_size": 1152,
-        "intermediate_size": 4304,
-        "num_hidden_layers": 27,
-        "num_attention_heads": 16,
-        "projector_hidden_act": "gelu_fast",
-        "vision_use_head": False,
-    }
-    final_config = PaliGemmaConfig(text_config=text_config, vision_config=vision_config, **config)
-    return final_config
-
-
-def get_gemma_config(precision: str):
-    config = {
-        "image_token_index": None,
-        "pad_token_id": 0,
-        "bos_token_id": 2,
-        "eos_token_id": 1,
-    }
-
-    config["image_token_index"] = 257152
-    text_config = {
-        "vocab_size": 257152,
-        "num_hidden_layers": 18,
-        "num_key_value_heads": 1,
-        "head_dim": 256,
-        "torch_dtype": precision,
-        "hidden_size": 1024,
-        "hidden_activation": "gelu_pytorch_tanh",
-        "num_attention_heads": 8,
-        "intermediate_size": 4096,
-        "is_encoder_decoder": False,
-    }
-    final_config = GemmaConfig()
-    final_config.update(text_config)
-    return final_config

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

@@ -1,423 +0,0 @@
-"""
-Convert pi0 parameters from Jax to Pytorch
-
-Follow [README of openpi](https://github.com/Physical-Intelligence/openpi) to create a new environment
-and install the required librairies.
-
-```bash
-cd ~/code/openpi
-source .venv/bin/activate
-```
-
-Example downloading parameters:
-```bash
-python
->>> import openpi.shared.download as download
->>> path='s3://openpi-assets/checkpoints/pi0_base/params'
->>> download.maybe_download(path)
-```
-
-Converting pi0_base:
-```python
-python lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py \
-    --checkpoint_dir /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_base/params \
-    --output_path /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_base_pytorch
-```
-
-```python
-python lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py \
-    --checkpoint_dir /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim/params \
-    --output_path /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim_pytorch
-```
-"""
-
-import argparse
-import pathlib
-
-import jax
-import numpy as np
-import orbax.checkpoint as ocp
-import torch
-from jax.sharding import SingleDeviceSharding
-
-from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
-from lerobot.common.policies.pi0.conversion_scripts.conversion_utils import (
-    get_gemma_config,
-    get_paligemma_config,
-)
-from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
-
-PRECISIONS = {"bfloat16": torch.bfloat16, "float32": torch.float32, "float16": torch.float16}
-
-
-def slice_paligemma_state_dict(state_dict, config):
-    suffix = "/value" if "img/embedding/kernel/value" in state_dict else ""
-
-    # fmt: off
-    # patch embeddings
-    state_dict["paligemma.vision_tower.vision_model.embeddings.patch_embedding.weight"] = state_dict.pop(f"img/embedding/kernel{suffix}").transpose(
-        3, 2, 0, 1
-    )
-    state_dict["paligemma.vision_tower.vision_model.embeddings.patch_embedding.bias"] = state_dict.pop(f"img/embedding/bias{suffix}")
-    # positional embeddings
-    state_dict["paligemma.vision_tower.vision_model.embeddings.position_embedding.weight"] = state_dict.pop(f"img/pos_embedding{suffix}").reshape(
-        -1, config.vision_config.hidden_size
-    )
-
-    # extract vision layers to be sliced at index 0. There are 27 layers in the base model.
-    encoderblock_layernorm0_scale = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_0/scale{suffix}")
-    encoderblock_layernorm0_bias = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_0/bias{suffix}")
-    encoderblock_layernorm1_scale = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_1/scale{suffix}")
-    encoderblock_layernorm1_bias = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_1/bias{suffix}")
-
-    encoderblock_mlp_dense0_kernel= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_0/kernel{suffix}")
-    encoderblock_mlp_dense0_bias= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_0/bias{suffix}")
-    encoderblock_mlp_dense1_kernel= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_1/kernel{suffix}")
-    encoderblock_mlp_dense1_bias= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_1/bias{suffix}")
-
-    encoderblock_attention_0_key_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/key/kernel{suffix}")
-    encoderblock_attention_0_key_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/key/bias{suffix}")
-    encoderblock_attention_0_value_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/value/kernel{suffix}")
-    encoderblock_attention_0_value_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/value/bias{suffix}")
-    encoderblock_attention_0_query_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/query/kernel{suffix}")
-    encoderblock_attention_0_query_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/query/bias{suffix}")
-    encoderblock_attention_0_out_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/out/kernel{suffix}")
-    encoderblock_attention_0_out_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/out/bias{suffix}")
-
-    for i in range(config.vision_config.num_hidden_layers):
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm1.weight"] = encoderblock_layernorm0_scale[i].transpose()
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm1.bias"] = encoderblock_layernorm0_bias[i]
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm2.weight"] = encoderblock_layernorm1_scale[i].transpose()
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm2.bias"] = encoderblock_layernorm1_bias[i]
-
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc1.weight"] = encoderblock_mlp_dense0_kernel[i].transpose()
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc1.bias"] = encoderblock_mlp_dense0_bias[i]
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc2.weight"] = encoderblock_mlp_dense1_kernel[i].transpose()
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc2.bias"] = encoderblock_mlp_dense1_bias[i]
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.k_proj.weight"] = encoderblock_attention_0_key_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.k_proj.bias"] = encoderblock_attention_0_key_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.v_proj.weight"] = encoderblock_attention_0_value_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.v_proj.bias"] = encoderblock_attention_0_value_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.q_proj.weight"] = encoderblock_attention_0_query_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.q_proj.bias"] = encoderblock_attention_0_query_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.out_proj.weight"] = encoderblock_attention_0_out_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
-        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.out_proj.bias"] = encoderblock_attention_0_out_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
-
-    state_dict["paligemma.vision_tower.vision_model.post_layernorm.weight"] = state_dict.pop(f"img/Transformer/encoder_norm/scale{suffix}").transpose()
-    state_dict["paligemma.vision_tower.vision_model.post_layernorm.bias"] = state_dict.pop(f"img/Transformer/encoder_norm/bias{suffix}")
-
-    # multimodal projector
-
-    state_dict['paligemma.multi_modal_projector.linear.weight'] = state_dict.pop(f"img/head/kernel{suffix}").transpose()
-    state_dict['paligemma.multi_modal_projector.linear.bias'] = state_dict.pop(f"img/head/bias{suffix}")
-
-    # text decoder (gemma)
-    embedding_vector = state_dict.pop(f"llm/embedder/input_embedding{suffix}")
-    state_dict["paligemma.language_model.model.embed_tokens.weight"] = embedding_vector
-
-    # pop the einsum attention + mlp representations. There are 18 layers in gemma-2b.
-
-    llm_attention_attn_vec_einsum = state_dict.pop(f"llm/layers/attn/attn_vec_einsum/w{suffix}")
-    llm_attention_kv_einsum = state_dict.pop(f"llm/layers/attn/kv_einsum/w{suffix}")
-    llm_attention_q_einsum = state_dict.pop(f"llm/layers/attn/q_einsum/w{suffix}")
-
-    llm_mlp_gating_einsum = state_dict.pop(f"llm/layers/mlp/gating_einsum{suffix}")
-    llm_mlp_linear = state_dict.pop(f"llm/layers/mlp/linear{suffix}")
-    # TODO verify correctness of layer norm loading
-
-    llm_input_layernorm = state_dict.pop(f"llm/layers/pre_attention_norm/scale{suffix}")
-    llm_post_attention_layernorm = state_dict.pop(f"llm/layers/pre_ffw_norm/scale{suffix}")
-
-    for i in range(config.text_config.num_hidden_layers):
-        # llm_attention_q_einsum[i].shape = (8, 2048, 256)
-        q_proj_weight_reshaped = llm_attention_q_einsum[i].transpose(0, 2, 1).reshape(config.text_config.num_attention_heads * config.text_config.head_dim, config.text_config.hidden_size)
-
-        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.q_proj.weight"] = q_proj_weight_reshaped
-
-        # llm_attention_kv_einsum[i, 0, 0].shape = (2048, 256)
-        k_proj_weight_reshaped = llm_attention_kv_einsum[i, 0, 0].transpose()
-        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.k_proj.weight"] = k_proj_weight_reshaped
-        # llm_attention_kv_einsum[i, 1, 0].shape = (2048, 256)
-        v_proj_weight_reshaped = llm_attention_kv_einsum[i, 1, 0].transpose()
-        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.v_proj.weight"] = v_proj_weight_reshaped
-
-        # output projection.
-
-        # llm_attention_attn_vec_einsum[i].shape = (8, 256, 2048)
-        o_proj_weight_reshaped = llm_attention_attn_vec_einsum[i].transpose(2, 0, 1).reshape(config.text_config.num_attention_heads * config.text_config.head_dim, config.text_config.hidden_size)
-
-        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.o_proj.weight"] = o_proj_weight_reshaped
-        # mlp layers
-        gate_proj_weight = llm_mlp_gating_einsum[i, 0]
-        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.gate_proj.weight"] = gate_proj_weight.transpose()
-        up_proj_weight = llm_mlp_gating_einsum[i, 1]
-        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.up_proj.weight"] = up_proj_weight.transpose()
-        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.down_proj.weight"] = llm_mlp_linear[i].transpose()
-        state_dict[f"paligemma.language_model.model.layers.{i}.input_layernorm.weight"] = llm_input_layernorm[i]
-        state_dict[f"paligemma.language_model.model.layers.{i}.post_attention_layernorm.weight"] = llm_post_attention_layernorm[i]
-
-    state_dict["paligemma.language_model.model.norm.weight"] = state_dict.pop(f"llm/final_norm/scale{suffix}")
-    state_dict["paligemma.language_model.lm_head.weight"] = embedding_vector # weights are tied.
-
-    # fmt: on
-    expert_dict = {}
-    final_state_dict = {}
-    for key, value in state_dict.items():
-        if key not in [
-            f"llm/final_norm_1/scale{suffix}",
-            f"llm/layers/attn/attn_vec_einsum_1/w{suffix}",
-            f"llm/layers/attn/kv_einsum_1/w{suffix}",
-            f"llm/layers/attn/q_einsum_1/w{suffix}",
-            f"llm/layers/mlp_1/gating_einsum{suffix}",
-            f"llm/layers/mlp_1/linear{suffix}",
-            f"llm/layers/pre_attention_norm_1/scale{suffix}",
-            f"llm/layers/pre_ffw_norm_1/scale{suffix}",
-        ]:
-            final_state_dict[key] = torch.from_numpy(value)
-        else:
-            expert_dict[key] = value
-
-    return final_state_dict, expert_dict
-
-
-def slice_gemma_state_dict(state_dict, config, num_expert=1):
-    # fmt: off
-    # text decoder (gemma)
-    # no embedding vector, the expert just has the decoder layers
-
-    embedding_vector = torch.zeros([config.vocab_size, config.hidden_size])
-    state_dict["gemma_expert.model.embed_tokens.weight"] = embedding_vector
-
-    # pop the einsum attention + mlp representations. There are 18 layers in gemma-2b.
-
-    suffix = "/value" if f"llm/layers/attn/attn_vec_einsum_{num_expert}/w/value" in state_dict else ""
-
-    llm_attention_attn_vec_einsum = state_dict.pop(f"llm/layers/attn/attn_vec_einsum_{num_expert}/w{suffix}")
-    llm_attention_kv_einsum = state_dict.pop(f"llm/layers/attn/kv_einsum_{num_expert}/w{suffix}")
-    llm_attention_q_einsum = state_dict.pop(f"llm/layers/attn/q_einsum_{num_expert}/w{suffix}")
-
-    llm_mlp_gating_einsum = state_dict.pop(f"llm/layers/mlp_{num_expert}/gating_einsum{suffix}")
-    llm_mlp_linear = state_dict.pop(f"llm/layers/mlp_{num_expert}/linear{suffix}")
-    # TODO verify correctness of layer norm loading
-
-    llm_input_layernorm = state_dict.pop(f"llm/layers/pre_attention_norm_{num_expert}/scale{suffix}")
-    llm_post_attention_layernorm = state_dict.pop(f"llm/layers/pre_ffw_norm_{num_expert}/scale{suffix}")
-
-    for i in range(config.num_hidden_layers):
-        q_proj_weight_reshaped = llm_attention_q_einsum[i].transpose(0, 2, 1).reshape(config.num_attention_heads * config.head_dim, config.hidden_size)
-
-        state_dict[f"gemma_expert.model.layers.{i}.self_attn.q_proj.weight"] = q_proj_weight_reshaped
-
-        k_proj_weight_reshaped = llm_attention_kv_einsum[i, 0, 0].transpose()
-        state_dict[f"gemma_expert.model.layers.{i}.self_attn.k_proj.weight"] = k_proj_weight_reshaped
-        v_proj_weight_reshaped = llm_attention_kv_einsum[i, 1, 0].transpose()
-        state_dict[f"gemma_expert.model.layers.{i}.self_attn.v_proj.weight"] = v_proj_weight_reshaped
-
-        # output projection.
-
-        # llm_attention_attn_vec_einsum[i].shape = (8, 256, 1024)
-        o_proj_weight_reshaped = llm_attention_attn_vec_einsum[i].reshape(config.num_attention_heads * config.head_dim, config.hidden_size).transpose(1,0)# .transpose(2, 0, 1).reshape(config.num_attention_heads * config.head_dim, config.hidden_size).transpose(1, 0)
-
-        state_dict[f"gemma_expert.model.layers.{i}.self_attn.o_proj.weight"] = o_proj_weight_reshaped
-        # mlp layers
-        gate_proj_weight = llm_mlp_gating_einsum[i, 0]
-        state_dict[f"gemma_expert.model.layers.{i}.mlp.gate_proj.weight"] = gate_proj_weight.transpose()
-        up_proj_weight = llm_mlp_gating_einsum[i, 1]
-        state_dict[f"gemma_expert.model.layers.{i}.mlp.up_proj.weight"] = up_proj_weight.transpose()
-        state_dict[f"gemma_expert.model.layers.{i}.mlp.down_proj.weight"] = llm_mlp_linear[i].transpose()
-        state_dict[f"gemma_expert.model.layers.{i}.input_layernorm.weight"] = llm_input_layernorm[i]
-        state_dict[f"gemma_expert.model.layers.{i}.post_attention_layernorm.weight"] = llm_post_attention_layernorm[i]
-
-    state_dict["gemma_expert.model.norm.weight"] = state_dict.pop(f"llm/final_norm_{num_expert}/scale{suffix}")
-    state_dict["gemma_expert.lm_head.weight"] = embedding_vector # weights are tied. (and zeros here)
-
-    # fmt: on
-    final_state_dict = {}
-    for key, value in state_dict.items():
-        if not isinstance(value, torch.Tensor):
-            final_state_dict[key] = torch.from_numpy(value)
-        else:
-            final_state_dict[key] = value
-    return final_state_dict
-
-
-def flatten_for_memory(tree, parent_key=""):
-    out = {}
-    for k, v in tree.items():
-        new_key = f"{parent_key}/{k}" if parent_key else k
-        if isinstance(v, dict):
-            out.update(flatten_for_memory(v, new_key))
-        else:
-            out[new_key] = np.array(v)  # Ensure conversion to np.array for consistency
-    return out
-
-
-def flatten_for_npz(tree, parent_key=""):
-    out = {}
-    for k, v in tree.items():
-        new_key = f"{parent_key}/{k}" if parent_key else k
-        if isinstance(v, dict):
-            out.update(flatten_for_npz(v, new_key))
-        else:
-            # bf16/f32 here?
-            out[new_key] = np.array(v)
-    return out
-
-
-def slice_initial_orbax_checkpoint(checkpoint_dir: str):
-    params_path = pathlib.Path(checkpoint_dir).resolve()
-    checkpointer = ocp.PyTreeCheckpointer()
-
-    metadata = checkpointer.metadata(params_path)
-    print("Metadata keys:", list(metadata.keys()))
-
-    params_name = "params"
-
-    item = {params_name: metadata[params_name]}
-    device = jax.local_devices()[0]  # Use the first local device
-    sharding = SingleDeviceSharding(device)
-    restored = checkpointer.restore(
-        params_path,
-        ocp.args.PyTreeRestore(
-            item=item,
-            restore_args=jax.tree_util.tree_map(
-                lambda _: ocp.ArrayRestoreArgs(
-                    restore_type=jax.Array,  # or np.ndarray, but bf16 is annoying about it
-                    sharding=sharding,
-                ),
-                item,
-            ),
-            transforms={},
-        ),
-    )
-    params = restored[params_name]
-
-    # get params for PaliGemma
-    pali_params = params["PaliGemma"]
-    del params["PaliGemma"]
-    pali_params_flat = flatten_for_npz(pali_params)
-    return {"paligemma_params": pali_params_flat, "projection_params": params}
-
-
-def update_keys_with_prefix(d: dict, prefix: str) -> dict:
-    """Update dictionary keys by adding a prefix."""
-    return {f"{prefix}{key}": value for key, value in d.items()}
-
-
-def convert_pi0_checkpoint(checkpoint_dir: str, precision: str, tokenizer_id: str, output_path: str):
-    # Break down orbax ckpts - they are in OCDBT
-    initial_params = slice_initial_orbax_checkpoint(checkpoint_dir=checkpoint_dir)
-    # process projection params
-    keys = [
-        "state_proj",
-        "action_in_proj",
-        "action_out_proj",
-        "action_time_mlp_in",
-        "action_time_mlp_out",
-    ]
-
-    projection_params = {}
-    for key in keys:
-        kernel_params = initial_params["projection_params"][key]["kernel"]
-        bias_params = initial_params["projection_params"][key]["bias"]
-        if isinstance(kernel_params, dict):
-            weight = kernel_params["value"]
-            bias = bias_params["value"]
-        else:
-            weight = kernel_params
-            bias = bias_params
-        projection_params[f"{key}.weight"] = torch.from_numpy(np.array(weight)).T
-        projection_params[f"{key}.bias"] = torch.from_numpy(np.array(bias))
-
-    # Process PaliGemma weights
-    paligemma_config = get_paligemma_config(precision)
-    paligemma_params, gemma_raw_dictionary = slice_paligemma_state_dict(
-        initial_params["paligemma_params"], paligemma_config
-    )
-
-    # Process Gemma  weights (at this stage they are unused)
-    gemma_config = get_gemma_config(precision)
-    gemma_params = slice_gemma_state_dict(gemma_raw_dictionary, config=gemma_config)
-
-    # Instantiate model from configs
-
-    if "pi0_aloha_sim" in checkpoint_dir:
-        pi0_config = PI0Config(
-            empty_cameras=2,
-            adapt_to_pi_aloha=True,
-            use_delta_joint_actions_aloha=False,
-        )
-    elif "pi0_aloha_towel" in checkpoint_dir:
-        pi0_config = PI0Config(
-            adapt_to_pi_aloha=True,
-            use_delta_joint_actions_aloha=True,
-        )
-    elif "pi0_base" in checkpoint_dir:
-        pi0_config = PI0Config(
-            empty_cameras=0,
-            adapt_to_pi_aloha=False,
-            use_delta_joint_actions_aloha=False,
-        )
-    else:
-        raise ValueError()
-
-    # gemma_config=gemma_config, paligemma_config=paligemma_config)
-    pi0_model = PI0Policy(pi0_config)
-
-    paligemma_params = update_keys_with_prefix(paligemma_params, "model.paligemma_with_expert.")
-    gemma_params = update_keys_with_prefix(gemma_params, "model.paligemma_with_expert.")
-    projection_params = update_keys_with_prefix(projection_params, "model.")
-
-    # load state dict
-    torch_dtype = PRECISIONS[precision]
-    pi0_model.load_state_dict({**paligemma_params, **gemma_params, **projection_params})
-    pi0_model = pi0_model.to(torch_dtype)
-    # pi0_tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
-
-    pi0_model.save_pretrained(output_path, safe_serialization=True)
-    # pi0_tokenizer.save_pretrained(output_path, dtype=torch_dtype)
-
-    # assert that model loads properly
-    del pi0_model
-    PI0Policy.from_pretrained(output_path)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--checkpoint_dir",
-        default="/raid/pablo/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim/params",
-        type=str,
-        help="Path to the ocdbt checkpoint",
-    )
-
-    parser.add_argument(
-        "--precision",
-        choices=["float32", "bfloat16", "float16"],
-        default="float32",
-        type=str,
-        help="Precision identifier for model conversion - should match the base checkpoint precision.",
-    )
-    # tokenizer is identical to paligemma, it appears
-
-    parser.add_argument(
-        "--tokenizer_hub_id",
-        default="google/paligemma-3b-pt-224",
-        type=str,
-        help="Hub path to the tokenizer to save",
-    )
-
-    parser.add_argument(
-        "--output_path",
-        required=True,
-        type=str,
-        help="Path to save converted weights to",
-    )
-
-    args = parser.parse_args()
-    convert_pi0_checkpoint(
-        checkpoint_dir=args.checkpoint_dir,
-        precision=args.precision,
-        tokenizer_id=args.tokenizer_hub_id,
-        output_path=args.output_path,
-    )

lerobot/common/policies/pi0/flex_attention.py

@@ -1,127 +0,0 @@
-import torch
-import torch.nn.functional as F  # noqa: N812
-from packaging.version import Version
-
-if Version(torch.__version__) > Version("2.5.0"):
-    # Ffex attention is only available from torch 2.5 onwards
-    from torch.nn.attention.flex_attention import (
-        _mask_mod_signature,
-        _round_up_to_multiple,
-        create_block_mask,
-        create_mask,
-        flex_attention,
-    )
-
-
-# @torch.compile(dynamic=False)
-def flex_attention_forward(
-    attention_mask: torch.Tensor,
-    batch_size: int,
-    head_dim: int,
-    query_states: torch.Tensor,
-    key_states: torch.Tensor,
-    value_states: torch.Tensor,
-    scaling=None,
-):
-    """
-    This is defined out of classes to make compile happy.
-    """
-
-    original_dtype = query_states.dtype
-    num_att_heads = 8
-    num_key_value_heads = 1
-    num_key_value_groups = num_att_heads // num_key_value_heads
-
-    key_states = key_states[:, :, :, None, :]
-    key_states = key_states.expand(
-        batch_size, key_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
-    )
-    key_states = key_states.reshape(
-        batch_size, key_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
-    )
-
-    value_states = value_states[:, :, :, None, :]
-    value_states = value_states.expand(
-        batch_size, value_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
-    )
-    value_states = value_states.reshape(
-        batch_size, value_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
-    )
-
-    query_states = query_states.transpose(1, 2)
-    key_states = key_states.transpose(1, 2)
-    value_states = value_states.transpose(1, 2)
-
-    query_states = query_states.to(torch.float32)
-    key_states = key_states.to(torch.float32)
-    value_states = value_states.to(torch.float32)
-
-    causal_mask = attention_mask
-    if causal_mask is not None:
-        causal_mask = causal_mask[:, None, :, : key_states.shape[2]]
-
-        if causal_mask.shape[1] == 1 and query_states.shape[1] > 1:
-            causal_mask = causal_mask.expand(-1, query_states.shape[1], -1, -1)
-
-    def precomputed_mask_factory(precomputed_mask: torch.Tensor) -> _mask_mod_signature:
-        def mask_mod(b, h, q_idx, kv_idx):
-            # Danger zone: if b,h,q_idx,kv_idx exceed the shape, device-side assert occurs.
-            return precomputed_mask[b][h][q_idx][kv_idx]
-
-        return mask_mod
-
-    b_mask, h_mask, q_len, kv_len = causal_mask.shape  # The shape of your mask
-
-    block_size = 128
-    q_len_rounded = _round_up_to_multiple(q_len, block_size)
-    kv_len_rounded = _round_up_to_multiple(kv_len, block_size)
-
-    # *CRITICAL* we do need to expand here, else we get a CUDA index error
-
-    pad_q = q_len_rounded - q_len
-    pad_k = kv_len_rounded - kv_len
-
-    padded_causal_mask = F.pad(causal_mask, (0, pad_k, 0, pad_q), value=0.0)
-    mask_mod_fn_orig = precomputed_mask_factory(padded_causal_mask)
-
-    mask_4d = create_mask(
-        mod_fn=mask_mod_fn_orig,
-        B=b_mask,
-        H=h_mask,
-        Q_LEN=q_len_rounded,
-        KV_LEN=kv_len_rounded,
-        device=causal_mask.device,
-        _compile=False,
-    )
-
-    mask_mod_fn_padded = precomputed_mask_factory(mask_4d)
-    block_mask = create_block_mask(
-        mask_mod=mask_mod_fn_padded,
-        B=b_mask,
-        H=h_mask,
-        Q_LEN=q_len_rounded,
-        KV_LEN=kv_len_rounded,
-        BLOCK_SIZE=block_size,
-        device=causal_mask.device,
-        _compile=False,
-    )
-
-    #  mask is applied inside the kernel, ideally more efficiently than score_mod.
-    attn_output, attention_weights = flex_attention(
-        query_states,
-        key_states,
-        value_states,
-        block_mask=block_mask,
-        enable_gqa=True,  # because we shaped query/key states for GQA
-        scale=head_dim**-0.5 if scaling is None else scaling,
-        return_lse=True,
-    )
-
-    attn_output = attn_output.to(dtype=original_dtype)
-    attn_output = attn_output.transpose(1, 2).contiguous()  # [B, Q_LEN, H, head_dim]
-    attn_output = attn_output.reshape(
-        batch_size,
-        -1,
-        attn_output.shape[2] * attn_output.shape[3],  # merges [H, head_dim]
-    )
-    return attn_output

lerobot/common/policies/pi0/modeling_pi0.py

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

lerobot/common/policies/pi0/paligemma_with_expert.py

@@ -1,403 +0,0 @@
-from typing import List, Optional, Union
-
-import torch
-import torch.version
-from pytest import Cache
-from torch import nn
-from transformers import (
-    AutoConfig,
-    GemmaForCausalLM,
-    PaliGemmaForConditionalGeneration,
-    PretrainedConfig,
-    PreTrainedModel,
-)
-from transformers.models.auto import CONFIG_MAPPING
-
-from lerobot.common.policies.pi0.flex_attention import flex_attention_forward
-
-
-def apply_rope(x, positions, max_wavelength=10_000):
-    """
-    Applies RoPE positions [B, L] to x [B, L, H, D].
-    """
-    d_half = x.shape[-1] // 2
-    device = x.device
-    dtype = x.dtype
-    x = x.to(torch.float32)
-
-    freq_exponents = (2.0 / x.shape[-1]) * torch.arange(d_half, dtype=torch.float32, device=device)
-    timescale = max_wavelength**freq_exponents
-    radians = positions[..., None].to(torch.float32) / timescale[None, None, :].to(torch.float32)
-
-    radians = radians[..., None, :]
-
-    sin = torch.sin(radians)  # .to(dtype=dtype)
-    cos = torch.cos(radians)  # .to(dtype=dtype)
-
-    x1, x2 = x.split(d_half, dim=-1)
-    res = torch.empty_like(x)
-    res[..., :d_half] = x1 * cos - x2 * sin
-    res[..., d_half:] = x2 * cos + x1 * sin
-
-    return res.to(dtype)
-
-
-class PaliGemmaWithExpertConfig(PretrainedConfig):
-    model_type = "PaliGemmaWithExpertModel"
-    sub_configs = {"paligemma_config": AutoConfig, "gemma_expert_config": AutoConfig}
-
-    def __init__(
-        self,
-        paligemma_config: dict | None = None,
-        gemma_expert_config: dict | None = None,
-        freeze_vision_encoder: bool = True,
-        train_expert_only: bool = True,
-        attention_implementation: str = "eager",
-        **kwargs,
-    ):
-        self.freeze_vision_encoder = freeze_vision_encoder
-        self.train_expert_only = train_expert_only
-        self.attention_implementation = attention_implementation
-
-        if paligemma_config is None:
-            # Default config from Pi0
-            self.paligemma_config = CONFIG_MAPPING["paligemma"](
-                transformers_version="4.48.1",
-                _vocab_size=257152,
-                bos_token_id=2,
-                eos_token_id=1,
-                hidden_size=2048,
-                image_token_index=257152,
-                model_type="paligemma",
-                pad_token_id=0,
-                projection_dim=2048,
-                text_config={
-                    "hidden_activation": "gelu_pytorch_tanh",
-                    "hidden_size": 2048,
-                    "intermediate_size": 16384,
-                    "model_type": "gemma",
-                    "num_attention_heads": 8,
-                    "num_hidden_layers": 18,
-                    "num_image_tokens": 256,
-                    "num_key_value_heads": 1,
-                    "torch_dtype": "float32",
-                    "vocab_size": 257152,
-                },
-                vision_config={
-                    "hidden_size": 1152,
-                    "intermediate_size": 4304,
-                    "model_type": "siglip_vision_model",
-                    "num_attention_heads": 16,
-                    "num_hidden_layers": 27,
-                    "num_image_tokens": 256,
-                    "patch_size": 14,
-                    "projection_dim": 2048,
-                    "projector_hidden_act": "gelu_fast",
-                    "torch_dtype": "float32",
-                    "vision_use_head": False,
-                },
-            )
-        elif isinstance(self.paligemma_config, dict):
-            # Override Pi0 default config for PaliGemma
-            if "model_type" not in gemma_expert_config:
-                paligemma_config["model_type"] = "paligemma"
-
-            cfg_cls = CONFIG_MAPPING[paligemma_config["model_type"]]
-            self.paligemma_config = cfg_cls(**paligemma_config)
-
-        if gemma_expert_config is None:
-            # Default config from Pi0
-            self.gemma_expert_config = CONFIG_MAPPING["gemma"](
-                attention_bias=False,
-                attention_dropout=0.0,
-                bos_token_id=2,
-                eos_token_id=1,
-                head_dim=256,
-                hidden_act="gelu_pytorch_tanh",
-                hidden_activation="gelu_pytorch_tanh",
-                hidden_size=1024,
-                initializer_range=0.02,
-                intermediate_size=4096,
-                max_position_embeddings=8192,
-                model_type="gemma",
-                num_attention_heads=8,
-                num_hidden_layers=18,
-                num_key_value_heads=1,
-                pad_token_id=0,
-                rms_norm_eps=1e-06,
-                rope_theta=10000.0,
-                torch_dtype="float32",
-                transformers_version="4.48.1",
-                use_cache=True,
-                vocab_size=257152,
-            )
-        elif isinstance(self.gemma_expert_config, dict):
-            # Override Pi0 default config for Gemma Expert
-            if "model_type" not in gemma_expert_config:
-                gemma_expert_config["model_type"] = "gemma"
-
-            cfg_cls = CONFIG_MAPPING[paligemma_config["model_type"]]
-            self.gemma_expert_config = cfg_cls(**gemma_expert_config)
-
-        super().__init__(**kwargs)
-
-    def __post_init__(self):
-        super().__post_init__()
-        if self.train_expert_only and not self.freeze_vision_encoder:
-            raise ValueError(
-                "You set `freeze_vision_encoder=False` and `train_expert_only=True` which are not compatible."
-            )
-
-        if self.attention_implementation not in ["eager", "fa2", "flex"]:
-            raise ValueError(
-                f"Wrong value provided for `attention_implementation` ({self.attention_implementation}). Expected 'eager', 'fa2' or 'flex'."
-            )
-
-
-class PaliGemmaWithExpertModel(PreTrainedModel):
-    config_class = PaliGemmaWithExpertConfig
-
-    def __init__(self, config: PaliGemmaWithExpertConfig):
-        super().__init__(config=config)
-        self.config = config
-        self.paligemma = PaliGemmaForConditionalGeneration(config=config.paligemma_config)
-        self.gemma_expert = GemmaForCausalLM(config=config.gemma_expert_config)
-        # Remove unused embed_tokens
-        self.gemma_expert.model.embed_tokens = None
-
-        self.to_bfloat16_like_physical_intelligence()
-        self.set_requires_grad()
-
-    def set_requires_grad(self):
-        if self.config.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-            for params in self.paligemma.vision_tower.parameters():
-                params.requires_grad = False
-
-        if self.config.train_expert_only:
-            self.paligemma.eval()
-            for params in self.paligemma.parameters():
-                params.requires_grad = False
-
-    def train(self, mode: bool = True):
-        super().train(mode)
-
-        if self.config.freeze_vision_encoder:
-            self.paligemma.vision_tower.eval()
-
-        if self.config.train_expert_only:
-            self.paligemma.eval()
-
-    def to_bfloat16_like_physical_intelligence(self):
-        self.paligemma = self.paligemma.to(dtype=torch.bfloat16)
-
-        params_to_change_dtype = [
-            "language_model.model.layers",
-            "gemma_expert.model.layers",
-            "vision_tower",
-            "multi_modal",
-        ]
-        for name, param in self.named_parameters():
-            if any(selector in name for selector in params_to_change_dtype):
-                param.data = param.data.to(dtype=torch.bfloat16)
-
-    def embed_image(self, image: torch.Tensor):
-        return self.paligemma.get_image_features(image)
-
-    def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.language_model.model.embed_tokens(tokens)
-
-    # TODO: break down this huge forward into modules or functions
-    def forward(
-        self,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None,
-        inputs_embeds: List[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        fill_kv_cache: Optional[bool] = None,
-    ):
-        models = [self.paligemma.language_model.model, self.gemma_expert.model]
-
-        for hidden_states in inputs_embeds:
-            # TODO this is very inefficient
-            # dtype is always the same, batch size too (if > 1 len)
-            # device could be trickier in multi gpu edge cases but that's it
-            if hidden_states is None:
-                continue
-            batch_size = hidden_states.shape[0]
-
-        # RMSNorm
-        num_layers = self.paligemma.config.text_config.num_hidden_layers
-        head_dim = self.paligemma.config.text_config.head_dim
-        for layer_idx in range(num_layers):
-            query_states = []
-            key_states = []
-            value_states = []
-            for i, hidden_states in enumerate(inputs_embeds):
-                if hidden_states is None:
-                    continue
-                layer = models[i].layers[layer_idx]
-                # normalizer = torch.tensor(models[i].config.hidden_size**0.5, dtype=hidden_states.dtype)
-                # hidden_states = hidden_states * normalizer
-                hidden_states = layer.input_layernorm(hidden_states)
-
-                input_shape = hidden_states.shape[:-1]
-                hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
-
-                hidden_states = hidden_states.to(dtype=torch.bfloat16)
-                query_state = layer.self_attn.q_proj(hidden_states).view(hidden_shape)
-                key_state = layer.self_attn.k_proj(hidden_states).view(hidden_shape)
-                value_state = layer.self_attn.v_proj(hidden_states).view(hidden_shape)
-
-                query_states.append(query_state)
-                key_states.append(key_state)
-                value_states.append(value_state)
-
-            # B,L,H,D with L sequence length, H number of heads, D head dim
-            # concatenate on the number of embeddings/tokens
-            query_states = torch.cat(query_states, dim=1)
-            key_states = torch.cat(key_states, dim=1)
-            value_states = torch.cat(value_states, dim=1)
-
-            query_states = apply_rope(query_states, position_ids)
-            key_states = apply_rope(key_states, position_ids)
-
-            if use_cache and past_key_values is None:
-                past_key_values = {}
-
-            if use_cache:
-                if fill_kv_cache:
-                    past_key_values[layer_idx] = {
-                        "key_states": key_states,
-                        "value_states": value_states,
-                    }
-                else:
-                    # TODO here, some optimization can be done - similar to a `StaticCache` we can declare the `max_len` before.
-                    # so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
-                    # the max len, then we (for instance) double the cache size. This implementation already exists
-                    # in `transformers`. (molbap)
-                    key_states = torch.cat([past_key_values[layer_idx]["key_states"], key_states], dim=1)
-                    value_states = torch.cat(
-                        [past_key_values[layer_idx]["value_states"], value_states], dim=1
-                    )
-
-            attention_interface = self.get_attention_interface()
-            att_output = attention_interface(
-                attention_mask, batch_size, head_dim, query_states, key_states, value_states
-            )
-            att_output = att_output.to(dtype=torch.bfloat16)
-
-            # first part of att_output is prefix (up to sequence length, [:, 0:prefix_seq_len])
-            outputs_embeds = []
-            start = 0
-            for i, hidden_states in enumerate(inputs_embeds):
-                layer = models[i].layers[layer_idx]
-
-                if hidden_states is not None:
-                    end = start + hidden_states.shape[1]
-
-                    if att_output.dtype != layer.self_attn.o_proj.weight.dtype:
-                        att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
-                    out_emb = layer.self_attn.o_proj(att_output[:, start:end])
-
-                    # TODO: first dropout (by default 0.0)
-
-                    # first residual
-                    out_emb += hidden_states
-                    after_first_residual = out_emb.clone()
-
-                    out_emb = layer.post_attention_layernorm(out_emb)
-                    out_emb = layer.mlp(out_emb)
-
-                    # TODO: second dropout (by default 0.0)
-
-                    # second residual
-                    out_emb += after_first_residual
-
-                    outputs_embeds.append(out_emb)
-
-                    start = end
-                else:
-                    outputs_embeds.append(None)
-
-            inputs_embeds = outputs_embeds
-
-        # final norm
-        outputs_embeds = []
-        for i, hidden_states in enumerate(inputs_embeds):
-            if hidden_states is not None:
-                out_emb = models[i].norm(hidden_states)
-                outputs_embeds.append(out_emb)
-            else:
-                outputs_embeds.append(None)
-
-        return outputs_embeds, past_key_values
-
-    def get_attention_interface(self):
-        if self.config.attention_implementation == "fa2":
-            attention_interface = self.flash_attention_forward
-        elif self.config.attention_implementation == "flex":
-            attention_interface = flex_attention_forward
-        else:
-            attention_interface = self.eager_attention_forward
-        return attention_interface
-
-    def flash_attention_forward(
-        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
-    ):
-        raise NotImplementedError("FA2 is not implemented (yet)")
-
-    def eager_attention_forward(
-        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
-    ):
-        num_att_heads = self.config.paligemma_config.text_config.num_attention_heads
-        num_key_value_heads = self.config.paligemma_config.text_config.num_key_value_heads
-        num_key_value_groups = num_att_heads // num_key_value_heads
-
-        # query_states: batch_size, sequence_length, num_att_head, head_dim
-        # key_states: batch_size, sequence_length, num_key_value_head, head_dim
-        # value_states: batch_size, sequence_length, num_key_value_head, head_dim
-        sequence_length = key_states.shape[1]
-
-        key_states = key_states[:, :, :, None, :].expand(
-            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
-        )
-        key_states = key_states.reshape(
-            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
-        )
-
-        value_states = value_states[:, :, :, None, :].expand(
-            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
-        )
-        value_states = value_states.reshape(
-            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
-        )
-
-        # Attention here is upcasted to float32 to match the original eager implementation.
-
-        query_states = query_states.to(dtype=torch.float32)
-        key_states = key_states.to(dtype=torch.float32)
-
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-
-        att_weights = torch.matmul(query_states, key_states.transpose(2, 3))
-        att_weights *= head_dim**-0.5
-        big_neg = -2.3819763e38  # See gemma/modules.py
-
-        masked_att_weights = torch.where(attention_mask[:, None, :, :], att_weights, big_neg)
-
-        probs = nn.functional.softmax(masked_att_weights, dim=-1)
-        probs = probs.to(dtype=value_states.dtype)
-
-        # probs: batch_size, num_key_value_head, num_att_head, sequence_length, sequence_length
-        # value_states: batch_size, sequence_length, num_att_heads, head_dim
-
-        att_output = torch.matmul(probs, value_states.permute(0, 2, 1, 3))
-
-        att_output = att_output.permute(0, 2, 1, 3)
-        # we use -1 because sequence length can change
-        att_output = att_output.reshape(batch_size, -1, num_key_value_heads * num_key_value_groups * head_dim)
-
-        return att_output

lerobot/common/policies/policy_protocol.py

@@ -0,0 +1,75 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""A protocol that all policies should follow.
+
+This provides a mechanism for type-hinting and isinstance checks without requiring the policies classes
+subclass a base class.
+
+The protocol structure, method signatures, and docstrings should be used by developers as a reference for
+how to implement new policies.
+"""
+
+from typing import Protocol, runtime_checkable
+
+from torch import Tensor
+
+
+@runtime_checkable
+class Policy(Protocol):
+    """The required interface for implementing a policy.
+
+    We also expect all policies to subclass torch.nn.Module and PyTorchModelHubMixin.
+    """
+
+    name: str
+
+    def __init__(self, cfg, dataset_stats: dict[str, dict[str, Tensor]] | None = None):
+        """
+        Args:
+            cfg: Policy configuration class instance or None, in which case the default instantiation of the
+                 configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization.
+        """
+
+    def reset(self):
+        """To be called whenever the environment is reset.
+
+        Does things like clearing caches.
+        """
+
+    def forward(self, batch: dict[str, Tensor]) -> dict:
+        """Run the batch through the model and compute the loss for training or validation.
+
+        Returns a dictionary with "loss" and potentially other information. Apart from "loss" which is a Tensor, all
+        other items should be logging-friendly, native Python types.
+        """
+
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Return one action to run in the environment (potentially in batch mode).
+
+        When the model uses a history of observations, or outputs a sequence of actions, this method deals
+        with caching.
+        """
+
+
+@runtime_checkable
+class PolicyWithUpdate(Policy, Protocol):
+    def update(self):
+        """An update method that is to be called after a training optimization step.
+
+        Implements an additional updates the model parameters may need (for example, doing an EMA step for a
+        target model, or incrementing an internal buffer).
+        """

lerobot/common/policies/pretrained.py

@@ -1,187 +0,0 @@
-import abc
-import logging
-import os
-from pathlib import Path
-from typing import Type, TypeVar
-
-import packaging
-import safetensors
-from huggingface_hub import hf_hub_download
-from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
-from huggingface_hub.errors import HfHubHTTPError
-from safetensors.torch import load_model as load_model_as_safetensor
-from safetensors.torch import save_model as save_model_as_safetensor
-from torch import Tensor, nn
-
-from lerobot.common.utils.hub import HubMixin
-from lerobot.configs.policies import PreTrainedConfig
-
-T = TypeVar("T", bound="PreTrainedPolicy")
-
-DEFAULT_POLICY_CARD = """
----
-# For reference on model card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/modelcard.md?plain=1
-# Doc / guide: https://huggingface.co/docs/hub/model-cards
-{{ card_data }}
----
-
-This policy has been pushed to the Hub using [LeRobot](https://github.com/huggingface/lerobot):
-- Docs: {{ docs_url | default("[More Information Needed]", true) }}
-"""
-
-
-class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
-    """
-    Base class for policy models.
-    """
-
-    config_class: None
-    name: None
-
-    def __init__(self, config: PreTrainedConfig, *inputs, **kwargs):
-        super().__init__()
-        if not isinstance(config, PreTrainedConfig):
-            raise ValueError(
-                f"Parameter config in `{self.__class__.__name__}(config)` should be an instance of class "
-                "`PreTrainedConfig`. To create a model from a pretrained model use "
-                f"`model = {self.__class__.__name__}.from_pretrained(PRETRAINED_MODEL_NAME)`"
-            )
-        self.config = config
-
-    def __init_subclass__(cls, **kwargs):
-        super().__init_subclass__(**kwargs)
-        if not getattr(cls, "config_class", None):
-            raise TypeError(f"Class {cls.__name__} must define 'config_class'")
-        if not getattr(cls, "name", None):
-            raise TypeError(f"Class {cls.__name__} must define 'name'")
-
-    def _save_pretrained(self, save_directory: Path) -> None:
-        self.config._save_pretrained(save_directory)
-        model_to_save = self.module if hasattr(self, "module") else self
-        save_model_as_safetensor(model_to_save, str(save_directory / SAFETENSORS_SINGLE_FILE))
-
-    @classmethod
-    def from_pretrained(
-        cls: Type[T],
-        pretrained_name_or_path: str | Path,
-        *,
-        config: PreTrainedConfig | None = None,
-        force_download: bool = False,
-        resume_download: bool | None = None,
-        proxies: dict | None = None,
-        token: str | bool | None = None,
-        cache_dir: str | Path | None = None,
-        local_files_only: bool = False,
-        revision: str | None = None,
-        map_location: str = "cpu",
-        strict: bool = False,
-        **kwargs,
-    ) -> T:
-        """
-        The policy is set in evaluation mode by default using `policy.eval()` (dropout modules are
-        deactivated). To train it, you should first set it back in training mode with `policy.train()`.
-        """
-        if config is None:
-            config = PreTrainedConfig.from_pretrained(
-                pretrained_name_or_path=pretrained_name_or_path,
-                force_download=force_download,
-                resume_download=resume_download,
-                proxies=proxies,
-                token=token,
-                cache_dir=cache_dir,
-                local_files_only=local_files_only,
-                revision=revision,
-                **kwargs,
-            )
-        model_id = str(pretrained_name_or_path)
-        instance = cls(config, **kwargs)
-        if os.path.isdir(model_id):
-            print("Loading weights from local directory")
-            model_file = os.path.join(model_id, SAFETENSORS_SINGLE_FILE)
-            policy = cls._load_as_safetensor(instance, model_file, map_location, strict)
-        else:
-            try:
-                model_file = hf_hub_download(
-                    repo_id=model_id,
-                    filename=SAFETENSORS_SINGLE_FILE,
-                    revision=revision,
-                    cache_dir=cache_dir,
-                    force_download=force_download,
-                    proxies=proxies,
-                    resume_download=resume_download,
-                    token=token,
-                    local_files_only=local_files_only,
-                )
-                policy = cls._load_as_safetensor(instance, model_file, map_location, strict)
-            except HfHubHTTPError as e:
-                raise FileNotFoundError(
-                    f"{SAFETENSORS_SINGLE_FILE} not found on the HuggingFace Hub in {model_id}"
-                ) from e
-
-        policy.to(map_location)
-        policy.eval()
-        return policy
-
-    @classmethod
-    def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
-        if packaging.version.parse(safetensors.__version__) < packaging.version.parse("0.4.3"):
-            load_model_as_safetensor(model, model_file, strict=strict)
-            if map_location != "cpu":
-                logging.warning(
-                    "Loading model weights on other devices than 'cpu' is not supported natively in your version of safetensors."
-                    " This means that the model is loaded on 'cpu' first and then copied to the device."
-                    " This leads to a slower loading time."
-                    " Please update safetensors to version 0.4.3 or above for improved performance."
-                )
-                model.to(map_location)
-        else:
-            safetensors.torch.load_model(model, model_file, strict=strict, device=map_location)
-        return model
-
-    # def generate_model_card(self, *args, **kwargs) -> ModelCard:
-    #     card = ModelCard.from_template(
-    #         card_data=self._hub_mixin_info.model_card_data,
-    #         template_str=self._hub_mixin_info.model_card_template,
-    #         repo_url=self._hub_mixin_info.repo_url,
-    #         docs_url=self._hub_mixin_info.docs_url,
-    #         **kwargs,
-    #     )
-    #     return card
-
-    @abc.abstractmethod
-    def get_optim_params(self) -> dict:
-        """
-        Returns the policy-specific parameters dict to be passed on to the optimizer.
-        """
-        raise NotImplementedError
-
-    @abc.abstractmethod
-    def reset(self):
-        """To be called whenever the environment is reset.
-
-        Does things like clearing caches.
-        """
-        raise NotImplementedError
-
-    # TODO(aliberts, rcadene): split into 'forward' and 'compute_loss'?
-    @abc.abstractmethod
-    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict | None]:
-        """_summary_
-
-        Args:
-            batch (dict[str, Tensor]): _description_
-
-        Returns:
-            tuple[Tensor, dict | None]: The loss and potentially other information. Apart from the loss which
-                is a Tensor, all other items should be logging-friendly, native Python types.
-        """
-        raise NotImplementedError
-
-    @abc.abstractmethod
-    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
-        """Return one action to run in the environment (potentially in batch mode).
-
-        When the model uses a history of observations, or outputs a sequence of actions, this method deals
-        with caching.
-        """
-        raise NotImplementedError

lerobot/common/policies/tdmpc/configuration_tdmpc.py

@@ -16,14 +16,9 @@
 # limitations under the License.
 from dataclasses import dataclass, field
 
-from lerobot.common.optim.optimizers import AdamConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
 
-
-@PreTrainedConfig.register_subclass("tdmpc")
 @dataclass
-class TDMPCConfig(PreTrainedConfig):
+class TDMPCConfig:
     """Configuration class for TDMPCPolicy.
 
     Defaults are configured for training with xarm_lift_medium_replay providing proprioceptive and single
@@ -107,19 +102,27 @@ class TDMPCConfig(PreTrainedConfig):
     """
 
     # Input / output structure.
-    n_obs_steps: int = 1
     n_action_repeats: int = 2
     horizon: int = 5
     n_action_steps: int = 1
 
-    normalization_mapping: dict[str, NormalizationMode] = field(
+    input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
-            "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.IDENTITY,
-            "ENV": NormalizationMode.IDENTITY,
-            "ACTION": NormalizationMode.MIN_MAX,
+            "observation.image": [3, 84, 84],
+            "observation.state": [4],
         }
     )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [4],
+        }
+    )
+
+    # Normalization / Unnormalization
+    input_normalization_modes: dict[str, str] | None = None
+    output_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {"action": "min_max"},
+    )
 
     # Architecture / modeling.
     # Neural networks.
@@ -156,27 +159,32 @@ class TDMPCConfig(PreTrainedConfig):
     # Target model.
     target_model_momentum: float = 0.995
 
-    # Training presets
-    optimizer_lr: float = 3e-4
-
     def __post_init__(self):
-        super().__post_init__()
-
         """Input validation (not exhaustive)."""
+        # There should only be one image key.
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
+        if len(image_keys) > 1:
+            raise ValueError(
+                f"{self.__class__.__name__} handles at most one image for now. Got image keys {image_keys}."
+            )
+        if len(image_keys) > 0:
+            image_key = next(iter(image_keys))
+            if self.input_shapes[image_key][-2] != self.input_shapes[image_key][-1]:
+                # TODO(alexander-soare): This limitation is solely because of code in the random shift
+                # augmentation. It should be able to be removed.
+                raise ValueError(
+                    f"Only square images are handled now. Got image shape {self.input_shapes[image_key]}."
+                )
         if self.n_gaussian_samples <= 0:
             raise ValueError(
                 f"The number of guassian samples for CEM should be non-zero. Got `{self.n_gaussian_samples=}`"
             )
-        if self.normalization_mapping["ACTION"] is not NormalizationMode.MIN_MAX:
+        if self.output_normalization_modes != {"action": "min_max"}:
             raise ValueError(
                 "TD-MPC assumes the action space dimensions to all be in [-1, 1]. Therefore it is strongly "
                 f"advised that you stick with the default. See {self.__class__.__name__} docstring for more "
                 "information."
             )
-        if self.n_obs_steps != 1:
-            raise ValueError(
-                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
-            )
         if self.n_action_steps > 1:
             if self.n_action_repeats != 1:
                 raise ValueError(
@@ -186,35 +194,3 @@ class TDMPCConfig(PreTrainedConfig):
                 raise ValueError("If `n_action_steps > 1`, `use_mpc` must be set to `True`.")
             if self.n_action_steps > self.horizon:
                 raise ValueError("`n_action_steps` must be less than or equal to `horizon`.")
-
-    def get_optimizer_preset(self) -> AdamConfig:
-        return AdamConfig(lr=self.optimizer_lr)
-
-    def get_scheduler_preset(self) -> None:
-        return None
-
-    def validate_features(self) -> None:
-        # There should only be one image key.
-        if len(self.image_features) > 1:
-            raise ValueError(
-                f"{self.__class__.__name__} handles at most one image for now. Got image keys {self.image_features}."
-            )
-
-        if len(self.image_features) > 0:
-            image_ft = next(iter(self.image_features.values()))
-            if image_ft.shape[-2] != image_ft.shape[-1]:
-                # TODO(alexander-soare): This limitation is solely because of code in the random shift
-                # augmentation. It should be able to be removed.
-                raise ValueError(f"Only square images are handled now. Got image shape {image_ft.shape}.")
-
-    @property
-    def observation_delta_indices(self) -> list:
-        return list(range(self.horizon + 1))
-
-    @property
-    def action_delta_indices(self) -> list:
-        return list(range(self.horizon))
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return list(range(self.horizon))

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -33,16 +33,21 @@ import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F  # noqa: N812
+from huggingface_hub import PyTorchModelHubMixin
 from torch import Tensor
 
-from lerobot.common.constants import OBS_ENV, OBS_ROBOT
 from lerobot.common.policies.normalize import Normalize, Unnormalize
-from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
-from lerobot.common.policies.utils import get_device_from_parameters, get_output_shape, populate_queues
+from lerobot.common.policies.utils import get_device_from_parameters, populate_queues
 
 
-class TDMPCPolicy(PreTrainedPolicy):
+class TDMPCPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "tdmpc"],
+):
     """Implementation of TD-MPC learning + inference.
 
     Please note several warnings for this policy.
@@ -60,10 +65,11 @@ class TDMPCPolicy(PreTrainedPolicy):
           match our xarm environment.
     """
 
-    config_class = TDMPCConfig
     name = "tdmpc"
 
-    def __init__(self, config: TDMPCConfig, dataset_stats: dict[str, dict[str, Tensor]] | None = None):
+    def __init__(
+        self, config: TDMPCConfig | None = None, dataset_stats: dict[str, dict[str, Tensor]] | None = None
+    ):
         """
         Args:
             config: Policy configuration class instance or None, in which case the default instantiation of
@@ -71,27 +77,41 @@ class TDMPCPolicy(PreTrainedPolicy):
             dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
-        super().__init__(config)
-        config.validate_features()
+        super().__init__()
+
+        if config is None:
+            config = TDMPCConfig()
         self.config = config
-
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
         self.model = TDMPCTOLD(config)
         self.model_target = deepcopy(self.model)
         for param in self.model_target.parameters():
             param.requires_grad = False
 
-        self.reset()
+        if config.input_normalization_modes is not None:
+            self.normalize_inputs = Normalize(
+                config.input_shapes, config.input_normalization_modes, dataset_stats
+            )
+        else:
+            self.normalize_inputs = nn.Identity()
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
 
-    def get_optim_params(self) -> dict:
-        return self.parameters()
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        # Note: This check is covered in the post-init of the config but have a sanity check just in case.
+        self._use_image = False
+        self._use_env_state = False
+        if len(image_keys) > 0:
+            assert len(image_keys) == 1
+            self._use_image = True
+            self.input_image_key = image_keys[0]
+        if "observation.environment_state" in config.input_shapes:
+            self._use_env_state = True
+
+        self.reset()
 
     def reset(self):
         """
@@ -102,9 +122,9 @@ class TDMPCPolicy(PreTrainedPolicy):
             "observation.state": deque(maxlen=1),
             "action": deque(maxlen=max(self.config.n_action_steps, self.config.n_action_repeats)),
         }
-        if self.config.image_features:
+        if self._use_image:
             self._queues["observation.image"] = deque(maxlen=1)
-        if self.config.env_state_feature:
+        if self._use_env_state:
             self._queues["observation.environment_state"] = deque(maxlen=1)
         # Previous mean obtained from the cross-entropy method (CEM) used during MPC. It is used to warm start
         # CEM for the next step.
@@ -114,9 +134,9 @@ class TDMPCPolicy(PreTrainedPolicy):
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select a single action given environment observations."""
         batch = self.normalize_inputs(batch)
-        if self.config.image_features:
+        if self._use_image:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.image"] = batch[next(iter(self.config.image_features))]
+            batch["observation.image"] = batch[self.input_image_key]
 
         self._queues = populate_queues(self._queues, batch)
 
@@ -131,9 +151,9 @@ class TDMPCPolicy(PreTrainedPolicy):
 
             # NOTE: Order of observations matters here.
             encode_keys = []
-            if self.config.image_features:
+            if self._use_image:
                 encode_keys.append("observation.image")
-            if self.config.env_state_feature:
+            if self._use_env_state:
                 encode_keys.append("observation.environment_state")
             encode_keys.append("observation.state")
             z = self.model.encode({k: batch[k] for k in encode_keys})
@@ -176,7 +196,7 @@ class TDMPCPolicy(PreTrainedPolicy):
             self.config.horizon,
             self.config.n_pi_samples,
             batch_size,
-            self.config.action_feature.shape[0],
+            self.config.output_shapes["action"][0],
             device=device,
         )
         if self.config.n_pi_samples > 0:
@@ -195,7 +215,7 @@ class TDMPCPolicy(PreTrainedPolicy):
         # algorithm.
         # The initial mean and standard deviation for the cross-entropy method (CEM).
         mean = torch.zeros(
-            self.config.horizon, batch_size, self.config.action_feature.shape[0], device=device
+            self.config.horizon, batch_size, self.config.output_shapes["action"][0], device=device
         )
         # Maybe warm start CEM with the mean from the previous step.
         if self._prev_mean is not None:
@@ -208,7 +228,7 @@ class TDMPCPolicy(PreTrainedPolicy):
                 self.config.horizon,
                 self.config.n_gaussian_samples,
                 batch_size,
-                self.config.action_feature.shape[0],
+                self.config.output_shapes["action"][0],
                 device=std.device,
             )
             gaussian_actions = torch.clamp(mean.unsqueeze(1) + std.unsqueeze(1) * std_normal_noise, -1, 1)
@@ -302,7 +322,7 @@ class TDMPCPolicy(PreTrainedPolicy):
             G -= running_discount * self.config.uncertainty_regularizer_coeff * terminal_values.std(0)
         return G
 
-    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor | float]:
         """Run the batch through the model and compute the loss.
 
         Returns a dictionary with loss as a tensor, and other information as native floats.
@@ -310,16 +330,16 @@ class TDMPCPolicy(PreTrainedPolicy):
         device = get_device_from_parameters(self)
 
         batch = self.normalize_inputs(batch)
-        if self.config.image_features:
+        if self._use_image:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-            batch["observation.image"] = batch[next(iter(self.config.image_features))]
+            batch["observation.image"] = batch[self.input_image_key]
         batch = self.normalize_targets(batch)
 
         info = {}
 
         # (b, t) -> (t, b)
         for key in batch:
-            if isinstance(batch[key], torch.Tensor) and batch[key].ndim > 1:
+            if batch[key].ndim > 1:
                 batch[key] = batch[key].transpose(1, 0)
 
         action = batch["action"]  # (t, b, action_dim)
@@ -327,7 +347,7 @@ class TDMPCPolicy(PreTrainedPolicy):
         observations = {k: v for k, v in batch.items() if k.startswith("observation.")}
 
         # Apply random image augmentations.
-        if self.config.image_features and self.config.max_random_shift_ratio > 0:
+        if self._use_image and self.config.max_random_shift_ratio > 0:
             observations["observation.image"] = flatten_forward_unflatten(
                 partial(random_shifts_aug, max_random_shift_ratio=self.config.max_random_shift_ratio),
                 observations["observation.image"],
@@ -340,7 +360,7 @@ class TDMPCPolicy(PreTrainedPolicy):
             current_observation[k] = observations[k][0]
             next_observations[k] = observations[k][1:]
         horizon, batch_size = next_observations[
-            "observation.image" if self.config.image_features else "observation.environment_state"
+            "observation.image" if self._use_image else "observation.environment_state"
         ].shape[:2]
 
         # Run latent rollout using the latent dynamics model and policy model.
@@ -495,16 +515,17 @@ class TDMPCPolicy(PreTrainedPolicy):
                 "Q_value_loss": q_value_loss.item(),
                 "V_value_loss": v_value_loss.item(),
                 "pi_loss": pi_loss.item(),
+                "loss": loss,
                 "sum_loss": loss.item() * self.config.horizon,
             }
         )
 
         # Undo (b, t) -> (t, b).
         for key in batch:
-            if isinstance(batch[key], torch.Tensor) and batch[key].ndim > 1:
+            if batch[key].ndim > 1:
                 batch[key] = batch[key].transpose(1, 0)
 
-        return loss, info
+        return info
 
     def update(self):
         """Update the target model's parameters with an EMA step."""
@@ -522,7 +543,7 @@ class TDMPCTOLD(nn.Module):
         self.config = config
         self._encoder = TDMPCObservationEncoder(config)
         self._dynamics = nn.Sequential(
-            nn.Linear(config.latent_dim + config.action_feature.shape[0], config.mlp_dim),
+            nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
             nn.LayerNorm(config.mlp_dim),
             nn.Mish(),
             nn.Linear(config.mlp_dim, config.mlp_dim),
@@ -533,7 +554,7 @@ class TDMPCTOLD(nn.Module):
             nn.Sigmoid(),
         )
         self._reward = nn.Sequential(
-            nn.Linear(config.latent_dim + config.action_feature.shape[0], config.mlp_dim),
+            nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
             nn.LayerNorm(config.mlp_dim),
             nn.Mish(),
             nn.Linear(config.mlp_dim, config.mlp_dim),
@@ -548,12 +569,12 @@ class TDMPCTOLD(nn.Module):
             nn.Linear(config.mlp_dim, config.mlp_dim),
             nn.LayerNorm(config.mlp_dim),
             nn.Mish(),
-            nn.Linear(config.mlp_dim, config.action_feature.shape[0]),
+            nn.Linear(config.mlp_dim, config.output_shapes["action"][0]),
         )
         self._Qs = nn.ModuleList(
             [
                 nn.Sequential(
-                    nn.Linear(config.latent_dim + config.action_feature.shape[0], config.mlp_dim),
+                    nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
                     nn.LayerNorm(config.mlp_dim),
                     nn.Tanh(),
                     nn.Linear(config.mlp_dim, config.mlp_dim),
@@ -594,9 +615,9 @@ class TDMPCTOLD(nn.Module):
 
         self.apply(_apply_fn)
         for m in [self._reward, *self._Qs]:
-            assert isinstance(m[-1], nn.Linear), (
-                "Sanity check. The last linear layer needs 0 initialization on weights."
-            )
+            assert isinstance(
+                m[-1], nn.Linear
+            ), "Sanity check. The last linear layer needs 0 initialization on weights."
             nn.init.zeros_(m[-1].weight)
             nn.init.zeros_(m[-1].bias)  # this has already been done, but keep this line here for good measure
 
@@ -693,13 +714,10 @@ class TDMPCObservationEncoder(nn.Module):
         super().__init__()
         self.config = config
 
-        if config.image_features:
+        if "observation.image" in config.input_shapes:
             self.image_enc_layers = nn.Sequential(
                 nn.Conv2d(
-                    next(iter(config.image_features.values())).shape[0],
-                    config.image_encoder_hidden_dim,
-                    7,
-                    stride=2,
+                    config.input_shapes["observation.image"][0], config.image_encoder_hidden_dim, 7, stride=2
                 ),
                 nn.ReLU(),
                 nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 5, stride=2),
@@ -709,8 +727,9 @@ class TDMPCObservationEncoder(nn.Module):
                 nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
                 nn.ReLU(),
             )
-            dummy_shape = (1, *next(iter(config.image_features.values())).shape)
-            out_shape = get_output_shape(self.image_enc_layers, dummy_shape)[1:]
+            dummy_batch = torch.zeros(1, *config.input_shapes["observation.image"])
+            with torch.inference_mode():
+                out_shape = self.image_enc_layers(dummy_batch).shape[1:]
             self.image_enc_layers.extend(
                 nn.Sequential(
                     nn.Flatten(),
@@ -719,19 +738,19 @@ class TDMPCObservationEncoder(nn.Module):
                     nn.Sigmoid(),
                 )
             )
-
-        if config.robot_state_feature:
+        if "observation.state" in config.input_shapes:
             self.state_enc_layers = nn.Sequential(
-                nn.Linear(config.robot_state_feature.shape[0], config.state_encoder_hidden_dim),
+                nn.Linear(config.input_shapes["observation.state"][0], config.state_encoder_hidden_dim),
                 nn.ELU(),
                 nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
                 nn.LayerNorm(config.latent_dim),
                 nn.Sigmoid(),
             )
-
-        if config.env_state_feature:
+        if "observation.environment_state" in config.input_shapes:
             self.env_state_enc_layers = nn.Sequential(
-                nn.Linear(config.env_state_feature.shape[0], config.state_encoder_hidden_dim),
+                nn.Linear(
+                    config.input_shapes["observation.environment_state"][0], config.state_encoder_hidden_dim
+                ),
                 nn.ELU(),
                 nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
                 nn.LayerNorm(config.latent_dim),
@@ -746,16 +765,12 @@ class TDMPCObservationEncoder(nn.Module):
         """
         feat = []
         # NOTE: Order of observations matters here.
-        if self.config.image_features:
-            feat.append(
-                flatten_forward_unflatten(
-                    self.image_enc_layers, obs_dict[next(iter(self.config.image_features))]
-                )
-            )
-        if self.config.env_state_feature:
-            feat.append(self.env_state_enc_layers(obs_dict[OBS_ENV]))
-        if self.config.robot_state_feature:
-            feat.append(self.state_enc_layers(obs_dict[OBS_ROBOT]))
+        if "observation.image" in self.config.input_shapes:
+            feat.append(flatten_forward_unflatten(self.image_enc_layers, obs_dict["observation.image"]))
+        if "observation.environment_state" in self.config.input_shapes:
+            feat.append(self.env_state_enc_layers(obs_dict["observation.environment_state"]))
+        if "observation.state" in self.config.input_shapes:
+            feat.append(self.state_enc_layers(obs_dict["observation.state"]))
         return torch.stack(feat, dim=0).mean(0)
 
 

lerobot/common/policies/tdmpc2/configuration_tdmpc2.py

@@ -0,0 +1,193 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Nicklas Hansen, Xiaolong Wang, Hao Su,
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass, field
+
+
+@dataclass
+class TDMPC2Config:
+    """Configuration class for TDMPC2Policy.
+
+    Defaults are configured for training with xarm_lift_medium_replay providing proprioceptive and single
+    camera observations.
+
+    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
+    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift_ratio`.
+
+    Args:
+        n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
+            action repeats in Q-learning or ask your favorite chatbot)
+        horizon: Horizon for model predictive control.
+        n_action_steps: Number of action steps to take from the plan given by model predictive control. This
+            is an alternative to using action repeats. If this is set to more than 1, then we require
+            `n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
+            approach of using multiple steps from the plan is not in the original implementation.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range. Note that here this defaults to None meaning inputs are not normalized. This is to
+            match the original implementation.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets. NOTE: Clipping
+            to [-1, +1] is used during MPPI/CEM. Therefore, it is recommended that you stick with "min_max"
+            normalization mode here.
+        image_encoder_hidden_dim: Number of channels for the convolutional layers used for image encoding.
+        state_encoder_hidden_dim: Hidden dimension for MLP used for state vector encoding.
+        latent_dim: Observation's latent embedding dimension.
+        q_ensemble_size: Number of Q function estimators to use in an ensemble for uncertainty estimation.
+        mlp_dim: Hidden dimension of MLPs used for modelling the dynamics encoder, reward function, policy
+            (π), Q ensemble, and V.
+        discount: Discount factor (γ) to use for the reinforcement learning formalism.
+        use_mpc: Whether to use model predictive control. The alternative is to just sample the policy model
+            (π) for each step.
+        cem_iterations: Number of iterations for the MPPI/CEM loop in MPC.
+        max_std: Maximum standard deviation for actions sampled from the gaussian PDF in CEM.
+        min_std: Minimum standard deviation for noise applied to actions sampled from the policy model (π).
+            Doubles up as the minimum standard deviation for actions sampled from the gaussian PDF in CEM.
+        n_gaussian_samples: Number of samples to draw from the gaussian distribution every CEM iteration. Must
+            be non-zero.
+        n_pi_samples: Number of samples to draw from the policy / world model rollout every CEM iteration. Can
+            be zero.
+        n_elites: The number of elite samples to use for updating the gaussian parameters every CEM iteration.
+        elite_weighting_temperature: The temperature to use for softmax weighting (by trajectory value) of the
+            elites, when updating the gaussian parameters for CEM.
+        max_random_shift_ratio: Maximum random shift (as a proportion of the image size) to apply to the
+            image(s) (in units of pixels) for training-time augmentation. If set to 0, no such augmentation
+            is applied. Note that the input images are assumed to be square for this augmentation.
+        reward_coeff: Loss weighting coefficient for the reward regression loss.
+        value_coeff: Loss weighting coefficient for both the state-action value (Q) TD loss, and the state
+            value (V) expectile regression loss.
+        consistency_coeff: Loss weighting coefficient for the consistency loss.
+        temporal_decay_coeff: Exponential decay coefficient for decaying the loss coefficient for future time-
+            steps. Hint: each loss computation involves `horizon` steps worth of actions starting from the
+            current time step.
+        target_model_momentum: Momentum (α) used for EMA updates of the target models. Updates are calculated
+            as ϕ ← αϕ + (1-α)θ where ϕ are the parameters of the target model and θ are the parameters of the
+            model being trained.
+    """
+
+    # Input / output structure.
+    n_action_repeats: int = 1
+    horizon: int = 3
+    n_action_steps: int = 1
+
+    input_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "observation.image": [3, 84, 84],
+            "observation.state": [4],
+        }
+    )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [4],
+        }
+    )
+
+    # Normalization / Unnormalization
+    input_normalization_modes: dict[str, str] | None = None
+    output_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {"action": "min_max"},
+    )
+
+    # Architecture / modeling.
+    # Neural networks.
+    image_encoder_hidden_dim: int = 32
+    state_encoder_hidden_dim: int = 256
+    latent_dim: int = 512
+    q_ensemble_size: int = 5
+    num_enc_layers: int = 2
+    mlp_dim: int = 512
+    # Reinforcement learning.
+    discount: float = 0.9
+    simnorm_dim: int = 8
+    dropout: float = 0.01
+
+    # actor
+    log_std_min: float = -10
+    log_std_max: float = 2
+
+    # critic
+    num_bins: int = 101
+    vmin: int = -10
+    vmax: int = +10
+
+    # Inference.
+    use_mpc: bool = True
+    cem_iterations: int = 6
+    max_std: float = 2.0
+    min_std: float = 0.05
+    n_gaussian_samples: int = 512
+    n_pi_samples: int = 24
+    n_elites: int = 64
+    elite_weighting_temperature: float = 0.5
+
+    # Training and loss computation.
+    max_random_shift_ratio: float = 0.0476
+    # Loss coefficients.
+    reward_coeff: float = 0.1
+    value_coeff: float = 0.1
+    consistency_coeff: float = 20.0
+    entropy_coef: float = 1e-4
+    temporal_decay_coeff: float = 0.5
+    # Target model. NOTE (michel_aractingi) this is equivelant to
+    # 1 - target_model_momentum of our TD-MPC1 implementation because
+    # of the use of `torch.lerp`
+    target_model_momentum: float = 0.01
+
+    def __post_init__(self):
+        """Input validation (not exhaustive)."""
+        # There should only be one image key.
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
+        if len(image_keys) > 1:
+            raise ValueError(
+                f"{self.__class__.__name__} handles at most one image for now. Got image keys {image_keys}."
+            )
+        if len(image_keys) > 0:
+            image_key = next(iter(image_keys))
+            if self.input_shapes[image_key][-2] != self.input_shapes[image_key][-1]:
+                # TODO(alexander-soare): This limitation is solely because of code in the random shift
+                # augmentation. It should be able to be removed.
+                raise ValueError(
+                    f"Only square images are handled now. Got image shape {self.input_shapes[image_key]}."
+                )
+        if self.n_gaussian_samples <= 0:
+            raise ValueError(
+                f"The number of guassian samples for CEM should be non-zero. Got `{self.n_gaussian_samples=}`"
+            )
+        if self.output_normalization_modes != {"action": "min_max"}:
+            raise ValueError(
+                "TD-MPC assumes the action space dimensions to all be in [-1, 1]. Therefore it is strongly "
+                f"advised that you stick with the default. See {self.__class__.__name__} docstring for more "
+                "information."
+            )
+        if self.n_action_steps > 1:
+            if self.n_action_repeats != 1:
+                raise ValueError(
+                    "If `n_action_steps > 1`, `n_action_repeats` must be left to its default value of 1."
+                )
+            if not self.use_mpc:
+                raise ValueError("If `n_action_steps > 1`, `use_mpc` must be set to `True`.")
+            if self.n_action_steps > self.horizon:
+                raise ValueError("`n_action_steps` must be less than or equal to `horizon`.")

lerobot/common/policies/tdmpc2/modeling_tdmpc2.py

@@ -0,0 +1,834 @@
+#!/usr/bin/env python
+
+# Copyright 2024 Nicklas Hansen and The HuggingFace Inc. team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Implementation of TD-MPC2: Scalable, Robust World Models for Continuous Control
+
+We refer to the main paper and codebase:
+    TD-MPC2 paper: (https://arxiv.org/abs/2310.16828)
+    TD-MPC2 code:  (https://github.com/nicklashansen/tdmpc2)
+"""
+
+# ruff: noqa: N806
+
+from collections import deque
+from copy import deepcopy
+from functools import partial
+from typing import Callable
+
+import einops
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from huggingface_hub import PyTorchModelHubMixin
+from torch import Tensor
+
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.tdmpc2.configuration_tdmpc2 import TDMPC2Config
+from lerobot.common.policies.tdmpc2.tdmpc2_utils import (
+    NormedLinear,
+    SimNorm,
+    gaussian_logprob,
+    soft_cross_entropy,
+    squash,
+    two_hot_inv,
+)
+from lerobot.common.policies.utils import get_device_from_parameters, populate_queues
+
+
+class TDMPC2Policy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "tdmpc2"],
+):
+    """Implementation of TD-MPC2 learning + inference."""
+
+    name = "tdmpc2"
+
+    def __init__(
+        self, config: TDMPC2Config | None = None, dataset_stats: dict[str, dict[str, Tensor]] | None = None
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+        super().__init__()
+
+        if config is None:
+            config = TDMPC2Config()
+        self.config = config
+        self.model = TDMPC2WorldModel(config)
+        # TODO (michel-aractingi) temp fix for gpu
+        self.model = self.model.to("cuda:0")
+
+        if config.input_normalization_modes is not None:
+            self.normalize_inputs = Normalize(
+                config.input_shapes, config.input_normalization_modes, dataset_stats
+            )
+        else:
+            self.normalize_inputs = nn.Identity()
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        # Note: This check is covered in the post-init of the config but have a sanity check just in case.
+        self._use_image = False
+        self._use_env_state = False
+        if len(image_keys) > 0:
+            assert len(image_keys) == 1
+            self._use_image = True
+            self.input_image_key = image_keys[0]
+        if "observation.environment_state" in config.input_shapes:
+            self._use_env_state = True
+
+        self.scale = RunningScale(self.config.target_model_momentum)
+        self.discount = (
+            self.config.discount
+        )  # TODO (michel-aractingi) downscale discount according to episode length
+
+        self.reset()
+
+    def reset(self):
+        """
+        Clear observation and action queues. Clear previous means for warm starting of MPPI/CEM. Should be
+        called on `env.reset()`
+        """
+        self._queues = {
+            "observation.state": deque(maxlen=1),
+            "action": deque(maxlen=max(self.config.n_action_steps, self.config.n_action_repeats)),
+        }
+        if self._use_image:
+            self._queues["observation.image"] = deque(maxlen=1)
+        if self._use_env_state:
+            self._queues["observation.environment_state"] = deque(maxlen=1)
+        # Previous mean obtained from the cross-entropy method (CEM) used during MPC. It is used to warm start
+        # CEM for the next step.
+        self._prev_mean: torch.Tensor | None = None
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations."""
+        batch = self.normalize_inputs(batch)
+        if self._use_image:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.image"] = batch[self.input_image_key]
+
+        self._queues = populate_queues(self._queues, batch)
+
+        # When the action queue is depleted, populate it again by querying the policy.
+        if len(self._queues["action"]) == 0:
+            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
+
+            # Remove the time dimensions as it is not handled yet.
+            for key in batch:
+                assert batch[key].shape[1] == 1
+                batch[key] = batch[key][:, 0]
+
+            # NOTE: Order of observations matters here.
+            encode_keys = []
+            if self._use_image:
+                encode_keys.append("observation.image")
+            if self._use_env_state:
+                encode_keys.append("observation.environment_state")
+            encode_keys.append("observation.state")
+            z = self.model.encode({k: batch[k] for k in encode_keys})
+            if self.config.use_mpc:  # noqa: SIM108
+                actions = self.plan(z)  # (horizon, batch, action_dim)
+            else:
+                # Plan with the policy (π) alone. This always returns one action so unsqueeze to get a
+                # sequence dimension like in the MPC branch.
+                actions = self.model.pi(z)[0].unsqueeze(0)
+
+            actions = torch.clamp(actions, -1, +1)
+
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.n_action_repeats > 1:
+                for _ in range(self.config.n_action_repeats):
+                    self._queues["action"].append(actions[0])
+            else:
+                # Action queue is (n_action_steps, batch_size, action_dim), so we transpose the action.
+                self._queues["action"].extend(actions[: self.config.n_action_steps])
+
+        action = self._queues["action"].popleft()
+        return action
+
+    @torch.no_grad()
+    def plan(self, z: Tensor) -> Tensor:
+        """Plan sequence of actions using TD-MPC inference.
+
+        Args:
+            z: (batch, latent_dim,) tensor for the initial state.
+        Returns:
+            (horizon, batch, action_dim,) tensor for the planned trajectory of actions.
+        """
+        device = get_device_from_parameters(self)
+
+        batch_size = z.shape[0]
+
+        # Sample Nπ trajectories from the policy.
+        pi_actions = torch.empty(
+            self.config.horizon,
+            self.config.n_pi_samples,
+            batch_size,
+            self.config.output_shapes["action"][0],
+            device=device,
+        )
+        if self.config.n_pi_samples > 0:
+            _z = einops.repeat(z, "b d -> n b d", n=self.config.n_pi_samples)
+            for t in range(self.config.horizon):
+                # Note: Adding a small amount of noise here doesn't hurt during inference and may even be
+                # helpful for CEM.
+                pi_actions[t] = self.model.pi(_z)[0]
+                _z = self.model.latent_dynamics(_z, pi_actions[t])
+
+        # In the CEM loop we will need this for a call to estimate_value with the gaussian sampled
+        # trajectories.
+        z = einops.repeat(z, "b d -> n b d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)
+
+        # Model Predictive Path Integral (MPPI) with the cross-entropy method (CEM) as the optimization
+        # algorithm.
+        # The initial mean and standard deviation for the cross-entropy method (CEM).
+        mean = torch.zeros(
+            self.config.horizon, batch_size, self.config.output_shapes["action"][0], device=device
+        )
+        # Maybe warm start CEM with the mean from the previous step.
+        if self._prev_mean is not None:
+            mean[:-1] = self._prev_mean[1:]
+        std = self.config.max_std * torch.ones_like(mean)
+
+        for _ in range(self.config.cem_iterations):
+            # Randomly sample action trajectories for the gaussian distribution.
+            std_normal_noise = torch.randn(
+                self.config.horizon,
+                self.config.n_gaussian_samples,
+                batch_size,
+                self.config.output_shapes["action"][0],
+                device=std.device,
+            )
+            gaussian_actions = torch.clamp(mean.unsqueeze(1) + std.unsqueeze(1) * std_normal_noise, -1, 1)
+
+            # Compute elite actions.
+            actions = torch.cat([gaussian_actions, pi_actions], dim=1)
+            value = self.estimate_value(z, actions).nan_to_num_(0).squeeze()
+            elite_idxs = torch.topk(value, self.config.n_elites, dim=0).indices  # (n_elites, batch)
+            elite_value = value.take_along_dim(elite_idxs, dim=0)  # (n_elites, batch)
+            # (horizon, n_elites, batch, action_dim)
+            elite_actions = actions.take_along_dim(einops.rearrange(elite_idxs, "n b -> 1 n b 1"), dim=1)
+
+            # Update gaussian PDF parameters to be the (weighted) mean and standard deviation of the elites.
+            max_value = elite_value.max(0, keepdim=True)[0]  # (1, batch)
+            # The weighting is a softmax over trajectory values. Note that this is not the same as the usage
+            # of Ω in eqn 4 of the TD-MPC paper. Instead it is the normalized version of it: s = Ω/ΣΩ. This
+            # makes the equations: μ = Σ(s⋅Γ), σ = Σ(s⋅(Γ-μ)²).
+            score = torch.exp(self.config.elite_weighting_temperature * (elite_value - max_value))
+            score /= score.sum(axis=0, keepdim=True)
+            # (horizon, batch, action_dim)
+            mean = torch.sum(einops.rearrange(score, "n b -> n b 1") * elite_actions, dim=1) / (
+                einops.rearrange(score.sum(0), "b -> 1 b 1") + 1e-9
+            )
+            std = torch.sqrt(
+                torch.sum(
+                    einops.rearrange(score, "n b -> n b 1")
+                    * (elite_actions - einops.rearrange(mean, "h b d -> h 1 b d")) ** 2,
+                    dim=1,
+                )
+                / (einops.rearrange(score.sum(0), "b -> 1 b 1") + 1e-9)
+            ).clamp_(self.config.min_std, self.config.max_std)
+
+        # Keep track of the mean for warm-starting subsequent steps.
+        self._prev_mean = mean
+
+        # Randomly select one of the elite actions from the last iteration of MPPI/CEM using the softmax
+        # scores from the last iteration.
+        actions = elite_actions[:, torch.multinomial(score.T, 1).squeeze(), torch.arange(batch_size)]
+        return actions
+
+    @torch.no_grad()
+    def estimate_value(self, z: Tensor, actions: Tensor):
+        """Estimates the value of a trajectory as per eqn 4 of the FOWM paper.
+
+        Args:
+            z: (batch, latent_dim) tensor of initial latent states.
+            actions: (horizon, batch, action_dim) tensor of action trajectories.
+        Returns:
+            (batch,) tensor of values.
+        """
+        # Initialize return and running discount factor.
+        G, running_discount = 0, 1
+        # Iterate over the actions in the trajectory to simulate the trajectory using the latent dynamics
+        # model. Keep track of return.
+        for t in range(actions.shape[0]):
+            # Estimate the next state (latent) and reward.
+            z, reward = self.model.latent_dynamics_and_reward(z, actions[t], discretize_reward=True)
+            # Update the return and running discount.
+            G += running_discount * reward
+            running_discount *= self.config.discount
+
+        # next_action = self.model.pi(z)[0]  # (batch, action_dim)
+        # terminal_values = self.model.Qs(z, next_action, return_type="avg")  # (ensemble, batch)
+
+        return G + running_discount * self.model.Qs(z, self.model.pi(z)[0], return_type="avg")
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor | float]:
+        """Run the batch through the model and compute the loss.
+
+        Returns a dictionary with loss as a tensor, and other information as native floats.
+        """
+        device = get_device_from_parameters(self)
+
+        batch = self.normalize_inputs(batch)
+        if self._use_image:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.image"] = batch[self.input_image_key]
+        batch = self.normalize_targets(batch)
+
+        info = {}
+
+        # (b, t) -> (t, b)
+        for key in batch:
+            if batch[key].ndim > 1:
+                batch[key] = batch[key].transpose(1, 0)
+
+        action = batch["action"]  # (t, b, action_dim)
+        reward = batch["next.reward"]  # (t, b)
+        observations = {k: v for k, v in batch.items() if k.startswith("observation.")}
+
+        # Apply random image augmentations.
+        if self._use_image and self.config.max_random_shift_ratio > 0:
+            observations["observation.image"] = flatten_forward_unflatten(
+                partial(random_shifts_aug, max_random_shift_ratio=self.config.max_random_shift_ratio),
+                observations["observation.image"],
+            )
+
+        # Get the current observation for predicting trajectories, and all future observations for use in
+        # the latent consistency loss and TD loss.
+        current_observation, next_observations = {}, {}
+        for k in observations:
+            current_observation[k] = observations[k][0]
+            next_observations[k] = observations[k][1:]
+        horizon, batch_size = next_observations[
+            "observation.image" if self._use_image else "observation.environment_state"
+        ].shape[:2]
+
+        # Run latent rollout using the latent dynamics model and policy model.
+        # Note this has shape `horizon+1` because there are `horizon` actions and a current `z`. Each action
+        # gives us a next `z`.
+        batch_size = batch["index"].shape[0]
+        z_preds = torch.empty(horizon + 1, batch_size, self.config.latent_dim, device=device)
+        z_preds[0] = self.model.encode(current_observation)
+        reward_preds = torch.empty(horizon, batch_size, self.config.num_bins, device=device)
+        for t in range(horizon):
+            z_preds[t + 1], reward_preds[t] = self.model.latent_dynamics_and_reward(z_preds[t], action[t])
+
+        # Compute Q value predictions based on the latent rollout.
+        q_preds_ensemble = self.model.Qs(
+            z_preds[:-1], action, return_type="all"
+        )  # (ensemble, horizon, batch)
+        info.update({"Q": q_preds_ensemble.mean().item()})
+
+        # Compute various targets with stopgrad.
+        with torch.no_grad():
+            # Latent state consistency targets for consistency loss.
+            z_targets = self.model.encode(next_observations)
+
+            # Compute the TD-target from a reward and the next observation
+            pi = self.model.pi(z_targets)[0]
+            td_targets = (
+                reward
+                + self.config.discount
+                * self.model.Qs(z_targets, pi, return_type="min", target=True).squeeze()
+            )
+
+        # Compute losses.
+        # Exponentially decay the loss weight with respect to the timestep. Steps that are more distant in the
+        # future have less impact on the loss. Note: unsqueeze will let us broadcast to (seq, batch).
+        temporal_loss_coeffs = torch.pow(
+            self.config.temporal_decay_coeff, torch.arange(horizon, device=device)
+        ).unsqueeze(-1)
+
+        # Compute consistency loss as MSE loss between latents predicted from the rollout and latents
+        # predicted from the (target model's) observation encoder.
+        consistency_loss = (
+            (
+                temporal_loss_coeffs
+                * F.mse_loss(z_preds[1:], z_targets, reduction="none").mean(dim=-1)
+                # `z_preds` depends on the current observation and the actions.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+                # `z_targets` depends on the next observation.
+                * ~batch["observation.state_is_pad"][1:]
+            )
+            .sum(0)
+            .mean()
+        )
+        # Compute the reward loss as MSE loss between rewards predicted from the rollout and the dataset
+        # rewards.
+        reward_loss = (
+            (
+                temporal_loss_coeffs
+                * soft_cross_entropy(reward_preds, reward, self.config).mean(1)
+                * ~batch["next.reward_is_pad"]
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+            )
+            .sum(0)
+            .mean()
+        )
+
+        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
+        ce_value_loss = 0.0
+        for i in range(self.config.q_ensemble_size):
+            ce_value_loss += soft_cross_entropy(q_preds_ensemble[i], td_targets, self.config).mean(1)
+
+        q_value_loss = (
+            (
+                temporal_loss_coeffs
+                * ce_value_loss
+                # `q_preds_ensemble` depends on the first observation and the actions.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+                # q_targets depends on the reward and the next observations.
+                * ~batch["next.reward_is_pad"]
+                * ~batch["observation.state_is_pad"][1:]
+            )
+            .sum(0)
+            .mean()
+        )
+
+        # Calculate the advantage weighted regression loss for π as detailed in FOWM 3.1.
+        # We won't need these gradients again so detach.
+        z_preds = z_preds.detach()
+        action_preds, _, log_pis, _ = self.model.pi(z_preds[:-1])
+
+        with torch.no_grad():
+            # avoid unnessecary computation of the gradients during policy optimization
+            # TODO (michel-aractingi): the same logic should be extended when adding task embeddings
+            qs = self.model.Qs(z_preds[:-1], action_preds, return_type="avg")
+            self.scale.update(qs[0])
+            qs = self.scale(qs)
+
+        pi_loss = (
+            (self.config.entropy_coef * log_pis - qs).mean(dim=2)
+            * temporal_loss_coeffs
+            # `action_preds` depends on the first observation and the actions.
+            * ~batch["observation.state_is_pad"][0]
+            * ~batch["action_is_pad"]
+        ).mean()
+
+        loss = (
+            self.config.consistency_coeff * consistency_loss
+            + self.config.reward_coeff * reward_loss
+            + self.config.value_coeff * q_value_loss
+            + pi_loss
+        )
+
+        info.update(
+            {
+                "consistency_loss": consistency_loss.item(),
+                "reward_loss": reward_loss.item(),
+                "Q_value_loss": q_value_loss.item(),
+                "pi_loss": pi_loss.item(),
+                "loss": loss,
+                "sum_loss": loss.item() * self.config.horizon,
+                "pi_scale": float(self.scale.value),
+            }
+        )
+
+        # Undo (b, t) -> (t, b).
+        for key in batch:
+            if batch[key].ndim > 1:
+                batch[key] = batch[key].transpose(1, 0)
+
+        return info
+
+    def update(self):
+        """Update the target model's using polyak averaging."""
+        self.model.update_target_Q()
+
+
+class TDMPC2WorldModel(nn.Module):
+    """Latent dynamics model used in TD-MPC2."""
+
+    def __init__(self, config: TDMPC2Config):
+        super().__init__()
+        self.config = config
+
+        self._encoder = TDMPC2ObservationEncoder(config)
+
+        # Define latent dynamics head
+        self._dynamics = nn.Sequential(
+            NormedLinear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
+            NormedLinear(config.mlp_dim, config.mlp_dim),
+            NormedLinear(config.mlp_dim, config.latent_dim, act=SimNorm(config.simnorm_dim)),
+        )
+
+        # Define reward head
+        self._reward = nn.Sequential(
+            NormedLinear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
+            NormedLinear(config.mlp_dim, config.mlp_dim),
+            nn.Linear(config.mlp_dim, max(config.num_bins, 1)),
+        )
+
+        # Define policy head
+        self._pi = nn.Sequential(
+            NormedLinear(config.latent_dim, config.mlp_dim),
+            NormedLinear(config.mlp_dim, config.mlp_dim),
+            nn.Linear(config.mlp_dim, 2 * config.output_shapes["action"][0]),
+        )
+
+        # Define ensemble of Q functions
+        self._Qs = nn.ModuleList(
+            [
+                nn.Sequential(
+                    NormedLinear(
+                        config.latent_dim + config.output_shapes["action"][0],
+                        config.mlp_dim,
+                        dropout=config.dropout,
+                    ),
+                    NormedLinear(config.mlp_dim, config.mlp_dim),
+                    nn.Linear(config.mlp_dim, max(config.num_bins, 1)),
+                )
+                for _ in range(config.q_ensemble_size)
+            ]
+        )
+
+        self._init_weights()
+
+        self._target_Qs = deepcopy(self._Qs).requires_grad_(False)
+
+        self.log_std_min = torch.tensor(config.log_std_min)
+        self.log_std_dif = torch.tensor(config.log_std_max) - self.log_std_min
+
+        self.bins = torch.linspace(config.vmin, config.vmax, config.num_bins)
+        self.config.bin_size = (config.vmax - config.vmin) / (config.num_bins - 1)
+
+    def _init_weights(self):
+        """Initialize model weights.
+        Custom weight initializations proposed in TD-MPC2.
+
+        """
+
+        def _apply_fn(m):
+            if isinstance(m, nn.Linear):
+                nn.init.trunc_normal_(m.weight, std=0.02)
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.ParameterList):
+                for i, p in enumerate(m):
+                    if p.dim() == 3:  # Linear
+                        nn.init.trunc_normal_(p, std=0.02)  # Weight
+                        nn.init.constant_(m[i + 1], 0)  # Bias
+
+        self.apply(_apply_fn)
+
+        # initialize parameters of the
+        for m in [self._reward, *self._Qs]:
+            assert isinstance(
+                m[-1], nn.Linear
+            ), "Sanity check. The last linear layer needs 0 initialization on weights."
+            nn.init.zeros_(m[-1].weight)
+
+    def to(self, *args, **kwargs):
+        """
+        Overriding `to` method to also move additional tensors to device.
+        """
+        super().to(*args, **kwargs)
+        self.log_std_min = self.log_std_min.to(*args, **kwargs)
+        self.log_std_dif = self.log_std_dif.to(*args, **kwargs)
+        self.bins = self.bins.to(*args, **kwargs)
+        return self
+
+    def train(self, mode):
+        super().train(mode)
+        self._target_Qs.train(False)
+        return self
+
+    def encode(self, obs: dict[str, Tensor]) -> Tensor:
+        """Encodes an observation into its latent representation."""
+        return self._encoder(obs)
+
+    def latent_dynamics_and_reward(
+        self, z: Tensor, a: Tensor, discretize_reward: bool = False
+    ) -> tuple[Tensor, Tensor, bool]:
+        """Predict the next state's latent representation and the reward given a current latent and action.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+        Returns:
+            A tuple containing:
+                - (*, latent_dim) tensor for the next state's latent representation.
+                - (*,) tensor for the estimated reward.
+        """
+        x = torch.cat([z, a], dim=-1)
+        reward = self._reward(x).squeeze(-1)
+        if discretize_reward:
+            reward = two_hot_inv(reward, self.bins)
+        return self._dynamics(x), reward
+
+    def latent_dynamics(self, z: Tensor, a: Tensor) -> Tensor:
+        """Predict the next state's latent representation given a current latent and action.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+        Returns:
+            (*, latent_dim) tensor for the next state's latent representation.
+        """
+        x = torch.cat([z, a], dim=-1)
+        return self._dynamics(x)
+
+    def pi(self, z: Tensor) -> Tensor:
+        """Samples an action from the learned policy.
+
+        The policy can also have added (truncated) Gaussian noise injected for encouraging exploration when
+        generating rollouts for online training.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            std: The standard deviation of the injected noise.
+        Returns:
+            (*, action_dim) tensor for the sampled action.
+        """
+        mu, log_std = self._pi(z).chunk(2, dim=-1)
+        log_std = self.log_std_min + 0.5 * self.log_std_dif * (torch.tanh(log_std) + 1)
+        eps = torch.randn_like(mu)
+
+        log_pi = gaussian_logprob(eps, log_std)
+        pi = mu + eps * log_std.exp()
+        mu, pi, log_pi = squash(mu, pi, log_pi)
+
+        return pi, mu, log_pi, log_std
+
+    def Qs(self, z: Tensor, a: Tensor, return_type: str = "min", target=False) -> Tensor:  # noqa: N802
+        """Predict state-action value for all of the learned Q functions.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+            return_type: either 'min' or 'all' otherwise the average is returned
+        Returns:
+            (q_ensemble, *) tensor for the value predictions of each learned Q function in the ensemble or the average or min
+        """
+        x = torch.cat([z, a], dim=-1)
+
+        if target:
+            out = torch.stack([q(x).squeeze(-1) for q in self._target_Qs], dim=0)
+        else:
+            out = torch.stack([q(x).squeeze(-1) for q in self._Qs], dim=0)
+
+        if return_type == "all":
+            return out
+
+        Q1, Q2 = out[np.random.choice(len(self._Qs), size=2, replace=False)]
+        Q1, Q2 = two_hot_inv(Q1, self.bins), two_hot_inv(Q2, self.bins)
+        return torch.min(Q1, Q2) if return_type == "min" else (Q1 + Q2) / 2
+
+    def update_target_Q(self):
+        """
+        Soft-update target Q-networks using Polyak averaging.
+        """
+        with torch.no_grad():
+            for p, p_target in zip(self._Qs.parameters(), self._target_Qs.parameters(), strict=False):
+                p_target.data.lerp_(p.data, self.config.target_model_momentum)
+
+
+class TDMPC2ObservationEncoder(nn.Module):
+    """Encode image and/or state vector observations."""
+
+    def __init__(self, config: TDMPC2Config):
+        """
+        Creates encoders for pixel and/or state modalities.
+        TODO(alexander-soare): The original work allows for multiple images by concatenating them along the
+            channel dimension. Re-implement this capability.
+        """
+        super().__init__()
+        self.config = config
+
+        # Define the observation encoder whether its pixels or states
+        encoder_dict = {}
+        for obs_key in config.input_shapes:
+            if "observation.image" in config.input_shapes:
+                encoder_module = nn.Sequential(
+                    nn.Conv2d(config.input_shapes[obs_key][0], config.image_encoder_hidden_dim, 7, stride=2),
+                    nn.ReLU(inplace=True),
+                    nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 5, stride=2),
+                    nn.ReLU(inplace=True),
+                    nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
+                    nn.ReLU(inplace=True),
+                    nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=1),
+                )
+                dummy_batch = torch.zeros(1, *config.input_shapes[obs_key])
+                with torch.inference_mode():
+                    out_shape = encoder_module(dummy_batch).shape[1:]
+                encoder_module.extend(
+                    nn.Sequential(
+                        nn.Flatten(),
+                        NormedLinear(np.prod(out_shape), config.latent_dim, act=SimNorm(config.simnorm_dim)),
+                    )
+                )
+
+            elif (
+                "observation.state" in config.input_shapes
+                or "observation.environment_state" in config.input_shapes
+            ):
+                encoder_module = nn.ModuleList()
+                encoder_module.append(
+                    NormedLinear(config.input_shapes[obs_key][0], config.state_encoder_hidden_dim)
+                )
+                assert config.num_enc_layers > 0
+                for _ in range(config.num_enc_layers - 1):
+                    encoder_module.append(
+                        NormedLinear(config.state_encoder_hidden_dim, config.state_encoder_hidden_dim)
+                    )
+                encoder_module.append(
+                    NormedLinear(
+                        config.state_encoder_hidden_dim, config.latent_dim, act=SimNorm(config.simnorm_dim)
+                    )
+                )
+                encoder_module = nn.Sequential(*encoder_module)
+
+            else:
+                raise NotImplementedError(f"No corresponding encoder module for key {obs_key}.")
+
+            encoder_dict[obs_key.replace(".", "")] = encoder_module
+
+        self.encoder = nn.ModuleDict(encoder_dict)
+
+    def forward(self, obs_dict: dict[str, Tensor]) -> Tensor:
+        """Encode the image and/or state vector.
+
+        Each modality is encoded into a feature vector of size (latent_dim,) and then a uniform mean is taken
+        over all features.
+        """
+        feat = []
+        for obs_key in self.config.input_shapes:
+            if "observation.image" in obs_key:
+                feat.append(
+                    flatten_forward_unflatten(self.encoder[obs_key.replace(".", "")], obs_dict[obs_key])
+                )
+            else:
+                feat.append(self.encoder[obs_key.replace(".", "")](obs_dict[obs_key]))
+        return torch.stack(feat, dim=0).mean(0)
+
+
+def random_shifts_aug(x: Tensor, max_random_shift_ratio: float) -> Tensor:
+    """Randomly shifts images horizontally and vertically.
+
+    Adapted from https://github.com/facebookresearch/drqv2
+    """
+    b, _, h, w = x.size()
+    assert h == w, "non-square images not handled yet"
+    pad = int(round(max_random_shift_ratio * h))
+    x = F.pad(x, tuple([pad] * 4), "replicate")
+    eps = 1.0 / (h + 2 * pad)
+    arange = torch.linspace(
+        -1.0 + eps,
+        1.0 - eps,
+        h + 2 * pad,
+        device=x.device,
+        dtype=torch.float32,
+    )[:h]
+    arange = einops.repeat(arange, "w -> h w 1", h=h)
+    base_grid = torch.cat([arange, arange.transpose(1, 0)], dim=2)
+    base_grid = einops.repeat(base_grid, "h w c -> b h w c", b=b)
+    # A random shift in units of pixels and within the boundaries of the padding.
+    shift = torch.randint(
+        0,
+        2 * pad + 1,
+        size=(b, 1, 1, 2),
+        device=x.device,
+        dtype=torch.float32,
+    )
+    shift *= 2.0 / (h + 2 * pad)
+    grid = base_grid + shift
+    return F.grid_sample(x, grid, padding_mode="zeros", align_corners=False)
+
+
+def flatten_forward_unflatten(fn: Callable[[Tensor], Tensor], image_tensor: Tensor) -> Tensor:
+    """Helper to temporarily flatten extra dims at the start of the image tensor.
+
+    Args:
+        fn: Callable that the image tensor will be passed to. It should accept (B, C, H, W) and return
+            (B, *), where * is any number of dimensions.
+        image_tensor: An image tensor of shape (**, C, H, W), where ** is any number of dimensions, generally
+            different from *.
+    Returns:
+        A return value from the callable reshaped to (**, *).
+    """
+    if image_tensor.ndim == 4:
+        return fn(image_tensor)
+    start_dims = image_tensor.shape[:-3]
+    inp = torch.flatten(image_tensor, end_dim=-4)
+    flat_out = fn(inp)
+    return torch.reshape(flat_out, (*start_dims, *flat_out.shape[1:]))
+
+
+class RunningScale:
+    """Running trimmed scale estimator."""
+
+    def __init__(self, tau):
+        self.tau = tau
+        self._value = torch.ones(1, dtype=torch.float32, device=torch.device("cuda"))
+        self._percentiles = torch.tensor([5, 95], dtype=torch.float32, device=torch.device("cuda"))
+
+    def state_dict(self):
+        return dict(value=self._value, percentiles=self._percentiles)
+
+    def load_state_dict(self, state_dict):
+        self._value.data.copy_(state_dict["value"])
+        self._percentiles.data.copy_(state_dict["percentiles"])
+
+    @property
+    def value(self):
+        return self._value.cpu().item()
+
+    def _percentile(self, x):
+        x_dtype, x_shape = x.dtype, x.shape
+        x = x.view(x.shape[0], -1)
+        in_sorted, _ = torch.sort(x, dim=0)
+        positions = self._percentiles * (x.shape[0] - 1) / 100
+        floored = torch.floor(positions)
+        ceiled = floored + 1
+        ceiled[ceiled > x.shape[0] - 1] = x.shape[0] - 1
+        weight_ceiled = positions - floored
+        weight_floored = 1.0 - weight_ceiled
+        d0 = in_sorted[floored.long(), :] * weight_floored[:, None]
+        d1 = in_sorted[ceiled.long(), :] * weight_ceiled[:, None]
+        return (d0 + d1).view(-1, *x_shape[1:]).type(x_dtype)
+
+    def update(self, x):
+        percentiles = self._percentile(x.detach())
+        value = torch.clamp(percentiles[1] - percentiles[0], min=1.0)
+        self._value.data.lerp_(value, self.tau)
+
+    def __call__(self, x, update=False):
+        if update:
+            self.update(x)
+        return x * (1 / self.value)
+
+    def __repr__(self):
+        return f"RunningScale(S: {self.value})"

lerobot/common/policies/tdmpc2/tdmpc2_utils.py

@@ -0,0 +1,164 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functorch import combine_state_for_ensemble
+
+
+class Ensemble(nn.Module):
+    """
+    Vectorized ensemble of modules.
+    """
+
+    def __init__(self, modules, **kwargs):
+        super().__init__()
+        modules = nn.ModuleList(modules)
+        fn, params, _ = combine_state_for_ensemble(modules)
+        self.vmap = torch.vmap(fn, in_dims=(0, 0, None), randomness="different", **kwargs)
+        self.params = nn.ParameterList([nn.Parameter(p) for p in params])
+        self._repr = str(modules)
+
+    def forward(self, *args, **kwargs):
+        return self.vmap([p for p in self.params], (), *args, **kwargs)
+
+    def __repr__(self):
+        return "Vectorized " + self._repr
+
+
+class SimNorm(nn.Module):
+    """
+    Simplicial normalization.
+    Adapted from https://arxiv.org/abs/2204.00616.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x):
+        shp = x.shape
+        x = x.view(*shp[:-1], -1, self.dim)
+        x = F.softmax(x, dim=-1)
+        return x.view(*shp)
+
+    def __repr__(self):
+        return f"SimNorm(dim={self.dim})"
+
+
+class NormedLinear(nn.Linear):
+    """
+    Linear layer with LayerNorm, activation, and optionally dropout.
+    """
+
+    def __init__(self, *args, dropout=0.0, act=nn.Mish(inplace=True), **kwargs):
+        super().__init__(*args, **kwargs)
+        self.ln = nn.LayerNorm(self.out_features)
+        self.act = act
+        self.dropout = nn.Dropout(dropout, inplace=True) if dropout else None
+
+    def forward(self, x):
+        x = super().forward(x)
+        if self.dropout:
+            x = self.dropout(x)
+        return self.act(self.ln(x))
+
+    def __repr__(self):
+        repr_dropout = f", dropout={self.dropout.p}" if self.dropout else ""
+        return (
+            f"NormedLinear(in_features={self.in_features}, "
+            f"out_features={self.out_features}, "
+            f"bias={self.bias is not None}{repr_dropout}, "
+            f"act={self.act.__class__.__name__})"
+        )
+
+
+def soft_cross_entropy(pred, target, cfg):
+    """Computes the cross entropy loss between predictions and soft targets."""
+    pred = F.log_softmax(pred, dim=-1)
+    target = two_hot(target, cfg)
+    return -(target * pred).sum(-1, keepdim=True)
+
+
+@torch.jit.script
+def log_std(x, low, dif):
+    return low + 0.5 * dif * (torch.tanh(x) + 1)
+
+
+@torch.jit.script
+def _gaussian_residual(eps, log_std):
+    return -0.5 * eps.pow(2) - log_std
+
+
+@torch.jit.script
+def _gaussian_logprob(residual):
+    return residual - 0.5 * torch.log(2 * torch.pi)
+
+
+def gaussian_logprob(eps, log_std, size=None):
+    """Compute Gaussian log probability."""
+    residual = _gaussian_residual(eps, log_std).sum(-1, keepdim=True)
+    if size is None:
+        size = eps.size(-1)
+    return _gaussian_logprob(residual) * size
+
+
+@torch.jit.script
+def _squash(pi):
+    return torch.log(F.relu(1 - pi.pow(2)) + 1e-6)
+
+
+def squash(mu, pi, log_pi):
+    """Apply squashing function."""
+    mu = torch.tanh(mu)
+    pi = torch.tanh(pi)
+    log_pi -= _squash(pi).sum(-1, keepdim=True)
+    return mu, pi, log_pi
+
+
+@torch.jit.script
+def symlog(x):
+    """
+    Symmetric logarithmic function.
+    Adapted from https://github.com/danijar/dreamerv3.
+    """
+    return torch.sign(x) * torch.log(1 + torch.abs(x))
+
+
+@torch.jit.script
+def symexp(x):
+    """
+    Symmetric exponential function.
+    Adapted from https://github.com/danijar/dreamerv3.
+    """
+    return torch.sign(x) * (torch.exp(torch.abs(x)) - 1)
+
+
+def two_hot(x, cfg):
+    """Converts a batch of scalars to soft two-hot encoded targets for discrete regression."""
+
+    # x shape [horizon, num_features]
+    if cfg.num_bins == 0:
+        return x
+    elif cfg.num_bins == 1:
+        return symlog(x)
+    x = torch.clamp(symlog(x), cfg.vmin, cfg.vmax)
+    bin_idx = torch.floor((x - cfg.vmin) / cfg.bin_size).long()  # shape [num_features]
+    bin_offset = ((x - cfg.vmin) / cfg.bin_size - bin_idx.float()).unsqueeze(-1)  # shape [num_features , 1]
+    soft_two_hot = torch.zeros(
+        *x.shape, cfg.num_bins, device=x.device
+    )  # shape [horizon, num_features, num_bins]
+    soft_two_hot.scatter_(2, bin_idx.unsqueeze(-1), 1 - bin_offset)
+    soft_two_hot.scatter_(2, (bin_idx.unsqueeze(-1) + 1) % cfg.num_bins, bin_offset)
+    return soft_two_hot
+
+
+def two_hot_inv(x, bins):
+    """Converts a batch of soft two-hot encoded vectors to scalars."""
+    num_bins = bins.shape[0]
+    if num_bins == 0:
+        return x
+    elif num_bins == 1:
+        return symexp(x)
+
+    x = F.softmax(x, dim=-1)
+    x = torch.sum(x * bins, dim=-1, keepdim=True)
+    return symexp(x)

lerobot/common/policies/utils.py

@@ -13,7 +13,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 import torch
 from torch import nn
 
@@ -48,20 +47,3 @@ def get_dtype_from_parameters(module: nn.Module) -> torch.dtype:
     Note: assumes that all parameters have the same dtype.
     """
     return next(iter(module.parameters())).dtype
-
-
-def get_output_shape(module: nn.Module, input_shape: tuple) -> tuple:
-    """
-    Calculates the output shape of a PyTorch module given an input shape.
-
-    Args:
-        module (nn.Module): a PyTorch module
-        input_shape (tuple): A tuple representing the input shape, e.g., (batch_size, channels, height, width)
-
-    Returns:
-        tuple: The output shape of the module.
-    """
-    dummy_input = torch.zeros(size=input_shape)
-    with torch.inference_mode():
-        output = module(dummy_input)
-    return tuple(output.shape)

lerobot/common/policies/vqbet/configuration_vqbet.py

@@ -18,15 +18,9 @@
 
 from dataclasses import dataclass, field
 
-from lerobot.common.optim.optimizers import AdamConfig
-from lerobot.common.optim.schedulers import VQBeTSchedulerConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
 
-
-@PreTrainedConfig.register_subclass("vqbet")
 @dataclass
-class VQBeTConfig(PreTrainedConfig):
+class VQBeTConfig:
     """Configuration class for VQ-BeT.
 
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
@@ -96,13 +90,26 @@ class VQBeTConfig(PreTrainedConfig):
     n_action_pred_token: int = 3
     action_chunk_size: int = 5
 
-    normalization_mapping: dict[str, NormalizationMode] = field(
+    input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
-            "VISUAL": NormalizationMode.IDENTITY,
-            "STATE": NormalizationMode.MIN_MAX,
-            "ACTION": NormalizationMode.MIN_MAX,
+            "observation.image": [3, 96, 96],
+            "observation.state": [2],
         }
     )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [2],
+        }
+    )
+
+    # Normalization / Unnormalization
+    input_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {
+            "observation.image": "mean_std",
+            "observation.state": "min_max",
+        }
+    )
+    output_normalization_modes: dict[str, str] = field(default_factory=lambda: {"action": "min_max"})
 
     # Architecture / modeling.
     # Vision backbone.
@@ -132,69 +139,29 @@ class VQBeTConfig(PreTrainedConfig):
     bet_softmax_temperature: float = 0.1
     sequentially_select: 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
-    optimizer_vqvae_lr: float = 1e-3
-    optimizer_vqvae_weight_decay: float = 1e-4
-    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}."
             )
-
-    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) -> VQBeTSchedulerConfig:
-        return VQBeTSchedulerConfig(
-            num_warmup_steps=self.scheduler_warmup_steps,
-            num_vqvae_training_steps=self.n_vqvae_training_steps,
-        )
-
-    def validate_features(self) -> None:
-        # Note: this check was previously performed inside VQBeTRgbEncoder in the form of
-        # assert len(image_keys) == 1
-        if not len(self.image_features) == 1:
-            raise ValueError("You must provide only one image among the inputs.")
-
+        image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
         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]:
+            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 the images shapes. Got {self.crop_shape} "
-                        f"for `crop_shape` and {image_ft.shape} for "
-                        f"`{key}`."
+                        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, 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:
+        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"`{key}` does not match `{first_image_key}`, but we expect all image shapes to match."
+                    f"`input_shapes[{image_key}]` does not match `input_shapes[{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, self.n_action_pred_token + self.action_chunk_size - 1))
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None

lerobot/common/policies/vqbet/modeling_vqbet.py

@@ -16,6 +16,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import math
 import warnings
 from collections import deque
 from typing import Callable, List
@@ -25,23 +26,29 @@ import numpy as np
 import torch
 import torch.nn.functional as F  # noqa: N812
 import torchvision
+from huggingface_hub import PyTorchModelHubMixin
 from torch import Tensor, nn
+from torch.optim.lr_scheduler import LambdaLR
 
 from lerobot.common.policies.normalize import Normalize, Unnormalize
-from lerobot.common.policies.pretrained import PreTrainedPolicy
-from lerobot.common.policies.utils import get_device_from_parameters, get_output_shape, populate_queues
+from lerobot.common.policies.utils import get_device_from_parameters, populate_queues
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
 from lerobot.common.policies.vqbet.vqbet_utils import GPT, ResidualVQ
 
 # ruff: noqa: N806
 
 
-class VQBeTPolicy(PreTrainedPolicy):
+class VQBeTPolicy(
+    nn.Module,
+    PyTorchModelHubMixin,
+    library_name="lerobot",
+    repo_url="https://github.com/huggingface/lerobot",
+    tags=["robotics", "vqbet"],
+):
     """
     VQ-BeT Policy as per "Behavior Generation with Latent Actions"
     """
 
-    config_class = VQBeTConfig
     name = "vqbet"
 
     def __init__(
@@ -56,62 +63,26 @@ class VQBeTPolicy(PreTrainedPolicy):
             dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
-        super().__init__(config)
-        config.validate_features()
+        super().__init__()
+        if config is None:
+            config = VQBeTConfig()
         self.config = config
-
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_inputs = Normalize(
+            config.input_shapes, config.input_normalization_modes, dataset_stats
+        )
         self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
+            config.output_shapes, config.output_normalization_modes, dataset_stats
         )
         self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
+            config.output_shapes, config.output_normalization_modes, dataset_stats
         )
 
         self.vqbet = VQBeTModel(config)
 
+        self.expected_image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+
         self.reset()
 
-    def get_optim_params(self) -> dict:
-        vqvae_params = (
-            list(self.vqbet.action_head.vqvae_model.encoder.parameters())
-            + list(self.vqbet.action_head.vqvae_model.decoder.parameters())
-            + list(self.vqbet.action_head.vqvae_model.vq_layer.parameters())
-        )
-        decay_params, no_decay_params = self.vqbet.policy.configure_parameters()
-        decay_params = (
-            decay_params
-            + list(self.vqbet.rgb_encoder.parameters())
-            + list(self.vqbet.state_projector.parameters())
-            + list(self.vqbet.rgb_feature_projector.parameters())
-            + [self.vqbet.action_token]
-            + list(self.vqbet.action_head.map_to_cbet_preds_offset.parameters())
-        )
-
-        if self.config.sequentially_select:
-            decay_params = (
-                decay_params
-                + list(self.vqbet.action_head.map_to_cbet_preds_primary_bin.parameters())
-                + list(self.vqbet.action_head.map_to_cbet_preds_secondary_bin.parameters())
-            )
-        else:
-            decay_params = decay_params + list(self.vqbet.action_head.map_to_cbet_preds_bin.parameters())
-
-        return [
-            {
-                "params": decay_params,
-            },
-            {
-                "params": vqvae_params,
-                "weight_decay": self.config.optimizer_vqvae_weight_decay,
-                "lr": self.config.optimizer_vqvae_lr,
-            },
-            {
-                "params": no_decay_params,
-                "weight_decay": 0.0,
-            },
-        ]
-
     def reset(self):
         """
         Clear observation and action queues. Should be called on `env.reset()`
@@ -134,7 +105,7 @@ class VQBeTPolicy(PreTrainedPolicy):
 
         batch = self.normalize_inputs(batch)
         batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-        batch["observation.images"] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
+        batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         # Note: It's important that this happens after stacking the images into a single key.
         self._queues = populate_queues(self._queues, batch)
 
@@ -156,11 +127,11 @@ class VQBeTPolicy(PreTrainedPolicy):
         action = self._queues["action"].popleft()
         return action
 
-    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         """Run the batch through the model and compute the loss for training or validation."""
         batch = self.normalize_inputs(batch)
         batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-        batch["observation.images"] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
+        batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
         batch = self.normalize_targets(batch)
         # VQ-BeT discretizes action using VQ-VAE before training BeT (please refer to section 3.2 in the VQ-BeT paper https://arxiv.org/pdf/2403.03181)
         if not self.vqbet.action_head.vqvae_model.discretized.item():
@@ -170,16 +141,16 @@ class VQBeTPolicy(PreTrainedPolicy):
             loss, n_different_codes, n_different_combinations, recon_l1_error = (
                 self.vqbet.action_head.discretize(self.config.n_vqvae_training_steps, batch["action"])
             )
-            return loss, {
+            return {
+                "loss": loss,
                 "n_different_codes": n_different_codes,
                 "n_different_combinations": n_different_combinations,
                 "recon_l1_error": recon_l1_error,
             }
         # if Residual VQ is already trained, VQ-BeT trains its GPT and bin prediction head / offset prediction head parts.
         _, loss_dict = self.vqbet(batch, rollout=False)
-        loss = loss_dict.pop("loss")
 
-        return loss, loss_dict
+        return loss_dict
 
 
 class SpatialSoftmax(nn.Module):
@@ -317,14 +288,14 @@ class VQBeTModel(nn.Module):
         self.config = config
 
         self.rgb_encoder = VQBeTRgbEncoder(config)
-        self.num_images = len(self.config.image_features)
+        self.num_images = len([k for k in config.input_shapes if k.startswith("observation.image")])
         # This action query token is used as a prompt for querying action chunks. Please refer to "A_Q" in the image above.
         # Note: During the forward pass, this token is repeated as many times as needed. The authors also experimented with initializing the necessary number of tokens independently and observed inferior results.
         self.action_token = nn.Parameter(torch.randn(1, 1, self.config.gpt_input_dim))
 
         # To input state and observation features into GPT layers, we first project the features to fit the shape of input size of GPT.
         self.state_projector = MLP(
-            config.robot_state_feature.shape[0], hidden_channels=[self.config.gpt_input_dim]
+            config.input_shapes["observation.state"][0], hidden_channels=[self.config.gpt_input_dim]
         )
         self.rgb_feature_projector = MLP(
             self.rgb_encoder.feature_dim, hidden_channels=[self.config.gpt_input_dim]
@@ -342,7 +313,7 @@ class VQBeTModel(nn.Module):
             torch.row_stack([torch.arange(i, i + self.config.action_chunk_size) for i in range(num_tokens)]),
         )
 
-    def forward(self, batch: dict[str, Tensor], rollout: bool) -> tuple[dict, dict]:
+    def forward(self, batch: dict[str, Tensor], rollout: bool) -> Tensor:
         # Input validation.
         assert set(batch).issuperset({"observation.state", "observation.images"})
         batch_size, n_obs_steps = batch["observation.state"].shape[:2]
@@ -379,10 +350,10 @@ class VQBeTModel(nn.Module):
 
         # get action features (pass through GPT)
         features = self.policy(input_tokens)
-        # len(self.config.input_features) is the number of different observation modes.
+        # len(self.config.input_shapes) is the number of different observation modes.
         # this line gets the index of action prompt tokens.
-        historical_act_pred_index = np.arange(0, n_obs_steps) * (len(self.config.input_features) + 1) + len(
-            self.config.input_features
+        historical_act_pred_index = np.arange(0, n_obs_steps) * (len(self.config.input_shapes) + 1) + len(
+            self.config.input_shapes
         )
 
         # only extract the output tokens at the position of action query:
@@ -421,7 +392,7 @@ class VQBeTHead(nn.Module):
 
         self.map_to_cbet_preds_offset: output the predicted offsets for all the codes in all the layers.
             The input dimension of ` self.map_to_cbet_preds_offset` is same with the output of GPT,
-            and the output dimension of ` self.map_to_cbet_preds_offset` is `self.vqvae_model.vqvae_num_layers (=fixed as 2) * self.config.vqvae_n_embed * config.action_chunk_size * config.action_feature.shape[0]`.
+            and the output dimension of ` self.map_to_cbet_preds_offset` is `self.vqvae_model.vqvae_num_layers (=fixed as 2) * self.config.vqvae_n_embed * config.action_chunk_size * config.output_shapes["action"][0]`.
         """
 
         super().__init__()
@@ -448,7 +419,7 @@ class VQBeTHead(nn.Module):
                 self.vqvae_model.vqvae_num_layers
                 * self.config.vqvae_n_embed
                 * config.action_chunk_size
-                * config.action_feature.shape[0],
+                * config.output_shapes["action"][0],
             ],
         )
         # loss
@@ -482,7 +453,7 @@ class VQBeTHead(nn.Module):
                 param.requires_grad = False
         return loss, n_different_codes, n_different_combinations, recon_l1_error
 
-    def forward(self, x, **kwargs) -> dict:
+    def forward(self, x, **kwargs):
         # N is the batch size, and T is number of action query tokens, which are process through same GPT
         N, T, _ = x.shape
         # we calculate N and T side parallely. Thus, the dimensions would be
@@ -652,6 +623,84 @@ class VQBeTHead(nn.Module):
         return loss_dict
 
 
+class VQBeTOptimizer(torch.optim.Adam):
+    def __init__(self, policy, cfg):
+        vqvae_params = (
+            list(policy.vqbet.action_head.vqvae_model.encoder.parameters())
+            + list(policy.vqbet.action_head.vqvae_model.decoder.parameters())
+            + list(policy.vqbet.action_head.vqvae_model.vq_layer.parameters())
+        )
+        decay_params, no_decay_params = policy.vqbet.policy.configure_parameters()
+        decay_params = (
+            decay_params
+            + list(policy.vqbet.rgb_encoder.parameters())
+            + list(policy.vqbet.state_projector.parameters())
+            + list(policy.vqbet.rgb_feature_projector.parameters())
+            + [policy.vqbet.action_token]
+            + list(policy.vqbet.action_head.map_to_cbet_preds_offset.parameters())
+        )
+
+        if cfg.policy.sequentially_select:
+            decay_params = (
+                decay_params
+                + list(policy.vqbet.action_head.map_to_cbet_preds_primary_bin.parameters())
+                + list(policy.vqbet.action_head.map_to_cbet_preds_secondary_bin.parameters())
+            )
+        else:
+            decay_params = decay_params + list(policy.vqbet.action_head.map_to_cbet_preds_bin.parameters())
+
+        optim_groups = [
+            {
+                "params": decay_params,
+                "weight_decay": cfg.training.adam_weight_decay,
+                "lr": cfg.training.lr,
+            },
+            {
+                "params": vqvae_params,
+                "weight_decay": 0.0001,
+                "lr": cfg.training.vqvae_lr,
+            },
+            {
+                "params": no_decay_params,
+                "weight_decay": 0.0,
+                "lr": cfg.training.lr,
+            },
+        ]
+        super().__init__(
+            optim_groups,
+            cfg.training.lr,
+            cfg.training.adam_betas,
+            cfg.training.adam_eps,
+        )
+
+
+class VQBeTScheduler(nn.Module):
+    def __init__(self, optimizer, cfg):
+        super().__init__()
+        n_vqvae_training_steps = cfg.training.n_vqvae_training_steps
+
+        num_warmup_steps = cfg.training.lr_warmup_steps
+        num_training_steps = cfg.training.offline_steps
+        num_cycles = 0.5
+
+        def lr_lambda(current_step):
+            if current_step < n_vqvae_training_steps:
+                return float(1)
+            else:
+                current_step = current_step - n_vqvae_training_steps
+                if current_step < num_warmup_steps:
+                    return float(current_step) / float(max(1, num_warmup_steps))
+                progress = float(current_step - num_warmup_steps) / float(
+                    max(1, num_training_steps - num_warmup_steps)
+                )
+                return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress)))
+
+        self.lr_scheduler = LambdaLR(optimizer, lr_lambda, -1)
+
+    def step(self):
+        self.lr_scheduler.step()
+
+
 class VQBeTRgbEncoder(nn.Module):
     """Encode an RGB image into a 1D feature vector.
 
@@ -694,15 +743,19 @@ class VQBeTRgbEncoder(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.image_features` and it should
+        # The dummy input should take the number of image channels from `config.input_shapes` and it should
         # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
-        # height and width from `config.image_features`.
-
-        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:]
-
+        # height and width from `config.input_shapes`.
+        image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
+        assert len(image_keys) == 1
+        image_key = image_keys[0]
+        dummy_input_h_w = (
+            config.crop_shape if config.crop_shape is not None else config.input_shapes[image_key][1:]
+        )
+        dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *dummy_input_h_w))
+        with torch.inference_mode():
+            dummy_feature_map = self.backbone(dummy_input)
+        feature_map_shape = tuple(dummy_feature_map.shape[1:])
         self.pool = SpatialSoftmax(feature_map_shape, num_kp=config.spatial_softmax_num_keypoints)
         self.feature_dim = config.spatial_softmax_num_keypoints * 2
         self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
@@ -791,7 +844,7 @@ class VqVae(nn.Module):
         )
 
         self.encoder = MLP(
-            in_channels=self.config.action_feature.shape[0] * self.config.action_chunk_size,
+            in_channels=self.config.output_shapes["action"][0] * self.config.action_chunk_size,
             hidden_channels=[
                 config.vqvae_enc_hidden_dim,
                 config.vqvae_enc_hidden_dim,
@@ -803,7 +856,7 @@ class VqVae(nn.Module):
             hidden_channels=[
                 config.vqvae_enc_hidden_dim,
                 config.vqvae_enc_hidden_dim,
-                self.config.action_feature.shape[0] * self.config.action_chunk_size,
+                self.config.output_shapes["action"][0] * self.config.action_chunk_size,
             ],
         )
 
@@ -819,9 +872,9 @@ class VqVae(nn.Module):
         # given latent vector, this function outputs the decoded action.
         output = self.decoder(latent)
         if self.config.action_chunk_size == 1:
-            return einops.rearrange(output, "N (T A) -> N T A", A=self.config.action_feature.shape[0])
+            return einops.rearrange(output, "N (T A) -> N T A", A=self.config.output_shapes["action"][0])
         else:
-            return einops.rearrange(output, "N (T A) -> N T A", A=self.config.action_feature.shape[0])
+            return einops.rearrange(output, "N (T A) -> N T A", A=self.config.output_shapes["action"][0])
 
     def get_code(self, state):
         # in phase 2 of VQ-BeT training, we need a `ground truth labels of action data` to calculate the Focal loss for code prediction head. (please refer to section 3.3 in the paper https://arxiv.org/pdf/2403.03181)

lerobot/common/policies/vqbet/vqbet_utils.py

@@ -203,9 +203,9 @@ class GPT(nn.Module):
     def forward(self, input, targets=None):
         device = input.device
         b, t, d = input.size()
-        assert t <= self.config.gpt_block_size, (
-            f"Cannot forward sequence of length {t}, block size is only {self.config.gpt_block_size}"
-        )
+        assert (
+            t <= self.config.gpt_block_size
+        ), f"Cannot forward sequence of length {t}, block size is only {self.config.gpt_block_size}"
 
         # positional encodings that are added to the input embeddings
         pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0)  # shape (1, t)
@@ -273,10 +273,10 @@ class GPT(nn.Module):
         assert len(inter_params) == 0, "parameters {} made it into both decay/no_decay sets!".format(
             str(inter_params)
         )
-        assert len(param_dict.keys() - union_params) == 0, (
-            "parameters {} were not separated into either decay/no_decay set!".format(
-                str(param_dict.keys() - union_params),
-            )
+        assert (
+            len(param_dict.keys() - union_params) == 0
+        ), "parameters {} were not separated into either decay/no_decay set!".format(
+            str(param_dict.keys() - union_params),
         )
 
         decay = [param_dict[pn] for pn in sorted(decay)]
@@ -419,9 +419,9 @@ class ResidualVQ(nn.Module):
         # and the network should be able to reconstruct
 
         if quantize_dim < self.num_quantizers:
-            assert self.quantize_dropout > 0.0, (
-                "quantize dropout must be greater than 0 if you wish to reconstruct from a signal with less fine quantizations"
-            )
+            assert (
+                self.quantize_dropout > 0.0
+            ), "quantize dropout must be greater than 0 if you wish to reconstruct from a signal with less fine quantizations"
             indices = F.pad(indices, (0, self.num_quantizers - quantize_dim), value=-1)
 
         # get ready for gathering
@@ -472,9 +472,9 @@ class ResidualVQ(nn.Module):
         all_indices = []
 
         if return_loss:
-            assert not torch.any(indices == -1), (
-                "some of the residual vq indices were dropped out. please use indices derived when the module is in eval mode to derive cross entropy loss"
-            )
+            assert not torch.any(
+                indices == -1
+            ), "some of the residual vq indices were dropped out. please use indices derived when the module is in eval mode to derive cross entropy loss"
             ce_losses = []
 
         should_quantize_dropout = self.training and self.quantize_dropout and not return_loss
@@ -887,9 +887,9 @@ class VectorQuantize(nn.Module):
                 # only calculate orthogonal loss for the activated codes for this batch
 
                 if self.orthogonal_reg_active_codes_only:
-                    assert not (is_multiheaded and self.separate_codebook_per_head), (
-                        "orthogonal regularization for only active codes not compatible with multi-headed with separate codebooks yet"
-                    )
+                    assert not (
+                        is_multiheaded and self.separate_codebook_per_head
+                    ), "orthogonal regularization for only active codes not compatible with multi-headed with separate codebooks yet"
                     unique_code_ids = torch.unique(embed_ind)
                     codebook = codebook[:, unique_code_ids]
 
@@ -999,9 +999,9 @@ def gumbel_sample(
     ind = sampling_logits.argmax(dim=dim)
     one_hot = F.one_hot(ind, size).type(dtype)
 
-    assert not (reinmax and not straight_through), (
-        "reinmax can only be turned on if using straight through gumbel softmax"
-    )
+    assert not (
+        reinmax and not straight_through
+    ), "reinmax can only be turned on if using straight through gumbel softmax"
 
     if not straight_through or temperature <= 0.0 or not training:
         return ind, one_hot
@@ -1209,9 +1209,9 @@ class EuclideanCodebook(nn.Module):
         self.gumbel_sample = gumbel_sample
         self.sample_codebook_temp = sample_codebook_temp
 
-        assert not (use_ddp and num_codebooks > 1 and kmeans_init), (
-            "kmeans init is not compatible with multiple codebooks in distributed environment for now"
-        )
+        assert not (
+            use_ddp and num_codebooks > 1 and kmeans_init
+        ), "kmeans init is not compatible with multiple codebooks in distributed environment for now"
 
         self.sample_fn = sample_vectors_distributed if use_ddp and sync_kmeans else batched_sample_vectors
         self.kmeans_all_reduce_fn = distributed.all_reduce if use_ddp and sync_kmeans else noop

lerobot/common/robot_devices/cameras/configs.py

@@ -1,100 +0,0 @@
-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__)
-
-
-@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})")
-
-
-@CameraConfig.register_subclass("intelrealsense")
-@dataclass
-class IntelRealSenseCameraConfig(CameraConfig):
-    """
-    Example of tested options for Intel Real Sense D405:
-
-    ```python
-    IntelRealSenseCameraConfig(128422271347, 30, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 60, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 90, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 30, 1280, 720)
-    IntelRealSenseCameraConfig(128422271347, 30, 640, 480, use_depth=True)
-    IntelRealSenseCameraConfig(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/robot_devices/cameras/intelrealsense.py

@@ -11,13 +11,13 @@ 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.cameras.configs import IntelRealSenseCameraConfig
 from lerobot.common.robot_devices.utils import (
     RobotDeviceAlreadyConnectedError,
     RobotDeviceNotConnectedError,
@@ -94,10 +94,7 @@ def save_images_from_cameras(
     cameras = []
     for cam_sn in serial_numbers:
         print(f"{cam_sn=}")
-        config = IntelRealSenseCameraConfig(
-            serial_number=cam_sn, fps=fps, width=width, height=height, mock=mock
-        )
-        camera = IntelRealSenseCamera(config)
+        camera = IntelRealSenseCamera(cam_sn, fps=fps, width=width, height=height, mock=mock)
         camera.connect()
         print(
             f"IntelRealSenseCamera({camera.serial_number}, fps={camera.fps}, width={camera.width}, height={camera.height}, color_mode={camera.color_mode})"
@@ -152,6 +149,48 @@ def save_images_from_cameras(
             camera.disconnect()
 
 
+@dataclass
+class IntelRealSenseCameraConfig:
+    """
+    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:
@@ -167,35 +206,33 @@ class IntelRealSenseCamera:
     When an IntelRealSenseCamera 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 instantiating with a serial number:
+    Example of usage:
     ```python
-    from lerobot.common.robot_devices.cameras.configs import IntelRealSenseCameraConfig
-
-    config = IntelRealSenseCameraConfig(serial_number=128422271347)
-    camera = IntelRealSenseCamera(config)
+    # Instantiate with its serial number
+    camera = IntelRealSenseCamera(128422271347)
+    # Or by its name if it's unique
+    camera = IntelRealSenseCamera.init_from_name("Intel RealSense D405")
     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 = IntelRealSenseCameraConfig(name="Intel RealSense D405")
-    ```
-
     Example of changing default fps, width, height and color_mode:
     ```python
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, fps=30, width=1280, height=720)
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480)
-    config = IntelRealSenseCameraConfig(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
+    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()
     ```
 
     Example of returning depth:
     ```python
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, use_depth=True)
-    camera = IntelRealSenseCamera(config)
+    camera = IntelRealSenseCamera(serial_number, use_depth=True)
     camera.connect()
     color_image, depth_map = camera.read()
     ```
@@ -203,17 +240,20 @@ class IntelRealSenseCamera:
 
     def __init__(
         self,
-        config: IntelRealSenseCameraConfig,
+        serial_number: int,
+        config: IntelRealSenseCameraConfig | None = None,
+        **kwargs,
     ):
-        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
+        if config is None:
+            config = IntelRealSenseCameraConfig()
+
+        # Overwrite the config arguments using kwargs
+        config = replace(config, **kwargs)
+
+        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
@@ -241,7 +281,8 @@ class IntelRealSenseCamera:
         elif config.rotation == 180:
             self.rotation = cv2.ROTATE_180
 
-    def find_serial_number_from_name(self, name):
+    @classmethod
+    def init_from_name(cls, name: str, config: IntelRealSenseCameraConfig | None = None, **kwargs):
         camera_infos = find_cameras()
         camera_names = [cam["name"] for cam in camera_infos]
         this_name_count = Counter(camera_names)[name]
@@ -254,7 +295,13 @@ class IntelRealSenseCamera:
         name_to_serial_dict = {cam["name"]: cam["serial_number"] for cam in camera_infos}
         cam_sn = name_to_serial_dict[name]
 
-        return cam_sn
+        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)
 
     def connect(self):
         if self.is_connected:

lerobot/common/robot_devices/cameras/opencv.py

@@ -9,13 +9,13 @@ 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.cameras.configs import OpenCVCameraConfig
 from lerobot.common.robot_devices.utils import (
     RobotDeviceAlreadyConnectedError,
     RobotDeviceNotConnectedError,
@@ -126,8 +126,7 @@ def save_images_from_cameras(
     print("Connecting cameras")
     cameras = []
     for cam_idx in camera_ids:
-        config = OpenCVCameraConfig(camera_index=cam_idx, fps=fps, width=width, height=height, mock=mock)
-        camera = OpenCVCamera(config)
+        camera = OpenCVCamera(cam_idx, fps=fps, width=width, height=height, mock=mock)
         camera.connect()
         print(
             f"OpenCVCamera({camera.camera_index}, fps={camera.fps}, width={camera.width}, "
@@ -176,6 +175,36 @@ 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:
     """
     The OpenCVCamera class allows to efficiently record images from cameras. It relies on opencv2 to communicate
@@ -195,10 +224,7 @@ class OpenCVCamera:
 
     Example of usage:
     ```python
-    from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
-
-    config = OpenCVCameraConfig(camera_index=0)
-    camera = OpenCVCamera(config)
+    camera = OpenCVCamera(camera_index=0)
     camera.connect()
     color_image = camera.read()
     # when done using the camera, consider disconnecting
@@ -207,16 +233,25 @@ class OpenCVCamera:
 
     Example of changing default fps, width, height and color_mode:
     ```python
-    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
+    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()
     ```
     """
 
-    def __init__(self, config: OpenCVCameraConfig):
-        self.config = config
-        self.camera_index = config.camera_index
+    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
         self.port = None
 
         # Linux uses ports for connecting to cameras
@@ -228,12 +263,11 @@ 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: {self.camera_index}")
+                raise ValueError(f"Please check the provided camera_index: {camera_index}")
 
         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
 

lerobot/common/robot_devices/cameras/utils.py

@@ -2,12 +2,6 @@ from typing import Protocol
 
 import numpy as np
 
-from lerobot.common.robot_devices.cameras.configs import (
-    CameraConfig,
-    IntelRealSenseCameraConfig,
-    OpenCVCameraConfig,
-)
-
 
 # Defines a camera type
 class Camera(Protocol):
@@ -15,39 +9,3 @@ class Camera(Protocol):
     def read(self, temporary_color: str | None = None) -> np.ndarray: ...
     def async_read(self) -> np.ndarray: ...
     def disconnect(self): ...
-
-
-def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> list[Camera]:
-    cameras = {}
-
-    for key, cfg in camera_configs.items():
-        if cfg.type == "opencv":
-            from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
-
-            cameras[key] = OpenCVCamera(cfg)
-
-        elif cfg.type == "intelrealsense":
-            from lerobot.common.robot_devices.cameras.intelrealsense import IntelRealSenseCamera
-
-            cameras[key] = IntelRealSenseCamera(cfg)
-        else:
-            raise ValueError(f"The motor type '{cfg.type}' is not valid.")
-
-    return cameras
-
-
-def make_camera(camera_type, **kwargs) -> Camera:
-    if camera_type == "opencv":
-        from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
-
-        config = OpenCVCameraConfig(**kwargs)
-        return OpenCVCamera(config)
-
-    elif camera_type == "intelrealsense":
-        from lerobot.common.robot_devices.cameras.intelrealsense import IntelRealSenseCamera
-
-        config = IntelRealSenseCameraConfig(**kwargs)
-        return IntelRealSenseCamera(config)
-
-    else:
-        raise ValueError(f"The camera type '{camera_type}' is not valid.")

lerobot/common/robot_devices/control_configs.py

@@ -1,146 +0,0 @@
-import logging
-from dataclasses import dataclass
-from pathlib import Path
-
-import draccus
-
-from lerobot.common.robot_devices.robots.configs import RobotConfig
-from lerobot.common.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
-from lerobot.configs import parser
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.train import TrainPipelineConfig
-
-
-@dataclass
-class ControlConfig(draccus.ChoiceRegistry):
-    pass
-
-
-@ControlConfig.register_subclass("calibrate")
-@dataclass
-class CalibrateControlConfig(ControlConfig):
-    # List of arms to calibrate (e.g. `--arms='["left_follower","right_follower"]' left_leader`)
-    arms: list[str] | None = None
-
-
-@ControlConfig.register_subclass("teleoperate")
-@dataclass
-class TeleoperateControlConfig(ControlConfig):
-    # Limit the maximum frames per second. By default, no limit.
-    fps: int | None = None
-    teleop_time_s: float | None = None
-    # Display all cameras on screen
-    display_cameras: bool = True
-
-
-@ControlConfig.register_subclass("record")
-@dataclass
-class RecordControlConfig(ControlConfig):
-    # Dataset identifier. By convention it should match '{hf_username}/{dataset_name}' (e.g. `lerobot/test`).
-    repo_id: str
-    # A short but accurate description of the task performed during the recording (e.g. "Pick the Lego block and drop it in the box on the right.")
-    single_task: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path').
-    root: str | Path | None = None
-    policy: PreTrainedConfig | None = None
-    # TODO(rcadene, aliberts): By default, use device and use_amp values from policy checkpoint.
-    device: str | None = None  # cuda | cpu | mps
-    # `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
-    # automatic gradient scaling is used.
-    use_amp: bool | None = None
-    # Limit the frames per second. By default, uses the policy fps.
-    fps: int | None = None
-    # Number of seconds before starting data collection. It allows the robot devices to warmup and synchronize.
-    warmup_time_s: int | float = 10
-    # Number of seconds for data recording for each episode.
-    episode_time_s: int | float = 60
-    # Number of seconds for resetting the environment after each episode.
-    reset_time_s: int | float = 60
-    # Number of episodes to record.
-    num_episodes: int = 50
-    # Encode frames in the dataset into video
-    video: bool = True
-    # By default, run the computation of the data statistics at the end of data collection. Compute intensive and not required to just replay an episode.
-    run_compute_stats: bool = True
-    # Upload dataset to Hugging Face hub.
-    push_to_hub: bool = True
-    # Upload on private repository on the Hugging Face hub.
-    private: bool = False
-    # Add tags to your dataset on the hub.
-    tags: list[str] | None = None
-    # Number of subprocesses handling the saving of frames as PNGs. Set to 0 to use threads only;
-    # set to ≥1 to use subprocesses, each using threads to write images. The best number of processes
-    # and threads depends on your system. We recommend 4 threads per camera with 0 processes.
-    # If fps is unstable, adjust the thread count. If still unstable, try using 1 or more subprocesses.
-    num_image_writer_processes: int = 0
-    # Number of threads writing the frames as png images on disk, per camera.
-    # Too many threads might cause unstable teleoperation fps due to main thread being blocked.
-    # Not enough threads might cause low camera fps.
-    num_image_writer_threads_per_camera: int = 4
-    # Display all cameras on screen
-    display_cameras: bool = True
-    # Use vocal synthesis to read events.
-    play_sounds: bool = True
-    # Resume recording on an existing dataset.
-    resume: bool = False
-    # TODO(rcadene, aliberts): remove local_files_only when refactor with dataset as argument
-    # Use local files only. By default, this script will try to fetch the dataset from the hub if it exists.
-    local_files_only: bool = False
-
-    def __post_init__(self):
-        # HACK: We parse again the cli args here to get the pretrained path if there was one.
-        policy_path = parser.get_path_arg("control.policy")
-        if policy_path:
-            cli_overrides = parser.get_cli_overrides("control.policy")
-            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            self.policy.pretrained_path = policy_path
-
-            # When no device or use_amp are given, use the one from training config.
-            if self.device is None or self.use_amp is None:
-                train_cfg = TrainPipelineConfig.from_pretrained(policy_path)
-                if self.device is None:
-                    self.device = train_cfg.device
-                if self.use_amp is None:
-                    self.use_amp = train_cfg.use_amp
-
-            # Automatically switch to available device if necessary
-            if not is_torch_device_available(self.device):
-                auto_device = auto_select_torch_device()
-                logging.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
-                self.device = auto_device
-
-            # Automatically deactivate AMP if necessary
-            if self.use_amp and not is_amp_available(self.device):
-                logging.warning(
-                    f"Automatic Mixed Precision (amp) is not available on device '{self.device}'. Deactivating AMP."
-                )
-                self.use_amp = False
-
-
-@ControlConfig.register_subclass("replay")
-@dataclass
-class ReplayControlConfig(ControlConfig):
-    # Dataset identifier. By convention it should match '{hf_username}/{dataset_name}' (e.g. `lerobot/test`).
-    repo_id: str
-    # Index of the episode to replay.
-    episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path').
-    root: str | Path | None = None
-    # Limit the frames per second. By default, uses the dataset fps.
-    fps: int | None = None
-    # Use vocal synthesis to read events.
-    play_sounds: bool = True
-    # TODO(rcadene, aliberts): remove local_files_only when refactor with dataset as argument
-    # Use local files only. By default, this script will try to fetch the dataset from the hub if it exists.
-    local_files_only: bool = False
-
-
-@dataclass
-class ControlPipelineConfig:
-    robot: RobotConfig
-    control: ControlConfig
-
-    @classmethod
-    def __get_path_fields__(cls) -> list[str]:
-        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
-        return ["control.policy"]

lerobot/common/robot_devices/control_utils.py

@@ -13,15 +13,14 @@ from functools import cache
 import cv2
 import torch
 import tqdm
-from deepdiff import DeepDiff
 from termcolor import colored
 
-from lerobot.common.datasets.image_writer import safe_stop_image_writer
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.common.datasets.utils import get_features_from_robot
+from lerobot.common.datasets.populate_dataset import add_frame, safe_stop_image_writer
+from lerobot.common.policies.factory import make_policy
 from lerobot.common.robot_devices.robots.utils import Robot
 from lerobot.common.robot_devices.utils import busy_wait
-from lerobot.common.utils.utils import get_safe_torch_device, has_method
+from lerobot.common.utils.utils import get_safe_torch_device, init_hydra_config, set_global_seed
+from lerobot.scripts.eval import get_pretrained_policy_path
 
 
 def log_control_info(robot: Robot, dt_s, episode_index=None, frame_index=None, fps=None):
@@ -33,7 +32,7 @@ def log_control_info(robot: Robot, dt_s, episode_index=None, frame_index=None, f
 
     def log_dt(shortname, dt_val_s):
         nonlocal log_items, fps
-        info_str = f"{shortname}:{dt_val_s * 1000:5.2f} ({1 / dt_val_s:3.1f}hz)"
+        info_str = f"{shortname}:{dt_val_s * 1000:5.2f} ({1/ dt_val_s:3.1f}hz)"
         if fps is not None:
             actual_fps = 1 / dt_val_s
             if actual_fps < fps - 1:
@@ -87,6 +86,10 @@ def is_headless():
         return True
 
 
+def has_method(_object: object, method_name: str):
+    return hasattr(_object, method_name) and callable(getattr(_object, method_name))
+
+
 def predict_action(observation, policy, device, use_amp):
     observation = copy(observation)
     with (
@@ -155,10 +158,30 @@ def init_keyboard_listener():
     return listener, events
 
 
+def init_policy(pretrained_policy_name_or_path, policy_overrides):
+    """Instantiate the policy and load fps, device and use_amp from config yaml"""
+    pretrained_policy_path = get_pretrained_policy_path(pretrained_policy_name_or_path)
+    hydra_cfg = init_hydra_config(pretrained_policy_path / "config.yaml", policy_overrides)
+    policy = make_policy(hydra_cfg=hydra_cfg, pretrained_policy_name_or_path=pretrained_policy_path)
+
+    # Check device is available
+    device = get_safe_torch_device(hydra_cfg.device, log=True)
+    use_amp = hydra_cfg.use_amp
+    policy_fps = hydra_cfg.env.fps
+
+    policy.eval()
+    policy.to(device)
+
+    torch.backends.cudnn.benchmark = True
+    torch.backends.cuda.matmul.allow_tf32 = True
+    set_global_seed(hydra_cfg.seed)
+    return policy, policy_fps, device, use_amp
+
+
 def warmup_record(
     robot,
     events,
-    enable_teleoperation,
+    enable_teloperation,
     warmup_time_s,
     display_cameras,
     fps,
@@ -169,7 +192,7 @@ def warmup_record(
         display_cameras=display_cameras,
         events=events,
         fps=fps,
-        teleoperate=enable_teleoperation,
+        teleoperate=enable_teloperation,
     )
 
 
@@ -204,12 +227,12 @@ def control_loop(
     control_time_s=None,
     teleoperate=False,
     display_cameras=False,
-    dataset: LeRobotDataset | None = None,
+    dataset=None,
     events=None,
     policy=None,
-    device: torch.device | str | None = None,
-    use_amp: bool | None = None,
-    fps: int | None = None,
+    device=None,
+    use_amp=None,
+    fps=None,
 ):
     # TODO(rcadene): Add option to record logs
     if not robot.is_connected:
@@ -224,12 +247,9 @@ def control_loop(
     if teleoperate and policy is not None:
         raise ValueError("When `teleoperate` is True, `policy` should be None.")
 
-    if dataset is not None and fps is not None and dataset.fps != fps:
+    if dataset is not None and fps is not None and dataset["fps"] != fps:
         raise ValueError(f"The dataset fps should be equal to requested fps ({dataset['fps']} != {fps}).")
 
-    if isinstance(device, str):
-        device = get_safe_torch_device(device)
-
     timestamp = 0
     start_episode_t = time.perf_counter()
     while timestamp < control_time_s:
@@ -248,8 +268,7 @@ def control_loop(
                 action = {"action": action}
 
         if dataset is not None:
-            frame = {**observation, **action}
-            dataset.add_frame(frame)
+            add_frame(dataset, observation, action)
 
         if display_cameras and not is_headless():
             image_keys = [key for key in observation if "image" in key]
@@ -301,40 +320,11 @@ def stop_recording(robot, listener, display_cameras):
             cv2.destroyAllWindows()
 
 
-def sanity_check_dataset_name(repo_id, policy_cfg):
+def sanity_check_dataset_name(repo_id, policy):
     _, dataset_name = repo_id.split("/")
     # either repo_id doesnt start with "eval_" and there is no policy
     # or repo_id starts with "eval_" and there is a policy
-
-    # Check if dataset_name starts with "eval_" but policy is missing
-    if dataset_name.startswith("eval_") and policy_cfg is None:
+    if dataset_name.startswith("eval_") == (policy is None):
         raise ValueError(
-            f"Your dataset name begins with 'eval_' ({dataset_name}), but no policy is provided ({policy_cfg.type})."
-        )
-
-    # Check if dataset_name does not start with "eval_" but policy is provided
-    if not dataset_name.startswith("eval_") and policy_cfg is not None:
-        raise ValueError(
-            f"Your dataset name does not begin with 'eval_' ({dataset_name}), but a policy is provided ({policy_cfg.type})."
-        )
-
-
-def sanity_check_dataset_robot_compatibility(
-    dataset: LeRobotDataset, robot: Robot, fps: int, use_videos: bool
-) -> None:
-    fields = [
-        ("robot_type", dataset.meta.robot_type, robot.robot_type),
-        ("fps", dataset.fps, fps),
-        ("features", dataset.features, get_features_from_robot(robot, use_videos)),
-    ]
-
-    mismatches = []
-    for field, dataset_value, present_value in fields:
-        diff = DeepDiff(dataset_value, present_value, exclude_regex_paths=[r".*\['info'\]$"])
-        if diff:
-            mismatches.append(f"{field}: expected {present_value}, got {dataset_value}")
-
-    if mismatches:
-        raise ValueError(
-            "Dataset metadata compatibility check failed with mismatches:\n" + "\n".join(mismatches)
+            f"Your dataset name begins by 'eval_' ({dataset_name}) but no policy is provided ({policy})."
         )

lerobot/common/robot_devices/motors/configs.py

@@ -1,27 +0,0 @@
-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