fix

- Changed the `tokenizer_max_length` from 48 to 200 to enhance the model's capability in handling longer sequences.
- This adjustment aims to improve the overall performance and flexibility of the PI05 configuration.

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -25,7 +25,7 @@ body:
     id: system-info
     attributes:
       label: System Info
-      description: Please share your LeRobot configuration by running `lerobot-info` (if installed) or `python -m lerobot.scripts.display_sys_info` (if not installed) and pasting the output below.
+      description: If needed, you can share your lerobot configuration with us by running `python -m lerobot.scripts.display_sys_info` and copy-pasting its outputs below
       render: Shell
       placeholder: lerobot version, OS, python version, numpy version, torch version, and lerobot's configuration
     validations:

.github/workflows/full_tests.yml

@@ -78,7 +78,7 @@ jobs:
           python-version: ${{ env.PYTHON_VERSION }}
 
       - name: Install lerobot with all extras
-        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
+        run: uv sync --all-extras
 
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10

.github/workflows/nightly.yml

@@ -119,7 +119,6 @@ jobs:
       TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
     container:
       image: ${{ needs.build-docker-cpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
-      options: --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
         password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
@@ -159,36 +158,3 @@ jobs:
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests
         run: make test-end-to-end
-
-  # This job runs multi-GPU training tests with 4 GPUs
-  nightly-multi-gpu-tests:
-    name: Nightly Multi-GPU Tests
-    needs: [build-docker-gpu-nightly]
-    runs-on:
-      group: aws-g4dn-12xlarge  # Instance with 4 GPUs
-    env:
-      HF_HOME: /home/user_lerobot/.cache/huggingface
-      HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
-      TORCH_HOME: /home/user_lerobot/.cache/torch
-      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
-      CUDA_VISIBLE_DEVICES: "0,1,2,3"
-    container:
-      image: ${{ needs.build-docker-gpu-nightly.outputs.image_tag }} # zizmor: ignore[unpinned-images]
-      options: --gpus all --shm-size "16gb"
-      credentials:
-        username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-        password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-    defaults:
-      run:
-        shell: bash
-        working-directory: /lerobot
-    steps:
-      - name: Verify GPU availability
-        run: |
-          nvidia-smi
-          python -c "import torch; print(f'PyTorch CUDA available: {torch.cuda.is_available()}'); print(f'Number of GPUs: {torch.cuda.device_count()}')"
-
-      - name: Run multi-GPU training tests
-      # TODO(Steven): Investigate why motors tests are failing in multi-GPU setup
-        run: pytest tests -vv --maxfail=10 --ignore=tests/motors/
-        timeout-minutes: 10

.github/workflows/release.yml

@@ -82,14 +82,6 @@ jobs:
             exit 1
           fi
 
-      - name: Remove Tags with Git dependencies
-        # TODO(Steven): Temporary patch to remove libero and pi from PyPi 0.4.0 release due to its reliance on git dependencies.
-        run: |
-          echo "::info:: Checking for Git dependencies to remove from pyproject.toml..."
-          grep -E '@ git\+https|lerobot\[pi\]|lerobot\[libero\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
-          sed -E -i '/@ git\+https|lerobot\[pi\]|lerobot\[libero\]/d' pyproject.toml
-          echo "::info:: Git dependencies removed. Proceeding with build."
-
       - name: Install build dependencies
         run: python -m pip install build
 
@@ -111,7 +103,7 @@ jobs:
       - name: Publish to TestPyPI for pre-releases
         # True for tags like 'v0.2.0-rc1'
         if: startsWith(github.ref, 'refs/tags/v') && contains(github.ref, '-')
-        uses: pypa/gh-action-pypi-publish@v1.13.0 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
+        uses: pypa/gh-action-pypi-publish@v1.12.4 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
         with:
           repository-url: https://test.pypi.org/legacy/
           verbose: true
@@ -119,7 +111,7 @@ jobs:
 
       - name: Publish to PyPI
         if: startsWith(github.ref, 'refs/tags/v') && !contains(github.ref, '-')
-        uses: pypa/gh-action-pypi-publish@v1.13.0 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
+        uses: pypa/gh-action-pypi-publish@v1.12.4 # zizmor: ignore[unpinned-uses, use-trusted-publishing]
         with:
           verbose: true
           print-hash: true
@@ -146,7 +138,7 @@ jobs:
       - name: Setup uv and Python
         uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
         with:
-          enable-cache: true # zizmor: ignore[cache-poisoning]
+          enable-cache: true
           version: ${{ env.UV_VERSION }}
           python-version: ${{ env.PYTHON_VERSION }}
       - name: Create uv virtual environment

.github/workflows/stale.yml

@@ -27,17 +27,15 @@ env:
     This issue was closed because it has been stalled for 14 days with no activity.
     Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
   CLOSE_PR_MESSAGE: >
-    This PR was closed because it has been stalled for 21 days with no activity.
+    This PR was closed because it has been stalled for 14 days with no activity.
     Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
   WARN_ISSUE_MESSAGE: >
     This issue has been automatically marked as stale because it has not had
-    recent activity (6 months). It will be closed if no further activity occurs.
-    Any change, comment or update to this issue will reset this count.
+    recent activity (1 year). It will be closed if no further activity occurs.
     Thank you for your contributions.
   WARN_PR_MESSAGE: >
     This PR has been automatically marked as stale because it has not had
     recent activity (1 year). It will be closed if no further activity occurs.
-    Any change, comment or update to this PR will reset this count.
     Thank you for your contributions.
 
 jobs:
@@ -58,10 +56,10 @@ jobs:
           stale-pr-label: stale
           exempt-issue-labels: never-stale
           exempt-pr-labels: never-stale
-          days-before-issue-stale: 180
+          days-before-issue-stale: 180 # TODO(Steven): Will modify this to 90 after initial cleanup
           days-before-issue-close: 14
-          days-before-pr-stale: 365
-          days-before-pr-close: 21
+          days-before-pr-stale: 180
+          days-before-pr-close: 14
           delete-branch: true
           close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
           close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}

.github/workflows/unbound_deps_tests.yml

@@ -1,183 +0,0 @@
-# Copyright 2025 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 workflow handles full testing with unboud dependencies versions.
-name: Unbound Dependency Tests
-
-on:
-  # Allows running this workflow manually from the Actions tab
-  workflow_dispatch:
-
-  # Run on the 1st and 15th of every month at 09:00 UTC
-  schedule:
-    - cron: '0 2 1,15 * *'
-
-permissions:
-  contents: read
-
-# Sets up the environment variables
-env:
-  UV_VERSION: "0.8.0"
-  PYTHON_VERSION: "3.10"
-  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
-
-# Ensures that only the latest action is built, canceling older runs.
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-
-  # This job runs the E2E tests + pytest with all unbound extras
-  full-tests:
-    name: Full Unbound Tests
-    runs-on: ubuntu-latest
-    env:
-      MUJOCO_GL: egl
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          lfs: true
-          persist-credentials: false
-
-      - name: Install apt dependencies
-        run: |
-          sudo apt-get update && sudo apt-get install -y build-essential \
-          git curl libglib2.0-0 libegl1-mesa-dev ffmpeg libusb-1.0-0-dev \
-          speech-dispatcher libgeos-dev portaudio19-dev
-
-      - name: Setup uv and Python
-        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          enable-cache: true
-          version: ${{ env.UV_VERSION }}
-          python-version: ${{ env.PYTHON_VERSION }}
-
-      - name: Unbound dependencies
-        run: |
-          sed -i 's/,[[:space:]]*<[0-9\.]*//g' pyproject.toml
-          echo "Dependencies unbound:" && cat pyproject.toml
-
-      - name: Install lerobot with all extras
-        run: uv sync --all-extras
-
-      - name: Run pytest (all extras)
-        run: uv run pytest tests -vv
-
-      - name: Run end-to-end tests
-        run: uv run make test-end-to-end
-
-  # This job builds a GPU enabled image for testing
-  build-and-push-docker:
-    name: Build and Push Docker
-    runs-on:
-      group: aws-general-8-plus
-    outputs:
-      image_tag: ${{ env.DOCKER_IMAGE_NAME }}
-    env:
-      GITHUB_REF: ${{ github.ref }}
-    steps:
-      - name: Install Git LFS
-        run: |
-          sudo apt-get update
-          sudo apt-get install git-lfs
-          git lfs install
-      - uses: actions/checkout@v4
-        with:
-          lfs: true
-          persist-credentials: false
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          cache-binary: false
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
-        with:
-          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-      - name: Build and push Docker image
-        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
-        with:
-          context: .
-          file: ./docker/Dockerfile.internal
-          push: true
-          tags: ${{ env.DOCKER_IMAGE_NAME }}
-          build-args: |
-            UNBOUND_DEPS=true
-
-  # This job runs pytest with all unbound extras in a GPU enabled host
-  # It runs everytime a test image is created
-  gpu-tests:
-    name: GPU Unbound Tests
-    needs: [build-and-push-docker]
-    runs-on:
-      group: aws-g6-4xlarge-plus
-    env:
-      HF_HOME: /home/user_lerobot/.cache/huggingface
-      HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
-      TORCH_HOME: /home/user_lerobot/.cache/torch
-      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
-    container:
-      image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
-      options: --gpus all --shm-size "16gb"
-      credentials:
-        username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
-        password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
-    defaults:
-      run:
-        shell: bash
-        working-directory: /lerobot
-    steps:
-      - name: Run pytest on GPU
-        run: pytest tests -vv
-      - name: Run end-to-end tests
-        run: make test-end-to-end
-
-  # This job deletes the test image recently created
-  # It runs everytime after the gpu-tests have finished
-  delete-unbound-image:
-    name: Delete Unbound Image
-    needs: [gpu-tests, build-and-push-docker]
-    if: always() && needs.build-and-push-docker.result == 'success'
-    runs-on: ubuntu-latest
-    steps:
-      - name: Get Docker Hub Token and Delete Image
-        # zizmor: ignore[template-injection]
-        run: |
-          IMAGE_NAME=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f1)
-          IMAGE_TAG=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f2)
-
-          echo "Attempting to delete image: $IMAGE_NAME:$IMAGE_TAG"
-
-          TOKEN=$(curl -s -H "Content-Type: application/json" \
-                       -X POST \
-                       -d '{"username": "${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}", "password": "${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}"}' \
-                       https://hub.docker.com/v2/users/login/ | jq -r .token)
-
-          if [ "$TOKEN" == "null" ] || [ -z "$TOKEN" ]; then
-            echo "::error::Failed to get Docker Hub token."
-            exit 1
-          fi
-
-          HTTP_RESPONSE=$(curl -s -o /dev/null -w "%{http_code}" \
-                               -H "Authorization: JWT ${TOKEN}" \
-                               -X DELETE \
-                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/${IMAGE_TAG}/)
-
-          if [ "$HTTP_RESPONSE" -eq 204 ]; then
-            echo "Successfully deleted Docker image tag: $IMAGE_NAME:$IMAGE_TAG"
-          else
-            echo "::error::Failed to delete Docker image. HTTP status: $HTTP_RESPONSE"
-            exit 1
-          fi

.pre-commit-config.yaml

@@ -26,7 +26,7 @@ repos:
 
    ##### General Code Quality & Formatting #####
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v6.0.0
+    rev: v5.0.0
     hooks:
       - id: check-added-large-files
         args: ['--maxkb=1024']
@@ -39,20 +39,20 @@ repos:
       - id: trailing-whitespace
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.14.1
+    rev: v0.12.4
     hooks:
       - id: ruff-format
       - id: ruff
         args: [--fix, --exit-non-zero-on-fix]
 
   - repo: https://github.com/adhtruong/mirrors-typos
-    rev: v1.38.1
+    rev: v1.34.0
     hooks:
       - id: typos
         args: [--force-exclude]
 
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.21.0
+    rev: v3.20.0
     hooks:
     -   id: pyupgrade
         args: [--py310-plus]
@@ -68,12 +68,12 @@ repos:
 
   ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.28.0
+    rev: v8.27.2
     hooks:
       - id: gitleaks
 
   - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.15.2
+    rev: v1.11.0
     hooks:
       - id: zizmor
 
@@ -86,12 +86,11 @@ repos:
 
   # TODO(Steven): Uncomment when ready to use
   ##### Static Analysis & Typing #####
-  - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.18.2
-    hooks:
-      - id: mypy
-        args: [--config-file=pyproject.toml]
-        exclude: ^(examples|benchmarks|tests)/
+  # - repo: https://github.com/pre-commit/mirrors-mypy
+  #   rev: v1.16.0
+  #   hooks:
+  #     - id: mypy
+  #       args: [--python-version=3.10]
 
   ##### Docstring Checks #####
   # - repo: https://github.com/akaihola/darglint2

CONTRIBUTING.md

@@ -72,6 +72,7 @@ post it.
 
 Look at our implementations for [datasets](./src/lerobot/datasets/), [policies](./src/lerobot/policies/),
 environments ([aloha](https://github.com/huggingface/gym-aloha),
+[xarm](https://github.com/huggingface/gym-xarm),
 [pusht](https://github.com/huggingface/gym-pusht))
 and follow the same api design.
 
@@ -137,7 +138,7 @@ Follow these steps to start contributing:
 4. for development, we advise to use a tool like `poetry` or `uv` instead of just `pip` to easily track our dependencies.
    Follow the instructions to [install poetry](https://python-poetry.org/docs/#installation) (use a version >=2.1.0) or to [install uv](https://docs.astral.sh/uv/getting-started/installation/#installation-methods) if you don't have one of them already.
 
-   Set up a development environment with conda:
+   Set up a development environment with conda or miniconda:
 
    ```bash
    conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev

Makefile

@@ -119,9 +119,10 @@ test-tdmpc-ete-train:
 		--policy.type=tdmpc \
 		--policy.device=$(DEVICE) \
 		--policy.push_to_hub=false \
-		--env.type=pusht \
+		--env.type=xarm \
+		--env.task=XarmLift-v0 \
 		--env.episode_length=5 \
-		--dataset.repo_id=lerobot/pusht_image \
+		--dataset.repo_id=lerobot/xarm_lift_medium \
 		--dataset.image_transforms.enable=true \
 		--dataset.episodes="[0]" \
 		--batch_size=2 \
@@ -139,10 +140,9 @@ test-tdmpc-ete-eval:
 	lerobot-eval \
 		--policy.path=tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
 		--policy.device=$(DEVICE) \
-		--env.type=pusht \
+		--env.type=xarm \
 		--env.episode_length=5 \
-		--env.observation_height=96 \
-        --env.observation_width=96 \
+		--env.task=XarmLift-v0 \
 		--eval.n_episodes=1 \
 		--eval.batch_size=1
 

README.md

@@ -104,14 +104,14 @@ LeRobot works with Python 3.10+ and PyTorch 2.2+.
 
 ### Environment Setup
 
-Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniforge`](https://conda-forge.org/download/):
+Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/free/miniconda/index.html):
 
 ```bash
 conda create -y -n lerobot python=3.10
 conda activate lerobot
 ```
 
-When using `conda`, install `ffmpeg` in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
 
 ```bash
 conda install ffmpeg -c conda-forge
@@ -185,11 +185,6 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
-> [!NOTE]
-> For lerobot 0.4.0, if you want to install libero or pi tags, you will have to do: `pip install "lerobot[pi,libero]@git+https://github.com/huggingface/lerobot.git"`.
->
-> This will be solved in the next patch release
-
 ### Weights & Biases
 
 To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
@@ -202,23 +197,23 @@ wandb login
 
 ### Visualize datasets
 
-Check out [example 1](https://github.com/huggingface/lerobot/blob/main/examples/dataset/load_lerobot_dataset.py) that illustrates how to use our dataset class which automatically downloads data from the Hugging Face hub.
+Check out [example 1](https://github.com/huggingface/lerobot/blob/main/examples/1_load_lerobot_dataset.py) that illustrates how to use our dataset class which automatically downloads data from the Hugging Face hub.
 
 You can also locally visualize episodes from a dataset on the hub by executing our script from the command line:
 
 ```bash
-lerobot-dataset-viz \
+python -m lerobot.scripts.visualize_dataset \
     --repo-id lerobot/pusht \
     --episode-index 0
 ```
 
-or from a dataset in a local folder with the `root` option and the `--mode local` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
+or from a dataset in a local folder with the `root` option and the `--local-files-only` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
 
 ```bash
-lerobot-dataset-viz \
+python -m lerobot.scripts.visualize_dataset \
     --repo-id lerobot/pusht \
     --root ./my_local_data_dir \
-    --mode local \
+    --local-files-only 1 \
     --episode-index 0
 ```
 
@@ -226,7 +221,7 @@ It will open `rerun.io` and display the camera streams, robot states and actions
 
 https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-fd46b787-b532-47e2-bb6f-fd536a55a7ed.mov?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20240505%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240505T172924Z&X-Amz-Expires=300&X-Amz-Signature=d680b26c532eeaf80740f08af3320d22ad0b8a4e4da1bcc4f33142c15b509eda&X-Amz-SignedHeaders=host&actor_id=24889239&key_id=0&repo_id=748713144
 
-Our script can also visualize datasets stored on a distant server. See `lerobot-dataset-viz --help` for more instructions.
+Our script can also visualize datasets stored on a distant server. See `python -m lerobot.scripts.visualize_dataset --help` for more instructions.
 
 ### The `LeRobotDataset` format
 
@@ -315,7 +310,7 @@ To upload these to the hub, run the following:
 huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
 ```
 
-See [lerobot_eval.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_eval.py) for an example of how other people may use your policy.
+See [eval.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/eval.py) for an example of how other people may use your policy.
 
 ### Acknowledgment
 
@@ -342,3 +337,7 @@ If you want, you can cite this work with:
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)
+
+```
+
+```

benchmarks/video/run_video_benchmark.py

@@ -35,13 +35,12 @@ import torch
 from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
 from tqdm import tqdm
 
-from benchmarks.video.benchmark import TimeBenchmark
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.video_utils import (
     decode_video_frames_torchvision,
     encode_video_frames,
 )
-from lerobot.utils.constants import OBS_IMAGE
+from lerobot.utils.benchmark import TimeBenchmark
 
 BASE_ENCODING = OrderedDict(
     [
@@ -118,7 +117,7 @@ def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
     hf_dataset = dataset.hf_dataset.with_format(None)
 
     # We only save images from the first camera
-    img_keys = [key for key in hf_dataset.features if key.startswith(OBS_IMAGE)]
+    img_keys = [key for key in hf_dataset.features if key.startswith("observation.image")]
     imgs_dataset = hf_dataset.select_columns(img_keys[0])
 
     for i, item in enumerate(

docker/Dockerfile.internal

@@ -75,14 +75,6 @@ RUN uv venv --python python${PYTHON_VERSION}
 # Install Python dependencies for caching
 COPY --chown=user_lerobot:user_lerobot pyproject.toml README.md MANIFEST.in ./
 COPY --chown=user_lerobot:user_lerobot src/ src/
-
-ARG UNBOUND_DEPS=false
-
-RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
-    sed -i 's/,[[:space:]]*<[0-9\.]*//g' pyproject.toml; \
-    echo "Dependencies unbound:" && cat pyproject.toml; \
-    fi
-
 RUN uv pip install --no-cache ".[all]"
 
 # Copy the rest of the application source code

docker/Dockerfile.user

@@ -61,14 +61,6 @@ RUN uv venv
 # Install Python dependencies for caching
 COPY --chown=user_lerobot:user_lerobot pyproject.toml README.md MANIFEST.in ./
 COPY --chown=user_lerobot:user_lerobot src/ src/
-
-ARG UNBOUND_DEPS=false
-
-RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
-    sed -i 's/,[[:space:]]*<[0-9\.]*//g' pyproject.toml; \
-    echo "Dependencies unbound:" && cat pyproject.toml; \
-    fi
-
 RUN uv pip install --no-cache ".[all]"
 
 # Copy the rest of the application code

docs/source/_toctree.yml

@@ -7,6 +7,8 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
+  - local: il_sim
+    title: Imitation Learning in Sim
   - local: cameras
     title: Cameras
   - local: integrate_hardware
@@ -17,39 +19,23 @@
     title: Train RL in Simulation
   - local: async
     title: Use Async Inference
-  - local: multi_gpu_training
-    title: Multi GPU training
   title: "Tutorials"
 - sections:
   - local: lerobot-dataset-v3
     title: Using LeRobotDataset
   - local: porting_datasets_v3
     title: Porting Large Datasets
-  - local: using_dataset_tools
-    title: Using the Dataset Tools
   title: "Datasets"
 - sections:
-  - local: act
-    title: ACT
   - local: smolvla
     title: SmolVLA
   - local: pi0
     title: π₀ (Pi0)
   - local: pi05
     title: π₀.₅ (Pi05)
-  - local: groot
-    title: NVIDIA GR00T N1.5
-  title: "Policies"
-- sections:
-  - local: envhub
-    title: Environments from the Hub
-  - local: il_sim
-    title: Imitation Learning in Sim
   - local: libero
     title: Using Libero
-  - local: metaworld
-    title: Using MetaWorld
-  title: "Simulation"
+  title: "Policies"
 - sections:
   - local: introduction_processors
     title: Introduction to Robot Processors

docs/source/act.mdx

@@ -1,92 +0,0 @@
-# ACT (Action Chunking with Transformers)
-
-ACT is a **lightweight and efficient policy for imitation learning**, especially well-suited for fine-grained manipulation tasks. It's the **first model we recommend when you're starting out** with LeRobot due to its fast training time, low computational requirements, and strong performance.
-
-<div class="video-container">
-  <iframe
-    width="100%"
-    height="415"
-    src="https://www.youtube.com/embed/ft73x0LfGpM"
-    title="LeRobot ACT Tutorial"
-    frameborder="0"
-    allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture"
-    allowfullscreen
-  ></iframe>
-</div>
-
-_Watch this tutorial from the LeRobot team to learn how ACT works: [LeRobot ACT Tutorial](https://www.youtube.com/watch?v=ft73x0LfGpM)_
-
-## Model Overview
-
-Action Chunking with Transformers (ACT) was introduced in the paper [Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware](https://arxiv.org/abs/2304.13705) by Zhao et al. The policy was designed to enable precise, contact-rich manipulation tasks using affordable hardware and minimal demonstration data.
-
-### Why ACT is Great for Beginners
-
-ACT stands out as an excellent starting point for several reasons:
-
-- **Fast Training**: Trains in a few hours on a single GPU
-- **Lightweight**: Only ~80M parameters, making it efficient and easy to work with
-- **Data Efficient**: Often achieves high success rates with just 50 demonstrations
-
-### Architecture
-
-ACT uses a transformer-based architecture with three main components:
-
-1. **Vision Backbone**: ResNet-18 processes images from multiple camera viewpoints
-2. **Transformer Encoder**: Synthesizes information from camera features, joint positions, and a learned latent variable
-3. **Transformer Decoder**: Generates coherent action sequences using cross-attention
-
-The policy takes as input:
-
-- Multiple RGB images (e.g., from wrist cameras, front/top cameras)
-- Current robot joint positions
-- A latent style variable `z` (learned during training, set to zero during inference)
-
-And outputs a chunk of `k` future action sequences.
-
-## Installation Requirements
-
-1. Install LeRobot by following our [Installation Guide](./installation).
-2. ACT is included in the base LeRobot installation, so no additional dependencies are needed!
-
-## Training ACT
-
-ACT works seamlessly with the standard LeRobot training pipeline. Here's a complete example for training ACT on your dataset:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=${HF_USER}/your_dataset \
-  --policy.type=act \
-  --output_dir=outputs/train/act_your_dataset \
-  --job_name=act_your_dataset \
-  --policy.device=cuda \
-  --wandb.enable=true \
-  --policy.repo_id=${HF_USER}/act_policy
-```
-
-### Training Tips
-
-1. **Start with defaults**: ACT's default hyperparameters work well for most tasks
-2. **Training duration**: Expect a few hours for 100k training steps on a single GPU
-3. **Batch size**: Start with batch size 8 and adjust based on your GPU memory
-
-### Train using Google Colab
-
-If your local computer doesn't have a powerful GPU, you can utilize Google Colab to train your model by following the [ACT training notebook](./notebooks#training-act).
-
-## Evaluating ACT
-
-Once training is complete, you can evaluate your ACT policy using the `lerobot-record` command with your trained policy. This will run inference and record evaluation episodes:
-
-```bash
-lerobot-record \
-  --robot.type=so100_follower \
-  --robot.port=/dev/ttyACM0 \
-  --robot.id=my_robot \
-  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
-  --display_data=true \
-  --dataset.repo_id=${HF_USER}/eval_act_your_dataset \
-  --dataset.num_episodes=10 \
-  --dataset.single_task="Your task description" \
-  --policy.path=${HF_USER}/act_policy
-```

docs/source/async.mdx

@@ -31,15 +31,15 @@ Then, spin up a policy server (in one terminal, or in a separate machine) specif
 You can spin up a policy server running:
 
 ```shell
-python -m lerobot.async_inference.policy_server \
-     --host=127.0.0.1 \
-     --port=8080
+python src/lerobot/scripts/server/policy_server.py \
+    --host=127.0.0.1 \
+    --port=8080 \
 ```
 
 This will start a policy server listening on `127.0.0.1:8080` (`localhost`, port 8080). At this stage, the policy server is empty, as all information related to which policy to run and with which parameters are specified during the first handshake with the client. Spin up a client with:
 
 ```shell
-python -m lerobot.async_inference.robot_client \
+python src/lerobot/scripts/server/robot_client.py \
     --server_address=127.0.0.1:8080 \ # SERVER: the host address and port of the policy server
     --robot.type=so100_follower \ # ROBOT: your robot type
     --robot.port=/dev/tty.usbmodem585A0076841 \ # ROBOT: your robot port
@@ -113,17 +113,17 @@ As such, spinning up a policy server is as easy as specifying the host address a
 <hfoptions id="start_policy_server">
 <hfoption id="Command">
 ```bash
-python -m lerobot.async_inference.policy_server \
-     --host=127.0.0.1 \
-     --port=8080
+python -m lerobot.scripts.server.policy_server \
+    --host="localhost" \
+    --port=8080
 ```
 </hfoption>
 <hfoption id="API example">
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.async_inference.configs import PolicyServerConfig
-from lerobot.async_inference.policy_server import serve
+from lerobot.scripts.server.configs import PolicyServerConfig
+from lerobot.scripts.server.policy_server import serve
 
 config = PolicyServerConfig(
     host="localhost",
@@ -148,7 +148,7 @@ The `RobotClient` streams observations to the `PolicyServer`, and receives actio
 <hfoptions id="start_robot_client">
 <hfoption id="Command">
 ```bash
-python -m lerobot.async_inference.robot_client \
+python src/lerobot/scripts/server/robot_client.py \
     --server_address=127.0.0.1:8080 \ # SERVER: the host address and port of the policy server
     --robot.type=so100_follower \ # ROBOT: your robot type
     --robot.port=/dev/tty.usbmodem585A0076841 \ # ROBOT: your robot port
@@ -171,9 +171,9 @@ python -m lerobot.async_inference.robot_client \
 import threading
 from lerobot.robots.so100_follower import SO100FollowerConfig
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.async_inference.configs import RobotClientConfig
-from lerobot.async_inference.robot_client import RobotClient
-from lerobot.async_inference.helpers import visualize_action_queue_size
+from lerobot.scripts.server.configs import RobotClientConfig
+from lerobot.scripts.server.robot_client import RobotClient
+from lerobot.scripts.server.helpers import visualize_action_queue_size
 
 # 1. Create the robot instance
 """Check out the cameras available in your setup by running `python lerobot/find_cameras.py`"""

docs/source/envhub.mdx

@@ -1,424 +0,0 @@
-# Loading Environments from the Hub
-
-The **EnvHub** feature allows you to load simulation environments directly from the Hugging Face Hub with a single line of code. This unlocks a powerful new model for collaboration: instead of environments being locked away inside monolithic libraries, anyone can publish custom environments and share them with the community.
-
-## Overview
-
-With EnvHub, you can:
-
-- Load environments from the Hub instantly
-- Share your custom simulation tasks with the community
-- Version control your environments using Git
-- Distribute complex physics simulations without packaging hassles
-
-## Quick Start
-
-Loading an environment from the Hub is as simple as:
-
-```python
-from lerobot.envs.factory import make_env
-
-# Load a hub environment (requires explicit consent to run remote code)
-env = make_env("lerobot/cartpole-env", trust_remote_code=True)
-```
-
-<Tip warning={true}>
-  **Security Notice**: Loading environments from the Hub executes Python code
-  from third-party repositories. Only use `trust_remote_code=True` with
-  repositories you trust. We strongly recommend pinning to a specific commit
-  hash for reproducibility and security.
-</Tip>
-
-## What is EnvHub?
-
-EnvHub is a framework that allows researchers and developers to:
-
-1. **Publish environments** to the Hugging Face Hub as Git repositories
-2. **Load environments** dynamically without installing them as packages
-3. **Version and track** environment changes using Git semantics
-4. **Discover** new simulation tasks shared by the community
-
-This design means you can go from discovering an interesting environment on the Hub to running experiments in seconds, without worrying about dependency conflicts or complex installation procedures.
-
-## Repository Structure
-
-To make your environment loadable from the Hub, your repository must contain at minimum:
-
-### Required Files
-
-**`env.py`** (or custom Python file)
-
-- Must expose a `make_env(n_envs: int, use_async_envs: bool)` function
-- This function should return one of:
-  - A `gym.vector.VectorEnv` (most common)
-  - A single `gym.Env` (will be automatically wrapped)
-  - A dict mapping `{suite_name: {task_id: VectorEnv}}` (for multi-task benchmarks)
-
-### Optional Files
-
-**`requirements.txt`**
-
-- List any additional dependencies your environment needs
-- Users will need to install these manually before loading your environment
-
-**`README.md`**
-
-- Document your environment: what task it implements, observation/action spaces, rewards, etc.
-- Include usage examples and any special setup instructions
-
-**`.gitignore`**
-
-- Exclude unnecessary files from your repository
-
-### Example Repository Structure
-
-```
-my-environment-repo/
-├── env.py                 # Main environment definition (required)
-├── requirements.txt       # Dependencies (optional)
-├── README.md             # Documentation (recommended)
-├── assets/               # Images, videos, etc. (optional)
-│   └── demo.gif
-└── configs/              # Config files if needed (optional)
-    └── task_config.yaml
-```
-
-## Creating Your Environment Repository
-
-### Step 1: Define Your Environment
-
-Create an `env.py` file with a `make_env` function:
-
-```python
-# env.py
-import gymnasium as gym
-
-def make_env(n_envs: int = 1, use_async_envs: bool = False):
-    """
-    Create vectorized environments for your custom task.
-
-    Args:
-        n_envs: Number of parallel environments
-        use_async_envs: Whether to use AsyncVectorEnv or SyncVectorEnv
-
-    Returns:
-        gym.vector.VectorEnv or dict mapping suite names to vectorized envs
-    """
-    def _make_single_env():
-        # Create your custom environment
-        return gym.make("CartPole-v1")
-
-    # Choose vector environment type
-    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
-
-    # Create vectorized environment
-    vec_env = env_cls([_make_single_env for _ in range(n_envs)])
-
-    return vec_env
-```
-
-### Step 2: Test Locally
-
-Before uploading, test your environment locally:
-
-```python
-from lerobot.envs.utils import _load_module_from_path, _call_make_env, _normalize_hub_result
-
-# Load your module
-module = _load_module_from_path("./env.py")
-
-# Test the make_env function
-result = _call_make_env(module, n_envs=2, use_async_envs=False)
-normalized = _normalize_hub_result(result)
-
-# Verify it works
-suite_name = next(iter(normalized))
-env = normalized[suite_name][0]
-obs, info = env.reset()
-print(f"Observation shape: {obs.shape if hasattr(obs, 'shape') else type(obs)}")
-env.close()
-```
-
-### Step 3: Upload to the Hub
-
-Upload your repository to Hugging Face:
-
-```bash
-# Install huggingface_hub if needed
-pip install huggingface_hub
-
-# Login to Hugging Face
-huggingface-cli login
-
-# Create a new repository
-huggingface-cli repo create my-custom-env --type space --org my-org
-
-# Initialize git and push
-git init
-git add .
-git commit -m "Initial environment implementation"
-git remote add origin https://huggingface.co/my-org/my-custom-env
-git push -u origin main
-```
-
-Alternatively, use the `huggingface_hub` Python API:
-
-```python
-from huggingface_hub import HfApi
-
-api = HfApi()
-
-# Create repository
-api.create_repo("my-custom-env", repo_type="space")
-
-# Upload files
-api.upload_folder(
-    folder_path="./my-env-folder",
-    repo_id="username/my-custom-env",
-    repo_type="space",
-)
-```
-
-## Loading Environments from the Hub
-
-### Basic Usage
-
-```python
-from lerobot.envs.factory import make_env
-
-# Load from the hub
-envs_dict = make_env(
-    "username/my-custom-env",
-    n_envs=4,
-    trust_remote_code=True
-)
-
-# Access the environment
-suite_name = next(iter(envs_dict))
-env = envs_dict[suite_name][0]
-
-# Use it like any gym environment
-obs, info = env.reset()
-action = env.action_space.sample()
-obs, reward, terminated, truncated, info = env.step(action)
-```
-
-### Advanced: Pinning to Specific Versions
-
-For reproducibility and security, pin to a specific Git revision:
-
-```python
-# Pin to a specific branch
-env = make_env("username/my-env@main", trust_remote_code=True)
-
-# Pin to a specific commit (recommended for papers/experiments)
-env = make_env("username/my-env@abc123def456", trust_remote_code=True)
-
-# Pin to a tag
-env = make_env("username/my-env@v1.0.0", trust_remote_code=True)
-```
-
-### Custom File Paths
-
-If your environment definition is not in `env.py`:
-
-```python
-# Load from a custom file
-env = make_env("username/my-env:custom_env.py", trust_remote_code=True)
-
-# Combine with version pinning
-env = make_env("username/my-env@v1.0:envs/task_a.py", trust_remote_code=True)
-```
-
-### Async Environments
-
-For better performance with multiple environments:
-
-```python
-envs_dict = make_env(
-    "username/my-env",
-    n_envs=8,
-    use_async_envs=True,  # Use AsyncVectorEnv for parallel execution
-    trust_remote_code=True
-)
-```
-
-## URL Format Reference
-
-The hub URL format supports several patterns:
-
-| Pattern              | Description                    | Example                                |
-| -------------------- | ------------------------------ | -------------------------------------- |
-| `user/repo`          | Load `env.py` from main branch | `make_env("lerobot/pusht-env")`        |
-| `user/repo@revision` | Load from specific revision    | `make_env("lerobot/pusht-env@main")`   |
-| `user/repo:path`     | Load custom file               | `make_env("lerobot/envs:pusht.py")`    |
-| `user/repo@rev:path` | Revision + custom file         | `make_env("lerobot/envs@v1:pusht.py")` |
-
-## Multi-Task Environments
-
-For benchmarks with multiple tasks (like LIBERO), return a nested dictionary:
-
-```python
-def make_env(n_envs: int = 1, use_async_envs: bool = False):
-    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
-
-    # Return dict: {suite_name: {task_id: VectorEnv}}
-    return {
-        "suite_1": {
-            0: env_cls([lambda: gym.make("Task1-v0") for _ in range(n_envs)]),
-            1: env_cls([lambda: gym.make("Task2-v0") for _ in range(n_envs)]),
-        },
-        "suite_2": {
-            0: env_cls([lambda: gym.make("Task3-v0") for _ in range(n_envs)]),
-        }
-    }
-```
-
-## Security Considerations
-
-<Tip warning={true}>
-  **Important**: The `trust_remote_code=True` flag is required to execute
-  environment code from the Hub. This is by design for security.
-</Tip>
-
-When loading environments from the Hub:
-
-1. **Review the code first**: Visit the repository and inspect `env.py` before loading
-2. **Pin to commits**: Use specific commit hashes for reproducibility
-3. **Check dependencies**: Review `requirements.txt` for suspicious packages
-4. **Use trusted sources**: Prefer official organizations or well-known researchers
-5. **Sandbox if needed**: Run untrusted code in isolated environments (containers, VMs)
-
-Example of safe usage:
-
-```python
-# ❌ BAD: Loading without inspection
-env = make_env("random-user/untrusted-env", trust_remote_code=True)
-
-# ✅ GOOD: Review code, then pin to specific commit
-# 1. Visit https://huggingface.co/trusted-org/verified-env
-# 2. Review the env.py file
-# 3. Copy the commit hash
-env = make_env("trusted-org/verified-env@a1b2c3d4", trust_remote_code=True)
-```
-
-## Example: CartPole from the Hub
-
-Here's a complete example using the reference CartPole environment:
-
-```python
-from lerobot.envs.factory import make_env
-import numpy as np
-
-# Load the environment
-envs_dict = make_env("lerobot/cartpole-env", n_envs=4, trust_remote_code=True)
-
-# Get the vectorized environment
-suite_name = next(iter(envs_dict))
-env = envs_dict[suite_name][0]
-
-# Run a simple episode
-obs, info = env.reset()
-done = np.zeros(env.num_envs, dtype=bool)
-total_reward = np.zeros(env.num_envs)
-
-while not done.all():
-    # Random policy
-    action = env.action_space.sample()
-    obs, reward, terminated, truncated, info = env.step(action)
-    total_reward += reward
-    done = terminated | truncated
-
-print(f"Average reward: {total_reward.mean():.2f}")
-env.close()
-```
-
-## Benefits of EnvHub
-
-### For Environment Authors
-
-- **Easy distribution**: No PyPI packaging required
-- **Version control**: Use Git for environment versioning
-- **Rapid iteration**: Push updates instantly
-- **Documentation**: Hub README renders beautifully
-- **Community**: Reach LeRobot users directly
-
-### For Researchers
-
-- **Quick experiments**: Load any environment in one line
-- **Reproducibility**: Pin to specific commits
-- **Discovery**: Browse environments on the Hub
-- **No conflicts**: No need to install conflicting packages
-
-### For the Community
-
-- **Growing ecosystem**: More diverse simulation tasks
-- **Standardization**: Common `make_env` API
-- **Collaboration**: Fork and improve existing environments
-- **Accessibility**: Lower barrier to sharing research
-
-## Troubleshooting
-
-### "Refusing to execute remote code"
-
-You must explicitly pass `trust_remote_code=True`:
-
-```python
-env = make_env("user/repo", trust_remote_code=True)
-```
-
-### "Module X not found"
-
-The hub environment has dependencies you need to install:
-
-```bash
-# Check the repo's requirements.txt and install dependencies
-pip install gymnasium numpy
-```
-
-### "make_env not found in module"
-
-Your `env.py` must expose a `make_env` function:
-
-```python
-def make_env(n_envs: int, use_async_envs: bool):
-    # Your implementation
-    pass
-```
-
-### Environment returns wrong type
-
-The `make_env` function must return:
-
-- A `gym.vector.VectorEnv`, or
-- A single `gym.Env`, or
-- A dict `{suite_name: {task_id: VectorEnv}}`
-
-## Best Practices
-
-1. **Document your environment**: Include observation/action space descriptions, reward structure, and termination conditions in your README
-2. **Add requirements.txt**: List all dependencies with versions
-3. **Test thoroughly**: Verify your environment works locally before pushing
-4. **Use semantic versioning**: Tag releases with version numbers
-5. **Add examples**: Include usage examples in your README
-6. **Keep it simple**: Minimize dependencies when possible
-7. **License your work**: Add a LICENSE file to clarify usage terms
-
-## Future Directions
-
-The EnvHub ecosystem enables exciting possibilities:
-
-- **GPU-accelerated physics**: Share Isaac Gym or Brax environments
-- **Photorealistic rendering**: Distribute environments with advanced graphics
-- **Multi-agent scenarios**: Complex interaction tasks
-- **Real-world simulators**: Digital twins of physical setups
-- **Procedural generation**: Infinite task variations
-- **Domain randomization**: Pre-configured DR pipelines
-
-As more researchers and developers contribute, the diversity and quality of available environments will grow, benefiting the entire robotics learning community.
-
-## See Also
-
-- [Hugging Face Hub Documentation](https://huggingface.co/docs/hub/en/index)
-- [Gymnasium Documentation](https://gymnasium.farama.org/index.html)
-- [Example Hub Environment](https://huggingface.co/lerobot/cartpole-env)

docs/source/groot.mdx

@@ -1,125 +0,0 @@
-# GR00T N1.5 Policy
-
-GR00T N1.5 is an open foundation model from NVIDIA designed for generalized humanoid robot reasoning and skills. It is a cross-embodiment model that accepts multimodal input, including language and images, to perform manipulation tasks in diverse environments.
-
-This document outlines the specifics of its integration and usage within the LeRobot framework.
-
-## Model Overview
-
-NVIDIA Isaac GR00T N1.5 is an upgraded version of the GR00T N1 foundation model. It is built to improve generalization and language-following abilities for humanoid robots.
-
-Developers and researchers can post-train GR00T N1.5 with their own real or synthetic data to adapt it for specific humanoid robots or tasks.
-
-GR00T N1.5 (specifically the GR00T-N1.5-3B model) is built using pre-trained vision and language encoders. It utilizes a flow matching action transformer to model a chunk of actions, conditioned on vision, language, and proprioception.
-
-Its strong performance comes from being trained on an expansive and diverse humanoid dataset, which includes:
-
-- Real captured data from robots.
-- Synthetic data generated using NVIDIA Isaac GR00T Blueprint.
-- Internet-scale video data.
-
-This approach allows the model to be highly adaptable through post-training for specific embodiments, tasks, and environments.
-
-## Installation Requirements
-
-As of today, GR00T N1.5 requires flash attention for it's internal working.
-
-We are working on making this optional, but in the meantime that means that we require an extra installation step and it can only be used in CUDA enabled devices.
-
-1. Following the Environment Setup of our [Installation Guide](./installation). **Attention** don't install `lerobot` in this step.
-2. Install [Flash Attention](https://github.com/Dao-AILab/flash-attention) by running:
-
-```bash
-# Check https://pytorch.org/get-started/locally/ for your system
-pip install "torch>=2.2.1,<2.8.0" "torchvision>=0.21.0,<0.23.0" # --index-url https://download.pytorch.org/whl/cu1XX
-pip install ninja "packaging>=24.2,<26.0" # flash attention dependencies
-pip install "flash-attn>=2.5.9,<3.0.0" --no-build-isolation
-python -c "import flash_attn; print(f'Flash Attention {flash_attn.__version__} imported successfully')"
-```
-
-3. Install LeRobot by running:
-
-```bash
-pip install lerobot[groot]
-```
-
-## Usage
-
-To use GR00T in your LeRobot configuration, specify the policy type as:
-
-```python
-policy.type=groot
-```
-
-## Training
-
-### Training Command Example
-
-Here's a complete training command for finetuning the base GR00T model on your own dataset:
-
-```bash
-# Using a multi-GPU setup
-accelerate launch \
-  --multi_gpu \
-  --num_processes=$NUM_GPUS \
-  $(which lerobot-train) \
-  --output_dir=$OUTPUT_DIR \
-  --save_checkpoint=true \
-  --batch_size=$BATCH_SIZE \
-  --steps=$NUM_STEPS \
-  --save_freq=$SAVE_FREQ \
-  --log_freq=$LOG_FREQ \
-  --policy.push_to_hub=true \
-  --policy.type=groot \
-  --policy.repo_id=$REPO_ID \
-  --policy.tune_diffusion_model=false \
-  --dataset.repo_id=$DATASET_ID \
-  --wandb.enable=true \
-  --wandb.disable_artifact=true \
-  --job_name=$JOB_NAME
-```
-
-## Performance Results
-
-### Libero Benchmark Results
-
-> [!NOTE]
-> Follow our instructions for Libero usage: [Libero](./libero)
-
-GR00T has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the GR00T N1.5 model for 30k steps on the Libero dataset and compared the results to the GR00T reference results.
-
-| Benchmark          | LeRobot Implementation | GR00T Reference |
-| ------------------ | ---------------------- | --------------- |
-| **Libero Spatial** | 82.0%                  | 92.0%           |
-| **Libero Object**  | 99.0%                  | 92.0%           |
-| **Libero Long**    | 82.0%                  | 76.0%           |
-| **Average**        | 87.0%                  | 87.0%           |
-
-These results demonstrate GR00T's strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the [Libero](https://huggingface.co/docs/lerobot/libero) section.
-
-### Evaluate in your hardware setup
-
-Once you have trained your model using your parameters you can run inference in your downstream task. Follow the instructions in [Imitation Learning for Robots](./il_robots). For example:
-
-```bash
-lerobot-record \
-  --robot.type=bi_so100_follower \
-  --robot.left_arm_port=/dev/ttyACM1 \
-  --robot.right_arm_port=/dev/ttyACM0 \
-  --robot.id=bimanual_follower \
-  --robot.cameras='{ right: {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30},
-    left: {"type": "opencv", "index_or_path": 2, "width": 640, "height": 480, "fps": 30},
-    top: {"type": "opencv", "index_or_path": 4, "width": 640, "height": 480, "fps": 30},
-  }' \
-  --display_data=true \
-  --dataset.repo_id=<user>/eval_groot-bimanual  \
-  --dataset.num_episodes=10 \
-  --dataset.single_task="Grab and handover the red cube to the other arm"
-  --policy.path=<user>/groot-bimanual # your trained model
-  --dataset.episode_time_s=30
-  --dataset.reset_time_s=10
-```
-
-## License
-
-This model follows the **Apache 2.0 License**, consistent with the original [GR00T repository](https://github.com/NVIDIA/Isaac-GR00T).

docs/source/hilserl.mdx

@@ -62,7 +62,7 @@ pip install -e ".[hilserl]"
 
 ### Understanding Configuration
 
-The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
+The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/scripts/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
 
 <!-- prettier-ignore-start -->
 ```python
@@ -95,6 +95,7 @@ class HILSerlProcessorConfig:
 class ObservationConfig:
     add_joint_velocity_to_observation: bool = False    # Add joint velocities to state
     add_current_to_observation: bool = False    # Add motor currents to state
+    add_ee_pose_to_observation: bool = False    # Add end-effector pose to state
     display_cameras: bool = False    # Display camera feeds during execution
 
 class ImagePreprocessingConfig:
@@ -104,6 +105,7 @@ class ImagePreprocessingConfig:
 class GripperConfig:
     use_gripper: bool = True    # Enable gripper control
     gripper_penalty: float = 0.0    # Penalty for inappropriate gripper usage
+    gripper_penalty_in_reward: bool = False    # Include gripper penalty in reward
 
 class ResetConfig:
     fixed_reset_joint_positions: Any | None = None    # Joint positions for reset
@@ -286,6 +288,7 @@ You can enable multiple observation processing features simultaneously:
       "observation": {
         "add_joint_velocity_to_observation": true,
         "add_current_to_observation": true,
+        "add_ee_pose_to_observation": false,
         "display_cameras": false
       }
     }
@@ -301,19 +304,19 @@ Before collecting demonstrations, you need to determine the appropriate operatio
 
 This helps simplify the problem of learning on the real robot in two ways: 1) by limiting the robot's operational space to a specific region that solves the task and avoids unnecessary or unsafe exploration, and 2) by allowing training in end-effector space rather than joint space. Empirically, learning in joint space for reinforcement learning in manipulation is often a harder problem - some tasks are nearly impossible to learn in joint space but become learnable when the action space is transformed to end-effector coordinates.
 
-**Using lerobot-find-joint-limits**
+**Using find_joint_limits.py**
 
 This script helps you find the safe operational bounds for your robot's end-effector. Given that you have a follower and leader arm, you can use the script to find the bounds for the follower arm that will be applied during training.
 Bounding the action space will reduce the redundant exploration of the agent and guarantees safety.
 
 ```bash
-lerobot-find-joint-limits \
-  --robot.type=so100_follower \
-  --robot.port=/dev/tty.usbmodem58760431541 \
-  --robot.id=black \
-  --teleop.type=so100_leader \
-  --teleop.port=/dev/tty.usbmodem58760431551 \
-  --teleop.id=blue
+python -m lerobot.scripts.find_joint_limits \
+    --robot.type=so100_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.id=black \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=blue
 ```
 
 **Workflow**
@@ -515,7 +518,7 @@ During the online training, press `space` to take over the policy and `space` ag
 Start the recording process, an example of the config file can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json):
 
 ```bash
-python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config_so100.json
+python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config_so100.json
 ```
 
 During recording:
@@ -546,7 +549,7 @@ Note: If you already know the crop parameters, you can skip this step and just s
 Use the `crop_dataset_roi.py` script to interactively select regions of interest in your camera images:
 
 ```bash
-python -m lerobot.rl.crop_dataset_roi --repo-id username/pick_lift_cube
+python -m lerobot.scripts.rl.crop_dataset_roi --repo-id username/pick_lift_cube
 ```
 
 1. For each camera view, the script will display the first frame
@@ -615,7 +618,7 @@ Before training, you need to collect a dataset with labeled examples. The `recor
 To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
 
 ```bash
-python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/reward_classifier_train_config.json
+python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/reward_classifier_train_config.json
 ```
 
 **Key Parameters for Data Collection**
@@ -761,7 +764,7 @@ or set the argument in the json config file.
 Run `gym_manipulator.py` to test the model.
 
 ```bash
-python -m lerobot.rl.gym_manipulator --config_path path/to/env_config.json
+python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config.json
 ```
 
 The reward classifier will automatically provide rewards based on the visual input from the robot's cameras.
@@ -774,7 +777,7 @@ The reward classifier will automatically provide rewards based on the visual inp
 2. **Collect a dataset**:
 
    ```bash
-   python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
+   python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
    ```
 
 3. **Train the classifier**:
@@ -785,7 +788,7 @@ The reward classifier will automatically provide rewards based on the visual inp
 
 4. **Test the classifier**:
    ```bash
-   python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
+   python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
    ```
 
 ### Training with Actor-Learner
@@ -807,7 +810,7 @@ Create a training configuration file (example available [here](https://huggingfa
 First, start the learner server process:
 
 ```bash
-python -m lerobot.rl.learner --config_path src/lerobot/configs/train_config_hilserl_so100.json
+python -m lerobot.scripts.rl.learner --config_path src/lerobot/configs/train_config_hilserl_so100.json
 ```
 
 The learner:
@@ -822,7 +825,7 @@ The learner:
 In a separate terminal, start the actor process with the same configuration:
 
 ```bash
-python -m lerobot.rl.actor --config_path src/lerobot/configs/train_config_hilserl_so100.json
+python -m lerobot.scripts.rl.actor --config_path src/lerobot/configs/train_config_hilserl_so100.json
 ```
 
 The actor:

docs/source/hilserl_sim.mdx

@@ -91,7 +91,7 @@ Important parameters:
 To run the environment, set mode to null:
 
 ```bash
-python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
+python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 ```
 
 ### Recording a Dataset
@@ -118,7 +118,7 @@ To collect a dataset, set the mode to `record` whilst defining the repo_id and n
 ```
 
 ```bash
-python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
+python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 ```
 
 ### Training a Policy
@@ -126,13 +126,13 @@ python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 To train a policy, checkout the configuration example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/gym_hil/train_config.json) and run the actor and learner servers:
 
 ```bash
-python -m lerobot.rl.actor --config_path path/to/train_gym_hil_env.json
+python -m lerobot.scripts.rl.actor --config_path path/to/train_gym_hil_env.json
 ```
 
 In a different terminal, run the learner server:
 
 ```bash
-python -m lerobot.rl.learner --config_path path/to/train_gym_hil_env.json
+python -m lerobot.scripts.rl.learner --config_path path/to/train_gym_hil_env.json
 ```
 
 The simulation environment provides a safe and repeatable way to develop and test your Human-In-the-Loop reinforcement learning components before deploying to real robots.

docs/source/il_robots.mdx

@@ -165,7 +165,7 @@ huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | head -n 1)
+HF_USER=$(huggingface-cli whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -200,7 +200,7 @@ from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderCo
 from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import _init_rerun
 from lerobot.record import record_loop
 
 NUM_EPISODES = 5
@@ -237,7 +237,7 @@ dataset = LeRobotDataset.create(
 
 # Initialize the keyboard listener and rerun visualization
 _, events = init_keyboard_listener()
-init_rerun(session_name="recording")
+_init_rerun(session_name="recording")
 
 # Connect the robot and teleoperator
 robot.connect()
@@ -513,14 +513,13 @@ from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
-
+from lerobot.utils.visualization_utils import _init_rerun
+from lerobot.record import record_loop
+from lerobot.policies.factory import make_processor
 
 NUM_EPISODES = 5
 FPS = 30
@@ -558,12 +557,12 @@ dataset = LeRobotDataset.create(
 
 # Initialize the keyboard listener and rerun visualization
 _, events = init_keyboard_listener()
-init_rerun(session_name="recording")
+_init_rerun(session_name="recording")
 
 # Connect the robot
 robot.connect()
 
-preprocessor, postprocessor = make_pre_post_processors(
+preprocessor, postprocessor = make_processor(
     policy_cfg=policy,
     pretrained_path=HF_MODEL_ID,
     dataset_stats=dataset.meta.stats,

docs/source/il_sim.mdx

@@ -61,14 +61,14 @@ Then we can run this command to start:
 <hfoption id="Linux">
 
 ```bash
-python -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
+python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
 ```
 
 </hfoption>
 <hfoption id="MacOS">
 
 ```bash
-mjpython -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
+mjpython -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
 ```
 
 </hfoption>
@@ -198,14 +198,14 @@ Then you can run this command to visualize your trained policy
 <hfoption id="Linux">
 
 ```bash
-python -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
+python -m lerobot.scripts.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
 ```
 
 </hfoption>
 <hfoption id="MacOS">
 
 ```bash
-mjpython -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
+mjpython -m lerobot.scripts.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
 ```
 
 </hfoption>

docs/source/installation.mdx

@@ -1,15 +1,8 @@
 # Installation
 
-## Install [`miniforge`](https://conda-forge.org/download/)
-
-```bash
-wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
-bash Miniforge3-$(uname)-$(uname -m).sh
-```
-
 ## Environment Setup
 
-Create a virtual environment with Python 3.10, using conda:
+Create a virtual environment with Python 3.10, using [`Miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install)
 
 ```bash
 conda create -y -n lerobot python=3.10
@@ -21,7 +14,7 @@ Then activate your conda environment, you have to do this each time you open a s
 conda activate lerobot
 ```
 
-When using `conda`, install `ffmpeg` in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
 
 ```bash
 conda install ffmpeg -c conda-forge
@@ -81,9 +74,6 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
-> [!NOTE]
-> For lerobot 0.4.0, if you want to install libero or pi, you will have to do: `pip install "lerobot[pi,libero]@git+https://github.com/huggingface/lerobot.git"`
-
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
@@ -101,7 +91,7 @@ LeRobot provides optional extras for specific functionalities. Multiple extras c
 
 ### Simulations
 
-Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht))
+Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), `xarm` ([gym-xarm](https://github.com/huggingface/gym-xarm)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht))
 Example:
 
 ```bash

docs/source/integrate_hardware.mdx

@@ -8,7 +8,7 @@ To that end, we provide the [`Robot`](https://github.com/huggingface/lerobot/blo
 
 - Your own robot which exposes a communication interface (e.g. serial, CAN, TCP)
 - A way to read sensor data and send motor commands programmatically, e.g. manufacturer's SDK or API, or your own protocol implementation.
-- LeRobot installed in your environment. Follow our [Installation Guide](./installation).
+- LeRobot installed in your environment. Follow our [Installation Guide](./installation.mdx).
 
 ## Choose your motors
 
@@ -65,7 +65,7 @@ class MyCoolRobotConfig(RobotConfig):
 ```
 <!-- prettier-ignore-end -->
 
-[Cameras tutorial](./cameras) to understand how to detect and add your camera.
+[Cameras tutorial](./cameras.mdx) to understand how to detect and add your camera.
 
 Next, we'll create our actual robot class which inherits from `Robot`. This abstract class defines a contract you must follow for your robot to be usable with the rest of the LeRobot tools.
 
@@ -208,36 +208,34 @@ LeRobot supports saving and loading calibration data automatically. This is usef
 
 <!-- prettier-ignore-start -->
 ```python
-@property
-def is_calibrated(self) -> bool:
-    return True
-
-def calibrate(self) -> None:
-    pass
-```
-<!-- prettier-ignore-end -->
+> @property
+> def is_calibrated(self) -> bool:
+>    return True
+>
+> def calibrate(self) -> None:
+>    pass
+> ```
 
 ### `is_calibrated`
 
 This should reflect whether your robot has the required calibration loaded.
 
-<!-- prettier-ignore-start -->
-```python
+```
+<!-- prettier-ignore-end -->python
 @property
 def is_calibrated(self) -> bool:
     return self.bus.is_calibrated
 ```
-<!-- prettier-ignore-end -->
 
 ### `calibrate()`
 
 The goal of the calibration is twofold:
-
-- Know the physical range of motion of each motors in order to only send commands within this range.
-- Normalize raw motors positions to sensible continuous values (e.g. percentages, degrees) instead of arbitrary discrete value dependant on the specific motor used that will not replicate elsewhere.
+    - Know the physical range of motion of each motors in order to only send commands within this range.
+    - Normalize raw motors positions to sensible continuous values (e.g. percentages, degrees) instead of arbitrary discrete value dependant on the specific motor used that will not replicate elsewhere.
 
 It should implement the logic for calibration (if relevant) and update the `self.calibration` dictionary. If you are using Feetech or Dynamixel motors, our bus interfaces already include methods to help with this.
 
+
 <!-- prettier-ignore-start -->
 ```python
 def calibrate(self) -> None:
@@ -337,134 +335,6 @@ For implementing teleoperation devices, we also provide a [`Teleoperator`](https
 
 The main differences are in the I/O functions: a teleoperator allows you to produce action via `get_action` and can receive feedback actions via `send_feedback`. Feedback could be anything controllable on the teleoperation device that could help the person controlling it understand the consequences of the actions sent. Think motion/force feedback on a leader arm, vibrations on a gamepad controller for example. To implement a teleoperator, you can follow this same tutorial and adapt it for these two methods.
 
-## Using Your Own `LeRobot` Devices 🔌
-
-You can easily extend `lerobot` with your own custom hardware—be it a camera, robot, or teleoperation device—by creating a separate, installable Python package. If you follow a few simple conventions, the `lerobot` command-line tools (like `lerobot-teleop` and `lerobot-record`) will **automatically discover and integrate your creations** without requiring any changes to the `lerobot` source code.
-
-This guide outlines the conventions your plugin must follow.
-
-### The 4 Core Conventions
-
-To ensure your custom device is discoverable, you must adhere to the following four rules.
-
-#### 1\. Create an Installable Package with a Specific Prefix
-
-Your project must be a standard, installable Python package. Crucially, the name of your package (as defined in `pyproject.toml` or `setup.py`) must begin with one of these prefixes:
-
-- `lerobot_robot_` for a robot.
-- `lerobot_camera_` for a camera.
-- `lerobot_teleoperator_` for a teleoperation device.
-
-This prefix system is how `lerobot` automatically finds your plugin in the Python environment.
-
-#### 2\. Follow the `SomethingConfig`/`Something` Naming Pattern
-
-Your device's implementation class must be named after its configuration class, simply by removing the `Config` suffix.
-
-- **Config Class:** `MyAwesomeTeleopConfig`
-- **Device Class:** `MyAwesomeTeleop`
-
-#### 3\. Place Your Files in a Predictable Structure
-
-The device class (`MyAwesomeTeleop`) must be located in a predictable module relative to its configuration class (`MyAwesomeTeleopConfig`). `lerobot` will automatically search in these locations:
-
-- In the **same module** as the config class.
-- In a **submodule named after the device** (e.g., `my_awesome_teleop.py`).
-
-The recommended and simplest structure is to place them in separate, clearly named files within the same directory.
-
-#### 4\. Expose Classes in `__init__.py`
-
-Your package's `__init__.py` file should import and expose both the configuration and the device classes, making them easily accessible.
-
-### Putting It All Together: A Complete Example
-
-Let's create a new teleoperator called `my_awesome_teleop`.
-
-#### Directory Structure
-
-Here is what the project folder should look like. The package name, `lerobot_teleoperator_my_awesome_teleop`, follows **Convention \#1**.
-
-```
-lerobot_teleoperator_my_awesome_teleop/
-├── pyproject.toml # (or setup.py) lists lerobot as a dependency
-└── lerobot_teleoperator_my_awesome_teleop/
-    ├── __init__.py
-    ├── config_my_awesome_teleop.py
-    └── my_awesome_teleop.py
-```
-
-#### File Contents
-
-- **`config_my_awesome_teleop.py`**: Defines the configuration class. Note the `Config` suffix (**Convention \#2**).
-
-  ```python
-  from dataclasses import dataclass
-
-  from lerobot.teleoperators.config import TeleoperatorConfig
-
-  @TeleoperatorConfig.register_subclass("my_awesome_teleop")
-  @dataclass
-  class MyAwesomeTeleopConfig(TeleoperatorConfig):
-      # Your configuration fields go here
-      port: str = "192.168.1.1"
-  ```
-
-- **`my_awesome_teleop.py`**: Implements the device. The class name `MyAwesomeTeleop` matches its config class name (**Convention \#2**). This file structure adheres to **Convention \#3**.
-
-  ```python
-  from lerobot.teleoperators.teleoperator import Teleoperator
-
-  from .config_my_awesome_teleop import MyAwesomeTeleopConfig
-
-  class MyAwesomeTeleop(Teleoperator):
-      config_class = MyAwesomeTeleopConfig
-      name = "my_awesome_teleop"
-
-      def __init__(self, config: MyAwesomeTeleopConfig):
-          super().__init__(config)
-          self.config = config
-
-      # Your device logic (e.g., connect) goes here
-  ```
-
-- **`__init__.py`**: Exposes the key classes (**Convention \#4**).
-
-  ```python
-  from .config_my_awesome_teleop import MyAwesomeTeleopConfig
-  from .my_awesome_teleop import MyAwesomeTeleop
-  ```
-
-### Installation and Usage
-
-1.  **Install your new plugin in your Python environment.** You can install your local plugin package using `pip`'s editable mode or from PyPi.
-
-    ```bash
-    # Locally
-    # Navigate to your plugin's root directory and install it
-    cd lerobot_teleoperator_my_awesome_teleop
-    pip install -e .
-
-    # From PyPi
-    pip install lerobot_teleoperator_my_awesome_teleop
-    ```
-
-2.  **Use it directly from the command line.** Now, you can use your custom device by referencing its type.
-
-    ```bash
-    lerobot-teleoperate --teleop.type=my_awesome_teleop \
-    # other arguments
-    ```
-
-And that's it\! Your custom device is now fully integrated.
-
-### Looking for an example ?
-
-Check out these two packages from the community:
-
-- https://github.com/SpesRobotics/lerobot-robot-xarm
-- https://github.com/SpesRobotics/lerobot-teleoperator-teleop
-
 ## Wrapping Up
 
 Once your robot class is complete, you can leverage the LeRobot ecosystem:

docs/source/introduction_processors.mdx

@@ -297,9 +297,9 @@ LeRobot provides many registered processor steps. Here are the most commonly use
 
 ### Next Steps
 
-- **[Implement Your Own Processor](./implement_your_own_processor)** - Create custom processor steps
-- **[Debug Your Pipeline](./debug_processor_pipeline)** - Troubleshoot and optimize pipelines
-- **[Processors for Robots and Teleoperators](./processors_robots_teleop)** - Real-world integration patterns
+- **[Implement Your Own Processor](implement_your_own_processor.mdx)** - Create custom processor steps
+- **[Debug Your Pipeline](debug_processor_pipeline.mdx)** - Troubleshoot and optimize pipelines
+- **[Processors for Robots and Teleoperators](processors_robots_teleop.mdx)** - Real-world integration patterns
 
 ## Summary
 

docs/source/koch.mdx

@@ -277,7 +277,7 @@ leader.disconnect()
 </hfoption>
 </hfoptions>
 
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
+Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)
 
 > [!TIP]
 > If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).

docs/source/lekiwi.mdx

@@ -323,7 +323,7 @@ To replay an episode run the API example below, make sure to change `remote_ip`,
 python examples/lekiwi/replay.py
 ```
 
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by the training part of this tutorial: [Getting started with real-world robots](./il_robots)
+Congrats 🎉, your robot is all set to learn a task on its own. Start training it by the training part of this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)
 
 ## Evaluate your policy
 

docs/source/lerobot-dataset-v3.mdx

@@ -246,7 +246,7 @@ You can also use any `torchvision.transforms.v2` transform by passing it directl
 Use the visualization script to preview how transforms affect your data:
 
 ```bash
-lerobot-imgtransform-viz \
+python -m lerobot.scripts.visualize_image_transforms \
   --repo-id=your-username/your-dataset \
   --output-dir=./transform_examples \
   --n-examples=5
@@ -279,36 +279,3 @@ python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=<HF_USER/DAT
 - Aggregates parquet files: `episode-0000.parquet`, `episode-0001.parquet`, … → **`file-0000.parquet`**, …
 - Aggregates mp4 files: `episode-0000.mp4`, `episode-0001.mp4`, … → **`file-0000.mp4`**, …
 - Updates `meta/episodes/*` (chunked Parquet) with per‑episode lengths, tasks, and byte/frame offsets.
-
-## Common Issues
-
-### Always call `finalize()` before pushing
-
-When creating or recording datasets, you **must** call `dataset.finalize()` to properly close parquet writers. See the [PR #1903](https://github.com/huggingface/lerobot/pull/1903) for more details.
-
-```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-# Create dataset and record episodes
-dataset = LeRobotDataset.create(...)
-
-for episode in range(num_episodes):
-    # Record frames
-    for frame in episode_data:
-        dataset.add_frame(frame)
-    dataset.save_episode()
-
-# Call finalize() when done recording and before push_to_hub()
-dataset.finalize()  # Closes parquet writers, writes metadata footers
-dataset.push_to_hub()
-```
-
-**Why is this necessary?**
-
-Dataset v3.0 uses incremental parquet writing with buffered metadata for efficiency. The `finalize()` method:
-
-- Flushes any buffered episode metadata to disk
-- Closes parquet writers to write footer metadata, otherwise the parquet files will be corrupt
-- Ensures the dataset is valid for loading
-
-Without calling `finalize()`, your parquet files will be incomplete and the dataset won't load properly.

docs/source/libero.mdx

@@ -28,17 +28,12 @@ LIBERO is now part of our **multi-eval supported simulation**, meaning you can b
 To Install LIBERO, after following LeRobot official instructions, just do:
 `pip install -e ".[libero]"`
 
-> [!NOTE]
-> For lerobot 0.4.0, if you want to install libero tag, you will have to do: `pip install "lerobot[libero]@git+https://github.com/huggingface/lerobot.git"`.
->
-> This will be solved in the next patch release
-
 ### Single-suite evaluation
 
 Evaluate a policy on one LIBERO suite:
 
 ```bash
-lerobot-eval \
+python src/lerobot/scripts/eval.py \
   --policy.path="your-policy-id" \
   --env.type=libero \
   --env.task=libero_object \
@@ -57,7 +52,7 @@ lerobot-eval \
 Benchmark a policy across multiple suites at once:
 
 ```bash
-lerobot-eval \
+python src/lerobot/scripts/eval.py \
   --policy.path="your-policy-id" \
   --env.type=libero \
   --env.task=libero_object,libero_spatial \
@@ -108,11 +103,10 @@ For reference, here is the **original dataset** published by Physical Intelligen
 ### Example training command
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/train.py \
   --policy.type=smolvla \
   --policy.repo_id=${HF_USER}/libero-test \
-  --policy.load_vlm_weights=true \
-  --dataset.repo_id=HuggingFaceVLA/libero \
+  --dataset.repo_id=jadechoghari/smol-libero3 \
   --env.type=libero \
   --env.task=libero_10 \
   --output_dir=./outputs/ \
@@ -131,41 +125,6 @@ LeRobot uses MuJoCo for simulation. You need to set the rendering backend before
 
 - `export MUJOCO_GL=egl` → for headless servers (e.g. HPC, cloud)
 
-## Reproducing π₀.₅ results
+## Reproducing π₀ and π₀.₅ results
 
-We reproduce the results of π₀.₅ on the LIBERO benchmark using the LeRobot implementation. We take the Physical Intelligence LIBERO base model (`pi05_libero`) and finetune for an additional 6k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
-
-The finetuned model can be found here:
-
-- **π₀.₅ LIBERO**: [lerobot/pi05_libero_finetuned](https://huggingface.co/lerobot/pi05_libero_finetuned)
-
-We then evaluate the finetuned model using the LeRobot LIBERO implementation, by running the following command:
-
-```bash
-lerobot-eval \
-  --output_dir=/logs/ \
-  --env.type=libero \
-  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=10 \
-  --policy.path=pi05_libero_finetuned \
-  --policy.n_action_steps=10 \
-  --output_dir=./eval_logs/ \
-  --env.max_parallel_tasks=1
-```
-
-**Note:** We set `n_action_steps=10`, similar to the original OpenPI implementation.
-
-### Results
-
-We obtain the following results on the LIBERO benchmark:
-
-| Model    | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average  |
-| -------- | -------------- | ------------- | ----------- | --------- | -------- |
-| **π₀.₅** | 97.0           | 99.0          | 98.0        | 96.0      | **97.5** |
-
-These results are consistent with the original [results](https://github.com/Physical-Intelligence/openpi/tree/main/examples/libero#results) reported by Physical Intelligence:
-
-| Model    | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average   |
-| -------- | -------------- | ------------- | ----------- | --------- | --------- |
-| **π₀.₅** | 98.8           | 98.2          | 98.0        | 92.4      | **96.85** |
+We can also reproduce the results of π₀ and π₀.₅ on the Libero benchmark by using the finetuned libero models.

docs/source/metaworld.mdx

@@ -1,80 +0,0 @@
-# Meta-World
-
-Meta-World is a well-designed, open-source simulation benchmark for multi-task and meta reinforcement learning in continuous-control robotic manipulation. It gives researchers a shared, realistic playground to test whether algorithms can _learn many different tasks_ and _generalize quickly to new ones_ — two central challenges for real-world robotics.
-
-- 📄 [MetaWorld paper](https://arxiv.org/pdf/1910.10897)
-- 💻 [Original MetaWorld repo](https://github.com/Farama-Foundation/Metaworld)
-
-![MetaWorld MT10 demo](https://meta-world.github.io/figures/ml45.gif)
-
-## Why Meta-World matters
-
-- **Diverse, realistic tasks.** Meta-World bundles a large suite of simulated manipulation tasks (50 in the MT50 suite) using everyday objects and a common tabletop Sawyer arm. This diversity exposes algorithms to a wide variety of dynamics, contacts and goal specifications while keeping a consistent control and observation structure.
-- **Focus on generalization and multi-task learning.** By evaluating across task distributions that share structure but differ in goals and objects, Meta-World reveals whether an agent truly learns transferable skills rather than overfitting to a narrow task.
-- **Standardized evaluation protocol.** It provides clear evaluation modes and difficulty splits, so different methods can be compared fairly across easy, medium, hard and very-hard regimes.
-- **Empirical insight.** Past evaluations on Meta-World show impressive progress on some fronts, but also highlight that current multi-task and meta-RL methods still struggle with large, diverse task sets. That gap points to important research directions.
-
-## What it enables in LeRobot
-
-In LeRobot, you can evaluate any policy or vision-language-action (VLA) model on Meta-World tasks and get a clear success-rate measure. The integration is designed to be straightforward:
-
-- We provide a LeRobot-ready dataset for Meta-World (MT50) on the HF Hub: `https://huggingface.co/datasets/lerobot/metaworld_mt50`.
-  - This dataset is formatted for the MT50 evaluation that uses all 50 tasks (the most challenging multi-task setting).
-  - MT50 gives the policy a one-hot task vector and uses fixed object/goal positions for consistency.
-
-- Task descriptions and the exact keys required for evaluation are available in the repo/dataset — use these to ensure your policy outputs the right success signals.
-
-## Quick start, train a SmolVLA policy on Meta-World
-
-Example command to train a SmolVLA policy on a subset of tasks:
-
-```bash
-lerobot-train \
-  --policy.type=smolvla \
-  --policy.repo_id=${HF_USER}/metaworld-test \
-  --policy.load_vlm_weights=true \
-  --dataset.repo_id=lerobot/metaworld_mt50 \
-  --env.type=metaworld \
-  --env.task=assembly-v3,dial-turn-v3,handle-press-side-v3 \
-  --output_dir=./outputs/ \
-  --steps=100000 \
-  --batch_size=4 \
-  --eval.batch_size=1 \
-  --eval.n_episodes=1 \
-  --eval_freq=1000
-```
-
-Notes:
-
-- `--env.task` accepts explicit task lists (comma separated) or difficulty groups (e.g., `env.task="hard"`).
-- Adjust `batch_size`, `steps`, and `eval_freq` to match your compute budget.
-- **Gymnasium Assertion Error**: if you encounter an error like
-  `AssertionError: ['human', 'rgb_array', 'depth_array']` when running MetaWorld environments, this comes from a mismatch between MetaWorld and your Gymnasium version.
-  We recommend using:
-
-```bash
-  pip install "gymnasium==1.1.0"
-```
-
-to ensure proper compatibility.
-
-## Quick start — evaluate a trained policy
-
-To evaluate a trained policy on the Meta-World medium difficulty split:
-
-```bash
-lerobot-eval \
-  --policy.path="your-policy-id" \
-  --env.type=metaworld \
-  --env.task=medium \
-  --eval.batch_size=1 \
-  --eval.n_episodes=2
-```
-
-This will run episodes and return per-task success rates using the standard Meta-World evaluation keys.
-
-## Practical tips
-
-- If you care about generalization, run on the full MT50 suite — it’s intentionally challenging and reveals strengths/weaknesses better than a few narrow tasks.
-- Use the one-hot task conditioning for multi-task training (MT10 / MT50 conventions) so policies have explicit task context.
-- Inspect the dataset task descriptions and the `info["is_success"]` keys when writing post-processing or logging so your success metrics line up with the benchmark.

docs/source/multi_gpu_training.mdx

@@ -1,125 +0,0 @@
-# Multi-GPU Training
-
-This guide shows you how to train policies on multiple GPUs using [Hugging Face Accelerate](https://huggingface.co/docs/accelerate).
-
-## Installation
-
-First, ensure you have accelerate installed:
-
-```bash
-pip install accelerate
-```
-
-## Training with Multiple GPUs
-
-You can launch training in two ways:
-
-### Option 1: Without config (specify parameters directly)
-
-You can specify all parameters directly in the command without running `accelerate config`:
-
-```bash
-accelerate launch \
-  --multi_gpu \
-  --num_processes=2 \
-  $(which lerobot-train) \
-  --dataset.repo_id=${HF_USER}/my_dataset \
-  --policy.type=act \
-  --policy.repo_id=${HF_USER}/my_trained_policy \
-  --output_dir=outputs/train/act_multi_gpu \
-  --job_name=act_multi_gpu \
-  --wandb.enable=true
-```
-
-**Key accelerate parameters:**
-
-- `--multi_gpu`: Enable multi-GPU training
-- `--num_processes=2`: Number of GPUs to use
-- `--mixed_precision=fp16`: Use fp16 mixed precision (or `bf16` if supported)
-
-### Option 2: Using accelerate config
-
-If you prefer to save your configuration, you can optionally configure accelerate for your hardware setup by running:
-
-```bash
-accelerate config
-```
-
-This interactive setup will ask you questions about your training environment (number of GPUs, mixed precision settings, etc.) and saves the configuration for future use. For a simple multi-GPU setup on a single machine, you can use these recommended settings:
-
-- Compute environment: This machine
-- Number of machines: 1
-- Number of processes: (number of GPUs you want to use)
-- GPU ids to use: (leave empty to use all)
-- Mixed precision: fp16 or bf16 (recommended for faster training)
-
-Then launch training with:
-
-```bash
-accelerate launch $(which lerobot-train) \
-  --dataset.repo_id=${HF_USER}/my_dataset \
-  --policy.type=act \
-  --policy.repo_id=${HF_USER}/my_trained_policy \
-  --output_dir=outputs/train/act_multi_gpu \
-  --job_name=act_multi_gpu \
-  --wandb.enable=true
-```
-
-## How It Works
-
-When you launch training with accelerate:
-
-1. **Automatic detection**: LeRobot automatically detects if it's running under accelerate
-2. **Data distribution**: Your batch is automatically split across GPUs
-3. **Gradient synchronization**: Gradients are synchronized across GPUs during backpropagation
-4. **Single process logging**: Only the main process logs to wandb and saves checkpoints
-
-## Learning Rate and Training Steps Scaling
-
-**Important:** LeRobot does **NOT** automatically scale learning rates or training steps based on the number of GPUs. This gives you full control over your training hyperparameters.
-
-### Why No Automatic Scaling?
-
-Many distributed training frameworks automatically scale the learning rate by the number of GPUs (e.g., `lr = base_lr × num_gpus`).
-However, LeRobot keeps the learning rate exactly as you specify it.
-
-### When and How to Scale
-
-If you want to scale your hyperparameters when using multiple GPUs, you should do it manually:
-
-**Learning Rate Scaling:**
-
-```bash
-# Example: 2 GPUs with linear LR scaling
-# Base LR: 1e-4, with 2 GPUs -> 2e-4
-accelerate launch --num_processes=2 $(which lerobot-train) \
-  --optimizer.lr=2e-4 \
-  --dataset.repo_id=lerobot/pusht \
-  --policy=act
-```
-
-**Training Steps Scaling:**
-
-Since the effective batch size `bs` increases with multiple GPUs (batch_size × num_gpus), you may want to reduce the number of training steps proportionally:
-
-```bash
-# Example: 2 GPUs with effective batch size 2x larger
-# Original: batch_size=8, steps=100000
-# With 2 GPUs: batch_size=8 (16 in total), steps=50000
-accelerate launch --num_processes=2 $(which lerobot-train) \
-  --batch_size=8 \
-  --steps=50000 \
-  --dataset.repo_id=lerobot/pusht \
-  --policy=act
-```
-
-## Notes
-
-- The `--policy.use_amp` flag in `lerobot-train` is only used when **not** running with accelerate. When using accelerate, mixed precision is controlled by accelerate's configuration.
-- Training logs, checkpoints, and hub uploads are only done by the main process to avoid conflicts. Non-main processes have console logging disabled to prevent duplicate output.
-- The effective batch size is `batch_size × num_gpus`. If you use 4 GPUs with `--batch_size=8`, your effective batch size is 32.
-- Learning rate scheduling is handled correctly across multiple processes—LeRobot sets `step_scheduler_with_optimizer=False` to prevent accelerate from adjusting scheduler steps based on the number of processes.
-- When saving or pushing models, LeRobot automatically unwraps the model from accelerate's distributed wrapper to ensure compatibility.
-- WandB integration automatically initializes only on the main process, preventing multiple runs from being created.
-
-For more advanced configurations and troubleshooting, see the [Accelerate documentation](https://huggingface.co/docs/accelerate). If you want to learn more about how to train on a large number of GPUs, checkout this awesome guide: [Ultrascale Playbook](https://huggingface.co/spaces/nanotron/ultrascale-playbook).

docs/source/phone_teleop.mdx

@@ -79,7 +79,7 @@ After running the example:
 - Android: after starting the script, open the printed local URL on your phone, tap Start, then press and hold Move.
 - iOS: open HEBI Mobile I/O first; B1 enables motion. A3 controls the gripper.
 
-Additionally you can customize mapping or safety limits by editing the processor steps shown in the examples. You can also remap inputs (e.g., use a different analog input) or adapt the pipeline to other robots (e.g., LeKiwi) by modifying the input and kinematics steps. More about this in the [Processors for Robots and Teleoperators](./processors_robots_teleop) guide.
+Additionally you can customize mapping or safety limits by editing the processor steps shown in the examples. You can also remap inputs (e.g., use a different analog input) or adapt the pipeline to other robots (e.g., LeKiwi) by modifying the input and kinematics steps. More about this in the [Processors for Robots and Teleoperators](./processors_robots_teleop.mdx) guide.
 
 - Run this example to record a dataset, which saves absolute end effector observations and actions:
 
@@ -136,12 +136,13 @@ Additionally you can customize mapping or safety limits by editing the processor
   ),
   ```
 
-- The `EEBoundsAndSafety` step clamps EE motion to a workspace and checks for large ee step jumps to ensure safety. The `end_effector_bounds` are the bounds for the EE pose and can be modified to change the workspace. The `max_ee_step_m` are the step limits for the EE pose and can be modified to change the safety limits.
+- The `EEBoundsAndSafety` step clamps EE motion to a workspace and checks for large ee step jumps to ensure safety. The `end_effector_bounds` are the bounds for the EE pose and can be modified to change the workspace. The `max_ee_step_m` and `max_ee_twist_step_rad` are the step limits for the EE pose and can be modified to change the safety limits.
 
   ```examples/phone_to_so100/teleoperate.py
   EEBoundsAndSafety(
       end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
       max_ee_step_m=0.10,
+      max_ee_twist_step_rad=0.50,
   )
   ```
 

docs/source/pi0.mdx

@@ -4,7 +4,7 @@
 
 ## Model Overview
 
-π₀ represents a breakthrough in robotics as the first general-purpose robot foundation model developed by [Physical Intelligence](https://www.physicalintelligence.company/blog/pi0). Unlike traditional robot programs that are narrow specialists programmed for repetitive motions, π₀ is designed to be a generalist policy that can understand visual inputs, interpret natural language instructions, and control a variety of different robots across diverse tasks.
+π₀ represents a breakthrough in robotics as the first general-purpose robot foundation model developed by [Physical Intelligence](https://www.physicalintelligence.company/blog/pi0). Unlike traditional robots that are narrow specialists programmed for repetitive motions, π₀ is designed to be a generalist policy that can understand visual inputs, interpret natural language instructions, and control a variety of different robots across diverse tasks.
 
 ### The Vision for Physical Intelligence
 
@@ -21,17 +21,43 @@ As described by Physical Intelligence, while AI has achieved remarkable success
 
 ## Installation Requirements
 
-1. Install LeRobot by following our [Installation Guide](./installation).
-2. Install Pi0 dependencies by running:
+⚠️ **Warning**: This policy requires patching the Hugging Face `transformers` library.
+
+### Prerequisites
+
+1. Ensure you have the exact version installed:
 
    ```bash
-   pip install -e ".[pi]"
+   pip show transformers
    ```
 
-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
-   >
-   > This will be solved in the next patch release
+   It must be version **4.53.2**.
+
+2. Apply the custom patches:
+   ```bash
+   cp -r ./src/lerobot/policies/pi0/transformers_replace/* \
+     $(python -c "import transformers, os; print(os.path.dirname(transformers.__file__))")
+   ```
+
+### What the patches do:
+
+- Support the **AdaRMS optimizer**
+- Correctly control the precision of activations
+- Allow the KV cache to be used without updates
+
+**Important Notes:**
+
+- This permanently modifies your `transformers` installation
+- The changes survive reinstalls unless you explicitly remove the patched files or recreate the environment
+
+### Restoring Clean State
+
+To undo the patches and restore a clean state:
+
+```bash
+pip uninstall transformers
+pip install transformers==4.53.2
+```
 
 ## Training Data and Capabilities
 
@@ -54,17 +80,18 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py \
+python src/lerobot/scripts/train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=pi0 \
     --output_dir=./outputs/pi0_training \
     --job_name=pi0_training \
-    --policy.pretrained_path=lerobot/pi0_base \
+    --policy.pretrained_path=pepijn223/pi0_base_fp32 \
     --policy.repo_id=your_repo_id \
     --policy.compile_model=true \
     --policy.gradient_checkpointing=true \
     --policy.dtype=bfloat16 \
     --steps=3000 \
+    --policy.scheduler_decay_steps=3000 \
     --policy.device=cuda \
     --batch_size=32
 ```
@@ -75,9 +102,10 @@ python src/lerobot/scripts/lerobot_train.py \
 - **`--policy.gradient_checkpointing=true`**: Reduces memory usage significantly during training
 - **`--policy.dtype=bfloat16`**: Use mixed precision training for efficiency
 - **`--batch_size=32`**: Batch size for training, adapt this based on your GPU memory
-- **`--policy.pretrained_path=lerobot/pi0_base`**: The base π₀ model you want to finetune, options are:
-  - [lerobot/pi0_base](https://huggingface.co/lerobot/pi0_base)
-  - [lerobot/pi0_libero](https://huggingface.co/lerobot/pi0_libero) (specifically trained on the Libero dataset)
+- **`--policy.pretrained_path=pepijn223/pi0_base_fp32`**: The base π₀ model you want to finetune, options are:
+  - [pepijn223/pi0_base_fp32](https://huggingface.co/pepijn223/pi0_base_fp32)
+  - [pepijn223/pi0_libero_fp32](https://huggingface.co/pepijn223/pi0_libero_fp32) (specifically trained on the Libero dataset)
+  - [pepijn223/pi0_droid_fp32](https://huggingface.co/pepijn223/pi0_droid_fp32) (specifically trained on the Droid dataset)
 
 ## License
 

docs/source/pi05.mdx

@@ -29,17 +29,43 @@ This diverse training mixture creates a "curriculum" that enables generalization
 
 ## Installation Requirements
 
-1. Install LeRobot by following our [Installation Guide](./installation).
-2. Install Pi0.5 dependencies by running:
+⚠️ **Warning**: This policy requires patching the Hugging Face `transformers` library.
+
+### Prerequisites
+
+1. Ensure you have the exact version installed:
 
    ```bash
-   pip install -e ".[pi]"
+   pip show transformers
    ```
 
-   > [!NOTE]
-   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
-   >
-   > This will be solved in the next patch release
+   It must be version **4.53.2**.
+
+2. Apply the custom patches:
+   ```bash
+   cp -r ./src/lerobot/policies/pi05/transformers_replace/* \
+     $(python -c "import transformers, os; print(os.path.dirname(transformers.__file__))")
+   ```
+
+### What the patches do:
+
+- Support the **AdaRMS optimizer**
+- Correctly control the precision of activations
+- Allow the KV cache to be used without updates
+
+**Important Notes:**
+
+- This permanently modifies your `transformers` installation
+- The changes survive reinstalls unless you explicitly remove the patched files or recreate the environment
+
+### Restoring Clean State
+
+To undo the patches and restore a clean state:
+
+```bash
+pip uninstall transformers
+pip install transformers==4.53.2
+```
 
 ## Usage
 
@@ -56,18 +82,19 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
 
 ```bash
-python src/lerobot/scripts/lerobot_train.py\
+python src/lerobot/scripts/train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=pi05 \
-    --output_dir=./outputs/pi05_training \
-    --job_name=pi05_training \
-    --policy.repo_id=your_repo_id \
-    --policy.pretrained_path=lerobot/pi05_base \
+    --output_dir=./outputs/pi0_training \
+    --job_name=pi0_training \
+    --policy.repo_id=pepijn223/pi05_base_fp32 \
+    --policy.pretrained_path=your_repo_id \
     --policy.compile_model=true \
     --policy.gradient_checkpointing=true \
     --wandb.enable=true \
     --policy.dtype=bfloat16 \
     --steps=3000 \
+    --policy.scheduler_decay_steps=3000 \
     --policy.device=cuda \
     --batch_size=32
 ```
@@ -78,18 +105,10 @@ python src/lerobot/scripts/lerobot_train.py\
 - **`--policy.gradient_checkpointing=true`**: Reduces memory usage significantly during training
 - **`--policy.dtype=bfloat16`**: Use mixed precision training for efficiency
 - **`--batch_size=32`**: Batch size for training, adapt this based on your GPU memory
-- **`--policy.pretrained_path=lerobot/pi05_base`**: The base π₀.₅ model you want to finetune, options are:
-  - [lerobot/pi05_base](https://huggingface.co/lerobot/pi05_base)
-  - [lerobot/pi05_libero](https://huggingface.co/lerobot/pi05_libero) (specifically trained on the Libero dataset)
-
-If your dataset is not converted with `quantiles`, you can convert it with the following command:
-
-```bash
-python src/lerobot/datasets/v30/augment_dataset_quantile_stats.py \
-    --repo-id=your_dataset \
-```
-
-Or train pi05 with this normalization mapping: `--policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'`
+- **`--policy.pretrained_path=pepijn223/pi05_base_fp32`**: The base π₀.₅ model you want to finetune, options are:
+  - [pepijn223/pi05_base_fp32](https://huggingface.co/pepijn223/pi05_base_fp32)
+  - [pepijn223/pi05_libero_fp32](https://huggingface.co/pepijn223/pi05_libero_fp32) (specifically trained on the Libero dataset)
+  - [pepijn223/pi05_droid_fp32](https://huggingface.co/pepijn223/pi05_droid_fp32) (specifically trained on the Droid dataset)
 
 ## Performance Results
 
@@ -99,13 +118,13 @@ Or train pi05 with this normalization mapping: `--policy.normalization_mapping='
 
 | Benchmark          | LeRobot Implementation | OpenPI Reference |
 | ------------------ | ---------------------- | ---------------- |
-| **Libero Spatial** | 97.0%                  | 98.8%            |
+| **Libero Spatial** | 98.0%                  | 98.8%            |
 | **Libero Object**  | 99.0%                  | 98.2%            |
-| **Libero Goal**    | 98.0%                  | 98.0%            |
-| **Libero 10**      | 96.0%                  | 92.4%            |
-| **Average**        | 97.5%                  | 96.85%           |
+| **Libero Goal**    | 97.0%                  | 98.0%            |
+| **Libero 10**      | 93.0%                  | 92.4%            |
+| **Average**        | 96.75%                 | 96.85%           |
 
-These results demonstrate π₀.₅'s strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the [Libero](https://huggingface.co/docs/lerobot/libero) section.
+These results demonstrate π₀.₅'s strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the [Libero](#libero) section.
 
 ## License
 

docs/source/policy_groot_README.md

@@ -1,27 +0,0 @@
-## Research Paper
-
-Paper: https://research.nvidia.com/labs/gear/gr00t-n1_5/
-
-## Repository
-
-Code: https://github.com/NVIDIA/Isaac-GR00T
-
-## Citation
-
-```bibtex
-@inproceedings{gr00tn1_2025,
-  archivePrefix = {arxiv},
-  eprint     = {2503.14734},
-  title      = {{GR00T} {N1}: An Open Foundation Model for Generalist Humanoid Robots},
-  author     = {NVIDIA and Johan Bjorck andFernando Castañeda, Nikita Cherniadev and Xingye Da and Runyu Ding and Linxi "Jim" Fan and Yu Fang and Dieter Fox and Fengyuan Hu and Spencer Huang and Joel Jang and Zhenyu Jiang and Jan Kautz and Kaushil Kundalia and Lawrence Lao and Zhiqi Li and Zongyu Lin and Kevin Lin and Guilin Liu and Edith Llontop and Loic Magne and Ajay Mandlekar and Avnish Narayan and Soroush Nasiriany and Scott Reed and You Liang Tan and Guanzhi Wang and Zu Wang and Jing Wang and Qi Wang and Jiannan Xiang and Yuqi Xie and Yinzhen Xu and Zhenjia Xu and Seonghyeon Ye and Zhiding Yu and Ao Zhang and Hao Zhang and Yizhou Zhao and Ruijie Zheng and Yuke Zhu},
-  month      = {March},
-  year       = {2025},
-  booktitle  = {ArXiv Preprint},
-}
-```
-
-## Additional Resources
-
-Blog: https://developer.nvidia.com/isaac/gr00t
-
-Hugging Face Model: https://huggingface.co/nvidia/GR00T-N1.5-3B

docs/source/processors_robots_teleop.mdx

@@ -38,7 +38,7 @@ phone_to_robot_ee_pose_processor = RobotProcessorPipeline[RobotAction, RobotActi
             kinematics=kinematics_solver, end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5}, motor_names=list(robot.bus.motors.keys()),
         ),
         EEBoundsAndSafety(
-            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]}, max_ee_step_m=0.20,
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]}, max_ee_step_m=0.20, max_ee_twist_step_rad=0.50,
         ),
         GripperVelocityToJoint(),
     ],

docs/source/smolvla.mdx

@@ -29,7 +29,7 @@ SmolVLA is Hugging Face’s lightweight foundation model for robotics. Designed
 ## Collect a dataset
 
 SmolVLA is a base model, so fine-tuning on your own data is required for optimal performance in your setup.
-We recommend recording ~50 episodes of your task as a starting point. Follow our guide to get started: [Recording a Dataset](./il_robots)
+We recommend recording ~50 episodes of your task as a starting point. Follow our guide to get started: [Recording a Dataset](https://huggingface.co/docs/lerobot/getting_started_real_world_robot#record-a-dataset)
 
 <Tip>
 
@@ -93,7 +93,7 @@ lerobot-train --help
 
 ## Evaluate the finetuned model and run it in real-time
 
-Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./il_robots).
+Similarly for when recording an episode, it is recommended that you are logged in to the HuggingFace Hub. You can follow the corresponding steps: [Record a dataset](./getting_started_real_world_robot#record-a-dataset).
 Once you are logged in, you can run inference in your setup by doing:
 
 ```bash

docs/source/so100.mdx

@@ -634,7 +634,7 @@ leader.disconnect()
 </hfoption>
 </hfoptions>
 
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
+Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)
 
 > [!TIP]
 > If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).

docs/source/so101.mdx

@@ -430,7 +430,7 @@ leader.disconnect()
 </hfoption>
 </hfoptions>
 
-Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./il_robots)
+Congrats 🎉, your robot is all set to learn a task on its own. Start training it by following this tutorial: [Getting started with real-world robots](./getting_started_real_world_robot)
 
 > [!TIP]
 > If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).

docs/source/using_dataset_tools.mdx

@@ -1,102 +0,0 @@
-# Using Dataset Tools
-
-This guide covers the dataset tools utilities available in LeRobot for modifying and editing existing datasets.
-
-## Overview
-
-LeRobot provides several utilities for manipulating datasets:
-
-1. **Delete Episodes** - Remove specific episodes from a dataset
-2. **Split Dataset** - Divide a dataset into multiple smaller datasets
-3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
-4. **Add Features** - Add new features to a dataset
-5. **Remove Features** - Remove features from a dataset
-
-The core implementation is in `lerobot.datasets.dataset_tools`.
-An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
-
-## Command-Line Tool: lerobot-edit-dataset
-
-`lerobot-edit-dataset` is a command-line script for editing datasets. It can be used to delete episodes, split datasets, merge datasets, add features, and remove features.
-
-Run `lerobot-edit-dataset --help` for more information on the configuration of each operation.
-
-### Usage Examples
-
-#### Delete Episodes
-
-Remove specific episodes from a dataset. This is useful for filtering out undesired data.
-
-```bash
-# Delete episodes 0, 2, and 5 (modifies original dataset)
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --operation.type delete_episodes \
-    --operation.episode_indices "[0, 2, 5]"
-
-# Delete episodes and save to a new dataset (preserves original dataset)
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --new_repo_id lerobot/pusht_after_deletion \
-    --operation.type delete_episodes \
-    --operation.episode_indices "[0, 2, 5]"
-```
-
-#### Split Dataset
-
-Divide a dataset into multiple subsets.
-
-```bash
-# Split by fractions (e.g. 80% train, 20% test, 20% val)
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --operation.type split \
-    --operation.splits '{"train": 0.8, "test": 0.2, "val": 0.2}'
-
-# Split by specific episode indices
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --operation.type split \
-    --operation.splits '{"task1": [0, 1, 2, 3], "task2": [4, 5]}'
-```
-
-There are no constraints on the split names, they can be determined by the user. Resulting datasets are saved under the repo id with the split name appended, e.g. `lerobot/pusht_train`, `lerobot/pusht_task1`, `lerobot/pusht_task2`.
-
-#### Merge Datasets
-
-Combine multiple datasets into a single dataset.
-
-```bash
-# Merge train and validation splits back into one dataset
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_merged \
-    --operation.type merge \
-    --operation.repo_ids "['lerobot/pusht_train', 'lerobot/pusht_val']"
-```
-
-#### Remove Features
-
-Remove features from a dataset.
-
-```bash
-# Remove a camera feature
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --operation.type remove_feature \
-    --operation.feature_names "['observation.images.top']"
-```
-
-### Push to Hub
-
-Add the `--push_to_hub` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:
-
-```bash
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht \
-    --new_repo_id lerobot/pusht_after_deletion \
-    --operation.type delete_episodes \
-    --operation.episode_indices "[0, 2, 5]" \
-    --push_to_hub
-```
-
-There is also a tool for adding features to a dataset that is not yet covered in `lerobot-edit-dataset`.

examples/backward_compatibility/replay.py

@@ -44,7 +44,6 @@ from lerobot.robots import (  # noqa: F401
     so100_follower,
     so101_follower,
 )
-from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import busy_wait
 from lerobot.utils.utils import (
     init_logging,
@@ -79,16 +78,16 @@ def replay(cfg: ReplayConfig):
 
     robot = make_robot_from_config(cfg.robot)
     dataset = LeRobotDataset(cfg.dataset.repo_id, root=cfg.dataset.root, episodes=[cfg.dataset.episode])
-    actions = dataset.hf_dataset.select_columns(ACTION)
+    actions = dataset.hf_dataset.select_columns("action")
     robot.connect()
 
     log_say("Replaying episode", cfg.play_sounds, blocking=True)
     for idx in range(dataset.num_frames):
         start_episode_t = time.perf_counter()
 
-        action_array = actions[idx][ACTION]
+        action_array = actions[idx]["action"]
         action = {}
-        for i, name in enumerate(dataset.features[ACTION]["names"]):
+        for i, name in enumerate(dataset.features["action"]["names"]):
             key = f"{name.removeprefix('main_')}.pos"
             action[key] = action_array[i].item()
 

examples/dataset/load_lerobot_dataset.py

@@ -132,15 +132,17 @@ 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)
 
-if __name__ == "__main__":
-    dataloader = torch.utils.data.DataLoader(
-        dataset,
-        num_workers=4,
-        batch_size=32,
-        shuffle=True,
-    )
-    for batch in dataloader:
-        print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
-        print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
-        print(f"{batch['action'].shape=}")  # (32, 64, c)
-        break
+# Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
+# PyTorch datasets.
+dataloader = torch.utils.data.DataLoader(
+    dataset,
+    num_workers=4,
+    batch_size=32,
+    shuffle=True,
+)
+
+for batch in dataloader:
+    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
+    print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
+    print(f"{batch['action'].shape=}")  # (32, 64, c)
+    break

examples/dataset/use_dataset_tools.py

@@ -1,124 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 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.
-
-"""
-Example script demonstrating dataset tools utilities.
-
-This script shows how to:
-1. Delete episodes from a dataset
-2. Split a dataset into train/val sets
-3. Add/remove features
-4. Merge datasets
-
-Usage:
-    python examples/dataset/use_dataset_tools.py
-"""
-
-import numpy as np
-
-from lerobot.datasets.dataset_tools import (
-    add_features,
-    delete_episodes,
-    merge_datasets,
-    modify_features,
-    remove_feature,
-    split_dataset,
-)
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-
-def main():
-    dataset = LeRobotDataset("lerobot/pusht")
-
-    print(f"Original dataset: {dataset.meta.total_episodes} episodes, {dataset.meta.total_frames} frames")
-    print(f"Features: {list(dataset.meta.features.keys())}")
-
-    print("\n1. Deleting episodes 0 and 2...")
-    filtered_dataset = delete_episodes(dataset, episode_indices=[0, 2], repo_id="lerobot/pusht_filtered")
-    print(f"Filtered dataset: {filtered_dataset.meta.total_episodes} episodes")
-
-    print("\n2. Splitting dataset into train/val...")
-    splits = split_dataset(
-        dataset,
-        splits={"train": 0.8, "val": 0.2},
-    )
-    print(f"Train split: {splits['train'].meta.total_episodes} episodes")
-    print(f"Val split: {splits['val'].meta.total_episodes} episodes")
-
-    print("\n3. Adding features...")
-
-    reward_values = np.random.randn(dataset.meta.total_frames).astype(np.float32)
-
-    def compute_success(row_dict, episode_index, frame_index):
-        episode_length = 10
-        return float(frame_index >= episode_length - 10)
-
-    dataset_with_features = add_features(
-        dataset,
-        features={
-            "reward": (
-                reward_values,
-                {"dtype": "float32", "shape": (1,), "names": None},
-            ),
-            "success": (
-                compute_success,
-                {"dtype": "float32", "shape": (1,), "names": None},
-            ),
-        },
-        repo_id="lerobot/pusht_with_features",
-    )
-
-    print(f"New features: {list(dataset_with_features.meta.features.keys())}")
-
-    print("\n4. Removing the success feature...")
-    dataset_cleaned = remove_feature(
-        dataset_with_features, feature_names="success", repo_id="lerobot/pusht_cleaned"
-    )
-    print(f"Features after removal: {list(dataset_cleaned.meta.features.keys())}")
-
-    print("\n5. Using modify_features to add and remove features simultaneously...")
-    dataset_modified = modify_features(
-        dataset_with_features,
-        add_features={
-            "discount": (
-                np.ones(dataset.meta.total_frames, dtype=np.float32) * 0.99,
-                {"dtype": "float32", "shape": (1,), "names": None},
-            ),
-        },
-        remove_features="reward",
-        repo_id="lerobot/pusht_modified",
-    )
-    print(f"Modified features: {list(dataset_modified.meta.features.keys())}")
-
-    print("\n6. Merging train and val splits back together...")
-    merged = merge_datasets([splits["train"], splits["val"]], output_repo_id="lerobot/pusht_merged")
-    print(f"Merged dataset: {merged.meta.total_episodes} episodes")
-
-    print("\n7. Complex workflow example...")
-
-    if len(dataset.meta.camera_keys) > 1:
-        camera_to_remove = dataset.meta.camera_keys[0]
-        print(f"Removing camera: {camera_to_remove}")
-        dataset_no_cam = remove_feature(
-            dataset, feature_names=camera_to_remove, repo_id="pusht_no_first_camera"
-        )
-        print(f"Remaining cameras: {dataset_no_cam.meta.camera_keys}")
-
-    print("\nDone! Check ~/.cache/huggingface/lerobot/ for the created datasets.")
-
-
-if __name__ == "__main__":
-    main()

examples/lekiwi/evaluate.py

@@ -19,12 +19,11 @@ from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
 from lerobot.processor import make_default_processors
+from lerobot.record import record_loop
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
-from lerobot.scripts.lerobot_record import record_loop
-from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import _init_rerun
 
 NUM_EPISODES = 2
 FPS = 30
@@ -42,8 +41,8 @@ robot = LeKiwiClient(robot_config)
 policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
 # Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, ACTION)
-obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
+action_features = hw_to_dataset_features(robot.action_features, "action")
+obs_features = hw_to_dataset_features(robot.observation_features, "observation")
 dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
@@ -74,7 +73,7 @@ teleop_action_processor, robot_action_processor, robot_observation_processor = m
 
 # Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
-init_rerun(session_name="lekiwi_evaluate")
+_init_rerun(session_name="lekiwi_evaluate")
 
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
@@ -133,6 +132,4 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
 log_say("Stop recording")
 robot.disconnect()
 listener.stop()
-
-dataset.finalize()
 dataset.push_to_hub()

examples/lekiwi/record.py

@@ -17,15 +17,14 @@
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.processor import make_default_processors
+from lerobot.record import record_loop
 from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
 from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
-from lerobot.utils.constants import ACTION, OBS_STR
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import _init_rerun
 
 NUM_EPISODES = 2
 FPS = 30
@@ -48,8 +47,8 @@ keyboard = KeyboardTeleop(keyboard_config)
 teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
 
 # Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, ACTION)
-obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
+action_features = hw_to_dataset_features(robot.action_features, "action")
+obs_features = hw_to_dataset_features(robot.observation_features, "observation")
 dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
@@ -70,7 +69,7 @@ keyboard.connect()
 
 # Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
-init_rerun(session_name="lekiwi_record")
+_init_rerun(session_name="lekiwi_record")
 
 if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
     raise ValueError("Robot or teleop is not connected!")
@@ -130,6 +129,4 @@ robot.disconnect()
 leader_arm.disconnect()
 keyboard.disconnect()
 listener.stop()
-
-dataset.finalize()
 dataset.push_to_hub()

examples/lekiwi/replay.py

@@ -19,7 +19,6 @@ import time
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
 from lerobot.robots.lekiwi.lekiwi_client import LeKiwiClient
-from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import busy_wait
 from lerobot.utils.utils import log_say
 
@@ -35,7 +34,7 @@ robot = LeKiwiClient(robot_config)
 dataset = LeRobotDataset("<hf_username>/<dataset_repo_id>", episodes=[EPISODE_IDX])
 # Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
 episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
-actions = episode_frames.select_columns(ACTION)
+actions = episode_frames.select_columns("action")
 
 # Connect to the robot
 robot.connect()
@@ -50,7 +49,7 @@ for idx in range(len(episode_frames)):
 
     # Get recorded action from dataset
     action = {
-        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        name: float(actions[idx]["action"][i]) for i, name in enumerate(dataset.features["action"]["names"])
     }
 
     # Send action to robot

examples/lekiwi/teleoperate.py

@@ -20,7 +20,7 @@ from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
 from lerobot.utils.robot_utils import busy_wait
-from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+from lerobot.utils.visualization_utils import _init_rerun, log_rerun_data
 
 FPS = 30
 
@@ -41,7 +41,7 @@ leader_arm.connect()
 keyboard.connect()
 
 # Init rerun viewer
-init_rerun(session_name="lekiwi_teleop")
+_init_rerun(session_name="lekiwi_teleop")
 
 if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
     raise ValueError("Robot or teleop is not connected!")

examples/phone_to_so100/evaluate.py

@@ -34,16 +34,16 @@ from lerobot.processor.converters import (
     transition_to_observation,
     transition_to_robot_action,
 )
+from lerobot.record import record_loop
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import _init_rerun
 
 NUM_EPISODES = 5
 FPS = 30
@@ -137,7 +137,7 @@ robot.connect()
 
 # Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
-init_rerun(session_name="phone_so100_evaluate")
+_init_rerun(session_name="phone_so100_evaluate")
 
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
@@ -194,6 +194,4 @@ for episode_idx in range(NUM_EPISODES):
 log_say("Stop recording")
 robot.disconnect()
 listener.stop()
-
-dataset.finalize()
 dataset.push_to_hub()

examples/phone_to_so100/record.py

@@ -26,6 +26,7 @@ from lerobot.processor.converters import (
     transition_to_observation,
     transition_to_robot_action,
 )
+from lerobot.record import record_loop
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.robot_kinematic_processor import (
     EEBoundsAndSafety,
@@ -35,13 +36,12 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import _init_rerun
 
 NUM_EPISODES = 2
 FPS = 30
@@ -84,6 +84,7 @@ phone_to_robot_ee_pose_processor = RobotProcessorPipeline[tuple[RobotAction, Rob
         EEBoundsAndSafety(
             end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
             max_ee_step_m=0.20,
+            max_ee_twist_step_rad=0.50,
         ),
         GripperVelocityToJoint(speed_factor=20.0),
     ],
@@ -142,7 +143,7 @@ phone.connect()
 
 # Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
-init_rerun(session_name="phone_so100_record")
+_init_rerun(session_name="phone_so100_record")
 
 if not robot.is_connected or not phone.is_connected:
     raise ValueError("Robot or teleop is not connected!")
@@ -200,6 +201,4 @@ log_say("Stop recording")
 robot.disconnect()
 phone.disconnect()
 listener.stop()
-
-dataset.finalize()
 dataset.push_to_hub()

examples/phone_to_so100/replay.py

@@ -28,7 +28,6 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import busy_wait
 from lerobot.utils.utils import log_say
 
@@ -67,7 +66,7 @@ robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotOb
 dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
 # Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
 episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
-actions = episode_frames.select_columns(ACTION)
+actions = episode_frames.select_columns("action")
 
 # Connect to the robot
 robot.connect()
@@ -82,7 +81,7 @@ for idx in range(len(episode_frames)):
 
     # Get recorded action from dataset
     ee_action = {
-        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        name: float(actions[idx]["action"][i]) for i, name in enumerate(dataset.features["action"]["names"])
     }
 
     # Get robot observation

examples/phone_to_so100/teleoperate.py

@@ -33,7 +33,7 @@ from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone
 from lerobot.utils.robot_utils import busy_wait
-from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+from lerobot.utils.visualization_utils import _init_rerun, log_rerun_data
 
 FPS = 30
 
@@ -67,6 +67,7 @@ phone_to_robot_joints_processor = RobotProcessorPipeline[tuple[RobotAction, Robo
         EEBoundsAndSafety(
             end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
             max_ee_step_m=0.10,
+            max_ee_twist_step_rad=0.50,
         ),
         GripperVelocityToJoint(
             speed_factor=20.0,
@@ -86,7 +87,7 @@ robot.connect()
 teleop_device.connect()
 
 # Init rerun viewer
-init_rerun(session_name="phone_so100_teleop")
+_init_rerun(session_name="phone_so100_teleop")
 
 if not robot.is_connected or not teleop_device.is_connected:
     raise ValueError("Robot or teleop is not connected!")

examples/port_datasets/port_droid.py

@@ -362,8 +362,6 @@ def port_droid(
         lerobot_dataset.save_episode()
         logging.info("Save_episode")
 
-    lerobot_dataset.finalize()
-
     if push_to_hub:
         lerobot_dataset.push_to_hub(
             # Add openx tag, since it belongs to the openx collection of datasets

examples/so100_to_so100_EE/evaluate.py

@@ -34,16 +34,16 @@ from lerobot.processor.converters import (
     transition_to_observation,
     transition_to_robot_action,
 )
+from lerobot.record import record_loop
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.robot_kinematic_processor import (
     ForwardKinematicsJointsToEE,
     InverseKinematicsEEToJoints,
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import _init_rerun
 
 NUM_EPISODES = 5
 FPS = 30
@@ -138,7 +138,7 @@ robot.connect()
 
 # Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
-init_rerun(session_name="so100_so100_evaluate")
+_init_rerun(session_name="so100_so100_evaluate")
 
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
@@ -195,6 +195,4 @@ for episode_idx in range(NUM_EPISODES):
 log_say("Stop recording")
 robot.disconnect()
 listener.stop()
-
-dataset.finalize()
 dataset.push_to_hub()

examples/so100_to_so100_EE/record.py

@@ -27,6 +27,7 @@ from lerobot.processor.converters import (
     transition_to_observation,
     transition_to_robot_action,
 )
+from lerobot.record import record_loop
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.robot_kinematic_processor import (
     EEBoundsAndSafety,
@@ -34,12 +35,11 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.scripts.lerobot_record import record_loop
 from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
 from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
 from lerobot.utils.control_utils import init_keyboard_listener
 from lerobot.utils.utils import log_say
-from lerobot.utils.visualization_utils import init_rerun
+from lerobot.utils.visualization_utils import _init_rerun
 
 NUM_EPISODES = 2
 FPS = 30
@@ -101,6 +101,7 @@ ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservati
         EEBoundsAndSafety(
             end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
             max_ee_step_m=0.10,
+            max_ee_twist_step_rad=0.50,
         ),
         InverseKinematicsEEToJoints(
             kinematics=follower_kinematics_solver,
@@ -142,7 +143,7 @@ follower.connect()
 
 # Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
-init_rerun(session_name="recording_phone")
+_init_rerun(session_name="recording_phone")
 
 if not leader.is_connected or not follower.is_connected:
     raise ValueError("Robot or teleop is not connected!")
@@ -199,6 +200,4 @@ log_say("Stop recording")
 leader.disconnect()
 follower.disconnect()
 listener.stop()
-
-dataset.finalize()
 dataset.push_to_hub()

examples/so100_to_so100_EE/replay.py

@@ -29,7 +29,6 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
     InverseKinematicsEEToJoints,
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.utils.constants import ACTION
 from lerobot.utils.robot_utils import busy_wait
 from lerobot.utils.utils import log_say
 
@@ -68,7 +67,7 @@ robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotOb
 dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
 # Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
 episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
-actions = episode_frames.select_columns(ACTION)
+actions = episode_frames.select_columns("action")
 
 # Connect to the robot
 robot.connect()
@@ -83,7 +82,7 @@ for idx in range(len(episode_frames)):
 
     # Get recorded action from dataset
     ee_action = {
-        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        name: float(actions[idx]["action"][i]) for i, name in enumerate(dataset.features["action"]["names"])
     }
 
     # Get robot observation

examples/so100_to_so100_EE/teleoperate.py

@@ -33,7 +33,7 @@ from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
 from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
 from lerobot.utils.robot_utils import busy_wait
-from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
+from lerobot.utils.visualization_utils import _init_rerun, log_rerun_data
 
 FPS = 30
 
@@ -78,6 +78,7 @@ ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservati
         EEBoundsAndSafety(
             end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
             max_ee_step_m=0.10,
+            max_ee_twist_step_rad=0.50,
         ),
         InverseKinematicsEEToJoints(
             kinematics=follower_kinematics_solver,
@@ -94,7 +95,7 @@ follower.connect()
 leader.connect()
 
 # Init rerun viewer
-init_rerun(session_name="so100_so100_EE_teleop")
+_init_rerun(session_name="so100_so100_EE_teleop")
 
 print("Starting teleop loop...")
 while True:

examples/training/train_with_streaming.py

@@ -20,13 +20,13 @@ from pathlib import Path
 import torch
 
 from lerobot.configs.types import FeatureType
+from lerobot.constants import ACTION
 from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.policies.act.modeling_act import ACTPolicy
 from lerobot.policies.factory import make_pre_post_processors
-from lerobot.utils.constants import ACTION
 
 
 def main():

examples/tutorial/act/act_training_example.py

@@ -1,98 +0,0 @@
-"""This script demonstrates how to train ACT Policy on a real-world dataset."""
-
-from pathlib import Path
-
-import torch
-
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.datasets.utils import dataset_to_policy_features
-from lerobot.policies.act.configuration_act import ACTConfig
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
-
-
-def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:
-    if delta_indices is None:
-        return [0]
-
-    return [i / fps for i in delta_indices]
-
-
-output_directory = Path("outputs/robot_learning_tutorial/act")
-output_directory.mkdir(parents=True, exist_ok=True)
-
-# Select your device
-device = torch.device("mps")  # or "cuda" or "cpu"
-
-dataset_id = "lerobot/svla_so101_pickplace"
-
-# This specifies the inputs the model will be expecting and the outputs it will produce
-dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-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}
-
-cfg = ACTConfig(input_features=input_features, output_features=output_features)
-policy = ACTPolicy(cfg)
-preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
-
-policy.train()
-policy.to(device)
-
-# To perform action chunking, ACT expects a given number of actions as targets
-delta_timestamps = {
-    "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
-}
-
-# add image features if they are present
-delta_timestamps |= {
-    k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps) for k in cfg.image_features
-}
-
-# Instantiate the dataset
-dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
-
-# Create the optimizer and dataloader for offline training
-optimizer = cfg.get_optimizer_preset().build(policy.parameters())
-batch_size = 32
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    batch_size=batch_size,
-    shuffle=True,
-    pin_memory=device.type != "cpu",
-    drop_last=True,
-)
-
-# Number of training steps and logging frequency
-training_steps = 1
-log_freq = 1
-
-# Run training loop
-step = 0
-done = False
-while not done:
-    for batch in dataloader:
-        batch = preprocessor(batch)
-        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 the policy checkpoint, alongside the pre/post processors
-policy.save_pretrained(output_directory)
-preprocessor.save_pretrained(output_directory)
-postprocessor.save_pretrained(output_directory)
-
-# Save all assets to the Hub
-policy.push_to_hub("fracapuano/robot_learning_tutorial_act")
-preprocessor.push_to_hub("fracapuano/robot_learning_tutorial_act")
-postprocessor.push_to_hub("fracapuano/robot_learning_tutorial_act")

examples/tutorial/act/act_using_example.py

@@ -1,57 +0,0 @@
-import torch
-
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.policies.act.modeling_act import ACTPolicy
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
-
-device = torch.device("mps")  # or "cuda" or "cpu"
-model_id = "fracapuano/robot_learning_tutorial_act"
-model = ACTPolicy.from_pretrained(model_id)
-
-dataset_id = "lerobot/svla_so101_pickplace"
-# This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
-dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-preprocess, postprocess = make_pre_post_processors(model.config, dataset_stats=dataset_metadata.stats)
-
-# # find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
-
-# # the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
-
-MAX_EPISODES = 5
-MAX_STEPS_PER_EPISODE = 20
-
-# Robot and environment configuration
-# Camera keys must match the name and resolutions of the ones used for training!
-# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
-camera_config = {
-    "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-}
-
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-robot = SO100Follower(robot_cfg)
-robot.connect()
-
-for _ in range(MAX_EPISODES):
-    for _ in range(MAX_STEPS_PER_EPISODE):
-        obs = robot.get_observation()
-        obs_frame = build_inference_frame(
-            observation=obs, ds_features=dataset_metadata.features, device=device
-        )
-
-        obs = preprocess(obs_frame)
-
-        action = model.select_action(obs)
-        action = postprocess(action)
-
-        action = make_robot_action(action, dataset_metadata.features)
-
-        robot.send_action(action)
-
-    print("Episode finished! Starting new episode...")

examples/tutorial/async-inf/policy_server.py

@@ -1,11 +0,0 @@
-from lerobot.async_inference.configs import PolicyServerConfig
-from lerobot.async_inference.policy_server import serve
-
-host = ...  # something like "127.0.0.1" if you're exposing to localhost
-port = ...  # something like 8080
-
-config = PolicyServerConfig(
-    host=host,
-    port=port,
-)
-serve(config)

examples/tutorial/async-inf/robot_client.py

@@ -1,55 +0,0 @@
-import threading
-
-from lerobot.async_inference.configs import RobotClientConfig
-from lerobot.async_inference.helpers import visualize_action_queue_size
-from lerobot.async_inference.robot_client import RobotClient
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.robots.so100_follower import SO100FollowerConfig
-
-# these cameras must match the ones expected by the policy - find your cameras with lerobot-find-cameras
-# check the config.json on the Hub for the policy you are using to see the expected camera specs
-camera_cfg = {
-    "up": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "side": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-}
-
-# # find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
-
-# # the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
-
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_cfg)
-
-server_address = ...  # something like "127.0.0.1:8080" if using localhost
-
-# 3. Create client configuration
-client_cfg = RobotClientConfig(
-    robot=robot_cfg,
-    server_address=server_address,
-    policy_device="mps",
-    policy_type="act",
-    pretrained_name_or_path="fracapuano/robot_learning_tutorial_act",
-    chunk_size_threshold=0.5,  # g
-    actions_per_chunk=50,  # make sure this is less than the max actions of the policy
-)
-
-# 4. Create and start client
-client = RobotClient(client_cfg)
-
-# 5. Provide a textual description of the task
-task = ...
-
-if client.start():
-    # Start action receiver thread
-    action_receiver_thread = threading.Thread(target=client.receive_actions, daemon=True)
-    action_receiver_thread.start()
-
-    try:
-        # Run the control loop
-        client.control_loop(task)
-    except KeyboardInterrupt:
-        client.stop()
-        action_receiver_thread.join()
-        # (Optionally) plot the action queue size
-        visualize_action_queue_size(client.action_queue_size)

examples/tutorial/diffusion/diffusion_training_example.py

@@ -1,99 +0,0 @@
-"""This script demonstrates how to train Diffusion Policy on a real-world dataset."""
-
-from pathlib import Path
-
-import torch
-
-from lerobot.configs.types import FeatureType
-from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
-from lerobot.datasets.utils import dataset_to_policy_features
-from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.policies.factory import make_pre_post_processors
-
-
-def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:
-    if delta_indices is None:
-        return [0]
-
-    return [i / fps for i in delta_indices]
-
-
-output_directory = Path("outputs/robot_learning_tutorial/diffusion")
-output_directory.mkdir(parents=True, exist_ok=True)
-
-# Select your device
-device = torch.device("mps")  # or "cuda" or "cpu"
-
-dataset_id = "lerobot/svla_so101_pickplace"
-
-# This specifies the inputs the model will be expecting and the outputs it will produce
-dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-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}
-
-cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
-policy = DiffusionPolicy(cfg)
-preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
-
-policy.train()
-policy.to(device)
-
-# To perform action chunking, ACT expects a given number of actions as targets
-delta_timestamps = {
-    "observation.state": make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps),
-    "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
-}
-
-# add image features if they are present
-delta_timestamps |= {
-    k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps) for k in cfg.image_features
-}
-
-# Instantiate the dataset
-dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
-
-# Create the optimizer and dataloader for offline training
-optimizer = cfg.get_optimizer_preset().build(policy.parameters())
-batch_size = 32
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    batch_size=batch_size,
-    shuffle=True,
-    pin_memory=device.type != "cpu",
-    drop_last=True,
-)
-
-# Number of training steps and logging frequency
-training_steps = 1
-log_freq = 1
-
-# Run training loop
-step = 0
-done = False
-while not done:
-    for batch in dataloader:
-        batch = preprocessor(batch)
-        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 the policy checkpoint, alongside the pre/post processors
-policy.save_pretrained(output_directory)
-preprocessor.save_pretrained(output_directory)
-postprocessor.save_pretrained(output_directory)
-
-# Save all assets to the Hub
-policy.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
-preprocessor.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
-postprocessor.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")

examples/tutorial/diffusion/diffusion_using_example.py

@@ -1,60 +0,0 @@
-import torch
-
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
-
-device = torch.device("mps")  # or "cuda" or "cpu"
-model_id = "fracapuano/robot_learning_tutorial_diffusion"
-
-model = DiffusionPolicy.from_pretrained(model_id)
-
-dataset_id = "lerobot/svla_so101_pickplace"
-# This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
-dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-preprocess, postprocess = make_pre_post_processors(
-    model.config, model_id, dataset_stats=dataset_metadata.stats
-)
-
-MAX_EPISODES = 5
-MAX_STEPS_PER_EPISODE = 20
-
-
-# # find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
-
-# # the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
-
-# Robot and environment configuration
-# Camera keys must match the name and resolutions of the ones used for training!
-# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
-camera_config = {
-    "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-}
-
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-robot = SO100Follower(robot_cfg)
-robot.connect()
-
-
-for _ in range(MAX_EPISODES):
-    for _ in range(MAX_STEPS_PER_EPISODE):
-        obs = robot.get_observation()
-        obs_frame = build_inference_frame(
-            observation=obs, ds_features=dataset_metadata.features, device=device
-        )
-
-        obs = preprocess(obs_frame)
-
-        action = model.select_action(obs)
-        action = postprocess(action)
-        action = make_robot_action(action, dataset_metadata.features)
-        robot.send_action(action)
-
-    print("Episode finished! Starting new episode...")

examples/tutorial/pi0/using_pi0_example.py

@@ -1,67 +0,0 @@
-import torch
-
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.policies.pi0.modeling_pi0 import PI0Policy
-from lerobot.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
-
-MAX_EPISODES = 5
-MAX_STEPS_PER_EPISODE = 20
-
-device = torch.device("mps")  # or "cuda" or "cpu"
-model_id = "lerobot/pi0_base"
-
-model = PI0Policy.from_pretrained(model_id)
-
-preprocess, postprocess = make_pre_post_processors(
-    model.config,
-    model_id,
-    # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
-    preprocessor_overrides={"device_processor": {"device": str(device)}},
-)
-
-# find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
-
-# the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
-
-# Robot and environment configuration
-# Camera keys must match the name and resolutions of the ones used for training!
-# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
-camera_config = {
-    "base_0_rgb": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "left_wrist_0_rgb": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-    "right_wrist_0_rgb": OpenCVCameraConfig(index_or_path=2, width=640, height=480, fps=30),
-}
-
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-robot = SO100Follower(robot_cfg)
-robot.connect()
-
-task = ""  # something like "pick the red block"
-robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
-
-# This is used to match the raw observation keys to the keys expected by the policy
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-dataset_features = {**action_features, **obs_features}
-
-for _ in range(MAX_EPISODES):
-    for _ in range(MAX_STEPS_PER_EPISODE):
-        obs = robot.get_observation()
-        obs_frame = build_inference_frame(
-            observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
-        )
-
-        obs = preprocess(obs_frame)
-
-        action = model.select_action(obs)
-        action = postprocess(action)
-        action = make_robot_action(action, dataset_features)
-        robot.send_action(action)
-
-    print("Episode finished! Starting new episode...")

examples/tutorial/rl/hilserl_example.py

@@ -1,345 +0,0 @@
-import multiprocessing as mp
-import signal
-from pathlib import Path
-from queue import Empty, Full
-
-import torch
-import torch.optim as optim
-
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
-from lerobot.policies.sac.configuration_sac import SACConfig
-from lerobot.policies.sac.modeling_sac import SACPolicy
-from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
-from lerobot.rl.buffer import ReplayBuffer
-from lerobot.rl.gym_manipulator import make_robot_env
-from lerobot.robots.so100_follower import SO100FollowerConfig
-from lerobot.teleoperators.so100_leader import SO100LeaderConfig
-from lerobot.teleoperators.utils import TeleopEvents
-
-LOG_EVERY = 10
-SEND_EVERY = 10
-
-
-def run_learner(
-    transitions_queue: mp.Queue,
-    parameters_queue: mp.Queue,
-    shutdown_event: mp.Event,
-    policy_learner: SACPolicy,
-    online_buffer: ReplayBuffer,
-    offline_buffer: ReplayBuffer,
-    lr: float = 3e-4,
-    batch_size: int = 32,
-    device: torch.device = "mps",
-):
-    """The learner process - trains SAC policy on transitions streamed from the actor, updating parameters
-    for the actor to adopt."""
-    policy_learner.train()
-    policy_learner.to(device)
-
-    # Create Adam optimizer from scratch - simple and clean
-    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
-
-    print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
-    print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
-
-    training_step = 0
-
-    while not shutdown_event.is_set():
-        # retrieve incoming transitions from the actor process
-        try:
-            transitions = transitions_queue.get(timeout=0.1)
-            for transition in transitions:
-                # HIL-SERL: Add ALL transitions to online buffer
-                online_buffer.add(**transition)
-
-                # HIL-SERL: Add ONLY human intervention transitions to offline buffer
-                is_intervention = transition.get("complementary_info", {}).get("is_intervention", False)
-                if is_intervention:
-                    offline_buffer.add(**transition)
-                    print(
-                        f"[LEARNER] Human intervention detected! Added to offline buffer (now {len(offline_buffer)} transitions)"
-                    )
-
-        except Empty:
-            pass  # No transitions available, continue
-
-        # Train if we have enough data
-        if len(online_buffer) >= policy_learner.config.online_step_before_learning:
-            # Sample from online buffer (autonomous + human data)
-            online_batch = online_buffer.sample(batch_size // 2)
-
-            # Sample from offline buffer (human demonstrations only, either precollected or at runtime)
-            offline_batch = offline_buffer.sample(batch_size // 2)
-
-            # Combine batches - this is the key HIL-SERL mechanism!
-            batch = {}
-            for key in online_batch:
-                if key in offline_batch:
-                    batch[key] = torch.cat([online_batch[key], offline_batch[key]], dim=0)
-                else:
-                    batch[key] = online_batch[key]
-
-            loss, _ = policy_learner.forward(batch)
-
-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
-            training_step += 1
-
-            if training_step % LOG_EVERY == 0:
-                print(
-                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
-                    f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
-                )
-
-            # Send updated parameters to actor every 10 training steps
-            if training_step % SEND_EVERY == 0:
-                try:
-                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
-                    parameters_queue.put_nowait(state_dict)
-                    print("[LEARNER] Sent updated parameters to actor")
-                except Full:
-                    # Missing write due to queue not being consumed (should happen rarely)
-                    pass
-
-    print("[LEARNER] Learner process finished")
-
-
-def run_actor(
-    transitions_queue: mp.Queue,
-    parameters_queue: mp.Queue,
-    shutdown_event: mp.Event,
-    policy_actor: SACPolicy,
-    reward_classifier: Classifier,
-    env_cfg: HILSerlRobotEnvConfig,
-    device: torch.device = "mps",
-    output_directory: Path | None = None,
-):
-    """The actor process - interacts with environment and collects data.
-    The policy is frozen and only the parameters are updated, popping the most recent ones from a queue."""
-    policy_actor.eval()
-    policy_actor.to(device)
-
-    reward_classifier.eval()
-    reward_classifier.to(device)
-
-    # Create robot environment inside the actor process
-    env, teleop_device = make_robot_env(env_cfg)
-
-    try:
-        for episode in range(MAX_EPISODES):
-            if shutdown_event.is_set():
-                break
-
-            obs, _info = env.reset()
-            episode_reward = 0.0
-            step = 0
-            episode_transitions = []
-
-            print(f"[ACTOR] Starting episode {episode + 1}")
-
-            while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
-                try:
-                    new_params = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_params)
-                    print("[ACTOR] Updated policy parameters from learner")
-                except Empty:  # No new updated parameters available from learner, waiting
-                    pass
-
-                # Get action from policy
-                policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
-                action = action_tensor.squeeze(0).cpu().numpy()
-
-                # Step environment
-                next_obs, _env_reward, terminated, truncated, _info = env.step(action)
-                done = terminated or truncated
-
-                # Predict reward
-                policy_next_obs = make_policy_obs(next_obs, device=device)
-                reward = reward_classifier.predict_reward(policy_next_obs)
-
-                if reward >= 1.0 and not done:  # success detected! halt episode
-                    terminated = True
-                    done = True
-
-                # In HIL-SERL, human interventions come from the teleop device
-                is_intervention = False
-                if hasattr(teleop_device, "get_teleop_events"):
-                    # Real intervention detection from teleop device
-                    teleop_events = teleop_device.get_teleop_events()
-                    is_intervention = teleop_events.get(TeleopEvents.IS_INTERVENTION, False)
-
-                # Store transition with intervention metadata
-                transition = {
-                    "state": policy_obs,
-                    "action": action,
-                    "reward": float(reward) if hasattr(reward, "item") else reward,
-                    "next_state": policy_next_obs,
-                    "done": done,
-                    "truncated": truncated,
-                    "complementary_info": {
-                        "is_intervention": is_intervention,
-                    },
-                }
-
-                episode_transitions.append(transition)
-
-                episode_reward += reward
-                step += 1
-
-                obs = next_obs
-
-                if done:
-                    break
-
-            # Send episode transitions to learner
-            transitions_queue.put_nowait(episode_transitions)
-
-    except KeyboardInterrupt:
-        print("[ACTOR] Interrupted by user")
-    finally:
-        # Clean up
-        if hasattr(env, "robot") and env.robot.is_connected:
-            env.robot.disconnect()
-        if teleop_device and hasattr(teleop_device, "disconnect"):
-            teleop_device.disconnect()
-        if output_directory is not None:
-            policy_actor.save_pretrained(output_directory)
-            print(f"[ACTOR] Latest actor policy saved at: {output_directory}")
-
-        print("[ACTOR] Actor process finished")
-
-
-def make_policy_obs(obs, device: torch.device = "cpu"):
-    return {
-        "observation.state": torch.from_numpy(obs["agent_pos"]).float().unsqueeze(0).to(device),
-        **{
-            f"observation.image.{k}": torch.from_numpy(obs["pixels"][k]).float().unsqueeze(0).to(device)
-            for k in obs["pixels"]
-        },
-    }
-
-
-"""Main function - coordinates actor and learner processes."""
-
-device = "mps"  # or "cuda" or "cpu"
-output_directory = Path("outputs/robot_learning_tutorial/hil_serl")
-output_directory.mkdir(parents=True, exist_ok=True)
-
-# find ports using lerobot-find-port
-follower_port = ...
-leader_port = ...
-
-# the robot ids are used the load the right calibration files
-follower_id = ...
-leader_id = ...
-
-# A pretrained model (to be used in-distribution!)
-reward_classifier_id = "fracapuano/reward_classifier_hil_serl_example"
-reward_classifier = Classifier.from_pretrained(reward_classifier_id)
-
-reward_classifier.to(device)
-reward_classifier.eval()
-
-MAX_EPISODES = 5
-MAX_STEPS_PER_EPISODE = 20
-
-# Robot and environment configuration
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id)
-teleop_cfg = SO100LeaderConfig(port=leader_port, id=leader_id)
-processor_cfg = HILSerlProcessorConfig(control_mode="leader")
-
-env_cfg = HILSerlRobotEnvConfig(robot=robot_cfg, teleop=teleop_cfg, processor=processor_cfg)
-
-# Create robot environment
-env, teleop_device = make_robot_env(env_cfg)
-
-obs_features = hw_to_dataset_features(env.robot.observation_features, "observation")
-action_features = hw_to_dataset_features(env.robot.action_features, "action")
-
-# Create SAC policy for action selection
-policy_cfg = SACConfig(
-    device=device,
-    input_features=obs_features,
-    output_features=action_features,
-)
-
-policy_actor = SACPolicy(policy_cfg)
-policy_learner = SACPolicy(policy_cfg)
-
-demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
-offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)
-
-# Online buffer: initialized from scratch
-online_replay_buffer = ReplayBuffer(device=device, state_keys=list(obs_features.keys()))
-# Offline buffer: Created from dataset (pre-populated it with demonstrations)
-offline_replay_buffer = ReplayBuffer.from_lerobot_dataset(
-    lerobot_dataset=offline_dataset, device=device, state_keys=list(obs_features.keys())
-)
-
-# Create communication channels between learner and actor processes
-transitions_queue = mp.Queue(maxsize=10)
-parameters_queue = mp.Queue(maxsize=2)
-shutdown_event = mp.Event()
-
-
-# Signal handler for graceful shutdown
-def signal_handler(sig):
-    print(f"\nSignal {sig} received, shutting down...")
-    shutdown_event.set()
-
-
-signal.signal(signal.SIGINT, signal_handler)
-signal.signal(signal.SIGTERM, signal_handler)
-
-# Create processes
-learner_process = mp.Process(
-    target=run_learner,
-    args=(
-        transitions_queue,
-        parameters_queue,
-        shutdown_event,
-        policy_learner,
-        online_replay_buffer,
-        offline_replay_buffer,
-    ),
-    kwargs={"device": device},  # can run on accelerated hardware for training
-)
-
-actor_process = mp.Process(
-    target=run_actor,
-    args=(
-        transitions_queue,
-        parameters_queue,
-        shutdown_event,
-        policy_actor,
-        reward_classifier,
-        env_cfg,
-        output_directory,
-    ),
-    kwargs={"device": "cpu"},  # actor is frozen, can run on CPU or accelerate for inference
-)
-
-learner_process.start()
-actor_process.start()
-
-try:
-    # Wait for actor to finish (it controls the episode loop)
-    actor_process.join()
-    shutdown_event.set()
-    learner_process.join(timeout=10)
-
-except KeyboardInterrupt:
-    print("Main process interrupted")
-    shutdown_event.set()
-    actor_process.join(timeout=5)
-    learner_process.join(timeout=10)
-
-finally:
-    if learner_process.is_alive():
-        learner_process.terminate()
-    if actor_process.is_alive():
-        actor_process.terminate()

examples/tutorial/rl/reward_classifier_example.py

@@ -1,62 +0,0 @@
-import torch
-
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.policies.factory import make_policy, make_pre_post_processors
-from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
-
-# Device to use for training
-device = "mps"  # or "cuda", or "cpu"
-
-# Load the dataset used for training
-repo_id = "lerobot/example_hil_serl_dataset"
-dataset = LeRobotDataset(repo_id)
-
-# Configure the policy to extract features from the image frames
-camera_keys = dataset.meta.camera_keys
-
-config = RewardClassifierConfig(
-    num_cameras=len(camera_keys),
-    device=device,
-    # backbone model to extract features from the image frames
-    model_name="microsoft/resnet-18",
-)
-
-# Make policy, preprocessor, and optimizer
-policy = make_policy(config, ds_meta=dataset.meta)
-optimizer = config.get_optimizer_preset().build(policy.parameters())
-preprocessor, _ = make_pre_post_processors(policy_cfg=config, dataset_stats=dataset.meta.stats)
-
-
-classifier_id = "fracapuano/reward_classifier_hil_serl_example"
-
-# Instantiate a dataloader
-dataloader = torch.utils.data.DataLoader(dataset, batch_size=16, shuffle=True)
-
-# Training loop
-num_epochs = 5
-for epoch in range(num_epochs):
-    total_loss = 0
-    total_accuracy = 0
-    for batch in dataloader:
-        # Preprocess the batch and move it to the correct device.
-        batch = preprocessor(batch)
-
-        # Forward pass
-        loss, output_dict = policy.forward(batch)
-
-        # Backward pass and optimization
-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
-
-        total_loss += loss.item()
-        total_accuracy += output_dict["accuracy"]
-
-    avg_loss = total_loss / len(dataloader)
-    avg_accuracy = total_accuracy / len(dataloader)
-    print(f"Epoch {epoch + 1}/{num_epochs}, Loss: {avg_loss:.4f}, Accuracy: {avg_accuracy:.2f}%")
-
-print("Training finished!")
-
-# You can now save the trained policy.
-policy.push_to_hub(classifier_id)

examples/tutorial/smolvla/using_smolvla_example.py

@@ -1,66 +0,0 @@
-import torch
-
-from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.policies.factory import make_pre_post_processors
-from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
-from lerobot.policies.utils import build_inference_frame, make_robot_action
-from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
-from lerobot.robots.so100_follower.so100_follower import SO100Follower
-
-MAX_EPISODES = 5
-MAX_STEPS_PER_EPISODE = 20
-
-device = torch.device("mps")  # or "cuda" or "cpu"
-model_id = "lerobot/smolvla_base"
-
-model = SmolVLAPolicy.from_pretrained(model_id)
-
-preprocess, postprocess = make_pre_post_processors(
-    model.config,
-    model_id,
-    # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
-    preprocessor_overrides={"device_processor": {"device": str(device)}},
-)
-
-# find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
-
-# the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
-
-# Robot and environment configuration
-# Camera keys must match the name and resolutions of the ones used for training!
-# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
-camera_config = {
-    "camera1": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "camera2": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-}
-
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-robot = SO100Follower(robot_cfg)
-robot.connect()
-
-task = ""  # something like "pick the red block"
-robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
-
-# This is used to match the raw observation keys to the keys expected by the policy
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-dataset_features = {**action_features, **obs_features}
-
-for _ in range(MAX_EPISODES):
-    for _ in range(MAX_STEPS_PER_EPISODE):
-        obs = robot.get_observation()
-        obs_frame = build_inference_frame(
-            observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
-        )
-
-        obs = preprocess(obs_frame)
-
-        action = model.select_action(obs)
-        action = postprocess(action)
-        action = make_robot_action(action, dataset_features)
-        robot.send_action(action)
-
-    print("Episode finished! Starting new episode...")

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.1"
+version = "0.3.4"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -59,30 +59,28 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<4.2.0",
-    "diffusers>=0.27.2,<0.36.0",
-    "huggingface-hub[hf-transfer,cli]>=0.34.2,<0.36.0",
-    "accelerate>=1.10.0,<2.0.0",
+    "datasets>=4.0.0",
+    "diffusers>=0.27.2",
+    "huggingface-hub[hf-transfer,cli]>=0.34.2",
 
     # Core dependencies
-    "setuptools>=71.0.0,<81.0.0",
-    "cmake>=3.29.0.1,<4.2.0",
-    "einops>=0.8.0,<0.9.0",
-    "opencv-python-headless>=4.9.0,<4.13.0",
-    "av>=15.0.0,<16.0.0",
-    "jsonlines>=4.0.0,<5.0.0",
-    "packaging>=24.2,<26.0",
-    "pynput>=1.7.7,<1.9.0",
-    "pyserial>=3.5,<4.0",
-    "wandb>=0.20.0,<0.22.0", # TODO: Bumb dependency (compatible with protobuf)
+    "cmake>=3.29.0.1",
+    "einops>=0.8.0",
+    "opencv-python-headless>=4.9.0",
+    "av>=14.2.0",
+    "jsonlines>=4.0.0",
+    "packaging>=24.2",
+    "pynput>=1.7.7",
+    "pyserial>=3.5",
+    "wandb>=0.20.0",
 
     "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
     "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
     "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
 
     "draccus==0.10.0", # TODO: Remove ==
-    "gymnasium>=1.1.1,<2.0.0",
-    "rerun-sdk>=0.24.0,<0.27.0",
+    "gymnasium>=0.29.1,<1.0.0", # TODO: Bumb dependency
+    "rerun-sdk>=0.21.0,<0.23.0", # TODO: Bumb dependency
 
     # Support dependencies
     "deepdiff>=7.0.1,<9.0.0",
@@ -94,56 +92,51 @@ dependencies = [
 [project.optional-dependencies]
 
 # Common
-pygame-dep = ["pygame>=2.5.1,<2.7.0"]
-placo-dep = ["placo>=0.9.6,<0.10.0"]
-transformers-dep = ["transformers>=4.53.0,<5.0.0"]
-grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"] # TODO: Bumb dependency (compatible with wandb)
+pygame-dep = ["pygame>=2.5.1"]
+placo-dep = ["placo>=0.9.6"]
+transformers-dep = ["transformers>=4.53.0"]
+grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"]
 
 # Motors
-feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
-dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
+feetech = ["feetech-servo-sdk>=1.0.0"]
+dynamixel = ["dynamixel-sdk>=3.7.31"]
 
 # Robots
-gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
+gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
-lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
-reachy2 = ["reachy2_sdk>=1.0.14,<1.1.0"]
+lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1"]
+reachy2 = ["reachy2_sdk>=1.0.14"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
-    "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
-    "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
+    "pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'",
+    "pyrealsense2-macosx>=2.54 ; sys_platform == 'darwin'",
 ]
-phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
+phone = ["hebi-py>=2.8.0", "teleop>=0.1.0"]
+# stretch = [
+#     "hello-robot-stretch-body>=0.7.27 ; sys_platform == 'linux'",
+#     "pyrender @ git+https://github.com/mmatl/pyrender.git ; sys_platform == 'linux'",
+#     "pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'"
+# ] # TODO: Currently not supported
 
 # Policies
-pi = ["transformers @ git+https://github.com/huggingface/transformers.git@fix/lerobot_openpi"]
-smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14,<0.6.0", "accelerate>=1.7.0,<2.0.0", "safetensors>=0.4.3,<1.0.0"]
-groot = [
-    "lerobot[transformers-dep]",
-    "peft>=0.13.0,<1.0.0",
-    "dm-tree>=0.1.8,<1.0.0",
-    "timm>=1.0.0,<1.1.0",
-    "safetensors>=0.4.3,<1.0.0",
-    "Pillow>=10.0.0,<13.0.0",
-    "decord>=0.6.0,<1.0.0; (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
-    "ninja>=1.11.1,<2.0.0",
-    "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
-]
-hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+pi = ["lerobot[transformers-dep]"]
+smolvla = ["lerobot[transformers-dep]", "num2words>=0.5.14", "accelerate>=1.7.0", "safetensors>=0.4.3"]
+hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.11", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
-async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
+async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3"]
 
 # Development
-dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1"]
-test = ["pytest>=8.1.0,<9.0.0", "pytest-timeout>=2.4.0,<3.0.0", "pytest-cov>=5.0.0,<8.0.0", "mock-serial>=0.0.1,<0.1.0 ; sys_platform != 'win32'"]
-video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
+dev = ["pre-commit>=3.7.0", "debugpy>=1.8.1", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1"]
+test = ["pytest>=8.1.0", "pytest-timeout>=2.4.0", "pytest-cov>=5.0.0", "mock-serial>=0.0.1 ; sys_platform != 'win32'"]
+video_benchmark = ["scikit-image>=0.23.2", "pandas>=2.2.2"]
 
 # Simulation
-aloha = ["gym-aloha>=0.1.2,<0.2.0"]
-pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
-metaworld = ["metaworld==3.0.0"]
+aloha = ["gym-aloha>=0.1.1"]
+pusht = ["gym-pusht>=0.1.5", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
+xarm = ["gym-xarm>=0.1.1"]
+libero = ["lerobot[transformers-dep]", "libero @ git+https://github.com/huggingface/lerobot-libero.git@main#egg=libero"]
+
 
 # All
 all = [
@@ -156,7 +149,6 @@ all = [
     "lerobot[intelrealsense]",
     "lerobot[pi]",
     "lerobot[smolvla]",
-    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[hilserl]",
     "lerobot[async]",
     "lerobot[dev]",
@@ -164,26 +156,21 @@ all = [
     "lerobot[video_benchmark]",
     "lerobot[aloha]",
     "lerobot[pusht]",
+    "lerobot[xarm]",
     "lerobot[phone]",
     "lerobot[libero]",
-    "lerobot[metaworld]",
 ]
 
 [project.scripts]
-lerobot-calibrate="lerobot.scripts.lerobot_calibrate:main"
-lerobot-find-cameras="lerobot.scripts.lerobot_find_cameras:main"
-lerobot-find-port="lerobot.scripts.lerobot_find_port:main"
-lerobot-record="lerobot.scripts.lerobot_record:main"
-lerobot-replay="lerobot.scripts.lerobot_replay:main"
-lerobot-setup-motors="lerobot.scripts.lerobot_setup_motors:main"
-lerobot-teleoperate="lerobot.scripts.lerobot_teleoperate:main"
-lerobot-eval="lerobot.scripts.lerobot_eval:main"
-lerobot-train="lerobot.scripts.lerobot_train:main"
-lerobot-dataset-viz="lerobot.scripts.lerobot_dataset_viz:main"
-lerobot-info="lerobot.scripts.lerobot_info:main"
-lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
-lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
-lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
+lerobot-calibrate="lerobot.calibrate:main"
+lerobot-find-cameras="lerobot.find_cameras:main"
+lerobot-find-port="lerobot.find_port:main"
+lerobot-record="lerobot.record:main"
+lerobot-replay="lerobot.replay:main"
+lerobot-setup-motors="lerobot.setup_motors:main"
+lerobot-teleoperate="lerobot.teleoperate:main"
+lerobot-eval="lerobot.scripts.eval:main"
+lerobot-train="lerobot.scripts.train:main"
 
 # ---------------- Tool Configurations ----------------
 [tool.setuptools.packages.find]
@@ -210,7 +197,7 @@ exclude = ["tests/artifacts/**/*.safetensors", "*_pb2.py", "*_pb2_grpc.py"]
 # N: pep8-naming
 # TODO: Uncomment rules when ready to use
 select = [
-    "E", "W", "F", "I", "B", "C4", "T20", "N", "UP", "SIM" #, "A", "S", "D", "RUF"
+    "E", "W", "F", "I", "B", "C4", "T20", "N" # "SIM", "A", "S", "D", "RUF", "UP"
 ]
 ignore = [
     "E501", # Line too long
@@ -241,6 +228,9 @@ exclude_dirs = [
     "tests",
     "benchmarks",
     "src/lerobot/datasets/push_dataset_to_hub",
+    "src/lerobot/datasets/v2/convert_dataset_v1_to_v2",
+    "src/lerobot/policies/pi0/conversion_scripts",
+    "src/lerobot/scripts/push_dataset_to_hub.py",
 ]
 skips = ["B101", "B311", "B404", "B603", "B615"]
 
@@ -255,8 +245,6 @@ default.extend-ignore-identifiers-re = [
     "pn",
     "ser",
     "ein",
-    "thw",
-    "inpt",
 ]
 
 # TODO: Uncomment when ready to use
@@ -275,91 +263,8 @@ default.extend-ignore-identifiers-re = [
 # color = true
 # paths = ["src/lerobot"]
 
-# TODO: Enable mypy gradually module by module across multiple PRs
-# Uncomment [tool.mypy] first, then uncomment individual module overrides as they get proper type annotations
-
-[tool.mypy]
-python_version = "3.10"
-ignore_missing_imports = true
-follow_imports = "skip"
+# [tool.mypy]
+# python_version = "3.10"
 # warn_return_any = true
 # warn_unused_configs = true
-# strict = true
-# disallow_untyped_defs = true
-# disallow_incomplete_defs = true
-# check_untyped_defs = true
-
-[[tool.mypy.overrides]]
-module = "lerobot.*"
-ignore_errors = true
-
-[[tool.mypy.overrides]]
-module = "lerobot.envs.*"
-ignore_errors = false
-
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.utils.*"
-# ignore_errors = false
-
-[[tool.mypy.overrides]]
-module = "lerobot.configs.*"
-ignore_errors = false
-
-# extra strictness for configs
-disallow_untyped_defs = true
-disallow_incomplete_defs = true
-check_untyped_defs = true
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.optim.*"
-# ignore_errors = false
-
-[[tool.mypy.overrides]]
-module = "lerobot.model.*"
-ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.processor.*"
-# ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.datasets.*"
-# ignore_errors = false
-
-[[tool.mypy.overrides]]
-module = "lerobot.cameras.*"
-ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.motors.*"
-# ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.robots.*"
-# ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.teleoperators.*"
-# ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.policies.*"
-# ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.rl.*"
-# ignore_errors = false
-
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.async_inference.*"
-# ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.transport.*"
-# ignore_errors = false
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.scripts.*"
-# ignore_errors = false
+# ignore_missing_imports = false

requirements-macos.txt

@@ -1,4 +1,3 @@
-#
 # This file is autogenerated by pip-compile with Python 3.10
 # by the following command:
 #
@@ -13,62 +12,47 @@ absl-py==2.3.1
     #   dm-tree
     #   labmaze
     #   mujoco
-    #   tensorboard
-accelerate==1.11.0
-    # via
-    #   lerobot
-    #   peft
+accelerate==1.9.0
+    # via lerobot
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.13.1
+aiohttp==3.12.15
     # via fsspec
 aiosignal==1.4.0
     # via aiohttp
 annotated-types==0.7.0
     # via pydantic
-antlr4-python3-runtime==4.9.3
-    # via
-    #   hydra-core
-    #   omegaconf
-anyio==4.11.0
-    # via
-    #   starlette
-    #   watchfiles
 asttokens==3.0.0
     # via stack-data
 async-timeout==5.0.1
     # via aiohttp
-attrs==25.4.0
+attrs==25.3.0
     # via
     #   aiohttp
     #   dm-tree
     #   jsonlines
-    #   jsonschema
-    #   referencing
     #   rerun-sdk
-av==15.1.0
+av==15.0.0
     # via lerobot
-bddl==1.0.1
-    # via libero
-certifi==2025.10.5
+blinker==1.9.0
+    # via flask
+certifi==2025.7.14
     # via
     #   requests
     #   sentry-sdk
-cffi==2.0.0
+cffi==1.17.1
     # via pymunk
 cfgv==3.4.0
     # via pre-commit
-charset-normalizer==3.4.4
+charset-normalizer==3.4.2
     # via requests
-click==8.3.0
+click==8.2.1
     # via
-    #   uvicorn
+    #   flask
     #   wandb
 cloudpickle==3.1.1
-    # via
-    #   gymnasium
-    #   libero
-cmake==4.1.0
+    # via gymnasium
+cmake==4.0.3
     # via lerobot
 cmeel==0.57.3
     # via
@@ -110,27 +94,27 @@ coal-library==3.0.1
     # via pin
 contourpy==1.3.2
     # via matplotlib
-coverage[toml]==7.11.0
+coverage[toml]==7.10.1
     # via pytest-cov
 cycler==0.12.1
     # via matplotlib
-datasets==4.1.1
+datasets==3.6.0
     # via lerobot
-debugpy==1.8.17
+debugpy==1.8.15
     # via lerobot
 decorator==5.2.1
     # via ipython
-deepdiff==8.6.1
+deepdiff==8.5.0
     # via lerobot
-diffusers==0.35.2
+diffusers==0.34.0
     # via lerobot
-dill==0.4.0
+dill==0.3.8
     # via
     #   datasets
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.34
+dm-control==1.0.14
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -138,45 +122,29 @@ dm-tree==0.1.9
     # via
     #   dm-control
     #   dm-env
-    #   lerobot
 docopt==0.6.2
     # via num2words
 draccus==0.10.0
     # via lerobot
-dynamixel-sdk==3.8.4
+dynamixel-sdk==3.7.31
     # via lerobot
-easydict==1.13
-    # via libero
-egl-probe @ git+https://github.com/huggingface/egl_probe.git
-    # via
-    #   libero
-    #   robomimic
 eigenpy==3.10.3
     # via coal-library
 einops==0.8.1
-    # via
-    #   lerobot
-    #   libero
+    # via lerobot
 eiquadprog==1.2.9
     # via placo
-etils[epath,epy]==1.13.0
-    # via mujoco
 exceptiongroup==1.3.0
     # via
-    #   anyio
     #   ipython
     #   pytest
-executing==2.2.1
+executing==2.2.0
     # via stack-data
 farama-notifications==0.0.4
     # via gymnasium
-fastapi==0.119.1
-    # via teleop
-fastjsonschema==2.21.2
-    # via nbformat
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.20.0
+filelock==3.18.0
     # via
     #   datasets
     #   diffusers
@@ -184,25 +152,24 @@ filelock==3.20.0
     #   torch
     #   transformers
     #   virtualenv
-fonttools==4.60.1
+flask==3.1.1
+    # via lerobot
+fonttools==4.59.0
     # via matplotlib
-frozenlist==1.8.0
+frozenlist==1.7.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2025.9.0
+fsspec[http]==2025.3.0
     # via
     #   datasets
-    #   etils
     #   huggingface-hub
     #   torch
-future==1.0.0
-    # via libero
 gitdb==4.0.12
     # via gitpython
 gitpython==3.1.45
     # via wandb
-glfw==2.10.0
+glfw==2.9.0
     # via
     #   dm-control
     #   mujoco
@@ -210,79 +177,61 @@ grpcio==1.73.1
     # via
     #   grpcio-tools
     #   lerobot
-    #   reachy2-sdk
-    #   reachy2-sdk-api
-    #   tensorboard
 grpcio-tools==1.73.1
-    # via
-    #   lerobot
-    #   reachy2-sdk-api
-gym-aloha==0.1.3
     # via lerobot
-gym-hil==0.1.13
+gym-aloha==0.1.1
     # via lerobot
-gym-pusht==0.1.6
+gym-hil==0.1.10
     # via lerobot
-gymnasium==1.2.1
+gym-pusht==0.1.5
+    # via lerobot
+gym-xarm==0.1.1
+    # via lerobot
+gymnasium==0.29.1
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
+    #   gym-xarm
+    #   gymnasium-robotics
     #   lerobot
-    #   libero
-    #   metaworld
-h11==0.16.0
-    # via uvicorn
-h5py==3.15.1
-    # via robomimic
-hebi-py==2.11.0
-    # via lerobot
+    #   pettingzoo
+gymnasium-robotics==1.2.4
+    # via gym-xarm
 hf-transfer==0.1.9
     # via huggingface-hub
-hf-xet==1.1.10
+hf-xet==1.1.5
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
-httptools==0.7.1
-    # via uvicorn
-huggingface-hub[cli,hf-transfer]==0.35.3
+huggingface-hub[cli,hf-transfer]==0.34.3
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
-    #   peft
-    #   timm
     #   tokenizers
     #   transformers
-hydra-core==1.3.2
-    # via libero
-identify==2.6.15
+identify==2.6.12
     # via pre-commit
-idna==3.11
+idna==3.10
     # via
-    #   anyio
     #   requests
     #   yarl
 imageio[ffmpeg]==2.37.0
     # via
     #   gym-aloha
     #   gym-hil
+    #   gymnasium-robotics
     #   lerobot
-    #   metaworld
-    #   robomimic
     #   scikit-image
 imageio-ffmpeg==0.6.0
-    # via
-    #   imageio
-    #   robomimic
+    # via imageio
 importlib-metadata==8.7.0
     # via diffusers
-importlib-resources==6.5.2
-    # via etils
-iniconfig==2.3.0
+iniconfig==2.1.0
     # via pytest
 inquirerpy==0.3.4
     # via huggingface-hub
@@ -290,71 +239,50 @@ ipython==8.37.0
     # via meshcat
 ischedule==1.2.7
     # via placo
+itsdangerous==2.2.0
+    # via flask
 jedi==0.19.2
     # via ipython
 jinja2==3.1.6
-    # via torch
+    # via
+    #   flask
+    #   gymnasium-robotics
+    #   torch
 jsonlines==4.0.0
     # via lerobot
-jsonschema==4.25.1
-    # via nbformat
-jsonschema-specifications==2025.9.1
-    # via jsonschema
-jupyter-core==5.9.1
-    # via nbformat
-jupytext==1.18.1
-    # via bddl
-kiwisolver==1.4.9
+kiwisolver==1.4.8
     # via matplotlib
 labmaze==1.0.6
     # via dm-control
 lazy-loader==0.4
     # via scikit-image
-libero @ git+https://github.com/huggingface/lerobot-libero.git@main
-    # via lerobot
-llvmlite==0.45.1
-    # via numba
-lxml==6.0.2
+lxml==6.0.0
     # via dm-control
-markdown==3.9
-    # via tensorboard
-markdown-it-py==4.0.0
-    # via
-    #   jupytext
-    #   mdit-py-plugins
-markupsafe==3.0.3
+markupsafe==3.0.2
     # via
+    #   flask
     #   jinja2
     #   werkzeug
-matplotlib==3.10.7
-    # via
-    #   lerobot
-    #   libero
-matplotlib-inline==0.2.1
+matplotlib==3.10.5
+    # via lerobot
+matplotlib-inline==0.1.7
     # via ipython
-mdit-py-plugins==0.5.0
-    # via jupytext
-mdurl==0.1.2
-    # via markdown-it-py
 mergedeep==1.3.4
     # via draccus
 meshcat==0.3.2
     # via placo
-metaworld==3.0.0
-    # via lerobot
 mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==3.3.7
+mujoco==2.3.7
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
-    #   libero
-    #   metaworld
-    #   robosuite
-multidict==6.7.0
+    #   gym-xarm
+    #   gymnasium-robotics
+multidict==6.6.3
     # via
     #   aiohttp
     #   yarl
@@ -362,25 +290,17 @@ multiprocess==0.70.16
     # via datasets
 mypy-extensions==1.1.0
     # via typing-inspect
-nbformat==5.10.4
-    # via jupytext
 networkx==3.4.2
     # via
-    #   bddl
     #   scikit-image
     #   torch
-ninja==1.13.0
-    # via lerobot
 nodeenv==1.9.1
     # via pre-commit
 num2words==0.5.14
     # via lerobot
-numba==0.62.1
-    # via robosuite
 numpy==2.2.6
     # via
     #   accelerate
-    #   bddl
     #   cmeel-boost
     #   contourpy
     #   datasets
@@ -389,43 +309,25 @@ numpy==2.2.6
     #   dm-env
     #   dm-tree
     #   gymnasium
-    #   h5py
-    #   hebi-py
+    #   gymnasium-robotics
     #   imageio
     #   labmaze
-    #   libero
     #   matplotlib
     #   meshcat
-    #   metaworld
     #   mujoco
-    #   numba
     #   opencv-python
     #   opencv-python-headless
     #   pandas
-    #   peft
-    #   pyquaternion
-    #   reachy2-sdk
+    #   pettingzoo
     #   rerun-sdk
-    #   robomimic
-    #   robosuite
     #   scikit-image
     #   scipy
     #   shapely
-    #   teleop
-    #   tensorboard
-    #   tensorboardx
     #   tifffile
     #   torchvision
     #   transformers
-    #   transforms3d
-omegaconf==2.3.0
-    # via hydra-core
 opencv-python==4.12.0.88
-    # via
-    #   gym-pusht
-    #   libero
-    #   reachy2-sdk
-    #   robosuite
+    # via gym-pusht
 opencv-python-headless==4.12.0.88
     # via lerobot
 orderly-set==5.5.0
@@ -435,63 +337,53 @@ packaging==25.0
     #   accelerate
     #   datasets
     #   huggingface-hub
-    #   hydra-core
-    #   jupytext
     #   lazy-loader
     #   lerobot
     #   matplotlib
-    #   peft
     #   pytest
-    #   reachy2-sdk
     #   scikit-image
-    #   tensorboard
-    #   tensorboardx
     #   transformers
     #   wandb
-pandas==2.3.3
+pandas==2.3.1
     # via
     #   datasets
     #   lerobot
-parso==0.8.5
+parso==0.8.4
     # via jedi
-peft==0.17.1
-    # via lerobot
+pettingzoo==1.24.3
+    # via gymnasium-robotics
 pexpect==4.9.0
     # via ipython
 pfzy==0.3.4
     # via inquirerpy
-pillow==12.0.0
+pillow==11.3.0
     # via
     #   diffusers
     #   imageio
-    #   lerobot
     #   matplotlib
     #   meshcat
     #   rerun-sdk
-    #   robosuite
     #   scikit-image
-    #   tensorboard
     #   torchvision
 pin==3.4.0
     # via placo
 placo==0.9.14
     # via lerobot
-platformdirs==4.5.0
+platformdirs==4.3.8
     # via
-    #   jupyter-core
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.3.0
+pre-commit==4.2.0
     # via lerobot
-prompt-toolkit==3.0.52
+prompt-toolkit==3.0.51
     # via
     #   inquirerpy
     #   ipython
-propcache==0.4.1
+propcache==0.3.2
     # via
     #   aiohttp
     #   yarl
@@ -500,17 +392,11 @@ protobuf==6.31.0
     #   dm-control
     #   grpcio-tools
     #   lerobot
-    #   reachy2-sdk
-    #   reachy2-sdk-api
-    #   tensorboard
-    #   tensorboardx
     #   wandb
-psutil==7.1.1
+psutil==7.0.0
     # via
     #   accelerate
     #   imageio
-    #   peft
-    #   robomimic
 ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
@@ -519,13 +405,11 @@ pyarrow==21.0.0
     # via
     #   datasets
     #   rerun-sdk
-pycparser==2.23
+pycparser==2.22
     # via cffi
-pydantic==2.12.3
-    # via
-    #   fastapi
-    #   wandb
-pydantic-core==2.41.4
+pydantic==2.11.7
+    # via wandb
+pydantic-core==2.33.2
     # via pydantic
 pygame==2.6.1
     # via
@@ -540,42 +424,40 @@ pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.4.1
+pyngrok==7.2.12
     # via meshcat
 pynput==1.8.1
     # via
     #   gym-hil
     #   lerobot
-pyobjc-core==12.0
+pyobjc-core==11.1
     # via
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-cocoa
     #   pyobjc-framework-coretext
     #   pyobjc-framework-quartz
-pyobjc-framework-applicationservices==12.0
+pyobjc-framework-applicationservices==11.1
     # via pynput
-pyobjc-framework-cocoa==12.0
+pyobjc-framework-cocoa==11.1
     # via
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-coretext
     #   pyobjc-framework-quartz
-pyobjc-framework-coretext==12.0
+pyobjc-framework-coretext==11.1
     # via pyobjc-framework-applicationservices
-pyobjc-framework-quartz==12.0
+pyobjc-framework-quartz==11.1
     # via
     #   pynput
     #   pyobjc-framework-applicationservices
     #   pyobjc-framework-coretext
-pyopengl==3.1.10
+pyopengl==3.1.9
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.2.5
+pyparsing==3.2.3
     # via
     #   dm-control
     #   matplotlib
-pyquaternion==0.9.9
-    # via reachy2-sdk
 pyrealsense2-macosx==2.54.2
     # via lerobot
 pyserial==3.5
@@ -583,14 +465,12 @@ pyserial==3.5
     #   dynamixel-sdk
     #   feetech-servo-sdk
     #   lerobot
-pytest==8.4.2
+pytest==8.4.1
     # via
-    #   bddl
     #   lerobot
     #   pytest-cov
     #   pytest-timeout
-    #   teleop
-pytest-cov==7.0.0
+pytest-cov==6.2.1
     # via lerobot
 pytest-timeout==2.4.0
     # via lerobot
@@ -598,73 +478,46 @@ python-dateutil==2.9.0.post0
     # via
     #   matplotlib
     #   pandas
-python-dotenv==1.1.1
-    # via uvicorn
 pytz==2025.2
     # via pandas
-pyyaml==6.0.3
+pyyaml==6.0.2
     # via
     #   accelerate
     #   datasets
     #   draccus
-    #   hebi-py
     #   huggingface-hub
-    #   jupytext
-    #   omegaconf
-    #   peft
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
-    #   timm
     #   transformers
-    #   uvicorn
     #   wandb
 pyyaml-include==1.4.1
     # via draccus
-pyzmq==27.1.0
+pyzmq==27.0.0
     # via
     #   lerobot
     #   meshcat
-reachy2-sdk==1.0.14
-    # via lerobot
-reachy2-sdk-api==1.0.21
-    # via reachy2-sdk
-referencing==0.37.0
-    # via
-    #   jsonschema
-    #   jsonschema-specifications
-regex==2025.10.23
+regex==2025.7.34
     # via
     #   diffusers
     #   transformers
-requests==2.32.5
+requests==2.32.4
     # via
     #   datasets
     #   diffusers
     #   dm-control
     #   huggingface-hub
-    #   teleop
     #   transformers
     #   wandb
-rerun-sdk==0.26.1
+rerun-sdk==0.22.1
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
-robomimic==0.2.0
-    # via libero
-robosuite==1.4.0
-    # via libero
-rpds-py==0.28.0
-    # via
-    #   jsonschema
-    #   referencing
-safetensors==0.6.2
+safetensors==0.5.3
     # via
     #   accelerate
     #   diffusers
     #   lerobot
-    #   peft
-    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
@@ -673,12 +526,10 @@ scikit-image==0.25.2
 scipy==1.15.3
     # via
     #   dm-control
-    #   metaworld
-    #   robosuite
     #   scikit-image
-sentry-sdk==2.42.1
+sentry-sdk==2.34.1
     # via wandb
-shapely==2.1.2
+shapely==2.1.1
     # via gym-pusht
 six==1.17.0
     # via
@@ -686,106 +537,64 @@ six==1.17.0
     #   python-dateutil
 smmap==5.0.2
     # via gitdb
-sniffio==1.3.1
-    # via anyio
 stack-data==0.6.3
     # via ipython
-starlette==0.48.0
-    # via fastapi
 sympy==1.14.0
     # via torch
-teleop==0.1.2
-    # via lerobot
-tensorboard==2.20.0
-    # via robomimic
-tensorboard-data-server==0.7.2
-    # via tensorboard
-tensorboardx==2.6.4
-    # via robomimic
 termcolor==3.1.0
-    # via
-    #   lerobot
-    #   robomimic
-thop==0.1.1.post2209072238
-    # via libero
+    # via lerobot
 tifffile==2025.5.10
     # via scikit-image
-timm==1.0.20
-    # via lerobot
-tokenizers==0.22.1
+tokenizers==0.21.4
     # via transformers
 toml==0.10.2
     # via draccus
-tomli==2.3.0
+tomli==2.2.1
     # via
     #   cmeel
     #   coverage
-    #   jupytext
     #   pytest
 torch==2.7.1
     # via
     #   accelerate
     #   lerobot
-    #   peft
-    #   robomimic
-    #   thop
-    #   timm
     #   torchvision
 torchcodec==0.5
     # via lerobot
 torchvision==0.22.1
-    # via
-    #   lerobot
-    #   robomimic
-    #   timm
-tornado==6.5.2
+    # via lerobot
+tornado==6.5.1
     # via meshcat
 tqdm==4.67.1
     # via
     #   datasets
     #   dm-control
     #   huggingface-hub
-    #   peft
-    #   robomimic
     #   transformers
 traitlets==5.14.3
     # via
     #   ipython
-    #   jupyter-core
     #   matplotlib-inline
-    #   nbformat
-transformers==4.57.1
-    # via
-    #   lerobot
-    #   libero
-    #   peft
-transforms3d==0.4.2
-    # via teleop
-typing-extensions==4.15.0
+transformers==4.51.3
+    # via lerobot
+typing-extensions==4.14.1
     # via
     #   aiosignal
-    #   anyio
-    #   etils
     #   exceptiongroup
-    #   fastapi
     #   gymnasium
     #   huggingface-hub
     #   ipython
     #   multidict
     #   pydantic
     #   pydantic-core
-    #   referencing
     #   rerun-sdk
-    #   starlette
     #   torch
     #   typing-inspect
     #   typing-inspection
-    #   uvicorn
-    #   virtualenv
     #   wandb
 typing-inspect==0.9.0
     # via draccus
-typing-inspection==0.4.2
+typing-inspection==0.4.1
     # via pydantic
 tzdata==2025.2
     # via pandas
@@ -795,36 +604,22 @@ urllib3==2.5.0
     # via
     #   requests
     #   sentry-sdk
-uvicorn[standard]==0.38.0
-    # via teleop
-uvloop==0.22.1
-    # via uvicorn
-virtualenv==20.35.3
+virtualenv==20.32.0
     # via pre-commit
-wandb==0.21.4
-    # via
-    #   lerobot
-    #   libero
-watchfiles==1.1.1
-    # via uvicorn
-wcwidth==0.2.14
+wandb==0.21.0
+    # via lerobot
+wcwidth==0.2.13
     # via prompt-toolkit
-websocket-client==1.9.0
-    # via teleop
-websockets==15.0.1
-    # via uvicorn
 werkzeug==3.1.3
-    # via tensorboard
-wrapt==2.0.0
+    # via flask
+wrapt==1.17.2
     # via dm-tree
-xxhash==3.6.0
+xxhash==3.5.0
     # via datasets
-yarl==1.22.0
+yarl==1.20.1
     # via aiohttp
 zipp==3.23.0
-    # via
-    #   etils
-    #   importlib-metadata
+    # via importlib-metadata
 
 # The following packages are considered to be unsafe in a requirements file:
 # setuptools

requirements-ubuntu.txt

@@ -13,62 +13,47 @@ absl-py==2.3.1
     #   dm-tree
     #   labmaze
     #   mujoco
-    #   tensorboard
-accelerate==1.11.0
-    # via
-    #   lerobot
-    #   peft
+accelerate==1.9.0
+    # via lerobot
 aiohappyeyeballs==2.6.1
     # via aiohttp
-aiohttp==3.13.1
+aiohttp==3.12.15
     # via fsspec
 aiosignal==1.4.0
     # via aiohttp
 annotated-types==0.7.0
     # via pydantic
-antlr4-python3-runtime==4.9.3
-    # via
-    #   hydra-core
-    #   omegaconf
-anyio==4.11.0
-    # via
-    #   starlette
-    #   watchfiles
 asttokens==3.0.0
     # via stack-data
 async-timeout==5.0.1
     # via aiohttp
-attrs==25.4.0
+attrs==25.3.0
     # via
     #   aiohttp
     #   dm-tree
     #   jsonlines
-    #   jsonschema
-    #   referencing
     #   rerun-sdk
-av==15.1.0
+av==15.0.0
     # via lerobot
-bddl==1.0.1
-    # via libero
-certifi==2025.10.5
+blinker==1.9.0
+    # via flask
+certifi==2025.7.14
     # via
     #   requests
     #   sentry-sdk
-cffi==2.0.0
+cffi==1.17.1
     # via pymunk
 cfgv==3.4.0
     # via pre-commit
-charset-normalizer==3.4.4
+charset-normalizer==3.4.2
     # via requests
-click==8.3.0
+click==8.2.1
     # via
-    #   uvicorn
+    #   flask
     #   wandb
 cloudpickle==3.1.1
-    # via
-    #   gymnasium
-    #   libero
-cmake==4.1.0
+    # via gymnasium
+cmake==4.0.3
     # via lerobot
 cmeel==0.57.3
     # via
@@ -110,29 +95,27 @@ coal-library==3.0.1
     # via pin
 contourpy==1.3.2
     # via matplotlib
-coverage[toml]==7.11.0
+coverage[toml]==7.10.1
     # via pytest-cov
 cycler==0.12.1
     # via matplotlib
-datasets==4.1.1
+datasets==3.6.0
     # via lerobot
-debugpy==1.8.17
+debugpy==1.8.15
     # via lerobot
 decorator==5.2.1
     # via ipython
-decord==0.6.0
+deepdiff==8.5.0
     # via lerobot
-deepdiff==8.6.1
+diffusers==0.34.0
     # via lerobot
-diffusers==0.35.2
-    # via lerobot
-dill==0.4.0
+dill==0.3.8
     # via
     #   datasets
     #   multiprocess
 distlib==0.4.0
     # via virtualenv
-dm-control==1.0.34
+dm-control==1.0.14
     # via gym-aloha
 dm-env==1.6
     # via dm-control
@@ -140,48 +123,31 @@ dm-tree==0.1.9
     # via
     #   dm-control
     #   dm-env
-    #   lerobot
 docopt==0.6.2
     # via num2words
 draccus==0.10.0
     # via lerobot
-dynamixel-sdk==3.8.4
+dynamixel-sdk==3.7.31
     # via lerobot
-easydict==1.13
-    # via libero
-egl-probe @ git+https://github.com/huggingface/egl_probe.git
-    # via
-    #   libero
-    #   robomimic
 eigenpy==3.10.3
     # via coal-library
 einops==0.8.1
-    # via
-    #   flash-attn
-    #   lerobot
-    #   libero
+    # via lerobot
 eiquadprog==1.2.9
     # via placo
-etils[epath,epy]==1.13.0
-    # via mujoco
 evdev==1.9.2
     # via pynput
 exceptiongroup==1.3.0
     # via
-    #   anyio
     #   ipython
     #   pytest
-executing==2.2.1
+executing==2.2.0
     # via stack-data
 farama-notifications==0.0.4
     # via gymnasium
-fastapi==0.119.1
-    # via teleop
-fastjsonschema==2.21.2
-    # via nbformat
 feetech-servo-sdk==1.0.0
     # via lerobot
-filelock==3.20.0
+filelock==3.18.0
     # via
     #   datasets
     #   diffusers
@@ -189,27 +155,24 @@ filelock==3.20.0
     #   torch
     #   transformers
     #   virtualenv
-flash-attn==2.8.3
+flask==3.1.1
     # via lerobot
-fonttools==4.60.1
+fonttools==4.59.0
     # via matplotlib
-frozenlist==1.8.0
+frozenlist==1.7.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec[http]==2025.9.0
+fsspec[http]==2025.3.0
     # via
     #   datasets
-    #   etils
     #   huggingface-hub
     #   torch
-future==1.0.0
-    # via libero
 gitdb==4.0.12
     # via gitpython
 gitpython==3.1.45
     # via wandb
-glfw==2.10.0
+glfw==2.9.0
     # via
     #   dm-control
     #   mujoco
@@ -217,79 +180,61 @@ grpcio==1.73.1
     # via
     #   grpcio-tools
     #   lerobot
-    #   reachy2-sdk
-    #   reachy2-sdk-api
-    #   tensorboard
 grpcio-tools==1.73.1
-    # via
-    #   lerobot
-    #   reachy2-sdk-api
-gym-aloha==0.1.3
     # via lerobot
-gym-hil==0.1.13
+gym-aloha==0.1.1
     # via lerobot
-gym-pusht==0.1.6
+gym-hil==0.1.10
     # via lerobot
-gymnasium==1.2.1
+gym-pusht==0.1.5
+    # via lerobot
+gym-xarm==0.1.1
+    # via lerobot
+gymnasium==0.29.1
     # via
     #   gym-aloha
     #   gym-hil
     #   gym-pusht
+    #   gym-xarm
+    #   gymnasium-robotics
     #   lerobot
-    #   libero
-    #   metaworld
-h11==0.16.0
-    # via uvicorn
-h5py==3.15.1
-    # via robomimic
-hebi-py==2.11.0
-    # via lerobot
+    #   pettingzoo
+gymnasium-robotics==1.2.4
+    # via gym-xarm
 hf-transfer==0.1.9
     # via huggingface-hub
-hf-xet==1.1.10
+hf-xet==1.1.5
     # via huggingface-hub
 hidapi==0.14.0.post4
     # via
     #   gym-hil
     #   lerobot
-httptools==0.7.1
-    # via uvicorn
-huggingface-hub[cli,hf-transfer]==0.35.3
+huggingface-hub[cli,hf-transfer]==0.34.3
     # via
     #   accelerate
     #   datasets
     #   diffusers
     #   lerobot
-    #   peft
-    #   timm
     #   tokenizers
     #   transformers
-hydra-core==1.3.2
-    # via libero
-identify==2.6.15
+identify==2.6.12
     # via pre-commit
-idna==3.11
+idna==3.10
     # via
-    #   anyio
     #   requests
     #   yarl
 imageio[ffmpeg]==2.37.0
     # via
     #   gym-aloha
     #   gym-hil
+    #   gymnasium-robotics
     #   lerobot
-    #   metaworld
-    #   robomimic
     #   scikit-image
 imageio-ffmpeg==0.6.0
-    # via
-    #   imageio
-    #   robomimic
+    # via imageio
 importlib-metadata==8.7.0
     # via diffusers
-importlib-resources==6.5.2
-    # via etils
-iniconfig==2.3.0
+iniconfig==2.1.0
     # via pytest
 inquirerpy==0.3.4
     # via huggingface-hub
@@ -297,71 +242,50 @@ ipython==8.37.0
     # via meshcat
 ischedule==1.2.7
     # via placo
+itsdangerous==2.2.0
+    # via flask
 jedi==0.19.2
     # via ipython
 jinja2==3.1.6
-    # via torch
+    # via
+    #   flask
+    #   gymnasium-robotics
+    #   torch
 jsonlines==4.0.0
     # via lerobot
-jsonschema==4.25.1
-    # via nbformat
-jsonschema-specifications==2025.9.1
-    # via jsonschema
-jupyter-core==5.9.1
-    # via nbformat
-jupytext==1.18.1
-    # via bddl
-kiwisolver==1.4.9
+kiwisolver==1.4.8
     # via matplotlib
 labmaze==1.0.6
     # via dm-control
 lazy-loader==0.4
     # via scikit-image
-libero @ git+https://github.com/huggingface/lerobot-libero.git@main
-    # via lerobot
-llvmlite==0.45.1
-    # via numba
-lxml==6.0.2
+lxml==6.0.0
     # via dm-control
-markdown==3.9
-    # via tensorboard
-markdown-it-py==4.0.0
-    # via
-    #   jupytext
-    #   mdit-py-plugins
-markupsafe==3.0.3
+markupsafe==3.0.2
     # via
+    #   flask
     #   jinja2
     #   werkzeug
-matplotlib==3.10.7
-    # via
-    #   lerobot
-    #   libero
-matplotlib-inline==0.2.1
+matplotlib==3.10.5
+    # via lerobot
+matplotlib-inline==0.1.7
     # via ipython
-mdit-py-plugins==0.5.0
-    # via jupytext
-mdurl==0.1.2
-    # via markdown-it-py
 mergedeep==1.3.4
     # via draccus
 meshcat==0.3.2
     # via placo
-metaworld==3.0.0
-    # via lerobot
 mock-serial==0.0.1
     # via lerobot
 mpmath==1.3.0
     # via sympy
-mujoco==3.3.7
+mujoco==2.3.7
     # via
     #   dm-control
     #   gym-aloha
     #   gym-hil
-    #   libero
-    #   metaworld
-    #   robosuite
-multidict==6.7.0
+    #   gym-xarm
+    #   gymnasium-robotics
+multidict==6.6.3
     # via
     #   aiohttp
     #   yarl
@@ -369,63 +293,42 @@ multiprocess==0.70.16
     # via datasets
 mypy-extensions==1.1.0
     # via typing-inspect
-nbformat==5.10.4
-    # via jupytext
 networkx==3.4.2
     # via
-    #   bddl
     #   scikit-image
     #   torch
-ninja==1.13.0
-    # via lerobot
 nodeenv==1.9.1
     # via pre-commit
 num2words==0.5.14
     # via lerobot
-numba==0.62.1
-    # via robosuite
 numpy==2.2.6
     # via
     #   accelerate
-    #   bddl
     #   cmeel-boost
     #   contourpy
     #   datasets
-    #   decord
     #   diffusers
     #   dm-control
     #   dm-env
     #   dm-tree
     #   gymnasium
-    #   h5py
-    #   hebi-py
+    #   gymnasium-robotics
     #   imageio
     #   labmaze
-    #   libero
     #   matplotlib
     #   meshcat
-    #   metaworld
     #   mujoco
-    #   numba
     #   opencv-python
     #   opencv-python-headless
     #   pandas
-    #   peft
-    #   pyquaternion
-    #   reachy2-sdk
+    #   pettingzoo
     #   rerun-sdk
-    #   robomimic
-    #   robosuite
     #   scikit-image
     #   scipy
     #   shapely
-    #   teleop
-    #   tensorboard
-    #   tensorboardx
     #   tifffile
     #   torchvision
     #   transformers
-    #   transforms3d
 nvidia-cublas-cu12==12.6.4.1
     # via
     #   nvidia-cudnn-cu12
@@ -463,14 +366,8 @@ nvidia-nvjitlink-cu12==12.6.85
     #   torch
 nvidia-nvtx-cu12==12.6.77
     # via torch
-omegaconf==2.3.0
-    # via hydra-core
 opencv-python==4.12.0.88
-    # via
-    #   gym-pusht
-    #   libero
-    #   reachy2-sdk
-    #   robosuite
+    # via gym-pusht
 opencv-python-headless==4.12.0.88
     # via lerobot
 orderly-set==5.5.0
@@ -480,63 +377,53 @@ packaging==25.0
     #   accelerate
     #   datasets
     #   huggingface-hub
-    #   hydra-core
-    #   jupytext
     #   lazy-loader
     #   lerobot
     #   matplotlib
-    #   peft
     #   pytest
-    #   reachy2-sdk
     #   scikit-image
-    #   tensorboard
-    #   tensorboardx
     #   transformers
     #   wandb
-pandas==2.3.3
+pandas==2.3.1
     # via
     #   datasets
     #   lerobot
-parso==0.8.5
+parso==0.8.4
     # via jedi
-peft==0.17.1
-    # via lerobot
+pettingzoo==1.24.3
+    # via gymnasium-robotics
 pexpect==4.9.0
     # via ipython
 pfzy==0.3.4
     # via inquirerpy
-pillow==12.0.0
+pillow==11.3.0
     # via
     #   diffusers
     #   imageio
-    #   lerobot
     #   matplotlib
     #   meshcat
     #   rerun-sdk
-    #   robosuite
     #   scikit-image
-    #   tensorboard
     #   torchvision
 pin==3.4.0
     # via placo
 placo==0.9.14
     # via lerobot
-platformdirs==4.5.0
+platformdirs==4.3.8
     # via
-    #   jupyter-core
     #   virtualenv
     #   wandb
 pluggy==1.6.0
     # via
     #   pytest
     #   pytest-cov
-pre-commit==4.3.0
+pre-commit==4.2.0
     # via lerobot
-prompt-toolkit==3.0.52
+prompt-toolkit==3.0.51
     # via
     #   inquirerpy
     #   ipython
-propcache==0.4.1
+propcache==0.3.2
     # via
     #   aiohttp
     #   yarl
@@ -545,17 +432,11 @@ protobuf==6.31.0
     #   dm-control
     #   grpcio-tools
     #   lerobot
-    #   reachy2-sdk
-    #   reachy2-sdk-api
-    #   tensorboard
-    #   tensorboardx
     #   wandb
-psutil==7.1.1
+psutil==7.0.0
     # via
     #   accelerate
     #   imageio
-    #   peft
-    #   robomimic
 ptyprocess==0.7.0
     # via pexpect
 pure-eval==0.2.3
@@ -564,13 +445,11 @@ pyarrow==21.0.0
     # via
     #   datasets
     #   rerun-sdk
-pycparser==2.23
+pycparser==2.22
     # via cffi
-pydantic==2.12.3
-    # via
-    #   fastapi
-    #   wandb
-pydantic-core==2.41.4
+pydantic==2.11.7
+    # via wandb
+pydantic-core==2.33.2
     # via pydantic
 pygame==2.6.1
     # via
@@ -585,22 +464,20 @@ pymunk==6.11.1
     # via
     #   gym-pusht
     #   lerobot
-pyngrok==7.4.1
+pyngrok==7.2.12
     # via meshcat
 pynput==1.8.1
     # via
     #   gym-hil
     #   lerobot
-pyopengl==3.1.10
+pyopengl==3.1.9
     # via
     #   dm-control
     #   mujoco
-pyparsing==3.2.5
+pyparsing==3.2.3
     # via
     #   dm-control
     #   matplotlib
-pyquaternion==0.9.9
-    # via reachy2-sdk
 pyrealsense2==2.56.5.9235
     # via lerobot
 pyserial==3.5
@@ -608,14 +485,12 @@ pyserial==3.5
     #   dynamixel-sdk
     #   feetech-servo-sdk
     #   lerobot
-pytest==8.4.2
+pytest==8.4.1
     # via
-    #   bddl
     #   lerobot
     #   pytest-cov
     #   pytest-timeout
-    #   teleop
-pytest-cov==7.0.0
+pytest-cov==6.2.1
     # via lerobot
 pytest-timeout==2.4.0
     # via lerobot
@@ -623,75 +498,48 @@ python-dateutil==2.9.0.post0
     # via
     #   matplotlib
     #   pandas
-python-dotenv==1.1.1
-    # via uvicorn
 python-xlib==0.33
     # via pynput
 pytz==2025.2
     # via pandas
-pyyaml==6.0.3
+pyyaml==6.0.2
     # via
     #   accelerate
     #   datasets
     #   draccus
-    #   hebi-py
     #   huggingface-hub
-    #   jupytext
-    #   omegaconf
-    #   peft
     #   pre-commit
     #   pyngrok
     #   pyyaml-include
-    #   timm
     #   transformers
-    #   uvicorn
     #   wandb
 pyyaml-include==1.4.1
     # via draccus
-pyzmq==27.1.0
+pyzmq==27.0.0
     # via
     #   lerobot
     #   meshcat
-reachy2-sdk==1.0.14
-    # via lerobot
-reachy2-sdk-api==1.0.21
-    # via reachy2-sdk
-referencing==0.37.0
-    # via
-    #   jsonschema
-    #   jsonschema-specifications
-regex==2025.10.23
+regex==2025.7.34
     # via
     #   diffusers
     #   transformers
-requests==2.32.5
+requests==2.32.4
     # via
     #   datasets
     #   diffusers
     #   dm-control
     #   huggingface-hub
-    #   teleop
     #   transformers
     #   wandb
-rerun-sdk==0.26.1
+rerun-sdk==0.22.1
     # via lerobot
 rhoban-cmeel-jsoncpp==1.9.4.9
     # via placo
-robomimic==0.2.0
-    # via libero
-robosuite==1.4.0
-    # via libero
-rpds-py==0.28.0
-    # via
-    #   jsonschema
-    #   referencing
-safetensors==0.6.2
+safetensors==0.5.3
     # via
     #   accelerate
     #   diffusers
     #   lerobot
-    #   peft
-    #   timm
     #   transformers
 scikit-image==0.25.2
     # via
@@ -700,12 +548,10 @@ scikit-image==0.25.2
 scipy==1.15.3
     # via
     #   dm-control
-    #   metaworld
-    #   robosuite
     #   scikit-image
-sentry-sdk==2.42.1
+sentry-sdk==2.34.1
     # via wandb
-shapely==2.1.2
+shapely==2.1.1
     # via gym-pusht
 six==1.17.0
     # via
@@ -714,109 +560,66 @@ six==1.17.0
     #   python-xlib
 smmap==5.0.2
     # via gitdb
-sniffio==1.3.1
-    # via anyio
 stack-data==0.6.3
     # via ipython
-starlette==0.48.0
-    # via fastapi
 sympy==1.14.0
     # via torch
-teleop==0.1.2
-    # via lerobot
-tensorboard==2.20.0
-    # via robomimic
-tensorboard-data-server==0.7.2
-    # via tensorboard
-tensorboardx==2.6.4
-    # via robomimic
 termcolor==3.1.0
-    # via
-    #   lerobot
-    #   robomimic
-thop==0.1.1.post2209072238
-    # via libero
+    # via lerobot
 tifffile==2025.5.10
     # via scikit-image
-timm==1.0.20
-    # via lerobot
-tokenizers==0.22.1
+tokenizers==0.21.4
     # via transformers
 toml==0.10.2
     # via draccus
-tomli==2.3.0
+tomli==2.2.1
     # via
     #   cmeel
     #   coverage
-    #   jupytext
     #   pytest
 torch==2.7.1
     # via
     #   accelerate
-    #   flash-attn
     #   lerobot
-    #   peft
-    #   robomimic
-    #   thop
-    #   timm
     #   torchvision
 torchcodec==0.5
     # via lerobot
 torchvision==0.22.1
-    # via
-    #   lerobot
-    #   robomimic
-    #   timm
-tornado==6.5.2
+    # via lerobot
+tornado==6.5.1
     # via meshcat
 tqdm==4.67.1
     # via
     #   datasets
     #   dm-control
     #   huggingface-hub
-    #   peft
-    #   robomimic
     #   transformers
 traitlets==5.14.3
     # via
     #   ipython
-    #   jupyter-core
     #   matplotlib-inline
-    #   nbformat
-transformers==4.57.1
-    # via
-    #   lerobot
-    #   libero
-    #   peft
-transforms3d==0.4.2
-    # via teleop
+transformers==4.51.3
+    # via lerobot
 triton==3.3.1
     # via torch
-typing-extensions==4.15.0
+typing-extensions==4.14.1
     # via
     #   aiosignal
-    #   anyio
-    #   etils
     #   exceptiongroup
-    #   fastapi
     #   gymnasium
     #   huggingface-hub
     #   ipython
     #   multidict
     #   pydantic
     #   pydantic-core
-    #   referencing
     #   rerun-sdk
-    #   starlette
     #   torch
     #   typing-inspect
     #   typing-inspection
-    #   uvicorn
-    #   virtualenv
     #   wandb
 typing-inspect==0.9.0
     # via draccus
-typing-inspection==0.4.2
+typing-inspection==0.4.1
     # via pydantic
 tzdata==2025.2
     # via pandas
@@ -826,36 +629,22 @@ urllib3==2.5.0
     # via
     #   requests
     #   sentry-sdk
-uvicorn[standard]==0.38.0
-    # via teleop
-uvloop==0.22.1
-    # via uvicorn
-virtualenv==20.35.3
+virtualenv==20.32.0
     # via pre-commit
-wandb==0.21.4
-    # via
-    #   lerobot
-    #   libero
-watchfiles==1.1.1
-    # via uvicorn
-wcwidth==0.2.14
+wandb==0.21.0
+    # via lerobot
+wcwidth==0.2.13
     # via prompt-toolkit
-websocket-client==1.9.0
-    # via teleop
-websockets==15.0.1
-    # via uvicorn
 werkzeug==3.1.3
-    # via tensorboard
-wrapt==2.0.0
+    # via flask
+wrapt==1.17.2
     # via dm-tree
-xxhash==3.6.0
+xxhash==3.5.0
     # via datasets
-yarl==1.22.0
+yarl==1.20.1
     # via aiohttp
 zipp==3.23.0
-    # via
-    #   etils
-    #   importlib-metadata
+    # via importlib-metadata
 
 # The following packages are considered to be unsafe in a requirements file:
 # setuptools

requirements.in

@@ -1,9 +1,9 @@
 # requirements.in
 
-# requirements-macos.txt was generated on macOS and is platform-specific (macOS 26.0.1 25A362 arm64).
-# Darwin MacBook-Pro.local 25.0.0 Darwin Kernel Version 25.0.0: Wed Sep 17 21:42:08 PDT 2025; root:xnu-12377.1.9~141/RELEASE_ARM64_T8132 arm64
+# requirements-macos.txt was generated on macOS and is platform-specific (macOS 15.5 24F74 arm64).
+# Darwin MacBook-Pro.local 24.5.0 Darwin Kernel Version 24.5.0: Tue Apr 22 19:54:43 PDT 2025; root:xnu-11417.121.6~2/RELEASE_ARM64_T8132 arm64
 
-# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.3 LTS x86_64).
-# Linux mlerobot-linux 6.14.0-33-generic #33~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Fri Sep 19 17:02:30 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
+# requirements-ubuntu.txt was generated on Linux and is platform-specific (Ubuntu 24.04.2 LTS x86_64).
+# Linux mlerobot-linux 6.14.0-27-generic #27~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Tue Jul 22 17:38:49 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
 
 -e .[all]

src/lerobot/__init__.py

@@ -57,6 +57,7 @@ available_tasks_per_env = {
         "AlohaTransferCube-v0",
     ],
     "pusht": ["PushT-v0"],
+    "xarm": ["XarmLift-v0"],
 }
 available_envs = list(available_tasks_per_env.keys())
 
@@ -74,6 +75,16 @@ available_datasets_per_env = {
     # TODO(alexander-soare): Add "lerobot/pusht_keypoints". Right now we can't because this is too tightly
     # coupled with tests.
     "pusht": ["lerobot/pusht", "lerobot/pusht_image"],
+    "xarm": [
+        "lerobot/xarm_lift_medium",
+        "lerobot/xarm_lift_medium_replay",
+        "lerobot/xarm_push_medium",
+        "lerobot/xarm_push_medium_replay",
+        "lerobot/xarm_lift_medium_image",
+        "lerobot/xarm_lift_medium_replay_image",
+        "lerobot/xarm_push_medium_image",
+        "lerobot/xarm_push_medium_replay_image",
+    ],
 }
 
 available_real_world_datasets = [
@@ -184,6 +195,7 @@ available_motors = [
 available_policies_per_env = {
     "aloha": ["act"],
     "pusht": ["diffusion", "vqbet"],
+    "xarm": ["tdmpc"],
     "koch_real": ["act_koch_real"],
     "aloha_real": ["act_aloha_real"],
 }

src/lerobot/calibrate.py

@@ -52,7 +52,6 @@ from lerobot.teleoperators import (  # noqa: F401
     so100_leader,
     so101_leader,
 )
-from lerobot.utils.import_utils import register_third_party_devices
 from lerobot.utils.utils import init_logging
 
 
@@ -84,7 +83,6 @@ def calibrate(cfg: CalibrateConfig):
 
 
 def main():
-    register_third_party_devices()
     calibrate()
 
 

src/lerobot/cameras/camera.py

@@ -17,7 +17,7 @@
 import abc
 from typing import Any
 
-from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
+import numpy as np
 
 from .configs import CameraConfig, ColorMode
 
@@ -89,7 +89,7 @@ class Camera(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
         """Capture and return a single frame from the camera.
 
         Args:
@@ -102,7 +102,7 @@ class Camera(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def async_read(self, timeout_ms: float = ...) -> NDArray[Any]:
+    def async_read(self, timeout_ms: float = ...) -> np.ndarray:
         """Asynchronously capture and return a single frame from the camera.
 
         Args:

src/lerobot/cameras/configs.py

@@ -18,7 +18,7 @@ import abc
 from dataclasses import dataclass
 from enum import Enum
 
-import draccus  # type: ignore  # TODO: add type stubs for draccus
+import draccus
 
 
 class ColorMode(str, Enum):
@@ -34,11 +34,11 @@ class Cv2Rotation(int, Enum):
 
 
 @dataclass(kw_only=True)
-class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus
+class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
     fps: int | None = None
     width: int | None = None
     height: int | None = None
 
     @property
     def type(self) -> str:
-        return str(self.get_choice_name(self.__class__))
+        return self.get_choice_name(self.__class__)

src/lerobot/cameras/opencv/__init__.py

@@ -14,5 +14,3 @@
 
 from .camera_opencv import OpenCVCamera
 from .configuration_opencv import OpenCVCameraConfig
-
-__all__ = ["OpenCVCamera", "OpenCVCameraConfig"]

src/lerobot/cameras/opencv/camera_opencv.py

@@ -23,18 +23,15 @@ import platform
 import time
 from pathlib import Path
 from threading import Event, Lock, Thread
-from multiprocessing import Process, Event as EventProcess, JoinableQueue as Queue
-from queue import Empty
 from typing import Any
 
-from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
-
 # Fix MSMF hardware transform compatibility for Windows before importing cv2
 if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS" not in os.environ:
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
-import cv2  # type: ignore  # TODO: add type stubs for OpenCV
+import cv2
+import numpy as np
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..utils import get_cv2_backend, get_cv2_rotation
@@ -121,10 +118,11 @@ class OpenCVCamera(Camera):
 
         self.videocapture: cv2.VideoCapture | None = None
 
-        self.process: Process | None = None
-        self.stop_event: EventProcess | None = None
-        self.frame_queue: Queue = Queue()
-        self.latest_frame: NDArray[Any] | None = None
+        self.thread: Thread | None = None
+        self.stop_event: Event | None = None
+        self.frame_lock: Lock = Lock()
+        self.latest_frame: np.ndarray | None = None
+        self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
         self.backend: int = get_cv2_backend()
@@ -142,7 +140,7 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
-    def connect(self, warmup: bool = True) -> None:
+    def connect(self, warmup: bool = True):
         """
         Connects to the OpenCV camera specified in the configuration.
 
@@ -182,14 +180,12 @@ class OpenCVCamera(Camera):
 
     def _configure_capture_settings(self) -> None:
         """
-        Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
+        Applies the specified FPS, width, and height settings to the connected camera.
 
         This method attempts to set the camera properties via OpenCV. It checks if
         the camera successfully applied the settings and raises an error if not.
-        FOURCC is set first (if specified) as it can affect the available FPS and resolution options.
 
         Args:
-            fourcc: The desired FOURCC code (e.g., "MJPG", "YUYV"). If None, auto-detect.
             fps: The desired frames per second. If None, the setting is skipped.
             width: The desired capture width. If None, the setting is skipped.
             height: The desired capture height. If None, the setting is skipped.
@@ -203,11 +199,10 @@ class OpenCVCamera(Camera):
         if not self.is_connected:
             raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
-        # Set FOURCC first (if specified) as it can affect available FPS/resolution options
-        if self.config.fourcc is not None:
-            self._validate_fourcc()
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+        if self.fps is None:
+            self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
+        else:
+            self._validate_fps()
 
         default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
         default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
@@ -221,56 +216,18 @@ class OpenCVCamera(Camera):
         else:
             self._validate_width_and_height()
 
-        if self.fps is None:
-            self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
-        else:
-            self._validate_fps()
-
     def _validate_fps(self) -> None:
         """Validates and sets the camera's frames per second (FPS)."""
 
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
-
-        if self.fps is None:
-            raise ValueError(f"{self} FPS is not set")
-
         success = self.videocapture.set(cv2.CAP_PROP_FPS, float(self.fps))
         actual_fps = self.videocapture.get(cv2.CAP_PROP_FPS)
         # Use math.isclose for robust float comparison
         if not success or not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
             raise RuntimeError(f"{self} failed to set fps={self.fps} ({actual_fps=}).")
 
-    def _validate_fourcc(self) -> None:
-        """Validates and sets the camera's FOURCC code."""
-
-        fourcc_code = cv2.VideoWriter_fourcc(*self.config.fourcc)
-
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
-
-        success = self.videocapture.set(cv2.CAP_PROP_FOURCC, fourcc_code)
-        actual_fourcc_code = self.videocapture.get(cv2.CAP_PROP_FOURCC)
-
-        # Convert actual FOURCC code back to string for comparison
-        actual_fourcc_code_int = int(actual_fourcc_code)
-        actual_fourcc = "".join([chr((actual_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
-
-        if not success or actual_fourcc != self.config.fourcc:
-            logger.warning(
-                f"{self} failed to set fourcc={self.config.fourcc} (actual={actual_fourcc}, success={success}). "
-                f"Continuing with default format."
-            )
-
     def _validate_width_and_height(self) -> None:
         """Validates and sets the camera's frame capture width and height."""
 
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
-
-        if self.capture_width is None or self.capture_height is None:
-            raise ValueError(f"{self} capture_width or capture_height is not set")
-
         width_success = self.videocapture.set(cv2.CAP_PROP_FRAME_WIDTH, float(self.capture_width))
         height_success = self.videocapture.set(cv2.CAP_PROP_FRAME_HEIGHT, float(self.capture_height))
 
@@ -301,12 +258,11 @@ class OpenCVCamera(Camera):
         """
         found_cameras_info = []
 
-        targets_to_scan: list[str | int]
         if platform.system() == "Linux":
             possible_paths = sorted(Path("/dev").glob("video*"), key=lambda p: p.name)
             targets_to_scan = [str(p) for p in possible_paths]
         else:
-            targets_to_scan = [int(i) for i in range(MAX_OPENCV_INDEX)]
+            targets_to_scan = list(range(MAX_OPENCV_INDEX))
 
         for target in targets_to_scan:
             camera = cv2.VideoCapture(target)
@@ -315,12 +271,6 @@ class OpenCVCamera(Camera):
                 default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
                 default_fps = camera.get(cv2.CAP_PROP_FPS)
                 default_format = camera.get(cv2.CAP_PROP_FORMAT)
-
-                # Get FOURCC code and convert to string
-                default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
-                default_fourcc_code_int = int(default_fourcc_code)
-                default_fourcc = "".join([chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
-
                 camera_info = {
                     "name": f"OpenCV Camera @ {target}",
                     "type": "OpenCV",
@@ -328,7 +278,6 @@ class OpenCVCamera(Camera):
                     "backend_api": camera.getBackendName(),
                     "default_stream_profile": {
                         "format": default_format,
-                        "fourcc": default_fourcc,
                         "width": default_width,
                         "height": default_height,
                         "fps": default_fps,
@@ -340,7 +289,7 @@ class OpenCVCamera(Camera):
 
         return found_cameras_info
 
-    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
         """
         Reads a single frame synchronously from the camera.
 
@@ -368,9 +317,6 @@ class OpenCVCamera(Camera):
 
         start_time = time.perf_counter()
 
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
-
         ret, frame = self.videocapture.read()
 
         if not ret or frame is None:
@@ -383,7 +329,7 @@ class OpenCVCamera(Camera):
 
         return processed_frame
 
-    def _postprocess_image(self, image: NDArray[Any], color_mode: ColorMode | None = None) -> NDArray[Any]:
+    def _postprocess_image(self, image: np.ndarray, color_mode: ColorMode | None = None) -> np.ndarray:
         """
         Applies color conversion, dimension validation, and rotation to a raw frame.
 
@@ -426,7 +372,7 @@ class OpenCVCamera(Camera):
 
         return processed_image
 
-    def _read_loop(self) -> None:
+    def _read_loop(self):
         """
         Internal loop run by the background thread for asynchronous reading.
 
@@ -437,45 +383,43 @@ class OpenCVCamera(Camera):
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
-        if self.stop_event is None:
-            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
-
         while not self.stop_event.is_set():
             try:
                 color_image = self.read()
-                self.frame_queue.put_nowait(color_image)
+
+                with self.frame_lock:
+                    self.latest_frame = color_image
+                self.new_frame_event.set()
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
                 logger.warning(f"Error reading frame in background thread for {self}: {e}")
 
-    def _start_read_process(self) -> None:
+    def _start_read_thread(self) -> None:
         """Starts or restarts the background read thread if it's not running."""
-        if self.process is not None and self.process.is_alive():
-            self.frame_queue.join()
-            self.process.join()
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=0.1)
         if self.stop_event is not None:
             self.stop_event.set()
 
         self.stop_event = Event()
-        self.process = Process(target=self._read_loop, args=(), name=f"{self}_read_loop")
-        self.process.daemon = True
-        self.process.start()
+        self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
+        self.thread.daemon = True
+        self.thread.start()
 
     def _stop_read_thread(self) -> None:
         """Signals the background read thread to stop and waits for it to join."""
         if self.stop_event is not None:
             self.stop_event.set()
 
-        if self.process is not None and self.process.is_alive():
-            self.frame_queue.join()
-            self.process.join()
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=2.0)
 
-        self.process = None
+        self.thread = None
         self.stop_event = None
 
-    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
+    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
         """
         Reads the latest available frame asynchronously.
 
@@ -499,34 +443,26 @@ class OpenCVCamera(Camera):
         if not self.is_connected:
             raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        if self.process is None or not self.process.is_alive():
-            self._start_read_process()
+        if self.thread is None or not self.thread.is_alive():
+            self._start_read_thread()
 
-        if self.latest_frame is None:
-            self.latest_frame = self.frame_queue.get()
-            self.frame_queue.task_done()
-            return self.latest_frame
+        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            thread_alive = self.thread is not None and self.thread.is_alive()
+            raise TimeoutError(
+                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
+                f"Read thread alive: {thread_alive}."
+            )
 
-        try:
-            frame = self.frame_queue.get(timeout=timeout_ms / 1000.0)
-            self.frame_queue.task_done()
-        except Empty:
-            process_alive = self.process is not None and self.process.is_alive()
-            if process_alive:
-                logger.warning(f"{self} async_read timed out after {timeout_ms} ms but camera is still running.")
-                return self.latest_frame
-            else:
-                raise TimeoutError(
-                    f"{self} async_read timed out after {timeout_ms} ms: camera is not responding !"
-                )
+        with self.frame_lock:
+            frame = self.latest_frame
+            self.new_frame_event.clear()
 
         if frame is None:
             raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
-        else:
-            self.latest_frame = frame
-            return self.latest_frame
 
-    def disconnect(self) -> None:
+        return frame
+
+    def disconnect(self):
         """
         Disconnects from the camera and cleans up resources.
 

src/lerobot/cameras/opencv/configuration_opencv.py

@@ -17,8 +17,6 @@ from pathlib import Path
 
 from ..configs import CameraConfig, ColorMode, Cv2Rotation
 
-__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
-
 
 @CameraConfig.register_subclass("opencv")
 @dataclass
@@ -35,9 +33,8 @@ class OpenCVCameraConfig(CameraConfig):
     OpenCVCameraConfig(0, 30, 1280, 720)   # 1280x720 @ 30FPS
     OpenCVCameraConfig(/dev/video4, 60, 640, 480)   # 640x480 @ 60FPS
 
-    # Advanced configurations with FOURCC format
-    OpenCVCameraConfig(128422271347, 30, 640, 480, rotation=Cv2Rotation.ROTATE_90, fourcc="MJPG")     # With 90° rotation and MJPG format
-    OpenCVCameraConfig(0, 30, 1280, 720, fourcc="YUYV")     # With YUYV format
+    # Advanced configurations
+    OpenCVCameraConfig(128422271347, 30, 640, 480, rotation=Cv2Rotation.ROTATE_90)     # With 90° rotation
     ```
 
     Attributes:
@@ -49,21 +46,17 @@ class OpenCVCameraConfig(CameraConfig):
         color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
-        fourcc: FOURCC code for video format (e.g., "MJPG", "YUYV", "I420"). Defaults to None (auto-detect).
 
     Note:
         - Only 3-channel color output (RGB/BGR) is currently supported.
-        - FOURCC codes must be 4-character strings (e.g., "MJPG", "YUYV"). Some common FOUCC codes: https://learn.microsoft.com/en-us/windows/win32/medfound/video-fourccs#fourcc-constants
-        - Setting FOURCC can help achieve higher frame rates on some cameras.
     """
 
     index_or_path: int | Path
     color_mode: ColorMode = ColorMode.RGB
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
-    fourcc: str | None = None
 
-    def __post_init__(self) -> None:
+    def __post_init__(self):
         if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
                 f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
@@ -78,8 +71,3 @@ class OpenCVCameraConfig(CameraConfig):
             raise ValueError(
                 f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
             )
-
-        if self.fourcc is not None and (not isinstance(self.fourcc, str) or len(self.fourcc) != 4):
-            raise ValueError(
-                f"`fourcc` must be a 4-character string (e.g., 'MJPG', 'YUYV'), but '{self.fourcc}' is provided."
-            )

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

@@ -16,8 +16,6 @@ from dataclasses import dataclass
 
 from ..configs import CameraConfig, ColorMode
 
-__all__ = ["CameraConfig", "ColorMode", "Reachy2CameraConfig"]
-
 
 @CameraConfig.register_subclass("reachy2_camera")
 @dataclass
@@ -64,7 +62,7 @@ class Reachy2CameraConfig(CameraConfig):
     port: int = 50065
     # use_depth: bool = False
 
-    def __post_init__(self) -> None:
+    def __post_init__(self):
         if self.name not in ["teleop", "depth"]:
             raise ValueError(f"`name` is expected to be 'teleop' or 'depth', but {self.name} is provided.")
         if (self.name == "teleop" and self.image_type not in ["left", "right"]) or (

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -23,19 +23,15 @@ import time
 from threading import Event, Lock, Thread
 from typing import Any
 
-from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
-
 # Fix MSMF hardware transform compatibility for Windows before importing cv2
 if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS" not in os.environ:
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
-import cv2  # type: ignore  # TODO: add type stubs for OpenCV
-import numpy as np  # type: ignore  # TODO: add type stubs for numpy
-from reachy2_sdk.media.camera import CameraView  # type: ignore  # TODO: add type stubs for reachy2_sdk
-from reachy2_sdk.media.camera_manager import (  # type: ignore  # TODO: add type stubs for reachy2_sdk
-    CameraManager,
-)
+import cv2
+import numpy as np
+from reachy2_sdk.media.camera import CameraView
+from reachy2_sdk.media.camera_manager import CameraManager
 
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.errors import DeviceNotConnectedError
 
 from ..camera import Camera
 from .configuration_reachy2_camera import ColorMode, Reachy2CameraConfig
@@ -77,7 +73,7 @@ class Reachy2Camera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: NDArray[Any] | None = None
+        self.latest_frame: np.ndarray | None = None
         self.new_frame_event: Event = Event()
 
     def __str__(self) -> str:
@@ -87,17 +83,13 @@ class Reachy2Camera(Camera):
     def is_connected(self) -> bool:
         """Checks if the camera is currently connected and opened."""
         if self.config.name == "teleop":
-            return bool(
-                self.cam_manager._grpc_connected and self.cam_manager.teleop if self.cam_manager else False
-            )
+            return self.cam_manager._grpc_connected and self.cam_manager.teleop if self.cam_manager else False
         elif self.config.name == "depth":
-            return bool(
-                self.cam_manager._grpc_connected and self.cam_manager.depth if self.cam_manager else False
-            )
+            return self.cam_manager._grpc_connected and self.cam_manager.depth if self.cam_manager else False
         else:
             raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")
 
-    def connect(self, warmup: bool = True) -> None:
+    def connect(self, warmup: bool = True):
         """
         Connects to the Reachy2 CameraManager as specified in the configuration.
         """
@@ -139,7 +131,7 @@ class Reachy2Camera(Camera):
         camera_manager.disconnect()
         return initialized_cameras
 
-    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
         """
         Reads a single frame synchronously from the camera.
 
@@ -160,7 +152,7 @@ class Reachy2Camera(Camera):
 
         start_time = time.perf_counter()
 
-        frame: NDArray[Any] = np.empty((0, 0, 3), dtype=np.uint8)
+        frame = None
 
         if self.cam_manager is None:
             raise DeviceNotConnectedError(f"{self} is not connected.")
@@ -187,7 +179,7 @@ class Reachy2Camera(Camera):
 
         return frame
 
-    def _read_loop(self) -> None:
+    def _read_loop(self):
         """
         Internal loop run by the background thread for asynchronous reading.
 
@@ -198,9 +190,6 @@ class Reachy2Camera(Camera):
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
-        if self.stop_event is None:
-            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
-
         while not self.stop_event.is_set():
             try:
                 color_image = self.read()
@@ -237,7 +226,7 @@ class Reachy2Camera(Camera):
         self.thread = None
         self.stop_event = None
 
-    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
+    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
         """
         Reads the latest available frame asynchronously.
 
@@ -280,7 +269,7 @@ class Reachy2Camera(Camera):
 
         return frame
 
-    def disconnect(self) -> None:
+    def disconnect(self):
         """
         Stops the background read thread (if running).
 

src/lerobot/cameras/realsense/camera_realsense.py

@@ -21,16 +21,15 @@ import time
 from threading import Event, Lock, Thread
 from typing import Any
 
-import cv2  # type: ignore  # TODO: add type stubs for OpenCV
-import numpy as np  # type: ignore  # TODO: add type stubs for numpy
-from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
+import cv2
+import numpy as np
 
 try:
-    import pyrealsense2 as rs  # type: ignore  # TODO: add type stubs for pyrealsense2
+    import pyrealsense2 as rs
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -133,7 +132,7 @@ class RealSenseCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: NDArray[Any] | None = None
+        self.latest_frame: np.ndarray | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -151,7 +150,7 @@ class RealSenseCamera(Camera):
         """Checks if the camera pipeline is started and streams are active."""
         return self.rs_pipeline is not None and self.rs_profile is not None
 
-    def connect(self, warmup: bool = True) -> None:
+    def connect(self, warmup: bool = True):
         """
         Connects to the RealSense camera specified in the configuration.
 
@@ -265,7 +264,7 @@ class RealSenseCamera(Camera):
         serial_number = str(found_devices[0]["serial_number"])
         return serial_number
 
-    def _configure_rs_pipeline_config(self, rs_config: Any) -> None:
+    def _configure_rs_pipeline_config(self, rs_config):
         """Creates and configures the RealSense pipeline configuration object."""
         rs.config.enable_device(rs_config, self.serial_number)
 
@@ -294,9 +293,6 @@ class RealSenseCamera(Camera):
         if not self.is_connected:
             raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
 
-        if self.rs_profile is None:
-            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
-
         stream = self.rs_profile.get_stream(rs.stream.color).as_video_stream_profile()
 
         if self.fps is None:
@@ -312,7 +308,7 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
-    def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
+    def read_depth(self, timeout_ms: int = 200) -> np.ndarray:
         """
         Reads a single frame (depth) synchronously from the camera.
 
@@ -340,9 +336,6 @@ class RealSenseCamera(Camera):
 
         start_time = time.perf_counter()
 
-        if self.rs_pipeline is None:
-            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
-
         ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
         if not ret or frame is None:
@@ -358,7 +351,7 @@ class RealSenseCamera(Camera):
 
         return depth_map_processed
 
-    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> NDArray[Any]:
+    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> np.ndarray:
         """
         Reads a single frame (color) synchronously from the camera.
 
@@ -383,9 +376,6 @@ class RealSenseCamera(Camera):
 
         start_time = time.perf_counter()
 
-        if self.rs_pipeline is None:
-            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
-
         ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
         if not ret or frame is None:
@@ -402,8 +392,8 @@ class RealSenseCamera(Camera):
         return color_image_processed
 
     def _postprocess_image(
-        self, image: NDArray[Any], color_mode: ColorMode | None = None, depth_frame: bool = False
-    ) -> NDArray[Any]:
+        self, image: np.ndarray, color_mode: ColorMode | None = None, depth_frame: bool = False
+    ) -> np.ndarray:
         """
         Applies color conversion, dimension validation, and rotation to a raw color frame.
 
@@ -448,7 +438,7 @@ class RealSenseCamera(Camera):
 
         return processed_image
 
-    def _read_loop(self) -> None:
+    def _read_loop(self):
         """
         Internal loop run by the background thread for asynchronous reading.
 
@@ -459,9 +449,6 @@ class RealSenseCamera(Camera):
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
-        if self.stop_event is None:
-            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
-
         while not self.stop_event.is_set():
             try:
                 color_image = self.read(timeout_ms=500)
@@ -487,7 +474,7 @@ class RealSenseCamera(Camera):
         self.thread.daemon = True
         self.thread.start()
 
-    def _stop_read_thread(self) -> None:
+    def _stop_read_thread(self):
         """Signals the background read thread to stop and waits for it to join."""
         if self.stop_event is not None:
             self.stop_event.set()
@@ -499,7 +486,7 @@ class RealSenseCamera(Camera):
         self.stop_event = None
 
     # NOTE(Steven): Missing implementation for depth for now
-    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
+    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
         """
         Reads the latest available frame data (color) asynchronously.
 
@@ -542,7 +529,7 @@ class RealSenseCamera(Camera):
 
         return frame
 
-    def disconnect(self) -> None:
+    def disconnect(self):
         """
         Disconnects from the camera, stops the pipeline, and cleans up resources.
 

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -59,7 +59,7 @@ class RealSenseCameraConfig(CameraConfig):
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
 
-    def __post_init__(self) -> None:
+    def __post_init__(self):
         if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
                 f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."

src/lerobot/cameras/utils.py

@@ -15,19 +15,19 @@
 # limitations under the License.
 
 import platform
-from typing import cast
-
-from lerobot.utils.import_utils import make_device_from_device_class
+from pathlib import Path
+from typing import TypeAlias
 
 from .camera import Camera
 from .configs import CameraConfig, Cv2Rotation
 
+IndexOrPath: TypeAlias = int | Path
+
 
 def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[str, Camera]:
-    cameras: dict[str, Camera] = {}
+    cameras = {}
 
     for key, cfg in camera_configs.items():
-        # TODO(Steven): Consider just using the make_device_from_device_class for all types
         if cfg.type == "opencv":
             from .opencv import OpenCVCamera
 
@@ -44,23 +44,20 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[s
             cameras[key] = Reachy2Camera(cfg)
 
         else:
-            try:
-                cameras[key] = cast(Camera, make_device_from_device_class(cfg))
-            except Exception as e:
-                raise ValueError(f"Error creating camera {key} with config {cfg}: {e}") from e
+            raise ValueError(f"The camera type '{cfg.type}' is not valid.")
 
     return cameras
 
 
 def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
-    import cv2  # type: ignore  # TODO: add type stubs for OpenCV
+    import cv2
 
     if rotation == Cv2Rotation.ROTATE_90:
-        return int(cv2.ROTATE_90_CLOCKWISE)
+        return cv2.ROTATE_90_CLOCKWISE
     elif rotation == Cv2Rotation.ROTATE_180:
-        return int(cv2.ROTATE_180)
+        return cv2.ROTATE_180
     elif rotation == Cv2Rotation.ROTATE_270:
-        return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
+        return cv2.ROTATE_90_COUNTERCLOCKWISE
     else:
         return None
 
@@ -69,8 +66,8 @@ def get_cv2_backend() -> int:
     import cv2
 
     if platform.system() == "Windows":
-        return int(cv2.CAP_MSMF)  # Use MSMF for Windows instead of AVFOUNDATION
+        return cv2.CAP_MSMF  # Use MSMF for Windows instead of AVFOUNDATION
     # elif platform.system() == "Darwin":  # macOS
     #     return cv2.CAP_AVFOUNDATION
     else:  # Linux and others
-        return int(cv2.CAP_ANY)
+        return cv2.CAP_ANY

src/lerobot/configs/default.py

@@ -57,7 +57,7 @@ class EvalConfig:
     # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
     use_async_envs: bool = False
 
-    def __post_init__(self) -> None:
+    def __post_init__(self):
         if self.batch_size > self.n_episodes:
             raise ValueError(
                 "The eval batch size is greater than the number of eval episodes "

src/lerobot/configs/eval.py

@@ -13,8 +13,8 @@
 # limitations under the License.
 
 import datetime as dt
+import logging
 from dataclasses import dataclass, field
-from logging import getLogger
 from pathlib import Path
 
 from lerobot import envs, policies  # noqa: F401
@@ -22,8 +22,6 @@ from lerobot.configs import parser
 from lerobot.configs.default import EvalConfig
 from lerobot.configs.policies import PreTrainedConfig
 
-logger = getLogger(__name__)
-
 
 @dataclass
 class EvalPipelineConfig:
@@ -36,31 +34,25 @@ class EvalPipelineConfig:
     output_dir: Path | None = None
     job_name: str | None = None
     seed: int | None = 1000
-    # Rename map for the observation to override the image and state keys
-    rename_map: dict[str, str] = field(default_factory=dict)
 
-    def __post_init__(self) -> None:
+    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("policy")
         if policy_path:
             cli_overrides = parser.get_cli_overrides("policy")
             self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            self.policy.pretrained_path = Path(policy_path)
+            self.policy.pretrained_path = policy_path
 
         else:
-            logger.warning(
+            logging.warning(
                 "No pretrained path was provided, evaluated policy will be built from scratch (random weights)."
             )
 
         if not self.job_name:
             if self.env is None:
-                self.job_name = f"{self.policy.type if self.policy is not None else 'scratch'}"
+                self.job_name = f"{self.policy.type}"
             else:
-                self.job_name = (
-                    f"{self.env.type}_{self.policy.type if self.policy is not None else 'scratch'}"
-                )
-
-            logger.warning(f"No job name provided, using '{self.job_name}' as job name.")
+                self.job_name = f"{self.env.type}_{self.policy.type}"
 
         if not self.output_dir:
             now = dt.datetime.now()

src/lerobot/configs/parser.py

@@ -16,19 +16,14 @@ import inspect
 import pkgutil
 import sys
 from argparse import ArgumentError
-from collections.abc import Callable, Iterable, Sequence
+from collections.abc import Sequence
 from functools import wraps
 from pathlib import Path
-from pkgutil import ModuleInfo
-from types import ModuleType
-from typing import Any, TypeVar, cast
 
 import draccus
 
 from lerobot.utils.utils import has_method
 
-F = TypeVar("F", bound=Callable[..., object])
-
 PATH_KEY = "path"
 PLUGIN_DISCOVERY_SUFFIX = "discover_packages_path"
 
@@ -65,7 +60,7 @@ def parse_arg(arg_name: str, args: Sequence[str] | None = None) -> str | None:
     return None
 
 
-def parse_plugin_args(plugin_arg_suffix: str, args: Sequence[str]) -> dict[str, str]:
+def parse_plugin_args(plugin_arg_suffix: str, args: Sequence[str]) -> dict:
     """Parse plugin-related arguments from command-line arguments.
 
     This function extracts arguments from command-line arguments that match a specified suffix pattern.
@@ -132,7 +127,7 @@ def load_plugin(plugin_path: str) -> None:
             f"Failed to load plugin '{plugin_path}'. Verify the path and installation: {str(e)}"
         ) from e
 
-    def iter_namespace(ns_pkg: ModuleType) -> Iterable[ModuleInfo]:
+    def iter_namespace(ns_pkg):
         return pkgutil.iter_modules(ns_pkg.__path__, ns_pkg.__name__ + ".")
 
     try:
@@ -153,8 +148,6 @@ def get_type_arg(field_name: str, args: Sequence[str] | None = None) -> str | No
 
 
 def filter_arg(field_to_filter: str, args: Sequence[str] | None = None) -> list[str]:
-    if args is None:
-        return []
     return [arg for arg in args if not arg.startswith(f"--{field_to_filter}=")]
 
 
@@ -178,8 +171,7 @@ def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | No
     if isinstance(fields_to_filter, str):
         fields_to_filter = [fields_to_filter]
 
-    filtered_args = [] if args is None else list(args)
-
+    filtered_args = args
     for field in fields_to_filter:
         if get_path_arg(field, args):
             if get_type_arg(field, args):
@@ -192,7 +184,7 @@ def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | No
     return filtered_args
 
 
-def wrap(config_path: Path | None = None) -> Callable[[F], F]:
+def wrap(config_path: Path | None = None):
     """
     HACK: Similar to draccus.wrap but does three additional things:
         - Will remove '.path' arguments from CLI in order to process them later on.
@@ -203,9 +195,9 @@ def wrap(config_path: Path | None = None) -> Callable[[F], F]:
             from the CLI '.type' arguments
     """
 
-    def wrapper_outer(fn: F) -> F:
+    def wrapper_outer(fn):
         @wraps(fn)
-        def wrapper_inner(*args: Any, **kwargs: Any) -> Any:
+        def wrapper_inner(*args, **kwargs):
             argspec = inspect.getfullargspec(fn)
             argtype = argspec.annotations[argspec.args[0]]
             if len(args) > 0 and type(args[0]) is argtype:
@@ -233,6 +225,6 @@ def wrap(config_path: Path | None = None) -> Callable[[F], F]:
             response = fn(cfg, *args, **kwargs)
             return response
 
-        return cast(F, wrapper_inner)
+        return wrapper_inner
 
-    return cast(Callable[[F], F], wrapper_outer)
+    return wrapper_outer

src/lerobot/configs/policies.py

@@ -14,12 +14,12 @@
 import abc
 import builtins
 import json
+import logging
 import os
 import tempfile
 from dataclasses import dataclass, field
-from logging import getLogger
 from pathlib import Path
-from typing import Any, TypeVar
+from typing import TypeVar
 
 import draccus
 from huggingface_hub import hf_hub_download
@@ -27,18 +27,17 @@ from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError
 
 from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.constants import ACTION, OBS_STATE
 from lerobot.optim.optimizers import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
-from lerobot.utils.constants import ACTION, OBS_STATE
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
 
 T = TypeVar("T", bound="PreTrainedConfig")
-logger = getLogger(__name__)
 
 
 @dataclass
-class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: ignore[misc,name-defined] #TODO: draccus issue
+class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     """
     Base configuration class for policy models.
 
@@ -58,12 +57,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     input_features: dict[str, PolicyFeature] = field(default_factory=dict)
     output_features: dict[str, PolicyFeature] = field(default_factory=dict)
 
-    device: str | None = None  # e.g. "cuda", "cuda:0", "cpu", or "mps"
+    device: str | None = None  # cuda | cpu | mp
     # `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
     # automatic gradient scaling is used.
     use_amp: bool = False
 
-    push_to_hub: bool = True  # type: ignore[assignment] # TODO: use a different name to avoid override
+    push_to_hub: bool = True
     repo_id: str | None = None
 
     # Upload on private repository on the Hugging Face hub.
@@ -74,41 +73,38 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     license: str | None = None
     # Either the repo ID of a model hosted on the Hub or a path to a directory containing weights
     # saved using `Policy.save_pretrained`. If not provided, the policy is initialized from scratch.
-    pretrained_path: Path | None = None
+    pretrained_path: str | None = None
 
-    def __post_init__(self) -> None:
+    def __post_init__(self):
         if not self.device or not is_torch_device_available(self.device):
             auto_device = auto_select_torch_device()
-            logger.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
+            logging.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
             self.device = auto_device.type
 
         # Automatically deactivate AMP if necessary
         if self.use_amp and not is_amp_available(self.device):
-            logger.warning(
+            logging.warning(
                 f"Automatic Mixed Precision (amp) is not available on device '{self.device}'. Deactivating AMP."
             )
             self.use_amp = False
 
     @property
     def type(self) -> str:
-        choice_name = self.get_choice_name(self.__class__)
-        if not isinstance(choice_name, str):
-            raise TypeError(f"Expected string from get_choice_name, got {type(choice_name)}")
-        return choice_name
+        return self.get_choice_name(self.__class__)
 
     @property
     @abc.abstractmethod
-    def observation_delta_indices(self) -> list | None:  # type: ignore[type-arg] #TODO: No implementation
+    def observation_delta_indices(self) -> list | None:
         raise NotImplementedError
 
     @property
     @abc.abstractmethod
-    def action_delta_indices(self) -> list | None:  # type: ignore[type-arg]    #TODO: No implementation
+    def action_delta_indices(self) -> list | None:
         raise NotImplementedError
 
     @property
     @abc.abstractmethod
-    def reward_delta_indices(self) -> list | None:  # type: ignore[type-arg]    #TODO: No implementation
+    def reward_delta_indices(self) -> list | None:
         raise NotImplementedError
 
     @abc.abstractmethod
@@ -158,13 +154,13 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
         pretrained_name_or_path: str | Path,
         *,
         force_download: bool = False,
-        resume_download: bool | None = None,
-        proxies: dict[Any, Any] | None = None,
+        resume_download: bool = 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,
-        **policy_kwargs: Any,
+        **policy_kwargs,
     ) -> T:
         model_id = str(pretrained_name_or_path)
         config_file: str | None = None
@@ -172,7 +168,7 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
             if CONFIG_NAME in os.listdir(model_id):
                 config_file = os.path.join(model_id, CONFIG_NAME)
             else:
-                logger.error(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
+                print(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
         else:
             try:
                 config_file = hf_hub_download(
@@ -198,17 +194,15 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
         with draccus.config_type("json"):
             orig_config = draccus.parse(cls, config_file, args=[])
 
-        if config_file is None:
-            raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
-
         with open(config_file) as f:
             config = json.load(f)
 
         config.pop("type")
-        with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
+        with tempfile.NamedTemporaryFile("w+") as f:
             json.dump(config, f)
             config_file = f.name
+            f.flush()
 
-        cli_overrides = policy_kwargs.pop("cli_overrides", [])
-        with draccus.config_type("json"):
-            return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)
+            cli_overrides = policy_kwargs.pop("cli_overrides", [])
+            with draccus.config_type("json"):
+                return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)

src/lerobot/configs/train.py

@@ -16,7 +16,6 @@ import datetime as dt
 import os
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any
 
 import draccus
 from huggingface_hub import hf_hub_download
@@ -64,18 +63,22 @@ class TrainPipelineConfig(HubMixin):
     scheduler: LRSchedulerConfig | None = None
     eval: EvalConfig = field(default_factory=EvalConfig)
     wandb: WandBConfig = field(default_factory=WandBConfig)
-    checkpoint_path: Path | None = field(init=False, default=None)
-    # Rename map for the observation to override the image and state keys
-    rename_map: dict[str, str] = field(default_factory=dict)
+    # Accelerate configuration for multi-GPU training
+    use_accelerate: bool = False
+    gradient_accumulation_steps: int = 1
+    mixed_precision: str = "no"  # Options: "no", "fp16", "bf16"
 
-    def validate(self) -> None:
+    def __post_init__(self):
+        self.checkpoint_path = None
+
+    def validate(self):
         # HACK: We parse again the cli args here to get the pretrained paths if there was some.
         policy_path = parser.get_path_arg("policy")
         if policy_path:
             # Only load the policy config
             cli_overrides = parser.get_cli_overrides("policy")
             self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            self.policy.pretrained_path = Path(policy_path)
+            self.policy.pretrained_path = policy_path
         elif self.resume:
             # The entire train config is already loaded, we just need to get the checkpoint dir
             config_path = parser.parse_arg("config_path")
@@ -83,22 +86,14 @@ class TrainPipelineConfig(HubMixin):
                 raise ValueError(
                     f"A config_path is expected when resuming a run. Please specify path to {TRAIN_CONFIG_NAME}"
                 )
-
             if not Path(config_path).resolve().exists():
                 raise NotADirectoryError(
                     f"{config_path=} is expected to be a local path. "
                     "Resuming from the hub is not supported for now."
                 )
-
-            policy_dir = Path(config_path).parent
-            if self.policy is not None:
-                self.policy.pretrained_path = policy_dir
-            self.checkpoint_path = policy_dir.parent
-
-        if self.policy is None:
-            raise ValueError(
-                "Policy is not configured. Please specify a pretrained policy with `--policy.path`."
-            )
+            policy_path = Path(config_path).parent
+            self.policy.pretrained_path = policy_path
+            self.checkpoint_path = policy_path.parent
 
         if not self.job_name:
             if self.env is None:
@@ -135,8 +130,8 @@ class TrainPipelineConfig(HubMixin):
         """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
         return ["policy"]
 
-    def to_dict(self) -> dict[str, Any]:
-        return draccus.encode(self)  # type: ignore[no-any-return]  # because of the third-party library draccus uses Any as the return type
+    def to_dict(self) -> dict:
+        return draccus.encode(self)
 
     def _save_pretrained(self, save_directory: Path) -> None:
         with open(save_directory / TRAIN_CONFIG_NAME, "w") as f, draccus.config_type("json"):
@@ -148,13 +143,13 @@ class TrainPipelineConfig(HubMixin):
         pretrained_name_or_path: str | Path,
         *,
         force_download: bool = False,
-        resume_download: bool | None = None,
-        proxies: dict[Any, Any] | None = None,
+        resume_download: bool = 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,
-        **kwargs: Any,
+        **kwargs,
     ) -> "TrainPipelineConfig":
         model_id = str(pretrained_name_or_path)
         config_file: str | None = None
@@ -190,6 +185,4 @@ class TrainPipelineConfig(HubMixin):
 
 @dataclass(kw_only=True)
 class TrainRLServerPipelineConfig(TrainPipelineConfig):
-    # NOTE: In RL, we don't need an offline dataset
-    # TODO: Make `TrainPipelineConfig.dataset` optional
-    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional
+    dataset: DatasetConfig | None = None  # NOTE: In RL, we don't need an offline dataset

src/lerobot/configs/types.py

@@ -15,6 +15,7 @@
 # https://stackoverflow.com/questions/24481852/serialising-an-enum-member-to-json
 from dataclasses import dataclass
 from enum import Enum
+from typing import Any, Protocol
 
 
 class FeatureType(str, Enum):
@@ -39,7 +40,11 @@ class NormalizationMode(str, Enum):
     QUANTILE10 = "QUANTILE10"
 
 
+class DictLike(Protocol):
+    def __getitem__(self, key: Any) -> Any: ...
+
+
 @dataclass
 class PolicyFeature:
     type: FeatureType
-    shape: tuple[int, ...]
+    shape: tuple

src/lerobot/constants.py

@@ -17,22 +17,21 @@ from pathlib import Path
 
 from huggingface_hub.constants import HF_HOME
 
-OBS_STR = "observation"
-OBS_PREFIX = OBS_STR + "."
-OBS_ENV_STATE = OBS_STR + ".environment_state"
-OBS_STATE = OBS_STR + ".state"
-OBS_IMAGE = OBS_STR + ".image"
-OBS_IMAGES = OBS_IMAGE + "s"
-OBS_LANGUAGE = OBS_STR + ".language"
-OBS_LANGUAGE_TOKENS = OBS_LANGUAGE + ".tokens"
-OBS_LANGUAGE_ATTENTION_MASK = OBS_LANGUAGE + ".attention_mask"
-
+OBS_ENV_STATE = "observation.environment_state"
+OBS_STATE = "observation.state"
+OBS_IMAGE = "observation.image"
+OBS_IMAGES = "observation.images"
+OBS_LANGUAGE = "observation.language"
 ACTION = "action"
 REWARD = "next.reward"
 TRUNCATED = "next.truncated"
 DONE = "next.done"
 
+OBS_LANGUAGE_TOKENS = OBS_LANGUAGE + ".tokens"
+OBS_LANGUAGE_ATTENTION_MASK = OBS_LANGUAGE + ".attention_mask"
+
 ROBOTS = "robots"
+ROBOT_TYPE = "robot_type"
 TELEOPERATORS = "teleoperators"
 
 # files & directories
@@ -67,6 +66,3 @@ HF_LEROBOT_CALIBRATION = Path(os.getenv("HF_LEROBOT_CALIBRATION", default_calibr
 # streaming datasets
 LOOKBACK_BACKTRACKTABLE = 100
 LOOKAHEAD_BACKTRACKTABLE = 100
-
-# openpi
-OPENPI_ATTENTION_MASK_VALUE = -2.3819763e38  # TODO(pepijn): Modify this when extending support to fp8 models

src/lerobot/datasets/aggregate.py

@@ -31,15 +31,15 @@ from lerobot.datasets.utils import (
     DEFAULT_EPISODES_PATH,
     DEFAULT_VIDEO_FILE_SIZE_IN_MB,
     DEFAULT_VIDEO_PATH,
-    get_file_size_in_mb,
     get_parquet_file_size_in_mb,
+    get_video_size_in_mb,
     to_parquet_with_hf_images,
     update_chunk_file_indices,
     write_info,
     write_stats,
     write_tasks,
 )
-from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
+from lerobot.datasets.video_utils import concatenate_video_files
 
 
 def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
@@ -93,13 +93,14 @@ def update_data_df(df, src_meta, dst_meta):
         pd.DataFrame: Updated DataFrame with adjusted indices.
     """
 
-    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
-    df["index"] = df["index"] + dst_meta.info["total_frames"]
+    def _update(row):
+        row["episode_index"] = row["episode_index"] + dst_meta.info["total_episodes"]
+        row["index"] = row["index"] + dst_meta.info["total_frames"]
+        task = src_meta.tasks.iloc[row["task_index"]].name
+        row["task_index"] = dst_meta.tasks.loc[task].task_index.item()
+        return row
 
-    src_task_names = src_meta.tasks.index.take(df["task_index"].to_numpy())
-    df["task_index"] = dst_meta.tasks.loc[src_task_names, "task_index"].to_numpy()
-
-    return df
+    return df.apply(_update, axis=1)
 
 
 def update_meta_data(
@@ -125,45 +126,27 @@ def update_meta_data(
         pd.DataFrame: Updated DataFrame with adjusted indices and timestamps.
     """
 
-    df["meta/episodes/chunk_index"] = df["meta/episodes/chunk_index"] + meta_idx["chunk"]
-    df["meta/episodes/file_index"] = df["meta/episodes/file_index"] + meta_idx["file"]
-    df["data/chunk_index"] = df["data/chunk_index"] + data_idx["chunk"]
-    df["data/file_index"] = df["data/file_index"] + data_idx["file"]
-    for key, video_idx in videos_idx.items():
-        # Store original video file indices before updating
-        orig_chunk_col = f"videos/{key}/chunk_index"
-        orig_file_col = f"videos/{key}/file_index"
-        df["_orig_chunk"] = df[orig_chunk_col].copy()
-        df["_orig_file"] = df[orig_file_col].copy()
-
-        # Update chunk and file indices to point to destination
-        df[orig_chunk_col] = video_idx["chunk"]
-        df[orig_file_col] = video_idx["file"]
-
-        # Apply per-source-file timestamp offsets
-        src_to_offset = video_idx.get("src_to_offset", {})
-        if src_to_offset:
-            # Apply offset based on original source file
-            for idx in df.index:
-                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
-                offset = src_to_offset.get(src_key, 0)
-                df.at[idx, f"videos/{key}/from_timestamp"] += offset
-                df.at[idx, f"videos/{key}/to_timestamp"] += offset
-        else:
-            # Fallback to simple offset (for backward compatibility)
-            df[f"videos/{key}/from_timestamp"] = (
-                df[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
+    def _update(row):
+        row["meta/episodes/chunk_index"] = row["meta/episodes/chunk_index"] + meta_idx["chunk"]
+        row["meta/episodes/file_index"] = row["meta/episodes/file_index"] + meta_idx["file"]
+        row["data/chunk_index"] = row["data/chunk_index"] + data_idx["chunk"]
+        row["data/file_index"] = row["data/file_index"] + data_idx["file"]
+        for key, video_idx in videos_idx.items():
+            row[f"videos/{key}/chunk_index"] = row[f"videos/{key}/chunk_index"] + video_idx["chunk"]
+            row[f"videos/{key}/file_index"] = row[f"videos/{key}/file_index"] + video_idx["file"]
+            row[f"videos/{key}/from_timestamp"] = (
+                row[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
+            )
+            row[f"videos/{key}/to_timestamp"] = (
+                row[f"videos/{key}/to_timestamp"] + video_idx["latest_duration"]
             )
-            df[f"videos/{key}/to_timestamp"] = df[f"videos/{key}/to_timestamp"] + video_idx["latest_duration"]
 
-        # Clean up temporary columns
-        df = df.drop(columns=["_orig_chunk", "_orig_file"])
+        row["dataset_from_index"] = row["dataset_from_index"] + dst_meta.info["total_frames"]
+        row["dataset_to_index"] = row["dataset_to_index"] + dst_meta.info["total_frames"]
+        row["episode_index"] = row["episode_index"] + dst_meta.info["total_episodes"]
+        return row
 
-    df["dataset_from_index"] = df["dataset_from_index"] + dst_meta.info["total_frames"]
-    df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info["total_frames"]
-    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
-
-    return df
+    return df.apply(_update, axis=1)
 
 
 def aggregate_datasets(
@@ -217,10 +200,6 @@ def aggregate_datasets(
         robot_type=robot_type,
         features=features,
         root=aggr_root,
-        use_videos=len(video_keys) > 0,
-        chunks_size=chunk_size,
-        data_files_size_in_mb=data_files_size_in_mb,
-        video_files_size_in_mb=video_files_size_in_mb,
     )
 
     logging.info("Find all tasks")
@@ -264,11 +243,6 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
     Returns:
         dict: Updated videos_idx with current chunk and file indices.
     """
-    for key in videos_idx:
-        videos_idx[key]["episode_duration"] = 0
-        # Track offset for each source (chunk, file) pair
-        videos_idx[key]["src_to_offset"] = {}
-
     for key, video_idx in videos_idx.items():
         unique_chunk_file_pairs = {
             (chunk, file)
@@ -282,7 +256,6 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
 
         chunk_idx = video_idx["chunk"]
         file_idx = video_idx["file"]
-        current_offset = video_idx["latest_duration"]
 
         for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
             src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
@@ -297,25 +270,21 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                 file_index=file_idx,
             )
 
-            src_duration = get_video_duration_in_s(src_path)
+            # If a new file is created, we don't want to increment the latest_duration
+            update_latest_duration = False
 
             if not dst_path.exists():
-                # Store offset before incrementing
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
+                # First write to this destination file
                 dst_path.parent.mkdir(parents=True, exist_ok=True)
                 shutil.copy(str(src_path), str(dst_path))
-                videos_idx[key]["episode_duration"] += src_duration
-                current_offset += src_duration
-                continue
+                continue  # not accumulating further, already copied the file in place
 
             # Check file sizes before appending
-            src_size = get_file_size_in_mb(src_path)
-            dst_size = get_file_size_in_mb(dst_path)
+            src_size = get_video_size_in_mb(src_path)
+            dst_size = get_video_size_in_mb(dst_path)
 
             if dst_size + src_size >= video_files_size_in_mb:
-                # Rotate to a new file, this source becomes start of new destination
-                # So its offset should be 0
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = 0
+                # Rotate to a new chunk/file
                 chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, chunk_size)
                 dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
                     video_key=key,
@@ -324,22 +293,25 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                 )
                 dst_path.parent.mkdir(parents=True, exist_ok=True)
                 shutil.copy(str(src_path), str(dst_path))
-                # Reset offset for next file
-                current_offset = src_duration
             else:
-                # Append to existing video file - use current accumulated offset
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
+                # Get the timestamps shift for this video
+                timestamps_shift_s = dst_meta.info["total_frames"] / dst_meta.info["fps"]
+
+                # Append to existing video file
                 concatenate_video_files(
                     [dst_path, src_path],
                     dst_path,
                 )
-                current_offset += src_duration
-
-            videos_idx[key]["episode_duration"] += src_duration
+                # Update the latest_duration when appending (shifts timestamps!)
+                update_latest_duration = not update_latest_duration
 
+        # Update the videos_idx with the final chunk and file indices for this key
         videos_idx[key]["chunk"] = chunk_idx
         videos_idx[key]["file"] = file_idx
 
+        if update_latest_duration:
+            videos_idx[key]["latest_duration"] += timestamps_shift_s
+
     return videos_idx
 
 
@@ -424,6 +396,9 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
             videos_idx,
         )
 
+        for k in videos_idx:
+            videos_idx[k]["latest_duration"] += videos_idx[k]["episode_duration"]
+
         meta_idx = append_or_create_parquet_file(
             df,
             src_path,
@@ -435,10 +410,6 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
             aggr_root=dst_meta.root,
         )
 
-    # Increment latest_duration by the total duration added from this source dataset
-    for k in videos_idx:
-        videos_idx[k]["latest_duration"] += videos_idx[k]["episode_duration"]
-
     return meta_idx
 
 

src/lerobot/datasets/backward_compatibility.py

@@ -23,9 +23,6 @@ Please, update your dataset to the new format using this command:
 python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id={repo_id}
 ```
 
-If you already have a converted version uploaded to the hub, then this error might be because of
-an older version in your local cache. Consider deleting the cached version and retrying.
-
 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).
 """

src/lerobot/datasets/compute_stats.py

@@ -21,14 +21,7 @@ DEFAULT_QUANTILES = [0.01, 0.10, 0.50, 0.90, 0.99]
 
 
 class RunningQuantileStats:
-    """
-    Maintains running statistics for batches of vectors, including mean,
-    standard deviation, min, max, and approximate quantiles.
-
-    Statistics are computed per feature dimension and updated incrementally
-    as new batches are observed. Quantiles are estimated using histograms,
-    which adapt dynamically if the observed data range expands.
-    """
+    """Compute running statistics including quantiles for a batch of vectors."""
 
     def __init__(self, quantile_list: list[float] | None = None, num_quantile_bins: int = 5000):
         self._count = 0
@@ -104,7 +97,6 @@ class RunningQuantileStats:
             raise ValueError("Cannot compute statistics for less than 2 vectors.")
 
         variance = self._mean_of_squares - self._mean**2
-
         stddev = np.sqrt(np.maximum(0, variance))
 
         stats = {
@@ -323,8 +315,9 @@ def _reshape_for_global_stats(
     if keepdims:
         target_shape = tuple(1 for _ in original_shape)
         return value.reshape(target_shape)
-    # Keep at least 1-D arrays to satisfy validator
-    return np.atleast_1d(value)
+    elif not keepdims and value.ndim > 0 and value.size == 1:
+        return value.item()
+    return value
 
 
 def _reshape_single_stat(
@@ -376,9 +369,10 @@ def _prepare_array_for_stats(array: np.ndarray, axis: int | tuple[int, ...] | No
         return reshaped, batch_size
 
     if axis == 0 or axis == (0,):  # Vector data
-        reshaped = array
         if array.ndim == 1:
             reshaped = array.reshape(-1, 1)
+        else:
+            reshaped = array
         return reshaped, array.shape[0]
 
     if axis == (1,):  # Feature-wise statistics
@@ -416,6 +410,12 @@ def _compute_basic_stats(
         "count": np.array([sample_count]),
     }
 
+    # For single-element arrays with shape (1,1), convert to scalar arrays
+    if array.shape == (1, 1):
+        for key in stats:
+            if key != "count" and stats[key].size == 1:
+                stats[key] = np.array(stats[key].item())
+
     for q in quantile_list_keys:
         stats[q] = stats["mean"].copy()
 
@@ -470,6 +470,12 @@ def get_feature_stats(
         stats = running_stats.get_statistics()
         stats["count"] = np.array([sample_count])
 
+    # For axis=None, the stats are computed as 1D arrays but should be 0-dimensional arrays
+    if axis is None and reshaped.shape[1] == 1:
+        for key in stats:
+            if key != "count" and stats[key].size == 1:
+                stats[key] = np.array(stats[key].item())
+
     stats = _reshape_stats_by_axis(stats, axis, keepdims, original_shape)
     return stats
 
@@ -591,7 +597,7 @@ def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, d
     }
 
     if stats_ft_list:
-        quantile_keys = [k for k in stats_ft_list[0] if k.startswith("q") and k[1:].isdigit()]
+        quantile_keys = [k for k in stats_ft_list[0].keys() if k.startswith("q") and k[1:].isdigit()]
 
         for q_key in quantile_keys:
             if all(q_key in s for s in stats_ft_list):

src/lerobot/datasets/factory.py

@@ -27,7 +27,6 @@ from lerobot.datasets.lerobot_dataset import (
 )
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.transforms import ImageTransforms
-from lerobot.utils.constants import ACTION, OBS_PREFIX, REWARD
 
 IMAGENET_STATS = {
     "mean": [[[0.485]], [[0.456]], [[0.406]]],  # (c,1,1)
@@ -55,11 +54,11 @@ def resolve_delta_timestamps(
     """
     delta_timestamps = {}
     for key in ds_meta.features:
-        if key == REWARD and cfg.reward_delta_indices is not None:
+        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:
+        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(OBS_PREFIX) and cfg.observation_delta_indices is not None:
+        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:

src/lerobot/datasets/image_writer.py

@@ -68,30 +68,7 @@ def image_array_to_pil_image(image_array: np.ndarray, range_check: bool = True)
     return PIL.Image.fromarray(image_array)
 
 
-def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1):
-    """
-    Saves a NumPy array or PIL Image to a file.
-
-    This function handles both NumPy arrays and PIL Image objects, converting
-    the former to a PIL Image before saving. It includes error handling for
-    the save operation.
-
-    Args:
-        image (np.ndarray | PIL.Image.Image): The image data to save.
-        fpath (Path): The destination file path for the image.
-        compress_level (int, optional): The compression level for the saved
-            image, as used by PIL.Image.save(). Defaults to 1.
-            Refer to: https://github.com/huggingface/lerobot/pull/2135
-            for more details on the default value rationale.
-
-    Raises:
-        TypeError: If the input 'image' is not a NumPy array or a
-            PIL.Image.Image object.
-
-    Side Effects:
-        Prints an error message to the console if the image writing process
-        fails for any reason.
-    """
+def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path):
     try:
         if isinstance(image, np.ndarray):
             img = image_array_to_pil_image(image)
@@ -99,7 +76,7 @@ def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path, compress_level
             img = image
         else:
             raise TypeError(f"Unsupported image type: {type(image)}")
-        img.save(fpath, compress_level=compress_level)
+        img.save(fpath)
     except Exception as e:
         print(f"Error writing image {fpath}: {e}")
 

src/lerobot/datasets/lerobot_dataset.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import contextlib
+import gc
 import logging
 import shutil
 import tempfile
@@ -25,13 +26,12 @@ import numpy as np
 import packaging.version
 import pandas as pd
 import PIL.Image
-import pyarrow as pa
-import pyarrow.parquet as pq
 import torch
 import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
+from lerobot.constants import HF_LEROBOT_HOME
 from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats
 from lerobot.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.datasets.utils import (
@@ -47,9 +47,13 @@ from lerobot.datasets.utils import (
     embed_images,
     flatten_dict,
     get_delta_indices,
-    get_file_size_in_mb,
+    get_hf_dataset_cache_dir,
+    get_hf_dataset_size_in_mb,
     get_hf_features_from_features,
+    get_parquet_file_size_in_mb,
+    get_parquet_num_frames,
     get_safe_version,
+    get_video_size_in_mb,
     hf_transform_to_torch,
     is_valid_version,
     load_episodes,
@@ -57,6 +61,7 @@ from lerobot.datasets.utils import (
     load_nested_dataset,
     load_stats,
     load_tasks,
+    to_parquet_with_hf_images,
     update_chunk_file_indices,
     validate_episode_buffer,
     validate_frame,
@@ -74,7 +79,6 @@ from lerobot.datasets.video_utils import (
     get_video_duration_in_s,
     get_video_info,
 )
-from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
 
@@ -86,15 +90,10 @@ class LeRobotDatasetMetadata:
         root: str | Path | None = None,
         revision: str | None = None,
         force_cache_sync: bool = False,
-        metadata_buffer_size: int = 10,
     ):
         self.repo_id = repo_id
         self.revision = revision if revision else CODEBASE_VERSION
         self.root = Path(root) if root is not None else HF_LEROBOT_HOME / repo_id
-        self.writer = None
-        self.latest_episode = None
-        self.metadata_buffer: list[dict] = []
-        self.metadata_buffer_size = metadata_buffer_size
 
         try:
             if force_cache_sync:
@@ -108,54 +107,6 @@ class LeRobotDatasetMetadata:
             self.pull_from_repo(allow_patterns="meta/")
             self.load_metadata()
 
-    def _flush_metadata_buffer(self) -> None:
-        """Write all buffered episode metadata to parquet file."""
-        if not hasattr(self, "metadata_buffer") or len(self.metadata_buffer) == 0:
-            return
-
-        combined_dict = {}
-        for episode_dict in self.metadata_buffer:
-            for key, value in episode_dict.items():
-                if key not in combined_dict:
-                    combined_dict[key] = []
-                # Extract value and serialize numpy arrays
-                # because PyArrow's from_pydict function doesn't support numpy arrays
-                val = value[0] if isinstance(value, list) else value
-                combined_dict[key].append(val.tolist() if isinstance(val, np.ndarray) else val)
-
-        first_ep = self.metadata_buffer[0]
-        chunk_idx = first_ep["meta/episodes/chunk_index"][0]
-        file_idx = first_ep["meta/episodes/file_index"][0]
-
-        table = pa.Table.from_pydict(combined_dict)
-
-        if not self.writer:
-            path = Path(self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx))
-            path.parent.mkdir(parents=True, exist_ok=True)
-            self.writer = pq.ParquetWriter(
-                path, schema=table.schema, compression="snappy", use_dictionary=True
-            )
-
-        self.writer.write_table(table)
-
-        self.latest_episode = self.metadata_buffer[-1]
-        self.metadata_buffer.clear()
-
-    def _close_writer(self) -> None:
-        """Close and cleanup the parquet writer if it exists."""
-        self._flush_metadata_buffer()
-
-        writer = getattr(self, "writer", None)
-        if writer is not None:
-            writer.close()
-            self.writer = None
-
-    def __del__(self):
-        """
-        Trust the user to call .finalize() but as an added safety check call the parquet writer to stop when calling the destructor
-        """
-        self._close_writer()
-
     def load_metadata(self):
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
@@ -187,12 +138,6 @@ class LeRobotDatasetMetadata:
         return packaging.version.parse(self.info["codebase_version"])
 
     def get_data_file_path(self, ep_index: int) -> Path:
-        if self.episodes is None:
-            self.episodes = load_episodes(self.root)
-        if ep_index >= len(self.episodes):
-            raise IndexError(
-                f"Episode index {ep_index} out of range. Episodes: {len(self.episodes) if self.episodes else 0}"
-            )
         ep = self.episodes[ep_index]
         chunk_idx = ep["data/chunk_index"]
         file_idx = ep["data/file_index"]
@@ -200,12 +145,6 @@ class LeRobotDatasetMetadata:
         return Path(fpath)
 
     def get_video_file_path(self, ep_index: int, vid_key: str) -> Path:
-        if self.episodes is None:
-            self.episodes = load_episodes(self.root)
-        if ep_index >= len(self.episodes):
-            raise IndexError(
-                f"Episode index {ep_index} out of range. Episodes: {len(self.episodes) if self.episodes else 0}"
-            )
         ep = self.episodes[ep_index]
         chunk_idx = ep[f"videos/{vid_key}/chunk_index"]
         file_idx = ep[f"videos/{vid_key}/file_index"]
@@ -321,75 +260,72 @@ class LeRobotDatasetMetadata:
             write_tasks(self.tasks, self.root)
 
     def _save_episode_metadata(self, episode_dict: dict) -> None:
-        """Buffer episode metadata and write to parquet in batches for efficiency.
+        """Save episode metadata to a parquet file and update the Hugging Face dataset of episodes metadata.
 
-        This function accumulates episode metadata in a buffer and flushes it when the buffer
-        reaches the configured size. This reduces I/O overhead by writing multiple episodes
-        at once instead of one row at a time.
+        This function processes episodes metadata from a dictionary, converts it into a Hugging Face dataset,
+        and saves it as a parquet file. It handles both the creation of new parquet files and the
+        updating of existing ones based on size constraints. After saving the metadata, it reloads
+        the Hugging Face dataset to ensure it is up-to-date.
 
         Notes: We both need to update parquet files and HF dataset:
         - `pandas` loads parquet file in RAM
         - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
           or loads directly from pyarrow cache.
         """
-        # Convert to list format for each value
+        # Convert buffer into HF Dataset
         episode_dict = {key: [value] for key, value in episode_dict.items()}
+        ep_dataset = datasets.Dataset.from_dict(episode_dict)
+        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
+        df = pd.DataFrame(ep_dataset)
         num_frames = episode_dict["length"][0]
 
-        if self.latest_episode is None:
+        if self.episodes is None:
             # Initialize indices and frame count for a new dataset made of the first episode data
             chunk_idx, file_idx = 0, 0
-            if self.episodes is not None and len(self.episodes) > 0:
-                # It means we are resuming recording, so we need to load the latest episode
-                # Update the indices to avoid overwriting the latest episode
-                chunk_idx = self.episodes[-1]["meta/episodes/chunk_index"]
-                file_idx = self.episodes[-1]["meta/episodes/file_index"]
-                latest_num_frames = self.episodes[-1]["dataset_to_index"]
-                episode_dict["dataset_from_index"] = [latest_num_frames]
-                episode_dict["dataset_to_index"] = [latest_num_frames + num_frames]
-
-                # When resuming, move to the next file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
-            else:
-                episode_dict["dataset_from_index"] = [0]
-                episode_dict["dataset_to_index"] = [num_frames]
-
-            episode_dict["meta/episodes/chunk_index"] = [chunk_idx]
-            episode_dict["meta/episodes/file_index"] = [file_idx]
+            df["meta/episodes/chunk_index"] = [chunk_idx]
+            df["meta/episodes/file_index"] = [file_idx]
+            df["dataset_from_index"] = [0]
+            df["dataset_to_index"] = [num_frames]
         else:
-            chunk_idx = self.latest_episode["meta/episodes/chunk_index"][0]
-            file_idx = self.latest_episode["meta/episodes/file_index"][0]
+            # Retrieve information from the latest parquet file
+            latest_ep = self.episodes[-1]
+            chunk_idx = latest_ep["meta/episodes/chunk_index"]
+            file_idx = latest_ep["meta/episodes/file_index"]
 
-            latest_path = (
-                self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-                if self.writer is None
-                else self.writer.where
-            )
+            latest_path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
 
-            if Path(latest_path).exists():
-                latest_size_in_mb = get_file_size_in_mb(Path(latest_path))
-                latest_num_frames = self.latest_episode["episode_index"][0]
-
-                av_size_per_frame = latest_size_in_mb / latest_num_frames if latest_num_frames > 0 else 0.0
-
-                if latest_size_in_mb + av_size_per_frame * num_frames >= self.data_files_size_in_mb:
-                    # Size limit is reached, flush buffer and prepare new parquet file
-                    self._flush_metadata_buffer()
-                    chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
-                    self._close_writer()
+            if latest_size_in_mb + ep_size_in_mb >= self.data_files_size_in_mb:
+                # Size limit is reached, prepare new parquet file
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
 
             # Update the existing pandas dataframe with new row
-            episode_dict["meta/episodes/chunk_index"] = [chunk_idx]
-            episode_dict["meta/episodes/file_index"] = [file_idx]
-            episode_dict["dataset_from_index"] = [self.latest_episode["dataset_to_index"][0]]
-            episode_dict["dataset_to_index"] = [self.latest_episode["dataset_to_index"][0] + num_frames]
+            df["meta/episodes/chunk_index"] = [chunk_idx]
+            df["meta/episodes/file_index"] = [file_idx]
+            df["dataset_from_index"] = [latest_ep["dataset_to_index"]]
+            df["dataset_to_index"] = [latest_ep["dataset_to_index"] + num_frames]
 
-        # Add to buffer
-        self.metadata_buffer.append(episode_dict)
-        self.latest_episode = episode_dict
+            if latest_size_in_mb + ep_size_in_mb < self.data_files_size_in_mb:
+                # Size limit wasnt reached, concatenate latest dataframe with new one
+                latest_df = pd.read_parquet(latest_path)
+                df = pd.concat([latest_df, df], ignore_index=True)
 
-        if len(self.metadata_buffer) >= self.metadata_buffer_size:
-            self._flush_metadata_buffer()
+                # Memort optimization
+                del latest_df
+                gc.collect()
+
+        # Write the resulting dataframe from RAM to disk
+        path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+        path.parent.mkdir(parents=True, exist_ok=True)
+        df.to_parquet(path, index=False)
+
+        if self.episodes is not None:
+            # Remove the episodes cache directory, necessary to avoid cache bloat
+            cached_dir = get_hf_dataset_cache_dir(self.episodes)
+            if cached_dir is not None:
+                shutil.rmtree(cached_dir)
+
+        self.episodes = load_episodes(self.root)
 
     def save_episode(
         self,
@@ -430,7 +366,9 @@ class LeRobotDatasetMetadata:
         video_keys = [video_key] if video_key is not None else self.video_keys
         for key in video_keys:
             if not self.features[key].get("info", None):
-                video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
+                video_path = self.root / self.video_path.format(
+                    video_key=video_key, chunk_index=0, file_index=0
+                )
                 self.info["features"][key]["info"] = get_video_info(video_path)
 
     def update_chunk_settings(
@@ -500,10 +438,6 @@ class LeRobotDatasetMetadata:
         robot_type: str | None = None,
         root: str | Path | None = None,
         use_videos: bool = True,
-        metadata_buffer_size: int = 10,
-        chunks_size: int | None = None,
-        data_files_size_in_mb: int | None = None,
-        video_files_size_in_mb: int | None = None,
     ) -> "LeRobotDatasetMetadata":
         """Creates metadata for a LeRobotDataset."""
         obj = cls.__new__(cls)
@@ -518,24 +452,11 @@ class LeRobotDatasetMetadata:
         obj.tasks = None
         obj.episodes = None
         obj.stats = None
-        obj.info = create_empty_dataset_info(
-            CODEBASE_VERSION,
-            fps,
-            features,
-            use_videos,
-            robot_type,
-            chunks_size,
-            data_files_size_in_mb,
-            video_files_size_in_mb,
-        )
+        obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, features, use_videos, robot_type)
         if len(obj.video_keys) > 0 and not use_videos:
             raise ValueError()
         write_json(obj.info, obj.root / INFO_PATH)
         obj.revision = None
-        obj.writer = None
-        obj.latest_episode = None
-        obj.metadata_buffer = []
-        obj.metadata_buffer_size = metadata_buffer_size
         return obj
 
 
@@ -682,9 +603,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # Unused attributes
         self.image_writer = None
         self.episode_buffer = None
-        self.writer = None
-        self.latest_episode = None
-        self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -693,11 +611,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             self.repo_id, self.root, self.revision, force_cache_sync=force_cache_sync
         )
 
-        # Track dataset state for efficient incremental writing
-        self._lazy_loading = False
-        self._recorded_frames = self.meta.total_frames
-        self._writer_closed_for_reading = False
-
         # Load actual data
         try:
             if force_cache_sync:
@@ -707,8 +620,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             if not self._check_cached_episodes_sufficient():
                 raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
         except (AssertionError, FileNotFoundError, NotADirectoryError):
-            if is_valid_version(self.revision):
-                self.revision = get_safe_version(self.repo_id, self.revision)
+            self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
             self.hf_dataset = self.load_hf_dataset()
 
@@ -717,19 +629,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
-    def _close_writer(self) -> None:
-        """Close and cleanup the parquet writer if it exists."""
-        writer = getattr(self, "writer", None)
-        if writer is not None:
-            writer.close()
-            self.writer = None
-
-    def __del__(self):
-        """
-        Trust the user to call .finalize() but as an added safety check call the parquet writer to stop when calling the destructor
-        """
-        self._close_writer()
-
     def push_to_hub(
         self,
         branch: str | None = None,
@@ -835,14 +734,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return hf_dataset
 
     def _check_cached_episodes_sufficient(self) -> bool:
-        """Check if the cached dataset contains all requested episodes and their video files."""
+        """Check if the cached dataset contains all requested episodes."""
         if self.hf_dataset is None or len(self.hf_dataset) == 0:
             return False
 
         # Get available episode indices from cached dataset
         available_episodes = {
             ep_idx.item() if isinstance(ep_idx, torch.Tensor) else ep_idx
-            for ep_idx in self.hf_dataset.unique("episode_index")
+            for ep_idx in self.hf_dataset["episode_index"]
         }
 
         # Determine requested episodes
@@ -854,18 +753,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             requested_episodes = set(self.episodes)
 
         # Check if all requested episodes are available in cached data
-        if not requested_episodes.issubset(available_episodes):
-            return False
-
-        # Check if all required video files exist
-        if len(self.meta.video_keys) > 0:
-            for ep_idx in requested_episodes:
-                for vid_key in self.meta.video_keys:
-                    video_path = self.root / self.meta.get_video_file_path(ep_idx, vid_key)
-                    if not video_path.exists():
-                        return False
-
-        return True
+        return requested_episodes.issubset(available_episodes)
 
     def create_hf_dataset(self) -> datasets.Dataset:
         features = get_hf_features_from_features(self.features)
@@ -881,15 +769,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     @property
     def num_frames(self) -> int:
-        """Number of frames in selected episodes.
-
-        Note: When episodes a subset of the full dataset is requested, we must return the
-        actual loaded data length (len(self.hf_dataset)) rather than metadata total_frames.
-        self.meta.total_frames is the total number of frames in the full dataset.
-        """
-        if self.episodes is not None and self.hf_dataset is not None:
-            return len(self.hf_dataset)
-        return self.meta.total_frames
+        """Number of frames in selected episodes."""
+        return len(self.hf_dataset) if self.hf_dataset is not None else self.meta.total_frames
 
     @property
     def num_episodes(self) -> int:
@@ -967,22 +848,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         return item
 
-    def _ensure_hf_dataset_loaded(self):
-        """Lazy load the HF dataset only when needed for reading."""
-        if self._lazy_loading or self.hf_dataset is None:
-            # Close the writer before loading to ensure parquet file is properly finalized
-            if self.writer is not None:
-                self._close_writer()
-                self._writer_closed_for_reading = True
-            self.hf_dataset = self.load_hf_dataset()
-            self._lazy_loading = False
+    def _add_padding_keys(self, item: dict, padding: dict[str, list[bool]]) -> dict:
+        for key, val in padding.items():
+            item[key] = torch.BoolTensor(val)
+        return item
 
     def __len__(self):
         return self.num_frames
 
     def __getitem__(self, idx) -> dict:
-        # Ensure dataset is loaded when we actually need to read from it
-        self._ensure_hf_dataset_loaded()
         item = self.hf_dataset[idx]
         ep_idx = item["episode_index"].item()
 
@@ -1021,14 +895,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             "})',\n"
         )
 
-    def finalize(self):
-        """
-        Close the parquet writers. This function needs to be called after data collection/conversion, else footer metadata won't be written to the parquet files.
-        The dataset won't be valid and can't be loaded as ds = LeRobotDataset(repo_id=repo, root=HF_LEROBOT_HOME.joinpath(repo))
-        """
-        self._close_writer()
-        self.meta._close_writer()
-
     def create_episode_buffer(self, episode_index: int | None = None) -> dict:
         current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
         ep_buffer = {}
@@ -1166,7 +1032,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Reset episode buffer and clean up temporary images (if not already deleted during video encoding)
             self.clear_episode_buffer(delete_images=len(self.meta.image_keys) > 0)
 
-    def _batch_save_episode_video(self, start_episode: int, end_episode: int | None = None) -> None:
+    def _batch_save_episode_video(self, start_episode: int, end_episode: int | None = None):
         """
         Batch save videos for multiple episodes.
 
@@ -1236,104 +1102,74 @@ class LeRobotDataset(torch.utils.data.Dataset):
         ep_dict = {key: episode_buffer[key] for key in self.hf_features}
         ep_dataset = datasets.Dataset.from_dict(ep_dict, features=self.hf_features, split="train")
         ep_dataset = embed_images(ep_dataset)
+        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
         ep_num_frames = len(ep_dataset)
+        df = pd.DataFrame(ep_dataset)
 
-        if self.latest_episode is None:
+        if self.meta.episodes is None:
             # Initialize indices and frame count for a new dataset made of the first episode data
             chunk_idx, file_idx = 0, 0
-            global_frame_index = 0
-            self._current_file_start_frame = 0
-            # However, if the episodes already exists
-            # It means we are resuming recording, so we need to load the latest episode
-            # Update the indices to avoid overwriting the latest episode
-            if self.meta.episodes is not None and len(self.meta.episodes) > 0:
-                latest_ep = self.meta.episodes[-1]
-                global_frame_index = latest_ep["dataset_to_index"]
-                chunk_idx = latest_ep["data/chunk_index"]
-                file_idx = latest_ep["data/file_index"]
-
-                # When resuming, move to the next file
-                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
-                self._current_file_start_frame = global_frame_index
+            latest_num_frames = 0
         else:
             # Retrieve information from the latest parquet file
-            latest_ep = self.latest_episode
+            latest_ep = self.meta.episodes[-1]
             chunk_idx = latest_ep["data/chunk_index"]
             file_idx = latest_ep["data/file_index"]
-            global_frame_index = latest_ep["index"][-1] + 1
 
             latest_path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
-            latest_size_in_mb = get_file_size_in_mb(latest_path)
-
-            frames_in_current_file = global_frame_index - self._current_file_start_frame
-            av_size_per_frame = (
-                latest_size_in_mb / frames_in_current_file if frames_in_current_file > 0 else 0
-            )
+            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
+            latest_num_frames = get_parquet_num_frames(latest_path)
 
             # Determine if a new parquet file is needed
-            if (
-                latest_size_in_mb + av_size_per_frame * ep_num_frames >= self.meta.data_files_size_in_mb
-                or self._writer_closed_for_reading
-            ):
-                # Size limit is reached or writer was closed for reading, prepare new parquet file
+            if latest_size_in_mb + ep_size_in_mb >= self.meta.data_files_size_in_mb:
+                # Size limit is reached, prepare new parquet file
                 chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
-                self._close_writer()
-                self._writer_closed_for_reading = False
-                self._current_file_start_frame = global_frame_index
+                latest_num_frames = 0
+            else:
+                # Update the existing parquet file with new rows
+                latest_df = pd.read_parquet(latest_path)
+                df = pd.concat([latest_df, df], ignore_index=True)
 
-        ep_dict["data/chunk_index"] = chunk_idx
-        ep_dict["data/file_index"] = file_idx
+                # Memort optimization
+                del latest_df
+                gc.collect()
 
         # Write the resulting dataframe from RAM to disk
         path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
         path.parent.mkdir(parents=True, exist_ok=True)
+        if len(self.meta.image_keys) > 0:
+            to_parquet_with_hf_images(df, path)
+        else:
+            df.to_parquet(path)
 
-        table = ep_dataset.with_format("arrow")[:]
-        if not self.writer:
-            self.writer = pq.ParquetWriter(
-                path, schema=table.schema, compression="snappy", use_dictionary=True
-            )
-        self.writer.write_table(table)
+        if self.hf_dataset is not None:
+            # Remove hf dataset cache directory, necessary to avoid cache bloat
+            cached_dir = get_hf_dataset_cache_dir(self.hf_dataset)
+            if cached_dir is not None:
+                shutil.rmtree(cached_dir)
+
+        self.hf_dataset = self.load_hf_dataset()
 
         metadata = {
             "data/chunk_index": chunk_idx,
             "data/file_index": file_idx,
-            "dataset_from_index": global_frame_index,
-            "dataset_to_index": global_frame_index + ep_num_frames,
+            "dataset_from_index": latest_num_frames,
+            "dataset_to_index": latest_num_frames + ep_num_frames,
         }
-
-        # Store metadata with episode data for next episode
-        self.latest_episode = {**ep_dict, **metadata}
-
-        # Mark that the HF dataset needs reloading (lazy loading approach)
-        # This avoids expensive reloading during sequential recording
-        self._lazy_loading = True
-        # Update recorded frames count for efficient length tracking
-        self._recorded_frames += ep_num_frames
-
         return metadata
 
-    def _save_episode_video(self, video_key: str, episode_index: int) -> dict:
+    def _save_episode_video(self, video_key: str, episode_index: int):
         # Encode episode frames into a temporary video
         ep_path = self._encode_temporary_episode_video(video_key, episode_index)
-        ep_size_in_mb = get_file_size_in_mb(ep_path)
+        ep_size_in_mb = get_video_size_in_mb(ep_path)
         ep_duration_in_s = get_video_duration_in_s(ep_path)
 
-        if (
-            episode_index == 0
-            or self.meta.latest_episode is None
-            or f"videos/{video_key}/chunk_index" not in self.meta.latest_episode
+        if self.meta.episodes is None or (
+            f"videos/{video_key}/chunk_index" not in self.meta.episodes.column_names
+            or f"videos/{video_key}/file_index" not in self.meta.episodes.column_names
         ):
             # Initialize indices for a new dataset made of the first episode data
             chunk_idx, file_idx = 0, 0
-            if self.meta.episodes is not None and len(self.meta.episodes) > 0:
-                # It means we are resuming recording, so we need to load the latest episode
-                # Update the indices to avoid overwriting the latest episode
-                old_chunk_idx = self.meta.episodes[-1][f"videos/{video_key}/chunk_index"]
-                old_file_idx = self.meta.episodes[-1][f"videos/{video_key}/file_index"]
-                chunk_idx, file_idx = update_chunk_file_indices(
-                    old_chunk_idx, old_file_idx, self.meta.chunks_size
-                )
             latest_duration_in_s = 0.0
             new_path = self.root / self.meta.video_path.format(
                 video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
@@ -1341,16 +1177,16 @@ class LeRobotDataset(torch.utils.data.Dataset):
             new_path.parent.mkdir(parents=True, exist_ok=True)
             shutil.move(str(ep_path), str(new_path))
         else:
-            # Retrieve information from the latest updated video file using latest_episode
-            latest_ep = self.meta.latest_episode
-            chunk_idx = latest_ep[f"videos/{video_key}/chunk_index"][0]
-            file_idx = latest_ep[f"videos/{video_key}/file_index"][0]
+            # Retrieve information from the latest updated video file (possibly several episodes ago)
+            latest_ep = self.meta.episodes[episode_index - 1]
+            chunk_idx = latest_ep[f"videos/{video_key}/chunk_index"]
+            file_idx = latest_ep[f"videos/{video_key}/file_index"]
 
             latest_path = self.root / self.meta.video_path.format(
                 video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
             )
-            latest_size_in_mb = get_file_size_in_mb(latest_path)
-            latest_duration_in_s = latest_ep[f"videos/{video_key}/to_timestamp"][0]
+            latest_size_in_mb = get_video_size_in_mb(latest_path)
+            latest_duration_in_s = get_video_duration_in_s(latest_path)
 
             if latest_size_in_mb + ep_size_in_mb >= self.meta.video_files_size_in_mb:
                 # Move temporary episode video to a new video file in the dataset
@@ -1427,7 +1263,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if self.image_writer is not None:
             self.image_writer.wait_until_done()
 
-    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
+    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> dict:
         """
         Use ffmpeg to convert frames stored as png into mp4 videos.
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
@@ -1484,13 +1320,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.delta_timestamps = None
         obj.delta_indices = None
         obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
-        obj.writer = None
-        obj.latest_episode = None
-        obj._current_file_start_frame = None
-        # Initialize tracking for incremental recording
-        obj._lazy_loading = False
-        obj._recorded_frames = 0
-        obj._writer_closed_for_reading = False
         return obj
 
 
@@ -1567,6 +1396,11 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         """
         return {repo_id: i for i, repo_id in enumerate(self.repo_ids)}
 
+    @property
+    def repo_index_to_id(self):
+        """Return the inverse mapping if repo_id_to_index."""
+        return {v: k for k, v in self.repo_id_to_index}
+
     @property
     def fps(self) -> int:
         """Frames per second used during data collection.
@@ -1597,7 +1431,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         """Keys to access image and video stream from cameras."""
         keys = []
         for key, feats in self.features.items():
-            if isinstance(feats, (datasets.Image | VideoFrame)):
+            if isinstance(feats, (datasets.Image, VideoFrame)):
                 keys.append(key)
         return keys
 

src/lerobot/datasets/pipeline_features.py

@@ -17,9 +17,9 @@ from collections.abc import Sequence
 from typing import Any
 
 from lerobot.configs.types import PipelineFeatureType
+from lerobot.constants import ACTION, OBS_IMAGES, OBS_STATE
 from lerobot.datasets.utils import hw_to_dataset_features
 from lerobot.processor import DataProcessorPipeline
-from lerobot.utils.constants import ACTION, OBS_IMAGES, OBS_STATE, OBS_STR
 
 
 def create_initial_features(
@@ -92,8 +92,8 @@ def aggregate_pipeline_dataset_features(
 
     # Intermediate storage for categorized and filtered features.
     processed_features: dict[str, dict[str, Any]] = {
-        ACTION: {},
-        OBS_STR: {},
+        "action": {},
+        "observation": {},
     }
     images_token = OBS_IMAGES.split(".")[-1]
 
@@ -125,15 +125,17 @@ def aggregate_pipeline_dataset_features(
             # 3. Add the feature to the appropriate group with a clean name.
             name = strip_prefix(key, PREFIXES_TO_STRIP)
             if is_action:
-                processed_features[ACTION][name] = value
+                processed_features["action"][name] = value
             else:
-                processed_features[OBS_STR][name] = value
+                processed_features["observation"][name] = value
 
     # Convert the processed features into the final dataset format.
     dataset_features = {}
-    if processed_features[ACTION]:
-        dataset_features.update(hw_to_dataset_features(processed_features[ACTION], ACTION, use_videos))
-    if processed_features[OBS_STR]:
-        dataset_features.update(hw_to_dataset_features(processed_features[OBS_STR], OBS_STR, use_videos))
+    if processed_features["action"]:
+        dataset_features.update(hw_to_dataset_features(processed_features["action"], ACTION, use_videos))
+    if processed_features["observation"]:
+        dataset_features.update(
+            hw_to_dataset_features(processed_features["observation"], "observation", use_videos)
+        )
 
     return dataset_features

src/lerobot/datasets/push_dataset_to_hub/utils.py

@@ -13,10 +13,67 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import inspect
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
 
 import datasets
+import numpy
+import PIL
 import torch
 
+from lerobot.datasets.video_utils import encode_video_frames
+
+
+def concatenate_episodes(ep_dicts):
+    data_dict = {}
+
+    keys = ep_dicts[0].keys()
+    for key in keys:
+        if torch.is_tensor(ep_dicts[0][key][0]):
+            data_dict[key] = torch.cat([ep_dict[key] for ep_dict in ep_dicts])
+        else:
+            if key not in data_dict:
+                data_dict[key] = []
+            for ep_dict in ep_dicts:
+                for x in ep_dict[key]:
+                    data_dict[key].append(x)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def save_images_concurrently(imgs_array: numpy.array, out_dir: Path, max_workers: int = 4):
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    def save_image(img_array, i, out_dir):
+        img = PIL.Image.fromarray(img_array)
+        img.save(str(out_dir / f"frame_{i:06d}.png"), quality=100)
+
+    num_images = len(imgs_array)
+    with ThreadPoolExecutor(max_workers=max_workers) as executor:
+        [executor.submit(save_image, imgs_array[i], i, out_dir) for i in range(num_images)]
+
+
+def get_default_encoding() -> dict:
+    """Returns the default ffmpeg encoding parameters used by `encode_video_frames`."""
+    signature = inspect.signature(encode_video_frames)
+    return {
+        k: v.default
+        for k, v in signature.parameters.items()
+        if v.default is not inspect.Parameter.empty and k in ["vcodec", "pix_fmt", "g", "crf"]
+    }
+
+
+def check_repo_id(repo_id: str) -> None:
+    if len(repo_id.split("/")) != 2:
+        raise ValueError(
+            f"""`repo_id` is expected to contain a community or user id `/` the name of the dataset
+            (e.g. 'lerobot/pusht'), but contains '{repo_id}'."""
+        )
+
 
 # TODO(aliberts): remove
 def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> dict[str, torch.Tensor]:

src/lerobot/datasets/streaming_dataset.py

@@ -21,6 +21,7 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
+from lerobot.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
 from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
     Backtrackable,
@@ -37,7 +38,6 @@ from lerobot.datasets.video_utils import (
     VideoDecoderCache,
     decode_video_frames_torchcodec,
 )
-from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
 
 
 class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
@@ -298,7 +298,9 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
 
         return padding_mask
 
-    def make_frame(self, dataset_iterator: Backtrackable) -> Generator:
+    def make_frame(
+        self, dataset_iterator: Backtrackable, previous_dataset_iterator: Backtrackable | None = None
+    ) -> Generator:
         """Makes a frame starting from a dataset iterator"""
         item = next(dataset_iterator)
         item = item_to_torch(item)