Merge branch 'main' into feat/add-multidataset-training

.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 }}
@@ -159,35 +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
-        run: pytest tests/training/test_multi_gpu.py -vv --maxfail=3
-        timeout-minutes: 10

.github/workflows/release.yml

@@ -103,7 +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
@@ -111,7 +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
@@ -138,7 +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
@@ -197,23 +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
 ```
 
@@ -221,7 +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
 
@@ -310,7 +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
 

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
@@ -17,35 +19,20 @@
     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)
-  title: "Policies"
-- sections:
-  - local: il_sim
-    title: Imitation Learning in Sim
+    title: Finetune SmolVLA
   - 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/hilserl.mdx

@@ -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**

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/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
@@ -98,7 +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

@@ -33,7 +33,7 @@ To Install LIBERO, after following LeRobot official instructions, just do:
 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 \
@@ -52,7 +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 \
@@ -103,11 +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/ \
@@ -125,42 +124,3 @@ lerobot-train \
 LeRobot uses MuJoCo for simulation. You need to set the rendering backend before training or evaluation:
 
 - `export MUJOCO_GL=egl` → for headless servers (e.g. HPC, cloud)
-
-## Reproducing 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** |

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

@@ -1,79 +0,0 @@
-# π₀ (Pi0)
-
-π₀ is a **Vision-Language-Action model for general robot control**, from Physical Intelligence. The LeRobot implementation is adapted from their open source [OpenPI](https://github.com/Physical-Intelligence/openpi) repository.
-
-## 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.
-
-### The Vision for Physical Intelligence
-
-As described by Physical Intelligence, while AI has achieved remarkable success in digital domains, from chess-playing to drug discovery, human intelligence still dramatically outpaces AI in the physical world. To paraphrase Moravec's paradox, winning a game of chess represents an "easy" problem for AI, but folding a shirt or cleaning up a table requires solving some of the most difficult engineering problems ever conceived. π₀ represents a first step toward developing artificial physical intelligence that enables users to simply ask robots to perform any task they want, just like they can with large language models.
-
-### Architecture and Approach
-
-π₀ combines several key innovations:
-
-- **Flow Matching**: Uses a novel method to augment pre-trained VLMs with continuous action outputs via flow matching (a variant of diffusion models)
-- **Cross-Embodiment Training**: Trained on data from 8 distinct robot platforms including UR5e, Bimanual UR5e, Franka, Bimanual Trossen, Bimanual ARX, Mobile Trossen, and Mobile Fibocom
-- **Internet-Scale Pre-training**: Inherits semantic knowledge from a pre-trained 3B parameter Vision-Language Model
-- **High-Frequency Control**: Outputs motor commands at up to 50 Hz for real-time dexterous manipulation
-
-## Installation Requirements
-
-1. Install LeRobot by following our [Installation Guide](./installation).
-2. Install Pi0 dependencies by running:
-
-   ```bash
-   pip install -e ".[pi]"
-   ```
-
-## Training Data and Capabilities
-
-π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:
-
-1. **Internet-Scale Pre-training**: Vision-language data from the web for semantic understanding
-2. **Open X-Embodiment Dataset**: Open-source robot manipulation datasets
-3. **Physical Intelligence Dataset**: Large and diverse dataset of dexterous tasks across 8 distinct robots
-
-## Usage
-
-To use π₀ in LeRobot, specify the policy type as:
-
-```python
-policy.type=pi0
-```
-
-## Training
-
-For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
-
-```bash
-python src/lerobot/scripts/lerobot_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.repo_id=your_repo_id \
-    --policy.compile_model=true \
-    --policy.gradient_checkpointing=true \
-    --policy.dtype=bfloat16 \
-    --steps=3000 \
-    --policy.device=cuda \
-    --batch_size=32
-```
-
-### Key Training Parameters
-
-- **`--policy.compile_model=true`**: Enables model compilation for faster training
-- **`--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)
-
-## License
-
-This model follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

docs/source/pi05.mdx

@@ -1,107 +0,0 @@
-# π₀.₅ (Pi05) Policy
-
-π₀.₅ is a **Vision-Language-Action model with open-world generalization**, from Physical Intelligence. The LeRobot implementation is adapted from their open source [OpenPI](https://github.com/Physical-Intelligence/openpi) repository.
-
-## Model Overview
-
-π₀.₅ represents a significant evolution from π₀, developed by [Physical Intelligence](https://www.physicalintelligence.company/blog/pi05) to address a big challenge in robotics: **open-world generalization**. While robots can perform impressive tasks in controlled environments, π₀.₅ is designed to generalize to entirely new environments and situations that were never seen during training.
-
-### The Generalization Challenge
-
-As Physical Intelligence explains, the fundamental challenge isn't performing tasks of agility or dexterity, but generalization, the ability to correctly perform tasks in new settings with new objects. Consider a robot cleaning different homes: each home has different objects in different places. Generalization must occur at multiple levels:
-
-- **Physical Level**: Understanding how to pick up a spoon (by the handle) or plate (by the edge), even with unseen objects in cluttered environments
-- **Semantic Level**: Understanding task semantics, where to put clothes and shoes (laundry hamper, not on the bed), and what tools are appropriate for cleaning spills
-- **Environmental Level**: Adapting to "messy" real-world environments like homes, grocery stores, offices, and hospitals
-
-### Co-Training on Heterogeneous Data
-
-The breakthrough innovation in π₀.₅ is **co-training on heterogeneous data sources**. The model learns from:
-
-1. **Multimodal Web Data**: Image captioning, visual question answering, object detection
-2. **Verbal Instructions**: Humans coaching robots through complex tasks step-by-step
-3. **Subtask Commands**: High-level semantic behavior labels (e.g., "pick up the pillow" for an unmade bed)
-4. **Cross-Embodiment Robot Data**: Data from various robot platforms with different capabilities
-5. **Multi-Environment Data**: Static robots deployed across many different homes
-6. **Mobile Manipulation Data**: ~400 hours of mobile robot demonstrations
-
-This diverse training mixture creates a "curriculum" that enables generalization across physical, visual, and semantic levels simultaneously.
-
-## Installation Requirements
-
-1. Install LeRobot by following our [Installation Guide](./installation).
-2. Install Pi0.5 dependencies by running:
-
-   ```bash
-   pip install -e ".[pi]"
-   ```
-
-## Usage
-
-To use π₀.₅ in your LeRobot configuration, specify the policy type as:
-
-```python
-policy.type=pi05
-```
-
-## Training
-
-### Training Command Example
-
-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\
-    --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 \
-    --policy.compile_model=true \
-    --policy.gradient_checkpointing=true \
-    --wandb.enable=true \
-    --policy.dtype=bfloat16 \
-    --steps=3000 \
-    --policy.device=cuda \
-    --batch_size=32
-```
-
-### Key Training Parameters
-
-- **`--policy.compile_model=true`**: Enables model compilation for faster training
-- **`--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"}'`
-
-## Performance Results
-
-### Libero Benchmark Results
-
-π₀.₅ has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the libero base model for an additional 6k steps on the Libero dataset and compared the results to the OpenPI reference results.
-
-| Benchmark          | LeRobot Implementation | OpenPI Reference |
-| ------------------ | ---------------------- | ---------------- |
-| **Libero Spatial** | 97.0%                  | 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%           |
-
-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.
-
-## License
-
-This model follows the **Apache 2.0 License**, consistent with the original [OpenPI repository](https://github.com/Physical-Intelligence/openpi).

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

@@ -1,4 +1,4 @@
-# SmolVLA
+# Finetune SmolVLA
 
 SmolVLA is Hugging Face’s lightweight foundation model for robotics. Designed for easy fine-tuning on LeRobot datasets, it helps accelerate your development!
 
@@ -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/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/evaluate/evaluate_libero.py

@@ -1,112 +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 script demonstrates how to evaluate pretrained vision-language-action (VLA) policies
-such as SmolVLA on Libero benchmark tasks using the LeRobot framework.
-
-It showcases the full evaluation pipeline — from environment creation to policy inference,
-visualization, and result logging — and is intended as a reference for benchmarking or
-integrating new robotic policies.
-
-Features included in this script:
-- loading Libero environments (e.g., libero_spatial, libero_object) via `make_env`.
-- initializing pretrained policies (e.g., SmolVLA) from Hugging Face using `make_policy`.
-- applying preprocessing and postprocessing transformations for model compatibility.
-- running evaluation rollouts and recording rendered frames from the simulator.
-- computing success metrics and saving rollout videos as MP4 for qualitative analysis.
-
-The script ends by saving a rollout video (`rollout.mp4`) and printing per-environment
-success indicators for quick visual and numerical evaluation.
-"""
-
-import numpy as np
-import torch
-import imageio.v2 as imageio
-from lerobot.envs.factory import make_env, make_env_config
-from lerobot.policies.factory import make_policy, make_pre_post_processors
-from lerobot.policies.factory import make_policy_config
-from lerobot.envs.utils import (
-    add_envs_task,
-    preprocess_observation,
-)
-import os 
-os.environ["MUJOCO_GL"] = "egl"
-
-SMOLVLA_LIBERO_PATH = "HuggingFaceVLA/smolvla_libero"
-LIBERO_CONFIG = make_env_config("libero", task="libero_spatial")
-breakpoint()
-POLICY_CONFIG = make_policy_config("smolvla", pretrained_path=SMOLVLA_LIBERO_PATH)
-policy = make_policy(
-        cfg=POLICY_CONFIG,
-        env_cfg=LIBERO_CONFIG,
-)
-breakpoint()
-libero_env = make_env(LIBERO_CONFIG)
-breakpoint()
-print(type(libero_env)) # <class 'dict'>
-print(libero_env.keys()) # dict_keys(['libero_spatial', 'libero_object'])
-
-# initilize your policy, here we use smolvla
-breakpoint()
-policy.eval()
-preprocessor, postprocessor = make_pre_post_processors(
-        policy_cfg=POLICY_CONFIG,
-        pretrained_path=SMOLVLA_LIBERO_PATH,
-        # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
-        preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
-    )
-policy.reset()
-# for the sake of this exemple we only use one env from each task
-libero_spatial_env = libero_env['libero_spatial'][0]
-# libero_object_env = libero_env['libero_object'][0]
-
-# let's first run an evaluation throgut the first task
-observation, info = libero_spatial_env.reset() # you can pass seeds
-max_steps = 220
-step = 0
-all_images = []
-done = np.array([False] * libero_spatial_env.num_envs)
-while not np.all(done) and step < max_steps:
-    observation = preprocess_observation(observation)
-    observation = add_envs_task(libero_spatial_env, observation)
-    observation = preprocessor(observation)
-    with torch.inference_mode():
-            action = policy.select_action(observation)
-    action = postprocessor(action)
-    # Convert to CPU / numpy.
-    action_numpy = action.to("cpu").numpy() 
-    # Apply the next action.
-    # let's render the video
-    image = libero_spatial_env.call("render")[0]
-    all_images.append(image)
-    observation, reward, terminated, truncated, info = libero_spatial_env.step(action_numpy)
-    if "final_info" in info:
-            final_info = info["final_info"]
-            if not isinstance(final_info, dict):
-                raise RuntimeError(
-                    "Unsupported `final_info` format: expected dict (Gymnasium >= 1.0). "
-                    "You're likely using an older version of gymnasium (< 1.0). Please upgrade."
-                )
-            successes = final_info["is_success"].tolist()
-    else:
-        successes = [False] * libero_spatial_env.num_envs
-    
-    done = terminated | truncated | done
-    if step + 1 == max_steps:
-        done = np.ones_like(done, dtype=bool)
-    step += 1
-
-print("The success: ", successes)
-

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/tester.py

@@ -0,0 +1,66 @@
+from lerobot.datasets.lerobot_dataset import MultiLeRobotDataset
+
+REPO_A = "lerobot/pusht"
+REPO_B = "lerobot/aloha_mobile_cabinet"  # replace with the actual repo id
+
+feature_keys_mapping = {
+    REPO_A: {  # pusht (1 camera, 2-dim)
+        "action": "actions",
+        "observation.state": "obs_state",
+        "observation.image": "obs_image.cam_high",
+    },
+    REPO_B: {  # dual arm (3 cameras, 14-dim)
+        "action": "actions",
+        "observation.state": "obs_state",
+        "observation.images.cam_high": "obs_image.cam_high",
+        "observation.images.cam_left_wrist": "obs_image.cam_left_wrist",
+        "observation.images.cam_right_wrist": "obs_image.cam_right_wrist",
+    },
+}
+
+from torchvision.transforms.v2 import Compose, ToImage, Resize
+image_tf = Compose([
+    ToImage(),          # converts to tensor if needed
+    Resize((224, 224)), # unify sizes across datasets (96x96 vs 480x640)
+])
+
+from torch.utils.data import DataLoader
+
+dataset = MultiLeRobotDataset(
+    repo_ids=[REPO_A, REPO_B],
+    image_transforms=image_tf,              # ensures same HxW
+    feature_keys_mapping=feature_keys_mapping,
+    train_on_all_features=True,             # keep union of cameras; zero-fill missing
+    # optional: override if you want fixed maxima; else inferred:
+    # max_action_dim=14,
+    # max_state_dim=14,
+    max_action_dim=14,
+    max_state_dim=14,
+    max_image_dim=224,
+    ignore_keys=[
+        "next.*", # drop reward/done/success
+        "index",
+        "timestamp",
+        "videos/*",               # drop all video metadata
+        "observation.effort",     # 👈 drop effort everywhere
+    ],
+)
+breakpoint()
+loader = DataLoader(dataset, batch_size=8, shuffle=True, num_workers=0, pin_memory=True)
+for _ in range(100):
+    batch = next(iter(loader))
+
+breakpoint()
+# vectors padded to maxima (pusht:2 -> 14; dual-arm:14 -> 14)
+assert batch["actions"].shape[-1] == 14
+assert batch["obs_state"].shape[-1] == 14
+assert batch["actions_padding_mask"].shape[-1] == 14
+assert batch["obs_state_padding_mask"].shape[-1] == 14
+
+# cameras: all canonical keys exist; pusht will have wrists zero-filled
+for cam in ["obs_image.cam_high", "obs_image.cam_left_wrist", "obs_image.cam_right_wrist"]:
+    assert cam in batch
+    assert f"{cam}_is_pad" in batch
+    # images should all be 3x224x224 (or your transform’s size)
+    img = batch[cam]
+    assert img.ndim in (4, 5)  # (B,C,H,W) or (B,T,C,H,W) depending on your loader

examples/tester.sh

@@ -0,0 +1,16 @@
+# storage / caches
+RAID=/raid/jade
+export TRANSFORMERS_CACHE=$RAID/.cache/huggingface/transformers
+export HF_HOME=$RAID/.cache/huggingface
+export HF_DATASETS_CACHE=$RAID/.cache/huggingface/datasets
+export HF_LEROBOT_HOME=$RAID/.cache/huggingface/lerobot
+export WANDB_CACHE_DIR=$RAID/.cache/wandb
+export TMPDIR=$RAID/.cache/tmp
+mkdir -p $TMPDIR
+export WANDB_MODE=offline
+# export HF_DATASETS_OFFLINE=1
+# export HF_HUB_OFFLINE=1
+export TOKENIZERS_PARALLELISM=false
+export MUJOCO_GL=egl
+
+python examples/tester.py

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():

pyproject.toml

@@ -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>=1.0.0",
-    "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",
@@ -94,45 +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>=4.52.0"]
+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"]
+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"]
-hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.11,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+pi0 = ["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
+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"]
-metaworld = ["metaworld>=3.0.0"]
+
 
 # All
 all = [
@@ -143,7 +147,7 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    "lerobot[pi]",
+    "lerobot[pi0]",
     "lerobot[smolvla]",
     "lerobot[hilserl]",
     "lerobot[async]",
@@ -152,26 +156,22 @@ 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-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"
 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"
 
 # ---------------- Tool Configurations ----------------
 [tool.setuptools.packages.find]
@@ -198,7 +198,7 @@ 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
@@ -229,6 +229,9 @@ 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"]
 
@@ -243,7 +246,6 @@ default.extend-ignore-identifiers-re = [
     "pn",
     "ser",
     "ein",
-    "inpt",
 ]
 
 # TODO: Uncomment when ready to use
@@ -262,86 +264,8 @@ 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
-
-# [[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

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/opencv/camera_opencv.py

@@ -31,7 +31,7 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
 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

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -31,7 +31,7 @@ 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

src/lerobot/cameras/realsense/camera_realsense.py

@@ -29,7 +29,7 @@ try:
 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

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
 
@@ -44,10 +44,7 @@ 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
 

src/lerobot/configs/default.py

@@ -16,6 +16,9 @@
 
 from dataclasses import dataclass, field
 
+from lerobot import (
+    policies,  # noqa: F401
+)
 from lerobot.datasets.transforms import ImageTransformsConfig
 from lerobot.datasets.video_utils import get_safe_default_codec
 

src/lerobot/configs/policies.py

@@ -27,9 +27,9 @@ 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
 
@@ -71,11 +71,9 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     tags: list[str] | None = None
     # Add tags to your policy on the hub.
     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: str | None = None
 
     def __post_init__(self):
+        self.pretrained_path = None
         if not self.device or not is_torch_device_available(self.device):
             auto_device = auto_select_torch_device()
             logging.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")

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):
@@ -35,8 +36,10 @@ class NormalizationMode(str, Enum):
     MIN_MAX = "MIN_MAX"
     MEAN_STD = "MEAN_STD"
     IDENTITY = "IDENTITY"
-    QUANTILES = "QUANTILES"
-    QUANTILE10 = "QUANTILE10"
+
+
+class DictLike(Protocol):
+    def __getitem__(self, key: Any) -> Any: ...
 
 
 @dataclass

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

@@ -17,179 +17,6 @@ import numpy as np
 
 from lerobot.datasets.utils import load_image_as_numpy
 
-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.
-    """
-
-    def __init__(self, quantile_list: list[float] | None = None, num_quantile_bins: int = 5000):
-        self._count = 0
-        self._mean = None
-        self._mean_of_squares = None
-        self._min = None
-        self._max = None
-        self._histograms = None
-        self._bin_edges = None
-        self._num_quantile_bins = num_quantile_bins
-
-        self._quantile_list = quantile_list
-        if self._quantile_list is None:
-            self._quantile_list = DEFAULT_QUANTILES
-        self._quantile_keys = [f"q{int(q * 100):02d}" for q in self._quantile_list]
-
-    def update(self, batch: np.ndarray) -> None:
-        """Update the running statistics with a batch of vectors.
-
-        Args:
-            batch: An array where all dimensions except the last are batch dimensions.
-        """
-        batch = batch.reshape(-1, batch.shape[-1])
-        num_elements, vector_length = batch.shape
-
-        if self._count == 0:
-            self._mean = np.mean(batch, axis=0)
-            self._mean_of_squares = np.mean(batch**2, axis=0)
-            self._min = np.min(batch, axis=0)
-            self._max = np.max(batch, axis=0)
-            self._histograms = [np.zeros(self._num_quantile_bins) for _ in range(vector_length)]
-            self._bin_edges = [
-                np.linspace(self._min[i] - 1e-10, self._max[i] + 1e-10, self._num_quantile_bins + 1)
-                for i in range(vector_length)
-            ]
-        else:
-            if vector_length != self._mean.size:
-                raise ValueError("The length of new vectors does not match the initialized vector length.")
-
-            new_max = np.max(batch, axis=0)
-            new_min = np.min(batch, axis=0)
-            max_changed = np.any(new_max > self._max)
-            min_changed = np.any(new_min < self._min)
-            self._max = np.maximum(self._max, new_max)
-            self._min = np.minimum(self._min, new_min)
-
-            if max_changed or min_changed:
-                self._adjust_histograms()
-
-        self._count += num_elements
-
-        batch_mean = np.mean(batch, axis=0)
-        batch_mean_of_squares = np.mean(batch**2, axis=0)
-
-        # Update running mean and mean of squares
-        self._mean += (batch_mean - self._mean) * (num_elements / self._count)
-        self._mean_of_squares += (batch_mean_of_squares - self._mean_of_squares) * (
-            num_elements / self._count
-        )
-
-        self._update_histograms(batch)
-
-    def get_statistics(self) -> dict[str, np.ndarray]:
-        """Compute and return the statistics of the vectors processed so far.
-
-        Args:
-            quantiles: List of quantiles to compute (e.g., [0.01, 0.10, 0.50, 0.90, 0.99]). If None, no quantiles computed.
-
-        Returns:
-            Dictionary containing the computed statistics.
-        """
-        if self._count < 2:
-            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 = {
-            "min": self._min.copy(),
-            "max": self._max.copy(),
-            "mean": self._mean.copy(),
-            "std": stddev,
-            "count": np.array([self._count]),
-        }
-
-        quantile_results = self._compute_quantiles()
-        for i, q in enumerate(self._quantile_keys):
-            stats[q] = quantile_results[i]
-
-        return stats
-
-    def _adjust_histograms(self):
-        """Adjust histograms when min or max changes."""
-        for i in range(len(self._histograms)):
-            old_edges = self._bin_edges[i]
-            old_hist = self._histograms[i]
-
-            # Create new edges with small padding to ensure range coverage
-            padding = (self._max[i] - self._min[i]) * 1e-10
-            new_edges = np.linspace(
-                self._min[i] - padding, self._max[i] + padding, self._num_quantile_bins + 1
-            )
-
-            # Redistribute existing histogram counts to new bins
-            # We need to map each old bin center to the new bins
-            old_centers = (old_edges[:-1] + old_edges[1:]) / 2
-            new_hist = np.zeros(self._num_quantile_bins)
-
-            for old_center, count in zip(old_centers, old_hist, strict=False):
-                if count > 0:
-                    # Find which new bin this old center belongs to
-                    bin_idx = np.searchsorted(new_edges, old_center) - 1
-                    bin_idx = max(0, min(bin_idx, self._num_quantile_bins - 1))
-                    new_hist[bin_idx] += count
-
-            self._histograms[i] = new_hist
-            self._bin_edges[i] = new_edges
-
-    def _update_histograms(self, batch: np.ndarray) -> None:
-        """Update histograms with new vectors."""
-        for i in range(batch.shape[1]):
-            hist, _ = np.histogram(batch[:, i], bins=self._bin_edges[i])
-            self._histograms[i] += hist
-
-    def _compute_quantiles(self) -> list[np.ndarray]:
-        """Compute quantiles based on histograms."""
-        results = []
-        for q in self._quantile_list:
-            target_count = q * self._count
-            q_values = []
-
-            for hist, edges in zip(self._histograms, self._bin_edges, strict=True):
-                q_value = self._compute_single_quantile(hist, edges, target_count)
-                q_values.append(q_value)
-
-            results.append(np.array(q_values))
-        return results
-
-    def _compute_single_quantile(self, hist: np.ndarray, edges: np.ndarray, target_count: float) -> float:
-        """Compute a single quantile value from histogram and bin edges."""
-        cumsum = np.cumsum(hist)
-        idx = np.searchsorted(cumsum, target_count)
-
-        if idx == 0:
-            return edges[0]
-        if idx >= len(cumsum):
-            return edges[-1]
-
-        # If not edge case, interpolate within the bin
-        count_before = cumsum[idx - 1]
-        count_in_bin = cumsum[idx] - count_before
-
-        # If no samples in this bin, use the bin edge
-        if count_in_bin == 0:
-            return edges[idx]
-
-        # Linear interpolation within the bin
-        fraction = (target_count - count_before) / count_in_bin
-        return edges[idx] + fraction * (edges[idx + 1] - edges[idx])
-
 
 def estimate_num_samples(
     dataset_len: int, min_num_samples: int = 100, max_num_samples: int = 10_000, power: float = 0.75
@@ -245,282 +72,33 @@ def sample_images(image_paths: list[str]) -> np.ndarray:
     return images
 
 
-def _reshape_stats_by_axis(
-    stats: dict[str, np.ndarray],
-    axis: int | tuple[int, ...] | None,
-    keepdims: bool,
-    original_shape: tuple[int, ...],
-) -> dict[str, np.ndarray]:
-    """Reshape all statistics to match NumPy's output conventions.
-
-    Applies consistent reshaping to all statistics (except 'count') based on the
-    axis and keepdims parameters. This ensures statistics have the correct shape
-    for broadcasting with the original data.
-
-    Args:
-        stats: Dictionary of computed statistics
-        axis: Axis or axes along which statistics were computed
-        keepdims: Whether to keep reduced dimensions as size-1 dimensions
-        original_shape: Shape of the original array
-
-    Returns:
-        Dictionary with reshaped statistics
-
-    Note:
-        The 'count' statistic is never reshaped as it represents metadata
-        rather than per-feature statistics.
-    """
-    if axis == (1,) and not keepdims:
-        return stats
-
-    result = {}
-    for key, value in stats.items():
-        if key == "count":
-            result[key] = value
-        else:
-            result[key] = _reshape_single_stat(value, axis, keepdims, original_shape)
-
-    return result
-
-
-def _reshape_for_image_stats(value: np.ndarray, keepdims: bool) -> np.ndarray:
-    """Reshape statistics for image data (axis=(0,2,3))."""
-    if keepdims and value.ndim == 1:
-        return value.reshape(1, -1, 1, 1)
-    return value
-
-
-def _reshape_for_vector_stats(
-    value: np.ndarray, keepdims: bool, original_shape: tuple[int, ...]
-) -> np.ndarray:
-    """Reshape statistics for vector data (axis=0 or axis=(0,))."""
-    if not keepdims:
-        return value
-
-    if len(original_shape) == 1 and value.ndim > 0:
-        return value.reshape(1)
-    elif len(original_shape) >= 2 and value.ndim == 1:
-        return value.reshape(1, -1)
-    return value
-
-
-def _reshape_for_feature_stats(value: np.ndarray, keepdims: bool) -> np.ndarray:
-    """Reshape statistics for feature-wise computation (axis=(1,))."""
-    if not keepdims:
-        return value
-
-    if value.ndim == 0:
-        return value.reshape(1, 1)
-    elif value.ndim == 1:
-        return value.reshape(-1, 1)
-    return value
-
-
-def _reshape_for_global_stats(
-    value: np.ndarray, keepdims: bool, original_shape: tuple[int, ...]
-) -> np.ndarray | float:
-    """Reshape statistics for global reduction (axis=None)."""
-    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)
-
-
-def _reshape_single_stat(
-    value: np.ndarray, axis: int | tuple[int, ...] | None, keepdims: bool, original_shape: tuple[int, ...]
-) -> np.ndarray | float:
-    """Apply appropriate reshaping to a single statistic array.
-
-    This function transforms statistic arrays to match expected output shapes
-    based on the axis configuration and keepdims parameter.
-
-    Args:
-        value: The statistic array to reshape
-        axis: Axis or axes that were reduced during computation
-        keepdims: Whether to maintain reduced dimensions as size-1 dimensions
-        original_shape: Shape of the original data before reduction
-
-    Returns:
-        Reshaped array following NumPy broadcasting conventions
-
-    """
-    if axis == (0, 2, 3):
-        return _reshape_for_image_stats(value, keepdims)
-
-    if axis in [0, (0,)]:
-        return _reshape_for_vector_stats(value, keepdims, original_shape)
-
-    if axis == (1,):
-        return _reshape_for_feature_stats(value, keepdims)
-
-    if axis is None:
-        return _reshape_for_global_stats(value, keepdims, original_shape)
-
-    return value
-
-
-def _prepare_array_for_stats(array: np.ndarray, axis: int | tuple[int, ...] | None) -> tuple[np.ndarray, int]:
-    """Prepare array for statistics computation by reshaping according to axis.
-
-    Args:
-        array: Input data array
-        axis: Axis or axes along which to compute statistics
-
-    Returns:
-        Tuple of (reshaped_array, sample_count)
-    """
-    if axis == (0, 2, 3):  # Image data
-        batch_size, channels, height, width = array.shape
-        reshaped = array.transpose(0, 2, 3, 1).reshape(-1, channels)
-        return reshaped, batch_size
-
-    if axis == 0 or axis == (0,):  # Vector data
-        reshaped = array
-        if array.ndim == 1:
-            reshaped = array.reshape(-1, 1)
-        return reshaped, array.shape[0]
-
-    if axis == (1,):  # Feature-wise statistics
-        return array.T, array.shape[1]
-
-    if axis is None:  # Global statistics
-        reshaped = array.reshape(-1, 1)
-        # For backward compatibility, count represents the first dimension size
-        return reshaped, array.shape[0] if array.ndim > 0 else 1
-
-    raise ValueError(f"Unsupported axis configuration: {axis}")
-
-
-def _compute_basic_stats(
-    array: np.ndarray, sample_count: int, quantile_list: list[float] | None = None
-) -> dict[str, np.ndarray]:
-    """Compute basic statistics for arrays with insufficient samples for quantiles.
-
-    Args:
-        array: Reshaped array ready for statistics computation
-        sample_count: Number of samples represented in the data
-
-    Returns:
-        Dictionary with basic statistics and quantiles set to mean values
-    """
-    if quantile_list is None:
-        quantile_list = DEFAULT_QUANTILES
-    quantile_list_keys = [f"q{int(q * 100):02d}" for q in quantile_list]
-
-    stats = {
-        "min": np.min(array, axis=0),
-        "max": np.max(array, axis=0),
-        "mean": np.mean(array, axis=0),
-        "std": np.std(array, axis=0),
-        "count": np.array([sample_count]),
+def get_feature_stats(array: np.ndarray, axis: tuple, keepdims: bool) -> dict[str, np.ndarray]:
+    return {
+        "min": np.min(array, axis=axis, keepdims=keepdims),
+        "max": np.max(array, axis=axis, keepdims=keepdims),
+        "mean": np.mean(array, axis=axis, keepdims=keepdims),
+        "std": np.std(array, axis=axis, keepdims=keepdims),
+        "count": np.array([len(array)]),
     }
 
-    for q in quantile_list_keys:
-        stats[q] = stats["mean"].copy()
-
-    return stats
-
-
-def get_feature_stats(
-    array: np.ndarray,
-    axis: int | tuple[int, ...] | None,
-    keepdims: bool,
-    quantile_list: list[float] | None = None,
-) -> dict[str, np.ndarray]:
-    """Compute comprehensive statistics for array features along specified axes.
-
-    This function calculates min, max, mean, std, and quantiles (1%, 10%, 50%, 90%, 99%)
-    for the input array along the specified axes. It handles different data layouts:
-    - Image data: axis=(0,2,3) computes per-channel statistics
-    - Vector data: axis=0 computes per-feature statistics
-    - Feature-wise: axis=1 computes statistics across features
-    - Global: axis=None computes statistics over entire array
-
-    Args:
-        array: Input data array with shape appropriate for the specified axis
-        axis: Axis or axes along which to compute statistics
-            - (0, 2, 3): For image data (batch, channels, height, width)
-            - 0 or (0,): For vector/tabular data (samples, features)
-            - (1,): For computing across features
-            - None: For global statistics over entire array
-        keepdims: If True, reduced axes are kept as dimensions with size 1
-
-    Returns:
-        Dictionary containing:
-            - 'min': Minimum values
-            - 'max': Maximum values
-            - 'mean': Mean values
-            - 'std': Standard deviation
-            - 'count': Number of samples (always shape (1,))
-            - 'q01', 'q10', 'q50', 'q90', 'q99': Quantile values
-
-    """
-    if quantile_list is None:
-        quantile_list = DEFAULT_QUANTILES
-
-    original_shape = array.shape
-    reshaped, sample_count = _prepare_array_for_stats(array, axis)
-
-    if reshaped.shape[0] < 2:
-        stats = _compute_basic_stats(reshaped, sample_count, quantile_list)
-    else:
-        running_stats = RunningQuantileStats()
-        running_stats.update(reshaped)
-        stats = running_stats.get_statistics()
-        stats["count"] = np.array([sample_count])
-
-    stats = _reshape_stats_by_axis(stats, axis, keepdims, original_shape)
-    return stats
-
-
-def compute_episode_stats(
-    episode_data: dict[str, list[str] | np.ndarray],
-    features: dict,
-    quantile_list: list[float] | None = None,
-) -> dict:
-    """Compute comprehensive statistics for all features in an episode.
-
-    Processes different data types appropriately:
-    - Images/videos: Samples from paths, computes per-channel stats, normalizes to [0,1]
-    - Numerical arrays: Computes per-feature statistics
-    - Strings: Skipped (no statistics computed)
-
-    Args:
-        episode_data: Dictionary mapping feature names to data
-            - For images/videos: list of file paths
-            - For numerical data: numpy arrays
-        features: Dictionary describing each feature's dtype and shape
-
-    Returns:
-        Dictionary mapping feature names to their statistics dictionaries.
-        Each statistics dictionary contains min, max, mean, std, count, and quantiles.
-
-    Note:
-        Image statistics are normalized to [0,1] range and have shape (3,1,1) for
-        per-channel values when dtype is 'image' or 'video'.
-    """
-    if quantile_list is None:
-        quantile_list = DEFAULT_QUANTILES
 
+def compute_episode_stats(episode_data: dict[str, list[str] | np.ndarray], features: dict) -> dict:
     ep_stats = {}
     for key, data in episode_data.items():
         if features[key]["dtype"] == "string":
-            continue
-
-        if features[key]["dtype"] in ["image", "video"]:
-            ep_ft_array = sample_images(data)
-            axes_to_reduce = (0, 2, 3)
+            continue  # HACK: we should receive np.arrays of strings
+        elif features[key]["dtype"] in ["image", "video"]:
+            ep_ft_array = sample_images(data)  # data is a list of image paths
+            axes_to_reduce = (0, 2, 3)  # keep channel dim
             keepdims = True
         else:
-            ep_ft_array = data
-            axes_to_reduce = 0
-            keepdims = data.ndim == 1
+            ep_ft_array = data  # data is already a np.ndarray
+            axes_to_reduce = 0  # compute stats over the first axis
+            keepdims = data.ndim == 1  # keep as np.array
 
-        ep_stats[key] = get_feature_stats(
-            ep_ft_array, axis=axes_to_reduce, keepdims=keepdims, quantile_list=quantile_list
-        )
+        ep_stats[key] = get_feature_stats(ep_ft_array, axis=axes_to_reduce, keepdims=keepdims)
 
+        # finally, we normalize and remove batch dim for images
         if features[key]["dtype"] in ["image", "video"]:
             ep_stats[key] = {
                 k: v if k == "count" else np.squeeze(v / 255.0, axis=0) for k, v in ep_stats[key].items()
@@ -529,37 +107,20 @@ def compute_episode_stats(
     return ep_stats
 
 
-def _validate_stat_value(value: np.ndarray, key: str, feature_key: str) -> None:
-    """Validate a single statistic value."""
-    if not isinstance(value, np.ndarray):
-        raise ValueError(
-            f"Stats must be composed of numpy array, but key '{key}' of feature '{feature_key}' "
-            f"is of type '{type(value)}' instead."
-        )
-
-    if value.ndim == 0:
-        raise ValueError("Number of dimensions must be at least 1, and is 0 instead.")
-
-    if key == "count" and value.shape != (1,):
-        raise ValueError(f"Shape of 'count' must be (1), but is {value.shape} instead.")
-
-    if "image" in feature_key and key != "count" and value.shape != (3, 1, 1):
-        raise ValueError(f"Shape of quantile '{key}' must be (3,1,1), but is {value.shape} instead.")
-
-
 def _assert_type_and_shape(stats_list: list[dict[str, dict]]):
-    """Validate that all statistics have correct types and shapes.
-
-    Args:
-        stats_list: List of statistics dictionaries to validate
-
-    Raises:
-        ValueError: If any statistic has incorrect type or shape
-    """
-    for stats in stats_list:
-        for feature_key, feature_stats in stats.items():
-            for stat_key, stat_value in feature_stats.items():
-                _validate_stat_value(stat_value, stat_key, feature_key)
+    for i in range(len(stats_list)):
+        for fkey in stats_list[i]:
+            for k, v in stats_list[i][fkey].items():
+                if not isinstance(v, np.ndarray):
+                    raise ValueError(
+                        f"Stats must be composed of numpy array, but key '{k}' of feature '{fkey}' is of type '{type(v)}' instead."
+                    )
+                if v.ndim == 0:
+                    raise ValueError("Number of dimensions must be at least 1, and is 0 instead.")
+                if k == "count" and v.shape != (1,):
+                    raise ValueError(f"Shape of 'count' must be (1), but is {v.shape} instead.")
+                if "image" in fkey and k != "count" and v.shape != (3, 1, 1):
+                    raise ValueError(f"Shape of '{k}' must be (3,1,1), but is {v.shape} instead.")
 
 
 def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
@@ -582,7 +143,7 @@ def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, d
     weighted_variances = (variances + delta_means**2) * counts
     total_variance = weighted_variances.sum(axis=0) / total_count
 
-    aggregated = {
+    return {
         "min": np.min(np.stack([s["min"] for s in stats_ft_list]), axis=0),
         "max": np.max(np.stack([s["max"] for s in stats_ft_list]), axis=0),
         "mean": total_mean,
@@ -590,17 +151,6 @@ def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, d
         "count": total_count,
     }
 
-    if stats_ft_list:
-        quantile_keys = [k for k in stats_ft_list[0] 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):
-                quantile_values = np.stack([s[q_key] for s in stats_ft_list])
-                weighted_quantiles = quantile_values * counts
-                aggregated[q_key] = weighted_quantiles.sum(axis=0) / total_count
-
-    return aggregated
-
 
 def aggregate_stats(stats_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
     """Aggregate stats from multiple compute_stats outputs into a single set of stats.
@@ -624,3 +174,79 @@ def aggregate_stats(stats_list: list[dict[str, dict]]) -> dict[str, dict[str, np
         aggregated_stats[key] = aggregate_feature_stats(stats_with_key)
 
     return aggregated_stats
+
+import numpy as np
+
+def aggregate_stats_multi(
+    stats_list: list[dict[str, dict]],
+    max_action_dim: int | None = None,
+    max_state_dim: int | None = None,
+) -> dict[str, dict[str, np.ndarray]]:
+    """Aggregate stats from multiple compute_stats outputs into a single set of stats.
+    
+    Supports heterogeneous robots by padding action/state stats to the max dim.
+    The final stats will have the union of all data keys from each of the stats dicts.
+
+    - new_min  = elementwise min across datasets
+    - new_max  = elementwise max across datasets
+    - new_mean = weighted mean (by count)
+    - new_std  = recomputed from total variance
+    """
+
+    data_keys = {key for stats in stats_list for key in stats}
+    aggregated_stats = {key: {} for key in data_keys}
+
+    def _pad(arr: np.ndarray, target: int) -> np.ndarray:
+        if arr.ndim == 0:  # scalar
+            return arr
+        if target is None or target <= 0 or arr.shape[-1] == target:
+            return arr
+        pad_width = [(0, 0)] * arr.ndim
+        pad_width[-1] = (0, target - arr.shape[-1])
+        return np.pad(arr, pad_width, mode="constant")
+
+    for key in data_keys:
+        stats_with_key = [stats[key] for stats in stats_list if key in stats]
+
+        # decide if this key should be padded
+        target_dim = None
+        if "action" in key and max_action_dim:
+            target_dim = max_action_dim
+        elif "state" in key and max_state_dim:
+            target_dim = max_state_dim
+
+        padded = []
+        counts = []
+        for s in stats_with_key:
+            mean = _pad(np.array(s["mean"]), target_dim)
+            std = _pad(np.array(s["std"]), target_dim)
+            min_ = _pad(np.array(s["min"]), target_dim)
+            max_ = _pad(np.array(s["max"]), target_dim)
+            count = s.get("count", 1)
+
+            padded.append(dict(mean=mean, std=std, min=min_, max=max_, count=count))
+            counts.append(count)
+
+        counts = np.array(counts, dtype=np.float64)
+        total_count = counts.sum()
+
+        means = np.stack([p["mean"] for p in padded])
+        stds  = np.stack([p["std"]  for p in padded])
+        mins  = np.stack([p["min"]  for p in padded])
+        maxs  = np.stack([p["max"]  for p in padded])
+
+        # weighted mean (broadcast weights properly)
+        new_mean = np.average(means, axis=0, weights=counts)
+        new_var  = np.average(stds**2 + (means - new_mean)**2, axis=0, weights=counts)
+
+        new_std = np.sqrt(new_var)
+
+        aggregated_stats[key] = {
+            "min": mins.min(axis=0),
+            "max": maxs.max(axis=0),
+            "mean": new_mean,
+            "std": new_std,
+            "count": int(total_count),
+        }
+
+    return aggregated_stats

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/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)

src/lerobot/datasets/transforms.py

@@ -120,7 +120,7 @@ class SharpnessJitter(Transform):
         self.sharpness = self._check_input(sharpness)
 
     def _check_input(self, sharpness):
-        if isinstance(sharpness, (int | float)):
+        if isinstance(sharpness, (int, float)):
             if sharpness < 0:
                 raise ValueError("If sharpness is a single number, it must be non negative.")
             sharpness = [1.0 - sharpness, 1.0 + sharpness]
@@ -206,11 +206,6 @@ class ImageTransformsConfig:
                 type="SharpnessJitter",
                 kwargs={"sharpness": (0.5, 1.5)},
             ),
-            "affine": ImageTransformConfig(
-                weight=1.0,
-                type="RandomAffine",
-                kwargs={"degrees": (-5.0, 5.0), "translate": (0.05, 0.05)},
-            ),
         }
     )
 
@@ -222,8 +217,6 @@ def make_transform_from_config(cfg: ImageTransformConfig):
         return v2.ColorJitter(**cfg.kwargs)
     elif cfg.type == "SharpnessJitter":
         return SharpnessJitter(**cfg.kwargs)
-    elif cfg.type == "RandomAffine":
-        return v2.RandomAffine(**cfg.kwargs)
     else:
         raise ValueError(f"Transform '{cfg.type}' is not valid.")
 

src/lerobot/datasets/utils.py

@@ -21,7 +21,7 @@ from collections import deque
 from collections.abc import Iterable, Iterator
 from pathlib import Path
 from pprint import pformat
-from typing import Any, Generic, TypeVar
+from typing import Any, Deque, Generic, TypeVar
 
 import datasets
 import numpy as np
@@ -30,7 +30,7 @@ import pandas
 import pandas as pd
 import pyarrow.parquet as pq
 import torch
-from datasets import Dataset
+from datasets import Dataset, concatenate_datasets
 from datasets.table import embed_table_storage
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
@@ -43,8 +43,7 @@ from lerobot.datasets.backward_compatibility import (
     BackwardCompatibilityError,
     ForwardCompatibilityError,
 )
-from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_STR
-from lerobot.utils.utils import SuppressProgressBars, is_valid_numpy_dtype_string
+from lerobot.utils.utils import is_valid_numpy_dtype_string
 
 DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
 DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
@@ -67,6 +66,18 @@ DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{fram
 LEGACY_EPISODES_PATH = "meta/episodes.jsonl"
 LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
 LEGACY_TASKS_PATH = "meta/tasks.jsonl"
+LEGACY_DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
+LEGACY_DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+
+DATASET_CARD_TEMPLATE = """
+---
+# Metadata will go there
+---
+This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
+
+## {}
+
+"""
 
 DEFAULT_FEATURES = {
     "timestamp": {"dtype": "float32", "shape": (1,), "names": None},
@@ -94,6 +105,12 @@ def get_hf_dataset_size_in_mb(hf_ds: Dataset) -> int:
     return hf_ds.data.nbytes // (1024**2)
 
 
+def get_hf_dataset_cache_dir(hf_ds: Dataset) -> Path | None:
+    if hf_ds.cache_files is None or len(hf_ds.cache_files) == 0:
+        return None
+    return Path(hf_ds.cache_files[0]["filename"]).parents[2]
+
+
 def update_chunk_file_indices(chunk_idx: int, file_idx: int, chunks_size: int) -> tuple[int, int]:
     if file_idx == chunks_size - 1:
         file_idx = 0
@@ -117,9 +134,8 @@ def load_nested_dataset(pq_dir: Path, features: datasets.Features | None = None)
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
     # TODO(rcadene): set num_proc to accelerate conversion to pyarrow
-    with SuppressProgressBars():
-        datasets = Dataset.from_parquet([str(path) for path in paths], features=features)
-    return datasets
+    datasets = [Dataset.from_parquet(str(path), features=features) for path in paths]
+    return concatenate_datasets(datasets)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -127,14 +143,10 @@ def get_parquet_num_frames(parquet_path: str | Path) -> int:
     return metadata.num_rows
 
 
-def get_file_size_in_mb(file_path: Path) -> float:
-    """Get file size on disk in megabytes.
-
-    Args:
-        file_path (Path): Path to the file.
-    """
-    file_size_bytes = file_path.stat().st_size
-    return file_size_bytes / (1024**2)
+def get_video_size_in_mb(mp4_path: Path) -> float:
+    file_size_bytes = mp4_path.stat().st_size
+    file_size_mb = file_size_bytes / (1024**2)
+    return file_size_mb
 
 
 def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
@@ -206,13 +218,13 @@ def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
     """
     serialized_dict = {}
     for key, value in flatten_dict(stats).items():
-        if isinstance(value, (torch.Tensor | np.ndarray)):
+        if isinstance(value, (torch.Tensor, np.ndarray)):
             serialized_dict[key] = value.tolist()
-        elif isinstance(value, list) and isinstance(value[0], (int | float | list)):
+        elif isinstance(value, list) and isinstance(value[0], (int, float, list)):
             serialized_dict[key] = value
         elif isinstance(value, np.generic):
             serialized_dict[key] = value.item()
-        elif isinstance(value, (int | float)):
+        elif isinstance(value, (int, float)):
             serialized_dict[key] = value
         else:
             raise NotImplementedError(f"The value '{value}' of type '{type(value)}' is not supported.")
@@ -370,6 +382,12 @@ def load_episodes(local_dir: Path) -> datasets.Dataset:
     return episodes
 
 
+def backward_compatible_episodes_stats(
+    stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
+) -> dict[int, dict[str, dict[str, np.ndarray]]]:
+    return dict.fromkeys(episodes, stats)
+
+
 def load_image_as_numpy(
     fpath: str | Path, dtype: np.dtype = np.float32, channel_first: bool = True
 ) -> np.ndarray:
@@ -627,14 +645,14 @@ def hw_to_dataset_features(
     }
     cam_fts = {key: shape for key, shape in hw_features.items() if isinstance(shape, tuple)}
 
-    if joint_fts and prefix == ACTION:
+    if joint_fts and prefix == "action":
         features[prefix] = {
             "dtype": "float32",
             "shape": (len(joint_fts),),
             "names": list(joint_fts),
         }
 
-    if joint_fts and prefix == OBS_STR:
+    if joint_fts and prefix == "observation":
         features[f"{prefix}.state"] = {
             "dtype": "float32",
             "shape": (len(joint_fts),),
@@ -710,11 +728,11 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
             # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
             if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
                 shape = (shape[2], shape[0], shape[1])
-        elif key == OBS_ENV_STATE:
+        elif key == "observation.environment_state":
             type = FeatureType.ENV
-        elif key.startswith(OBS_STR):
+        elif key.startswith("observation"):
             type = FeatureType.STATE
-        elif key.startswith(ACTION):
+        elif key.startswith("action"):
             type = FeatureType.ACTION
         else:
             continue
@@ -1178,7 +1196,7 @@ def item_to_torch(item: dict) -> dict:
         dict: Dictionary with all tensor-like items converted to torch.Tensor.
     """
     for key, val in item.items():
-        if isinstance(val, (np.ndarray | list)) and key not in ["task"]:
+        if isinstance(val, (np.ndarray, list)) and key not in ["task"]:
             # Convert numpy arrays and lists to torch tensors
             item[key] = torch.tensor(val)
     return item
@@ -1252,8 +1270,8 @@ class Backtrackable(Generic[T]):
             raise ValueError("lookahead must be > 0")
 
         self._source: Iterator[T] = iter(iterable)
-        self._back_buf: deque[T] = deque(maxlen=history)
-        self._ahead_buf: deque[T] = deque(maxlen=lookahead) if lookahead > 0 else deque()
+        self._back_buf: Deque[T] = deque(maxlen=history)
+        self._ahead_buf: Deque[T] = deque(maxlen=lookahead) if lookahead > 0 else deque()
         self._cursor: int = 0
         self._history = history
         self._lookahead = lookahead
@@ -1327,6 +1345,12 @@ class Backtrackable(Generic[T]):
         # When cursor<0, slice so the order remains chronological
         return list(self._back_buf)[: self._cursor or None]
 
+    def lookahead_buffer(self) -> list[T]:
+        """
+        Return a copy of the current lookahead buffer.
+        """
+        return list(self._ahead_buf)
+
     def can_peek_back(self, steps: int = 1) -> bool:
         """
         Check if we can go back `steps` items without raising an IndexError.
@@ -1352,6 +1376,31 @@ class Backtrackable(Generic[T]):
         except StopIteration:
             return False
 
+    def reset_cursor(self) -> None:
+        """
+        Reset cursor to the most recent position (equivalent to calling next()
+        until you're back to the latest item).
+        """
+        self._cursor = 0
+
+    def clear_ahead_buffer(self) -> None:
+        """
+        Clear the ahead buffer, discarding any pre-fetched items.
+        """
+        self._ahead_buf.clear()
+
+    def switch_source_iterable(self, new_source: Iterable[T]) -> None:
+        """
+        Switch the source of the backtrackable to a new iterable, keeping the history.
+
+        This is useful when iterating over a sequence of datasets. The history from the
+        previous source is kept, but the lookahead buffer is cleared. The cursor is reset
+        to the present.
+        """
+        self._source = iter(new_source)
+        self.clear_ahead_buffer()
+        self.reset_cursor()
+
 
 def safe_shard(dataset: datasets.IterableDataset, index: int, num_shards: int) -> datasets.Dataset:
     """

src/lerobot/datasets/v30/augment_dataset_quantile_stats.py

@@ -1,260 +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.
-
-"""
-This script augments existing LeRobot datasets with quantile statistics.
-
-Most datasets created before the quantile feature was added do not contain
-quantile statistics (q01, q10, q50, q90, q99) in their metadata. This script:
-
-1. Loads an existing LeRobot dataset in v3.0 format
-2. Checks if it already contains quantile statistics
-3. If missing, computes quantile statistics for all features
-4. Updates the dataset metadata with the new quantile statistics
-
-Usage:
-
-```bash
-python src/lerobot/datasets/v30/augment_dataset_quantile_stats.py \
-    --repo-id=lerobot/pusht \
-```
-"""
-
-import argparse
-import concurrent.futures
-import logging
-from pathlib import Path
-
-import numpy as np
-import torch
-from huggingface_hub import HfApi
-from requests import HTTPError
-from tqdm import tqdm
-
-from lerobot.datasets.compute_stats import DEFAULT_QUANTILES, aggregate_stats, get_feature_stats
-from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.datasets.utils import write_stats
-from lerobot.utils.utils import init_logging
-
-
-def has_quantile_stats(stats: dict[str, dict] | None, quantile_list_keys: list[str] | None = None) -> bool:
-    """Check if dataset statistics already contain quantile information.
-
-    Args:
-        stats: Dataset statistics dictionary
-
-    Returns:
-        True if quantile statistics are present, False otherwise
-    """
-    if quantile_list_keys is None:
-        quantile_list_keys = [f"q{int(q * 100):02d}" for q in DEFAULT_QUANTILES]
-
-    if stats is None:
-        return False
-
-    for feature_stats in stats.values():
-        if any(q_key in feature_stats for q_key in quantile_list_keys):
-            return True
-
-    return False
-
-
-def process_single_episode(dataset: LeRobotDataset, episode_idx: int) -> dict:
-    """Process a single episode and return its statistics.
-
-    Args:
-        dataset: The LeRobot dataset
-        episode_idx: Index of the episode to process
-
-    Returns:
-        Dictionary containing episode statistics
-    """
-    logging.info(f"Computing stats for episode {episode_idx}")
-
-    start_idx = dataset.meta.episodes[episode_idx]["dataset_from_index"]
-    end_idx = dataset.meta.episodes[episode_idx]["dataset_to_index"]
-
-    collected_data: dict[str, list] = {}
-    for idx in range(start_idx, end_idx):
-        item = dataset[idx]
-        for key, value in item.items():
-            if key not in dataset.features:
-                continue
-
-            if key not in collected_data:
-                collected_data[key] = []
-            collected_data[key].append(value)
-
-    ep_stats = {}
-    for key, data_list in collected_data.items():
-        if dataset.features[key]["dtype"] == "string":
-            continue
-
-        data = torch.stack(data_list).cpu().numpy()
-        if dataset.features[key]["dtype"] in ["image", "video"]:
-            if data.dtype == np.uint8:
-                data = data.astype(np.float32) / 255.0
-
-            axes_to_reduce = (0, 2, 3)
-            keepdims = True
-        else:
-            axes_to_reduce = 0
-            keepdims = data.ndim == 1
-
-        ep_stats[key] = get_feature_stats(
-            data, axis=axes_to_reduce, keepdims=keepdims, quantile_list=DEFAULT_QUANTILES
-        )
-
-        if dataset.features[key]["dtype"] in ["image", "video"]:
-            ep_stats[key] = {
-                k: v if k == "count" else np.squeeze(v, axis=0) for k, v in ep_stats[key].items()
-            }
-
-    return ep_stats
-
-
-def compute_quantile_stats_for_dataset(dataset: LeRobotDataset) -> dict[str, dict]:
-    """Compute quantile statistics for all episodes in the dataset.
-
-    Args:
-        dataset: The LeRobot dataset to compute statistics for
-
-    Returns:
-        Dictionary containing aggregated statistics with quantiles
-
-    Note:
-        Video decoding operations are not thread-safe, so we process episodes sequentially
-        when video keys are present. For datasets without videos, we use parallel processing
-        with ThreadPoolExecutor for better performance.
-    """
-    logging.info(f"Computing quantile statistics for dataset with {dataset.num_episodes} episodes")
-
-    episode_stats_list = []
-    has_videos = len(dataset.meta.video_keys) > 0
-
-    if has_videos:
-        logging.info("Dataset contains video keys - using sequential processing for thread safety")
-        for episode_idx in tqdm(range(dataset.num_episodes), desc="Processing episodes"):
-            ep_stats = process_single_episode(dataset, episode_idx)
-            episode_stats_list.append(ep_stats)
-    else:
-        logging.info("Dataset has no video keys - using parallel processing for better performance")
-        max_workers = min(dataset.num_episodes, 16)
-
-        with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
-            future_to_episode = {
-                executor.submit(process_single_episode, dataset, episode_idx): episode_idx
-                for episode_idx in range(dataset.num_episodes)
-            }
-
-            episode_results = {}
-            with tqdm(total=dataset.num_episodes, desc="Processing episodes") as pbar:
-                for future in concurrent.futures.as_completed(future_to_episode):
-                    episode_idx = future_to_episode[future]
-                    ep_stats = future.result()
-                    episode_results[episode_idx] = ep_stats
-                    pbar.update(1)
-
-        for episode_idx in range(dataset.num_episodes):
-            if episode_idx in episode_results:
-                episode_stats_list.append(episode_results[episode_idx])
-
-    if not episode_stats_list:
-        raise ValueError("No episode data found for computing statistics")
-
-    logging.info(f"Aggregating statistics from {len(episode_stats_list)} episodes")
-    return aggregate_stats(episode_stats_list)
-
-
-def augment_dataset_with_quantile_stats(
-    repo_id: str,
-    root: str | Path | None = None,
-    overwrite: bool = False,
-) -> None:
-    """Augment a dataset with quantile statistics if they are missing.
-
-    Args:
-        repo_id: Repository ID of the dataset
-        root: Local root directory for the dataset
-        overwrite: Overwrite existing quantile statistics if they already exist
-    """
-    logging.info(f"Loading dataset: {repo_id}")
-    dataset = LeRobotDataset(
-        repo_id=repo_id,
-        root=root,
-    )
-
-    if not overwrite and has_quantile_stats(dataset.meta.stats):
-        logging.info("Dataset already contains quantile statistics. No action needed.")
-        return
-
-    logging.info("Dataset does not contain quantile statistics. Computing them now...")
-
-    new_stats = compute_quantile_stats_for_dataset(dataset)
-
-    logging.info("Updating dataset metadata with new quantile statistics")
-    dataset.meta.stats = new_stats
-
-    write_stats(new_stats, dataset.meta.root)
-
-    logging.info("Successfully updated dataset with quantile statistics")
-    dataset.push_to_hub()
-
-    hub_api = HfApi()
-    try:
-        hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
-    except HTTPError as e:
-        logging.info(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
-        pass
-    hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=None, repo_type="dataset")
-
-
-def main():
-    """Main function to run the augmentation script."""
-    parser = argparse.ArgumentParser(description="Augment LeRobot dataset with quantile statistics")
-
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        required=True,
-        help="Repository ID of the dataset (e.g., 'lerobot/pusht')",
-    )
-
-    parser.add_argument(
-        "--root",
-        type=str,
-        help="Local root directory for the dataset",
-    )
-    parser.add_argument(
-        "--overwrite",
-        action="store_true",
-        help="Overwrite existing quantile statistics if they already exist",
-    )
-
-    args = parser.parse_args()
-    root = Path(args.root) if args.root else None
-
-    init_logging()
-
-    augment_dataset_with_quantile_stats(
-        repo_id=args.repo_id,
-        root=root,
-        overwrite=args.overwrite,
-    )
-
-
-if __name__ == "__main__":
-    main()

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -26,24 +26,14 @@ This script will help you convert any LeRobot dataset already pushed to the hub
 
 Usage:
 
-Convert a dataset from the hub:
 ```bash
 python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
     --repo-id=lerobot/pusht
 ```
 
-Convert a local dataset (works in place):
-```bash
-python src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py \
-    --repo-id=lerobot/pusht \
-    --root=/path/to/local/dataset/directory
-    --push-to-hub=false
-```
-
 """
 
 import argparse
-import logging
 import shutil
 from pathlib import Path
 from typing import Any
@@ -56,6 +46,7 @@ from datasets import Dataset, Features, Image
 from huggingface_hub import HfApi, snapshot_download
 from requests import HTTPError
 
+from lerobot.constants import HF_LEROBOT_HOME
 from lerobot.datasets.compute_stats import aggregate_stats
 from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
 from lerobot.datasets.utils import (
@@ -69,9 +60,9 @@ from lerobot.datasets.utils import (
     LEGACY_TASKS_PATH,
     cast_stats_to_numpy,
     flatten_dict,
-    get_file_size_in_mb,
     get_parquet_file_size_in_mb,
     get_parquet_num_frames,
+    get_video_size_in_mb,
     load_info,
     update_chunk_file_indices,
     write_episodes,
@@ -80,11 +71,9 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
-from lerobot.utils.constants import HF_LEROBOT_HOME
-from lerobot.utils.utils import init_logging
 
 V21 = "v2.1"
-V30 = "v3.0"
+
 
 """
 -------------------------
@@ -98,7 +87,7 @@ OLD
 videos/chunk-000/CAMERA/episode_000000.mp4
 
 NEW
-videos/CAMERA/chunk-000/file_000.mp4
+videos/chunk-000/file_000.mp4
 -------------------------
 OLD
 episodes.jsonl
@@ -154,19 +143,7 @@ def legacy_load_tasks(local_dir: Path) -> tuple[dict, dict]:
     return tasks, task_to_task_index
 
 
-def validate_local_dataset_version(local_path: Path) -> None:
-    """Validate that the local dataset has the expected v2.1 version."""
-    info = load_info(local_path)
-    dataset_version = info.get("codebase_version", "unknown")
-    if dataset_version != V21:
-        raise ValueError(
-            f"Local dataset has codebase version '{dataset_version}', expected '{V21}'. "
-            f"This script is specifically for converting v2.1 datasets to v3.0."
-        )
-
-
 def convert_tasks(root, new_root):
-    logging.info(f"Converting tasks from {root} to {new_root}")
     tasks, _ = legacy_load_tasks(root)
     task_indices = tasks.keys()
     task_strings = tasks.values()
@@ -208,10 +185,7 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
     num_frames = 0
     paths_to_cat = []
     episodes_metadata = []
-
-    logging.info(f"Converting data files from {len(ep_paths)} episodes")
-
-    for ep_path in tqdm.tqdm(ep_paths, desc="convert data files"):
+    for ep_path in ep_paths:
         ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
         ep_num_frames = get_parquet_num_frames(ep_path)
         ep_metadata = {
@@ -235,6 +209,7 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
 
         # Reset for the next file
         size_in_mb = ep_size_in_mb
+        num_frames = ep_num_frames
         paths_to_cat = [ep_path]
 
         chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
@@ -261,8 +236,6 @@ def get_image_keys(root):
 
 
 def convert_videos(root: Path, new_root: Path, video_file_size_in_mb: int):
-    logging.info(f"Converting videos from {root} to {new_root}")
-
     video_keys = get_video_keys(root)
     if len(video_keys) == 0:
         return None
@@ -281,7 +254,7 @@ def convert_videos(root: Path, new_root: Path, video_file_size_in_mb: int):
     episods_metadata = []
     num_cameras = len(video_keys)
     num_episodes = num_eps_per_cam[0]
-    for ep_idx in tqdm.tqdm(range(num_episodes), desc="convert videos"):
+    for ep_idx in range(num_episodes):
         # Sanity check
         ep_ids = [eps_metadata_per_cam[cam_idx][ep_idx]["episode_index"] for cam_idx in range(num_cameras)]
         ep_ids += [ep_idx]
@@ -308,9 +281,8 @@ def convert_videos_of_camera(root: Path, new_root: Path, video_key: str, video_f
     duration_in_s = 0.0
     paths_to_cat = []
     episodes_metadata = []
-
     for ep_path in tqdm.tqdm(ep_paths, desc=f"convert videos of {video_key}"):
-        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)
 
         # Check if adding this episode would exceed the limit
@@ -402,8 +374,6 @@ def generate_episode_metadata_dict(
 
 
 def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_metadata=None):
-    logging.info(f"Converting episodes metadata from {root} to {new_root}")
-
     episodes_legacy_metadata = legacy_load_episodes(root)
     episodes_stats = legacy_load_episodes_stats(root)
 
@@ -427,15 +397,14 @@ def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_
 
 def convert_info(root, new_root, data_file_size_in_mb, video_file_size_in_mb):
     info = load_info(root)
-    info["codebase_version"] = V30
+    info["codebase_version"] = "v3.0"
     del info["total_chunks"]
     del info["total_videos"]
     info["data_files_size_in_mb"] = data_file_size_in_mb
     info["video_files_size_in_mb"] = video_file_size_in_mb
     info["data_path"] = DEFAULT_DATA_PATH
-    info["video_path"] = DEFAULT_VIDEO_PATH if info["video_path"] is not None else None
+    info["video_path"] = DEFAULT_VIDEO_PATH
     info["fps"] = int(info["fps"])
-    logging.info(f"Converting info from {root} to {new_root}")
     for key in info["features"]:
         if info["features"][key]["dtype"] == "video":
             # already has fps in video_info
@@ -449,36 +418,16 @@ def convert_dataset(
     branch: str | None = None,
     data_file_size_in_mb: int | None = None,
     video_file_size_in_mb: int | None = None,
-    root: str | Path | None = None,
-    push_to_hub: bool = True,
-    force_conversion: bool = False,
 ):
+    root = HF_LEROBOT_HOME / repo_id
+    old_root = HF_LEROBOT_HOME / f"{repo_id}_old"
+    new_root = HF_LEROBOT_HOME / f"{repo_id}_v30"
+
     if data_file_size_in_mb is None:
         data_file_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
     if video_file_size_in_mb is None:
         video_file_size_in_mb = DEFAULT_VIDEO_FILE_SIZE_IN_MB
 
-    # First check if the dataset already has a v3.0 version
-    if root is None and not force_conversion:
-        try:
-            print("Trying to download v3.0 version of the dataset from the hub...")
-            snapshot_download(repo_id, repo_type="dataset", revision=V30, local_dir=HF_LEROBOT_HOME / repo_id)
-            return
-        except Exception:
-            print("Dataset does not have an uploaded v3.0 version. Continuing with conversion.")
-
-    # Set root based on whether local dataset path is provided
-    use_local_dataset = False
-    root = HF_LEROBOT_HOME / repo_id if root is None else Path(root) / repo_id
-    if root.exists():
-        validate_local_dataset_version(root)
-        use_local_dataset = True
-        print(f"Using local dataset at {root}")
-
-    old_root = root.parent / f"{root.name}_old"
-    new_root = root.parent / f"{root.name}_v30"
-
-    # Handle old_root cleanup if both old_root and root exist
     if old_root.is_dir() and root.is_dir():
         shutil.rmtree(str(root))
         shutil.move(str(old_root), str(root))
@@ -486,13 +435,12 @@ def convert_dataset(
     if new_root.is_dir():
         shutil.rmtree(new_root)
 
-    if not use_local_dataset:
-        snapshot_download(
-            repo_id,
-            repo_type="dataset",
-            revision=V21,
-            local_dir=root,
-        )
+    snapshot_download(
+        repo_id,
+        repo_type="dataset",
+        revision=V21,
+        local_dir=root,
+    )
 
     convert_info(root, new_root, data_file_size_in_mb, video_file_size_in_mb)
     convert_tasks(root, new_root)
@@ -503,26 +451,24 @@ def convert_dataset(
     shutil.move(str(root), str(old_root))
     shutil.move(str(new_root), str(root))
 
-    if push_to_hub:
-        hub_api = HfApi()
-        try:
-            hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
-        except HTTPError as e:
-            print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
-            pass
-        hub_api.delete_files(
-            delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*"],
-            repo_id=repo_id,
-            revision=branch,
-            repo_type="dataset",
-        )
-        hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
+    hub_api = HfApi()
+    try:
+        hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
+    except HTTPError as e:
+        print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
+        pass
+    hub_api.delete_files(
+        delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*"],
+        repo_id=repo_id,
+        revision=branch,
+        repo_type="dataset",
+    )
+    hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
 
-        LeRobotDataset(repo_id).push_to_hub()
+    LeRobotDataset(repo_id).push_to_hub()
 
 
 if __name__ == "__main__":
-    init_logging()
     parser = argparse.ArgumentParser()
     parser.add_argument(
         "--repo-id",
@@ -549,23 +495,6 @@ if __name__ == "__main__":
         default=None,
         help="File size in MB. Defaults to 100 for data and 500 for videos.",
     )
-    parser.add_argument(
-        "--root",
-        type=str,
-        default=None,
-        help="Local directory to use for downloading/writing the dataset.",
-    )
-    parser.add_argument(
-        "--push-to-hub",
-        type=lambda input: input.lower() == "true",
-        default=True,
-        help="Push the converted dataset to the hub.",
-    )
-    parser.add_argument(
-        "--force-conversion",
-        action="store_true",
-        help="Force conversion even if the dataset already has a v3.0 version.",
-    )
 
     args = parser.parse_args()
     convert_dataset(**vars(args))

src/lerobot/datasets/video_utils.py

@@ -342,8 +342,8 @@ def encode_video_frames(
     # Define video output frame size (assuming all input frames are the same size)
     if len(input_list) == 0:
         raise FileNotFoundError(f"No images found in {imgs_dir}.")
-    with Image.open(input_list[0]) as dummy_image:
-        width, height = dummy_image.size
+    dummy_image = Image.open(input_list[0])
+    width, height = dummy_image.size
 
     # Define video codec options
     video_options = {}
@@ -373,12 +373,11 @@ def encode_video_frames(
 
         # Loop through input frames and encode them
         for input_data in input_list:
-            with Image.open(input_data) as input_image:
-                input_image = input_image.convert("RGB")
-                input_frame = av.VideoFrame.from_image(input_image)
-                packet = output_stream.encode(input_frame)
-                if packet:
-                    output.mux(packet)
+            input_image = Image.open(input_data).convert("RGB")
+            input_frame = av.VideoFrame.from_image(input_image)
+            packet = output_stream.encode(input_frame)
+            if packet:
+                output.mux(packet)
 
         # Flush the encoder
         packet = output_stream.encode()
@@ -429,7 +428,7 @@ def concatenate_video_files(
     with tempfile.NamedTemporaryFile(mode="w", suffix=".ffconcat", delete=False) as tmp_concatenate_file:
         tmp_concatenate_file.write("ffconcat version 1.0\n")
         for input_path in input_video_paths:
-            tmp_concatenate_file.write(f"file '{str(input_path.resolve())}'\n")
+            tmp_concatenate_file.write(f"file '{str(input_path)}'\n")
         tmp_concatenate_file.flush()
         tmp_concatenate_path = tmp_concatenate_file.name
 
@@ -438,9 +437,7 @@ def concatenate_video_files(
         tmp_concatenate_path, mode="r", format="concat", options={"safe": "0"}
     )  # safe = 0 allows absolute paths as well as relative paths
 
-    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_named_file:
-        tmp_output_video_path = tmp_named_file.name
-
+    tmp_output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False).name
     output_container = av.open(
         tmp_output_video_path, mode="w", options={"movflags": "faststart"}
     )  # faststart is to move the metadata to the beginning of the file to speed up loading
@@ -452,9 +449,11 @@ def concatenate_video_files(
             stream_map[input_stream.index] = output_container.add_stream_from_template(
                 template=input_stream, opaque=True
             )
-
-            # set the time base to the input stream time base (missing in the codec context)
-            stream_map[input_stream.index].time_base = input_stream.time_base
+            stream_map[
+                input_stream.index
+            ].time_base = (
+                input_stream.time_base
+            )  # set the time base to the input stream time base (missing in the codec context)
 
     # Demux + remux packets (no re-encode)
     for packet in input_container.demux():
@@ -586,6 +585,19 @@ def get_video_pixel_channels(pix_fmt: str) -> int:
         raise ValueError("Unknown format")
 
 
+def get_image_pixel_channels(image: Image):
+    if image.mode == "L":
+        return 1  # Grayscale
+    elif image.mode == "LA":
+        return 2  # Grayscale + Alpha
+    elif image.mode == "RGB":
+        return 3  # RGB
+    elif image.mode == "RGBA":
+        return 4  # RGBA
+    else:
+        raise ValueError("Unknown format")
+
+
 def get_video_duration_in_s(video_path: Path | str) -> float:
     """
     Get the duration of a video file in seconds using PyAV.
@@ -643,9 +655,6 @@ class VideoEncodingManager:
             )
             self.dataset._batch_save_episode_video(start_ep, end_ep)
 
-        # Finalize the dataset to properly close all writers
-        self.dataset.finalize()
-
         # Clean up episode images if recording was interrupted
         if exc_type is not None:
             interrupted_episode_index = self.dataset.num_episodes

src/lerobot/envs/__init__.py

@@ -12,4 +12,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .configs import AlohaEnv, EnvConfig, PushtEnv  # noqa: F401
+from .configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv  # noqa: F401

src/lerobot/envs/configs.py

@@ -19,9 +19,9 @@ from typing import Any
 import draccus
 
 from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 from lerobot.robots import RobotConfig
 from lerobot.teleoperators.config import TeleoperatorConfig
-from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 
 
 @dataclass
@@ -37,16 +37,6 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
     def type(self) -> str:
         return self.get_choice_name(self.__class__)
 
-    @property
-    def package_name(self) -> str:
-        """Package name to import if environment not found in gym registry"""
-        return f"gym_{self.type}"
-
-    @property
-    def gym_id(self) -> str:
-        """ID string used in gym.make() to instantiate the environment"""
-        return f"{self.package_name}/{self.task}"
-
     @property
     @abc.abstractmethod
     def gym_kwargs(self) -> dict:
@@ -60,17 +50,15 @@ class AlohaEnv(EnvConfig):
     fps: int = 50
     episode_length: int = 400
     obs_type: str = "pixels_agent_pos"
-    observation_height: int = 480
-    observation_width: int = 640
     render_mode: str = "rgb_array"
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(14,)),
         }
     )
     features_map: dict[str, str] = field(
         default_factory=lambda: {
-            ACTION: ACTION,
+            "action": ACTION,
             "agent_pos": OBS_STATE,
             "top": f"{OBS_IMAGE}.top",
             "pixels/top": f"{OBS_IMAGES}.top",
@@ -79,14 +67,10 @@ class AlohaEnv(EnvConfig):
 
     def __post_init__(self):
         if self.obs_type == "pixels":
-            self.features["top"] = PolicyFeature(
-                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
-            )
+            self.features["top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
         elif self.obs_type == "pixels_agent_pos":
             self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(14,))
-            self.features["pixels/top"] = PolicyFeature(
-                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
-            )
+            self.features["pixels/top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
 
     @property
     def gym_kwargs(self) -> dict:
@@ -107,17 +91,15 @@ class PushtEnv(EnvConfig):
     render_mode: str = "rgb_array"
     visualization_width: int = 384
     visualization_height: int = 384
-    observation_height: int = 384
-    observation_width: int = 384
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(2,)),
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(2,)),
             "agent_pos": PolicyFeature(type=FeatureType.STATE, shape=(2,)),
         }
     )
     features_map: dict[str, str] = field(
         default_factory=lambda: {
-            ACTION: ACTION,
+            "action": ACTION,
             "agent_pos": OBS_STATE,
             "environment_state": OBS_ENV_STATE,
             "pixels": OBS_IMAGE,
@@ -126,9 +108,7 @@ class PushtEnv(EnvConfig):
 
     def __post_init__(self):
         if self.obs_type == "pixels_agent_pos":
-            self.features["pixels"] = PolicyFeature(
-                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
-            )
+            self.features["pixels"] = PolicyFeature(type=FeatureType.VISUAL, shape=(384, 384, 3))
         elif self.obs_type == "environment_state_agent_pos":
             self.features["environment_state"] = PolicyFeature(type=FeatureType.ENV, shape=(16,))
 
@@ -143,6 +123,45 @@ class PushtEnv(EnvConfig):
         }
 
 
+@EnvConfig.register_subclass("xarm")
+@dataclass
+class XarmEnv(EnvConfig):
+    task: str | None = "XarmLift-v0"
+    fps: int = 15
+    episode_length: int = 200
+    obs_type: str = "pixels_agent_pos"
+    render_mode: str = "rgb_array"
+    visualization_width: int = 384
+    visualization_height: int = 384
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
+            "pixels": PolicyFeature(type=FeatureType.VISUAL, shape=(84, 84, 3)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            "action": ACTION,
+            "agent_pos": OBS_STATE,
+            "pixels": OBS_IMAGE,
+        }
+    )
+
+    def __post_init__(self):
+        if self.obs_type == "pixels_agent_pos":
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(4,))
+
+    @property
+    def gym_kwargs(self) -> dict:
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "visualization_width": self.visualization_width,
+            "visualization_height": self.visualization_height,
+            "max_episode_steps": self.episode_length,
+        }
+
+
 @dataclass
 class ImagePreprocessingConfig:
     crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None
@@ -174,6 +193,7 @@ class ObservationConfig:
 
     add_joint_velocity_to_observation: bool = False
     add_current_to_observation: bool = False
+    add_ee_pose_to_observation: bool = False
     display_cameras: bool = False
 
 
@@ -183,6 +203,7 @@ class GripperConfig:
 
     use_gripper: bool = True
     gripper_penalty: float = 0.0
+    gripper_penalty_in_reward: bool = False
 
 
 @dataclass
@@ -235,17 +256,15 @@ class LiberoEnv(EnvConfig):
     render_mode: str = "rgb_array"
     camera_name: str = "agentview_image,robot0_eye_in_hand_image"
     init_states: bool = True
-    camera_name_mapping: dict[str, str] | None = None
-    observation_height: int = 360
-    observation_width: int = 360
+    camera_name_mapping: dict[str, str] | None = (None,)
     features: dict[str, PolicyFeature] = field(
         default_factory=lambda: {
-            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
+            "action": PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
         }
     )
     features_map: dict[str, str] = field(
         default_factory=lambda: {
-            ACTION: ACTION,
+            "action": ACTION,
             "agent_pos": OBS_STATE,
             "pixels/agentview_image": f"{OBS_IMAGES}.image",
             "pixels/robot0_eye_in_hand_image": f"{OBS_IMAGES}.image2",
@@ -255,18 +274,18 @@ class LiberoEnv(EnvConfig):
     def __post_init__(self):
         if self.obs_type == "pixels":
             self.features["pixels/agentview_image"] = PolicyFeature(
-                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
+                type=FeatureType.VISUAL, shape=(360, 360, 3)
             )
             self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
-                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
+                type=FeatureType.VISUAL, shape=(360, 360, 3)
             )
         elif self.obs_type == "pixels_agent_pos":
             self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(8,))
             self.features["pixels/agentview_image"] = PolicyFeature(
-                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
+                type=FeatureType.VISUAL, shape=(360, 360, 3)
             )
             self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
-                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
+                type=FeatureType.VISUAL, shape=(360, 360, 3)
             )
         else:
             raise ValueError(f"Unsupported obs_type: {self.obs_type}")
@@ -277,45 +296,3 @@ class LiberoEnv(EnvConfig):
             "obs_type": self.obs_type,
             "render_mode": self.render_mode,
         }
-
-
-@EnvConfig.register_subclass("metaworld")
-@dataclass
-class MetaworldEnv(EnvConfig):
-    task: str = "metaworld-push-v2"  # add all tasks
-    fps: int = 80
-    episode_length: int = 400
-    obs_type: str = "pixels_agent_pos"
-    render_mode: str = "rgb_array"
-    multitask_eval: bool = True
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": ACTION,
-            "agent_pos": OBS_STATE,
-            "top": f"{OBS_IMAGE}",
-            "pixels/top": f"{OBS_IMAGE}",
-        }
-    )
-
-    def __post_init__(self):
-        if self.obs_type == "pixels":
-            self.features["top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 480, 3))
-
-        elif self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(4,))
-            self.features["pixels/top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 480, 3))
-
-        else:
-            raise ValueError(f"Unsupported obs_type: {self.obs_type}")
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-        }

src/lerobot/envs/factory.py

@@ -16,9 +16,8 @@
 import importlib
 
 import gymnasium as gym
-from gymnasium.envs.registration import registry as gym_registry
 
-from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv
+from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv, XarmEnv
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -26,6 +25,8 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
         return AlohaEnv(**kwargs)
     elif env_type == "pusht":
         return PushtEnv(**kwargs)
+    elif env_type == "xarm":
+        return XarmEnv(**kwargs)
     elif env_type == "libero":
         return LiberoEnv(**kwargs)
     else:
@@ -62,9 +63,6 @@ def make_env(
     if "libero" in cfg.type:
         from lerobot.envs.libero import create_libero_envs
 
-        if cfg.task is None:
-            raise ValueError("LiberoEnv requires a task to be specified")
-
         return create_libero_envs(
             task=cfg.task,
             n_envs=n_envs,
@@ -73,38 +71,20 @@ def make_env(
             gym_kwargs=cfg.gym_kwargs,
             env_cls=env_cls,
         )
-    elif "metaworld" in cfg.type:
-        from lerobot.envs.metaworld import create_metaworld_envs
 
-        if cfg.task is None:
-            raise ValueError("MetaWorld requires a task to be specified")
+    package_name = f"gym_{cfg.type}"
+    try:
+        importlib.import_module(package_name)
+    except ModuleNotFoundError as e:
+        print(f"{package_name} is not installed. Please install it with `pip install 'lerobot[{cfg.type}]'`")
+        raise e
 
-        return create_metaworld_envs(
-            task=cfg.task,
-            n_envs=n_envs,
-            gym_kwargs=cfg.gym_kwargs,
-            env_cls=env_cls,
-        )
-
-    if cfg.gym_id not in gym_registry:
-        print(f"gym id '{cfg.gym_id}' not found, attempting to import '{cfg.package_name}'...")
-        try:
-            importlib.import_module(cfg.package_name)
-        except ModuleNotFoundError as e:
-            raise ModuleNotFoundError(
-                f"Package '{cfg.package_name}' required for env '{cfg.type}' not found. "
-                f"Please install it or check PYTHONPATH."
-            ) from e
-
-        if cfg.gym_id not in gym_registry:
-            raise gym.error.NameNotFound(
-                f"Environment '{cfg.gym_id}' not registered even after importing '{cfg.package_name}'."
-            )
+    gym_handle = f"{package_name}/{cfg.task}"
 
     def _make_one():
-        return gym.make(cfg.gym_id, disable_env_checker=cfg.disable_env_checker, **(cfg.gym_kwargs or {}))
+        return gym.make(gym_handle, disable_env_checker=cfg.disable_env_checker, **(cfg.gym_kwargs or {}))
 
-    vec = env_cls([_make_one for _ in range(n_envs)], autoreset_mode=gym.vector.AutoresetMode.SAME_STEP)
+    vec = env_cls([_make_one for _ in range(n_envs)])
 
     # normalize to {suite: {task_id: vec_env}} for consistency
     suite_name = cfg.type  # e.g., "pusht", "aloha"

src/lerobot/envs/libero.py

@@ -35,7 +35,7 @@ def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
     """Normalize camera_name into a non-empty list of strings."""
     if isinstance(camera_name, str):
         cams = [c.strip() for c in camera_name.split(",") if c.strip()]
-    elif isinstance(camera_name, (list | tuple)):
+    elif isinstance(camera_name, (list, tuple)):
         cams = [str(c).strip() for c in camera_name if str(c).strip()]
     else:
         raise TypeError(f"camera_name must be str or sequence[str], got {type(camera_name).__name__}")
@@ -260,23 +260,19 @@ class LiberoEnv(gym.Env):
 
         is_success = self._env.check_success()
         terminated = done or is_success
-        info.update(
-            {
-                "task": self.task,
-                "task_id": self.task_id,
-                "done": done,
-                "is_success": is_success,
-            }
-        )
+        info["is_success"] = is_success
+
         observation = self._format_raw_obs(raw_obs)
-        if terminated:
-            info["final_info"] = {
-                "task": self.task,
-                "task_id": self.task_id,
-                "done": bool(done),
-                "is_success": bool(is_success),
-            }
+        if done:
             self.reset()
+            info.update(
+                {
+                    "task": self.task,
+                    "task_id": self.task_id,
+                    "done": done,
+                    "is_success": is_success,
+                }
+            )
         truncated = False
         return observation, reward, terminated, truncated, info
 

src/lerobot/envs/metaworld.py

@@ -1,313 +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.
-import json
-from collections import defaultdict
-from collections.abc import Callable, Sequence
-from pathlib import Path
-from typing import Any
-
-import gymnasium as gym
-import metaworld
-import metaworld.policies as policies
-import numpy as np
-from gymnasium import spaces
-
-# ---- Load configuration data from the external JSON file ----
-CONFIG_PATH = Path(__file__).parent / "metaworld_config.json"
-try:
-    with open(CONFIG_PATH) as f:
-        data = json.load(f)
-except FileNotFoundError as err:
-    raise FileNotFoundError(
-        "Could not find 'metaworld_config.json'. "
-        "Please ensure the configuration file is in the same directory as the script."
-    ) from err
-except json.JSONDecodeError as err:
-    raise ValueError(
-        "Failed to decode 'metaworld_config.json'. Please ensure it is a valid JSON file."
-    ) from err
-
-# ---- Process the loaded data ----
-
-# extract and type-check top-level dicts
-task_descriptions_obj = data.get("TASK_DESCRIPTIONS")
-if not isinstance(task_descriptions_obj, dict):
-    raise TypeError("Expected TASK_DESCRIPTIONS to be a dict[str, str]")
-TASK_DESCRIPTIONS: dict[str, str] = task_descriptions_obj
-
-task_name_to_id_obj = data.get("TASK_NAME_TO_ID")
-if not isinstance(task_name_to_id_obj, dict):
-    raise TypeError("Expected TASK_NAME_TO_ID to be a dict[str, int]")
-TASK_NAME_TO_ID: dict[str, int] = task_name_to_id_obj
-
-# difficulty -> tasks mapping
-difficulty_to_tasks = data.get("DIFFICULTY_TO_TASKS")
-if not isinstance(difficulty_to_tasks, dict):
-    raise TypeError("Expected 'DIFFICULTY_TO_TASKS' to be a dict[str, list[str]]")
-DIFFICULTY_TO_TASKS: dict[str, list[str]] = difficulty_to_tasks
-
-# convert policy strings -> actual policy classes
-task_policy_mapping = data.get("TASK_POLICY_MAPPING")
-if not isinstance(task_policy_mapping, dict):
-    raise TypeError("Expected 'TASK_POLICY_MAPPING' to be a dict[str, str]")
-TASK_POLICY_MAPPING: dict[str, Any] = {
-    task_name: getattr(policies, policy_class_name)
-    for task_name, policy_class_name in task_policy_mapping.items()
-}
-ACTION_DIM = 4
-OBS_DIM = 4
-
-
-class MetaworldEnv(gym.Env):
-    metadata = {"render_modes": ["rgb_array"], "render_fps": 80}
-
-    def __init__(
-        self,
-        task,
-        camera_name="corner2",
-        obs_type="pixels",
-        render_mode="rgb_array",
-        observation_width=480,
-        observation_height=480,
-        visualization_width=640,
-        visualization_height=480,
-    ):
-        super().__init__()
-        self.task = task.replace("metaworld-", "")
-        self.obs_type = obs_type
-        self.render_mode = render_mode
-        self.observation_width = observation_width
-        self.observation_height = observation_height
-        self.visualization_width = visualization_width
-        self.visualization_height = visualization_height
-        self.camera_name = camera_name
-
-        self._env = self._make_envs_task(self.task)
-        self._max_episode_steps = self._env.max_path_length
-        self.task_description = TASK_DESCRIPTIONS[self.task]
-
-        self.expert_policy = TASK_POLICY_MAPPING[self.task]()
-
-        if self.obs_type == "state":
-            raise NotImplementedError()
-        elif self.obs_type == "pixels":
-            self.observation_space = spaces.Dict(
-                {
-                    "pixels": spaces.Box(
-                        low=0,
-                        high=255,
-                        shape=(self.observation_height, self.observation_width, 3),
-                        dtype=np.uint8,
-                    )
-                }
-            )
-        elif self.obs_type == "pixels_agent_pos":
-            self.observation_space = spaces.Dict(
-                {
-                    "pixels": spaces.Box(
-                        low=0,
-                        high=255,
-                        shape=(self.observation_height, self.observation_width, 3),
-                        dtype=np.uint8,
-                    ),
-                    "agent_pos": spaces.Box(
-                        low=-1000.0,
-                        high=1000.0,
-                        shape=(OBS_DIM,),
-                        dtype=np.float64,
-                    ),
-                }
-            )
-
-        self.action_space = spaces.Box(low=-1, high=1, shape=(ACTION_DIM,), dtype=np.float32)
-
-    def render(self) -> np.ndarray:
-        """
-        Render the current environment frame.
-
-        Returns:
-            np.ndarray: The rendered RGB image from the environment.
-        """
-        image = self._env.render()
-        if self.camera_name == "corner2":
-            # Images from this camera are flipped — correct them
-            image = np.flip(image, (0, 1))
-        return image
-
-    def _make_envs_task(self, env_name: str):
-        mt1 = metaworld.MT1(env_name, seed=42)
-        env = mt1.train_classes[env_name](render_mode="rgb_array", camera_name=self.camera_name)
-        env.set_task(mt1.train_tasks[0])
-        if self.camera_name == "corner2":
-            env.model.cam_pos[2] = [
-                0.75,
-                0.075,
-                0.7,
-            ]  # corner2 position, similar to https://arxiv.org/pdf/2206.14244
-        env.reset()
-        env._freeze_rand_vec = False  # otherwise no randomization
-        return env
-
-    def _format_raw_obs(self, raw_obs: np.ndarray) -> dict[str, Any]:
-        image = None
-        if self._env is not None:
-            image = self._env.render()
-            if self.camera_name == "corner2":
-                # NOTE: The "corner2" camera in MetaWorld environments outputs images with both axes inverted.
-                image = np.flip(image, (0, 1))
-        agent_pos = raw_obs[:4]
-        if self.obs_type == "state":
-            raise NotImplementedError(
-                "'state' obs_type not implemented for MetaWorld. Use pixel modes instead."
-            )
-
-        elif self.obs_type in ("pixels", "pixels_agent_pos"):
-            assert image is not None, (
-                "Expected `image` to be rendered before constructing pixel-based observations. "
-                "This likely means `env.render()` returned None or the environment was not provided."
-            )
-
-            if self.obs_type == "pixels":
-                obs = {"pixels": image.copy()}
-
-            else:  # pixels_agent_pos
-                obs = {
-                    "pixels": image.copy(),
-                    "agent_pos": agent_pos,
-                }
-        else:
-            raise ValueError(f"Unknown obs_type: {self.obs_type}")
-        return obs
-
-    def reset(
-        self,
-        seed: int | None = None,
-        **kwargs,
-    ) -> tuple[dict[str, Any], dict[str, Any]]:
-        """
-        Reset the environment to its initial state.
-
-        Args:
-            seed (Optional[int]): Random seed for environment initialization.
-
-        Returns:
-            observation (Dict[str, Any]): The initial formatted observation.
-            info (Dict[str, Any]): Additional info about the reset state.
-        """
-        super().reset(seed=seed)
-
-        raw_obs, info = self._env.reset(seed=seed)
-
-        observation = self._format_raw_obs(raw_obs)
-
-        info = {"is_success": False}
-        return observation, info
-
-    def step(self, action: np.ndarray) -> tuple[dict[str, Any], float, bool, bool, dict[str, Any]]:
-        """
-        Perform one environment step.
-
-        Args:
-            action (np.ndarray): The action to execute, must be 1-D with shape (action_dim,).
-
-        Returns:
-            observation (Dict[str, Any]): The formatted observation after the step.
-            reward (float): The scalar reward for this step.
-            terminated (bool): Whether the episode terminated successfully.
-            truncated (bool): Whether the episode was truncated due to a time limit.
-            info (Dict[str, Any]): Additional environment info.
-        """
-        if action.ndim != 1:
-            raise ValueError(
-                f"Expected action to be 1-D (shape (action_dim,)), "
-                f"but got shape {action.shape} with ndim={action.ndim}"
-            )
-        raw_obs, reward, done, truncated, info = self._env.step(action)
-
-        # Determine whether the task was successful
-        is_success = bool(info.get("success", 0))
-        terminated = done or is_success
-        info.update(
-            {
-                "task": self.task,
-                "done": done,
-                "is_success": is_success,
-            }
-        )
-
-        # Format the raw observation into the expected structure
-        observation = self._format_raw_obs(raw_obs)
-        if terminated:
-            info["final_info"] = {
-                "task": self.task,
-                "done": bool(done),
-                "is_success": bool(is_success),
-            }
-            self.reset()
-
-        return observation, reward, terminated, truncated, info
-
-    def close(self):
-        self._env.close()
-
-
-# ---- Main API ----------------------------------------------------------------
-
-
-def create_metaworld_envs(
-    task: str,
-    n_envs: int,
-    gym_kwargs: dict[str, Any] | None = None,
-    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
-) -> dict[str, dict[int, Any]]:
-    """
-    Create vectorized Meta-World environments with a consistent return shape.
-
-    Returns:
-        dict[task_group][task_id] -> vec_env (env_cls([...]) with exactly n_envs factories)
-    Notes:
-        - n_envs is the number of rollouts *per task* (episode_index = 0..n_envs-1).
-        - `task` can be a single difficulty group (e.g., "easy", "medium", "hard") or a comma-separated list.
-        - If a task name is not in DIFFICULTY_TO_TASKS, we treat it as a single custom task.
-    """
-    if env_cls is None or not callable(env_cls):
-        raise ValueError("env_cls must be a callable that wraps a list of environment factory callables.")
-    if not isinstance(n_envs, int) or n_envs <= 0:
-        raise ValueError(f"n_envs must be a positive int; got {n_envs}.")
-
-    gym_kwargs = dict(gym_kwargs or {})
-    task_groups = [t.strip() for t in task.split(",") if t.strip()]
-    if not task_groups:
-        raise ValueError("`task` must contain at least one Meta-World task or difficulty group.")
-
-    print(f"Creating Meta-World envs | task_groups={task_groups} | n_envs(per task)={n_envs}")
-
-    out: dict[str, dict[int, Any]] = defaultdict(dict)
-
-    for group in task_groups:
-        # if not in difficulty presets, treat it as a single custom task
-        tasks = DIFFICULTY_TO_TASKS.get(group, [group])
-
-        for tid, task_name in enumerate(tasks):
-            print(f"Building vec env | group={group} | task_id={tid} | task={task_name}")
-
-            # build n_envs factories
-            fns = [(lambda tn=task_name: MetaworldEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]
-
-            out[group][tid] = env_cls(fns)
-
-    # return a plain dict for consistency
-    return {group: dict(task_map) for group, task_map in out.items()}

src/lerobot/envs/metaworld_config.json

@@ -1,121 +0,0 @@
-{
-  "TASK_DESCRIPTIONS": {
-    "assembly-v3": "Pick up a nut and place it onto a peg",
-    "basketball-v3": "Dunk the basketball into the basket",
-    "bin-picking-v3": "Grasp the puck from one bin and place it into another bin",
-    "box-close-v3": "Grasp the cover and close the box with it",
-    "button-press-topdown-v3": "Press a button from the top",
-    "button-press-topdown-wall-v3": "Bypass a wall and press a button from the top",
-    "button-press-v3": "Press a button",
-    "button-press-wall-v3": "Bypass a wall and press a button",
-    "coffee-button-v3": "Push a button on the coffee machine",
-    "coffee-pull-v3": "Pull a mug from a coffee machine",
-    "coffee-push-v3": "Push a mug under a coffee machine",
-    "dial-turn-v3": "Rotate a dial 180 degrees",
-    "disassemble-v3": "Pick a nut out of a peg",
-    "door-close-v3": "Close a door with a revolving joint",
-    "door-lock-v3": "Lock the door by rotating the lock clockwise",
-    "door-open-v3": "Open a door with a revolving joint",
-    "door-unlock-v3": "Unlock the door by rotating the lock counter-clockwise",
-    "hand-insert-v3": "Insert the gripper into a hole",
-    "drawer-close-v3": "Push and close a drawer",
-    "drawer-open-v3": "Open a drawer",
-    "faucet-open-v3": "Rotate the faucet counter-clockwise",
-    "faucet-close-v3": "Rotate the faucet clockwise",
-    "hammer-v3": "Hammer a screw on the wall",
-    "handle-press-side-v3": "Press a handle down sideways",
-    "handle-press-v3": "Press a handle down",
-    "handle-pull-side-v3": "Pull a handle up sideways",
-    "handle-pull-v3": "Pull a handle up",
-    "lever-pull-v3": "Pull a lever down 90 degrees",
-    "peg-insert-side-v3": "Insert a peg sideways",
-    "pick-place-wall-v3": "Pick a puck, bypass a wall and place the puck",
-    "pick-out-of-hole-v3": "Pick up a puck from a hole",
-    "reach-v3": "Reach a goal position",
-    "push-back-v3": "Push the puck to a goal",
-    "push-v3": "Push the puck to a goal",
-    "pick-place-v3": "Pick and place a puck to a goal",
-    "plate-slide-v3": "Slide a plate into a cabinet",
-    "plate-slide-side-v3": "Slide a plate into a cabinet sideways",
-    "plate-slide-back-v3": "Get a plate from the cabinet",
-    "plate-slide-back-side-v3": "Get a plate from the cabinet sideways",
-    "peg-unplug-side-v3": "Unplug a peg sideways",
-    "soccer-v3": "Kick a soccer into the goal",
-    "stick-push-v3": "Grasp a stick and push a box using the stick",
-    "stick-pull-v3": "Grasp a stick and pull a box with the stick",
-    "push-wall-v3": "Bypass a wall and push a puck to a goal",
-    "reach-wall-v3": "Bypass a wall and reach a goal",
-    "shelf-place-v3": "Pick and place a puck onto a shelf",
-    "sweep-into-v3": "Sweep a puck into a hole",
-    "sweep-v3": "Sweep a puck off the table",
-    "window-open-v3": "Push and open a window",
-    "window-close-v3": "Push and close a window"
-  },
-  "TASK_NAME_TO_ID": {
-    "assembly-v3": 0, "basketball-v3": 1, "bin-picking-v3": 2, "box-close-v3": 3,
-    "button-press-topdown-v3": 4, "button-press-topdown-wall-v3": 5, "button-press-v3": 6,
-    "button-press-wall-v3": 7, "coffee-button-v3": 8, "coffee-pull-v3": 9, "coffee-push-v3": 10,
-    "dial-turn-v3": 11, "disassemble-v3": 12, "door-close-v3": 13, "door-lock-v3": 14,
-    "door-open-v3": 15, "door-unlock-v3": 16, "drawer-close-v3": 17, "drawer-open-v3": 18,
-    "faucet-close-v3": 19, "faucet-open-v3": 20, "hammer-v3": 21, "hand-insert-v3": 22,
-    "handle-press-side-v3": 23, "handle-press-v3": 24, "handle-pull-side-v3": 25,
-    "handle-pull-v3": 26, "lever-pull-v3": 27, "peg-insert-side-v3": 28, "peg-unplug-side-v3": 29,
-    "pick-out-of-hole-v3": 30, "pick-place-v3": 31, "pick-place-wall-v3": 32,
-    "plate-slide-back-side-v3": 33, "plate-slide-back-v3": 34, "plate-slide-side-v3": 35,
-    "plate-slide-v3": 36, "push-back-v3": 37, "push-v3": 38, "push-wall-v3": 39, "reach-v3": 40,
-    "reach-wall-v3": 41, "shelf-place-v3": 42, "soccer-v3": 43, "stick-pull-v3": 44,
-    "stick-push-v3": 45, "sweep-into-v3": 46, "sweep-v3": 47, "window-open-v3": 48,
-    "window-close-v3": 49
-  },
-  "DIFFICULTY_TO_TASKS": {
-    "easy": [
-      "button-press-v3", "button-press-topdown-v3", "button-press-topdown-wall-v3",
-      "button-press-wall-v3", "coffee-button-v3", "dial-turn-v3", "door-close-v3",
-      "door-lock-v3", "door-open-v3", "door-unlock-v3", "drawer-close-v3", "drawer-open-v3",
-      "faucet-close-v3", "faucet-open-v3", "handle-press-v3", "handle-press-side-v3",
-      "handle-pull-v3", "handle-pull-side-v3", "lever-pull-v3", "plate-slide-v3",
-      "plate-slide-back-v3", "plate-slide-back-side-v3", "plate-slide-side-v3", "reach-v3",
-      "reach-wall-v3", "window-close-v3", "window-open-v3", "peg-unplug-side-v3"
-    ],
-    "medium": [
-      "basketball-v3", "bin-picking-v3", "box-close-v3", "coffee-pull-v3", "coffee-push-v3",
-      "hammer-v3", "peg-insert-side-v3", "push-wall-v3", "soccer-v3", "sweep-v3", "sweep-into-v3"
-    ],
-    "hard": [
-      "assembly-v3", "hand-insert-v3", "pick-out-of-hole-v3", "pick-place-v3", "push-v3", "push-back-v3"
-    ],
-    "very_hard": [
-      "shelf-place-v3", "disassemble-v3", "stick-pull-v3", "stick-push-v3", "pick-place-wall-v3"
-    ]
-  },
-  "TASK_POLICY_MAPPING": {
-    "assembly-v3": "SawyerAssemblyV3Policy", "basketball-v3": "SawyerBasketballV3Policy",
-    "bin-picking-v3": "SawyerBinPickingV3Policy", "box-close-v3": "SawyerBoxCloseV3Policy",
-    "button-press-topdown-v3": "SawyerButtonPressTopdownV3Policy",
-    "button-press-topdown-wall-v3": "SawyerButtonPressTopdownWallV3Policy",
-    "button-press-v3": "SawyerButtonPressV3Policy", "button-press-wall-v3": "SawyerButtonPressWallV3Policy",
-    "coffee-button-v3": "SawyerCoffeeButtonV3Policy", "coffee-pull-v3": "SawyerCoffeePullV3Policy",
-    "coffee-push-v3": "SawyerCoffeePushV3Policy", "dial-turn-v3": "SawyerDialTurnV3Policy",
-    "disassemble-v3": "SawyerDisassembleV3Policy", "door-close-v3": "SawyerDoorCloseV3Policy",
-    "door-lock-v3": "SawyerDoorLockV3Policy", "door-open-v3": "SawyerDoorOpenV3Policy",
-    "door-unlock-v3": "SawyerDoorUnlockV3Policy", "drawer-close-v3": "SawyerDrawerCloseV3Policy",
-    "drawer-open-v3": "SawyerDrawerOpenV3Policy", "faucet-close-v3": "SawyerFaucetCloseV3Policy",
-    "faucet-open-v3": "SawyerFaucetOpenV3Policy", "hammer-v3": "SawyerHammerV3Policy",
-    "hand-insert-v3": "SawyerHandInsertV3Policy", "handle-press-side-v3": "SawyerHandlePressSideV3Policy",
-    "handle-press-v3": "SawyerHandlePressV3Policy", "handle-pull-side-v3": "SawyerHandlePullSideV3Policy",
-    "handle-pull-v3": "SawyerHandlePullV3Policy", "lever-pull-v3": "SawyerLeverPullV3Policy",
-    "peg-insert-side-v3": "SawyerPegInsertionSideV3Policy", "peg-unplug-side-v3": "SawyerPegUnplugSideV3Policy",
-    "pick-out-of-hole-v3": "SawyerPickOutOfHoleV3Policy", "pick-place-v3": "SawyerPickPlaceV3Policy",
-    "pick-place-wall-v3": "SawyerPickPlaceWallV3Policy",
-    "plate-slide-back-side-v3": "SawyerPlateSlideBackSideV3Policy",
-    "plate-slide-back-v3": "SawyerPlateSlideBackV3Policy",
-    "plate-slide-side-v3": "SawyerPlateSlideSideV3Policy", "plate-slide-v3": "SawyerPlateSlideV3Policy",
-    "push-back-v3": "SawyerPushBackV3Policy", "push-v3": "SawyerPushV3Policy",
-    "push-wall-v3": "SawyerPushWallV3Policy", "reach-v3": "SawyerReachV3Policy",
-    "reach-wall-v3": "SawyerReachWallV3Policy", "shelf-place-v3": "SawyerShelfPlaceV3Policy",
-    "soccer-v3": "SawyerSoccerV3Policy", "stick-pull-v3": "SawyerStickPullV3Policy",
-    "stick-push-v3": "SawyerStickPushV3Policy", "sweep-into-v3": "SawyerSweepIntoV3Policy",
-    "sweep-v3": "SawyerSweepV3Policy", "window-open-v3": "SawyerWindowOpenV3Policy",
-    "window-close-v3": "SawyerWindowCloseV3Policy"
-  }
-}

src/lerobot/envs/utils.py

@@ -26,7 +26,6 @@ from torch import Tensor
 
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.envs.configs import EnvConfig
-from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 from lerobot.utils.utils import get_channel_first_image_shape
 
 
@@ -42,44 +41,44 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
     return_observations = {}
     if "pixels" in observations:
         if isinstance(observations["pixels"], dict):
-            imgs = {f"{OBS_IMAGES}.{key}": img for key, img in observations["pixels"].items()}
+            imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
         else:
-            imgs = {OBS_IMAGE: observations["pixels"]}
+            imgs = {"observation.image": observations["pixels"]}
 
         for imgkey, img in imgs.items():
             # TODO(aliberts, rcadene): use transforms.ToTensor()?
-            img_tensor = torch.from_numpy(img)
+            img = torch.from_numpy(img)
 
             # When preprocessing observations in a non-vectorized environment, we need to add a batch dimension.
             # This is the case for human-in-the-loop RL where there is only one environment.
-            if img_tensor.ndim == 3:
-                img_tensor = img_tensor.unsqueeze(0)
+            if img.ndim == 3:
+                img = img.unsqueeze(0)
             # sanity check that images are channel last
-            _, h, w, c = img_tensor.shape
-            assert c < h and c < w, f"expect channel last images, but instead got {img_tensor.shape=}"
+            _, h, w, c = img.shape
+            assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
 
             # sanity check that images are uint8
-            assert img_tensor.dtype == torch.uint8, f"expect torch.uint8, but instead {img_tensor.dtype=}"
+            assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
 
             # convert to channel first of type float32 in range [0,1]
-            img_tensor = einops.rearrange(img_tensor, "b h w c -> b c h w").contiguous()
-            img_tensor = img_tensor.type(torch.float32)
-            img_tensor /= 255
+            img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
+            img = img.type(torch.float32)
+            img /= 255
 
-            return_observations[imgkey] = img_tensor
+            return_observations[imgkey] = img
 
     if "environment_state" in observations:
         env_state = torch.from_numpy(observations["environment_state"]).float()
         if env_state.dim() == 1:
             env_state = env_state.unsqueeze(0)
 
-        return_observations[OBS_ENV_STATE] = env_state
+        return_observations["observation.environment_state"] = env_state
 
     # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
     agent_pos = torch.from_numpy(observations["agent_pos"]).float()
     if agent_pos.dim() == 1:
         agent_pos = agent_pos.unsqueeze(0)
-    return_observations[OBS_STATE] = agent_pos
+    return_observations["observation.state"] = agent_pos
 
     return return_observations
 
@@ -183,10 +182,10 @@ def _(env: Mapping) -> None:
 
 @close_envs.register
 def _(envs: Sequence) -> None:
-    if isinstance(envs, (str | bytes)):
+    if isinstance(envs, (str, bytes)):
         return
     for v in envs:
-        if isinstance(v, Mapping) or isinstance(v, Sequence) and not isinstance(v, (str | bytes)):
+        if isinstance(v, Mapping) or isinstance(v, Sequence) and not isinstance(v, (str, bytes)):
             close_envs(v)
         elif hasattr(v, "close"):
             _close_single_env(v)

src/lerobot/errors.py

@@ -30,3 +30,14 @@ class DeviceAlreadyConnectedError(ConnectionError):
     ):
         self.message = message
         super().__init__(self.message)
+
+
+class InvalidActionError(ValueError):
+    """Exception raised when an action is already invalid."""
+
+    def __init__(
+        self,
+        message="The action is invalid. Check the value follows what it is expected from the action space.",
+    ):
+        self.message = message
+        super().__init__(self.message)

src/lerobot/find_cameras.py

@@ -24,7 +24,7 @@ lerobot-find-cameras
 ```
 """
 
-# NOTE(Steven): RealSense can also be identified/opened as OpenCV cameras. If you know the camera is a RealSense, use the `lerobot-find-cameras realsense` flag to avoid confusion.
+# NOTE(Steven): RealSense can also be identified/opened as OpenCV cameras. If you know the camera is a RealSense, use the `lerobot.find_cameras realsense` flag to avoid confusion.
 # NOTE(Steven): macOS cameras sometimes report different FPS at init time, not an issue here as we don't specify FPS when opening the cameras, but the information displayed might not be truthful.
 
 import argparse
@@ -180,7 +180,7 @@ def create_camera_instance(cam_meta: dict[str, Any]) -> dict[str, Any] | None:
 
         if instance:
             logger.info(f"Connecting to {cam_type} camera: {cam_id}...")
-            instance.connect(warmup=True)
+            instance.connect(warmup=False)
             return {"instance": instance, "meta": cam_meta}
     except Exception as e:
         logger.error(f"Failed to connect or configure {cam_type} camera {cam_id}: {e}")

src/lerobot/model/kinematics.py

@@ -22,18 +22,18 @@ class RobotKinematics:
         self,
         urdf_path: str,
         target_frame_name: str = "gripper_frame_link",
-        joint_names: list[str] | None = None,
+        joint_names: list[str] = None,
     ):
         """
         Initialize placo-based kinematics solver.
 
         Args:
-            urdf_path (str): Path to the robot URDF file
-            target_frame_name (str): Name of the end-effector frame in the URDF
-            joint_names (list[str] | None): List of joint names to use for the kinematics solver
+            urdf_path: Path to the robot URDF file
+            target_frame_name: Name of the end-effector frame in the URDF
+            joint_names: List of joint names to use for the kinematics solver
         """
         try:
-            import placo  # type: ignore[import-not-found] # C++ library with Python bindings, no type stubs available. TODO: Create stub file or request upstream typing support.
+            import placo
         except ImportError as e:
             raise ImportError(
                 "placo is required for RobotKinematics. "
@@ -52,7 +52,7 @@ class RobotKinematics:
         # Initialize frame task for IK
         self.tip_frame = self.solver.add_frame_task(self.target_frame_name, np.eye(4))
 
-    def forward_kinematics(self, joint_pos_deg: np.ndarray) -> np.ndarray:
+    def forward_kinematics(self, joint_pos_deg):
         """
         Compute forward kinematics for given joint configuration given the target frame name in the constructor.
 
@@ -77,12 +77,8 @@ class RobotKinematics:
         return self.robot.get_T_world_frame(self.target_frame_name)
 
     def inverse_kinematics(
-        self,
-        current_joint_pos: np.ndarray,
-        desired_ee_pose: np.ndarray,
-        position_weight: float = 1.0,
-        orientation_weight: float = 0.01,
-    ) -> np.ndarray:
+        self, current_joint_pos, desired_ee_pose, position_weight=1.0, orientation_weight=0.01
+    ):
         """
         Compute inverse kinematics using placo solver.
 

src/lerobot/motors/__init__.py

@@ -1,17 +1 @@
-#!/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.
-
 from .motors_bus import Motor, MotorCalibration, MotorNormMode, MotorsBus

src/lerobot/motors/dynamixel/dynamixel.py

@@ -22,7 +22,7 @@ import logging
 from copy import deepcopy
 from enum import Enum
 
-from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement
+from lerobot.utils.encoding_utils import decode_twos_complement, encode_twos_complement
 
 from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
 from .tables import (
@@ -60,7 +60,7 @@ class OperatingMode(Enum):
 
     # This mode controls position. This mode is identical to the Multi-turn Position Control from existing
     # DYNAMIXEL. 512 turns are supported(-256[rev] ~ 256[rev]). This mode is ideal for multi-turn wrists or
-    # conveyor systems or a system that requires an additional reduction gear. Note that Max Position
+    # conveyer systems or a system that requires an additional reduction gear. Note that Max Position
     # Limit(48), Min Position Limit(52) are not used on Extended Position Control Mode.
     EXTENDED_POSITION = 4
 

src/lerobot/motors/feetech/feetech.py

@@ -17,7 +17,7 @@ from copy import deepcopy
 from enum import Enum
 from pprint import pformat
 
-from lerobot.motors.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
+from lerobot.utils.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
 
 from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
 from .tables import (

src/lerobot/motors/feetech/tables.py

@@ -206,12 +206,8 @@ MODEL_BAUDRATE_TABLE = {
 # Sign-Magnitude encoding bits
 STS_SMS_SERIES_ENCODINGS_TABLE = {
     "Homing_Offset": 11,
-    "Goal_Position": 15,
     "Goal_Velocity": 15,
-    "Goal_Speed": 15,
-    "Present_Position": 15,
     "Present_Velocity": 15,
-    "Present_Speed": 15,
 }
 
 MODEL_ENCODING_TABLE = {

src/lerobot/motors/motors_bus.py

@@ -32,7 +32,7 @@ import serial
 from deepdiff import DeepDiff
 from tqdm import tqdm
 
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 from lerobot.utils.utils import enter_pressed, move_cursor_up
 
 NameOrID: TypeAlias = str | int
@@ -99,6 +99,12 @@ class Motor:
     norm_mode: MotorNormMode
 
 
+class JointOutOfRangeError(Exception):
+    def __init__(self, message="Joint is out of range"):
+        self.message = message
+        super().__init__(self.message)
+
+
 class PortHandler(Protocol):
     def __init__(self, port_name):
         self.is_open: bool
@@ -342,7 +348,7 @@ class MotorsBus(abc.ABC):
             raise TypeError(motors)
 
     def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> list[str]:
-        if isinstance(values, (int | float)):
+        if isinstance(values, (int, float)):
             return dict.fromkeys(self.ids, values)
         elif isinstance(values, dict):
             return {self.motors[motor].id: val for motor, val in values.items()}
@@ -669,7 +675,7 @@ class MotorsBus(abc.ABC):
         """
         if motors is None:
             motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
+        elif isinstance(motors, (str, int)):
             motors = [motors]
         elif not isinstance(motors, list):
             raise TypeError(motors)
@@ -697,7 +703,7 @@ class MotorsBus(abc.ABC):
         """
         if motors is None:
             motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
+        elif isinstance(motors, (str, int)):
             motors = [motors]
         elif not isinstance(motors, list):
             raise TypeError(motors)
@@ -733,7 +739,7 @@ class MotorsBus(abc.ABC):
         """
         if motors is None:
             motors = list(self.motors)
-        elif isinstance(motors, (str | int)):
+        elif isinstance(motors, (str, int)):
             motors = [motors]
         elif not isinstance(motors, list):
             raise TypeError(motors)

src/lerobot/optim/optimizers.py

@@ -22,11 +22,11 @@ import draccus
 import torch
 from safetensors.torch import load_file, save_file
 
-from lerobot.datasets.utils import flatten_dict, unflatten_dict, write_json
-from lerobot.utils.constants import (
+from lerobot.constants import (
     OPTIMIZER_PARAM_GROUPS,
     OPTIMIZER_STATE,
 )
+from lerobot.datasets.utils import flatten_dict, unflatten_dict, write_json
 from lerobot.utils.io_utils import deserialize_json_into_object
 
 

src/lerobot/optim/schedulers.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import abc
-import logging
 import math
 from dataclasses import asdict, dataclass
 from pathlib import Path
@@ -23,8 +22,8 @@ import draccus
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LambdaLR, LRScheduler
 
+from lerobot.constants import SCHEDULER_STATE
 from lerobot.datasets.utils import write_json
-from lerobot.utils.constants import SCHEDULER_STATE
 from lerobot.utils.io_utils import deserialize_json_into_object
 
 
@@ -80,11 +79,7 @@ class VQBeTSchedulerConfig(LRSchedulerConfig):
 @LRSchedulerConfig.register_subclass("cosine_decay_with_warmup")
 @dataclass
 class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
-    """Used by Physical Intelligence to train Pi0.
-
-    Automatically scales warmup and decay steps if num_training_steps < num_decay_steps.
-    This ensures the learning rate schedule completes properly even with shorter training runs.
-    """
+    """Used by Physical Intelligence to train Pi0"""
 
     num_warmup_steps: int
     num_decay_steps: int
@@ -92,39 +87,23 @@ class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
     decay_lr: float
 
     def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
-        # Auto-scale scheduler parameters if training steps are shorter than configured decay steps
-        actual_warmup_steps = self.num_warmup_steps
-        actual_decay_steps = self.num_decay_steps
-
-        if num_training_steps < self.num_decay_steps:
-            # Calculate scaling factor to fit the schedule into the available training steps
-            scale_factor = num_training_steps / self.num_decay_steps
-            actual_warmup_steps = int(self.num_warmup_steps * scale_factor)
-            actual_decay_steps = num_training_steps
-
-            logging.info(
-                f"Auto-scaling LR scheduler: "
-                f"num_training_steps ({num_training_steps}) < num_decay_steps ({self.num_decay_steps}). "
-                f"Scaling warmup: {self.num_warmup_steps} → {actual_warmup_steps}, "
-                f"decay: {self.num_decay_steps} → {actual_decay_steps} "
-                f"(scale factor: {scale_factor:.3f})"
-            )
+        del num_training_steps
 
         def lr_lambda(current_step):
             def linear_warmup_schedule(current_step):
                 if current_step <= 0:
-                    return 1 / (actual_warmup_steps + 1)
-                frac = 1 - current_step / actual_warmup_steps
-                return (1 / (actual_warmup_steps + 1) - 1) * frac + 1
+                    return 1 / (self.num_warmup_steps + 1)
+                frac = 1 - current_step / self.num_warmup_steps
+                return (1 / (self.num_warmup_steps + 1) - 1) * frac + 1
 
             def cosine_decay_schedule(current_step):
-                step = min(current_step, actual_decay_steps)
-                cosine_decay = 0.5 * (1 + math.cos(math.pi * step / actual_decay_steps))
+                step = min(current_step, self.num_decay_steps)
+                cosine_decay = 0.5 * (1 + math.cos(math.pi * step / self.num_decay_steps))
                 alpha = self.decay_lr / self.peak_lr
                 decayed = (1 - alpha) * cosine_decay + alpha
                 return decayed
 
-            if current_step < actual_warmup_steps:
+            if current_step < self.num_warmup_steps:
                 return linear_warmup_schedule(current_step)
 
             return cosine_decay_schedule(current_step)

src/lerobot/policies/__init__.py

@@ -15,7 +15,7 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
-from .pi05.configuration_pi05 import PI05Config as PI05Config
+from .pi0.processor_pi0 import Pi0NewLineProcessor
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .smolvla.processor_smolvla import SmolVLANewLineProcessor
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
@@ -25,7 +25,6 @@ __all__ = [
     "ACTConfig",
     "DiffusionConfig",
     "PI0Config",
-    "PI05Config",
     "SmolVLAConfig",
     "TDMPCConfig",
     "VQBeTConfig",

src/lerobot/policies/act/modeling_act.py

@@ -33,9 +33,9 @@ from torch import Tensor, nn
 from torchvision.models._utils import IntermediateLayerGetter
 from torchvision.ops.misc import FrozenBatchNorm2d
 
+from lerobot.constants import ACTION, OBS_IMAGES
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.policies.pretrained import PreTrainedPolicy
-from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 
 
 class ACTPolicy(PreTrainedPolicy):
@@ -394,22 +394,25 @@ class ACT(nn.Module):
             latent dimension.
         """
         if self.config.use_vae and self.training:
-            assert ACTION in batch, (
+            assert "action" in batch, (
                 "actions must be provided when using the variational objective in training mode."
             )
 
-        batch_size = batch[OBS_IMAGES][0].shape[0] if OBS_IMAGES in batch else batch[OBS_ENV_STATE].shape[0]
+        if "observation.images" in batch:
+            batch_size = batch["observation.images"][0].shape[0]
+        else:
+            batch_size = batch["observation.environment_state"].shape[0]
 
         # Prepare the latent for input to the transformer encoder.
-        if self.config.use_vae and ACTION in batch and self.training:
+        if self.config.use_vae and "action" in batch and self.training:
             # Prepare the input to the VAE encoder: [cls, *joint_space_configuration, *action_sequence].
             cls_embed = einops.repeat(
                 self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
             )  # (B, 1, D)
             if self.config.robot_state_feature:
-                robot_state_embed = self.vae_encoder_robot_state_input_proj(batch[OBS_STATE])
+                robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
                 robot_state_embed = robot_state_embed.unsqueeze(1)  # (B, 1, D)
-            action_embed = self.vae_encoder_action_input_proj(batch[ACTION])  # (B, S, D)
+            action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
 
             if self.config.robot_state_feature:
                 vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
@@ -427,7 +430,7 @@ class ACT(nn.Module):
             cls_joint_is_pad = torch.full(
                 (batch_size, 2 if self.config.robot_state_feature else 1),
                 False,
-                device=batch[OBS_STATE].device,
+                device=batch["observation.state"].device,
             )
             key_padding_mask = torch.cat(
                 [cls_joint_is_pad, batch["action_is_pad"]], axis=1
@@ -451,7 +454,7 @@ class ACT(nn.Module):
             mu = log_sigma_x2 = None
             # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
             latent_sample = torch.zeros([batch_size, self.config.latent_dim], dtype=torch.float32).to(
-                batch[OBS_STATE].device
+                batch["observation.state"].device
             )
 
         # Prepare transformer encoder inputs.
@@ -459,16 +462,18 @@ class ACT(nn.Module):
         encoder_in_pos_embed = list(self.encoder_1d_feature_pos_embed.weight.unsqueeze(1))
         # Robot state token.
         if self.config.robot_state_feature:
-            encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch[OBS_STATE]))
+            encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch["observation.state"]))
         # Environment state token.
         if self.config.env_state_feature:
-            encoder_in_tokens.append(self.encoder_env_state_input_proj(batch[OBS_ENV_STATE]))
+            encoder_in_tokens.append(
+                self.encoder_env_state_input_proj(batch["observation.environment_state"])
+            )
 
         if self.config.image_features:
             # For a list of images, the H and W may vary but H*W is constant.
             # NOTE: If modifying this section, verify on MPS devices that
             # gradients remain stable (no explosions or NaNs).
-            for img in batch[OBS_IMAGES]:
+            for img in batch["observation.images"]:
                 cam_features = self.backbone(img)["feature_map"]
                 cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
                 cam_features = self.encoder_img_feat_input_proj(cam_features)
@@ -626,8 +631,8 @@ class ACTDecoderLayer(nn.Module):
             x: (Decoder Sequence, Batch, Channel) tensor of input tokens.
             encoder_out: (Encoder Sequence, B, C) output features from the last layer of the encoder we are
                 cross-attending with.
-            encoder_pos_embed: (ES, 1, C) positional embedding for keys (from the encoder).
-            decoder_pos_embed: (DS, 1, C) positional embedding for the queries (from the decoder).
+            decoder_pos_embed: (ES, 1, C) positional embedding for keys (from the encoder).
+            encoder_pos_embed: (DS, 1, C) Positional_embedding for the queries (from the decoder).
         Returns:
             (DS, B, C) tensor of decoder output features.
         """

src/lerobot/policies/act/processor_act.py

@@ -17,6 +17,7 @@ from typing import Any
 
 import torch
 
+from lerobot.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
@@ -28,7 +29,6 @@ from lerobot.processor import (
     UnnormalizerProcessorStep,
 )
 from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 
 def make_act_pre_post_processors(

src/lerobot/policies/diffusion/modeling_diffusion.py

@@ -33,6 +33,7 @@ from diffusers.schedulers.scheduling_ddim import DDIMScheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 from torch import Tensor, nn
 
+from lerobot.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.policies.utils import (
@@ -41,7 +42,6 @@ from lerobot.policies.utils import (
     get_output_shape,
     populate_queues,
 )
-from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGES, OBS_STATE
 
 
 class DiffusionPolicy(PreTrainedPolicy):
@@ -81,25 +81,25 @@ class DiffusionPolicy(PreTrainedPolicy):
     def reset(self):
         """Clear observation and action queues. Should be called on `env.reset()`"""
         self._queues = {
-            OBS_STATE: deque(maxlen=self.config.n_obs_steps),
-            ACTION: deque(maxlen=self.config.n_action_steps),
+            "observation.state": deque(maxlen=self.config.n_obs_steps),
+            "action": deque(maxlen=self.config.n_action_steps),
         }
         if self.config.image_features:
-            self._queues[OBS_IMAGES] = deque(maxlen=self.config.n_obs_steps)
+            self._queues["observation.images"] = deque(maxlen=self.config.n_obs_steps)
         if self.config.env_state_feature:
-            self._queues[OBS_ENV_STATE] = deque(maxlen=self.config.n_obs_steps)
+            self._queues["observation.environment_state"] = deque(maxlen=self.config.n_obs_steps)
 
     @torch.no_grad()
-    def predict_action_chunk(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
         """Predict a chunk of actions given environment observations."""
         # stack n latest observations from the queue
         batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
-        actions = self.diffusion.generate_actions(batch, noise=noise)
+        actions = self.diffusion.generate_actions(batch)
 
         return actions
 
     @torch.no_grad()
-    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select a single action given environment observations.
 
         This method handles caching a history of observations and an action trajectory generated by the
@@ -131,7 +131,7 @@ class DiffusionPolicy(PreTrainedPolicy):
         self._queues = populate_queues(self._queues, batch)
 
         if len(self._queues[ACTION]) == 0:
-            actions = self.predict_action_chunk(batch, noise=noise)
+            actions = self.predict_action_chunk(batch)
             self._queues[ACTION].extend(actions.transpose(0, 1))
 
         action = self._queues[ACTION].popleft()
@@ -199,25 +199,17 @@ class DiffusionModel(nn.Module):
 
     # ========= inference  ============
     def conditional_sample(
-        self,
-        batch_size: int,
-        global_cond: Tensor | None = None,
-        generator: torch.Generator | None = None,
-        noise: Tensor | None = None,
+        self, batch_size: int, global_cond: Tensor | None = None, generator: torch.Generator | None = None
     ) -> Tensor:
         device = get_device_from_parameters(self)
         dtype = get_dtype_from_parameters(self)
 
         # Sample prior.
-        sample = (
-            noise
-            if noise is not None
-            else torch.randn(
-                size=(batch_size, self.config.horizon, self.config.action_feature.shape[0]),
-                dtype=dtype,
-                device=device,
-                generator=generator,
-            )
+        sample = torch.randn(
+            size=(batch_size, self.config.horizon, self.config.action_feature.shape[0]),
+            dtype=dtype,
+            device=device,
+            generator=generator,
         )
 
         self.noise_scheduler.set_timesteps(self.num_inference_steps)
@@ -242,7 +234,7 @@ class DiffusionModel(nn.Module):
         if self.config.image_features:
             if self.config.use_separate_rgb_encoder_per_camera:
                 # Combine batch and sequence dims while rearranging to make the camera index dimension first.
-                images_per_camera = einops.rearrange(batch[OBS_IMAGES], "b s n ... -> n (b s) ...")
+                images_per_camera = einops.rearrange(batch["observation.images"], "b s n ... -> n (b s) ...")
                 img_features_list = torch.cat(
                     [
                         encoder(images)
@@ -257,7 +249,7 @@ class DiffusionModel(nn.Module):
             else:
                 # Combine batch, sequence, and "which camera" dims before passing to shared encoder.
                 img_features = self.rgb_encoder(
-                    einops.rearrange(batch[OBS_IMAGES], "b s n ... -> (b s n) ...")
+                    einops.rearrange(batch["observation.images"], "b s n ... -> (b s n) ...")
                 )
                 # Separate batch dim and sequence dim back out. The camera index dim gets absorbed into the
                 # feature dim (effectively concatenating the camera features).
@@ -272,7 +264,7 @@ class DiffusionModel(nn.Module):
         # Concatenate features then flatten to (B, global_cond_dim).
         return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)
 
-    def generate_actions(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+    def generate_actions(self, batch: dict[str, Tensor]) -> Tensor:
         """
         This function expects `batch` to have:
         {
@@ -283,14 +275,14 @@ class DiffusionModel(nn.Module):
             "observation.environment_state": (B, n_obs_steps, environment_dim)
         }
         """
-        batch_size, n_obs_steps = batch[OBS_STATE].shape[:2]
+        batch_size, n_obs_steps = batch["observation.state"].shape[:2]
         assert n_obs_steps == self.config.n_obs_steps
 
         # Encode image features and concatenate them all together along with the state vector.
         global_cond = self._prepare_global_conditioning(batch)  # (B, global_cond_dim)
 
         # run sampling
-        actions = self.conditional_sample(batch_size, global_cond=global_cond, noise=noise)
+        actions = self.conditional_sample(batch_size, global_cond=global_cond)
 
         # Extract `n_action_steps` steps worth of actions (from the current observation).
         start = n_obs_steps - 1
@@ -314,10 +306,10 @@ class DiffusionModel(nn.Module):
         }
         """
         # Input validation.
-        assert set(batch).issuperset({OBS_STATE, ACTION, "action_is_pad"})
-        assert OBS_IMAGES in batch or OBS_ENV_STATE in batch
-        n_obs_steps = batch[OBS_STATE].shape[1]
-        horizon = batch[ACTION].shape[1]
+        assert set(batch).issuperset({"observation.state", "action", "action_is_pad"})
+        assert "observation.images" in batch or "observation.environment_state" in batch
+        n_obs_steps = batch["observation.state"].shape[1]
+        horizon = batch["action"].shape[1]
         assert horizon == self.config.horizon
         assert n_obs_steps == self.config.n_obs_steps
 
@@ -325,7 +317,7 @@ class DiffusionModel(nn.Module):
         global_cond = self._prepare_global_conditioning(batch)  # (B, global_cond_dim)
 
         # Forward diffusion.
-        trajectory = batch[ACTION]
+        trajectory = batch["action"]
         # Sample noise to add to the trajectory.
         eps = torch.randn(trajectory.shape, device=trajectory.device)
         # Sample a random noising timestep for each item in the batch.
@@ -346,7 +338,7 @@ class DiffusionModel(nn.Module):
         if self.config.prediction_type == "epsilon":
             target = eps
         elif self.config.prediction_type == "sample":
-            target = batch[ACTION]
+            target = batch["action"]
         else:
             raise ValueError(f"Unsupported prediction type {self.config.prediction_type}")