Merge branch 'main' into feat/add_macos_ci

* add metaworld

* smol update

Signed-off-by: Jade Choghari <chogharijade@gmail.com>

* update design

* Update src/lerobot/envs/metaworld.py

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Signed-off-by: Jade Choghari <chogharijade@gmail.com>

* update

* small changes

* iterate on review

* small fix

* small fix

* add docs

* update doc

* add better gif

* smol doc fix

* updage gymnasium

* add note

* depreciate gym-xarm

* more changes

* update doc

* comply with mypy

* more fixes

* update readme

* precommit

* update pusht

* add pusht instead

* changes

* style

* add changes

* update

* revert

* update v2

* chore(envs): move metaworld config to its own file + remove comments + simplify _format_raw_obs (#2200)

* update final changes

---------

Signed-off-by: Jade Choghari <chogharijade@gmail.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

.github/ISSUE_TEMPLATE/bug-report.yml

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

.github/workflows/fast_tests.yml

@@ -57,7 +57,11 @@ jobs:
   # It runs everytime we commit to a PR or push to main
   fast-pytest-tests:
     name: Fast Pytest Tests
-    runs-on: ubuntu-latest
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest]
     env:
       MUJOCO_GL: egl
     steps:
@@ -67,12 +71,21 @@ jobs:
           lfs: true
 
       # TODO(Steven): Evaluate the need of these dependencies
-      - name: Install apt dependencies
+      - name: Install 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
-
+          if [[ "${{ matrix.os }}" == 'ubuntu-latest' ]]; then
+            sudo apt-get update && sudo apt-get install -y build-essential \
+            git curl libglib2.0-0 libegl1-mesa-dev ffmpeg libusb-1.0-0-dev \
+            speech-dispatcher libgeos-dev portaudio19-dev
+          elif [[ "${{ matrix.os }}" == 'macos-latest' ]]; then
+            brew update && brew install git geos portaudio ffmpeg@7
+            # Add ffmpeg@7 paths for subsequent steps
+            echo "PATH=/opt/homebrew/opt/ffmpeg@7/bin:$PATH" >> $GITHUB_ENV
+            echo "LDFLAGS=-L/opt/homebrew/opt/ffmpeg@7/lib" >> $GITHUB_ENV
+            echo "CPPFLAGS=-I/opt/homebrew/opt/ffmpeg@7/include" >> $GITHUB_ENV
+            echo "PKG_CONFIG_PATH=/opt/homebrew/opt/ffmpeg@7/lib/pkgconfig" >> $GITHUB_ENV
+            echo "DYLD_LIBRARY_PATH=/opt/homebrew/opt/ffmpeg@7/lib:/opt/homebrew/lib:/usr/local/lib:$DYLD_LIBRARY_PATH" >> $GITHUB_ENV
+          fi
       - name: Setup uv and Python
         uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
         with:

.github/workflows/full_tests.yml

@@ -51,7 +51,11 @@ jobs:
   # It runs everytime a PR is approved or a push to main
   full-tests:
     name: Full Tests
-    runs-on: ubuntu-latest
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest]
     if: |
       (github.event_name == 'pull_request_review' && github.event.review.state == 'approved') ||
       github.event_name == 'push' ||
@@ -64,11 +68,16 @@ jobs:
           lfs: true
           persist-credentials: false
 
-      - name: Install apt dependencies
+      - name: Install 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
+          if [[ "${{ matrix.os }}" == 'ubuntu-latest' ]]; then
+            sudo apt-get update && sudo apt-get install -y build-essential \
+            git curl libglib2.0-0 libegl1-mesa-dev ffmpeg libusb-1.0-0-dev \
+            speech-dispatcher libgeos-dev portaudio19-dev
+          elif [[ "${{ matrix.os }}" == 'macos-latest' ]]; then
+            brew update && brew install git geos portaudio ffmpeg@7
+            echo "DYLD_LIBRARY_PATH=/opt/homebrew/opt/ffmpeg@7/lib:/opt/homebrew/lib:/usr/local/lib:$DYLD_LIBRARY_PATH" >> $GITHUB_ENV
+          fi
 
       - name: Setup uv and Python
         uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]

.github/workflows/release.yml

@@ -120,7 +120,11 @@ jobs:
   test-release:
     name: Test Release
     needs: [build-and-publish]
-    runs-on: ubuntu-latest
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest]
     permissions:
       contents: read
     env:
@@ -130,11 +134,16 @@ jobs:
         with:
           lfs: true
           persist-credentials: false
-      - name: Install apt dependencies
+      - name: Install dependencies
         run: |
-          sudo apt-get update && sudo apt-get install -y build-essential \
-          git curl libglib2.0-0 libegl1-mesa-dev ffmpeg libusb-1.0-0-dev \
-          speech-dispatcher libgeos-dev portaudio19-dev
+          if [[ "${{ matrix.os }}" == 'ubuntu-latest' ]]; then
+            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
+          elif [[ "${{ matrix.os }}" == 'macos-latest' ]]; then
+            brew update && brew install git geos portaudio ffmpeg@7
+            echo "DYLD_LIBRARY_PATH=/opt/homebrew/opt/ffmpeg@7/lib:/opt/homebrew/lib:/usr/local/lib:$DYLD_LIBRARY_PATH" >> $GITHUB_ENV
+          fi
       - name: Setup uv and Python
         uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
         with:

.github/workflows/stale.yml

@@ -0,0 +1,68 @@
+# 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 closing stale issues and PRs.
+name: Stale
+on:
+  # Allows running this workflow manually from the Actions tab
+  workflow_dispatch:
+
+  # Runs at 02:00
+  schedule:
+    - cron: "0 2 * * *"
+
+env:
+  CLOSE_ISSUE_MESSAGE: >
+    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 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.
+    Thank you for your contributions.
+  WARN_PR_MESSAGE: >
+    This PR has been automatically marked as stale because it has not had
+    recent activity (6 months). It will be closed if no further activity occurs.
+    Thank you for your contributions.
+
+jobs:
+  # This job runs the actions/stale action to close stale issues and PRs.
+  stale:
+    name: Close Stale Issues and PRs
+    runs-on: ubuntu-latest
+    permissions:
+      actions: write
+      contents: write # only for delete-branch option
+      issues: write
+      pull-requests: write
+    steps:
+      - uses: actions/stale@v10
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
+          stale-issue-label: stale
+          stale-pr-label: stale
+          exempt-issue-labels: never-stale
+          exempt-pr-labels: never-stale
+          days-before-issue-stale: 180 # TODO(Steven): Will modify this to 90 after initial cleanup
+          days-before-issue-close: 14
+          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 }}
+          stale-issue-message: ${{ env.WARN_ISSUE_MESSAGE }}
+          stale-pr-message: ${{ env.WARN_PR_MESSAGE }}
+          operations-per-run: 500

.github/workflows/unbound_deps_tests.yml

@@ -0,0 +1,192 @@
+# 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: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+    env:
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          lfs: true
+          persist-credentials: false
+
+      - name: Install dependencies
+        run: |
+          if [[ "${{ matrix.os }}" == 'ubuntu-latest' ]]; then
+            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
+          elif [[ "${{ matrix.os }}" == 'macos-latest' ]]; then
+            brew update && brew install git geos portaudio ffmpeg@7
+            echo "DYLD_LIBRARY_PATH=/opt/homebrew/opt/ffmpeg@7/lib:/opt/homebrew/lib:/usr/local/lib:$DYLD_LIBRARY_PATH" >> $GITHUB_ENV
+          fi
+
+      - 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

.gitignore

@@ -173,3 +173,7 @@ outputs/
 
 # Dev folders
 .cache/*
+*.stl
+*.urdf
+*.xml
+*.part

.pre-commit-config.yaml

@@ -86,11 +86,12 @@ repos:
 
   # TODO(Steven): Uncomment when ready to use
   ##### Static Analysis & Typing #####
-  # - repo: https://github.com/pre-commit/mirrors-mypy
-  #   rev: v1.16.0
-  #   hooks:
-  #     - id: mypy
-  #       args: [--python-version=3.10]
+  - repo: https://github.com/pre-commit/mirrors-mypy
+    rev: v1.16.0
+    hooks:
+      - id: mypy
+        args: [--config-file=pyproject.toml]
+        exclude: ^(examples|benchmarks|tests)/
 
   ##### Docstring Checks #####
   # - repo: https://github.com/akaihola/darglint2

CONTRIBUTING.md

@@ -72,7 +72,6 @@ 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.
 

Makefile

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

README.md

@@ -197,12 +197,12 @@ wandb login
 
 ### Visualize datasets
 
-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.
+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.
 
 You can also locally visualize episodes from a dataset on the hub by executing our script from the command line:
 
 ```bash
-python -m lerobot.scripts.visualize_dataset \
+lerobot-dataset-viz \
     --repo-id lerobot/pusht \
     --episode-index 0
 ```
@@ -210,7 +210,7 @@ python -m lerobot.scripts.visualize_dataset \
 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
-python -m lerobot.scripts.visualize_dataset \
+lerobot-dataset-viz \
     --repo-id lerobot/pusht \
     --root ./my_local_data_dir \
     --local-files-only 1 \
@@ -221,19 +221,19 @@ 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 `python -m lerobot.scripts.visualize_dataset --help` for more instructions.
+Our script can also visualize datasets stored on a distant server. See `lerobot-dataset-viz --help` for more instructions.
 
 ### The `LeRobotDataset` format
 
 A dataset in `LeRobotDataset` format is very simple to use. It can be loaded from a repository on the Hugging Face hub or a local folder simply with e.g. `dataset = LeRobotDataset("lerobot/aloha_static_coffee")` and can be indexed into like any Hugging Face and PyTorch dataset. For instance `dataset[0]` will retrieve a single temporal frame from the dataset containing observation(s) and an action as PyTorch tensors ready to be fed to a model.
 
-A specificity of `LeRobotDataset` is that, rather than retrieving a single frame by its index, we can retrieve several frames based on their temporal relationship with the indexed frame, by setting `delta_timestamps` to a list of relative times with respect to the indexed frame. For example, with `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}` one can retrieve, for a given index, 4 frames: 3 "previous" frames 1 second, 0.5 seconds, and 0.2 seconds before the indexed frame, and the indexed frame itself (corresponding to the 0 entry). See example [1_load_lerobot_dataset.py](https://github.com/huggingface/lerobot/blob/main/examples/1_load_lerobot_dataset.py) for more details on `delta_timestamps`.
+A specificity of `LeRobotDataset` is that, rather than retrieving a single frame by its index, we can retrieve several frames based on their temporal relationship with the indexed frame, by setting `delta_timestamps` to a list of relative times with respect to the indexed frame. For example, with `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}` one can retrieve, for a given index, 4 frames: 3 "previous" frames 1 second, 0.5 seconds, and 0.2 seconds before the indexed frame, and the indexed frame itself (corresponding to the 0 entry). See example [1_load_lerobot_dataset.py](https://github.com/huggingface/lerobot/blob/main/examples/dataset/load_lerobot_dataset.py) for more details on `delta_timestamps`.
 
 Under the hood, the `LeRobotDataset` format makes use of several ways to serialize data which can be useful to understand if you plan to work more closely with this format. We tried to make a flexible yet simple dataset format that would cover most type of features and specificities present in reinforcement learning and robotics, in simulation and in real-world, with a focus on cameras and robot states but easily extended to other types of sensory inputs as long as they can be represented by a tensor.
 
 Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
 
-````
+```
 dataset attributes:
   ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
   │  ├ observation.images.cam_high (VideoFrame):
@@ -269,7 +269,7 @@ dataset attributes:
   ├ root (Path): local directory where the dataset is stored
   ├ image_transforms (Callable): optional image transformations to apply to visual modalities
   └ delta_timestamps (dict): optional delta timestamps for temporal queries
-decoding videos (e.g., 'pyav', 'torchcodec')
+```
 
 A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
 
@@ -279,42 +279,6 @@ A `LeRobotDataset` is serialised using several widespread file formats for each
 
 Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
 
-### Evaluate a pretrained policy
-
-Check out [example 2](https://github.com/huggingface/lerobot/blob/main/examples/2_evaluate_pretrained_policy.py) that illustrates how to download a pretrained policy from Hugging Face hub, and run an evaluation on its corresponding environment.
-
-We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):
-
-```bash
-lerobot-eval \
-    --policy.path=lerobot/diffusion_pusht \
-    --env.type=pusht \
-    --eval.batch_size=10 \
-    --eval.n_episodes=10 \
-    --policy.use_amp=false \
-    --policy.device=cuda
-````
-
-Note: After training your own policy, you can re-evaluate the checkpoints with:
-
-```bash
-lerobot-eval --policy.path={OUTPUT_DIR}/checkpoints/last/pretrained_model
-```
-
-See `lerobot-eval --help` for more instructions.
-
-### Train your own policy
-
-Check out [example 3](https://github.com/huggingface/lerobot/blob/main/examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](https://github.com/huggingface/lerobot/blob/main/examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
-
-To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding `--wandb.enable=true`.
-
-A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser. Please also check [here](https://github.com/huggingface/lerobot/blob/main/examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explanation of some commonly used metrics in logs.
-
-\<img src="https://raw.githubusercontent.com/huggingface/lerobot/main/media/wandb.png" alt="WandB logs example"\>
-
-Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `--eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `lerobot-eval --help` for more instructions.
-
 #### Reproduce state-of-the-art (SOTA)
 
 We provide some pretrained policies on our [hub page](https://huggingface.co/lerobot) that can achieve state-of-the-art performances.
@@ -373,3 +337,7 @@ If you want, you can cite this work with:
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)
+
+```
+
+```

benchmarks/video/run_video_benchmark.py

@@ -35,12 +35,13 @@ 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.benchmark import TimeBenchmark
+from lerobot.utils.constants import OBS_IMAGE
 
 BASE_ENCODING = OrderedDict(
     [
@@ -117,7 +118,7 @@ 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("observation.image")]
+    img_keys = [key for key in hf_dataset.features if key.startswith(OBS_IMAGE)]
     imgs_dataset = hf_dataset.select_columns(img_keys[0])
 
     for i, item in enumerate(

docker/Dockerfile.internal

@@ -39,6 +39,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     software-properties-common build-essential git curl \
     libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
     libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
+    cmake pkg-config ninja-build \
     && add-apt-repository -y ppa:deadsnakes/ppa \
     && apt-get update \
     && apt-get install -y --no-install-recommends \
@@ -74,6 +75,14 @@ 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

@@ -31,6 +31,7 @@ ENV DEBIAN_FRONTEND=noninteractive \
 RUN apt-get update && apt-get install -y --no-install-recommends \
     build-essential git curl libglib2.0-0 libegl1-mesa-dev ffmpeg \
     libusb-1.0-0-dev speech-dispatcher libgeos-dev portaudio19-dev \
+    cmake pkg-config ninja-build \
     && curl -LsSf https://astral.sh/uv/install.sh | sh \
     && mv /root/.local/bin/uv /usr/local/bin/uv \
     && useradd --create-home --shell /bin/bash user_lerobot \
@@ -60,6 +61,14 @@ 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

@@ -7,8 +7,6 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
-  - local: il_sim
-    title: Imitation Learning in Sim
   - local: cameras
     title: Cameras
   - local: integrate_hardware
@@ -25,14 +23,38 @@
     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: Finetune SmolVLA
+    title: SmolVLA
+  - local: pi0
+    title: π₀ (Pi0)
+  - local: pi05
+    title: π₀.₅ (Pi05)
   title: "Policies"
 - sections:
-  - local: hope_jr
-    title: Hope Jr
+  - local: il_sim
+    title: Imitation Learning in Sim
+  - local: libero
+    title: Using Libero
+  - local: metaworld
+    title: Using MetaWorld
+  title: "Simulation"
+- sections:
+  - local: introduction_processors
+    title: Introduction to Robot Processors
+  - local: debug_processor_pipeline
+    title: Debug your processor pipeline
+  - local: implement_your_own_processor
+    title: Implement your own processor
+  - local: processors_robots_teleop
+    title: Processors for Robots and Teleoperators
+  title: "Robot Processors"
+- sections:
   - local: so101
     title: SO-101
   - local: so100
@@ -41,9 +63,15 @@
     title: Koch v1.1
   - local: lekiwi
     title: LeKiwi
+  - local: hope_jr
+    title: Hope Jr
   - local: reachy2
     title: Reachy 2
   title: "Robots"
+- sections:
+  - local: phone_teleop
+    title: Phone
+  title: "Teleoperators"
 - sections:
   - local: notebooks
     title: Notebooks

docs/source/act.mdx

@@ -0,0 +1,92 @@
+# 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 src/lerobot/scripts/server/policy_server.py \
-    --host=127.0.0.1 \
-    --port=8080 \
+python -m lerobot.async_inference.policy_server \
+     --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 src/lerobot/scripts/server/robot_client.py \
+python -m lerobot.async_inference.robot_client \
     --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.scripts.server.policy_server \
-    --host="localhost" \
-    --port=8080
+python -m lerobot.async_inference.policy_server \
+     --host=127.0.0.1 \
+     --port=8080
 ```
 </hfoption>
 <hfoption id="API example">
 
 <!-- prettier-ignore-start -->
 ```python
-from lerobot.scripts.server.configs import PolicyServerConfig
-from lerobot.scripts.server.policy_server import serve
+from lerobot.async_inference.configs import PolicyServerConfig
+from lerobot.async_inference.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 src/lerobot/scripts/server/robot_client.py \
+python -m lerobot.async_inference.robot_client \
     --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 src/lerobot/scripts/server/robot_client.py \
 import threading
 from lerobot.robots.so100_follower import SO100FollowerConfig
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-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
+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
 
 # 1. Create the robot instance
 """Check out the cameras available in your setup by running `python lerobot/find_cameras.py`"""

docs/source/backwardcomp.mdx

@@ -1,5 +1,61 @@
 # Backward compatibility
 
+## Policy Normalization Migration (PR #1452)
+
+**Breaking Change**: LeRobot policies no longer have built-in normalization layers embedded in their weights. Normalization is now handled by external `PolicyProcessorPipeline` components.
+
+### What changed?
+
+|                            | Before PR #1452                                  | After PR #1452                                               |
+| -------------------------- | ------------------------------------------------ | ------------------------------------------------------------ |
+| **Normalization Location** | Embedded in model weights (`normalize_inputs.*`) | External `PolicyProcessorPipeline` components                |
+| **Model State Dict**       | Contains normalization statistics                | **Clean weights only** - no normalization parameters         |
+| **Usage**                  | `policy(batch)` handles everything               | `preprocessor(batch)` → `policy(...)` → `postprocessor(...)` |
+
+### Impact on existing models
+
+- Models trained **before** PR #1452 have normalization embedded in their weights
+- These models need migration to work with the new `PolicyProcessorPipeline` system
+- The migration extracts normalization statistics and creates separate processor pipelines
+
+### Migrating old models
+
+Use the migration script to convert models with embedded normalization:
+
+```shell
+python src/lerobot/processor/migrate_policy_normalization.py \
+    --pretrained-path lerobot/act_aloha_sim_transfer_cube_human \
+    --push-to-hub \
+    --branch migrated
+```
+
+The script:
+
+1. **Extracts** normalization statistics from model weights
+2. **Creates** external preprocessor and postprocessor pipelines
+3. **Removes** normalization layers from model weights
+4. **Saves** clean model + processor pipelines
+5. **Pushes** to Hub with automatic PR creation
+
+### Using migrated models
+
+```python
+# New usage pattern (after migration)
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+
+# Load model and processors separately
+policy = make_policy(config, ds_meta=dataset.meta)
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=config,
+    dataset_stats=dataset.meta.stats
+)
+
+# Process data through pipeline
+processed_batch = preprocessor(raw_batch)
+action = policy.select_action(processed_batch)
+final_action = postprocessor(action)
+```
+
 ## Hardware API redesign
 
 PR [#777](https://github.com/huggingface/lerobot/pull/777) improves the LeRobot calibration but is **not backward-compatible**. Below is a overview of what changed and how you can continue to work with datasets created before this pull request.

docs/source/debug_processor_pipeline.mdx

@@ -0,0 +1,299 @@
+# Debug Your Processor Pipeline
+
+Processor pipelines can be complex, especially when chaining multiple transformation steps.
+Unlike simple function calls, pipelines lack natural observability, you can't easily see what happens
+between each step or where things go wrong.
+This guide provides debugging tools and techniques specifically designed to address these challenges
+and help you understand data flow through your pipelines.
+
+We'll explore three complementary debugging approaches: **hooks** for runtime monitoring, **step-through debugging** for detailed inspection, and **feature validation** for catching structural mismatches. Each serves a different purpose and together they provide complete visibility into your pipeline's behavior.
+
+## Understanding Hooks
+
+Hooks are functions that get called at specific points during pipeline execution.
+They provide a way to inspect, monitor, or modify data without changing your pipeline code.
+Think of them as "event listeners" for your pipeline.
+
+### What is a Hook?
+
+A hook is a callback function that gets automatically invoked at specific moments during pipeline execution.
+The concept comes from event-driven programming, imagine you could "hook into" the pipeline's execution flow to observe or react to what's happening.
+
+Think of hooks like inserting checkpoints into your pipeline. Every time the pipeline reaches one of these checkpoints, it pauses briefly to call your hook function, giving you a chance to inspect the current state, log information, and validate data.
+
+A hook is simply a function that accepts two parameters:
+
+- `step_idx: int` - The index of the current processing step (0, 1, 2, etc.)
+- `transition: EnvTransition` - The data transition at that point in the pipeline
+
+The beauty of hooks is their non-invasive nature: you can add monitoring, validation, or debugging logic without changing a single line of your pipeline code. The pipeline remains clean and focused on its core logic, while hooks handle the cross-cutting concerns like logging, monitoring, and debugging.
+
+### Before vs After Hooks
+
+The pipeline supports two types of hooks:
+
+- **Before hooks** (`register_before_step_hook`) - Called before each step executes
+- **After hooks** (`register_after_step_hook`) - Called after each step completes
+
+```python
+def before_hook(step_idx: int, transition: EnvTransition):
+    """Called before step processes the transition."""
+    print(f"About to execute step {step_idx}")
+    # Useful for: logging, validation, setup
+
+def after_hook(step_idx: int, transition: EnvTransition):
+    """Called after step has processed the transition."""
+    print(f"Completed step {step_idx}")
+    # Useful for: monitoring results, cleanup, debugging
+
+processor.register_before_step_hook(before_hook)
+processor.register_after_step_hook(after_hook)
+```
+
+### Implementing a NaN Detection Hook
+
+Here's a practical example of a hook that detects NaN values:
+
+```python
+def check_nans(step_idx: int, transition: EnvTransition):
+    """Check for NaN values in observations."""
+    obs = transition.get(TransitionKey.OBSERVATION)
+    if obs:
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor) and torch.isnan(value).any():
+                print(f"NaN detected in {key} at step {step_idx}")
+
+# Register the hook to run after each step
+processor.register_after_step_hook(check_nans)
+
+# Process your data - the hook will be called automatically
+output = processor(input_data)
+
+# Remove the hook when done debugging
+processor.unregister_after_step_hook(check_nans)
+```
+
+### How Hooks Work Internally
+
+Understanding the internal mechanism helps you use hooks more effectively. The pipeline maintains two separate lists: one for before-step hooks and another for after-step hooks. When you register a hook, it's simply appended to the appropriate list.
+
+During execution, the pipeline follows a strict sequence: for each processing step, it first calls all before-hooks in registration order, then executes the actual step transformation, and finally calls all after-hooks in registration order. This creates a predictable, sandwich-like structure around each step.
+
+The key insight is that hooks don't change the core pipeline logic—they're purely additive. The pipeline's `_forward` method orchestrates this dance between hooks and processing steps, ensuring that your debugging or monitoring code runs at exactly the right moments without interfering with the main data flow.
+
+Here's a simplified view of how the pipeline executes hooks:
+
+```python
+class DataProcessorPipeline:
+    def __init__(self):
+        self.steps = [...]
+        self.before_step_hooks = []  # List of before hooks
+        self.after_step_hooks = []   # List of after hooks
+
+    def _forward(self, transition):
+        """Internal method that processes the transition through all steps."""
+        for step_idx, processor_step in enumerate(self.steps):
+            # 1. Call all BEFORE hooks
+            for hook in self.before_step_hooks:
+                hook(step_idx, transition)
+
+            # 2. Execute the actual processing step
+            transition = processor_step(transition)
+
+            # 3. Call all AFTER hooks
+            for hook in self.after_step_hooks:
+                hook(step_idx, transition)
+
+        return transition
+
+    def register_before_step_hook(self, hook_fn):
+        self.before_step_hooks.append(hook_fn)
+
+    def register_after_step_hook(self, hook_fn):
+        self.after_step_hooks.append(hook_fn)
+```
+
+### Execution Flow
+
+The execution flow looks like this:
+
+```
+Input → Before Hook → Step 0 → After Hook → Before Hook → Step 1 → After Hook → ... → Output
+```
+
+For example, with 3 steps and both hook types:
+
+```python
+def timing_before(step_idx, transition):
+    print(f"⏱️  Starting step {step_idx}")
+
+def validation_after(step_idx, transition):
+    print(f"✅ Completed step {step_idx}")
+
+processor.register_before_step_hook(timing_before)
+processor.register_after_step_hook(validation_after)
+
+# This will output:
+# ⏱️  Starting step 0
+# ✅ Completed step 0
+# ⏱️  Starting step 1
+# ✅ Completed step 1
+# ⏱️  Starting step 2
+# ✅ Completed step 2
+```
+
+### Multiple Hooks
+
+You can register multiple hooks of the same type - they execute in the order registered:
+
+```python
+def log_shapes(step_idx: int, transition: EnvTransition):
+    obs = transition.get(TransitionKey.OBSERVATION)
+    if obs:
+        print(f"Step {step_idx} observation shapes:")
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor):
+                print(f"  {key}: {value.shape}")
+
+processor.register_after_step_hook(check_nans)      # Executes first
+processor.register_after_step_hook(log_shapes)     # Executes second
+
+# Both hooks will be called after each step in registration order
+output = processor(input_data)
+```
+
+While hooks are excellent for monitoring specific issues (like NaN detection) or gathering metrics during normal pipeline execution, sometimes you need to dive deeper. When you want to understand exactly what happens at each step or debug complex transformation logic, step-through debugging provides the detailed inspection you need.
+
+## Step-Through Debugging
+
+Step-through debugging is like having a slow-motion replay for your pipeline. Instead of watching your data get transformed in one quick blur from input to output, you can pause and examine what happens after each individual step.
+
+This approach is particularly valuable when you're trying to understand a complex pipeline, debug unexpected behavior, or verify that each transformation is working as expected. Unlike hooks, which are great for automated monitoring, step-through debugging gives you manual, interactive control over the inspection process.
+
+The `step_through()` method is a generator that yields the transition state after each processing step, allowing you to inspect intermediate results. Think of it as creating a series of snapshots of your data as it flows through the pipeline—each snapshot shows you exactly what your data looks like after one more transformation has been applied.
+
+### How Step-Through Works
+
+The `step_through()` method fundamentally changes how the pipeline executes. Instead of running all steps in sequence and only returning the final result, it transforms the pipeline into an iterator that yields intermediate results.
+
+Here's what happens internally: the method starts by converting your input data into the pipeline's internal transition format, then yields this initial state. Next, it applies the first processing step and yields the result. Then it applies the second step to that result and yields again, and so on. Each `yield` gives you a complete snapshot of the transition at that point.
+
+This generator pattern is powerful because it's lazy—the pipeline only computes the next step when you ask for it. This means you can stop at any point, inspect the current state thoroughly, and decide whether to continue. You're not forced to run the entire pipeline just to debug one problematic step.
+
+Instead of running the entire pipeline and only seeing the final result, `step_through()` pauses after each step and gives you the intermediate transition:
+
+```python
+# This creates a generator that yields intermediate states
+for i, intermediate_result in enumerate(processor.step_through(input_data)):
+    print(f"=== After step {i} ===")
+
+    # Inspect the observation at this stage
+    obs = intermediate_result.get(TransitionKey.OBSERVATION)
+    if obs:
+        for key, value in obs.items():
+            if isinstance(value, torch.Tensor):
+                print(f"{key}: shape={value.shape}, dtype={value.dtype}")
+```
+
+### Interactive Debugging with Breakpoints
+
+You can add breakpoints in the step-through loop to interactively debug:
+
+```python
+# Step through the pipeline with debugging
+for i, intermediate in enumerate(processor.step_through(data)):
+    print(f"Step {i}: {processor.steps[i].__class__.__name__}")
+
+    # Set a breakpoint to inspect the current state
+    breakpoint()  # Debugger will pause here
+
+    # You can now inspect 'intermediate' in the debugger:
+    # - Check tensor shapes and values
+    # - Verify expected transformations
+    # - Look for unexpected changes
+```
+
+During the debugger session, you can:
+
+- Examine `intermediate[TransitionKey.OBSERVATION]` to see observation data
+- Check `intermediate[TransitionKey.ACTION]` for action transformations
+- Inspect any part of the transition to understand what each step does
+
+Step-through debugging is perfect for understanding the _data_ transformations, but what about the _structure_ of that data? While hooks and step-through help you debug runtime behavior, you also need to ensure your pipeline produces data in the format expected by downstream components. This is where feature contract validation comes in.
+
+## Validating Feature Contracts
+
+Feature contracts define what data structure your pipeline expects as input and produces as output.
+Validating these contracts helps catch mismatches early.
+
+### Understanding Feature Contracts
+
+Each processor step has a `transform_features()` method that describes how it changes the data structure:
+
+```python
+# Get the expected output features from your pipeline
+initial_features = {
+    PipelineFeatureType.OBSERVATION: {
+        "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(7,)),
+        "observation.image": PolicyFeature(type=FeatureType.IMAGE, shape=(3, 224, 224))
+    },
+    PipelineFeatureType.ACTION: {
+        "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,))
+    }
+}
+
+# Check what your pipeline will output
+output_features = processor.transform_features(initial_features)
+
+print("Input features:")
+for feature_type, features in initial_features.items():
+    print(f"  {feature_type}:")
+    for key, feature in features.items():
+        print(f"    {key}: {feature.type.value}, shape={feature.shape}")
+
+print("\nOutput features:")
+for feature_type, features in output_features.items():
+    print(f"  {feature_type}:")
+    for key, feature in features.items():
+        print(f"    {key}: {feature.type.value}, shape={feature.shape}")
+```
+
+### Verifying Expected Features
+
+Check that your pipeline produces the features you expect:
+
+```python
+# Define what features you expect the pipeline to produce
+expected_keys = ["observation.state", "observation.image", "action"]
+
+print("Validating feature contract...")
+for expected_key in expected_keys:
+    found = False
+    for feature_type, features in output_features.items():
+        if expected_key in features:
+            feature = features[expected_key]
+            print(f"✅ {expected_key}: {feature.type.value}, shape={feature.shape}")
+            found = True
+            break
+
+    if not found:
+        print(f"❌ Missing expected feature: {expected_key}")
+```
+
+This validation helps ensure your pipeline will work correctly with downstream components that expect specific data structures.
+
+## Summary
+
+Now that you understand the three debugging approaches, you can tackle any pipeline issue systematically:
+
+1. **Hooks** - For runtime monitoring and validation without modifying pipeline code
+2. **Step-through** - For inspecting intermediate states and understanding transformations
+3. **Feature validation** - For ensuring data structure contracts are met
+
+**When to use each approach:**
+
+- Start with **step-through debugging** when you need to understand what your pipeline does or when something unexpected happens
+- Add **hooks** for continuous monitoring during development and production to catch issues automatically
+- Use **feature validation** before deployment to ensure your pipeline works with downstream components
+
+These three tools work together to give you the complete observability that complex pipelines naturally lack. With hooks watching for issues, step-through helping you understand behavior, and feature validation ensuring compatibility, you'll be able to debug any pipeline confidently and efficiently.

docs/source/hilserl.mdx

@@ -4,7 +4,13 @@ In this tutorial you will go through the full Human-in-the-Loop Sample-Efficient
 
 HIL-SERL is a sample-efficient reinforcement learning algorithm that combines human demonstrations with online learning and human interventions. The approach starts from a small set of human demonstrations, uses them to train a reward classifier, and then employs an actor-learner architecture where humans can intervene during policy execution to guide exploration and correct unsafe behaviors. In this tutorial, you'll use a gamepad to provide interventions and control the robot during the learning process.
 
-It combines three key ingredients: 1. **Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point. 2. **On-robot actor / learner loop with human interventions:** a distributed Soft Actor Critic (SAC) learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour. 3. **Safety & efficiency tools:** joint/end-effector (EE) bounds, crop region of interest (ROI) preprocessing and WandB monitoring keep the data useful and the hardware safe.
+It combines three key ingredients:
+
+1. **Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point.
+
+2. **On-robot actor / learner loop with human interventions:** a distributed Soft Actor Critic (SAC) learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour.
+
+3. **Safety & efficiency tools:** joint/end-effector (EE) bounds, crop region of interest (ROI) preprocessing and WandB monitoring keep the data useful and the hardware safe.
 
 Together these elements let HIL-SERL reach near-perfect task success and faster cycle times than imitation-only baselines.
 
@@ -56,49 +62,258 @@ pip install -e ".[hilserl]"
 
 ### Understanding Configuration
 
-The training process begins with proper configuration for the HILSerl environment. The configuration class of interest is `HILSerlRobotEnvConfig` in `lerobot/envs/configs.py`. Which is defined as:
+The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
 
 <!-- prettier-ignore-start -->
 ```python
+class GymManipulatorConfig:
+    env: HILSerlRobotEnvConfig    # Environment configuration (nested)
+    dataset: DatasetConfig    # Dataset recording/replay configuration (nested)
+    mode: str | None = None    # "record", "replay", or None (for training)
+    device: str = "cpu"    # Compute device
+
 class HILSerlRobotEnvConfig(EnvConfig):
     robot: RobotConfig | None = None    # Main robot agent (defined in `lerobot/robots`)
-    teleop: TeleoperatorConfig | None = None    # Teleoperator agent, e.g., gamepad or leader arm, (defined in `lerobot/teleoperators`)
-    wrapper: EnvTransformConfig | None = None    # Environment wrapper settings; check `lerobot/scripts/server/gym_manipulator.py`
-    fps: int = 10    # Control frequency
+    teleop: TeleoperatorConfig | None = None    # Teleoperator agent, e.g., gamepad or leader arm
+    processor: HILSerlProcessorConfig    # Processing pipeline configuration (nested)
     name: str = "real_robot"    # Environment name
-    mode: str = None    # "record", "replay", or None (for training)
-    repo_id: str | None = None    # LeRobot dataset repository ID
-    dataset_root: str | None = None    # Local dataset root (optional)
-    task: str = ""    # Task identifier
-    num_episodes: int = 10    # Number of episodes for recording
-    episode: int = 0    # episode index for replay
-    device: str = "cuda"    # Compute device
-    push_to_hub: bool = True    # Whether to push the recorded datasets to Hub
-    pretrained_policy_name_or_path: str | None = None    # For policy loading
-    reward_classifier_pretrained_path: str | None = None    # For reward model
-    number_of_steps_after_success: int = 0    # For reward classifier, collect more positive examples after a success to train a classifier
+    task: str | None = None    # Task identifier
+    fps: int = 10    # Control frequency
+
+# Nested processor configuration
+class HILSerlProcessorConfig:
+    control_mode: str = "gamepad"    # Control mode
+    observation: ObservationConfig | None = None    # Observation processing settings
+    image_preprocessing: ImagePreprocessingConfig | None = None    # Image crop/resize settings
+    gripper: GripperConfig | None = None    # Gripper control and penalty settings
+    reset: ResetConfig | None = None    # Environment reset and timing settings
+    inverse_kinematics: InverseKinematicsConfig | None = None    # IK processing settings
+    reward_classifier: RewardClassifierConfig | None = None    # Reward classifier settings
+    max_gripper_pos: float | None = 100.0    # Maximum gripper position
+
+# Sub-configuration classes
+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
+    display_cameras: bool = False    # Display camera feeds during execution
+
+class ImagePreprocessingConfig:
+    crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None    # Image cropping parameters
+    resize_size: tuple[int, int] | None = None    # Target image size
+
+class GripperConfig:
+    use_gripper: bool = True    # Enable gripper control
+    gripper_penalty: float = 0.0    # Penalty for inappropriate gripper usage
+
+class ResetConfig:
+    fixed_reset_joint_positions: Any | None = None    # Joint positions for reset
+    reset_time_s: float = 5.0    # Time to wait during reset
+    control_time_s: float = 20.0    # Maximum episode duration
+    terminate_on_success: bool = True    # Whether to terminate episodes on success detection
+
+class InverseKinematicsConfig:
+    urdf_path: str | None = None    # Path to robot URDF file
+    target_frame_name: str | None = None    # End-effector frame name
+    end_effector_bounds: dict[str, list[float]] | None = None    # EE workspace bounds
+    end_effector_step_sizes: dict[str, float] | None = None    # EE step sizes per axis
+
+class RewardClassifierConfig:
+    pretrained_path: str | None = None    # Path to pretrained reward classifier
+    success_threshold: float = 0.5    # Success detection threshold
+    success_reward: float = 1.0    # Reward value for successful episodes
+
+# Dataset configuration
+class DatasetConfig:
+    repo_id: str    # LeRobot dataset repository ID
+    task: str    # Task identifier
+    root: str | None = None    # Local dataset root directory
+    num_episodes_to_record: int = 5    # Number of episodes for recording
+    replay_episode: int | None = None    # Episode index for replay
+    push_to_hub: bool = False    # Whether to push datasets to Hub
 ```
 <!-- prettier-ignore-end -->
 
+### Processor Pipeline Architecture
+
+HIL-SERL uses a modular processor pipeline architecture that processes robot observations and actions through a series of composable steps. The pipeline is divided into two main components:
+
+#### Environment Processor Pipeline
+
+The environment processor (`env_processor`) handles incoming observations and environment state:
+
+1. **VanillaObservationProcessorStep**: Converts raw robot observations into standardized format
+2. **JointVelocityProcessorStep** (optional): Adds joint velocity information to observations
+3. **MotorCurrentProcessorStep** (optional): Adds motor current readings to observations
+4. **ForwardKinematicsJointsToEE** (optional): Computes end-effector pose from joint positions
+5. **ImageCropResizeProcessorStep** (optional): Crops and resizes camera images
+6. **TimeLimitProcessorStep** (optional): Enforces episode time limits
+7. **GripperPenaltyProcessorStep** (optional): Applies penalties for inappropriate gripper usage
+8. **RewardClassifierProcessorStep** (optional): Automated reward detection using vision models
+9. **AddBatchDimensionProcessorStep**: Converts data to batch format for neural network processing
+10. **DeviceProcessorStep**: Moves data to the specified compute device (CPU/GPU)
+
+#### Action Processor Pipeline
+
+The action processor (`action_processor`) handles outgoing actions and human interventions:
+
+1. **AddTeleopActionAsComplimentaryDataStep**: Captures teleoperator actions for logging
+2. **AddTeleopEventsAsInfoStep**: Records intervention events and episode control signals
+3. **InterventionActionProcessorStep**: Handles human interventions and episode termination
+4. **Inverse Kinematics Pipeline** (when enabled):
+   - **MapDeltaActionToRobotActionStep**: Converts delta actions to robot action format
+   - **EEReferenceAndDelta**: Computes end-effector reference and delta movements
+   - **EEBoundsAndSafety**: Enforces workspace safety bounds
+   - **InverseKinematicsEEToJoints**: Converts end-effector actions to joint targets
+   - **GripperVelocityToJoint**: Handles gripper control commands
+
+#### Configuration Examples
+
+**Basic Observation Processing**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true,
+        "add_current_to_observation": false,
+        "display_cameras": false
+      }
+    }
+  }
+}
+```
+
+**Image Processing**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "image_preprocessing": {
+        "crop_params_dict": {
+          "observation.images.front": [180, 250, 120, 150],
+          "observation.images.side": [180, 207, 180, 200]
+        },
+        "resize_size": [128, 128]
+      }
+    }
+  }
+}
+```
+
+**Inverse Kinematics Setup**:
+
+```json
+{
+  "env": {
+    "processor": {
+      "inverse_kinematics": {
+        "urdf_path": "path/to/robot.urdf",
+        "target_frame_name": "end_effector",
+        "end_effector_bounds": {
+          "min": [0.16, -0.08, 0.03],
+          "max": [0.24, 0.2, 0.1]
+        },
+        "end_effector_step_sizes": {
+          "x": 0.02,
+          "y": 0.02,
+          "z": 0.02
+        }
+      }
+    }
+  }
+}
+```
+
+### Advanced Observation Processing
+
+The HIL-SERL framework supports additional observation processing features that can improve policy learning:
+
+#### Joint Velocity Processing
+
+Enable joint velocity estimation to provide the policy with motion information:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true
+      }
+    }
+  }
+}
+```
+
+This processor:
+
+- Estimates joint velocities using finite differences between consecutive joint position readings
+- Adds velocity information to the observation state vector
+- Useful for policies that need motion awareness for dynamic tasks
+
+#### Motor Current Processing
+
+Monitor motor currents to detect contact forces and load conditions:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_current_to_observation": true
+      }
+    }
+  }
+}
+```
+
+This processor:
+
+- Reads motor current values from the robot's control system
+- Adds current measurements to the observation state vector
+- Helps detect contact events, object weights, and mechanical resistance
+- Useful for contact-rich manipulation tasks
+
+#### Combined Observation Processing
+
+You can enable multiple observation processing features simultaneously:
+
+```json
+{
+  "env": {
+    "processor": {
+      "observation": {
+        "add_joint_velocity_to_observation": true,
+        "add_current_to_observation": true,
+        "display_cameras": false
+      }
+    }
+  }
+}
+```
+
+**Note**: Enabling additional observation features increases the state space dimensionality, which may require adjusting your policy network architecture and potentially collecting more training data.
+
 ### Finding Robot Workspace Bounds
 
 Before collecting demonstrations, you need to determine the appropriate operational bounds for your robot.
 
 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 find_joint_limits.py**
+**Using lerobot-find-joint-limits**
 
 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
-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
+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
 ```
 
 **Workflow**
@@ -128,24 +343,58 @@ With the bounds defined, you can safely collect demonstrations for training. Tra
 
 **Setting Up Record Mode**
 
-Create a configuration file for recording demonstrations (or edit an existing one like [env_config_so100.json](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json)):
+Create a configuration file for recording demonstrations (or edit an existing one like [env_config.json](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/env_config.json)):
 
-1. Set `mode` to `"record"`
-2. Specify a unique `repo_id` for your dataset (e.g., "username/task_name")
-3. Set `num_episodes` to the number of demonstrations you want to collect
-4. Set `crop_params_dict` to `null` initially (we'll determine crops later)
-5. Configure `robot`, `cameras`, and other hardware settings
+1. Set `mode` to `"record"` at the root level
+2. Specify a unique `repo_id` for your dataset in the `dataset` section (e.g., "username/task_name")
+3. Set `num_episodes_to_record` in the `dataset` section to the number of demonstrations you want to collect
+4. Set `env.processor.image_preprocessing.crop_params_dict` to `{}` initially (we'll determine crops later)
+5. Configure `env.robot`, `env.teleop`, and other hardware settings in the `env` section
 
 Example configuration section:
 
 ```json
-"mode": "record",
-"repo_id": "username/pick_lift_cube",
-"dataset_root": null,
-"task": "pick_and_lift",
-"num_episodes": 15,
-"episode": 0,
-"push_to_hub": true
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "real_robot",
+    "fps": 10,
+    "processor": {
+      "control_mode": "gamepad",
+      "observation": {
+        "display_cameras": false
+      },
+      "image_preprocessing": {
+        "crop_params_dict": {},
+        "resize_size": [128, 128]
+      },
+      "gripper": {
+        "use_gripper": true,
+        "gripper_penalty": 0.0
+      },
+      "reset": {
+        "reset_time_s": 5.0,
+        "control_time_s": 20.0
+      }
+    },
+    "robot": {
+      // ... robot configuration ...
+    },
+    "teleop": {
+      // ... teleoperator configuration ...
+    }
+  },
+  "dataset": {
+    "repo_id": "username/pick_lift_cube",
+    "root": null,
+    "task": "pick_and_lift",
+    "num_episodes_to_record": 15,
+    "replay_episode": 0,
+    "push_to_hub": true
+  },
+  "mode": "record",
+  "device": "cpu"
+}
 ```
 
 ### Using a Teleoperation Device
@@ -191,10 +440,20 @@ The gamepad provides a very convenient way to control the robot and the episode
 To setup the gamepad, you need to set the `control_mode` to `"gamepad"` and define the `teleop` section in the configuration file.
 
 ```json
+{
+  "env": {
     "teleop": {
-        "type": "gamepad",
-        "use_gripper": true
+      "type": "gamepad",
+      "use_gripper": true
     },
+    "processor": {
+      "control_mode": "gamepad",
+      "gripper": {
+        "use_gripper": true
+      }
+    }
+  }
+}
 ```
 
 <p align="center">
@@ -216,11 +475,21 @@ The SO101 leader arm has reduced gears that allows it to move and track the foll
 To setup the SO101 leader, you need to set the `control_mode` to `"leader"` and define the `teleop` section in the configuration file.
 
 ```json
+{
+  "env": {
     "teleop": {
-        "type": "so101_leader",
-        "port": "/dev/tty.usbmodem585A0077921", # check your port number
-        "use_degrees": true
+      "type": "so101_leader",
+      "port": "/dev/tty.usbmodem585A0077921",
+      "use_degrees": true
     },
+    "processor": {
+      "control_mode": "leader",
+      "gripper": {
+        "use_gripper": true
+      }
+    }
+  }
+}
 ```
 
 In order to annotate the success/failure of the episode, **you will need** to use a keyboard to press `s` for success, `esc` for failure.
@@ -246,12 +515,12 @@ During the online training, press `space` to take over the policy and `space` ag
 Start the recording process, an example of the config file can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_so100.json):
 
 ```bash
-python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config_so100.json
+python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config_so100.json
 ```
 
 During recording:
 
-1. The robot will reset to the initial position defined in the configuration file `fixed_reset_joint_positions`
+1. The robot will reset to the initial position defined in the configuration file `env.processor.reset.fixed_reset_joint_positions`
 2. Complete the task successfully
 3. The episode ends with a reward of 1 when you press the "success" button
 4. If the time limit is reached, or the fail button is pressed, the episode ends with a reward of 0
@@ -277,7 +546,7 @@ Note: If you already know the crop parameters, you can skip this step and just s
 Use the `crop_dataset_roi.py` script to interactively select regions of interest in your camera images:
 
 ```bash
-python -m lerobot.scripts.rl.crop_dataset_roi --repo-id username/pick_lift_cube
+python -m lerobot.rl.crop_dataset_roi --repo-id username/pick_lift_cube
 ```
 
 1. For each camera view, the script will display the first frame
@@ -310,11 +579,19 @@ observation.images.front: [180, 250, 120, 150]
 Add these crop parameters to your training configuration:
 
 ```json
-"crop_params_dict": {
-    "observation.images.side": [180, 207, 180, 200],
-    "observation.images.front": [180, 250, 120, 150]
-},
-"resize_size": [128, 128]
+{
+  "env": {
+    "processor": {
+      "image_preprocessing": {
+        "crop_params_dict": {
+          "observation.images.side": [180, 207, 180, 200],
+          "observation.images.front": [180, 250, 120, 150]
+        },
+        "resize_size": [128, 128]
+      }
+    }
+  }
+}
 ```
 
 **Recommended image resolution**
@@ -338,31 +615,57 @@ Before training, you need to collect a dataset with labeled examples. The `recor
 To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
 
 ```bash
-python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/reward_classifier_train_config.json
+python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/reward_classifier_train_config.json
 ```
 
 **Key Parameters for Data Collection**
 
-- **mode**: set it to `"record"` to collect a dataset
-- **repo_id**: `"hf_username/dataset_name"`, name of the dataset and repo on the hub
-- **num_episodes**: Number of episodes to record
-- **number_of_steps_after_success**: Number of additional frames to record after a success (reward=1) is detected
-- **fps**: Number of frames per second to record
-- **push_to_hub**: Whether to push the dataset to the hub
+- **mode**: set it to `"record"` to collect a dataset (at root level)
+- **dataset.repo_id**: `"hf_username/dataset_name"`, name of the dataset and repo on the hub
+- **dataset.num_episodes_to_record**: Number of episodes to record
+- **env.processor.reset.terminate_on_success**: Whether to automatically terminate episodes when success is detected (default: `true`)
+- **env.fps**: Number of frames per second to record
+- **dataset.push_to_hub**: Whether to push the dataset to the hub
 
-The `number_of_steps_after_success` parameter is crucial as it allows you to collect more positive examples. When a success is detected, the system will continue recording for the specified number of steps while maintaining the reward=1 label. Otherwise, there won't be enough states in the dataset labeled to 1 to train a good classifier.
+The `env.processor.reset.terminate_on_success` parameter allows you to control episode termination behavior. When set to `false`, episodes will continue even after success is detected, allowing you to collect more positive examples with the reward=1 label. This is crucial for training reward classifiers as it provides more success state examples in your dataset. When set to `true` (default), episodes terminate immediately upon success detection.
+
+**Important**: For reward classifier training, set `terminate_on_success: false` to collect sufficient positive examples. For regular HIL-SERL training, keep it as `true` to enable automatic episode termination when the task is completed successfully.
 
 Example configuration section for data collection:
 
 ```json
 {
+  "env": {
+    "type": "gym_manipulator",
+    "name": "real_robot",
+    "fps": 10,
+    "processor": {
+      "reset": {
+        "reset_time_s": 5.0,
+        "control_time_s": 20.0,
+        "terminate_on_success": false
+      },
+      "gripper": {
+        "use_gripper": true
+      }
+    },
+    "robot": {
+      // ... robot configuration ...
+    },
+    "teleop": {
+      // ... teleoperator configuration ...
+    }
+  },
+  "dataset": {
+    "repo_id": "hf_username/dataset_name",
+    "dataset_root": "data/your_dataset",
+    "task": "reward_classifier_task",
+    "num_episodes_to_record": 20,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
   "mode": "record",
-  "repo_id": "hf_username/dataset_name",
-  "dataset_root": "data/your_dataset",
-  "num_episodes": 20,
-  "push_to_hub": true,
-  "fps": 10,
-  "number_of_steps_after_success": 15
+  "device": "cpu"
 }
 ```
 
@@ -421,9 +724,17 @@ To use your trained reward classifier, configure the `HILSerlRobotEnvConfig` to
 
 <!-- prettier-ignore-start -->
 ```python
-env_config = HILSerlRobotEnvConfig(
-    reward_classifier_pretrained_path="path_to_your_pretrained_trained_model",
-    # Other environment parameters
+config = GymManipulatorConfig(
+    env=HILSerlRobotEnvConfig(
+        processor=HILSerlProcessorConfig(
+            reward_classifier=RewardClassifierConfig(
+                pretrained_path="path_to_your_pretrained_trained_model"
+            )
+        ),
+        # Other environment parameters
+    ),
+    dataset=DatasetConfig(...),
+    mode=None  # For training
 )
 ```
 <!-- prettier-ignore-end -->
@@ -432,14 +743,25 @@ or set the argument in the json config file.
 
 ```json
 {
-  "reward_classifier_pretrained_path": "path_to_your_pretrained_model"
+  "env": {
+    "processor": {
+      "reward_classifier": {
+        "pretrained_path": "path_to_your_pretrained_model",
+        "success_threshold": 0.7,
+        "success_reward": 1.0
+      },
+      "reset": {
+        "terminate_on_success": true
+      }
+    }
+  }
 }
 ```
 
 Run `gym_manipulator.py` to test the model.
 
 ```bash
-python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config.json
+python -m lerobot.rl.gym_manipulator --config_path path/to/env_config.json
 ```
 
 The reward classifier will automatically provide rewards based on the visual input from the robot's cameras.
@@ -447,12 +769,12 @@ The reward classifier will automatically provide rewards based on the visual inp
 **Example Workflow for training the reward classifier**
 
 1. **Create the configuration files**:
-   Create the necessary json configuration files for the reward classifier and the environment. Check the examples [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/tree/main).
+   Create the necessary json configuration files for the reward classifier and the environment. Check the examples [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/reward_classifier/config.json).
 
 2. **Collect a dataset**:
 
    ```bash
-   python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
+   python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
    ```
 
 3. **Train the classifier**:
@@ -463,7 +785,7 @@ The reward classifier will automatically provide rewards based on the visual inp
 
 4. **Test the classifier**:
    ```bash
-   python -m lerobot.scripts.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
+   python -m lerobot.rl.gym_manipulator --config_path src/lerobot/configs/env_config.json
    ```
 
 ### Training with Actor-Learner
@@ -472,7 +794,7 @@ The LeRobot system uses a distributed actor-learner architecture for training. T
 
 **Configuration Setup**
 
-Create a training configuration file (example available [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/train_config_hilserl_so100.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.
+Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.
 
 1. Configure the policy settings (`type="sac"`, `device`, etc.)
 2. Set `dataset` to your cropped dataset
@@ -485,7 +807,7 @@ Create a training configuration file (example available [here](https://huggingfa
 First, start the learner server process:
 
 ```bash
-python -m lerobot.scripts.rl.learner --config_path src/lerobot/configs/train_config_hilserl_so100.json
+python -m lerobot.rl.learner --config_path src/lerobot/configs/train_config_hilserl_so100.json
 ```
 
 The learner:
@@ -500,7 +822,7 @@ The learner:
 In a separate terminal, start the actor process with the same configuration:
 
 ```bash
-python -m lerobot.scripts.rl.actor --config_path src/lerobot/configs/train_config_hilserl_so100.json
+python -m lerobot.rl.actor --config_path src/lerobot/configs/train_config_hilserl_so100.json
 ```
 
 The actor:

docs/source/hilserl_sim.mdx

@@ -26,15 +26,18 @@ pip install -e ".[hilserl]"
 
 ## Configuration
 
-To use `gym_hil` with LeRobot, you need to create a configuration file. An example is provided [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/gym_hil_env.json). Key configuration sections include:
+To use `gym_hil` with LeRobot, you need to create a configuration file. An example is provided [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/gym_hil/env_config.json). Key configuration sections include:
 
 ### Environment Type and Task
 
 ```json
 {
-  "type": "hil",
-  "name": "franka_sim",
-  "task": "PandaPickCubeGamepad-v0",
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
   "device": "cuda"
 }
 ```
@@ -45,28 +48,40 @@ Available tasks:
 - `PandaPickCubeGamepad-v0`: With gamepad control
 - `PandaPickCubeKeyboard-v0`: With keyboard control
 
-### Gym Wrappers Configuration
+### Processor Configuration
 
 ```json
-"wrapper": {
-    "gripper_penalty": -0.02,
-    "control_time_s": 15.0,
-    "use_gripper": true,
-    "fixed_reset_joint_positions": [0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785],
-    "end_effector_step_sizes": {
-        "x": 0.025,
-        "y": 0.025,
-        "z": 0.025
-    },
-    "control_mode": "gamepad"
+{
+  "env": {
+    "processor": {
+      "control_mode": "gamepad",
+      "gripper": {
+        "use_gripper": true,
+        "gripper_penalty": -0.02
+      },
+      "reset": {
+        "control_time_s": 15.0,
+        "fixed_reset_joint_positions": [
+          0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785
+        ]
+      },
+      "inverse_kinematics": {
+        "end_effector_step_sizes": {
+          "x": 0.025,
+          "y": 0.025,
+          "z": 0.025
+        }
+      }
     }
+  }
+}
 ```
 
 Important parameters:
 
-- `gripper_penalty`: Penalty for excessive gripper movement
-- `use_gripper`: Whether to enable gripper control
-- `end_effector_step_sizes`: Size of the steps in the x,y,z axes of the end-effector
+- `gripper.gripper_penalty`: Penalty for excessive gripper movement
+- `gripper.use_gripper`: Whether to enable gripper control
+- `inverse_kinematics.end_effector_step_sizes`: Size of the steps in the x,y,z axes of the end-effector
 - `control_mode`: Set to `"gamepad"` to use a gamepad controller
 
 ## Running with HIL RL of LeRobot
@@ -75,39 +90,50 @@ Important parameters:
 
 To run the environment, set mode to null:
 
-<!-- prettier-ignore-start -->
-```python
-python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/gym_hil_env.json
+```bash
+python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 ### Recording a Dataset
 
 To collect a dataset, set the mode to `record` whilst defining the repo_id and number of episodes to record:
 
-<!-- prettier-ignore-start -->
-```python
-python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/gym_hil_env.json
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0"
+  },
+  "dataset": {
+    "repo_id": "username/sim_dataset",
+    "root": null,
+    "task": "pick_cube",
+    "num_episodes_to_record": 10,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
+  "mode": "record"
+}
+```
+
+```bash
+python -m lerobot.rl.gym_manipulator --config_path path/to/gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 ### Training a Policy
 
-To train a policy, checkout the configuration example available [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/train_gym_hil_env.json) and run the actor and learner servers:
+To train a policy, checkout the configuration example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/gym_hil/train_config.json) and run the actor and learner servers:
 
-<!-- prettier-ignore-start -->
-```python
-python -m lerobot.scripts.rl.actor --config_path path/to/train_gym_hil_env.json
+```bash
+python -m lerobot.rl.actor --config_path path/to/train_gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 In a different terminal, run the learner server:
 
-<!-- prettier-ignore-start -->
-```python
-python -m lerobot.scripts.rl.learner --config_path path/to/train_gym_hil_env.json
+```bash
+python -m lerobot.rl.learner --config_path path/to/train_gym_hil_env.json
 ```
-<!-- prettier-ignore-end -->
 
 The simulation environment provides a safe and repeatable way to develop and test your Human-In-the-Loop reinforcement learning components before deploying to real robots.
 

docs/source/il_robots.mdx

@@ -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()
@@ -513,17 +513,21 @@ 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.record import record_loop
+from lerobot.utils.visualization_utils import init_rerun
+
 
 NUM_EPISODES = 5
 FPS = 30
 EPISODE_TIME_SEC = 60
 TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
 # Create the robot configuration
 camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
@@ -535,7 +539,7 @@ robot_config = SO100FollowerConfig(
 robot = SO100Follower(robot_config)
 
 # Initialize the policy
-policy = ACTPolicy.from_pretrained("<hf_username>/<my_policy_repo_id>")
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
 # Configure the dataset features
 action_features = hw_to_dataset_features(robot.action_features, "action")
@@ -544,7 +548,7 @@ dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/eval_<dataset_repo_id>",
+    repo_id=HF_DATASET_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -554,11 +558,17 @@ 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(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+)
+
 for episode_idx in range(NUM_EPISODES):
     log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
 
@@ -568,6 +578,8 @@ for episode_idx in range(NUM_EPISODES):
         events=events,
         fps=FPS,
         policy=policy,
+        preprocessor=preprocessor,
+        postprocessor=postprocessor,
         dataset=dataset,
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,

docs/source/il_sim.mdx

@@ -22,13 +22,38 @@ pip install -e ".[hilserl]"
 
 ## Teleoperate and Record a Dataset
 
-To use `gym_hil` with LeRobot, you need to use a configuration file. An example config file can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/env_config_gym_hil_il.json).
+To use `gym_hil` with LeRobot, you need to use a configuration file. An example config file can be found [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/sim_il/env_config.json).
 
-To teleoperate and collect a dataset, we need to modify this config file and you should add your `repo_id` here: `"repo_id": "il_gym",` and `"num_episodes": 30,` and make sure you set `mode` to `record`, "mode": "record".
+To teleoperate and collect a dataset, we need to modify this config file. Here's an example configuration for imitation learning data collection:
 
-If you do not have a Nvidia GPU also change `"device": "cuda"` parameter in the config file (for example to `mps` for MacOS).
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
+  "dataset": {
+    "repo_id": "your_username/il_gym",
+    "root": null,
+    "task": "pick_cube",
+    "num_episodes_to_record": 30,
+    "replay_episode": null,
+    "push_to_hub": true
+  },
+  "mode": "record",
+  "device": "cuda"
+}
+```
 
-By default the config file assumes you use a controller. To use your keyboard please change the envoirment specified at `"task"` in the config file and set it to `"PandaPickCubeKeyboard-v0"`.
+Key configuration points:
+
+- Set your `repo_id` in the `dataset` section: `"repo_id": "your_username/il_gym"`
+- Set `num_episodes_to_record: 30` to collect 30 demonstration episodes
+- Ensure `mode` is set to `"record"`
+- If you don't have an NVIDIA GPU, change `"device": "cuda"` to `"mps"` for macOS or `"cpu"`
+- To use keyboard instead of gamepad, change `"task"` to `"PandaPickCubeKeyboard-v0"`
 
 Then we can run this command to start:
 
@@ -36,14 +61,14 @@ Then we can run this command to start:
 <hfoption id="Linux">
 
 ```bash
-python -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
+python -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
 ```
 
 </hfoption>
 <hfoption id="MacOS">
 
 ```bash
-mjpython -m lerobot.scripts.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
+mjpython -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
 ```
 
 </hfoption>
@@ -140,9 +165,32 @@ huggingface-cli upload ${HF_USER}/il_sim_test${CKPT} \
 
 ## Evaluate your policy in Sim
 
-To evaluate your policy we have to use the config file that can be found [here](https://huggingface.co/datasets/aractingi/lerobot-example-config-files/blob/main/eval_config_gym_hil.json).
+To evaluate your policy we have to use a configuration file. An example can be found [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/sim_il/eval_config.json).
 
-Make sure to replace the `repo_id` with the dataset you trained on, for example `pepijn223/il_sim_dataset` and replace the `pretrained_policy_name_or_path` with your model id, for example `pepijn223/il_sim_model`
+Here's an example evaluation configuration:
+
+```json
+{
+  "env": {
+    "type": "gym_manipulator",
+    "name": "gym_hil",
+    "task": "PandaPickCubeGamepad-v0",
+    "fps": 10
+  },
+  "dataset": {
+    "repo_id": "your_username/il_sim_dataset",
+    "dataset_root": null,
+    "task": "pick_cube"
+  },
+  "pretrained_policy_name_or_path": "your_username/il_sim_model",
+  "device": "cuda"
+}
+```
+
+Make sure to replace:
+
+- `repo_id` with the dataset you trained on (e.g., `your_username/il_sim_dataset`)
+- `pretrained_policy_name_or_path` with your model ID (e.g., `your_username/il_sim_model`)
 
 Then you can run this command to visualize your trained policy
 
@@ -150,14 +198,14 @@ Then you can run this command to visualize your trained policy
 <hfoption id="Linux">
 
 ```bash
-python -m lerobot.scripts.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
+python -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
 ```
 
 </hfoption>
 <hfoption id="MacOS">
 
 ```bash
-mjpython -m lerobot.scripts.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
+mjpython -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
 ```
 
 </hfoption>

docs/source/implement_your_own_processor.mdx

@@ -0,0 +1,273 @@
+# Implement your own Robot Processor
+
+In this tutorial, you'll learn how to implement your own Robot Processor.
+It begins by exploring the need for a custom processor, then uses the `NormalizerProcessorStep` as the running example to explain how to implement, configure, and serialize a processor. Finally, it lists all helper processors that ship with LeRobot.
+
+## Why would you need a custom processor?
+
+In most cases, when reading raw data from sensors or when models output actions, you need to process this data to make it compatible with your target system. For example, a common need is normalizing data ranges to make them suitable for neural networks.
+
+LeRobot's `NormalizerProcessorStep` handles this crucial task:
+
+```python
+# Input: raw joint positions in [0, 180] degrees
+raw_action = torch.tensor([90.0, 45.0, 135.0])
+
+# After processing: normalized to [-1, 1] range for model training
+normalizer = NormalizerProcessorStep(features=features, norm_map=norm_map, stats=dataset_stats)
+normalized_result = normalizer(transition)
+# ...
+```
+
+Other common processing needs include:
+
+- **Device placement**: Moving tensors between CPU/GPU and converting data types
+- **Format conversion**: Transforming between different data structures
+- **Batching**: Adding/removing batch dimensions for model compatibility
+- **Safety constraints**: Applying limits to robot commands
+
+```python
+# Example pipeline combining multiple processors
+pipeline = PolicyProcessorPipeline([
+    RenameObservationsProcessorStep(rename_map={}),
+    AddBatchDimensionProcessorStep(),
+    NormalizerProcessorStep(features=features, stats=stats),
+    DeviceProcessorStep(device="cuda"),
+    # ...
+])
+```
+
+LeRobot provides a pipeline mechanism to implement sequences of processing steps for both input data and output actions, making it easy to compose these transformations in the right order for optimal performance.
+
+## How to implement your own processor?
+
+We'll use the `NormalizerProcessorStep` as our main example because it demonstrates essential processor patterns including state management, configuration serialization, and tensor handling that you'll commonly need.
+
+Prepare the sequence of processing steps necessary for your problem. A processor step is a class that implements the following methods:
+
+- `__call__`: implements the processing step for the input transition.
+- `get_config`: gets the configuration of the processor step.
+- `state_dict`: gets the state of the processor step.
+- `load_state_dict`: loads the state of the processor step.
+- `reset`: resets the state of the processor step.
+- `feature_contract`: displays the modification to the feature space during the processor step.
+
+### Implement the `__call__` method
+
+The `__call__` method is the core of your processor step. It takes an `EnvTransition` and returns a modified `EnvTransition`. Here's how the `NormalizerProcessorStep` works:
+
+```python
+@dataclass
+@ProcessorStepRegistry.register("normalizer_processor")
+class NormalizerProcessorStep(ProcessorStep):
+    """Normalize observations/actions using dataset statistics."""
+
+    features: dict[str, PolicyFeature]
+    norm_map: dict[FeatureType, NormalizationMode]
+    stats: dict[str, dict[str, Any]] | None = None
+    eps: float = 1e-8
+    _tensor_stats: dict = field(default_factory=dict, init=False, repr=False)
+
+    def __post_init__(self):
+        """Convert stats to tensors for efficient computation."""
+        self.stats = self.stats or {}
+        self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=torch.float32)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        new_transition = transition.copy()
+        # Normalize observations
+        # ...
+        # Normalize action
+        # ...
+        return new_transition
+
+```
+
+See the full implementation in `src/lerobot/processor/normalize_processor.py` for complete details.
+
+**Key principles:**
+
+- **Always use `transition.copy()`** to avoid side effects
+- **Handle both observations and actions** consistently
+- **Separate config from state**: `get_config()` returns JSON-serializable params, `state_dict()` returns tensors
+- **Convert stats to tensors** in `__post_init__()` for efficient computation
+
+### Configuration and State Management
+
+Processors support serialization through three methods that separate configuration from tensor state. The `NormalizerProcessorStep` demonstrates this perfectly - it carries dataset statistics (tensors) in its state, and hyperparameters in its config:
+
+```python
+# Continuing the NormalizerProcessorStep example...
+
+def get_config(self) -> dict[str, Any]:
+    """JSON-serializable configuration (no tensors)."""
+    return {
+        "eps": self.eps,
+        "features": {k: {"type": v.type.value, "shape": v.shape} for k, v in self.features.items()},
+        "norm_map": {ft.value: nm.value for ft, nm in self.norm_map.items()},
+        # ...
+    }
+
+def state_dict(self) -> dict[str, torch.Tensor]:
+    """Tensor state only (e.g., dataset statistics)."""
+    flat: dict[str, torch.Tensor] = {}
+    for key, sub in self._tensor_stats.items():
+        for stat_name, tensor in sub.items():
+            flat[f"{key}.{stat_name}"] = tensor.cpu()  # Always save to CPU
+    return flat
+
+def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+    """Restore tensor state at runtime."""
+    self._tensor_stats.clear()
+    for flat_key, tensor in state.items():
+        key, stat_name = flat_key.rsplit(".", 1)
+        # Load to processor's configured device
+        self._tensor_stats.setdefault(key, {})[stat_name] = tensor.to(
+            dtype=torch.float32, device=self.device
+        )
+        # ...
+```
+
+**Usage:**
+
+```python
+# Save (e.g., inside a policy)
+config = normalizer.get_config()
+tensors = normalizer.state_dict()
+
+# Restore (e.g., loading a pretrained policy)
+new_normalizer = NormalizerProcessorStep(**config)
+new_normalizer.load_state_dict(tensors)
+# Now new_normalizer has the same stats and configuration
+```
+
+### Transform features
+
+The `transform_features` method defines how your processor transforms feature names and shapes. This is crucial for policy configuration and debugging.
+
+For `NormalizerProcessorStep`, features are typically preserved unchanged since normalization doesn't alter keys or shapes:
+
+```python
+def transform_features(self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+    """Normalization preserves all feature definitions."""
+    return features  # No changes to feature structure
+    # ...
+```
+
+When your processor renames or reshapes data, implement this method to reflect the mapping for downstream components. For example, a simple rename processor:
+
+```python
+def transform_features(self, features: dict[str, PolicyFeature]) -> dict[str, PolicyFeature]:
+    # Simple renaming
+    if "pixels" in features:
+        features["observation.image"] = features.pop("pixels")
+
+    # Pattern-based renaming
+    for key in list(features.keys()):
+        if key.startswith("env_state."):
+            suffix = key[len("env_state."):]
+            features[f"observation.{suffix}"] = features.pop(key)
+            # ...
+
+    return features
+```
+
+**Key principles:**
+
+- Use `features.pop(old_key)` to remove and get the old feature
+- Use `features[new_key] = old_feature` to add the renamed feature
+- Always return the modified features dictionary
+- Document transformations clearly in the docstring
+
+### Using overrides
+
+You can override step parameters at load-time using `overrides`. This is handy for non-serializable objects or site-specific settings. It works both in policy factories and with `DataProcessorPipeline.from_pretrained(...)`.
+
+**Foundational model adaptation**: This is particularly useful when working with foundational pretrained policies where you rarely have access to the original training statistics. You can inject your own dataset statistics to adapt the normalizer to your specific robot or environment data.
+
+Example: during policy evaluation on the robot, override the device and rename map.
+Use this to run a policy trained on CUDA on a CPU-only robot, or to remap camera keys when the robot uses different names than the dataset.
+
+Direct usage with `from_pretrained`:
+
+```python
+from lerobot.processor import RobotProcessorPipeline
+
+# Load a foundational policy trained on diverse robot data
+# but adapt normalization to your specific robot/environment
+new_stats = LeRobotDataset(repo_id="username/my-dataset").meta.stats
+processor = RobotProcessorPipeline.from_pretrained(
+    "huggingface/foundational-robot-policy",  # Pretrained foundation model
+    overrides={
+        "normalizer_processor": {"stats": new_stats},     # Inject your robot's statistics
+        "device_processor": {"device": "cuda:0"},         # registry name for registered steps
+        "rename_processor": {"rename_map": robot_key_map}, # Map your robot's observation keys
+        # ...
+    },
+)
+```
+
+## Best Practices
+
+Based on analysis of all LeRobot processor implementations, here are the key patterns and practices:
+
+### 1. **Safe Data Handling**
+
+Always create copies of input data to avoid unintended side effects. Use `transition.copy()` and `observation.copy()` rather than modifying data in-place. This prevents your processor from accidentally affecting other components in the pipeline.
+
+Check for required data before processing and handle missing data gracefully. If your processor expects certain keys (like `"pixels"` for image processing), validate their presence first. For optional data, use safe access patterns like `transition.get()` and handle `None` values appropriately.
+
+When data validation fails, provide clear, actionable error messages that help users understand what went wrong and how to fix it.
+
+### 2. **Choose Appropriate Base Classes**
+
+LeRobot provides specialized base classes that reduce boilerplate code and ensure consistency. Use `ObservationProcessorStep` when you only need to modify observations, `ActionProcessorStep` for action-only processing, and `RobotActionProcessorStep` specifically for dictionary-based robot actions.
+
+Only inherit directly from `ProcessorStep` when you need full control over the entire transition or when processing multiple transition components simultaneously. The specialized base classes handle the transition management for you and provide type safety.
+
+### 3. **Registration and Naming**
+
+Register your processors with descriptive, namespaced names using `@ProcessorStepRegistry.register()`. Use organization prefixes like `"robotics_lab/safety_clipper"` or `"acme_corp/vision_enhancer"` to avoid naming conflicts. Avoid generic names like `"processor"` or `"step"` that could clash with other implementations.
+
+Good registration makes your processors discoverable and enables clean serialization/deserialization when saving and loading pipelines.
+
+### 4. **State Management Patterns**
+
+Distinguish between configuration parameters (JSON-serializable values) and internal state (tensors, buffers). Use dataclass fields with `init=False, repr=False` for internal state that shouldn't appear in the constructor or string representation.
+
+Implement the `reset()` method to clear internal state between episodes. This is crucial for stateful processors that accumulate data over time, like moving averages or temporal filters.
+
+Remember that `get_config()` should only return JSON-serializable configuration, while `state_dict()` handles tensor state separately.
+
+### 5. **Input Validation and Error Handling**
+
+Validate input types and shapes before processing. Check tensor properties like `dtype` and dimensions to ensure compatibility with your algorithms. For robot actions, verify that required pose components or joint values are present and within expected ranges.
+
+Use early returns for edge cases where no processing is needed. Provide clear, descriptive error messages that include the expected vs. actual data types or shapes. This makes debugging much easier for users.
+
+### 6. **Device and Dtype Awareness**
+
+Design your processors to automatically adapt to the device and dtype of input tensors. Internal tensors (like normalization statistics) should match the input tensor's device and dtype to ensure compatibility with multi-GPU training, mixed precision, and distributed setups.
+
+Implement a `to()` method that moves your processor's internal state to the specified device. Check device/dtype compatibility at runtime and automatically migrate internal state when needed. This pattern enables seamless operation across different hardware configurations without manual intervention.
+
+## Conclusion
+
+You now have all the tools to implement custom processors in LeRobot! The key steps are:
+
+1. **Define your processor** as a dataclass with the required methods (`__call__`, `get_config`, `state_dict`, `load_state_dict`, `reset`, `transform_features`)
+2. **Register it** using `@ProcessorStepRegistry.register("name")` for discoverability
+3. **Integrate it** into a `DataProcessorPipeline` with other processing steps
+4. **Use base classes** like `ObservationProcessorStep` when possible to reduce boilerplate
+5. **Implement device/dtype awareness** to support multi-GPU and mixed precision setups
+
+The processor system is designed to be modular and composable, allowing you to build complex data processing pipelines from simple, focused components. Whether you're preprocessing sensor data for training or post-processing model outputs for robot execution, custom processors give you the flexibility to handle any data transformation your robotics application requires.
+
+Key principles for robust processors:
+
+- **Device/dtype adaptation**: Internal tensors should match input tensors
+- **Clear error messages**: Help users understand what went wrong
+- **Base class usage**: Leverage specialized base classes to reduce boilerplate
+- **Feature contracts**: Declare data structure changes with `transform_features()`
+
+Start simple, test thoroughly, and ensure your processors work seamlessly across different hardware configurations!

docs/source/installation.mdx

@@ -91,7 +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)), `xarm` ([gym-xarm](https://github.com/huggingface/gym-xarm)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht))
+Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), 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.mdx).
+- LeRobot installed in your environment. Follow our [Installation Guide](./installation).
 
 ## Choose your motors
 
@@ -65,7 +65,7 @@ class MyCoolRobotConfig(RobotConfig):
 ```
 <!-- prettier-ignore-end -->
 
-[Cameras tutorial](./cameras.mdx) to understand how to detect and add your camera.
+[Cameras tutorial](./cameras) 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.
 
@@ -335,6 +335,134 @@ 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

@@ -0,0 +1,314 @@
+# Introduction to Processors
+
+In robotics, there's a fundamental mismatch between the data that robots and humans produce and what machine learning models expect.
+Robots output raw sensor data like camera images and joint positions that need normalization, batching, and device placement before models can process them.
+Language instructions from humans must be tokenized into numerical representations, and different robots use different coordinate systems that need standardization.
+
+The challenge extends to model outputs as well.
+Models might output end-effector positions while robots need joint-space commands, or teleoperators produce relative movements while robots expect absolute commands.
+Model predictions are often normalized and need conversion back to real-world scales.
+
+Cross-domain translation adds another layer of complexity.
+Training data from one robot setup needs adaptation for deployment on different hardware, models trained with specific camera configurations must work with new arrangements, and datasets with different naming conventions need harmonization.
+
+**That's where processors come in.** They serve as universal translators that bridge these gaps, ensuring seamless data flow from sensors to models to actuators.
+Processors handle all the preprocessing and postprocessing steps needed to convert raw environment data into model-ready inputs and vice versa.
+
+This means that your favorite policy can be used like this:
+
+```python
+import torch
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.your_policy import YourPolicy
+from lerobot.processor.pipeline import RobotProcessorPipeline, PolicyProcessorPipeline
+dataset = LeRobotDataset("hf_user/dataset", episodes=[0])
+sample = dataset[10]
+
+model = YourPolicy.from_pretrained(
+    "hf_user/model",
+)
+model.eval()
+model.to("cuda")
+preprocessor, postprocessor = make_pre_post_processors(model.config, pretrained_path="hf_user/model", dataset_stats=dataset.meta.stats)
+
+preprocessed_sample = preprocessor(sample)
+action = model.select_action(preprocessed_sample)
+postprocessed_action = postprocessor(action)
+```
+
+## What are Processors?
+
+In robotics, data comes in many forms: images from cameras, joint positions from sensors, text instructions from users, and more. Each type of data requires specific transformations before a model can use it effectively. Models need this data to be:
+
+- **Normalized**: Scaled to appropriate ranges for neural network processing
+- **Batched**: Organized with proper dimensions for batch processing
+- **Tokenized**: Text converted to numerical representations
+- **Device-placed**: Moved to the right hardware (CPU/GPU)
+- **Type-converted**: Cast to appropriate data types
+
+Processors handle these transformations through composable, reusable steps that can be chained together into pipelines. Think of them as a modular assembly line where each station performs a specific transformation on your data.
+
+## Core Concepts
+
+### EnvTransition: The Universal Data Container
+
+The `EnvTransition` is the fundamental data structure that flows through all processors.
+It's a typed dictionary that represents a complete robot-environment interaction:
+
+- **OBSERVATION**: All sensor data (images, states, proprioception)
+- **ACTION**: The action to execute or that was executed
+- **REWARD**: Reinforcement learning signal
+- **DONE/TRUNCATED**: Episode boundary indicators
+- **INFO**: Arbitrary metadata
+- **COMPLEMENTARY_DATA**: Task descriptions, indices, padding flags, inter-step data
+
+### ProcessorStep: The Building Block
+
+A `ProcessorStep` is a single transformation unit that processes transitions. It's an abstract base class with two required methods:
+
+```python
+from lerobot.processor import ProcessorStep, EnvTransition
+
+class MyProcessorStep(ProcessorStep):
+    """Example processor step - inherit and implement abstract methods."""
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Transform the transition - REQUIRED abstract method."""
+        # Your processing logic here
+        return transition
+
+    def transform_features(self, features):
+        """Declare how this step transforms feature shapes/types - REQUIRED abstract method."""
+        return features  # Most processors return features unchanged
+```
+
+`__call__` is the core of your processor step. It takes an `EnvTransition` and returns a modified `EnvTransition`.
+
+`transform_features` is used to declare how this step transforms feature shapes/types.
+
+### DataProcessorPipeline: The Generic Orchestrator
+
+The `DataProcessorPipeline[TInput, TOutput]` chains multiple `ProcessorStep` instances together:
+
+```python
+from lerobot.processor import RobotProcessorPipeline, PolicyProcessorPipeline
+
+# For robot hardware (unbatched data)
+robot_processor = RobotProcessorPipeline[RobotAction, RobotAction](
+    steps=[step1, step2, step3],
+    name="robot_pipeline"
+)
+
+# For model training/inference (batched data)
+policy_processor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+    steps=[step1, step2, step3],
+    name="policy_pipeline"
+)
+```
+
+## RobotProcessorPipeline vs PolicyProcessorPipeline
+
+The key distinction is in the data structures they handle:
+
+| Aspect          | RobotProcessorPipeline                       | PolicyProcessorPipeline                  |
+| --------------- | -------------------------------------------- | ---------------------------------------- |
+| **Input**       | `dict[str, Any]` - Individual robot values   | `dict[str, Any]` - Batched tensors       |
+| **Output**      | `dict[str, Any]` - Individual robot commands | `torch.Tensor` - Policy predictions      |
+| **Use Case**    | Real-time robot control                      | Model training/inference                 |
+| **Data Format** | Unbatched, heterogeneous                     | Batched, homogeneous                     |
+| **Examples**    | `{"joint_1": 0.5}`                           | `{"observation.state": tensor([[0.5]])}` |
+
+**Use `RobotProcessorPipeline`** for robot hardware interfaces:
+
+```python
+# Robot data structures: dict[str, Any] for observations and actions
+robot_obs: dict[str, Any] = {
+    "joint_1": 0.5,           # Individual joint values
+    "joint_2": -0.3,
+    "camera_0": image_array   # Raw camera data
+}
+
+robot_action: dict[str, Any] = {
+    "joint_1": 0.2,          # Target joint positions
+    "joint_2": 0.1,
+    "gripper": 0.8
+}
+```
+
+**Use `PolicyProcessorPipeline`** for model training and batch processing:
+
+```python
+# Policy data structures: batch dicts and tensors
+policy_batch: dict[str, Any] = {
+    "observation.state": torch.tensor([[0.5, -0.3]]),      # Batched states
+    "observation.images.camera0": torch.tensor(...),        # Batched images
+    "action": torch.tensor([[0.2, 0.1, 0.8]])              # Batched actions
+}
+
+policy_action: torch.Tensor = torch.tensor([[0.2, 0.1, 0.8]])  # Model output tensor
+```
+
+## Converter Functions
+
+LeRobot provides converter functions to bridge different data formats in `lerobot.processor.converters`. These functions handle the crucial translations between robot hardware data structures, policy model formats, and the internal `EnvTransition` representation that flows through processor pipelines.
+
+| Category                       | Function                      | Description                     |
+| ------------------------------ | ----------------------------- | ------------------------------- |
+| **Robot Hardware Converters**  | `robot_action_to_transition`  | Robot dict → EnvTransition      |
+|                                | `observation_to_transition`   | Robot obs → EnvTransition       |
+|                                | `transition_to_robot_action`  | EnvTransition → Robot dict      |
+| **Policy/Training Converters** | `batch_to_transition`         | Batch dict → EnvTransition      |
+|                                | `transition_to_batch`         | EnvTransition → Batch dict      |
+|                                | `policy_action_to_transition` | Policy tensor → EnvTransition   |
+|                                | `transition_to_policy_action` | EnvTransition → Policy tensor   |
+| **Utilities**                  | `create_transition`           | Build transitions with defaults |
+|                                | `identity_transition`         | Pass-through converter          |
+
+The key insight is that **robot hardware converters** work with individual values and dictionaries, while **policy/training converters** work with batched tensors and model outputs. The converter functions automatically handle the structural differences, so your processor steps can focus on the core transformations without worrying about data format compatibility.
+
+## Processor Examples
+
+The following examples demonstrate real-world processor configurations for policy training and inference.
+
+Here is an example processor for policy training and inference:
+
+```python
+# Training data preprocessing (optimized order for GPU performance)
+training_preprocessor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+    steps=[
+        RenameObservationsProcessorStep(rename_map={}),     # Standardize keys
+        AddBatchDimensionProcessorStep(),                   # Add batch dims
+        TokenizerProcessorStep(tokenizer_name="...", ...),  # Tokenize language
+        DeviceProcessorStep(device="cuda"),                 # Move to GPU first
+        NormalizerProcessorStep(features=..., stats=...),   # Normalize on GPU
+    ]
+)
+
+# Model output postprocessing
+training_postprocessor = PolicyProcessorPipeline[torch.Tensor, torch.Tensor](
+    steps=[
+        DeviceProcessorStep(device="cpu"),                  # Move to CPU
+        UnnormalizerProcessorStep(features=..., stats=...), # Denormalize
+    ]
+    to_transition=policy_action_to_transition,
+    to_output=transition_to_policy_action,
+)
+```
+
+### An interaction between a robot and a policy with processors
+
+The most common real-world scenario combines both pipeline types robot hardware generates observations that need policy processing, and policy outputs need robot-compatible postprocessing:
+
+```python
+# Real deployment: Robot sensors → Model → Robot commands
+with torch.no_grad():
+    while not done:
+        raw_obs = robot.get_observation()  # dict[str, Any]
+
+        # Add your robot observation to policy observation processor
+
+        policy_input = policy_preprocessor(raw_obs)  # Batched dict
+
+        policy_output = policy.select_action(policy_input)  # Policy tensor
+
+        policy_action = policy_postprocessor(policy_output)
+
+        # Add your robot action to policy action processor
+
+        robot.send_action(policy_action)
+```
+
+## Feature Contracts: Shape and Type Transformation
+
+Processors don't just transform data - they can also **change the data structure itself**. The `transform_features()` method declares these changes, which is crucial for dataset recording and policy creation.
+
+### Why Feature Contracts Matter
+
+When building datasets or policies, LeRobot needs to know:
+
+- **What data fields will exist** after processing
+- **What shapes and types** each field will have
+- **How to configure models** for the expected data structure
+
+```python
+# Example: A processor that adds velocity to observations
+class VelocityProcessor(ObservationProcessorStep):
+    def observation(self, obs):
+        new_obs = obs.copy()
+        if "observation.state" in obs:
+            # concatenate computed velocity field to the state
+            new_obs["observation.state"] = self._compute_velocity(obs["observation.state"])
+        return new_obs
+
+    def transform_features(self, features):
+        """Declare the new velocity field we're adding."""
+        state_feature = features[PipelineFeatureType.OBSERVATION].get("observation.state")
+        if state_feature:
+            double_shape = (state_feature.shape[0] * 2,) if state_feature.shape else (2,)
+            features[PipelineFeatureType.OBSERVATION]["observation.state"] = PolicyFeature(
+                type=FeatureType.STATE, shape=double_shape
+            )
+        return features
+```
+
+### Feature Specification Functions
+
+`create_initial_features()` and `aggregate_pipeline_dataset_features()` solve a critical dataset creation problem: determining the exact final data structure before any data is processed.
+Since processor pipelines can add new features (like velocity fields), change tensor shapes (like cropping images), or rename keys, datasets need to know the complete output specification upfront to allocate proper storage and define schemas.
+These functions work together by starting with robot hardware specifications (`create_initial_features()`) then simulating the entire pipeline transformation (`aggregate_pipeline_dataset_features()`) to compute the final feature dictionary that gets passed to `LeRobotDataset.create()`, ensuring perfect alignment between what processors output and what datasets expect to store.
+
+```python
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features
+
+# Start with robot's raw features
+initial_features = create_initial_features(
+    observation=robot.observation_features,  # {"joint_1.pos": float, "camera_0": (480,640,3)}
+    action=robot.action_features            # {"joint_1.pos": float, "gripper.pos": float}
+)
+
+# Apply processor pipeline to compute final features
+final_features = aggregate_pipeline_dataset_features(
+    pipeline=my_processor_pipeline,
+    initial_features=initial_features,
+    use_videos=True
+)
+
+# Use for dataset creation
+dataset = LeRobotDataset.create(
+    repo_id="my_dataset",
+    features=final_features,  # Knows exactly what data to expect
+    ...
+)
+```
+
+## Common Processor Steps
+
+LeRobot provides many registered processor steps. Here are the most commonly used core processors:
+
+### Essential Processors
+
+- **`normalizer_processor`**: Normalize observations/actions using dataset statistics (mean/std or min/max)
+- **`device_processor`**: Move tensors to CPU/GPU with optional dtype conversion
+- **`to_batch_processor`**: Add batch dimensions to transitions for model compatibility
+- **`rename_observations_processor`**: Rename observation keys using mapping dictionaries
+- **`tokenizer_processor`**: Tokenize natural language task descriptions into tokens and attention masks
+
+### 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
+
+## Summary
+
+Processors solve the data translation problem in robotics by providing:
+
+- **Modular transformations**: Composable, reusable processing steps
+- **Type safety**: Generic pipelines with compile-time checking
+- **Performance optimization**: GPU-accelerated operations
+- **Robot/Policy distinction**: Separate pipelines for different data structures
+- **Comprehensive ecosystem**: 30+ registered processors for common tasks
+
+The key insight: `RobotProcessorPipeline` handles unbatched robot hardware data, while `PolicyProcessorPipeline` handles batched model data. Choose the right tool for your data structure!

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](./getting_started_real_world_robot)
+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)
 
 > [!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](./getting_started_real_world_robot)
+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)
 
 ## Evaluate your policy
 

docs/source/lerobot-dataset-v3.mdx

@@ -8,6 +8,7 @@ This docs will guide you to:
 - Record a dataset and push it to the Hub
 - Load datasets for training with `LeRobotDataset`
 - Stream datasets without downloading using `StreamingLeRobotDataset`
+- Apply image transforms for data augmentation during training
 - Migrate existing `v2.1` datasets to `v3.0`
 
 ## What’s new in `v3`
@@ -150,6 +151,117 @@ dataset = StreamingLeRobotDataset(repo_id)  # streams directly from the Hub
   </figure>
 </div>
 
+## Image transforms
+
+Image transforms are data augmentations applied to camera frames during training to improve model robustness and generalization. LeRobot supports various transforms including brightness, contrast, saturation, hue, and sharpness adjustments.
+
+### Using transforms during dataset creation/recording
+
+Currently, transforms are applied during **training time only**, not during recording. When you create or record a dataset, the raw images are stored without transforms. This allows you to experiment with different augmentations later without re-recording data.
+
+### Adding transforms to existing datasets (API)
+
+Use the `image_transforms` parameter when loading a dataset for training:
+
+```python
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.transforms import ImageTransforms, ImageTransformsConfig, ImageTransformConfig
+
+# Option 1: Use default transform configuration (disabled by default)
+transforms_config = ImageTransformsConfig(
+    enable=True,  # Enable transforms
+    max_num_transforms=3,  # Apply up to 3 transforms per frame
+    random_order=False,  # Apply in standard order
+)
+transforms = ImageTransforms(transforms_config)
+
+dataset = LeRobotDataset(
+    repo_id="your-username/your-dataset",
+    image_transforms=transforms
+)
+
+# Option 2: Create custom transform configuration
+custom_transforms_config = ImageTransformsConfig(
+    enable=True,
+    max_num_transforms=2,
+    random_order=True,
+    tfs={
+        "brightness": ImageTransformConfig(
+            weight=1.0,
+            type="ColorJitter",
+            kwargs={"brightness": (0.7, 1.3)}  # Adjust brightness range
+        ),
+        "contrast": ImageTransformConfig(
+            weight=2.0,  # Higher weight = more likely to be selected
+            type="ColorJitter",
+            kwargs={"contrast": (0.8, 1.2)}
+        ),
+        "sharpness": ImageTransformConfig(
+            weight=0.5,  # Lower weight = less likely to be selected
+            type="SharpnessJitter",
+            kwargs={"sharpness": (0.3, 2.0)}
+        ),
+    }
+)
+
+dataset = LeRobotDataset(
+    repo_id="your-username/your-dataset",
+    image_transforms=ImageTransforms(custom_transforms_config)
+)
+
+# Option 3: Use pure torchvision transforms
+from torchvision.transforms import v2
+
+torchvision_transforms = v2.Compose([
+    v2.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
+    v2.GaussianBlur(kernel_size=3, sigma=(0.1, 2.0)),
+])
+
+dataset = LeRobotDataset(
+    repo_id="your-username/your-dataset",
+    image_transforms=torchvision_transforms
+)
+```
+
+### Available transform types
+
+LeRobot provides several transform types:
+
+- **`ColorJitter`**: Adjusts brightness, contrast, saturation, and hue
+- **`SharpnessJitter`**: Randomly adjusts image sharpness
+- **`Identity`**: No transformation (useful for testing)
+
+You can also use any `torchvision.transforms.v2` transform by passing it directly to the `image_transforms` parameter.
+
+### Configuration options
+
+- **`enable`**: Enable/disable transforms (default: `False`)
+- **`max_num_transforms`**: Maximum number of transforms applied per frame (default: `3`)
+- **`random_order`**: Apply transforms in random order vs. standard order (default: `False`)
+- **`weight`**: Sampling probability for each transform (higher = more likely, if sum of weights is not 1, they will be normalized)
+- **`kwargs`**: Transform-specific parameters (e.g., brightness range)
+
+### Visualizing transforms
+
+Use the visualization script to preview how transforms affect your data:
+
+```bash
+lerobot-imgtransform-viz \
+  --repo-id=your-username/your-dataset \
+  --output-dir=./transform_examples \
+  --n-examples=5
+```
+
+This saves example images showing the effect of each transform, helping you tune parameters.
+
+### Best practices
+
+- **Start conservative**: Begin with small ranges (e.g., brightness 0.9-1.1) and increase gradually
+- **Test first**: Use the visualization script to ensure transforms look reasonable
+- **Monitor training**: Strong augmentations can hurt performance if too aggressive
+- **Match your domain**: If your robot operates in varying lighting, use brightness/contrast transforms
+- **Combine wisely**: Using too many transforms simultaneously can make training unstable
+
 ## Migrate `v2.1` → `v3.0`
 
 A converter aggregates per‑episode files into larger shards and writes episode offsets/metadata. Convert your dataset using the instructions below.
@@ -167,3 +279,36 @@ 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

@@ -0,0 +1,166 @@
+# LIBERO
+
+**LIBERO** is a benchmark designed to study **lifelong robot learning**. The idea is that robots won’t just be pretrained once in a factory, they’ll need to keep learning and adapting with their human users over time. This ongoing adaptation is called **lifelong learning in decision making (LLDM)**, and it’s a key step toward building robots that become truly personalized helpers.
+
+- 📄 [LIBERO paper](https://arxiv.org/abs/2306.03310)
+- 💻 [Original LIBERO repo](https://github.com/Lifelong-Robot-Learning/LIBERO)
+
+To make progress on this challenge, LIBERO provides a set of standardized tasks that focus on **knowledge transfer**: how well a robot can apply what it has already learned to new situations. By evaluating on LIBERO, different algorithms can be compared fairly and researchers can build on each other’s work.
+
+LIBERO includes **five task suites**:
+
+- **LIBERO-Spatial (`libero_spatial`)** – tasks that require reasoning about spatial relations.
+- **LIBERO-Object (`libero_object`)** – tasks centered on manipulating different objects.
+- **LIBERO-Goal (`libero_goal`)** – goal-conditioned tasks where the robot must adapt to changing targets.
+- **LIBERO-90 (`libero_90`)** – 90 short-horizon tasks from the LIBERO-100 collection.
+- **LIBERO-Long (`libero_10`)** – 10 long-horizon tasks from the LIBERO-100 collection.
+
+Together, these suites cover **130 tasks**, ranging from simple object manipulations to complex multi-step scenarios. LIBERO is meant to grow over time, and to serve as a shared benchmark where the community can test and improve lifelong learning algorithms.
+
+![An overview of the LIBERO benchmark](https://libero-project.github.io/assets/img/libero/fig1.png)
+
+## Evaluating with LIBERO
+
+At **LeRobot**, we ported [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) into our framework and used it mainly to **evaluate [SmolVLA](https://huggingface.co/docs/lerobot/en/smolvla)**, our lightweight Vision-Language-Action model.
+
+LIBERO is now part of our **multi-eval supported simulation**, meaning you can benchmark your policies either on a **single suite of tasks** or across **multiple suites at once** with just a flag.
+
+To Install LIBERO, after following LeRobot official instructions, just do:
+`pip install -e ".[libero]"`
+
+### Single-suite evaluation
+
+Evaluate a policy on one LIBERO suite:
+
+```bash
+lerobot-eval \
+  --policy.path="your-policy-id" \
+  --env.type=libero \
+  --env.task=libero_object \
+  --eval.batch_size=2 \
+  --eval.n_episodes=3
+```
+
+- `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
+- `--eval.batch_size` controls how many environments run in parallel.
+- `--eval.n_episodes` sets how many episodes to run in total.
+
+---
+
+### Multi-suite evaluation
+
+Benchmark a policy across multiple suites at once:
+
+```bash
+lerobot-eval \
+  --policy.path="your-policy-id" \
+  --env.type=libero \
+  --env.task=libero_object,libero_spatial \
+  --eval.batch_size=1 \
+  --eval.n_episodes=2
+```
+
+- Pass a comma-separated list to `--env.task` for multi-suite evaluation.
+
+### Policy inputs and outputs
+
+When using LIBERO through LeRobot, policies interact with the environment via **observations** and **actions**:
+
+- **Observations**
+  - `observation.state` – proprioceptive features (agent state).
+  - `observation.images.image` – main camera view (`agentview_image`).
+  - `observation.images.image2` – wrist camera view (`robot0_eye_in_hand_image`).
+
+  ⚠️ **Note:** LeRobot enforces the `.images.*` prefix for any multi-modal visual features. Always ensure that your policy config `input_features` use the same naming keys, and that your dataset metadata keys follow this convention during evaluation.
+  If your data contains different keys, you must rename the observations to match what the policy expects, since naming keys are encoded inside the normalization statistics layer.
+  This will be fixed with the upcoming Pipeline PR.
+
+- **Actions**
+  - Continuous control values in a `Box(-1, 1, shape=(7,))` space.
+
+We also provide a notebook for quick testing:
+Training with LIBERO
+
+## Training with LIBERO
+
+When training on LIBERO tasks, make sure your dataset parquet and metadata keys follow the LeRobot convention.
+
+The environment expects:
+
+- `observation.state` → 8-dim agent state
+- `observation.images.image` → main camera (`agentview_image`)
+- `observation.images.image2` → wrist camera (`robot0_eye_in_hand_image`)
+
+⚠️ Cleaning the dataset upfront is **cleaner and more efficient** than remapping keys inside the code.
+To avoid potential mismatches and key errors, we provide a **preprocessed LIBERO dataset** that is fully compatible with the current LeRobot codebase and requires no additional manipulation:
+👉 [HuggingFaceVLA/libero](https://huggingface.co/datasets/HuggingFaceVLA/libero)
+
+For reference, here is the **original dataset** published by Physical Intelligence:
+👉 [physical-intelligence/libero](https://huggingface.co/datasets/physical-intelligence/libero)
+
+---
+
+### Example training command
+
+```bash
+lerobot-train \
+  --policy.type=smolvla \
+  --policy.repo_id=${HF_USER}/libero-test \
+  --policy.load_vlm_weights=true \
+  --dataset.repo_id=HuggingFaceVLA/libero \
+  --env.type=libero \
+  --env.task=libero_10 \
+  --output_dir=./outputs/ \
+  --steps=100000 \
+  --batch_size=4 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1 \
+  --eval_freq=1000 \
+```
+
+---
+
+### Note on rendering
+
+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

@@ -0,0 +1,80 @@
+# 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/phone_teleop.mdx

@@ -0,0 +1,191 @@
+# Phone
+
+Use your phone (iOS or Android) to control your robot.
+
+**In this guide you'll learn:**
+
+- How to connect an iOS/Android phone
+- How phone pose is mapped to robot end‑effector (EE) targets
+- How to tweak safety limits, gripper control, and IK settings
+
+To use phone to control your robot, install the relevant dependencies with:
+
+```bash
+pip install lerobot[phone]
+```
+
+## Get started
+
+### Supported platforms
+
+- iOS: Uses the HEBI Mobile I/O app (ARKit pose + buttons). Download the app first, open it and the examples will discover it on your network and stream the phone pose and inputs.
+- Android: Uses the `teleop` package (WebXR). When you start the Python process, it prints a local URL. Open the link on your phone, tap Start, then use Move to stream pose.
+
+Links:
+
+- Android WebXR library: [`teleop` on PyPI](https://pypi.org/project/teleop/)
+- iOS app: [HEBI Mobile I/O](https://docs.hebi.us/tools.html#mobile-io)
+
+### Phone orientation and controls
+
+- Orientation: hold the phone with the screen facing up and the top edge pointing in the same direction as the robot gripper. This ensures calibration aligns the phone’s frame with the robot frame so motion feels natural, see the image below for reference.
+- Enable/disable:
+  - iOS: Hold `B1` to enable teleoperation, release to stop. The first press captures a reference pose.
+  - Android: Press and hold the `Move` button, release to stop. The first press captures a reference pose.
+- Gripper control:
+  - iOS: Analog input `A3` controls the gripper as velocity input.
+  - Android: Buttons `A` and `B` act like increment/decrement (A opens, B closes). You can tune velocity in the `GripperVelocityToJoint` step.
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/phone_teleop.webp" alt="Phone teleop orientation" title="Phone teleop orientation" width="40%">
+
+### Step 1: Choose the platform
+
+Modify the examples to use `PhoneOS.IOS` or `PhoneOS.ANDROID` in `PhoneConfig`. The API is identical across platforms, only the input source differs. All examples are under `examples/` and have `phone_so100_*.py` variants.
+
+Teleoperation example:
+
+```36:43:examples/phone_so100_teleop.py
+from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
+
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+teleop_device = Phone(teleop_config)
+```
+
+### Step 2: Connect and calibrate
+
+When `Phone(teleop_config)` is created and `connect()` is called, calibration is prompted automatically. Hold the phone in the orientation described above, then:
+
+- iOS: press and hold `B1` to capture the reference pose.
+- Android: press `Move` button on the WebXR page to capture the reference pose.
+
+Why calibrate? We capture the current pose so subsequent poses are expressed in a robot aligned frame. When you again press the button to enable control, the position is recaptured to avoid drift when your phone is repositioned while it was disabled.
+
+### Step 3: Run an example
+
+Run on of the examples scripts to teleoperate, record a dataset, replay a dataset or evaluate a policy.
+
+All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
+
+Additionally you need to **copy the urdf of the robot to the examples folder**. For the examples in this tutorial (Using SO100/SO101) it is highly recommended to use the urdf in the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf)
+
+- Run this example to teleoperate:
+
+  ```bash
+  python examples/phone_to_so100/teleoperate.py
+  ```
+
+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.
+
+- Run this example to record a dataset, which saves absolute end effector observations and actions:
+
+  ```bash
+  python examples/phone_to_so100/record.py
+  ```
+
+- Run this example to replay recorded episodes:
+
+  ```bash
+  python examples/phone_to_so100/replay.py
+  ```
+
+- Run this example to evaluate a pretrained policy:
+
+  ```bash
+  python examples/phone_to_so100/evaluate.py
+  ```
+
+### Important pipeline steps and options
+
+- Kinematics are used in multiple steps. We use [Placo](https://github.com/Rhoban/placo) which is a wrapper around Pinocchio for handling our kinematics. We construct the kinematics object by passing the robot's URDF and target frame. We set `target_frame_name` to the gripper frame.
+
+  ```examples/phone_to_so100/teleoperate.py
+  kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+  )
+
+  ```
+
+- The `MapPhoneActionToRobotAction` step converts the calibrated phone pose and inputs into target deltas and gripper commands, below is shown what the step outputs.
+
+  ```src/lerobot/teleoperators/phone/phone_processor.py
+  action["enabled"] = enabled
+        action["target_x"] = -pos[1] if enabled else 0.0
+        action["target_y"] = pos[0] if enabled else 0.0
+        action["target_z"] = pos[2] if enabled else 0.0
+        action["target_wx"] = rotvec[1] if enabled else 0.0
+        action["target_wy"] = rotvec[0] if enabled else 0.0
+        action["target_wz"] = -rotvec[2] if enabled else 0.0
+        action["gripper_vel"] = gripper_vel  # Still send gripper action when disabled
+  ```
+
+- The `EEReferenceAndDelta` step converts target deltas to an absolute desired EE pose, storing a reference on enable, the `end_effector_step_sizes` are the step sizes for the EE pose and can be modified to change the motion speed.
+
+  ```examples/phone_to_so100/teleoperate.py
+  EEReferenceAndDelta(
+      kinematics=kinematics_solver,
+      end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+      motor_names=list(robot.bus.motors.keys()),
+      use_latched_reference=True,
+  ),
+  ```
+
+- 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.
+
+  ```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,
+  )
+  ```
+
+- The `GripperVelocityToJoint` step turns a velocity‑like gripper input into absolute gripper position using the current measured state. The `speed_factor` is the factor by which the velocity is multiplied.
+
+  ```examples/phone_to_so100/teleoperate.py
+  GripperVelocityToJoint(speed_factor=20.0)
+  ```
+
+#### Different IK initial guesses
+
+We use different IK initial guesses in the kinematic steps. As initial guess either the current measured joints or the previous IK solution is used.
+
+- Closed loop (used in record/eval): sets `initial_guess_current_joints=True` so IK starts from the measured joints each frame.
+
+  ```examples/phone_to_so100/record.py
+  InverseKinematicsEEToJoints(
+      kinematics=kinematics_solver,
+      motor_names=list(robot.bus.motors.keys()),
+      initial_guess_current_joints=True,  # closed loop
+  )
+  ```
+
+- Open loop (used in replay): sets `initial_guess_current_joints=False` so IK continues from the previous IK solution rather than the measured state. This preserves action stability when we replay without feedback.
+
+  ```examples/phone_to_so100/replay.py
+  InverseKinematicsEEToJoints(
+      kinematics=kinematics_solver,
+      motor_names=list(robot.bus.motors.keys()),
+      initial_guess_current_joints=False,  # open loop
+  )
+  ```
+
+### Pipeline steps explained
+
+- MapPhoneActionToRobotAction: converts calibrated phone pose and inputs into target deltas and a gripper command. Motion is gated by an enable signal (B1 on iOS, Move on Android).
+- EEReferenceAndDelta: latches a reference EE pose on enable and combines it with target deltas to produce an absolute desired EE pose each frame. When disabled, it keeps sending the last commanded pose.
+- EEBoundsAndSafety: clamps the EE pose to a workspace and rate‑limits jumps for safety. Also declares `action.ee.*` features.
+- InverseKinematicsEEToJoints: turns an EE pose into joint positions with IK. `initial_guess_current_joints=True` is recommended for closed‑loop control; set `False` for open‑loop replay for stability.
+- GripperVelocityToJoint: integrates a velocity‑like gripper input into an absolute gripper position using the current measured state.
+- ForwardKinematicsJointsToEE: computes `observation.state.ee.*` from observed joints for logging and training on EE state.
+
+### Troubleshooting
+
+- iOS not discovered: ensure HEBI Mobile I/O is open and your laptop/phone are on the same network.
+- Android URL not reachable: check local you used `https` instead of `http`, use the exact IP printed by the script and allow your browser to enter and ignore the certificate issue.
+- Motion feels inverted: adjust the sign flips in `MapPhoneActionToRobotAction` or swap axes to match your setup.

docs/source/pi0.mdx

@@ -0,0 +1,79 @@
+# 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

@@ -0,0 +1,107 @@
+# 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

@@ -0,0 +1,151 @@
+# Processors for Robots and Teleoperators
+
+This guide shows how to build and modify processing pipelines that connect teleoperators (e.g., phone) to robots and datasets. Pipelines standardize conversions between different action/observation spaces so you can swap teleops and robots without rewriting glue code.
+
+We use the Phone to SO‑100 follower examples for concreteness, but the same patterns apply to other robots.
+
+**What you'll learn**
+
+- Absolute vs. relative EE control: What each means, trade‑offs, and how to choose for your task.
+- Three-pipeline pattern: How to map teleop actions → dataset actions → robot commands, and robot observations → dataset observations.
+- Adapters (`to_transition` / `to_output`): How these convert raw dicts to `EnvTransition` and back to reduce boilerplate.
+- Dataset feature contracts: How steps declare features via `transform_features(...)`, and how to aggregate/merge them for recording.
+- Choosing a representation: When to store joints, absolute EE poses, or relative EE deltas—and how that affects training.
+- Pipeline customization guidance: How to swap robots/URDFs safely and tune bounds, step sizes, and options like IK initialization.
+
+### Absolute vs relative EE control
+
+The examples in this guide use absolute end effector (EE) poses because they are easy to reason about. In practice, relative EE deltas or joint position are often preferred as learning features.
+
+With processors, you choose the learning features you want to use for your policy. This could be joints positions/velocities, absolute EE, or relative EE positions. You can also choose to store other features, such as joint torques, motor currents, etc.
+
+## Three pipelines
+
+We often compose three pipelines. Depending on your setup, some can be empty if action and observation spaces already match.
+Each of these pipelines handle different conversions between different action and observation spaces. Below is a quick explanation of each pipeline.
+
+1. Pipeline 1: Teleop action space → dataset action space (phone pose → EE targets)
+2. Pipeline 2: Dataset action space → robot command space (EE targets → joints)
+3. Pipeline 3: Robot observation space → dataset observation space (joints → EE pose)
+
+Below is an example of the three pipelines that we use in the phone to SO-100 follower examples:
+
+```69:90:examples/phone_so100_record.py
+phone_to_robot_ee_pose_processor = RobotProcessorPipeline[RobotAction, RobotAction]( # teleop -> dataset action
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        EEReferenceAndDelta(
+            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,
+        ),
+        GripperVelocityToJoint(),
+    ],
+    to_transition=robot_action_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+robot_ee_to_joints_processor = RobotProcessorPipeline[RobotAction, RobotAction]( # dataset action -> robot
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()), initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation]( # robot obs -> dataset obs
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+```
+
+## Why to_transition / to_output
+
+To convert from robot/teleoperator to pipeline and back, we use the `to_transition` and `to_output` pipeline adapters.
+They standardize conversions to reduce boilerplate code, and form the bridge between the robot and teleoperators raw dictionaries and the pipeline’s `EnvTransition` format.
+In the phone to SO-100 follower examples we use the following adapters:
+
+- `robot_action_to_transition`: transforms the teleop action dict to a pipeline transition.
+- `transition_to_robot_action`: transforms the pipeline transition to a robot action dict.
+- `observation_to_transition`: transforms the robot observation dict to a pipeline transition.
+- `transition_to_observation`: transforms the pipeline transition to a observation dict.
+
+Checkout [src/lerobot/processor/converters.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/converters.py) for more details.
+
+## Dataset feature contracts
+
+Dataset features are determined by the keys saved in the dataset. Each step can declare what features it modifies in a contract called `transform_features(...)`. Once you build a processor, the processor can then aggregate all of these features with `aggregate_pipeline_dataset_features()` and merge multiple feature dicts with `combine_feature_dicts(...)`.
+
+Below is and example of how we declare features with the `transform_features` method in the phone to SO-100 follower examples:
+
+```src/lerobot/robots/so100_follower/robot_kinematic_processor.py
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        # We only use the ee pose in the dataset, so we don't need the joint positions
+        for n in self.motor_names:
+            features[PipelineFeatureType.ACTION].pop(f"{n}.pos", None)
+        # We specify the dataset features of this step that we want to be stored in the dataset
+        for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]:
+            features[PipelineFeatureType.ACTION][f"ee.{k}"] = PolicyFeature(
+                type=FeatureType.STATE, shape=(1,)
+            )
+        return features
+```
+
+Here we declare what PolicyFeatures we modify in this step, so we know what features we can expect when we run the processor. These features can then be aggregated and used to create the dataset features.
+
+Below is an example of how we aggregate and merge features in the phone to SO-100 record example:
+
+```121:145:examples/phone_so100_record.py
+features=combine_feature_dicts(
+        # Run the feature contract of the pipelines
+        # This tells you how the features would look like after the pipeline steps
+        aggregate_pipeline_dataset_features(
+            pipeline=phone_to_robot_ee_pose_processor,
+            initial_features=create_initial_features(action=phone.action_features), # <- Action features we can expect, these come from our teleop device (phone) and action processor
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_joints_to_ee_pose,
+            initial_features=create_initial_features(observation=robot.observation_features), # <- Observation features we can expect, these come from our robot and observation processor
+            use_videos=True,
+            patterns=["observation.state.ee"], # <- Here you could optionally filter the features we want to store in the dataset, with a specific pattern
+
+        ),
+    ),
+```
+
+How it works:
+
+- `aggregate_pipeline_dataset_features(...)`: applies `transform_features` across the pipeline and filters by patterns (images included when `use_videos=True`, and state features included when `patterns` is specified).
+- `combine_feature_dicts(...)`: combine multiple feature dicts.
+- Recording with `record_loop(...)` uses `build_dataset_frame(...)` to build frames consistent with `dataset.features` before we call `add_frame(...)` to add the frame to the dataset.
+
+## Guidance when customizing robot pipelines
+
+You can store any of the following features as your action/observation space:
+
+- Joint positions
+- Absolute EE poses
+- Relative EE deltas
+- Other features: joint velocity, torques, etc.
+
+Pick what you want to use for your policy action and observation space and configure/modify the pipelines and steps accordingly.
+
+### Different robots
+
+- You can easily reuse pipelines, for example to use another robot with phone teleop, modify the examples and swap the robot `RobotKinematics` (URDF) and `motor_names` to use your own robot with Phone teleop. Additionally you should ensure `target_frame_name` points to your gripper/wrist.
+
+### Safety first
+
+- When changing pipelines, start with tight bounds, implement safety steps when working with real robots.
+- Its advised to start with simulation first and then move to real robots.
+
+Thats it! We hope this guide helps you get started with customizing your robot pipelines, If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.

docs/source/smolvla.mdx

@@ -1,4 +1,4 @@
-# Finetune SmolVLA
+# 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](https://huggingface.co/docs/lerobot/getting_started_real_world_robot#record-a-dataset)
+We recommend recording ~50 episodes of your task as a starting point. Follow our guide to get started: [Recording a Dataset](./il_robots)
 
 <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](./getting_started_real_world_robot#record-a-dataset).
+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).
 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](./getting_started_real_world_robot)
+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)
 
 > [!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](./getting_started_real_world_robot)
+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)
 
 > [!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

@@ -0,0 +1,102 @@
+# 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/2_evaluate_pretrained_policy.py

@@ -1,139 +0,0 @@
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-This script demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
-training outputs directory. In the latter case, you might want to run examples/3_train_policy.py first.
-
-It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
-```bash
-pip install -e ".[pusht]"
-```
-"""
-
-from pathlib import Path
-
-import gym_pusht  # noqa: F401
-import gymnasium as gym
-import imageio
-import numpy
-import torch
-
-from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
-
-# Create a directory to store the video of the evaluation
-output_directory = Path("outputs/eval/example_pusht_diffusion")
-output_directory.mkdir(parents=True, exist_ok=True)
-
-# Select your device
-device = "cuda"
-
-# Provide the [hugging face repo id](https://huggingface.co/lerobot/diffusion_pusht):
-pretrained_policy_path = "lerobot/diffusion_pusht"
-# OR a path to a local outputs/train folder.
-# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
-
-policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
-
-# Initialize evaluation environment to render two observation types:
-# an image of the scene and state/position of the agent. The environment
-# also automatically stops running after 300 interactions/steps.
-env = gym.make(
-    "gym_pusht/PushT-v0",
-    obs_type="pixels_agent_pos",
-    max_episode_steps=300,
-)
-
-# We can verify that the shapes of the features expected by the policy match the ones from the observations
-# produced by the environment
-print(policy.config.input_features)
-print(env.observation_space)
-
-# Similarly, we can check that the actions produced by the policy will match the actions expected by the
-# environment
-print(policy.config.output_features)
-print(env.action_space)
-
-# Reset the policy and environments to prepare for rollout
-policy.reset()
-numpy_observation, info = env.reset(seed=42)
-
-# Prepare to collect every rewards and all the frames of the episode,
-# from initial state to final state.
-rewards = []
-frames = []
-
-# Render frame of the initial state
-frames.append(env.render())
-
-step = 0
-done = False
-while not done:
-    # Prepare observation for the policy running in Pytorch
-    state = torch.from_numpy(numpy_observation["agent_pos"])
-    image = torch.from_numpy(numpy_observation["pixels"])
-
-    # Convert to float32 with image from channel first in [0,255]
-    # to channel last in [0,1]
-    state = state.to(torch.float32)
-    image = image.to(torch.float32) / 255
-    image = image.permute(2, 0, 1)
-
-    # Send data tensors from CPU to GPU
-    state = state.to(device, non_blocking=True)
-    image = image.to(device, non_blocking=True)
-
-    # Add extra (empty) batch dimension, required to forward the policy
-    state = state.unsqueeze(0)
-    image = image.unsqueeze(0)
-
-    # Create the policy input dictionary
-    observation = {
-        "observation.state": state,
-        "observation.image": image,
-    }
-
-    # Predict the next action with respect to the current observation
-    with torch.inference_mode():
-        action = policy.select_action(observation)
-
-    # Prepare the action for the environment
-    numpy_action = action.squeeze(0).to("cpu").numpy()
-
-    # Step through the environment and receive a new observation
-    numpy_observation, reward, terminated, truncated, info = env.step(numpy_action)
-    print(f"{step=} {reward=} {terminated=}")
-
-    # Keep track of all the rewards and frames
-    rewards.append(reward)
-    frames.append(env.render())
-
-    # The rollout is considered done when the success state is reached (i.e. terminated is True),
-    # or the maximum number of iterations is reached (i.e. truncated is True)
-    done = terminated | truncated | done
-    step += 1
-
-if terminated:
-    print("Success!")
-else:
-    print("Failure!")
-
-# Get the speed of environment (i.e. its number of frames per second).
-fps = env.metadata["render_fps"]
-
-# Encode all frames into a mp4 video.
-video_path = output_directory / "rollout.mp4"
-imageio.mimsave(str(video_path), numpy.stack(frames), fps=fps)
-
-print(f"Video of the evaluation is available in '{video_path}'.")

examples/4_train_policy_with_script.md

@@ -1,311 +0,0 @@
-This tutorial will explain the training script, how to use it, and particularly how to configure everything needed for the training run.
-
-> **Note:** The following assumes you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
-
-## The training script
-
-LeRobot offers a training script at [`lerobot/scripts/train.py`](../src/lerobot/scripts/train.py). At a high level it does the following:
-
-- Initialize/load a configuration for the following steps using.
-- Instantiates a dataset.
-- (Optional) Instantiates a simulation environment corresponding to that dataset.
-- Instantiates a policy.
-- Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
-
-## Overview of the configuration system
-
-In the training script, the main function `train` expects a `TrainPipelineConfig` object:
-
-<!-- prettier-ignore-start -->
-```python
-# train.py
-@parser.wrap()
-def train(cfg: TrainPipelineConfig):
-```
-<!-- prettier-ignore-end -->
-
-You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../src/lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
-
-When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated to this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
-
-Let's have a look at a simplified example. Amongst other attributes, the training config has the following attributes:
-
-<!-- prettier-ignore-start -->
-```python
-@dataclass
-class TrainPipelineConfig:
-    dataset: DatasetConfig
-    env: envs.EnvConfig | None = None
-    policy: PreTrainedConfig | None = None
-```
-<!-- prettier-ignore-end -->
-
-in which `DatasetConfig` for example is defined as such:
-
-<!-- prettier-ignore-start -->
-```python
-@dataclass
-class DatasetConfig:
-    repo_id: str
-    episodes: list[int] | None = None
-    video_backend: str = "pyav"
-```
-<!-- prettier-ignore-end -->
-
-This creates a hierarchical relationship where, for example assuming we have a `cfg` instance of `TrainPipelineConfig`, we can access the `repo_id` value with `cfg.dataset.repo_id`.
-From the command line, we can specify this value by using a very similar syntax `--dataset.repo_id=repo/id`.
-
-By default, every field takes its default value specified in the dataclass. If a field doesn't have a default value, it needs to be specified either from the command line or from a config file – which path is also given in the command line (more in this below). In the example above, the `dataset` field doesn't have a default value which means it must be specified.
-
-## Specifying values from the CLI
-
-Let's say that we want to train [Diffusion Policy](../src/lerobot/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
-
-```bash
-lerobot-train \
-    --dataset.repo_id=lerobot/pusht \
-    --policy.type=diffusion \
-    --env.type=pusht
-```
-
-Let's break this down:
-
-- To specify the dataset, we just need to specify its `repo_id` on the hub which is the only required argument in the `DatasetConfig`. The rest of the fields have default values and in this case we are fine with those so we can just add the option `--dataset.repo_id=lerobot/pusht`.
-- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/policies](../src/lerobot/policies)
-- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/envs/configs.py`](../src/lerobot/envs/configs.py)
-
-Let's see another example. Let's say you've been training [ACT](../src/lerobot/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
-
-```bash
-lerobot-train \
-    --policy.type=act \
-    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
-    --env.type=aloha \
-    --output_dir=outputs/train/act_aloha_insertion
-```
-
-> Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.
-
-We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
-Looking at the [`AlohaEnv`](../src/lerobot/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
-
-```bash
-lerobot-train \
-    --policy.type=act \
-    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
-    --env.type=aloha \
-    --env.task=AlohaTransferCube-v0 \
-    --output_dir=outputs/train/act_aloha_transfer
-```
-
-## Loading from a config file
-
-Now, let's assume that we want to reproduce the run just above. That run has produced a `train_config.json` file in its checkpoints, which serializes the `TrainPipelineConfig` instance it used:
-
-```json
-{
-    "dataset": {
-        "repo_id": "lerobot/aloha_sim_transfer_cube_human",
-        "episodes": null,
-        ...
-    },
-    "env": {
-        "type": "aloha",
-        "task": "AlohaTransferCube-v0",
-        "fps": 50,
-        ...
-    },
-    "policy": {
-        "type": "act",
-        "n_obs_steps": 1,
-        ...
-    },
-    ...
-}
-```
-
-We can then simply load the config values from this file using:
-
-```bash
-lerobot-train \
-    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
-    --output_dir=outputs/train/act_aloha_transfer_2
-```
-
-`--config_path` is also a special argument which allows to initialize the config from a local config file. It can point to a directory that contains `train_config.json` or to the config file itself directly.
-
-Similarly to Hydra, we can still override some parameters in the CLI if we want to, e.g.:
-
-```bash
-lerobot-train \
-    --config_path=outputs/train/act_aloha_transfer/checkpoints/last/pretrained_model/ \
-    --output_dir=outputs/train/act_aloha_transfer_2
-    --policy.n_action_steps=80
-```
-
-> Note: While `--output_dir` is not required in general, in this case we need to specify it since it will otherwise take the value from the `train_config.json` (which is `outputs/train/act_aloha_transfer`). In order to prevent accidental deletion of previous run checkpoints, we raise an error if you're trying to write in an existing directory. This is not the case when resuming a run, which is what you'll learn next.
-
-`--config_path` can also accept the repo_id of a repo on the hub that contains a `train_config.json` file, e.g. running:
-
-```bash
-lerobot-train --config_path=lerobot/diffusion_pusht
-```
-
-will start a training run with the same configuration used for training [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)
-
-## Resume training
-
-Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to do that here.
-
-Let's reuse the command from the previous run and add a few more options:
-
-```bash
-lerobot-train \
-    --policy.type=act \
-    --dataset.repo_id=lerobot/aloha_sim_transfer_cube_human \
-    --env.type=aloha \
-    --env.task=AlohaTransferCube-v0 \
-    --log_freq=25 \
-    --save_freq=100 \
-    --output_dir=outputs/train/run_resumption
-```
-
-Here we've taken care to set up the log frequency and checkpointing frequency to low numbers so we can showcase resumption. You should be able to see some logging and have a first checkpoint within 1 minute (depending on hardware). Wait for the first checkpoint to happen, you should see a line that looks like this in your terminal:
-
-```
-INFO 2025-01-24 16:10:56 ts/train.py:263 Checkpoint policy after step 100
-```
-
-Now let's simulate a crash by killing the process (hit `ctrl`+`c`). We can then simply resume this run from the last checkpoint available with:
-
-```bash
-lerobot-train \
-    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
-    --resume=true
-```
-
-You should see from the logging that your training picks up from where it left off.
-
-Another reason for which you might want to resume a run is simply to extend training and add more training steps. The number of training steps is set by the option `--steps`, which is 100 000 by default.
-You could double the number of steps of the previous run with:
-
-```bash
-lerobot-train \
-    --config_path=outputs/train/run_resumption/checkpoints/last/pretrained_model/ \
-    --resume=true \
-    --steps=200000
-```
-
-## Outputs of a run
-
-In the output directory, there will be a folder called `checkpoints` with the following structure:
-
-```bash
-outputs/train/run_resumption/checkpoints
-├── 000100  # checkpoint_dir for training step 100
-│   ├── pretrained_model/
-│   │   ├── config.json  # policy config
-│   │   ├── model.safetensors  # policy weights
-│   │   └── train_config.json  # train config
-│   └── training_state/
-│       ├── optimizer_param_groups.json  #  optimizer param groups
-│       ├── optimizer_state.safetensors  # optimizer state
-│       ├── rng_state.safetensors  # rng states
-│       ├── scheduler_state.json  # scheduler state
-│       └── training_step.json  # training step
-├── 000200
-└── last -> 000200  # symlink to the last available checkpoint
-```
-
-## Fine-tuning a pre-trained policy
-
-In addition to the features currently in Draccus, we've added a special `.path` argument for the policy, which allows to load a policy as you would with `PreTrainedPolicy.from_pretrained()`. In that case, `path` can be a local directory that contains a checkpoint or a repo_id pointing to a pretrained policy on the hub.
-
-For example, we could fine-tune a [policy pre-trained on the aloha transfer task](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) on the aloha insertion task. We can achieve this with:
-
-```bash
-lerobot-train \
-    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \
-    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
-    --env.type=aloha \
-    --env.task=AlohaInsertion-v0
-```
-
-When doing so, keep in mind that the features of the fine-tuning dataset would have to match the input/output features of the pretrained policy.
-
-## Typical logs and metrics
-
-When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you configured your run correctly. The final configuration will also be saved with the checkpoint.
-
-After that, you will see training log like this one:
-
-```
-INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
-```
-
-or evaluation log:
-
-```
-INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
-```
-
-These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
-
-- `smpl`: number of samples seen during training.
-- `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
-- `epch`: number of time all unique samples are seen (epoch).
-- `grdn`: gradient norm.
-- `∑rwrd`: compute the sum of rewards in every evaluation episode and then take an average of them.
-- `success`: average success rate of eval episodes. Reward and success are usually different except for the sparsing reward setting, where reward=1 only when the task is completed successfully.
-- `eval_s`: time to evaluate the policy in the environment, in second.
-- `updt_s`: time to update the network parameters, in second.
-- `data_s`: time to load a batch of data, in second.
-
-Some metrics are useful for initial performance profiling. For example, if you find the current GPU utilization is low via the `nvidia-smi` command and `data_s` sometimes is too high, you may need to modify batch size or number of dataloading workers to accelerate dataloading. We also recommend [pytorch profiler](https://github.com/huggingface/lerobot?tab=readme-ov-file#improve-your-code-with-profiling) for detailed performance probing.
-
-## In short
-
-We'll summarize here the main use cases to remember from this tutorial.
-
-#### Train a policy from scratch – CLI
-
-```bash
-lerobot-train \
-    --policy.type=act \  # <- select 'act' policy
-    --env.type=pusht \  # <- select 'pusht' environment
-    --dataset.repo_id=lerobot/pusht  # <- train on this dataset
-```
-
-#### Train a policy from scratch - config file + CLI
-
-```bash
-lerobot-train \
-    --config_path=path/to/pretrained_model \  # <- can also be a repo_id
-    --policy.n_action_steps=80  # <- you may still override values
-```
-
-#### Resume/continue a training run
-
-```bash
-lerobot-train \
-    --config_path=checkpoint/pretrained_model/ \
-    --resume=true \
-    --steps=200000  # <- you can change some training parameters
-```
-
-#### Fine-tuning
-
-```bash
-lerobot-train \
-    --policy.path=lerobot/act_aloha_sim_transfer_cube_human \  # <- can also be a local path to a checkpoint
-    --dataset.repo_id=lerobot/aloha_sim_insertion_human \
-    --env.type=aloha \
-    --env.task=AlohaInsertion-v0
-```
-
----
-
-Now that you know the basics of how to train a policy, you might want to know how to apply this knowledge to actual robots, or how to record your own datasets and train policies on your specific task?
-If that's the case, head over to the next tutorial [`7_get_started_with_real_robot.md`](./7_get_started_with_real_robot.md).
-
-Or in the meantime, happy training! 🤗

examples/backward_compatibility/replay.py

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

examples/dataset/load_lerobot_dataset.py

@@ -136,7 +136,7 @@ print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
 # PyTorch datasets.
 dataloader = torch.utils.data.DataLoader(
     dataset,
-    num_workers=0,
+    num_workers=4,
     batch_size=32,
     shuffle=True,
 )

examples/dataset/use_dataset_image_transforms.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This example demonstrates how to use image transforms with LeRobot datasets for data augmentation during training.
+
+Image transforms are applied to camera frames to improve model robustness and generalization. They are applied
+at training time only, not during dataset recording, allowing you to experiment with different augmentations
+without re-recording data.
+"""
+
+import torch
+from torchvision.transforms import v2
+from torchvision.transforms.functional import to_pil_image
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.transforms import ImageTransformConfig, ImageTransforms, ImageTransformsConfig
+
+
+def save_image(tensor, filename):
+    """Helper function to save a tensor as an image file."""
+    if tensor.dim() == 3:  # [C, H, W]
+        if tensor.max() > 1.0:
+            tensor = tensor / 255.0
+        tensor = torch.clamp(tensor, 0.0, 1.0)
+        pil_image = to_pil_image(tensor)
+        pil_image.save(filename)
+        print(f"Saved: {filename}")
+    else:
+        print(f"Skipped {filename}: unexpected tensor shape {tensor.shape}")
+
+
+def example_1_default_transforms():
+    """Example 1: Use default transform configuration and save original vs transformed images"""
+    print("\n Example 1: Default Transform Configuration with Image Saving")
+
+    repo_id = "pepijn223/record_main_0"  # Example dataset
+
+    try:
+        # Load dataset without transforms (original)
+        dataset_original = LeRobotDataset(repo_id=repo_id)
+
+        # Load dataset with transforms enabled
+        transforms_config = ImageTransformsConfig(
+            enable=True,  # Enable transforms (disabled by default)
+            max_num_transforms=2,  # Apply up to 2 transforms per frame
+            random_order=False,  # Apply in standard order
+        )
+        dataset_with_transforms = LeRobotDataset(
+            repo_id=repo_id, image_transforms=ImageTransforms(transforms_config)
+        )
+
+        # Save original and transformed images for comparison
+        if len(dataset_original) > 0:
+            frame_idx = 0  # Use first frame
+            original_sample = dataset_original[frame_idx]
+            transformed_sample = dataset_with_transforms[frame_idx]
+
+            print(f"Saving comparison images (frame {frame_idx}):")
+
+            for cam_key in dataset_original.meta.camera_keys:
+                if cam_key in original_sample and cam_key in transformed_sample:
+                    cam_name = cam_key.replace(".", "_").replace("/", "_")
+
+                    # Save original and transformed images
+                    save_image(original_sample[cam_key], f"{cam_name}_original.png")
+                    save_image(transformed_sample[cam_key], f"{cam_name}_transformed.png")
+
+    except Exception as e:
+        print(f"Could not load dataset '{repo_id}': {e}")
+
+
+def example_2_custom_transforms():
+    """Example 2: Create custom transform configuration and save examples"""
+    print("\n Example 2: Custom Transform Configuration")
+
+    repo_id = "pepijn223/record_main_0"  # Example dataset
+
+    try:
+        # Create custom transform configuration with strong effects
+        custom_transforms_config = ImageTransformsConfig(
+            enable=True,
+            max_num_transforms=2,  # Apply up to 2 transforms per frame
+            random_order=True,  # Apply transforms in random order
+            tfs={
+                "brightness": ImageTransformConfig(
+                    weight=1.0,
+                    type="ColorJitter",
+                    kwargs={"brightness": (0.5, 1.5)},  # Strong brightness range
+                ),
+                "contrast": ImageTransformConfig(
+                    weight=1.0,  # Higher weight = more likely to be selected
+                    type="ColorJitter",
+                    kwargs={"contrast": (0.6, 1.4)},  # Strong contrast
+                ),
+                "sharpness": ImageTransformConfig(
+                    weight=0.5,  # Lower weight = less likely to be selected
+                    type="SharpnessJitter",
+                    kwargs={"sharpness": (0.2, 2.0)},  # Strong sharpness variation
+                ),
+            },
+        )
+
+        dataset_with_custom_transforms = LeRobotDataset(
+            repo_id=repo_id, image_transforms=ImageTransforms(custom_transforms_config)
+        )
+
+        # Save examples with strong transforms
+        if len(dataset_with_custom_transforms) > 0:
+            sample = dataset_with_custom_transforms[0]
+            print("Saving custom transform examples:")
+
+            for cam_key in dataset_with_custom_transforms.meta.camera_keys:
+                if cam_key in sample:
+                    cam_name = cam_key.replace(".", "_").replace("/", "_")
+                    save_image(sample[cam_key], f"{cam_name}_custom_transforms.png")
+
+    except Exception as e:
+        print(f"Could not load dataset '{repo_id}': {e}")
+
+
+def example_3_torchvision_transforms():
+    """Example 3: Use pure torchvision transforms and save examples"""
+    print("\n Example 3: Pure Torchvision Transforms")
+
+    repo_id = "pepijn223/record_main_0"  # Example dataset
+
+    try:
+        # Create torchvision transform pipeline
+        torchvision_transforms = v2.Compose(
+            [
+                v2.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.1),
+                v2.GaussianBlur(kernel_size=3, sigma=(0.1, 2.0)),
+                v2.RandomRotation(degrees=10),  # Small rotation
+            ]
+        )
+
+        dataset_with_torchvision = LeRobotDataset(repo_id=repo_id, image_transforms=torchvision_transforms)
+
+        # Save examples with torchvision transforms
+        if len(dataset_with_torchvision) > 0:
+            sample = dataset_with_torchvision[0]
+            print("Saving torchvision transform examples:")
+
+            for cam_key in dataset_with_torchvision.meta.camera_keys:
+                if cam_key in sample:
+                    cam_name = cam_key.replace(".", "_").replace("/", "_")
+                    save_image(sample[cam_key], f"{cam_name}_torchvision.png")
+
+    except Exception as e:
+        print(f"Could not load dataset '{repo_id}': {e}")
+
+
+def main():
+    """Run all examples"""
+    print("LeRobot Dataset Image Transforms Examples")
+
+    example_1_default_transforms()
+    example_2_custom_transforms()
+    example_3_torchvision_transforms()
+
+
+if __name__ == "__main__":
+    main()

examples/dataset/use_dataset_tools.py

@@ -0,0 +1,124 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Example script demonstrating dataset tools utilities.
+
+This script shows how to:
+1. Delete episodes from a dataset
+2. Split a dataset into train/val sets
+3. Add/remove features
+4. Merge datasets
+
+Usage:
+    python examples/dataset/use_dataset_tools.py
+"""
+
+import numpy as np
+
+from lerobot.datasets.dataset_tools import (
+    add_features,
+    delete_episodes,
+    merge_datasets,
+    modify_features,
+    remove_feature,
+    split_dataset,
+)
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+
+def main():
+    dataset = LeRobotDataset("lerobot/pusht")
+
+    print(f"Original dataset: {dataset.meta.total_episodes} episodes, {dataset.meta.total_frames} frames")
+    print(f"Features: {list(dataset.meta.features.keys())}")
+
+    print("\n1. Deleting episodes 0 and 2...")
+    filtered_dataset = delete_episodes(dataset, episode_indices=[0, 2], repo_id="lerobot/pusht_filtered")
+    print(f"Filtered dataset: {filtered_dataset.meta.total_episodes} episodes")
+
+    print("\n2. Splitting dataset into train/val...")
+    splits = split_dataset(
+        dataset,
+        splits={"train": 0.8, "val": 0.2},
+    )
+    print(f"Train split: {splits['train'].meta.total_episodes} episodes")
+    print(f"Val split: {splits['val'].meta.total_episodes} episodes")
+
+    print("\n3. Adding features...")
+
+    reward_values = np.random.randn(dataset.meta.total_frames).astype(np.float32)
+
+    def compute_success(row_dict, episode_index, frame_index):
+        episode_length = 10
+        return float(frame_index >= episode_length - 10)
+
+    dataset_with_features = add_features(
+        dataset,
+        features={
+            "reward": (
+                reward_values,
+                {"dtype": "float32", "shape": (1,), "names": None},
+            ),
+            "success": (
+                compute_success,
+                {"dtype": "float32", "shape": (1,), "names": None},
+            ),
+        },
+        repo_id="lerobot/pusht_with_features",
+    )
+
+    print(f"New features: {list(dataset_with_features.meta.features.keys())}")
+
+    print("\n4. Removing the success feature...")
+    dataset_cleaned = remove_feature(
+        dataset_with_features, feature_names="success", repo_id="lerobot/pusht_cleaned"
+    )
+    print(f"Features after removal: {list(dataset_cleaned.meta.features.keys())}")
+
+    print("\n5. Using modify_features to add and remove features simultaneously...")
+    dataset_modified = modify_features(
+        dataset_with_features,
+        add_features={
+            "discount": (
+                np.ones(dataset.meta.total_frames, dtype=np.float32) * 0.99,
+                {"dtype": "float32", "shape": (1,), "names": None},
+            ),
+        },
+        remove_features="reward",
+        repo_id="lerobot/pusht_modified",
+    )
+    print(f"Modified features: {list(dataset_modified.meta.features.keys())}")
+
+    print("\n6. Merging train and val splits back together...")
+    merged = merge_datasets([splits["train"], splits["val"]], output_repo_id="lerobot/pusht_merged")
+    print(f"Merged dataset: {merged.meta.total_episodes} episodes")
+
+    print("\n7. Complex workflow example...")
+
+    if len(dataset.meta.camera_keys) > 1:
+        camera_to_remove = dataset.meta.camera_keys[0]
+        print(f"Removing camera: {camera_to_remove}")
+        dataset_no_cam = remove_feature(
+            dataset, feature_names=camera_to_remove, repo_id="pusht_no_first_camera"
+        )
+        print(f"Remaining cameras: {dataset_no_cam.meta.camera_keys}")
+
+    print("\nDone! Check ~/.cache/huggingface/lerobot/ for the created datasets.")
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/evaluate.py

@@ -1,31 +1,54 @@
+# !/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 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.record import record_loop
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.processor import make_default_processors
 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
 EPISODE_TIME_SEC = 60
 TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
-# Create the robot and teleoperator configurations
+# Create the robot configuration & robot
 robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+
 robot = LeKiwiClient(robot_config)
 
-policy = ACTPolicy.from_pretrained("<hf_username>/<policy_repo_id>")
+# Create policy
+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, "observation")
+action_features = hw_to_dataset_features(robot.action_features, ACTION)
+obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
 dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/<eval_dataset_repo_id>",
+    repo_id=HF_DATASET_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -33,33 +56,52 @@ dataset = LeRobotDataset.create(
     image_writer_threads=4,
 )
 
+# Build Policy Processors
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+    # 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)}},
+)
+
+# Connect the robot
 # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
 robot.connect()
 
-_init_rerun(session_name="recording")
+# TODO(Steven): Update this example to use pipelines
+teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
 
+# Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
+init_rerun(session_name="lekiwi_evaluate")
 
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
 
+print("Starting evaluate loop...")
 recorded_episodes = 0
 while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
     log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
 
-    # Run the policy inference loop
+    # Main record loop
     record_loop(
         robot=robot,
         events=events,
         fps=FPS,
         policy=policy,
+        preprocessor=preprocessor,  # Pass the pre and post policy processors
+        postprocessor=postprocessor,
         dataset=dataset,
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,
         display_data=True,
+        teleop_action_processor=teleop_action_processor,
+        robot_action_processor=robot_action_processor,
+        robot_observation_processor=robot_observation_processor,
     )
 
-    # Logic for reset env
+    # Reset the environment if not stopping or re-recording
     if not events["stop_recording"] and (
         (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
     ):
@@ -71,6 +113,9 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
+            teleop_action_processor=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            robot_observation_processor=robot_observation_processor,
         )
 
     if events["rerecord_episode"]:
@@ -80,11 +125,14 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
         dataset.clear_episode_buffer()
         continue
 
+    # Save episode
     dataset.save_episode()
     recorded_episodes += 1
 
-# Upload to hub and clean up
-dataset.push_to_hub()
-
+# Clean up
+log_say("Stop recording")
 robot.disconnect()
 listener.stop()
+
+dataset.finalize()
+dataset.push_to_hub()

examples/lekiwi/record.py

@@ -1,37 +1,60 @@
+# !/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 lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.datasets.utils import hw_to_dataset_features
-from lerobot.record import record_loop
+from lerobot.processor import make_default_processors
 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 = 3
+NUM_EPISODES = 2
 FPS = 30
 EPISODE_TIME_SEC = 30
 RESET_TIME_SEC = 10
 TASK_DESCRIPTION = "My task description"
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
 
 # Create the robot and teleoperator configurations
 robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 leader_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
 keyboard_config = KeyboardTeleopConfig()
 
+# Initialize the robot and teleoperator
 robot = LeKiwiClient(robot_config)
 leader_arm = SO100Leader(leader_arm_config)
 keyboard = KeyboardTeleop(keyboard_config)
 
+# TODO(Steven): Update this example to use pipelines
+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, "observation")
+action_features = hw_to_dataset_features(robot.action_features, ACTION)
+obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
 dataset_features = {**action_features, **obs_features}
 
 # Create the dataset
 dataset = LeRobotDataset.create(
-    repo_id="<hf_username>/<dataset_repo_id>",
+    repo_id=HF_REPO_ID,
     fps=FPS,
     features=dataset_features,
     robot_type=robot.name,
@@ -39,23 +62,25 @@ dataset = LeRobotDataset.create(
     image_writer_threads=4,
 )
 
+# Connect the robot and teleoperator
 # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
 robot.connect()
 leader_arm.connect()
 keyboard.connect()
 
-_init_rerun(session_name="lekiwi_record")
-
+# Initialize the keyboard listener and rerun visualization
 listener, events = init_keyboard_listener()
+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, leader arm of keyboard is not connected!")
+    raise ValueError("Robot or teleop is not connected!")
 
+print("Starting record loop...")
 recorded_episodes = 0
 while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
     log_say(f"Recording episode {recorded_episodes}")
 
-    # Run the record loop
+    # Main record loop
     record_loop(
         robot=robot,
         events=events,
@@ -65,9 +90,12 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
         control_time_s=EPISODE_TIME_SEC,
         single_task=TASK_DESCRIPTION,
         display_data=True,
+        teleop_action_processor=teleop_action_processor,
+        robot_action_processor=robot_action_processor,
+        robot_observation_processor=robot_observation_processor,
     )
 
-    # Logic for reset env
+    # Reset the environment if not stopping or re-recording
     if not events["stop_recording"] and (
         (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
     ):
@@ -80,6 +108,9 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             control_time_s=RESET_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
+            teleop_action_processor=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            robot_observation_processor=robot_observation_processor,
         )
 
     if events["rerecord_episode"]:
@@ -89,13 +120,16 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
         dataset.clear_episode_buffer()
         continue
 
+    # Save episode
     dataset.save_episode()
     recorded_episodes += 1
 
-# Upload to hub and clean up
-dataset.push_to_hub()
-
+# Clean up
+log_say("Stop recording")
 robot.disconnect()
 leader_arm.disconnect()
 keyboard.disconnect()
 listener.stop()
+
+dataset.finalize()
+dataset.push_to_hub()

examples/lekiwi/replay.py

@@ -1,32 +1,60 @@
+# !/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 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
 
 EPISODE_IDX = 0
 
+# Initialize the robot config
 robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+
+# Initialize the robot
 robot = LeKiwiClient(robot_config)
 
+# Fetch the dataset to replay
 dataset = LeRobotDataset("<hf_username>/<dataset_repo_id>", episodes=[EPISODE_IDX])
-actions = dataset.hf_dataset.select_columns("action")
+# 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)
 
+# Connect to the robot
 robot.connect()
 
 if not robot.is_connected:
     raise ValueError("Robot is not connected!")
 
+print("Starting replay loop...")
 log_say(f"Replaying episode {EPISODE_IDX}")
-for idx in range(dataset.num_frames):
+for idx in range(len(episode_frames)):
     t0 = time.perf_counter()
 
+    # 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"])
     }
-    robot.send_action(action)
+
+    # Send action to robot
+    _ = robot.send_action(action)
 
     busy_wait(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 

examples/lekiwi/teleoperate.py

@@ -1,10 +1,26 @@
+# !/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 time
 
 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
 
@@ -13,35 +29,44 @@ robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="my_lekiwi")
 teleop_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
 keyboard_config = KeyboardTeleopConfig(id="my_laptop_keyboard")
 
+# Initialize the robot and teleoperator
 robot = LeKiwiClient(robot_config)
 leader_arm = SO100Leader(teleop_arm_config)
 keyboard = KeyboardTeleop(keyboard_config)
 
+# Connect to the robot and teleoperator
 # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
 robot.connect()
 leader_arm.connect()
 keyboard.connect()
 
-_init_rerun(session_name="lekiwi_teleop")
+# Init rerun viewer
+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, leader arm of keyboard is not connected!")
+    raise ValueError("Robot or teleop is not connected!")
 
+print("Starting teleop loop...")
 while True:
     t0 = time.perf_counter()
 
+    # Get robot observation
     observation = robot.get_observation()
 
+    # Get teleop action
+    # Arm
     arm_action = leader_arm.get_action()
     arm_action = {f"arm_{k}": v for k, v in arm_action.items()}
-
+    # Keyboard
     keyboard_keys = keyboard.get_action()
     base_action = robot._from_keyboard_to_base_action(keyboard_keys)
 
-    log_rerun_data(observation, {**arm_action, **base_action})
-
     action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
 
-    robot.send_action(action)
+    # Send action to robot
+    _ = robot.send_action(action)
+
+    # Visualize
+    log_rerun_data(observation=observation, action=action)
 
     busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/phone_to_so100/evaluate.py

@@ -0,0 +1,199 @@
+# !/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 lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.processor import (
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+    make_default_teleop_action_processor,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+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
+
+NUM_EPISODES = 5
+FPS = 30
+EPISODE_TIME_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
+
+# Create the robot configuration & robot
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem58760434471",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+
+robot = SO100Follower(robot_config)
+
+# Create policy
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert joints observation to EE observation
+robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_DATASET_ID,
+    fps=FPS,
+    features=combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_joints_to_ee_pose_processor,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=True,
+        ),
+        # User for now should be explicit on the feature keys that were used for record
+        # Alternatively, the user can pass the processor step that has the right features
+        aggregate_pipeline_dataset_features(
+            pipeline=make_default_teleop_action_processor(),
+            initial_features=create_initial_features(
+                action={
+                    f"ee.{k}": PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+                    for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]
+                }
+            ),
+            use_videos=True,
+        ),
+    ),
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Build Policy Processors
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+    # 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)}},
+)
+
+# Connect the robot
+robot.connect()
+
+# Initialize the keyboard listener and rerun visualization
+listener, events = init_keyboard_listener()
+init_rerun(session_name="phone_so100_evaluate")
+
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
+
+print("Starting evaluate loop...")
+episode_idx = 0
+for episode_idx in range(NUM_EPISODES):
+    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    # Main record loop
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        policy=policy,
+        preprocessor=preprocessor,  # Pass the pre and post policy processors
+        postprocessor=postprocessor,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=make_default_teleop_action_processor(),
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=make_default_teleop_action_processor(),
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose_processor,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-record episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+listener.stop()
+
+dataset.finalize()
+dataset.push_to_hub()

examples/phone_to_so100/record.py

@@ -0,0 +1,205 @@
+# !/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 lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    EEReferenceAndDelta,
+    ForwardKinematicsJointsToEE,
+    GripperVelocityToJoint,
+    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
+
+NUM_EPISODES = 2
+FPS = 30
+EPISODE_TIME_SEC = 60
+RESET_TIME_SEC = 30
+TASK_DESCRIPTION = "My task description"
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Create the robot and teleoperator configurations
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+phone = Phone(teleop_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert phone action to EE action
+phone_to_robot_ee_pose_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver,
+            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+            motor_names=list(robot.bus.motors.keys()),
+            use_latched_reference=True,
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.20,
+        ),
+        GripperVelocityToJoint(speed_factor=20.0),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert joint observation to EE observation
+robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation](
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_REPO_ID,
+    fps=FPS,
+    features=combine_feature_dicts(
+        # Run the feature contract of the pipelines
+        # This tells you how the features would look like after the pipeline steps
+        aggregate_pipeline_dataset_features(
+            pipeline=phone_to_robot_ee_pose_processor,
+            initial_features=create_initial_features(action=phone.action_features),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_joints_to_ee_pose,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=True,
+        ),
+    ),
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Connect the robot and teleoperator
+robot.connect()
+phone.connect()
+
+# Initialize the keyboard listener and rerun visualization
+listener, events = init_keyboard_listener()
+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!")
+
+
+print("Starting record loop. Move your phone to teleoperate the robot...")
+episode_idx = 0
+while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    # Main record loop
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        teleop=phone,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=phone_to_robot_ee_pose_processor,
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            teleop=phone,
+            control_time_s=RESET_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=phone_to_robot_ee_pose_processor,
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+phone.disconnect()
+listener.stop()
+
+dataset.finalize()
+dataset.push_to_hub()

examples/phone_to_so100/replay.py

@@ -0,0 +1,100 @@
+# !/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 time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    robot_action_observation_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+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
+
+EPISODE_IDX = 0
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Initialize the robot config
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+)
+
+# Initialize the robot
+robot = SO100Follower(robot_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=False,  # Because replay is open loop
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Fetch the dataset to replay
+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)
+
+# Connect to the robot
+robot.connect()
+
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
+
+print("Starting replay loop...")
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(len(episode_frames)):
+    t0 = time.perf_counter()
+
+    # Get recorded action from dataset
+    ee_action = {
+        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+    }
+
+    # Get robot observation
+    robot_obs = robot.get_observation()
+
+    # Dataset EE -> robot joints
+    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+
+    # Send action to robot
+    _ = robot.send_action(joint_action)
+
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+# Clean up
+robot.disconnect()

examples/phone_to_so100/teleoperate.py

@@ -0,0 +1,113 @@
+# !/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 specif
+
+import time
+
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    robot_action_observation_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    EEReferenceAndDelta,
+    GripperVelocityToJoint,
+    InverseKinematicsEEToJoints,
+)
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+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
+
+FPS = 30
+
+# Initialize the robot and teleoperator
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+)
+teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
+
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+teleop_device = Phone(teleop_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert phone action to ee pose action to joint action
+phone_to_robot_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+        EEReferenceAndDelta(
+            kinematics=kinematics_solver,
+            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+            motor_names=list(robot.bus.motors.keys()),
+            use_latched_reference=True,
+        ),
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.10,
+        ),
+        GripperVelocityToJoint(
+            speed_factor=20.0,
+        ),
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Connect to the robot and teleoperator
+robot.connect()
+teleop_device.connect()
+
+# Init rerun viewer
+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!")
+
+print("Starting teleop loop. Move your phone to teleoperate the robot...")
+while True:
+    t0 = time.perf_counter()
+
+    # Get robot observation
+    robot_obs = robot.get_observation()
+
+    # Get teleop action
+    phone_obs = teleop_device.get_action()
+
+    # Phone -> EE pose -> Joints transition
+    joint_action = phone_to_robot_joints_processor((phone_obs, robot_obs))
+
+    # Send action to robot
+    _ = robot.send_action(joint_action)
+
+    # Visualize
+    log_rerun_data(observation=phone_obs, action=joint_action)
+
+    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/port_datasets/port_droid.py

@@ -362,6 +362,8 @@ 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

@@ -0,0 +1,200 @@
+# !/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 lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.configs.types import FeatureType, PolicyFeature
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.processor import (
+    RobotAction,
+    RobotObservation,
+    RobotProcessorPipeline,
+    make_default_teleop_action_processor,
+)
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+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
+
+NUM_EPISODES = 5
+FPS = 30
+EPISODE_TIME_SEC = 60
+TASK_DESCRIPTION = "My task description"
+HF_MODEL_ID = "<hf_username>/<model_repo_id>"
+HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
+
+# Create the robot configuration & robot
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411",
+    id="my_awesome_follower_arm",
+    cameras=camera_config,
+    use_degrees=True,
+)
+
+robot = SO100Follower(robot_config)
+
+# Create policy
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert joints observation to EE observation
+robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+    steps=[
+        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_DATASET_ID,
+    fps=FPS,
+    features=combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_joints_to_ee_pose_processor,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=True,
+        ),
+        # User for now should be explicit on the feature keys that were used for record
+        # Alternatively, the user can pass the processor step that has the right features
+        aggregate_pipeline_dataset_features(
+            pipeline=make_default_teleop_action_processor(),
+            initial_features=create_initial_features(
+                action={
+                    f"ee.{k}": PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+                    for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]
+                }
+            ),
+            use_videos=True,
+        ),
+    ),
+    robot_type=robot.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+# Build Policy Processors
+preprocessor, postprocessor = make_pre_post_processors(
+    policy_cfg=policy,
+    pretrained_path=HF_MODEL_ID,
+    dataset_stats=dataset.meta.stats,
+    # 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)}},
+)
+
+# Connect the robot and teleoperator
+robot.connect()
+
+# Initialize the keyboard listener and rerun visualization
+listener, events = init_keyboard_listener()
+init_rerun(session_name="so100_so100_evaluate")
+
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
+
+print("Starting evaluate loop...")
+episode_idx = 0
+for episode_idx in range(NUM_EPISODES):
+    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    # Main record loop
+    record_loop(
+        robot=robot,
+        events=events,
+        fps=FPS,
+        policy=policy,
+        preprocessor=preprocessor,  # Pass the pre and post policy processors
+        postprocessor=postprocessor,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=make_default_teleop_action_processor(),
+        robot_action_processor=robot_ee_to_joints_processor,
+        robot_observation_processor=robot_joints_to_ee_pose_processor,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=make_default_teleop_action_processor(),
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose_processor,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-record episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+listener.stop()
+
+dataset.finalize()
+dataset.push_to_hub()

examples/so100_to_so100_EE/record.py

@@ -0,0 +1,204 @@
+# !/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 lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from lerobot.datasets.utils import combine_feature_dicts
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    observation_to_transition,
+    robot_action_observation_to_transition,
+    transition_to_observation,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    ForwardKinematicsJointsToEE,
+    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
+
+NUM_EPISODES = 2
+FPS = 30
+EPISODE_TIME_SEC = 60
+RESET_TIME_SEC = 30
+TASK_DESCRIPTION = "My task description"
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Create the robot and teleoperator configurations
+camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+follower_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", cameras=camera_config, use_degrees=True
+)
+leader_config = SO100LeaderConfig(port="/dev/tty.usbmodem5A460819811", id="my_awesome_leader_arm")
+
+# Initialize the robot and teleoperator
+follower = SO100Follower(follower_config)
+leader = SO100Leader(leader_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+follower_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(follower.bus.motors.keys()),
+)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+leader_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(leader.bus.motors.keys()),
+)
+
+# Build pipeline to convert follower joints to EE observation
+follower_joints_to_ee = RobotProcessorPipeline[RobotObservation, RobotObservation](
+    steps=[
+        ForwardKinematicsJointsToEE(
+            kinematics=follower_kinematics_solver, motor_names=list(follower.bus.motors.keys())
+        ),
+    ],
+    to_transition=observation_to_transition,
+    to_output=transition_to_observation,
+)
+
+# Build pipeline to convert leader joints to EE action
+leader_joints_to_ee = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        ForwardKinematicsJointsToEE(
+            kinematics=leader_kinematics_solver, motor_names=list(leader.bus.motors.keys())
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Build pipeline to convert EE action to follower joints
+ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    [
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.10,
+        ),
+        InverseKinematicsEEToJoints(
+            kinematics=follower_kinematics_solver,
+            motor_names=list(follower.bus.motors.keys()),
+            initial_guess_current_joints=True,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Create the dataset
+dataset = LeRobotDataset.create(
+    repo_id=HF_REPO_ID,
+    fps=FPS,
+    features=combine_feature_dicts(
+        # Run the feature contract of the pipelines
+        # This tells you how the features would look like after the pipeline steps
+        aggregate_pipeline_dataset_features(
+            pipeline=leader_joints_to_ee,
+            initial_features=create_initial_features(action=leader.action_features),
+            use_videos=True,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=follower_joints_to_ee,
+            initial_features=create_initial_features(observation=follower.observation_features),
+            use_videos=True,
+        ),
+    ),
+    robot_type=follower.name,
+    use_videos=True,
+    image_writer_threads=4,
+)
+
+
+# Connect the robot and teleoperator
+leader.connect()
+follower.connect()
+
+# Initialize the keyboard listener and rerun visualization
+listener, events = init_keyboard_listener()
+init_rerun(session_name="recording_phone")
+
+if not leader.is_connected or not follower.is_connected:
+    raise ValueError("Robot or teleop is not connected!")
+
+print("Starting record loop...")
+episode_idx = 0
+while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+    # Main record loop
+    record_loop(
+        robot=follower,
+        events=events,
+        fps=FPS,
+        teleop=leader,
+        dataset=dataset,
+        control_time_s=EPISODE_TIME_SEC,
+        single_task=TASK_DESCRIPTION,
+        display_data=True,
+        teleop_action_processor=leader_joints_to_ee,
+        robot_action_processor=ee_to_follower_joints,
+        robot_observation_processor=follower_joints_to_ee,
+    )
+
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
+        record_loop(
+            robot=follower,
+            events=events,
+            fps=FPS,
+            teleop=leader,
+            control_time_s=RESET_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=leader_joints_to_ee,
+            robot_action_processor=ee_to_follower_joints,
+            robot_observation_processor=follower_joints_to_ee,
+        )
+
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
+
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
+
+# Clean up
+log_say("Stop recording")
+leader.disconnect()
+follower.disconnect()
+listener.stop()
+
+dataset.finalize()
+dataset.push_to_hub()

examples/so100_to_so100_EE/replay.py

@@ -0,0 +1,101 @@
+# !/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 time
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    robot_action_observation_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+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
+
+EPISODE_IDX = 0
+HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
+
+# Initialize the robot config
+robot_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+)
+
+# Initialize the robot
+robot = SO100Follower(robot_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.bus.motors.keys()),
+)
+
+# Build pipeline to convert EE action to joints action
+robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    steps=[
+        InverseKinematicsEEToJoints(
+            kinematics=kinematics_solver,
+            motor_names=list(robot.bus.motors.keys()),
+            initial_guess_current_joints=False,  # Because replay is open loop
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Fetch the dataset to replay
+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)
+
+# Connect to the robot
+robot.connect()
+
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
+
+print("Starting replay loop...")
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(len(episode_frames)):
+    t0 = time.perf_counter()
+
+    # Get recorded action from dataset
+    ee_action = {
+        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+    }
+
+    # Get robot observation
+    robot_obs = robot.get_observation()
+
+    # Dataset EE -> robot joints
+    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+
+    # Send action to robot
+    _ = robot.send_action(joint_action)
+
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+# Clean up
+robot.disconnect()

examples/so100_to_so100_EE/teleoperate.py

@@ -0,0 +1,121 @@
+# !/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 time
+
+from lerobot.model.kinematics import RobotKinematics
+from lerobot.processor import RobotAction, RobotObservation, RobotProcessorPipeline
+from lerobot.processor.converters import (
+    robot_action_observation_to_transition,
+    robot_action_to_transition,
+    transition_to_robot_action,
+)
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.robot_kinematic_processor import (
+    EEBoundsAndSafety,
+    ForwardKinematicsJointsToEE,
+    InverseKinematicsEEToJoints,
+)
+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
+
+FPS = 30
+
+# Initialize the robot and teleoperator config
+follower_config = SO100FollowerConfig(
+    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+)
+leader_config = SO100LeaderConfig(port="/dev/tty.usbmodem5A460819811", id="my_awesome_leader_arm")
+
+# Initialize the robot and teleoperator
+follower = SO100Follower(follower_config)
+leader = SO100Leader(leader_config)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+follower_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(follower.bus.motors.keys()),
+)
+
+# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+leader_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(leader.bus.motors.keys()),
+)
+
+# Build pipeline to convert teleop joints to EE action
+leader_to_ee = RobotProcessorPipeline[RobotAction, RobotAction](
+    steps=[
+        ForwardKinematicsJointsToEE(
+            kinematics=leader_kinematics_solver, motor_names=list(leader.bus.motors.keys())
+        ),
+    ],
+    to_transition=robot_action_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# build pipeline to convert EE action to robot joints
+ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+    [
+        EEBoundsAndSafety(
+            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+            max_ee_step_m=0.10,
+        ),
+        InverseKinematicsEEToJoints(
+            kinematics=follower_kinematics_solver,
+            motor_names=list(follower.bus.motors.keys()),
+            initial_guess_current_joints=False,
+        ),
+    ],
+    to_transition=robot_action_observation_to_transition,
+    to_output=transition_to_robot_action,
+)
+
+# Connect to the robot and teleoperator
+follower.connect()
+leader.connect()
+
+# Init rerun viewer
+init_rerun(session_name="so100_so100_EE_teleop")
+
+print("Starting teleop loop...")
+while True:
+    t0 = time.perf_counter()
+
+    # Get robot observation
+    robot_obs = follower.get_observation()
+
+    # Get teleop observation
+    leader_joints_obs = leader.get_action()
+
+    # teleop joints -> teleop EE action
+    leader_ee_act = leader_to_ee(leader_joints_obs)
+
+    # teleop EE -> robot joints
+    follower_joints_act = ee_to_follower_joints((leader_ee_act, robot_obs))
+
+    # Send action to robot
+    _ = follower.send_action(follower_joints_act)
+
+    # Visualize
+    log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
+
+    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/training/train_policy.py

@@ -12,11 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""This script demonstrates how to train Diffusion Policy on the PushT environment.
-
-Once you have trained a model with this script, you can try to evaluate it on
-examples/2_evaluate_pretrained_policy.py
-"""
+"""This script demonstrates how to train Diffusion Policy on the PushT environment."""
 
 from pathlib import Path
 
@@ -27,6 +23,7 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetad
 from lerobot.datasets.utils import dataset_to_policy_features
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
+from lerobot.policies.factory import make_pre_post_processors
 
 
 def main():
@@ -56,9 +53,10 @@ def main():
     cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
 
     # We can now instantiate our policy with this config and the dataset stats.
-    policy = DiffusionPolicy(cfg, dataset_stats=dataset_metadata.stats)
+    policy = DiffusionPolicy(cfg)
     policy.train()
     policy.to(device)
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
 
     # Another policy-dataset interaction is with the delta_timestamps. Each policy expects a given number frames
     # which can differ for inputs, outputs and rewards (if there are some).
@@ -99,7 +97,7 @@ def main():
     done = False
     while not done:
         for batch in dataloader:
-            batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
+            batch = preprocessor(batch)
             loss, _ = policy.forward(batch)
             loss.backward()
             optimizer.step()
@@ -114,6 +112,8 @@ def main():
 
     # Save a policy checkpoint.
     policy.save_pretrained(output_directory)
+    preprocessor.save_pretrained(output_directory)
+    postprocessor.save_pretrained(output_directory)
 
 
 if __name__ == "__main__":

examples/training/train_with_streaming.py

@@ -13,23 +13,20 @@
 # limitations under the License.
 
 """This script demonstrates how to train a Diffusion Policy on the PushT environment,
-using a dataset processed in streaming mode.
-
-Once you have trained a model with this script, you can try to evaluate it on
-examples/2_evaluate_pretrained_policy.py
-"""
+using a dataset processed in streaming mode."""
 
 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():
@@ -50,9 +47,7 @@ def main():
     training_steps = 10
     log_freq = 1
 
-    dataset_id = (
-        "aractingi/droid_1.0.1"  # 26M frames! Would require 4TB of disk space if installed locally (:
-    )
+    dataset_id = "lerobot/droid_1.0.1"  # 26M frames! Would require 4TB of disk space if installed locally (:
     dataset_metadata = LeRobotDatasetMetadata(dataset_id)
     features = dataset_to_policy_features(dataset_metadata.features)
     output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
@@ -60,9 +55,10 @@ def main():
 
     # We can now instantiate our policy with this config and the dataset stats.
     cfg = ACTConfig(input_features=input_features, output_features=output_features)
-    policy = ACTPolicy(cfg, dataset_stats=dataset_metadata.stats)
+    policy = ACTPolicy(cfg)
     policy.train()
     policy.to(device)
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
 
     # Delta timestamps are used to (1) augment frames used during training and (2) supervise the policy.
     # Here, we use delta-timestamps to only provide ground truth actions for supervision
@@ -89,13 +85,7 @@ def main():
     done = False
     while not done:
         for batch in dataloader:
-            batch = {
-                k: (v.type(torch.float32) if isinstance(v, torch.Tensor) and v.dtype != torch.bool else v)
-                for k, v in batch.items()
-            }
-            batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
-
-            # batch = {k: (v.to(device) if isinstance(v, torch.Tensor) else v) for k, v in batch.items()}
+            batch = preprocessor(batch)
             loss, _ = policy.forward(batch)
             loss.backward()
             optimizer.step()
@@ -110,6 +100,8 @@ def main():
 
     # Save a policy checkpoint.
     policy.save_pretrained(output_directory)
+    preprocessor.save_pretrained(output_directory)
+    postprocessor.save_pretrained(output_directory)
 
 
 if __name__ == "__main__":

pyproject.toml

@@ -59,27 +59,28 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=2.19.0,<=3.6.0", # TODO: Bumb dependency
-    "diffusers>=0.27.2",
-    "huggingface-hub[hf-transfer,cli]>=0.34.2",
+    "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",
 
     # Core dependencies
-    "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",
+    "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,<0.23.0",
 
     "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>=0.29.1,<1.0.0", # TODO: Bumb dependency
+    "gymnasium>=1.0.0",
     "rerun-sdk>=0.21.0,<0.23.0", # TODO: Bumb dependency
 
     # Support dependencies
@@ -92,25 +93,26 @@ dependencies = [
 [project.optional-dependencies]
 
 # Common
-pygame-dep = ["pygame>=2.5.1"]
-placo-dep = ["placo>=0.9.6"]
-transformers-dep = ["transformers>=4.50.3,<4.52.0"] # TODO: Bumb dependency
+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"]
 
 # Motors
-feetech = ["feetech-servo-sdk>=1.0.0"]
-dynamixel = ["dynamixel-sdk>=3.7.31"]
+feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
+dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
 
 # Robots
-gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0"]
+gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
-lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1"]
-reachy2 = ["reachy2_sdk>=1.0.14"]
+lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
+reachy2 = ["reachy2_sdk>=1.0.14,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
-    "pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'",
-    "pyrealsense2-macosx>=2.54 ; sys_platform == 'darwin'",
+    "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
+    "pyrealsense2-macosx>=2.54,<2.55.0 ; sys_platform == 'darwin'",
 ]
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0"]
 # stretch = [
 #     "hello-robot-stretch-body>=0.7.27 ; sys_platform == 'linux'",
 #     "pyrender @ git+https://github.com/mmatl/pyrender.git ; sys_platform == 'linux'",
@@ -118,22 +120,23 @@ intelrealsense = [
 # ] # TODO: Currently not supported
 
 # Policies
-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>=0.1.9", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+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]"]
 
 # Features
-async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3"]
+async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
 
 # Development
-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"]
+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"]
 
 # Simulation
-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"]
+aloha = ["gym-aloha>=0.1.2,<0.2.0"]
+pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
+libero = ["lerobot[transformers-dep]", "libero @ git+https://github.com/huggingface/lerobot-libero.git@main#egg=libero"]
+metaworld = ["metaworld>=3.0.0"]
 
 # All
 all = [
@@ -144,7 +147,7 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    "lerobot[pi0]",
+    "lerobot[pi]",
     "lerobot[smolvla]",
     "lerobot[hilserl]",
     "lerobot[async]",
@@ -153,19 +156,26 @@ all = [
     "lerobot[video_benchmark]",
     "lerobot[aloha]",
     "lerobot[pusht]",
-    "lerobot[xarm]"
+    "lerobot[phone]",
+    "lerobot[libero]",
+    "lerobot[metaworld]",
 ]
 
 [project.scripts]
-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-calibrate="lerobot.scripts.lerobot_calibrate:main"
+lerobot-find-cameras="lerobot.scripts.lerobot_find_cameras:main"
+lerobot-find-port="lerobot.scripts.lerobot_find_port:main"
+lerobot-record="lerobot.scripts.lerobot_record:main"
+lerobot-replay="lerobot.scripts.lerobot_replay:main"
+lerobot-setup-motors="lerobot.scripts.lerobot_setup_motors:main"
+lerobot-teleoperate="lerobot.scripts.lerobot_teleoperate:main"
+lerobot-eval="lerobot.scripts.lerobot_eval:main"
+lerobot-train="lerobot.scripts.lerobot_train:main"
+lerobot-dataset-viz="lerobot.scripts.lerobot_dataset_viz:main"
+lerobot-info="lerobot.scripts.lerobot_info:main"
+lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
+lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
+lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
 
 # ---------------- Tool Configurations ----------------
 [tool.setuptools.packages.find]
@@ -192,7 +202,7 @@ 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" # "SIM", "A", "S", "D", "RUF", "UP"
+    "E", "W", "F", "I", "B", "C4", "T20", "N", "UP", "SIM" #, "A", "S", "D", "RUF"
 ]
 ignore = [
     "E501", # Line too long
@@ -258,8 +268,87 @@ default.extend-ignore-identifiers-re = [
 # color = true
 # paths = ["src/lerobot"]
 
-# [tool.mypy]
-# python_version = "3.10"
+# 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"
 # warn_return_any = true
 # warn_unused_configs = true
-# ignore_missing_imports = false
+# 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.*"
+# Enable type checking only for the envs module
+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

src/lerobot/__init__.py

@@ -57,7 +57,6 @@ available_tasks_per_env = {
         "AlohaTransferCube-v0",
     ],
     "pusht": ["PushT-v0"],
-    "xarm": ["XarmLift-v0"],
 }
 available_envs = list(available_tasks_per_env.keys())
 
@@ -75,16 +74,6 @@ 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 = [
@@ -195,7 +184,6 @@ 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/async_inference/configs.py

@@ -18,7 +18,8 @@ from dataclasses import dataclass, field
 import torch
 
 from lerobot.robots.config import RobotConfig
-from lerobot.scripts.server.constants import (
+
+from .constants import (
     DEFAULT_FPS,
     DEFAULT_INFERENCE_LATENCY,
     DEFAULT_OBS_QUEUE_TIMEOUT,
@@ -141,11 +142,6 @@ class RobotClientConfig:
         default=False, metadata={"help": "Visualize the action queue size"}
     )
 
-    # Verification configuration
-    verify_robot_cameras: bool = field(
-        default=True, metadata={"help": "Verify that the robot cameras match the policy cameras"}
-    )
-
     @property
     def environment_dt(self) -> float:
         """Environment time step, in seconds"""

src/lerobot/async_inference/constants.py

@@ -23,7 +23,7 @@ DEFAULT_INFERENCE_LATENCY = 1 / DEFAULT_FPS
 DEFAULT_OBS_QUEUE_TIMEOUT = 2
 
 # All action chunking policies
-SUPPORTED_POLICIES = ["act", "smolvla", "diffusion", "pi0", "tdmpc", "vqbet"]
+SUPPORTED_POLICIES = ["act", "smolvla", "diffusion", "tdmpc", "vqbet", "pi0", "pi05"]
 
 # TODO: Add all other robots
-SUPPORTED_ROBOTS = ["so100_follower", "so101_follower"]
+SUPPORTED_ROBOTS = ["so100_follower", "so101_follower", "bi_so100_follower"]

src/lerobot/async_inference/helpers.py

@@ -22,16 +22,22 @@ from pathlib import Path
 import torch
 
 from lerobot.configs.types import PolicyFeature
-from lerobot.constants import OBS_IMAGES, OBS_STATE
 from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
 
 # NOTE: Configs need to be loaded for the client to be able to instantiate the policy config
-from lerobot.policies import ACTConfig, DiffusionConfig, PI0Config, SmolVLAConfig, VQBeTConfig  # noqa: F401
+from lerobot.policies import (  # noqa: F401
+    ACTConfig,
+    DiffusionConfig,
+    PI0Config,
+    PI05Config,
+    SmolVLAConfig,
+    VQBeTConfig,
+)
 from lerobot.robots.robot import Robot
+from lerobot.utils.constants import OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.utils import init_logging
 
 Action = torch.Tensor
-ActionChunk = torch.Tensor
 
 # observation as received from the robot
 RawObservation = dict[str, torch.Tensor]
@@ -46,7 +52,7 @@ Observation = dict[str, torch.Tensor]
 def visualize_action_queue_size(action_queue_size: list[int]) -> None:
     import matplotlib.pyplot as plt
 
-    fig, ax = plt.subplots()
+    _, ax = plt.subplots()
     ax.set_title("Action Queue Size Over Time")
     ax.set_xlabel("Environment steps")
     ax.set_ylabel("Action Queue Size")
@@ -56,17 +62,8 @@ def visualize_action_queue_size(action_queue_size: list[int]) -> None:
     plt.show()
 
 
-def validate_robot_cameras_for_policy(
-    lerobot_observation_features: dict[str, dict], policy_image_features: dict[str, PolicyFeature]
-) -> None:
-    image_keys = list(filter(is_image_key, lerobot_observation_features))
-    assert set(image_keys) == set(policy_image_features.keys()), (
-        f"Policy image features must match robot cameras! Received {list(policy_image_features.keys())} != {image_keys}"
-    )
-
-
 def map_robot_keys_to_lerobot_features(robot: Robot) -> dict[str, dict]:
-    return hw_to_dataset_features(robot.observation_features, "observation", use_video=False)
+    return hw_to_dataset_features(robot.observation_features, OBS_STR, use_video=False)
 
 
 def is_image_key(k: str) -> bool:
@@ -86,11 +83,11 @@ def resize_robot_observation_image(image: torch.tensor, resize_dims: tuple[int,
     return resized.squeeze(0)
 
 
+# TODO(Steven): Consider implementing a pipeline step for this
 def raw_observation_to_observation(
     raw_observation: RawObservation,
     lerobot_features: dict[str, dict],
     policy_image_features: dict[str, PolicyFeature],
-    device: str,
 ) -> Observation:
     observation = {}
 
@@ -99,9 +96,7 @@ def raw_observation_to_observation(
         if isinstance(v, torch.Tensor):  # VLAs present natural-language instructions in observations
             if "image" in k:
                 # Policy expects images in shape (B, C, H, W)
-                observation[k] = prepare_image(v).unsqueeze(0).to(device)
-            else:
-                observation[k] = v.to(device)
+                observation[k] = prepare_image(v).unsqueeze(0)
         else:
             observation[k] = v
 
@@ -141,7 +136,7 @@ def make_lerobot_observation(
     lerobot_features: dict[str, dict],
 ) -> LeRobotObservation:
     """Make a lerobot observation from a raw observation."""
-    return build_dataset_frame(lerobot_features, robot_obs, prefix="observation")
+    return build_dataset_frame(lerobot_features, robot_obs, prefix=OBS_STR)
 
 
 def prepare_raw_observation(

src/lerobot/async_inference/policy_server.py

@@ -15,7 +15,7 @@
 """
 Example:
 ```shell
-python src/lerobot/scripts/server/policy_server.py \
+python -m lerobot.async_inference.policy_server \
      --host=127.0.0.1 \
      --port=8080 \
      --fps=30 \
@@ -32,15 +32,26 @@ from concurrent import futures
 from dataclasses import asdict
 from pprint import pformat
 from queue import Empty, Queue
+from typing import Any
 
 import draccus
 import grpc
 import torch
 
-from lerobot.policies.factory import get_policy_class
-from lerobot.scripts.server.configs import PolicyServerConfig
-from lerobot.scripts.server.constants import SUPPORTED_POLICIES
-from lerobot.scripts.server.helpers import (
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.processor import (
+    PolicyAction,
+    PolicyProcessorPipeline,
+)
+from lerobot.transport import (
+    services_pb2,  # type: ignore
+    services_pb2_grpc,  # type: ignore
+)
+from lerobot.transport.utils import receive_bytes_in_chunks
+
+from .configs import PolicyServerConfig
+from .constants import SUPPORTED_POLICIES
+from .helpers import (
     FPSTracker,
     Observation,
     RemotePolicyConfig,
@@ -50,11 +61,6 @@ from lerobot.scripts.server.helpers import (
     observations_similar,
     raw_observation_to_observation,
 )
-from lerobot.transport import (
-    services_pb2,  # type: ignore
-    services_pb2_grpc,  # type: ignore
-)
-from lerobot.transport.utils import receive_bytes_in_chunks
 
 
 class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
@@ -81,6 +87,8 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
         self.lerobot_features = None
         self.actions_per_chunk = None
         self.policy = None
+        self.preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]] | None = None
+        self.postprocessor: PolicyProcessorPipeline[PolicyAction, PolicyAction] | None = None
 
     @property
     def running(self):
@@ -145,6 +153,16 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
         start = time.perf_counter()
         self.policy = policy_class.from_pretrained(policy_specs.pretrained_name_or_path)
         self.policy.to(self.device)
+
+        # Load preprocessor and postprocessor, overriding device to match requested device
+        device_override = {"device": self.device}
+        self.preprocessor, self.postprocessor = make_pre_post_processors(
+            self.policy.config,
+            pretrained_path=policy_specs.pretrained_name_or_path,
+            preprocessor_overrides={"device_processor": device_override},
+            postprocessor_overrides={"device_processor": device_override},
+        )
+
         end = time.perf_counter()
 
         self.logger.info(f"Time taken to put policy on {self.device}: {end - start:.4f} seconds")
@@ -172,7 +190,7 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
         # Calculate FPS metrics
         fps_metrics = self.fps_tracker.calculate_fps_metrics(obs_timestamp)
 
-        self.logger.info(
+        self.logger.debug(
             f"Received observation #{obs_timestep} | "
             f"Avg FPS: {fps_metrics['avg_fps']:.2f} | "  # fps at which observations are received from client
             f"Target: {fps_metrics['target_fps']:.2f} | "
@@ -188,7 +206,7 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
         if not self._enqueue_observation(
             timed_observation  # wrapping a RawObservation
         ):
-            self.logger.info(f"Observation #{obs_timestep} has been filtered out")
+            self.logger.debug(f"Observation #{obs_timestep} has been filtered out")
 
         return services_pb2.Empty()
 
@@ -300,23 +318,6 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
             for i, action in enumerate(action_chunk)
         ]
 
-    def _prepare_observation(self, observation_t: TimedObservation) -> Observation:
-        """
-        Prepare observation, ready for policy inference.
-        E.g.: To keep observation sampling rate high (and network packet tiny) we send int8 [0,255] images from the
-        client and then convert them to float32 [0,1] images here, before running inference.
-        """
-        # RawObservation from robot.get_observation() - wrong keys, wrong dtype, wrong image shape
-        observation: Observation = raw_observation_to_observation(
-            observation_t.get_observation(),
-            self.lerobot_features,
-            self.policy_image_features,
-            self.device,
-        )
-        # processed Observation - right keys, right dtype, right image shape
-
-        return observation
-
     def _get_action_chunk(self, observation: dict[str, torch.Tensor]) -> torch.Tensor:
         """Get an action chunk from the policy. The chunk contains only"""
         chunk = self.policy.predict_action_chunk(observation)
@@ -326,44 +327,76 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
         return chunk[:, : self.actions_per_chunk, :]
 
     def _predict_action_chunk(self, observation_t: TimedObservation) -> list[TimedAction]:
-        """Predict an action chunk based on an observation"""
-        inference_starts = time.perf_counter()
+        """Predict an action chunk based on an observation.
 
+        Pipeline:
+        1. Convert raw observation to LeRobot format
+        2. Apply preprocessor (tokenization, normalization, batching, device placement)
+        3. Run policy inference to get action chunk
+        4. Apply postprocessor (unnormalization, device movement)
+        5. Convert to TimedAction list
+        """
         """1. Prepare observation"""
-        start_time = time.perf_counter()
-        observation = self._prepare_observation(observation_t)
-        preprocessing_time = time.perf_counter() - start_time
+        start_prepare = time.perf_counter()
+        observation: Observation = raw_observation_to_observation(
+            observation_t.get_observation(),
+            self.lerobot_features,
+            self.policy_image_features,
+        )
+        prepare_time = time.perf_counter() - start_prepare
 
+        """2. Apply preprocessor"""
+        start_preprocess = time.perf_counter()
+        observation = self.preprocessor(observation)
         self.last_processed_obs: TimedObservation = observation_t
+        preprocessing_time = time.perf_counter() - start_preprocess
 
-        """2. Get action chunk"""
-        start_time = time.perf_counter()
+        """3. Get action chunk"""
+        start_inference = time.perf_counter()
         action_tensor = self._get_action_chunk(observation)
-        inference_time = time.perf_counter() - start_time
+        inference_time = time.perf_counter() - start_inference
+        self.logger.info(
+            f"Preprocessing and inference took {inference_time:.4f}s, action shape: {action_tensor.shape}"
+        )
 
-        """3. Post-inference processing"""
-        start_time = time.perf_counter()
-        # Move to CPU before serializing
-        action_tensor = action_tensor.cpu().squeeze(0)
+        """4. Apply postprocessor"""
+        # Apply postprocessor (handles unnormalization and device movement)
+        # Postprocessor expects (B, action_dim) per action, but we have (B, chunk_size, action_dim)
+        # So we process each action in the chunk individually
+        start_postprocess = time.perf_counter()
+        _, chunk_size, _ = action_tensor.shape
 
+        # Process each action in the chunk
+        processed_actions = []
+        for i in range(chunk_size):
+            # Extract action at timestep i: (B, action_dim)
+            single_action = action_tensor[:, i, :]
+            processed_action = self.postprocessor(single_action)
+            processed_actions.append(processed_action)
+
+        # Stack back to (B, chunk_size, action_dim), then remove batch dim
+        action_tensor = torch.stack(processed_actions, dim=1).squeeze(0)
+        self.logger.debug(f"Postprocessed action shape: {action_tensor.shape}")
+
+        """5. Convert to TimedAction list"""
         action_chunk = self._time_action_chunk(
             observation_t.get_timestamp(), list(action_tensor), observation_t.get_timestep()
         )
-        postprocessing_time = time.perf_counter() - start_time
-        inference_stops = time.perf_counter()
+        postprocess_stops = time.perf_counter()
+        postprocessing_time = postprocess_stops - start_postprocess
 
         self.logger.info(
-            f"Observation {observation_t.get_timestep()} |"
-            f"Inference time: {1000 * (inference_stops - inference_starts):.2f}ms"
+            f"Observation {observation_t.get_timestep()} | "
+            f"Total time: {1000 * (postprocess_stops - start_prepare):.2f}ms"
         )
 
-        # full-process latency breakdown for debugging purposes
         self.logger.debug(
             f"Observation {observation_t.get_timestep()} | "
-            f"Preprocessing time: {1000 * (preprocessing_time - inference_starts):.2f}ms | "
-            f"Inference time: {1000 * (inference_time - preprocessing_time):.2f}ms | "
-            f"Postprocessing time: {1000 * (postprocessing_time - inference_time):.2f}ms | "
-            f"Total time: {1000 * (postprocessing_time - inference_starts):.2f}ms"
+            f"Prepare time: {1000 * prepare_time:.2f}ms | "
+            f"Preprocessing time: {1000 * preprocessing_time:.2f}ms | "
+            f"Inference time: {1000 * inference_time:.2f}ms | "
+            f"Postprocessing time: {1000 * postprocessing_time:.2f}ms | "
+            f"Total time: {1000 * (postprocess_stops - start_prepare):.2f}ms"
         )
 
         return action_chunk

src/lerobot/async_inference/robot_client.py

@@ -15,7 +15,7 @@
 """
 Example command:
 ```shell
-python src/lerobot/scripts/server/robot_client.py \
+python src/lerobot/async_inference/robot_client.py \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
@@ -48,18 +48,24 @@ import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.configs.policies import PreTrainedConfig
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
+    bi_so100_follower,
     koch_follower,
     make_robot_from_config,
     so100_follower,
     so101_follower,
 )
-from lerobot.scripts.server.configs import RobotClientConfig
-from lerobot.scripts.server.constants import SUPPORTED_ROBOTS
-from lerobot.scripts.server.helpers import (
+from lerobot.transport import (
+    services_pb2,  # type: ignore
+    services_pb2_grpc,  # type: ignore
+)
+from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
+
+from .configs import RobotClientConfig
+from .constants import SUPPORTED_ROBOTS
+from .helpers import (
     Action,
     FPSTracker,
     Observation,
@@ -69,14 +75,8 @@ from lerobot.scripts.server.helpers import (
     TimedObservation,
     get_logger,
     map_robot_keys_to_lerobot_features,
-    validate_robot_cameras_for_policy,
     visualize_action_queue_size,
 )
-from lerobot.transport import (
-    services_pb2,  # type: ignore
-    services_pb2_grpc,  # type: ignore
-)
-from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
 
 
 class RobotClient:
@@ -96,14 +96,6 @@ class RobotClient:
 
         lerobot_features = map_robot_keys_to_lerobot_features(self.robot)
 
-        if config.verify_robot_cameras:
-            # Load policy config for validation
-            policy_config = PreTrainedConfig.from_pretrained(config.pretrained_name_or_path)
-            policy_image_features = policy_config.image_features
-
-            # The cameras specified for inference must match the one supported by the policy chosen
-            validate_robot_cameras_for_policy(lerobot_features, policy_image_features)
-
         # Use environment variable if server_address is not provided in config
         self.server_address = config.server_address
 
@@ -213,7 +205,7 @@ class RobotClient:
             )
             _ = self.stub.SendObservations(observation_iterator)
             obs_timestep = obs.get_timestep()
-            self.logger.info(f"Sent observation #{obs_timestep} | ")
+            self.logger.debug(f"Sent observation #{obs_timestep} | ")
 
             return True
 
@@ -466,7 +458,7 @@ class RobotClient:
             if self._ready_to_send_observation():
                 _captured_observation = self.control_loop_observation(task, verbose)
 
-            self.logger.info(f"Control loop (ms): {(time.perf_counter() - control_loop_start) * 1000:.2f}")
+            self.logger.debug(f"Control loop (ms): {(time.perf_counter() - control_loop_start) * 1000:.2f}")
             # Dynamically adjust sleep time to maintain the desired control frequency
             time.sleep(max(0, self.config.environment_dt - (time.perf_counter() - control_loop_start)))
 

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.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.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.errors import DeviceNotConnectedError
+from lerobot.utils.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.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+from lerobot.utils.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 pathlib import Path
-from typing import TypeAlias
+from typing import cast
+
+from lerobot.utils.import_utils import make_device_from_device_class
 
 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 = {}
+    cameras: dict[str, Camera] = {}
 
     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
 
@@ -44,7 +44,10 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[s
             cameras[key] = Reachy2Camera(cfg)
 
         else:
-            raise ValueError(f"The camera type '{cfg.type}' is not valid.")
+            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
 
     return cameras
 

src/lerobot/configs/default.py

@@ -16,9 +16,6 @@
 
 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

@@ -26,10 +26,10 @@ from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError
 
-from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
-from lerobot.constants import ACTION, OBS_STATE
+from lerobot.configs.types import FeatureType, PolicyFeature
 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
 
@@ -53,7 +53,6 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     """
 
     n_obs_steps: int = 1
-    normalization_mapping: dict[str, NormalizationMode] = field(default_factory=dict)
 
     input_features: dict[str, PolicyFeature] = field(default_factory=dict)
     output_features: dict[str, PolicyFeature] = field(default_factory=dict)
@@ -72,9 +71,11 @@ 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}'.")
@@ -197,11 +198,10 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
             config = json.load(f)
 
         config.pop("type")
-        with tempfile.NamedTemporaryFile("w+") as f:
+        with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
             json.dump(config, f)
             config_file = f.name
-            f.flush()
 
-            cli_overrides = policy_kwargs.pop("cli_overrides", [])
-            with draccus.config_type("json"):
-                return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)
+        cli_overrides = policy_kwargs.pop("cli_overrides", [])
+        with draccus.config_type("json"):
+            return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)

src/lerobot/configs/types.py

@@ -15,7 +15,6 @@
 # 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):
@@ -24,16 +23,20 @@ class FeatureType(str, Enum):
     ENV = "ENV"
     ACTION = "ACTION"
     REWARD = "REWARD"
+    LANGUAGE = "LANGUAGE"
+
+
+class PipelineFeatureType(str, Enum):
+    ACTION = "ACTION"
+    OBSERVATION = "OBSERVATION"
 
 
 class NormalizationMode(str, Enum):
     MIN_MAX = "MIN_MAX"
     MEAN_STD = "MEAN_STD"
     IDENTITY = "IDENTITY"
-
-
-class DictLike(Protocol):
-    def __getitem__(self, key: Any) -> Any: ...
+    QUANTILES = "QUANTILES"
+    QUANTILE10 = "QUANTILE10"
 
 
 @dataclass

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
+from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
 
 
 def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
@@ -93,14 +93,13 @@ def update_data_df(df, src_meta, dst_meta):
         pd.DataFrame: Updated DataFrame with adjusted indices.
     """
 
-    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
+    df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
+    df["index"] = df["index"] + dst_meta.info["total_frames"]
 
-    return df.apply(_update, axis=1)
+    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
 
 
 def update_meta_data(
@@ -126,27 +125,45 @@ def update_meta_data(
         pd.DataFrame: Updated DataFrame with adjusted indices and timestamps.
     """
 
-    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["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()
 
-        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
+        # Update chunk and file indices to point to destination
+        df[orig_chunk_col] = video_idx["chunk"]
+        df[orig_file_col] = video_idx["file"]
 
-    return df.apply(_update, axis=1)
+        # 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"]
+            )
+            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"])
+
+    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
 
 
 def aggregate_datasets(
@@ -200,6 +217,10 @@ 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")
@@ -243,6 +264,11 @@ 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)
@@ -256,6 +282,7 @@ 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(
@@ -270,21 +297,25 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                 file_index=file_idx,
             )
 
-            # If a new file is created, we don't want to increment the latest_duration
-            update_latest_duration = False
+            src_duration = get_video_duration_in_s(src_path)
 
             if not dst_path.exists():
-                # First write to this destination file
+                # Store offset before incrementing
+                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
                 dst_path.parent.mkdir(parents=True, exist_ok=True)
                 shutil.copy(str(src_path), str(dst_path))
-                continue  # not accumulating further, already copied the file in place
+                videos_idx[key]["episode_duration"] += src_duration
+                current_offset += src_duration
+                continue
 
             # Check file sizes before appending
-            src_size = get_video_size_in_mb(src_path)
-            dst_size = get_video_size_in_mb(dst_path)
+            src_size = get_file_size_in_mb(src_path)
+            dst_size = get_file_size_in_mb(dst_path)
 
             if dst_size + src_size >= video_files_size_in_mb:
-                # Rotate to a new chunk/file
+                # 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
                 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,
@@ -293,25 +324,22 @@ 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:
-                # Get the timestamps shift for this video
-                timestamps_shift_s = dst_meta.info["total_frames"] / dst_meta.info["fps"]
-
-                # Append to existing video file
+                # Append to existing video file - use current accumulated offset
+                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
                 concatenate_video_files(
                     [dst_path, src_path],
                     dst_path,
                 )
-                # Update the latest_duration when appending (shifts timestamps!)
-                update_latest_duration = not update_latest_duration
+                current_offset += src_duration
+
+            videos_idx[key]["episode_duration"] += src_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
 
 
@@ -396,9 +424,6 @@ 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,
@@ -410,6 +435,10 @@ 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

@@ -23,6 +23,9 @@ 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

@@ -17,6 +17,179 @@ 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
@@ -72,33 +245,282 @@ def sample_images(image_paths: list[str]) -> np.ndarray:
     return images
 
 
-def get_feature_stats(array: np.ndarray, axis: tuple, keepdims: bool) -> dict[str, np.ndarray]:
-    return {
-        "min": np.min(array, axis=axis, keepdims=keepdims),
-        "max": np.max(array, axis=axis, keepdims=keepdims),
-        "mean": np.mean(array, axis=axis, keepdims=keepdims),
-        "std": np.std(array, axis=axis, keepdims=keepdims),
-        "count": np.array([len(array)]),
+def _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]),
     }
 
+    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  # 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
+            continue
+
+        if features[key]["dtype"] in ["image", "video"]:
+            ep_ft_array = sample_images(data)
+            axes_to_reduce = (0, 2, 3)
             keepdims = True
         else:
-            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_ft_array = data
+            axes_to_reduce = 0
+            keepdims = data.ndim == 1
 
-        ep_stats[key] = get_feature_stats(ep_ft_array, axis=axes_to_reduce, keepdims=keepdims)
+        ep_stats[key] = get_feature_stats(
+            ep_ft_array, axis=axes_to_reduce, keepdims=keepdims, quantile_list=quantile_list
+        )
 
-        # 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()
@@ -107,20 +529,37 @@ def compute_episode_stats(episode_data: dict[str, list[str] | np.ndarray], featu
     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]]):
-    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.")
+    """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)
 
 
 def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
@@ -143,7 +582,7 @@ 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
 
-    return {
+    aggregated = {
         "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,
@@ -151,6 +590,17 @@ 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.

src/lerobot/datasets/factory.py

@@ -27,6 +27,7 @@ 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)
@@ -54,11 +55,11 @@ def resolve_delta_timestamps(
     """
     delta_timestamps = {}
     for key in ds_meta.features:
-        if key == "next.reward" and cfg.reward_delta_indices is not None:
+        if key == 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("observation.") and cfg.observation_delta_indices is not None:
+        if key.startswith(OBS_PREFIX) 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

@@ -68,7 +68,30 @@ 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):
+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.
+    """
     try:
         if isinstance(image, np.ndarray):
             img = image_array_to_pil_image(image)
@@ -76,7 +99,7 @@ def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path):
             img = image
         else:
             raise TypeError(f"Unsupported image type: {type(image)}")
-        img.save(fpath)
+        img.save(fpath, compress_level=compress_level)
     except Exception as e:
         print(f"Error writing image {fpath}: {e}")
 

src/lerobot/datasets/lerobot_dataset.py

@@ -14,7 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import contextlib
-import gc
 import logging
 import shutil
 import tempfile
@@ -26,12 +25,13 @@ import numpy as np
 import packaging.version
 import pandas as pd
 import PIL.Image
+import pyarrow as pa
+import pyarrow.parquet as pq
 import torch
 import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
-from lerobot.constants import HF_LEROBOT_HOME
 from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats
 from lerobot.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.datasets.utils import (
@@ -47,13 +47,9 @@ from lerobot.datasets.utils import (
     embed_images,
     flatten_dict,
     get_delta_indices,
-    get_hf_dataset_cache_dir,
-    get_hf_dataset_size_in_mb,
+    get_file_size_in_mb,
     get_hf_features_from_features,
-    get_parquet_file_size_in_mb,
-    get_parquet_num_frames,
     get_safe_version,
-    get_video_size_in_mb,
     hf_transform_to_torch,
     is_valid_version,
     load_episodes,
@@ -61,7 +57,6 @@ from lerobot.datasets.utils import (
     load_nested_dataset,
     load_stats,
     load_tasks,
-    to_parquet_with_hf_images,
     update_chunk_file_indices,
     validate_episode_buffer,
     validate_frame,
@@ -79,6 +74,7 @@ from lerobot.datasets.video_utils import (
     get_video_duration_in_s,
     get_video_info,
 )
+from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
 
@@ -90,10 +86,15 @@ class LeRobotDatasetMetadata:
         root: str | Path | None = None,
         revision: str | None = None,
         force_cache_sync: bool = False,
+        metadata_buffer_size: int = 10,
     ):
         self.repo_id = repo_id
         self.revision = revision if revision else CODEBASE_VERSION
         self.root = Path(root) if root is not None else HF_LEROBOT_HOME / repo_id
+        self.writer = None
+        self.latest_episode = None
+        self.metadata_buffer: list[dict] = []
+        self.metadata_buffer_size = metadata_buffer_size
 
         try:
             if force_cache_sync:
@@ -107,6 +108,54 @@ class LeRobotDatasetMetadata:
             self.pull_from_repo(allow_patterns="meta/")
             self.load_metadata()
 
+    def _flush_metadata_buffer(self) -> None:
+        """Write all buffered episode metadata to parquet file."""
+        if not hasattr(self, "metadata_buffer") or len(self.metadata_buffer) == 0:
+            return
+
+        combined_dict = {}
+        for episode_dict in self.metadata_buffer:
+            for key, value in episode_dict.items():
+                if key not in combined_dict:
+                    combined_dict[key] = []
+                # Extract value and serialize numpy arrays
+                # because PyArrow's from_pydict function doesn't support numpy arrays
+                val = value[0] if isinstance(value, list) else value
+                combined_dict[key].append(val.tolist() if isinstance(val, np.ndarray) else val)
+
+        first_ep = self.metadata_buffer[0]
+        chunk_idx = first_ep["meta/episodes/chunk_index"][0]
+        file_idx = first_ep["meta/episodes/file_index"][0]
+
+        table = pa.Table.from_pydict(combined_dict)
+
+        if not self.writer:
+            path = Path(self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx))
+            path.parent.mkdir(parents=True, exist_ok=True)
+            self.writer = pq.ParquetWriter(
+                path, schema=table.schema, compression="snappy", use_dictionary=True
+            )
+
+        self.writer.write_table(table)
+
+        self.latest_episode = self.metadata_buffer[-1]
+        self.metadata_buffer.clear()
+
+    def _close_writer(self) -> None:
+        """Close and cleanup the parquet writer if it exists."""
+        self._flush_metadata_buffer()
+
+        writer = getattr(self, "writer", None)
+        if writer is not None:
+            writer.close()
+            self.writer = None
+
+    def __del__(self):
+        """
+        Trust the user to call .finalize() but as an added safety check call the parquet writer to stop when calling the destructor
+        """
+        self._close_writer()
+
     def load_metadata(self):
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
@@ -138,6 +187,12 @@ class LeRobotDatasetMetadata:
         return packaging.version.parse(self.info["codebase_version"])
 
     def get_data_file_path(self, ep_index: int) -> Path:
+        if self.episodes is None:
+            self.episodes = load_episodes(self.root)
+        if ep_index >= len(self.episodes):
+            raise IndexError(
+                f"Episode index {ep_index} out of range. Episodes: {len(self.episodes) if self.episodes else 0}"
+            )
         ep = self.episodes[ep_index]
         chunk_idx = ep["data/chunk_index"]
         file_idx = ep["data/file_index"]
@@ -145,6 +200,12 @@ class LeRobotDatasetMetadata:
         return Path(fpath)
 
     def get_video_file_path(self, ep_index: int, vid_key: str) -> Path:
+        if self.episodes is None:
+            self.episodes = load_episodes(self.root)
+        if ep_index >= len(self.episodes):
+            raise IndexError(
+                f"Episode index {ep_index} out of range. Episodes: {len(self.episodes) if self.episodes else 0}"
+            )
         ep = self.episodes[ep_index]
         chunk_idx = ep[f"videos/{vid_key}/chunk_index"]
         file_idx = ep[f"videos/{vid_key}/file_index"]
@@ -260,72 +321,75 @@ class LeRobotDatasetMetadata:
             write_tasks(self.tasks, self.root)
 
     def _save_episode_metadata(self, episode_dict: dict) -> None:
-        """Save episode metadata to a parquet file and update the Hugging Face dataset of episodes metadata.
+        """Buffer episode metadata and write to parquet in batches for efficiency.
 
-        This function processes episodes metadata from a dictionary, converts it into a Hugging Face dataset,
-        and saves it as a parquet file. It handles both the creation of new parquet files and the
-        updating of existing ones based on size constraints. After saving the metadata, it reloads
-        the Hugging Face dataset to ensure it is up-to-date.
+        This function accumulates episode metadata in a buffer and flushes it when the buffer
+        reaches the configured size. This reduces I/O overhead by writing multiple episodes
+        at once instead of one row at a time.
 
         Notes: We both need to update parquet files and HF dataset:
         - `pandas` loads parquet file in RAM
         - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
           or loads directly from pyarrow cache.
         """
-        # Convert buffer into HF Dataset
+        # Convert to list format for each value
         episode_dict = {key: [value] for key, value in episode_dict.items()}
-        ep_dataset = datasets.Dataset.from_dict(episode_dict)
-        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
-        df = pd.DataFrame(ep_dataset)
         num_frames = episode_dict["length"][0]
 
-        if self.episodes is None:
+        if self.latest_episode is None:
             # Initialize indices and frame count for a new dataset made of the first episode data
             chunk_idx, file_idx = 0, 0
-            df["meta/episodes/chunk_index"] = [chunk_idx]
-            df["meta/episodes/file_index"] = [file_idx]
-            df["dataset_from_index"] = [0]
-            df["dataset_to_index"] = [num_frames]
-        else:
-            # Retrieve information from the latest parquet file
-            latest_ep = self.episodes[-1]
-            chunk_idx = latest_ep["meta/episodes/chunk_index"]
-            file_idx = latest_ep["meta/episodes/file_index"]
+            if self.episodes is not None and len(self.episodes) > 0:
+                # It means we are resuming recording, so we need to load the latest episode
+                # Update the indices to avoid overwriting the latest episode
+                chunk_idx = self.episodes[-1]["meta/episodes/chunk_index"]
+                file_idx = self.episodes[-1]["meta/episodes/file_index"]
+                latest_num_frames = self.episodes[-1]["dataset_to_index"]
+                episode_dict["dataset_from_index"] = [latest_num_frames]
+                episode_dict["dataset_to_index"] = [latest_num_frames + num_frames]
 
-            latest_path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
-
-            if latest_size_in_mb + ep_size_in_mb >= self.data_files_size_in_mb:
-                # Size limit is reached, prepare new parquet file
+                # When resuming, move to the next file
                 chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
+            else:
+                episode_dict["dataset_from_index"] = [0]
+                episode_dict["dataset_to_index"] = [num_frames]
+
+            episode_dict["meta/episodes/chunk_index"] = [chunk_idx]
+            episode_dict["meta/episodes/file_index"] = [file_idx]
+        else:
+            chunk_idx = self.latest_episode["meta/episodes/chunk_index"][0]
+            file_idx = self.latest_episode["meta/episodes/file_index"][0]
+
+            latest_path = (
+                self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+                if self.writer is None
+                else self.writer.where
+            )
+
+            if Path(latest_path).exists():
+                latest_size_in_mb = get_file_size_in_mb(Path(latest_path))
+                latest_num_frames = self.latest_episode["episode_index"][0]
+
+                av_size_per_frame = latest_size_in_mb / latest_num_frames if latest_num_frames > 0 else 0.0
+
+                if latest_size_in_mb + av_size_per_frame * num_frames >= self.data_files_size_in_mb:
+                    # Size limit is reached, flush buffer and prepare new parquet file
+                    self._flush_metadata_buffer()
+                    chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
+                    self._close_writer()
 
             # Update the existing pandas dataframe with new row
-            df["meta/episodes/chunk_index"] = [chunk_idx]
-            df["meta/episodes/file_index"] = [file_idx]
-            df["dataset_from_index"] = [latest_ep["dataset_to_index"]]
-            df["dataset_to_index"] = [latest_ep["dataset_to_index"] + num_frames]
+            episode_dict["meta/episodes/chunk_index"] = [chunk_idx]
+            episode_dict["meta/episodes/file_index"] = [file_idx]
+            episode_dict["dataset_from_index"] = [self.latest_episode["dataset_to_index"][0]]
+            episode_dict["dataset_to_index"] = [self.latest_episode["dataset_to_index"][0] + num_frames]
 
-            if latest_size_in_mb + ep_size_in_mb < self.data_files_size_in_mb:
-                # Size limit wasnt reached, concatenate latest dataframe with new one
-                latest_df = pd.read_parquet(latest_path)
-                df = pd.concat([latest_df, df], ignore_index=True)
+        # Add to buffer
+        self.metadata_buffer.append(episode_dict)
+        self.latest_episode = episode_dict
 
-                # Memort optimization
-                del latest_df
-                gc.collect()
-
-        # Write the resulting dataframe from RAM to disk
-        path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
-        path.parent.mkdir(parents=True, exist_ok=True)
-        df.to_parquet(path, index=False)
-
-        if self.episodes is not None:
-            # Remove the episodes cache directory, necessary to avoid cache bloat
-            cached_dir = get_hf_dataset_cache_dir(self.episodes)
-            if cached_dir is not None:
-                shutil.rmtree(cached_dir)
-
-        self.episodes = load_episodes(self.root)
+        if len(self.metadata_buffer) >= self.metadata_buffer_size:
+            self._flush_metadata_buffer()
 
     def save_episode(
         self,
@@ -438,6 +502,10 @@ class LeRobotDatasetMetadata:
         robot_type: str | None = None,
         root: str | Path | None = None,
         use_videos: bool = True,
+        metadata_buffer_size: int = 10,
+        chunks_size: int | None = None,
+        data_files_size_in_mb: int | None = None,
+        video_files_size_in_mb: int | None = None,
     ) -> "LeRobotDatasetMetadata":
         """Creates metadata for a LeRobotDataset."""
         obj = cls.__new__(cls)
@@ -452,11 +520,24 @@ class LeRobotDatasetMetadata:
         obj.tasks = None
         obj.episodes = None
         obj.stats = None
-        obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, features, use_videos, robot_type)
+        obj.info = create_empty_dataset_info(
+            CODEBASE_VERSION,
+            fps,
+            features,
+            use_videos,
+            robot_type,
+            chunks_size,
+            data_files_size_in_mb,
+            video_files_size_in_mb,
+        )
         if len(obj.video_keys) > 0 and not use_videos:
             raise ValueError()
         write_json(obj.info, obj.root / INFO_PATH)
         obj.revision = None
+        obj.writer = None
+        obj.latest_episode = None
+        obj.metadata_buffer = []
+        obj.metadata_buffer_size = metadata_buffer_size
         return obj
 
 
@@ -603,6 +684,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # Unused attributes
         self.image_writer = None
         self.episode_buffer = None
+        self.writer = None
+        self.latest_episode = None
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -611,6 +694,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
             self.repo_id, self.root, self.revision, force_cache_sync=force_cache_sync
         )
 
+        # Track dataset state for efficient incremental writing
+        self._lazy_loading = False
+        self._recorded_frames = self.meta.total_frames
+        self._writer_closed_for_reading = False
+
         # Load actual data
         try:
             if force_cache_sync:
@@ -629,6 +717,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
+    def _close_writer(self) -> None:
+        """Close and cleanup the parquet writer if it exists."""
+        writer = getattr(self, "writer", None)
+        if writer is not None:
+            writer.close()
+            self.writer = None
+
+    def __del__(self):
+        """
+        Trust the user to call .finalize() but as an added safety check call the parquet writer to stop when calling the destructor
+        """
+        self._close_writer()
+
     def push_to_hub(
         self,
         branch: str | None = None,
@@ -769,8 +870,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     @property
     def num_frames(self) -> int:
-        """Number of frames in selected episodes."""
-        return len(self.hf_dataset) if self.hf_dataset is not None else self.meta.total_frames
+        """Number of frames in selected episodes.
+
+        Note: When episodes a subset of the full dataset is requested, we must return the
+        actual loaded data length (len(self.hf_dataset)) rather than metadata total_frames.
+        self.meta.total_frames is the total number of frames in the full dataset.
+        """
+        if self.episodes is not None and self.hf_dataset is not None:
+            return len(self.hf_dataset)
+        return self.meta.total_frames
 
     @property
     def num_episodes(self) -> int:
@@ -848,15 +956,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         return item
 
-    def _add_padding_keys(self, item: dict, padding: dict[str, list[bool]]) -> dict:
-        for key, val in padding.items():
-            item[key] = torch.BoolTensor(val)
-        return item
+    def _ensure_hf_dataset_loaded(self):
+        """Lazy load the HF dataset only when needed for reading."""
+        if self._lazy_loading or self.hf_dataset is None:
+            # Close the writer before loading to ensure parquet file is properly finalized
+            if self.writer is not None:
+                self._close_writer()
+                self._writer_closed_for_reading = True
+            self.hf_dataset = self.load_hf_dataset()
+            self._lazy_loading = False
 
     def __len__(self):
         return self.num_frames
 
     def __getitem__(self, idx) -> dict:
+        # Ensure dataset is loaded when we actually need to read from it
+        self._ensure_hf_dataset_loaded()
         item = self.hf_dataset[idx]
         ep_idx = item["episode_index"].item()
 
@@ -895,6 +1010,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
             "})',\n"
         )
 
+    def finalize(self):
+        """
+        Close the parquet writers. This function needs to be called after data collection/conversion, else footer metadata won't be written to the parquet files.
+        The dataset won't be valid and can't be loaded as ds = LeRobotDataset(repo_id=repo, root=HF_LEROBOT_HOME.joinpath(repo))
+        """
+        self._close_writer()
+        self.meta._close_writer()
+
     def create_episode_buffer(self, episode_index: int | None = None) -> dict:
         current_ep_idx = self.meta.total_episodes if episode_index is None else episode_index
         ep_buffer = {}
@@ -1032,7 +1155,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Reset episode buffer and clean up temporary images (if not already deleted during video encoding)
             self.clear_episode_buffer(delete_images=len(self.meta.image_keys) > 0)
 
-    def _batch_save_episode_video(self, start_episode: int, end_episode: int | None = None):
+    def _batch_save_episode_video(self, start_episode: int, end_episode: int | None = None) -> None:
         """
         Batch save videos for multiple episodes.
 
@@ -1102,74 +1225,101 @@ class LeRobotDataset(torch.utils.data.Dataset):
         ep_dict = {key: episode_buffer[key] for key in self.hf_features}
         ep_dataset = datasets.Dataset.from_dict(ep_dict, features=self.hf_features, split="train")
         ep_dataset = embed_images(ep_dataset)
-        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
         ep_num_frames = len(ep_dataset)
-        df = pd.DataFrame(ep_dataset)
 
-        if self.meta.episodes is None:
+        if self.latest_episode is None:
             # Initialize indices and frame count for a new dataset made of the first episode data
             chunk_idx, file_idx = 0, 0
-            latest_num_frames = 0
+            global_frame_index = 0
+            # However, if the episodes already exists
+            # It means we are resuming recording, so we need to load the latest episode
+            # Update the indices to avoid overwriting the latest episode
+            if self.meta.episodes is not None and len(self.meta.episodes) > 0:
+                latest_ep = self.meta.episodes[-1]
+                global_frame_index = latest_ep["dataset_to_index"]
+                chunk_idx = latest_ep["data/chunk_index"]
+                file_idx = latest_ep["data/file_index"]
+
+                # When resuming, move to the next file
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
         else:
             # Retrieve information from the latest parquet file
-            latest_ep = self.meta.episodes[-1]
+            latest_ep = self.latest_episode
             chunk_idx = latest_ep["data/chunk_index"]
             file_idx = latest_ep["data/file_index"]
+            global_frame_index = latest_ep["index"][-1] + 1
 
             latest_path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
-            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
-            latest_num_frames = get_parquet_num_frames(latest_path)
+            latest_size_in_mb = get_file_size_in_mb(latest_path)
+
+            frames_in_current_file = global_frame_index - latest_ep["dataset_from_index"]
+            av_size_per_frame = (
+                latest_size_in_mb / frames_in_current_file if frames_in_current_file > 0 else 0
+            )
 
             # Determine if a new parquet file is needed
-            if latest_size_in_mb + ep_size_in_mb >= self.meta.data_files_size_in_mb:
-                # Size limit is reached, prepare new parquet file
+            if (
+                latest_size_in_mb + av_size_per_frame * ep_num_frames >= self.meta.data_files_size_in_mb
+                or self._writer_closed_for_reading
+            ):
+                # Size limit is reached or writer was closed for reading, prepare new parquet file
                 chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
-                latest_num_frames = 0
-            else:
-                # Update the existing parquet file with new rows
-                latest_df = pd.read_parquet(latest_path)
-                df = pd.concat([latest_df, df], ignore_index=True)
+                self._close_writer()
+                self._writer_closed_for_reading = False
 
-                # Memort optimization
-                del latest_df
-                gc.collect()
+        ep_dict["data/chunk_index"] = chunk_idx
+        ep_dict["data/file_index"] = file_idx
 
         # Write the resulting dataframe from RAM to disk
         path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
         path.parent.mkdir(parents=True, exist_ok=True)
-        if len(self.meta.image_keys) > 0:
-            to_parquet_with_hf_images(df, path)
-        else:
-            df.to_parquet(path)
 
-        if self.hf_dataset is not None:
-            # Remove hf dataset cache directory, necessary to avoid cache bloat
-            cached_dir = get_hf_dataset_cache_dir(self.hf_dataset)
-            if cached_dir is not None:
-                shutil.rmtree(cached_dir)
-
-        self.hf_dataset = self.load_hf_dataset()
+        table = ep_dataset.with_format("arrow")[:]
+        if not self.writer:
+            self.writer = pq.ParquetWriter(
+                path, schema=table.schema, compression="snappy", use_dictionary=True
+            )
+        self.writer.write_table(table)
 
         metadata = {
             "data/chunk_index": chunk_idx,
             "data/file_index": file_idx,
-            "dataset_from_index": latest_num_frames,
-            "dataset_to_index": latest_num_frames + ep_num_frames,
+            "dataset_from_index": global_frame_index,
+            "dataset_to_index": global_frame_index + ep_num_frames,
         }
+
+        # Store metadata with episode data for next episode
+        self.latest_episode = {**ep_dict, **metadata}
+
+        # Mark that the HF dataset needs reloading (lazy loading approach)
+        # This avoids expensive reloading during sequential recording
+        self._lazy_loading = True
+        # Update recorded frames count for efficient length tracking
+        self._recorded_frames += ep_num_frames
+
         return metadata
 
-    def _save_episode_video(self, video_key: str, episode_index: int):
+    def _save_episode_video(self, video_key: str, episode_index: int) -> dict:
         # Encode episode frames into a temporary video
         ep_path = self._encode_temporary_episode_video(video_key, episode_index)
-        ep_size_in_mb = get_video_size_in_mb(ep_path)
+        ep_size_in_mb = get_file_size_in_mb(ep_path)
         ep_duration_in_s = get_video_duration_in_s(ep_path)
 
-        if self.meta.episodes is None or (
-            f"videos/{video_key}/chunk_index" not in self.meta.episodes.column_names
-            or f"videos/{video_key}/file_index" not in self.meta.episodes.column_names
+        if (
+            episode_index == 0
+            or self.meta.latest_episode is None
+            or f"videos/{video_key}/chunk_index" not in self.meta.latest_episode
         ):
             # Initialize indices for a new dataset made of the first episode data
             chunk_idx, file_idx = 0, 0
+            if self.meta.episodes is not None and len(self.meta.episodes) > 0:
+                # It means we are resuming recording, so we need to load the latest episode
+                # Update the indices to avoid overwriting the latest episode
+                old_chunk_idx = self.meta.episodes[-1][f"videos/{video_key}/chunk_index"]
+                old_file_idx = self.meta.episodes[-1][f"videos/{video_key}/file_index"]
+                chunk_idx, file_idx = update_chunk_file_indices(
+                    old_chunk_idx, old_file_idx, self.meta.chunks_size
+                )
             latest_duration_in_s = 0.0
             new_path = self.root / self.meta.video_path.format(
                 video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
@@ -1177,16 +1327,16 @@ class LeRobotDataset(torch.utils.data.Dataset):
             new_path.parent.mkdir(parents=True, exist_ok=True)
             shutil.move(str(ep_path), str(new_path))
         else:
-            # Retrieve information from the latest updated video file (possibly several episodes ago)
-            latest_ep = self.meta.episodes[episode_index - 1]
-            chunk_idx = latest_ep[f"videos/{video_key}/chunk_index"]
-            file_idx = latest_ep[f"videos/{video_key}/file_index"]
+            # Retrieve information from the latest updated video file using latest_episode
+            latest_ep = self.meta.latest_episode
+            chunk_idx = latest_ep[f"videos/{video_key}/chunk_index"][0]
+            file_idx = latest_ep[f"videos/{video_key}/file_index"][0]
 
             latest_path = self.root / self.meta.video_path.format(
                 video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
             )
-            latest_size_in_mb = get_video_size_in_mb(latest_path)
-            latest_duration_in_s = get_video_duration_in_s(latest_path)
+            latest_size_in_mb = get_file_size_in_mb(latest_path)
+            latest_duration_in_s = latest_ep[f"videos/{video_key}/to_timestamp"][0]
 
             if latest_size_in_mb + ep_size_in_mb >= self.meta.video_files_size_in_mb:
                 # Move temporary episode video to a new video file in the dataset
@@ -1263,7 +1413,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if self.image_writer is not None:
             self.image_writer.wait_until_done()
 
-    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> dict:
+    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
         """
         Use ffmpeg to convert frames stored as png into mp4 videos.
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
@@ -1320,6 +1470,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.delta_timestamps = None
         obj.delta_indices = None
         obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
+        obj.writer = None
+        obj.latest_episode = None
+        # Initialize tracking for incremental recording
+        obj._lazy_loading = False
+        obj._recorded_frames = 0
+        obj._writer_closed_for_reading = False
         return obj
 
 
@@ -1396,11 +1552,6 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         """
         return {repo_id: i for i, repo_id in enumerate(self.repo_ids)}
 
-    @property
-    def repo_index_to_id(self):
-        """Return the inverse mapping if repo_id_to_index."""
-        return {v: k for k, v in self.repo_id_to_index}
-
     @property
     def fps(self) -> int:
         """Frames per second used during data collection.
@@ -1431,7 +1582,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         """Keys to access image and video stream from cameras."""
         keys = []
         for key, feats in self.features.items():
-            if isinstance(feats, (datasets.Image, VideoFrame)):
+            if isinstance(feats, (datasets.Image | VideoFrame)):
                 keys.append(key)
         return keys
 

src/lerobot/datasets/pipeline_features.py

@@ -0,0 +1,139 @@
+# 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 re
+from collections.abc import Sequence
+from typing import Any
+
+from lerobot.configs.types import PipelineFeatureType
+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(
+    action: dict[str, Any] | None = None, observation: dict[str, Any] | None = None
+) -> dict[PipelineFeatureType, dict[str, Any]]:
+    """
+    Creates the initial features dict for the dataset from action and observation specs.
+
+    Args:
+        action: A dictionary of action feature names to their types/shapes.
+        observation: A dictionary of observation feature names to their types/shapes.
+
+    Returns:
+        The initial features dictionary structured by PipelineFeatureType.
+    """
+    features = {PipelineFeatureType.ACTION: {}, PipelineFeatureType.OBSERVATION: {}}
+    if action:
+        features[PipelineFeatureType.ACTION] = action
+    if observation:
+        features[PipelineFeatureType.OBSERVATION] = observation
+    return features
+
+
+# Helper to filter state/action keys based on regex patterns.
+def should_keep(key: str, patterns: tuple[str]) -> bool:
+    if patterns is None:
+        return True
+    return any(re.search(pat, key) for pat in patterns)
+
+
+def strip_prefix(key: str, prefixes_to_strip: tuple[str]) -> str:
+    for prefix in prefixes_to_strip:
+        if key.startswith(prefix):
+            return key[len(prefix) :]
+    return key
+
+
+# Define prefixes to strip from feature keys for clean names.
+# Handles both fully qualified (e.g., "action.state") and short (e.g., "state") forms.
+PREFIXES_TO_STRIP = tuple(
+    f"{token}." for const in (ACTION, OBS_STATE, OBS_IMAGES) for token in (const, const.split(".")[-1])
+)
+
+
+def aggregate_pipeline_dataset_features(
+    pipeline: DataProcessorPipeline,
+    initial_features: dict[PipelineFeatureType, dict[str, Any]],
+    *,
+    use_videos: bool = True,
+    patterns: Sequence[str] | None = None,
+) -> dict[str, dict]:
+    """
+    Aggregates and filters pipeline features to create a dataset-ready features dictionary.
+
+    This function transforms initial features using the pipeline, categorizes them as action or observations
+    (image or state), filters them based on `use_videos` and `patterns`, and finally
+    formats them for use with a Hugging Face LeRobot Dataset.
+
+    Args:
+        pipeline: The DataProcessorPipeline to apply.
+        initial_features: A dictionary of raw feature specs for actions and observations.
+        use_videos: If False, image features are excluded.
+        patterns: A sequence of regex patterns to filter action and state features.
+                  Image features are not affected by this filter.
+
+    Returns:
+        A dictionary of features formatted for a Hugging Face LeRobot Dataset.
+    """
+    all_features = pipeline.transform_features(initial_features)
+
+    # Intermediate storage for categorized and filtered features.
+    processed_features: dict[str, dict[str, Any]] = {
+        ACTION: {},
+        OBS_STR: {},
+    }
+    images_token = OBS_IMAGES.split(".")[-1]
+
+    # Iterate through all features transformed by the pipeline.
+    for ptype, feats in all_features.items():
+        if ptype not in [PipelineFeatureType.ACTION, PipelineFeatureType.OBSERVATION]:
+            continue
+
+        for key, value in feats.items():
+            # 1. Categorize the feature.
+            is_action = ptype == PipelineFeatureType.ACTION
+            # Observations are classified as images if their key matches image-related tokens or if the shape of the feature is 3.
+            # All other observations are treated as state.
+            is_image = not is_action and (
+                (isinstance(value, tuple) and len(value) == 3)
+                or (
+                    key.startswith(f"{OBS_IMAGES}.")
+                    or key.startswith(f"{images_token}.")
+                    or f".{images_token}." in key
+                )
+            )
+
+            # 2. Apply filtering rules.
+            if is_image and not use_videos:
+                continue
+            if not is_image and not should_keep(key, patterns):
+                continue
+
+            # 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
+            else:
+                processed_features[OBS_STR][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))
+
+    return dataset_features

src/lerobot/datasets/push_dataset_to_hub/utils.py

@@ -13,67 +13,10 @@
 # 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

@@ -21,7 +21,6 @@ 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,
@@ -38,6 +37,7 @@ 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):
@@ -298,9 +298,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
 
         return padding_mask
 
-    def make_frame(
-        self, dataset_iterator: Backtrackable, previous_dataset_iterator: Backtrackable | None = None
-    ) -> Generator:
+    def make_frame(self, dataset_iterator: Backtrackable) -> 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]

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, Deque, Generic, TypeVar
+from typing import Any, 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, concatenate_datasets
+from datasets import Dataset
 from datasets.table import embed_table_storage
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
@@ -43,7 +43,8 @@ from lerobot.datasets.backward_compatibility import (
     BackwardCompatibilityError,
     ForwardCompatibilityError,
 )
-from lerobot.utils.utils import is_valid_numpy_dtype_string
+from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_STR
+from lerobot.utils.utils import SuppressProgressBars, 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
@@ -66,18 +67,6 @@ 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},
@@ -105,12 +94,6 @@ 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
@@ -134,8 +117,9 @@ 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
-    datasets = [Dataset.from_parquet(str(path), features=features) for path in paths]
-    return concatenate_datasets(datasets)
+    with SuppressProgressBars():
+        datasets = Dataset.from_parquet([str(path) for path in paths], features=features)
+    return datasets
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -143,21 +127,31 @@ def get_parquet_num_frames(parquet_path: str | Path) -> int:
     return metadata.num_rows
 
 
-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 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 flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
-    """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
+    """Flatten a nested dictionary by joining keys with a separator.
 
-    For example:
-    ```
-    >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}`
-    >>> print(flatten_dict(dct))
-    {"a/b": 1, "a/c/d": 2, "e": 3}
-    ```
+    Example:
+        >>> dct = {"a": {"b": 1, "c": {"d": 2}}, "e": 3}
+        >>> print(flatten_dict(dct))
+        {'a/b': 1, 'a/c/d': 2, 'e': 3}
+
+    Args:
+        d (dict): The dictionary to flatten.
+        parent_key (str): The base key to prepend to the keys in this level.
+        sep (str): The separator to use between keys.
+
+    Returns:
+        dict: A flattened dictionary.
     """
     items = []
     for k, v in d.items():
@@ -170,6 +164,20 @@ def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
 
 
 def unflatten_dict(d: dict, sep: str = "/") -> dict:
+    """Unflatten a dictionary with delimited keys into a nested dictionary.
+
+    Example:
+        >>> flat_dct = {"a/b": 1, "a/c/d": 2, "e": 3}
+        >>> print(unflatten_dict(flat_dct))
+        {'a': {'b': 1, 'c': {'d': 2}}, 'e': 3}
+
+    Args:
+        d (dict): A dictionary with flattened keys.
+        sep (str): The separator used in the keys.
+
+    Returns:
+        dict: A nested dictionary.
+    """
     outdict = {}
     for key, value in d.items():
         parts = key.split(sep)
@@ -183,15 +191,28 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
 
 
 def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
+    """Serialize a dictionary containing tensors or numpy arrays to be JSON-compatible.
+
+    Converts torch.Tensor, np.ndarray, and np.generic types to lists or native Python types.
+
+    Args:
+        stats (dict): A dictionary that may contain non-serializable numeric types.
+
+    Returns:
+        dict: A dictionary with all values converted to JSON-serializable types.
+
+    Raises:
+        NotImplementedError: If a value has an unsupported type.
+    """
     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.")
@@ -199,6 +220,17 @@ def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
 
 
 def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
+    """Embed image bytes into the dataset table before saving to Parquet.
+
+    This function prepares a Hugging Face dataset for serialization by converting
+    image objects into an embedded format that can be stored in Arrow/Parquet.
+
+    Args:
+        dataset (datasets.Dataset): The input dataset, possibly containing image features.
+
+    Returns:
+        datasets.Dataset: The dataset with images embedded in the table storage.
+    """
     # Embed image bytes into the table before saving to parquet
     format = dataset.format
     dataset = dataset.with_format("arrow")
@@ -208,11 +240,27 @@ def embed_images(dataset: datasets.Dataset) -> datasets.Dataset:
 
 
 def load_json(fpath: Path) -> Any:
+    """Load data from a JSON file.
+
+    Args:
+        fpath (Path): Path to the JSON file.
+
+    Returns:
+        Any: The data loaded from the JSON file.
+    """
     with open(fpath) as f:
         return json.load(f)
 
 
 def write_json(data: dict, fpath: Path) -> None:
+    """Write data to a JSON file.
+
+    Creates parent directories if they don't exist.
+
+    Args:
+        data (dict): The dictionary to write.
+        fpath (Path): The path to the output JSON file.
+    """
     fpath.parent.mkdir(exist_ok=True, parents=True)
     with open(fpath, "w") as f:
         json.dump(data, f, indent=4, ensure_ascii=False)
@@ -223,6 +271,16 @@ def write_info(info: dict, local_dir: Path) -> None:
 
 
 def load_info(local_dir: Path) -> dict:
+    """Load dataset info metadata from its standard file path.
+
+    Also converts shape lists to tuples for consistency.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        dict: The dataset information dictionary.
+    """
     info = load_json(local_dir / INFO_PATH)
     for ft in info["features"].values():
         ft["shape"] = tuple(ft["shape"])
@@ -230,16 +288,40 @@ def load_info(local_dir: Path) -> dict:
 
 
 def write_stats(stats: dict, local_dir: Path) -> None:
+    """Serialize and write dataset statistics to their standard file path.
+
+    Args:
+        stats (dict): The statistics dictionary (can contain tensors/numpy arrays).
+        local_dir (Path): The root directory of the dataset.
+    """
     serialized_stats = serialize_dict(stats)
     write_json(serialized_stats, local_dir / STATS_PATH)
 
 
 def cast_stats_to_numpy(stats: dict) -> dict[str, dict[str, np.ndarray]]:
+    """Recursively cast numerical values in a stats dictionary to numpy arrays.
+
+    Args:
+        stats (dict): The statistics dictionary.
+
+    Returns:
+        dict: The statistics dictionary with values cast to numpy arrays.
+    """
     stats = {key: np.array(value) for key, value in flatten_dict(stats).items()}
     return unflatten_dict(stats)
 
 
 def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]] | None:
+    """Load dataset statistics and cast numerical values to numpy arrays.
+
+    Returns None if the stats file doesn't exist.
+
+    Args:
+        local_dir (Path): The root directory of the dataset.
+
+    Returns:
+        A dictionary of statistics or None if the file is not found.
+    """
     if not (local_dir / STATS_PATH).exists():
         return None
     stats = load_json(local_dir / STATS_PATH)
@@ -288,15 +370,21 @@ 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:
+    """Load an image from a file into a numpy array.
+
+    Args:
+        fpath (str | Path): Path to the image file.
+        dtype (np.dtype): The desired data type of the output array. If floating,
+            pixels are scaled to [0, 1].
+        channel_first (bool): If True, converts the image to (C, H, W) format.
+            Otherwise, it remains in (H, W, C) format.
+
+    Returns:
+        np.ndarray: The image as a numpy array.
+    """
     img = PILImage.open(fpath).convert("RGB")
     img_array = np.array(img, dtype=dtype)
     if channel_first:  # (H, W, C) -> (C, H, W)
@@ -307,10 +395,19 @@ def load_image_as_numpy(
 
 
 def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[torch.Tensor | str]]:
-    """Get a transform function that convert items from Hugging Face dataset (pyarrow)
-    to torch tensors. Importantly, images are converted from PIL, which corresponds to
-    a channel last representation (h w c) of uint8 type, to a torch image representation
-    with channel first (c h w) of float32 type in range [0,1].
+    """Convert a batch from a Hugging Face dataset to torch tensors.
+
+    This transform function converts items from Hugging Face dataset format (pyarrow)
+    to torch tensors. Importantly, images are converted from PIL objects (H, W, C, uint8)
+    to a torch image representation (C, H, W, float32) in the range [0, 1]. Other
+    types are converted to torch.tensor.
+
+    Args:
+        items_dict (dict): A dictionary representing a batch of data from a
+            Hugging Face dataset.
+
+    Returns:
+        dict: The batch with items converted to torch tensors.
     """
     for key in items_dict:
         first_item = items_dict[key][0]
@@ -325,6 +422,14 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
 
 
 def is_valid_version(version: str) -> bool:
+    """Check if a string is a valid PEP 440 version.
+
+    Args:
+        version (str): The version string to check.
+
+    Returns:
+        bool: True if the version string is valid, False otherwise.
+    """
     try:
         packaging.version.parse(version)
         return True
@@ -338,6 +443,18 @@ def check_version_compatibility(
     current_version: str | packaging.version.Version,
     enforce_breaking_major: bool = True,
 ) -> None:
+    """Check for version compatibility between a dataset and the current codebase.
+
+    Args:
+        repo_id (str): The repository ID for logging purposes.
+        version_to_check (str | packaging.version.Version): The version of the dataset.
+        current_version (str | packaging.version.Version): The current version of the codebase.
+        enforce_breaking_major (bool): If True, raise an error on major version mismatch.
+
+    Raises:
+        BackwardCompatibilityError: If the dataset version is from a newer, incompatible
+            major version of the codebase.
+    """
     v_check = (
         packaging.version.parse(version_to_check)
         if not isinstance(version_to_check, packaging.version.Version)
@@ -355,7 +472,14 @@ def check_version_compatibility(
 
 
 def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
-    """Returns available valid versions (branches and tags) on given repo."""
+    """Return available valid versions (branches and tags) on a given Hub repo.
+
+    Args:
+        repo_id (str): The repository ID on the Hugging Face Hub.
+
+    Returns:
+        list[packaging.version.Version]: A list of valid versions found.
+    """
     api = HfApi()
     repo_refs = api.list_repo_refs(repo_id, repo_type="dataset")
     repo_refs = [b.name for b in repo_refs.branches + repo_refs.tags]
@@ -368,9 +492,22 @@ def get_repo_versions(repo_id: str) -> list[packaging.version.Version]:
 
 
 def get_safe_version(repo_id: str, version: str | packaging.version.Version) -> str:
-    """
-    Returns the version if available on repo or the latest compatible one.
-    Otherwise, will throw a `CompatibilityError`.
+    """Return the specified version if available on repo, or the latest compatible one.
+
+    If the exact version is not found, it looks for the latest version with the
+    same major version number that is less than or equal to the target minor version.
+
+    Args:
+        repo_id (str): The repository ID on the Hugging Face Hub.
+        version (str | packaging.version.Version): The target version.
+
+    Returns:
+        str: The safe version string (e.g., "v1.2.3") to use as a revision.
+
+    Raises:
+        RevisionNotFoundError: If the repo has no version tags.
+        BackwardCompatibilityError: If only older major versions are available.
+        ForwardCompatibilityError: If only newer major versions are available.
     """
     target_version = (
         packaging.version.parse(version) if not isinstance(version, packaging.version.Version) else version
@@ -412,6 +549,17 @@ def get_safe_version(repo_id: str, version: str | packaging.version.Version) ->
 
 
 def get_hf_features_from_features(features: dict) -> datasets.Features:
+    """Convert a LeRobot features dictionary to a `datasets.Features` object.
+
+    Args:
+        features (dict): A LeRobot-style feature dictionary.
+
+    Returns:
+        datasets.Features: The corresponding Hugging Face `datasets.Features` object.
+
+    Raises:
+        ValueError: If a feature has an unsupported shape.
+    """
     hf_features = {}
     for key, ft in features.items():
         if ft["dtype"] == "video":
@@ -439,6 +587,14 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
 
 
 def _validate_feature_names(features: dict[str, dict]) -> None:
+    """Validate that feature names do not contain invalid characters.
+
+    Args:
+        features (dict): The LeRobot features dictionary.
+
+    Raises:
+        ValueError: If any feature name contains '/'.
+    """
     invalid_features = {name: ft for name, ft in features.items() if "/" in name}
     if invalid_features:
         raise ValueError(f"Feature names should not contain '/'. Found '/' in '{invalid_features}'.")
@@ -447,18 +603,38 @@ def _validate_feature_names(features: dict[str, dict]) -> None:
 def hw_to_dataset_features(
     hw_features: dict[str, type | tuple], prefix: str, use_video: bool = True
 ) -> dict[str, dict]:
+    """Convert hardware-specific features to a LeRobot dataset feature dictionary.
+
+    This function takes a dictionary describing hardware outputs (like joint states
+    or camera image shapes) and formats it into the standard LeRobot feature
+    specification.
+
+    Args:
+        hw_features (dict): Dictionary mapping feature names to their type (float for
+            joints) or shape (tuple for images).
+        prefix (str): The prefix to add to the feature keys (e.g., "observation"
+            or "action").
+        use_video (bool): If True, image features are marked as "video", otherwise "image".
+
+    Returns:
+        dict: A LeRobot features dictionary.
+    """
     features = {}
-    joint_fts = {key: ftype for key, ftype in hw_features.items() if ftype is float}
+    joint_fts = {
+        key: ftype
+        for key, ftype in hw_features.items()
+        if ftype is float or (isinstance(ftype, PolicyFeature) and ftype.type != FeatureType.VISUAL)
+    }
     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 == "observation":
+    if joint_fts and prefix == OBS_STR:
         features[f"{prefix}.state"] = {
             "dtype": "float32",
             "shape": (len(joint_fts),),
@@ -479,6 +655,20 @@ def hw_to_dataset_features(
 def build_dataset_frame(
     ds_features: dict[str, dict], values: dict[str, Any], prefix: str
 ) -> dict[str, np.ndarray]:
+    """Construct a single data frame from raw values based on dataset features.
+
+    A "frame" is a dictionary containing all the data for a single timestep,
+    formatted as numpy arrays according to the feature specification.
+
+    Args:
+        ds_features (dict): The LeRobot dataset features dictionary.
+        values (dict): A dictionary of raw values from the hardware/environment.
+        prefix (str): The prefix to filter features by (e.g., "observation"
+            or "action").
+
+    Returns:
+        dict: A dictionary representing a single frame of data.
+    """
     frame = {}
     for key, ft in ds_features.items():
         if key in DEFAULT_FEATURES or not key.startswith(prefix):
@@ -492,6 +682,21 @@ def build_dataset_frame(
 
 
 def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFeature]:
+    """Convert dataset features to policy features.
+
+    This function transforms the dataset's feature specification into a format
+    that a policy can use, classifying features by type (e.g., visual, state,
+    action) and ensuring correct shapes (e.g., channel-first for images).
+
+    Args:
+        features (dict): The LeRobot dataset features dictionary.
+
+    Returns:
+        dict: A dictionary mapping feature keys to `PolicyFeature` objects.
+
+    Raises:
+        ValueError: If an image feature does not have a 3D shape.
+    """
     # TODO(aliberts): Implement "type" in dataset features and simplify this
     policy_features = {}
     for key, ft in features.items():
@@ -505,11 +710,11 @@ 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 == "observation.environment_state":
+        elif key == OBS_ENV_STATE:
             type = FeatureType.ENV
-        elif key.startswith("observation"):
+        elif key.startswith(OBS_STR):
             type = FeatureType.STATE
-        elif key.startswith("action"):
+        elif key.startswith(ACTION):
             type = FeatureType.ACTION
         else:
             continue
@@ -522,6 +727,58 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
     return policy_features
 
 
+def combine_feature_dicts(*dicts: dict) -> dict:
+    """Merge LeRobot grouped feature dicts.
+
+    - For 1D numeric specs (dtype not image/video/string) with "names": we merge the names and recompute the shape.
+    - For others (e.g. `observation.images.*`), the last one wins (if they are identical).
+
+    Args:
+        *dicts: A variable number of LeRobot feature dictionaries to merge.
+
+    Returns:
+        dict: A single merged feature dictionary.
+
+    Raises:
+        ValueError: If there's a dtype mismatch for a feature being merged.
+    """
+    out: dict = {}
+    for d in dicts:
+        for key, value in d.items():
+            if not isinstance(value, dict):
+                out[key] = value
+                continue
+
+            dtype = value.get("dtype")
+            shape = value.get("shape")
+            is_vector = (
+                dtype not in ("image", "video", "string")
+                and isinstance(shape, tuple)
+                and len(shape) == 1
+                and "names" in value
+            )
+
+            if is_vector:
+                # Initialize or retrieve the accumulating dict for this feature key
+                target = out.setdefault(key, {"dtype": dtype, "names": [], "shape": (0,)})
+                # Ensure consistent data types across merged entries
+                if "dtype" in target and dtype != target["dtype"]:
+                    raise ValueError(f"dtype mismatch for '{key}': {target['dtype']} vs {dtype}")
+
+                # Merge feature names: append only new ones to preserve order without duplicates
+                seen = set(target["names"])
+                for n in value["names"]:
+                    if n not in seen:
+                        target["names"].append(n)
+                        seen.add(n)
+                # Recompute the shape to reflect the updated number of features
+                target["shape"] = (len(target["names"]),)
+            else:
+                # For images/videos and non-1D entries: override with the latest definition
+                out[key] = value
+    return out
+
+
 def create_empty_dataset_info(
     codebase_version: str,
     fps: int,
@@ -532,6 +789,18 @@ def create_empty_dataset_info(
     data_files_size_in_mb: int | None = None,
     video_files_size_in_mb: int | None = None,
 ) -> dict:
+    """Create a template dictionary for a new dataset's `info.json`.
+
+    Args:
+        codebase_version (str): The version of the LeRobot codebase.
+        fps (int): The frames per second of the data.
+        features (dict): The LeRobot features dictionary for the dataset.
+        use_videos (bool): Whether the dataset will store videos.
+        robot_type (str | None): The type of robot used, if any.
+
+    Returns:
+        dict: A dictionary with the initial dataset metadata.
+    """
     return {
         "codebase_version": codebase_version,
         "robot_type": robot_type,
@@ -552,9 +821,23 @@ def create_empty_dataset_info(
 def check_delta_timestamps(
     delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
 ) -> bool:
-    """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
-    This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be
-    actual timestamps from the dataset.
+    """Check if delta timestamps are multiples of 1/fps +/- tolerance.
+
+    This ensures that adding these delta timestamps to any existing timestamp in
+    the dataset will result in a value that aligns with the dataset's frame rate.
+
+    Args:
+        delta_timestamps (dict): A dictionary where values are lists of time
+            deltas in seconds.
+        fps (int): The frames per second of the dataset.
+        tolerance_s (float): The allowed tolerance in seconds.
+        raise_value_error (bool): If True, raises an error on failure.
+
+    Returns:
+        bool: True if all deltas are valid, False otherwise.
+
+    Raises:
+        ValueError: If any delta is outside the tolerance and `raise_value_error` is True.
     """
     outside_tolerance = {}
     for key, delta_ts in delta_timestamps.items():
@@ -580,6 +863,15 @@ def check_delta_timestamps(
 
 
 def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
+    """Convert delta timestamps in seconds to delta indices in frames.
+
+    Args:
+        delta_timestamps (dict): A dictionary of time deltas in seconds.
+        fps (int): The frames per second of the dataset.
+
+    Returns:
+        dict: A dictionary of frame delta indices.
+    """
     delta_indices = {}
     for key, delta_ts in delta_timestamps.items():
         delta_indices[key] = [round(d * fps) for d in delta_ts]
@@ -588,9 +880,17 @@ def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dic
 
 
 def cycle(iterable: Any) -> Iterator[Any]:
-    """The equivalent of itertools.cycle, but safe for Pytorch dataloaders.
+    """Create a dataloader-safe cyclical iterator.
 
-    See https://github.com/pytorch/pytorch/issues/23900 for information on why itertools.cycle is not safe.
+    This is an equivalent of `itertools.cycle` but is safe for use with
+    PyTorch DataLoaders with multiple workers.
+    See https://github.com/pytorch/pytorch/issues/23900 for details.
+
+    Args:
+        iterable: The iterable to cycle over.
+
+    Yields:
+        Items from the iterable, restarting from the beginning when exhausted.
     """
     iterator = iter(iterable)
     while True:
@@ -601,8 +901,14 @@ def cycle(iterable: Any) -> Iterator[Any]:
 
 
 def create_branch(repo_id: str, *, branch: str, repo_type: str | None = None) -> None:
-    """Create a branch on a existing Hugging Face repo. Delete the branch if it already
-    exists before creating it.
+    """Create a branch on an existing Hugging Face repo.
+
+    Deletes the branch if it already exists before creating it.
+
+    Args:
+        repo_id (str): The ID of the repository.
+        branch (str): The name of the branch to create.
+        repo_type (str | None): The type of the repository (e.g., "dataset").
     """
     api = HfApi()
 
@@ -620,9 +926,20 @@ def create_lerobot_dataset_card(
     dataset_info: dict | None = None,
     **kwargs,
 ) -> DatasetCard:
-    """
-    Keyword arguments will be used to replace values in src/lerobot/datasets/card_template.md.
-    Note: If specified, license must be one of https://huggingface.co/docs/hub/repositories-licenses.
+    """Create a `DatasetCard` for a LeRobot dataset.
+
+    Keyword arguments are used to replace values in the card template.
+    Note: If specified, `license` must be a valid license identifier from
+    https://huggingface.co/docs/hub/repositories-licenses.
+
+    Args:
+        tags (list | None): A list of tags to add to the dataset card.
+        dataset_info (dict | None): The dataset's info dictionary, which will
+            be displayed on the card.
+        **kwargs: Additional keyword arguments to populate the card template.
+
+    Returns:
+        DatasetCard: The generated dataset card object.
     """
     card_tags = ["LeRobot"]
 
@@ -675,6 +992,15 @@ def validate_frame(frame: dict, features: dict) -> None:
 
 
 def validate_features_presence(actual_features: set[str], expected_features: set[str]) -> str:
+    """Check for missing or extra features in a frame.
+
+    Args:
+        actual_features (set[str]): The set of feature names present in the frame.
+        expected_features (set[str]): The set of feature names expected in the frame.
+
+    Returns:
+        str: An error message string if there's a mismatch, otherwise an empty string.
+    """
     error_message = ""
     missing_features = expected_features - actual_features
     extra_features = actual_features - expected_features
@@ -692,6 +1018,19 @@ def validate_features_presence(actual_features: set[str], expected_features: set
 def validate_feature_dtype_and_shape(
     name: str, feature: dict, value: np.ndarray | PILImage.Image | str
 ) -> str:
+    """Validate the dtype and shape of a single feature's value.
+
+    Args:
+        name (str): The name of the feature.
+        feature (dict): The feature specification from the LeRobot features dictionary.
+        value: The value of the feature to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+
+    Raises:
+        NotImplementedError: If the feature dtype is not supported for validation.
+    """
     expected_dtype = feature["dtype"]
     expected_shape = feature["shape"]
     if is_valid_numpy_dtype_string(expected_dtype):
@@ -707,6 +1046,17 @@ def validate_feature_dtype_and_shape(
 def validate_feature_numpy_array(
     name: str, expected_dtype: str, expected_shape: list[int], value: np.ndarray
 ) -> str:
+    """Validate a feature that is expected to be a numpy array.
+
+    Args:
+        name (str): The name of the feature.
+        expected_dtype (str): The expected numpy dtype as a string.
+        expected_shape (list[int]): The expected shape.
+        value (np.ndarray): The numpy array to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     error_message = ""
     if isinstance(value, np.ndarray):
         actual_dtype = value.dtype
@@ -726,6 +1076,18 @@ def validate_feature_numpy_array(
 def validate_feature_image_or_video(
     name: str, expected_shape: list[str], value: np.ndarray | PILImage.Image
 ) -> str:
+    """Validate a feature that is expected to be an image or video frame.
+
+    Accepts `np.ndarray` (channel-first or channel-last) or `PIL.Image.Image`.
+
+    Args:
+        name (str): The name of the feature.
+        expected_shape (list[str]): The expected shape (C, H, W).
+        value: The image data to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     # Note: The check of pixels range ([0,1] for float and [0,255] for uint8) is done by the image writer threads.
     error_message = ""
     if isinstance(value, np.ndarray):
@@ -742,12 +1104,35 @@ def validate_feature_image_or_video(
 
 
 def validate_feature_string(name: str, value: str) -> str:
+    """Validate a feature that is expected to be a string.
+
+    Args:
+        name (str): The name of the feature.
+        value (str): The value to validate.
+
+    Returns:
+        str: An error message if validation fails, otherwise an empty string.
+    """
     if not isinstance(value, str):
         return f"The feature '{name}' is expected to be of type 'str', but type '{type(value)}' provided instead.\n"
     return ""
 
 
 def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict) -> None:
+    """Validate the episode buffer before it's written to disk.
+
+    Ensures the buffer has the required keys, contains at least one frame, and
+    has features consistent with the dataset's specification.
+
+    Args:
+        episode_buffer (dict): The buffer containing data for a single episode.
+        total_episodes (int): The current total number of episodes in the dataset.
+        features (dict): The LeRobot features dictionary for the dataset.
+
+    Raises:
+        ValueError: If the buffer is invalid.
+        NotImplementedError: If the episode index is manually set and doesn't match.
+    """
     if "size" not in episode_buffer:
         raise ValueError("size key not found in episode_buffer")
 
@@ -793,7 +1178,7 @@ 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
@@ -867,8 +1252,8 @@ 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
@@ -942,12 +1327,6 @@ 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.
@@ -973,31 +1352,6 @@ 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

@@ -0,0 +1,260 @@
+#!/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

@@ -26,14 +26,24 @@ 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
@@ -46,7 +56,6 @@ 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 (
@@ -60,9 +69,9 @@ 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,
@@ -71,9 +80,11 @@ 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"
 
 """
 -------------------------
@@ -143,7 +154,19 @@ 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()
@@ -185,7 +208,10 @@ def convert_data(root: Path, new_root: Path, data_file_size_in_mb: int):
     num_frames = 0
     paths_to_cat = []
     episodes_metadata = []
-    for ep_path in ep_paths:
+
+    logging.info(f"Converting data files from {len(ep_paths)} episodes")
+
+    for ep_path in tqdm.tqdm(ep_paths, desc="convert data files"):
         ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
         ep_num_frames = get_parquet_num_frames(ep_path)
         ep_metadata = {
@@ -209,7 +235,6 @@ 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)
@@ -236,6 +261,8 @@ 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
@@ -254,7 +281,7 @@ 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 range(num_episodes):
+    for ep_idx in tqdm.tqdm(range(num_episodes), desc="convert videos"):
         # 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]
@@ -281,8 +308,9 @@ 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_video_size_in_mb(ep_path)
+        ep_size_in_mb = get_file_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
@@ -374,6 +402,8 @@ 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)
 
@@ -397,14 +427,15 @@ 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"] = "v3.0"
+    info["codebase_version"] = V30
     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
-    info["fps"] = float(info["fps"])
+    info["video_path"] = DEFAULT_VIDEO_PATH if info["video_path"] is not None else None
+    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
@@ -418,16 +449,36 @@ 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))
@@ -435,12 +486,13 @@ def convert_dataset(
     if new_root.is_dir():
         shutil.rmtree(new_root)
 
-    snapshot_download(
-        repo_id,
-        repo_type="dataset",
-        revision=V21,
-        local_dir=root,
-    )
+    if not use_local_dataset:
+        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)
@@ -451,24 +503,26 @@ def convert_dataset(
     shutil.move(str(root), str(old_root))
     shutil.move(str(new_root), str(root))
 
-    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")
+    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")
 
-    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",
@@ -495,6 +549,23 @@ 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

@@ -428,7 +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)}'\n")
+            tmp_concatenate_file.write(f"file '{str(input_path.resolve())}'\n")
         tmp_concatenate_file.flush()
         tmp_concatenate_path = tmp_concatenate_file.name
 
@@ -437,7 +437,9 @@ def concatenate_video_files(
         tmp_concatenate_path, mode="r", format="concat", options={"safe": "0"}
     )  # safe = 0 allows absolute paths as well as relative paths
 
-    tmp_output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False).name
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_named_file:
+        tmp_output_video_path = tmp_named_file.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
@@ -449,11 +451,9 @@ def concatenate_video_files(
             stream_map[input_stream.index] = output_container.add_stream_from_template(
                 template=input_stream, opaque=True
             )
-            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)
+
+            # 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
 
     # Demux + remux packets (no re-encode)
     for packet in input_container.demux():
@@ -585,19 +585,6 @@ 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.
@@ -655,6 +642,9 @@ 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, XarmEnv  # noqa: F401
+from .configs import AlohaEnv, EnvConfig, PushtEnv  # 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
@@ -30,6 +30,8 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
     fps: int = 30
     features: dict[str, PolicyFeature] = field(default_factory=dict)
     features_map: dict[str, str] = field(default_factory=dict)
+    max_parallel_tasks: int = 1
+    disable_env_checker: bool = True
 
     @property
     def type(self) -> str:
@@ -48,15 +50,17 @@ 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",
@@ -65,10 +69,14 @@ class AlohaEnv(EnvConfig):
 
     def __post_init__(self):
         if self.obs_type == "pixels":
-            self.features["top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
+            self.features["top"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
+            )
         elif self.obs_type == "pixels_agent_pos":
             self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(14,))
-            self.features["pixels/top"] = PolicyFeature(type=FeatureType.VISUAL, shape=(480, 640, 3))
+            self.features["pixels/top"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
+            )
 
     @property
     def gym_kwargs(self) -> dict:
@@ -89,15 +97,17 @@ 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,
@@ -106,7 +116,9 @@ class PushtEnv(EnvConfig):
 
     def __post_init__(self):
         if self.obs_type == "pixels_agent_pos":
-            self.features["pixels"] = PolicyFeature(type=FeatureType.VISUAL, shape=(384, 384, 3))
+            self.features["pixels"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
+            )
         elif self.obs_type == "environment_state_agent_pos":
             self.features["environment_state"] = PolicyFeature(type=FeatureType.ENV, shape=(16,))
 
@@ -121,73 +133,70 @@ 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 VideoRecordConfig:
-    """Configuration for video recording in ManiSkill environments."""
-
-    enabled: bool = False
-    record_dir: str = "videos"
-    trajectory_name: str = "trajectory"
-
-
-@dataclass
-class EnvTransformConfig:
-    """Configuration for environment wrappers."""
-
-    # ee_action_space_params: EEActionSpaceConfig = field(default_factory=EEActionSpaceConfig)
-    control_mode: str = "gamepad"
-    display_cameras: bool = False
-    add_joint_velocity_to_observation: bool = False
-    add_current_to_observation: bool = False
-    add_ee_pose_to_observation: bool = False
+class ImagePreprocessingConfig:
     crop_params_dict: dict[str, tuple[int, int, int, int]] | None = None
     resize_size: tuple[int, int] | None = None
-    control_time_s: float = 20.0
+
+
+@dataclass
+class RewardClassifierConfig:
+    """Configuration for reward classification."""
+
+    pretrained_path: str | None = None
+    success_threshold: float = 0.5
+    success_reward: float = 1.0
+
+
+@dataclass
+class InverseKinematicsConfig:
+    """Configuration for inverse kinematics processing."""
+
+    urdf_path: str | None = None
+    target_frame_name: str | None = None
+    end_effector_bounds: dict[str, list[float]] | None = None
+    end_effector_step_sizes: dict[str, float] | None = None
+
+
+@dataclass
+class ObservationConfig:
+    """Configuration for observation processing."""
+
+    add_joint_velocity_to_observation: bool = False
+    add_current_to_observation: bool = False
+    display_cameras: bool = False
+
+
+@dataclass
+class GripperConfig:
+    """Configuration for gripper control and penalties."""
+
+    use_gripper: bool = True
+    gripper_penalty: float = 0.0
+
+
+@dataclass
+class ResetConfig:
+    """Configuration for environment reset behavior."""
+
     fixed_reset_joint_positions: Any | None = None
     reset_time_s: float = 5.0
-    use_gripper: bool = True
-    gripper_quantization_threshold: float | None = 0.8
-    gripper_penalty: float = 0.0
-    gripper_penalty_in_reward: bool = False
+    control_time_s: float = 20.0
+    terminate_on_success: bool = True
+
+
+@dataclass
+class HILSerlProcessorConfig:
+    """Configuration for environment processing pipeline."""
+
+    control_mode: str = "gamepad"
+    observation: ObservationConfig | None = None
+    image_preprocessing: ImagePreprocessingConfig | None = None
+    gripper: GripperConfig | None = None
+    reset: ResetConfig | None = None
+    inverse_kinematics: InverseKinematicsConfig | None = None
+    reward_classifier: RewardClassifierConfig | None = None
+    max_gripper_pos: float | None = 100.0
 
 
 @EnvConfig.register_subclass(name="gym_manipulator")
@@ -197,77 +206,106 @@ class HILSerlRobotEnvConfig(EnvConfig):
 
     robot: RobotConfig | None = None
     teleop: TeleoperatorConfig | None = None
-    wrapper: EnvTransformConfig | None = None
-    fps: int = 10
+    processor: HILSerlProcessorConfig = field(default_factory=HILSerlProcessorConfig)
+
     name: str = "real_robot"
-    mode: str | None = None  # Either "record", "replay", None
-    repo_id: str | None = None
-    dataset_root: str | None = None
-    task: str | None = ""
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: str | None = None
-    reward_classifier_pretrained_path: str | None = None
-    # For the reward classifier, to record more positive examples after a success
-    number_of_steps_after_success: int = 0
 
     @property
     def gym_kwargs(self) -> dict:
         return {}
 
 
-@EnvConfig.register_subclass("hil")
+@EnvConfig.register_subclass("libero")
 @dataclass
-class HILEnvConfig(EnvConfig):
-    """Configuration for the HIL environment."""
+class LiberoEnv(EnvConfig):
+    task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
+    fps: int = 30
+    episode_length: int = 520
+    obs_type: str = "pixels_agent_pos"
+    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
+    features: dict[str, PolicyFeature] = field(
+        default_factory=lambda: {
+            ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
+        }
+    )
+    features_map: dict[str, str] = field(
+        default_factory=lambda: {
+            ACTION: ACTION,
+            "agent_pos": OBS_STATE,
+            "pixels/agentview_image": f"{OBS_IMAGES}.image",
+            "pixels/robot0_eye_in_hand_image": f"{OBS_IMAGES}.image2",
+        }
+    )
 
-    name: str = "PandaPickCube"
-    task: str | None = "PandaPickCubeKeyboard-v0"
-    use_viewer: bool = True
-    gripper_penalty: float = 0.0
-    use_gamepad: bool = True
-    state_dim: int = 18
-    action_dim: int = 4
-    fps: int = 100
-    episode_length: int = 100
-    video_record: VideoRecordConfig = field(default_factory=VideoRecordConfig)
+    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)
+            )
+            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 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)
+            )
+            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
+                type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 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,
+        }
+
+
+@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,)),
-            "observation.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 128, 128)),
-            "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(18,)),
         }
     )
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             "action": ACTION,
-            "observation.image": OBS_IMAGE,
-            "observation.state": OBS_STATE,
+            "agent_pos": OBS_STATE,
+            "top": f"{OBS_IMAGE}",
+            "pixels/top": f"{OBS_IMAGE}",
         }
     )
-    ################# args from hilserlrobotenv
-    reward_classifier_pretrained_path: str | None = None
-    robot_config: RobotConfig | None = None
-    teleop_config: TeleoperatorConfig | None = None
-    wrapper: EnvTransformConfig | None = None
-    mode: str | None = None  # Either "record", "replay", None
-    repo_id: str | None = None
-    dataset_root: str | None = None
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: str | None = None
-    # For the reward classifier, to record more positive examples after a success
-    number_of_steps_after_success: int = 0
-    ############################
+
+    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 {
-            "use_viewer": self.use_viewer,
-            "use_gamepad": self.use_gamepad,
-            "gripper_penalty": self.gripper_penalty,
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
         }

src/lerobot/envs/factory.py

@@ -17,7 +17,7 @@ import importlib
 
 import gymnasium as gym
 
-from lerobot.envs.configs import AlohaEnv, EnvConfig, HILEnvConfig, PushtEnv, XarmEnv
+from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -25,15 +25,15 @@ 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 == "hil":
-        return HILEnvConfig(**kwargs)
+    elif env_type == "libero":
+        return LiberoEnv(**kwargs)
     else:
         raise ValueError(f"Policy type '{env_type}' is not available.")
 
 
-def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> gym.vector.VectorEnv | None:
+def make_env(
+    cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False
+) -> dict[str, dict[int, gym.vector.VectorEnv]]:
     """Makes a gym vector environment according to the config.
 
     Args:
@@ -47,13 +47,44 @@ def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> g
         ModuleNotFoundError: If the requested env package is not installed
 
     Returns:
-        gym.vector.VectorEnv: The parallelized gym.env instance.
+        dict[str, dict[int, gym.vector.VectorEnv]]:
+            A mapping from suite name to indexed vectorized environments.
+            - For multi-task benchmarks (e.g., LIBERO): one entry per suite, and one vec env per task_id.
+            - For single-task environments: a single suite entry (cfg.type) with task_id=0.
+
     """
     if n_envs < 1:
-        raise ValueError("`n_envs must be at least 1")
+        raise ValueError("`n_envs` must be at least 1")
 
+    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+
+    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,
+            camera_name=cfg.camera_name,
+            init_states=cfg.init_states,
+            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")
+
+        return create_metaworld_envs(
+            task=cfg.task,
+            n_envs=n_envs,
+            gym_kwargs=cfg.gym_kwargs,
+            env_cls=env_cls,
+        )
     package_name = f"gym_{cfg.type}"
-
     try:
         importlib.import_module(package_name)
     except ModuleNotFoundError as e:
@@ -62,10 +93,11 @@ def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> g
 
     gym_handle = f"{package_name}/{cfg.task}"
 
-    # batched version of the env that returns an observation of shape (b, c)
-    env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
-    env = env_cls(
-        [lambda: gym.make(gym_handle, disable_env_checker=True, **cfg.gym_kwargs) for _ in range(n_envs)]
-    )
+    def _make_one():
+        return gym.make(gym_handle, disable_env_checker=cfg.disable_env_checker, **(cfg.gym_kwargs or {}))
 
-    return env
+    vec = env_cls([_make_one for _ in range(n_envs)], autoreset_mode=gym.vector.AutoresetMode.SAME_STEP)
+
+    # normalize to {suite: {task_id: vec_env}} for consistency
+    suite_name = cfg.type  # e.g., "pusht", "aloha"
+    return {suite_name: {0: vec}}

src/lerobot/envs/libero.py

@@ -0,0 +1,381 @@
+#!/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 __future__ import annotations
+
+import os
+from collections import defaultdict
+from collections.abc import Callable, Iterable, Mapping, Sequence
+from functools import partial
+from pathlib import Path
+from typing import Any
+
+import gymnasium as gym
+import numpy as np
+import torch
+from gymnasium import spaces
+from libero.libero import benchmark, get_libero_path
+from libero.libero.envs import OffScreenRenderEnv
+from robosuite.utils.transform_utils import quat2axisangle
+
+
+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)):
+        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__}")
+    if not cams:
+        raise ValueError("camera_name resolved to an empty list.")
+    return cams
+
+
+def _get_suite(name: str) -> benchmark.Benchmark:
+    """Instantiate a LIBERO suite by name with clear validation."""
+    bench = benchmark.get_benchmark_dict()
+    if name not in bench:
+        raise ValueError(f"Unknown LIBERO suite '{name}'. Available: {', '.join(sorted(bench.keys()))}")
+    suite = bench[name]()
+    if not getattr(suite, "tasks", None):
+        raise ValueError(f"Suite '{name}' has no tasks.")
+    return suite
+
+
+def _select_task_ids(total_tasks: int, task_ids: Iterable[int] | None) -> list[int]:
+    """Validate/normalize task ids. If None → all tasks."""
+    if task_ids is None:
+        return list(range(total_tasks))
+    ids = sorted({int(t) for t in task_ids})
+    for t in ids:
+        if t < 0 or t >= total_tasks:
+            raise ValueError(f"task_id {t} out of range [0, {total_tasks - 1}].")
+    return ids
+
+
+def get_task_init_states(task_suite: Any, i: int) -> np.ndarray:
+    init_states_path = (
+        Path(get_libero_path("init_states"))
+        / task_suite.tasks[i].problem_folder
+        / task_suite.tasks[i].init_states_file
+    )
+    init_states = torch.load(init_states_path, weights_only=False)  # nosec B614
+    return init_states
+
+
+def get_libero_dummy_action():
+    """Get dummy/no-op action, used to roll out the simulation while the robot does nothing."""
+    return [0, 0, 0, 0, 0, 0, -1]
+
+
+OBS_STATE_DIM = 8
+ACTION_DIM = 7
+AGENT_POS_LOW = -1000.0
+AGENT_POS_HIGH = 1000.0
+ACTION_LOW = -1.0
+ACTION_HIGH = 1.0
+TASK_SUITE_MAX_STEPS: dict[str, int] = {
+    "libero_spatial": 280,  # longest training demo has 193 steps
+    "libero_object": 280,  # longest training demo has 254 steps
+    "libero_goal": 300,  # longest training demo has 270 steps
+    "libero_10": 520,  # longest training demo has 505 steps
+    "libero_90": 400,  # longest training demo has 373 steps
+}
+
+
+class LiberoEnv(gym.Env):
+    metadata = {"render_modes": ["rgb_array"], "render_fps": 80}
+
+    def __init__(
+        self,
+        task_suite: Any,
+        task_id: int,
+        task_suite_name: str,
+        camera_name: str | Sequence[str] = "agentview_image,robot0_eye_in_hand_image",
+        obs_type: str = "pixels",
+        render_mode: str = "rgb_array",
+        observation_width: int = 256,
+        observation_height: int = 256,
+        visualization_width: int = 640,
+        visualization_height: int = 480,
+        init_states: bool = True,
+        episode_index: int = 0,
+        camera_name_mapping: dict[str, str] | None = None,
+        num_steps_wait: int = 10,
+    ):
+        super().__init__()
+        self.task_id = task_id
+        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.init_states = init_states
+        self.camera_name = _parse_camera_names(
+            camera_name
+        )  # agentview_image (main) or robot0_eye_in_hand_image (wrist)
+
+        # Map raw camera names to "image1" and "image2".
+        # The preprocessing step `preprocess_observation` will then prefix these with `.images.*`,
+        # following the LeRobot convention (e.g., `observation.images.image`, `observation.images.image2`).
+        # This ensures the policy consistently receives observations in the
+        # expected format regardless of the original camera naming.
+        if camera_name_mapping is None:
+            camera_name_mapping = {
+                "agentview_image": "image",
+                "robot0_eye_in_hand_image": "image2",
+            }
+        self.camera_name_mapping = camera_name_mapping
+        self.num_steps_wait = num_steps_wait
+        self.episode_index = episode_index
+        # Load once and keep
+        self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
+        self._init_state_id = self.episode_index  # tie each sub-env to a fixed init state
+
+        self._env = self._make_envs_task(task_suite, self.task_id)
+        default_steps = 500
+        self._max_episode_steps = TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
+
+        images = {}
+        for cam in self.camera_name:
+            images[self.camera_name_mapping[cam]] = spaces.Box(
+                low=0,
+                high=255,
+                shape=(self.observation_height, self.observation_width, 3),
+                dtype=np.uint8,
+            )
+
+        if self.obs_type == "state":
+            raise NotImplementedError(
+                "The 'state' observation type is not supported in LiberoEnv. "
+                "Please switch to an image-based obs_type (e.g. 'pixels', 'pixels_agent_pos')."
+            )
+
+        elif self.obs_type == "pixels":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Dict(images),
+                }
+            )
+        elif self.obs_type == "pixels_agent_pos":
+            self.observation_space = spaces.Dict(
+                {
+                    "pixels": spaces.Dict(images),
+                    "agent_pos": spaces.Box(
+                        low=AGENT_POS_LOW,
+                        high=AGENT_POS_HIGH,
+                        shape=(OBS_STATE_DIM,),
+                        dtype=np.float64,
+                    ),
+                }
+            )
+
+        self.action_space = spaces.Box(
+            low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32
+        )
+
+    def render(self):
+        raw_obs = self._env.env._get_observations()
+        image = self._format_raw_obs(raw_obs)["pixels"]["image"]
+        return image
+
+    def _make_envs_task(self, task_suite: Any, task_id: int = 0):
+        task = task_suite.get_task(task_id)
+        self.task = task.name
+        self.task_description = task.language
+        task_bddl_file = os.path.join(get_libero_path("bddl_files"), task.problem_folder, task.bddl_file)
+
+        env_args = {
+            "bddl_file_name": task_bddl_file,
+            "camera_heights": self.observation_height,
+            "camera_widths": self.observation_width,
+        }
+        env = OffScreenRenderEnv(**env_args)
+        env.reset()
+        return env
+
+    def _format_raw_obs(self, raw_obs: dict[str, Any]) -> dict[str, Any]:
+        images = {}
+        for camera_name in self.camera_name:
+            image = raw_obs[camera_name]
+            image = image[::-1, ::-1]  # rotate 180 degrees
+            images[self.camera_name_mapping[camera_name]] = image
+        state = np.concatenate(
+            (
+                raw_obs["robot0_eef_pos"],
+                quat2axisangle(raw_obs["robot0_eef_quat"]),
+                raw_obs["robot0_gripper_qpos"],
+            )
+        )
+        agent_pos = state
+        if self.obs_type == "pixels":
+            return {"pixels": images.copy()}
+        if self.obs_type == "pixels_agent_pos":
+            return {
+                "pixels": images.copy(),
+                "agent_pos": agent_pos,
+            }
+        raise NotImplementedError(
+            f"The observation type '{self.obs_type}' is not supported in LiberoEnv. "
+            "Please switch to an image-based obs_type (e.g. 'pixels', 'pixels_agent_pos')."
+        )
+
+    def reset(self, seed=None, **kwargs):
+        super().reset(seed=seed)
+        self._env.seed(seed)
+        if self.init_states and self._init_states is not None:
+            self._env.set_init_state(self._init_states[self._init_state_id])
+        raw_obs = self._env.reset()
+
+        # After reset, objects may be unstable (slightly floating, intersecting, etc.).
+        # Step the simulator with a no-op action for a few frames so everything settles.
+        # Increasing this value can improve determinism and reproducibility across resets.
+        for _ in range(self.num_steps_wait):
+            raw_obs, _, _, _ = self._env.step(get_libero_dummy_action())
+        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]]:
+        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, info = self._env.step(action)
+
+        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,
+            }
+        )
+        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),
+            }
+            self.reset()
+        truncated = False
+        return observation, reward, terminated, truncated, info
+
+    def close(self):
+        self._env.close()
+
+
+def _make_env_fns(
+    *,
+    suite,
+    suite_name: str,
+    task_id: int,
+    n_envs: int,
+    camera_names: list[str],
+    init_states: bool,
+    gym_kwargs: Mapping[str, Any],
+) -> list[Callable[[], LiberoEnv]]:
+    """Build n_envs factory callables for a single (suite, task_id)."""
+
+    def _make_env(episode_index: int, **kwargs) -> LiberoEnv:
+        local_kwargs = dict(kwargs)
+        return LiberoEnv(
+            task_suite=suite,
+            task_id=task_id,
+            task_suite_name=suite_name,
+            camera_name=camera_names,
+            init_states=init_states,
+            episode_index=episode_index,
+            **local_kwargs,
+        )
+
+    fns: list[Callable[[], LiberoEnv]] = []
+    for episode_index in range(n_envs):
+        fns.append(partial(_make_env, episode_index, **gym_kwargs))
+    return fns
+
+
+# ---- Main API ----------------------------------------------------------------
+
+
+def create_libero_envs(
+    task: str,
+    n_envs: int,
+    gym_kwargs: dict[str, Any] | None = None,
+    camera_name: str | Sequence[str] = "agentview_image,robot0_eye_in_hand_image",
+    init_states: bool = True,
+    env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+) -> dict[str, dict[int, Any]]:
+    """
+    Create vectorized LIBERO environments with a consistent return shape.
+
+    Returns:
+        dict[suite_name][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 suite or a comma-separated list of suites.
+        - You may pass `task_ids` (list[int]) inside `gym_kwargs` to restrict tasks per suite.
+    """
+    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_ids_filter = gym_kwargs.pop("task_ids", None)  # optional: limit to specific tasks
+
+    camera_names = _parse_camera_names(camera_name)
+    suite_names = [s.strip() for s in str(task).split(",") if s.strip()]
+    if not suite_names:
+        raise ValueError("`task` must contain at least one LIBERO suite name.")
+
+    print(
+        f"Creating LIBERO envs | suites={suite_names} | n_envs(per task)={n_envs} | init_states={init_states}"
+    )
+    if task_ids_filter is not None:
+        print(f"Restricting to task_ids={task_ids_filter}")
+
+    out: dict[str, dict[int, Any]] = defaultdict(dict)
+
+    for suite_name in suite_names:
+        suite = _get_suite(suite_name)
+        total = len(suite.tasks)
+        selected = _select_task_ids(total, task_ids_filter)
+
+        if not selected:
+            raise ValueError(f"No tasks selected for suite '{suite_name}' (available: {total}).")
+
+        for tid in selected:
+            fns = _make_env_fns(
+                suite=suite,
+                suite_name=suite_name,
+                task_id=tid,
+                n_envs=n_envs,
+                camera_names=camera_names,
+                init_states=init_states,
+                gym_kwargs=gym_kwargs,
+            )
+            out[suite_name][tid] = env_cls(fns)
+            print(f"Built vec env | suite={suite_name} | task_id={tid} | n_envs={n_envs}")
+
+    # return plain dicts for predictability
+    return {suite: dict(task_map) for suite, task_map in out.items()}

src/lerobot/envs/metaworld.py

@@ -0,0 +1,313 @@
+#!/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

@@ -0,0 +1,121 @@
+{
+  "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

@@ -14,6 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import warnings
+from collections.abc import Mapping, Sequence
+from functools import singledispatch
 from typing import Any
 
 import einops
@@ -24,6 +26,7 @@ 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
 
 
@@ -39,44 +42,44 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
     return_observations = {}
     if "pixels" in observations:
         if isinstance(observations["pixels"], dict):
-            imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
+            imgs = {f"{OBS_IMAGES}.{key}": img for key, img in observations["pixels"].items()}
         else:
-            imgs = {"observation.image": observations["pixels"]}
+            imgs = {OBS_IMAGE: observations["pixels"]}
 
         for imgkey, img in imgs.items():
             # TODO(aliberts, rcadene): use transforms.ToTensor()?
-            img = torch.from_numpy(img)
+            img_tensor = 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.ndim == 3:
-                img = img.unsqueeze(0)
+            if img_tensor.ndim == 3:
+                img_tensor = img_tensor.unsqueeze(0)
             # sanity check that images are channel last
-            _, h, w, c = img.shape
-            assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
+            _, h, w, c = img_tensor.shape
+            assert c < h and c < w, f"expect channel last images, but instead got {img_tensor.shape=}"
 
             # sanity check that images are uint8
-            assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
+            assert img_tensor.dtype == torch.uint8, f"expect torch.uint8, but instead {img_tensor.dtype=}"
 
             # convert to channel first of type float32 in range [0,1]
-            img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
-            img = img.type(torch.float32)
-            img /= 255
+            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
 
-            return_observations[imgkey] = img
+            return_observations[imgkey] = img_tensor
 
     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["observation.environment_state"] = env_state
+        return_observations[OBS_ENV_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["observation.state"] = agent_pos
+    return_observations[OBS_STATE] = agent_pos
 
     return return_observations
 
@@ -127,10 +130,68 @@ def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
 def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
     """Adds task feature to the observation dict with respect to the first environment attribute."""
     if hasattr(env.envs[0], "task_description"):
-        observation["task"] = env.call("task_description")
+        task_result = env.call("task_description")
+
+        if isinstance(task_result, tuple):
+            task_result = list(task_result)
+
+        if not isinstance(task_result, list):
+            raise TypeError(f"Expected task_description to return a list, got {type(task_result)}")
+        if not all(isinstance(item, str) for item in task_result):
+            raise TypeError("All items in task_description result must be strings")
+
+        observation["task"] = task_result
     elif hasattr(env.envs[0], "task"):
-        observation["task"] = env.call("task")
+        task_result = env.call("task")
+
+        if isinstance(task_result, tuple):
+            task_result = list(task_result)
+
+        if not isinstance(task_result, list):
+            raise TypeError(f"Expected task to return a list, got {type(task_result)}")
+        if not all(isinstance(item, str) for item in task_result):
+            raise TypeError("All items in task result must be strings")
+
+        observation["task"] = task_result
     else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
         num_envs = observation[list(observation.keys())[0]].shape[0]
         observation["task"] = ["" for _ in range(num_envs)]
     return observation
+
+
+def _close_single_env(env: Any) -> None:
+    try:
+        env.close()
+    except Exception as exc:
+        print(f"Exception while closing env {env}: {exc}")
+
+
+@singledispatch
+def close_envs(obj: Any) -> None:
+    """Default: raise if the type is not recognized."""
+    raise NotImplementedError(f"close_envs not implemented for type {type(obj).__name__}")
+
+
+@close_envs.register
+def _(env: Mapping) -> None:
+    for v in env.values():
+        if isinstance(v, Mapping):
+            close_envs(v)
+        elif hasattr(v, "close"):
+            _close_single_env(v)
+
+
+@close_envs.register
+def _(envs: Sequence) -> None:
+    if isinstance(envs, (str | bytes)):
+        return
+    for v in envs:
+        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)
+
+
+@close_envs.register
+def _(env: gym.Env) -> None:
+    _close_single_env(env)

src/lerobot/motors/__init__.py

@@ -1 +1,17 @@
+#!/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.utils.encoding_utils import decode_twos_complement, encode_twos_complement
+from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement
 
 from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
 from .tables import (