fix(scripts): better prints teleop (#2538)

* fix(time benchmark): removing deprecated TimeBenchmark dependency

* fix(typo): renaming frames in an up-to-date fashion

* feat(duets): rearanging crf and g parameters in a proper unique combination manner

* fix(segfault): fixing segfault by adding a lock in ThreadPoolExecutor

* chore(update) : update datasets, codecs and backends to the latest versions

* chore(unused files): removing unused files

* fix(dataset paths): fix datasets paths to live among lerobot datasets

.github/workflows/full_tests.yml

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

.github/workflows/nightly.yml

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

.github/workflows/release.yml

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

.github/workflows/stale.yml

@@ -27,15 +27,17 @@ env:
     This issue was closed because it has been stalled for 14 days with no activity.
     Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
   CLOSE_PR_MESSAGE: >
-    This PR was closed because it has been stalled for 14 days with no activity.
+    This PR was closed because it has been stalled for 21 days with no activity.
     Feel free to reopen if is still relevant, or to ping a collaborator if you have any questions.
   WARN_ISSUE_MESSAGE: >
     This issue has been automatically marked as stale because it has not had
     recent activity (6 months). It will be closed if no further activity occurs.
+    Any change, comment or update to this issue will reset this count.
     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.
+    recent activity (1 year). It will be closed if no further activity occurs.
+    Any change, comment or update to this PR will reset this count.
     Thank you for your contributions.
 
 jobs:
@@ -56,10 +58,10 @@ jobs:
           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-stale: 180
           days-before-issue-close: 14
-          days-before-pr-stale: 180
-          days-before-pr-close: 14
+          days-before-pr-stale: 365
+          days-before-pr-close: 21
           delete-branch: true
           close-issue-message: ${{ env.CLOSE_ISSUE_MESSAGE }}
           close-pr-message: ${{ env.CLOSE_PR_MESSAGE }}

.github/workflows/unbound_deps_tests.yml

@@ -0,0 +1,183 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This workflow handles full testing with unboud dependencies versions.
+name: Unbound Dependency Tests
+
+on:
+  # Allows running this workflow manually from the Actions tab
+  workflow_dispatch:
+
+  # Run on the 1st and 15th of every month at 09:00 UTC
+  schedule:
+    - cron: '0 2 1,15 * *'
+
+permissions:
+  contents: read
+
+# Sets up the environment variables
+env:
+  UV_VERSION: "0.8.0"
+  PYTHON_VERSION: "3.10"
+  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu:unbound
+
+# Ensures that only the latest action is built, canceling older runs.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+
+  # This job runs the E2E tests + pytest with all unbound extras
+  full-tests:
+    name: Full Unbound Tests
+    runs-on: ubuntu-latest
+    env:
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          lfs: true
+          persist-credentials: false
+
+      - name: Install apt dependencies
+        run: |
+          sudo apt-get update && sudo apt-get install -y build-essential \
+          git curl libglib2.0-0 libegl1-mesa-dev ffmpeg libusb-1.0-0-dev \
+          speech-dispatcher libgeos-dev portaudio19-dev
+
+      - name: Setup uv and Python
+        uses: astral-sh/setup-uv@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          enable-cache: true
+          version: ${{ env.UV_VERSION }}
+          python-version: ${{ env.PYTHON_VERSION }}
+
+      - name: Unbound dependencies
+        run: |
+          sed -i 's/,[[:space:]]*<[0-9\.]*//g' pyproject.toml
+          echo "Dependencies unbound:" && cat pyproject.toml
+
+      - name: Install lerobot with all extras
+        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
+
+      - name: Run pytest (all extras)
+        run: uv run pytest tests -vv
+
+      - name: Run end-to-end tests
+        run: uv run make test-end-to-end
+
+  # This job builds a GPU enabled image for testing
+  build-and-push-docker:
+    name: Build and Push Docker
+    runs-on:
+      group: aws-general-8-plus
+    outputs:
+      image_tag: ${{ env.DOCKER_IMAGE_NAME }}
+    env:
+      GITHUB_REF: ${{ github.ref }}
+    steps:
+      - name: Install Git LFS
+        run: |
+          sudo apt-get update
+          sudo apt-get install git-lfs
+          git lfs install
+      - uses: actions/checkout@v4
+        with:
+          lfs: true
+          persist-credentials: false
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          cache-binary: false
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3 # zizmor: ignore[unpinned-uses]
+        with:
+          username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+      - name: Build and push Docker image
+        uses: docker/build-push-action@v6 # zizmor: ignore[unpinned-uses]
+        with:
+          context: .
+          file: ./docker/Dockerfile.internal
+          push: true
+          tags: ${{ env.DOCKER_IMAGE_NAME }}
+          build-args: |
+            UNBOUND_DEPS=true
+
+  # This job runs pytest with all unbound extras in a GPU enabled host
+  # It runs everytime a test image is created
+  gpu-tests:
+    name: GPU Unbound Tests
+    needs: [build-and-push-docker]
+    runs-on:
+      group: aws-g6-4xlarge-plus
+    env:
+      HF_HOME: /home/user_lerobot/.cache/huggingface
+      HF_LEROBOT_HOME: /home/user_lerobot/.cache/huggingface/lerobot
+      TORCH_HOME: /home/user_lerobot/.cache/torch
+      TRITON_CACHE_DIR: /home/user_lerobot/.cache/triton
+    container:
+      image: ${{ needs.build-and-push-docker.outputs.image_tag }} # zizmor: ignore[unpinned-images]
+      options: --gpus all --shm-size "16gb"
+      credentials:
+        username: ${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}
+        password: ${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}
+    defaults:
+      run:
+        shell: bash
+        working-directory: /lerobot
+    steps:
+      - name: Run pytest on GPU
+        run: pytest tests -vv
+      - name: Run end-to-end tests
+        run: make test-end-to-end
+
+  # This job deletes the test image recently created
+  # It runs everytime after the gpu-tests have finished
+  delete-unbound-image:
+    name: Delete Unbound Image
+    needs: [gpu-tests, build-and-push-docker]
+    if: always() && needs.build-and-push-docker.result == 'success'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Get Docker Hub Token and Delete Image
+        # zizmor: ignore[template-injection]
+        run: |
+          IMAGE_NAME=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f1)
+          IMAGE_TAG=$(echo "${{ needs.build-and-push-docker.outputs.image_tag }}" | cut -d':' -f2)
+
+          echo "Attempting to delete image: $IMAGE_NAME:$IMAGE_TAG"
+
+          TOKEN=$(curl -s -H "Content-Type: application/json" \
+                       -X POST \
+                       -d '{"username": "${{ secrets.DOCKERHUB_LEROBOT_USERNAME }}", "password": "${{ secrets.DOCKERHUB_LEROBOT_PASSWORD }}"}' \
+                       https://hub.docker.com/v2/users/login/ | jq -r .token)
+
+          if [ "$TOKEN" == "null" ] || [ -z "$TOKEN" ]; then
+            echo "::error::Failed to get Docker Hub token."
+            exit 1
+          fi
+
+          HTTP_RESPONSE=$(curl -s -o /dev/null -w "%{http_code}" \
+                               -H "Authorization: JWT ${TOKEN}" \
+                               -X DELETE \
+                               https://hub.docker.com/v2/repositories/${IMAGE_NAME}/tags/${IMAGE_TAG}/)
+
+          if [ "$HTTP_RESPONSE" -eq 204 ]; then
+            echo "Successfully deleted Docker image tag: $IMAGE_NAME:$IMAGE_TAG"
+          else
+            echo "::error::Failed to delete Docker image. HTTP status: $HTTP_RESPONSE"
+            exit 1
+          fi

.pre-commit-config.yaml

@@ -26,7 +26,7 @@ repos:
 
    ##### General Code Quality & Formatting #####
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+    rev: v6.0.0
     hooks:
       - id: check-added-large-files
         args: ['--maxkb=1024']
@@ -39,20 +39,20 @@ repos:
       - id: trailing-whitespace
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.12.4
+    rev: v0.14.1
     hooks:
       - id: ruff-format
       - id: ruff
         args: [--fix, --exit-non-zero-on-fix]
 
   - repo: https://github.com/adhtruong/mirrors-typos
-    rev: v1.34.0
+    rev: v1.38.1
     hooks:
       - id: typos
         args: [--force-exclude]
 
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.20.0
+    rev: v3.21.0
     hooks:
     -   id: pyupgrade
         args: [--py310-plus]
@@ -68,12 +68,12 @@ repos:
 
   ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.27.2
+    rev: v8.28.0
     hooks:
       - id: gitleaks
 
   - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.11.0
+    rev: v1.15.2
     hooks:
       - id: zizmor
 
@@ -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: v1.18.2
+    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.
 
@@ -138,7 +137,7 @@ Follow these steps to start contributing:
 4. for development, we advise to use a tool like `poetry` or `uv` instead of just `pip` to easily track our dependencies.
    Follow the instructions to [install poetry](https://python-poetry.org/docs/#installation) (use a version >=2.1.0) or to [install uv](https://docs.astral.sh/uv/getting-started/installation/#installation-methods) if you don't have one of them already.
 
-   Set up a development environment with conda or miniconda:
+   Set up a development environment with conda:
 
    ```bash
    conda create -y -n lerobot-dev python=3.10 && conda activate lerobot-dev

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

@@ -104,14 +104,14 @@ LeRobot works with Python 3.10+ and PyTorch 2.2+.
 
 ### Environment Setup
 
-Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/free/miniconda/index.html):
+Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniforge`](https://conda-forge.org/download/):
 
 ```bash
 conda create -y -n lerobot python=3.10
 conda activate lerobot
 ```
 
-When using `miniconda`, install `ffmpeg` in your environment:
+When using `conda`, install `ffmpeg` in your environment:
 
 ```bash
 conda install ffmpeg -c conda-forge
@@ -185,6 +185,11 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
+> [!NOTE]
+> For lerobot 0.4.0, if you want to install pi tags, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+>
+> This will be solved in the next patch release
+
 ### Weights & Biases
 
 To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
@@ -197,7 +202,7 @@ 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:
 
@@ -207,13 +212,13 @@ lerobot-dataset-viz \
     --episode-index 0
 ```
 
-or from a dataset in a local folder with the `root` option and the `--local-files-only` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
+or from a dataset in a local folder with the `root` option and the `--mode local` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
 
 ```bash
 lerobot-dataset-viz \
     --repo-id lerobot/pusht \
     --root ./my_local_data_dir \
-    --local-files-only 1 \
+    --mode local \
     --episode-index 0
 ```
 
@@ -310,7 +315,7 @@ To upload these to the hub, run the following:
 huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
 ```
 
-See [eval.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/eval.py) for an example of how other people may use your policy.
+See [lerobot_eval.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_eval.py) for an example of how other people may use your policy.
 
 ### Acknowledgment
 
@@ -337,7 +342,3 @@ 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/benchmark.py

@@ -1,94 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import threading
-import time
-from contextlib import ContextDecorator
-
-
-class TimeBenchmark(ContextDecorator):
-    """
-    Measures execution time using a context manager or decorator.
-
-    This class supports both context manager and decorator usage, and is thread-safe for multithreaded
-    environments.
-
-    Args:
-        print: If True, prints the elapsed time upon exiting the context or completing the function. Defaults
-        to False.
-
-    Examples:
-
-        Using as a context manager:
-
-        >>> benchmark = TimeBenchmark()
-        >>> with benchmark:
-        ...     time.sleep(1)
-        >>> print(f"Block took {benchmark.result:.4f} seconds")
-        Block took approximately 1.0000 seconds
-
-        Using with multithreading:
-
-        ```python
-        import threading
-
-        benchmark = TimeBenchmark()
-
-
-        def context_manager_example():
-            with benchmark:
-                time.sleep(0.01)
-            print(f"Block took {benchmark.result_ms:.2f} milliseconds")
-
-
-        threads = []
-        for _ in range(3):
-            t1 = threading.Thread(target=context_manager_example)
-            threads.append(t1)
-
-        for t in threads:
-            t.start()
-
-        for t in threads:
-            t.join()
-        ```
-        Expected output:
-        Block took approximately 10.00 milliseconds
-        Block took approximately 10.00 milliseconds
-        Block took approximately 10.00 milliseconds
-    """
-
-    def __init__(self, print=False):
-        self.local = threading.local()
-        self.print_time = print
-
-    def __enter__(self):
-        self.local.start_time = time.perf_counter()
-        return self
-
-    def __exit__(self, *exc):
-        self.local.end_time = time.perf_counter()
-        self.local.elapsed_time = self.local.end_time - self.local.start_time
-        if self.print_time:
-            print(f"Elapsed time: {self.local.elapsed_time:.4f} seconds")
-        return False
-
-    @property
-    def result(self):
-        return getattr(self.local, "elapsed_time", None)
-
-    @property
-    def result_ms(self):
-        return self.result * 1e3

benchmarks/video/capture_camera_feed.py

@@ -1,102 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Capture video feed from a camera as raw images."""
-
-import argparse
-import datetime as dt
-import os
-import time
-from pathlib import Path
-
-import cv2
-import rerun as rr
-
-# see https://rerun.io/docs/howto/visualization/limit-ram
-RERUN_MEMORY_LIMIT = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")
-
-
-def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int, duration: int):
-    rr.init("lerobot_capture_camera_feed")
-    rr.spawn(memory_limit=RERUN_MEMORY_LIMIT)
-
-    now = dt.datetime.now()
-    capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
-    if not capture_dir.exists():
-        capture_dir.mkdir(parents=True, exist_ok=True)
-
-    # Opens the default webcam
-    cap = cv2.VideoCapture(0)
-    if not cap.isOpened():
-        print("Error: Could not open video stream.")
-        return
-
-    cap.set(cv2.CAP_PROP_FPS, fps)
-    cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
-    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
-
-    frame_index = 0
-    start_time = time.time()
-    while time.time() - start_time < duration:
-        ret, frame = cap.read()
-
-        if not ret:
-            print("Error: Could not read frame.")
-            break
-        rr.log("video/stream", rr.Image(frame), static=True)
-        cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
-        frame_index += 1
-
-    # Release the capture
-    cap.release()
-
-    # TODO(Steven): Add a graceful shutdown via a close() method for the Viewer context, though not currently supported in the Rerun API.
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-
-    parser.add_argument(
-        "--output-dir",
-        type=Path,
-        default=Path("outputs/cam_capture/"),
-        help="Directory where the capture images are written. A subfolder named with the current date & time will be created inside it for each capture.",
-    )
-    parser.add_argument(
-        "--fps",
-        type=int,
-        default=30,
-        help="Frames Per Second of the capture.",
-    )
-    parser.add_argument(
-        "--width",
-        type=int,
-        default=1280,
-        help="Width of the captured images.",
-    )
-    parser.add_argument(
-        "--height",
-        type=int,
-        default=720,
-        help="Height of the captured images.",
-    )
-    parser.add_argument(
-        "--duration",
-        type=int,
-        default=20,
-        help="Duration in seconds for which the video stream should be captured.",
-    )
-    args = parser.parse_args()
-    display_and_save_video_stream(**vars(args))

benchmarks/video/run_video_benchmark.py

@@ -21,11 +21,13 @@ See the provided README.md or run `python benchmark/video/run_video_benchmark.py
 
 import argparse
 import datetime as dt
+import itertools
 import random
 import shutil
 from collections import OrderedDict
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
+from threading import Lock
 
 import einops
 import numpy as np
@@ -35,13 +37,13 @@ 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,
+    decode_video_frames,
     encode_video_frames,
 )
 from lerobot.utils.constants import OBS_IMAGE
+from lerobot.utils.utils import TimerManager
 
 BASE_ENCODING = OrderedDict(
     [
@@ -86,7 +88,7 @@ def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> t
     frames = []
     for ts in timestamps:
         idx = int(ts * fps)
-        frame = PIL.Image.open(imgs_dir / f"frame_{idx:06d}.png")
+        frame = PIL.Image.open(imgs_dir / f"frame-{idx:06d}.png")
         frame = torch.from_numpy(np.array(frame))
         frame = frame.type(torch.float32) / 255
         frame = einops.rearrange(frame, "h w c -> c h w")
@@ -97,21 +99,21 @@ def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> t
 def save_decoded_frames(
     imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
 ) -> None:
-    if save_dir.exists() and len(list(save_dir.glob("frame_*.png"))) == len(timestamps):
+    if save_dir.exists() and len(list(save_dir.glob("frame-*.png"))) == len(timestamps):
         return
 
     save_dir.mkdir(parents=True, exist_ok=True)
     for i, ts in enumerate(timestamps):
         idx = int(ts * fps)
         frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
-        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame_{idx:06d}_decoded.png")
-        shutil.copyfile(imgs_dir / f"frame_{idx:06d}.png", save_dir / f"frame_{idx:06d}_original.png")
+        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame-{idx:06d}_decoded.png")
+        shutil.copyfile(imgs_dir / f"frame-{idx:06d}.png", save_dir / f"frame-{idx:06d}_original.png")
 
 
 def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
     episode_index = 0
     ep_num_images = dataset.meta.episodes["length"][episode_index]
-    if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
+    if imgs_dir.exists() and len(list(imgs_dir.glob("frame-*.png"))) == ep_num_images:
         return
 
     imgs_dir.mkdir(parents=True, exist_ok=True)
@@ -125,7 +127,7 @@ def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
         tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
     ):
         img = item[img_keys[0]]
-        img.save(str(imgs_dir / f"frame_{i:06d}.png"), quality=100)
+        img.save(str(imgs_dir / f"frame-{i:06d}.png"), quality=100)
 
         if i >= ep_num_images - 1:
             break
@@ -149,18 +151,6 @@ def sample_timestamps(timestamps_mode: str, ep_num_images: int, fps: int) -> lis
     return [idx / fps for idx in frame_indexes]
 
 
-def decode_video_frames(
-    video_path: str,
-    timestamps: list[float],
-    tolerance_s: float,
-    backend: str,
-) -> torch.Tensor:
-    if backend in ["pyav", "video_reader"]:
-        return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
-    else:
-        raise NotImplementedError(backend)
-
-
 def benchmark_decoding(
     imgs_dir: Path,
     video_path: Path,
@@ -172,8 +162,8 @@ def benchmark_decoding(
     num_workers: int = 4,
     save_frames: bool = False,
 ) -> dict:
-    def process_sample(sample: int):
-        time_benchmark = TimeBenchmark()
+    def process_sample(sample: int, lock: Lock):
+        time_benchmark = TimerManager(log=False)
         timestamps = sample_timestamps(timestamps_mode, ep_num_images, fps)
         num_frames = len(timestamps)
         result = {
@@ -182,13 +172,13 @@ def benchmark_decoding(
             "mse_values": [],
         }
 
-        with time_benchmark:
+        with time_benchmark, lock:
             frames = decode_video_frames(video_path, timestamps=timestamps, tolerance_s=5e-1, backend=backend)
-        result["load_time_video_ms"] = time_benchmark.result_ms / num_frames
+        result["load_time_video_ms"] = (time_benchmark.last * 1000) / num_frames
 
         with time_benchmark:
             original_frames = load_original_frames(imgs_dir, timestamps, fps)
-        result["load_time_images_ms"] = time_benchmark.result_ms / num_frames
+        result["load_time_images_ms"] = (time_benchmark.last * 1000) / num_frames
 
         frames_np, original_frames_np = frames.numpy(), original_frames.numpy()
         for i in range(num_frames):
@@ -215,8 +205,10 @@ def benchmark_decoding(
     # A sample is a single set of decoded frames specified by timestamps_mode (e.g. a single frame, 2 frames, etc.).
     # For each sample, we record metrics (loading time and quality metrics) which are then averaged over all samples.
     # As these samples are independent, we run them in parallel threads to speed up the benchmark.
+    # Use a single shared lock for all worker threads
+    shared_lock = Lock()
     with ThreadPoolExecutor(max_workers=num_workers) as executor:
-        futures = [executor.submit(process_sample, i) for i in range(num_samples)]
+        futures = [executor.submit(process_sample, i, shared_lock) for i in range(num_samples)]
         for future in tqdm(as_completed(futures), total=num_samples, desc="samples", leave=False):
             result = future.result()
             load_times_video_ms.append(result["load_time_video_ms"])
@@ -358,24 +350,27 @@ def main(
                 imgs_dir = output_dir / "images" / dataset.repo_id.replace("/", "_")
                 # We only use the first episode
                 save_first_episode(imgs_dir, dataset)
-                for key, values in tqdm(encoding_benchmarks.items(), desc="encodings (g, crf)", leave=False):
-                    for value in tqdm(values, desc=f"encodings ({key})", leave=False):
-                        encoding_cfg = BASE_ENCODING.copy()
-                        encoding_cfg["vcodec"] = video_codec
-                        encoding_cfg["pix_fmt"] = pixel_format
+                for duet in [
+                    dict(zip(encoding_benchmarks.keys(), unique_combination, strict=False))
+                    for unique_combination in itertools.product(*encoding_benchmarks.values())
+                ]:
+                    encoding_cfg = BASE_ENCODING.copy()
+                    encoding_cfg["vcodec"] = video_codec
+                    encoding_cfg["pix_fmt"] = pixel_format
+                    for key, value in duet.items():
                         encoding_cfg[key] = value
-                        args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
-                        video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
-                        benchmark_table += benchmark_encoding_decoding(
-                            dataset,
-                            video_path,
-                            imgs_dir,
-                            encoding_cfg,
-                            decoding_benchmarks,
-                            num_samples,
-                            num_workers,
-                            save_frames,
-                        )
+                    args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
+                    video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
+                    benchmark_table += benchmark_encoding_decoding(
+                        dataset,
+                        video_path,
+                        imgs_dir,
+                        encoding_cfg,
+                        decoding_benchmarks,
+                        num_samples,
+                        num_workers,
+                        save_frames,
+                    )
 
             # Save intermediate results
             benchmark_df = pd.DataFrame(benchmark_table, columns=headers)
@@ -409,9 +404,9 @@ if __name__ == "__main__":
         nargs="*",
         default=[
             "lerobot/pusht_image",
-            "aliberts/aloha_mobile_shrimp_image",
-            "aliberts/paris_street",
-            "aliberts/kitchen",
+            "lerobot/aloha_mobile_shrimp_image",
+            "lerobot/paris_street",
+            "lerobot/kitchen",
         ],
         help="Datasets repo-ids to test against. First episodes only are used. Must be images.",
     )
@@ -419,7 +414,7 @@ if __name__ == "__main__":
         "--vcodec",
         type=str,
         nargs="*",
-        default=["libx264", "hevc", "libsvtav1"],
+        default=["h264", "hevc", "libsvtav1"],
         help="Video codecs to be tested",
     )
     parser.add_argument(
@@ -468,7 +463,7 @@ if __name__ == "__main__":
         "--backends",
         type=str,
         nargs="*",
-        default=["pyav", "video_reader"],
+        default=["torchcodec", "pyav"],
         help="Torchvision decoding backend to be tested.",
     )
     parser.add_argument(

docker/Dockerfile.internal

@@ -75,6 +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

@@ -61,6 +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
@@ -17,22 +15,45 @@
     title: Train a Robot with RL
   - local: hilserl_sim
     title: Train RL in Simulation
-  - local: async
-    title: Use Async Inference
+  - local: multi_gpu_training
+    title: Multi GPU training
   title: "Tutorials"
 - sections:
   - local: lerobot-dataset-v3
     title: Using LeRobotDataset
   - local: porting_datasets_v3
     title: Porting Large Datasets
+  - local: using_dataset_tools
+    title: Using the Dataset Tools
   title: "Datasets"
 - sections:
+  - local: act
+    title: ACT
   - local: smolvla
-    title: Finetune SmolVLA
+    title: SmolVLA
+  - local: pi0
+    title: π₀ (Pi0)
+  - local: pi05
+    title: π₀.₅ (Pi05)
+  - local: groot
+    title: NVIDIA GR00T N1.5
+  title: "Policies"
+- sections:
+  - local: async
+    title: Use Async Inference
+  - local: rtc
+    title: Real-Time Chunking (RTC)
+  title: "Inference"
+- sections:
+  - local: envhub
+    title: Environments from the Hub
+  - local: envhub_leisaac
+    title: Control & Train Robots in Sim (LeIsaac)
   - local: libero
     title: Using Libero
-  title: "Policies"
-
+  - local: metaworld
+    title: Using MetaWorld
+  title: "Simulation"
 - sections:
   - local: introduction_processors
     title: Introduction to Robot Processors
@@ -42,6 +63,8 @@
     title: Implement your own processor
   - local: processors_robots_teleop
     title: Processors for Robots and Teleoperators
+  - local: env_processor
+    title: Environment Processors
   title: "Robot Processors"
 - sections:
   - local: so101

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/async_inference/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/async_inference/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
@@ -113,9 +113,9 @@ 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">
@@ -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/async_inference/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
@@ -196,7 +196,7 @@ client_cfg = RobotClientConfig(
     server_address="localhost:8080",
     policy_device="mps",
     policy_type="smolvla",
-    pretrained_name_or_path="fracapuano/smolvla_async",
+    pretrained_name_or_path="<user>/smolvla_async",
     chunk_size_threshold=0.5,
     actions_per_chunk=50,  # make sure this is less than the max actions of the policy
 )

docs/source/env_processor.mdx

@@ -0,0 +1,418 @@
+# Environment Processors
+
+Environment processors are a critical layer in LeRobot's data processing architecture that handle **environment-specific** transformations, separate from policy-specific processing. This separation of concerns enables cleaner code, better modularity, and easier experimentation with different environments and policies.
+
+## Why Environment Processors?
+
+When working with different robot environments (LIBERO, MetaWorld, Aloha, etc.), each environment often has unique data formats, coordinate systems, and conventions that need standardization **before** policy processing. Without environment processors, these transformations would be:
+
+1. **Hardcoded in environment code** - Making it difficult to experiment with different state representations
+2. **Duplicated across policies** - Each policy would need to handle environment-specific quirks
+3. **Mixed with policy logic** - Violating separation of concerns and making debugging harder
+
+Environment processors solve this by providing a **dedicated processing layer** between raw environment observations and policy inputs.
+
+## The Processing Pipeline
+
+Here's how data flows through the complete processing pipeline during evaluation:
+
+```python
+# In lerobot_eval.py rollout() function:
+
+# 1. Raw environment observation (numpy arrays, various formats)
+raw_observation = env.step(action)
+
+# 2. Convert numpy to torch, normalize images [0,1]
+observation = preprocess_observation(raw_observation)
+
+# 3. Add task metadata (for multi-task environments)
+observation = add_envs_task(env, observation)
+
+# 4. ENVIRONMENT-SPECIFIC preprocessing (NEW!)
+#    - Flatten robot states
+#    - Rotate images to match dataset conventions
+#    - Handle environment-specific coordinate systems
+observation = env_preprocessor(observation)
+
+# 5. POLICY-SPECIFIC preprocessing
+#    - Normalize with dataset statistics
+#    - Add batch dimensions
+#    - Move to GPU
+#    - Tokenize language instructions
+observation = preprocessor(observation)
+
+# 6. Policy inference
+action = policy.select_action(observation)
+
+# 7. POLICY-SPECIFIC postprocessing
+#    - Unnormalize actions
+#    - Remove batch dimensions
+action = postprocessor(action)
+
+# 8. ENVIRONMENT-SPECIFIC postprocessing (NEW!)
+#    - Convert action formats if needed
+#    - Apply environment-specific constraints
+action_transition = {"action": action}
+action_transition = env_postprocessor(action_transition)
+action = action_transition["action"]
+
+# 9. Execute in environment
+env.step(action)
+```
+
+## The Benefits
+
+### 1. **Separation of Concerns**
+
+Environment processors handle transformations specific to the **environment's data format**, while policy processors handle transformations specific to the **model's requirements**.
+
+```python
+# ❌ Before: Mixed concerns
+class LiberoVLAPolicy:
+    def preprocess(self, obs):
+        # Environment-specific: Flatten robot state (shouldn't be in policy!)
+        state = self._flatten_robot_state(obs["robot_state"])
+        # Policy-specific: Normalize with dataset stats
+        state = self.normalizer(state)
+        return state
+
+# ✅ After: Clear separation
+# Environment processor: Handles LIBERO's nested robot state
+env_preprocessor = LiberoProcessorStep()  # Flattens robot_state
+
+# Policy processor: Handles model requirements
+policy_preprocessor = NormalizerProcessorStep(stats=dataset_stats)
+```
+
+### 2. **Flexibility and Reusability**
+
+The same policy can work with different environment processors, and the same environment processor can work with different policies:
+
+```python
+# Use SmolVLA policy with LIBERO environment
+libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(libero_cfg)
+smolvla_preprocessor, smolvla_postprocessor = make_pre_post_processors(smolvla_cfg)
+
+# Or use ACT policy with the same LIBERO environment
+libero_preprocessor, libero_postprocessor = make_env_pre_post_processors(libero_cfg)
+act_preprocessor, act_postprocessor = make_pre_post_processors(act_cfg)
+```
+
+### 3. **Easier Experimentation**
+
+Want to try different state representations for LIBERO? Just create a new processor:
+
+```python
+# Original: 8D state (pos + quat→axisangle + gripper)
+@ProcessorStepRegistry.register("libero_processor")
+class LiberoProcessorStep(ObservationProcessorStep):
+    def _process_observation(self, obs):
+        eef_pos = robot_state["eef"]["pos"]          # 3D
+        eef_axisangle = quat2axisangle(quat)         # 3D
+        gripper = robot_state["gripper"]["qpos"]     # 2D
+        state = torch.cat([eef_pos, eef_axisangle, gripper], dim=-1)  # 8D
+        return state
+
+# Experiment: Add velocity for better control
+@ProcessorStepRegistry.register("libero_velocity_processor")
+class LiberoVelocityProcessorStep(ObservationProcessorStep):
+    def _process_observation(self, obs):
+        # Include velocities for 14D state
+        eef_pos = robot_state["eef"]["pos"]          # 3D
+        eef_axisangle = quat2axisangle(quat)         # 3D
+        eef_vel = robot_state["eef"]["vel"]          # 3D  (NEW)
+        gripper_pos = robot_state["gripper"]["qpos"] # 2D
+        gripper_vel = robot_state["gripper"]["qvel"] # 3D  (NEW)
+        state = torch.cat([eef_pos, eef_axisangle, eef_vel,
+                          gripper_pos, gripper_vel], dim=-1)  # 14D
+        return state
+```
+
+### 4. **Cleaner Environment Code**
+
+Environments expose **all available data** without needing to know what downstream models will use:
+
+```python
+# LIBERO environment exposes full robot state
+observation = {
+    "pixels": {"image": img, "image2": img2},
+    "robot_state": {
+        "eef": {"pos": ..., "quat": ..., "vel": ..., "mat": ..., "axisangle": ...},
+        "gripper": {"qpos": ..., "qvel": ...},
+        "joints": {"pos": ..., "vel": ...}
+    }
+}
+
+# Environment processor decides what to use
+# Policy processor handles model-specific transformations
+```
+
+## Using Environment Processors
+
+### Factory Function
+
+The `make_env_pre_post_processors` function follows the same pattern as `make_pre_post_processors` for policies:
+
+```python
+from lerobot.envs.factory import make_env_pre_post_processors
+from lerobot.envs.configs import LiberoEnv, PushtEnv
+
+# For LIBERO: Returns LiberoProcessorStep in preprocessor
+libero_cfg = LiberoEnv(task="libero_spatial", camera_name=["agentview"])
+env_preprocessor, env_postprocessor = make_env_pre_post_processors(libero_cfg)
+
+# For other environments: Returns identity processors (no-op)
+pusht_cfg = PushtEnv()
+env_preprocessor, env_postprocessor = make_env_pre_post_processors(pusht_cfg)
+```
+
+### Implementation in `envs/factory.py`
+
+```python
+def make_env_pre_post_processors(
+    env_cfg: EnvConfig,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+]:
+    """
+    Create preprocessor and postprocessor pipelines for environment observations.
+
+    Args:
+        env_cfg: The configuration of the environment.
+
+    Returns:
+        A tuple containing:
+            - preprocessor: Pipeline that processes environment observations
+            - postprocessor: Pipeline that processes environment outputs
+    """
+    # For LIBERO environments, add the LiberoProcessorStep to preprocessor
+    if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
+        preprocessor = PolicyProcessorPipeline(steps=[LiberoProcessorStep()])
+    else:
+        # For all other environments, return an identity preprocessor
+        preprocessor = PolicyProcessorPipeline(steps=[])
+
+    # Postprocessor is currently identity for all environments
+    # Future: Could add environment-specific action transformations
+    postprocessor = PolicyProcessorPipeline(steps=[])
+
+    return preprocessor, postprocessor
+```
+
+### Integration in Evaluation
+
+In `lerobot_eval.py`, the environment processors are created once and used throughout:
+
+```python
+def eval_main(cfg: EvalPipelineConfig):
+    # Create environment
+    envs = make_env(cfg.env, n_envs=cfg.eval.batch_size)
+
+    # Create policy
+    policy = make_policy(cfg=cfg.policy, env_cfg=cfg.env)
+
+    # Create policy processors
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=cfg.policy,
+        pretrained_path=cfg.policy.pretrained_path,
+    )
+
+    # Create environment processors (NEW!)
+    env_preprocessor, env_postprocessor = make_env_pre_post_processors(env_cfg=cfg.env)
+
+    # Run evaluation with both processor types
+    eval_policy_all(
+        envs=envs,
+        policy=policy,
+        env_preprocessor=env_preprocessor,      # Environment-specific
+        env_postprocessor=env_postprocessor,    # Environment-specific
+        preprocessor=preprocessor,              # Policy-specific
+        postprocessor=postprocessor,            # Policy-specific
+        n_episodes=cfg.eval.n_episodes,
+    )
+```
+
+## Example: LIBERO Environment Processor
+
+The `LiberoProcessorStep` demonstrates a real-world environment processor:
+
+```python
+from lerobot.processor.pipeline import ObservationProcessorStep
+
+@dataclass
+@ProcessorStepRegistry.register(name="libero_processor")
+class LiberoProcessorStep(ObservationProcessorStep):
+    """
+    Processes LIBERO observations into the LeRobot format.
+
+    **State Processing:**
+    - Extracts end-effector position (3D)
+    - Converts quaternion to axis-angle representation (3D)
+    - Extracts gripper joint positions (2D)
+    - Concatenates into 8D state vector
+
+    **Image Processing:**
+    - Rotates images 180° to match HuggingFaceVLA/libero convention
+    """
+
+    def _process_observation(self, observation):
+        processed_obs = observation.copy()
+
+        # Process images: Flip 180° for camera convention
+        for key in list(processed_obs.keys()):
+            if key.startswith("observation.images."):
+                img = processed_obs[key]
+                img = torch.flip(img, dims=[2, 3])  # Flip H and W
+                processed_obs[key] = img
+
+        # Process robot_state: Flatten to 8D vector
+        if "observation.robot_state" in processed_obs:
+            robot_state = processed_obs.pop("observation.robot_state")
+
+            eef_pos = robot_state["eef"]["pos"]           # (B, 3)
+            eef_quat = robot_state["eef"]["quat"]         # (B, 4)
+            gripper_qpos = robot_state["gripper"]["qpos"] # (B, 2)
+
+            # Convert quaternion to axis-angle
+            eef_axisangle = self._quat2axisangle(eef_quat)  # (B, 3)
+
+            # Concatenate into single state vector
+            state = torch.cat((eef_pos, eef_axisangle, gripper_qpos), dim=-1)
+            state = state.float()
+
+            processed_obs["observation.state"] = state
+
+        return processed_obs
+```
+
+### Why These Transformations?
+
+1. **Image Rotation**: The HuggingFaceVLA/libero dataset has images rotated 180° from the raw LIBERO simulator. The processor handles this convention mismatch so policies trained on the dataset work seamlessly.
+
+2. **State Flattening**: The raw LIBERO environment exposes nested dictionaries with all available state information (position, quaternion, velocity, matrix representation, etc.). The processor:
+   - Selects the relevant components (pos, quat, gripper)
+   - Converts quaternion to axis-angle (more suitable for learning)
+   - Flattens to a single 8D vector that policies expect
+
+3. **Flexibility**: The environment still exposes **all** raw data. If you want to try different state representations (e.g., including velocities, using matrix representation instead of axis-angle), you can create a new processor without modifying the environment code.
+
+## Adding Environment Processors for New Environments
+
+To add environment processors for a new environment:
+
+### 1. Create the Processor Step
+
+```python
+# In src/lerobot/processor/env_processor.py
+
+@dataclass
+@ProcessorStepRegistry.register(name="myenv_processor")
+class MyEnvProcessorStep(ObservationProcessorStep):
+    """Process observations from MyEnv."""
+
+    def _process_observation(self, observation):
+        processed = observation.copy()
+
+        # Your environment-specific transformations
+        if "myenv.specific.state" in processed:
+            state = processed.pop("myenv.specific.state")
+            # Transform to standard format
+            processed["observation.state"] = self._transform_state(state)
+
+        return processed
+```
+
+### 2. Update the Factory
+
+```python
+# In src/lerobot/envs/factory.py
+
+def make_env_pre_post_processors(env_cfg: EnvConfig):
+    if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
+        preprocessor = PolicyProcessorPipeline(steps=[LiberoProcessorStep()])
+    elif isinstance(env_cfg, MyEnvConfig) or "myenv" in env_cfg.type:
+        preprocessor = PolicyProcessorPipeline(steps=[MyEnvProcessorStep()])
+    else:
+        preprocessor = PolicyProcessorPipeline(steps=[])
+
+    postprocessor = PolicyProcessorPipeline(steps=[])
+    return preprocessor, postprocessor
+```
+
+### 3. Use in Evaluation
+
+No changes needed! The evaluation script automatically uses the appropriate processor:
+
+```bash
+lerobot-eval \
+    --policy.path=lerobot/my_policy \
+    --env.type=myenv \  # Automatically uses MyEnvProcessorStep
+    --eval.n_episodes=10
+```
+
+## Future: Environment Postprocessors
+
+Currently, postprocessors are identity (no-op) for all environments. Future use cases include:
+
+### Action Space Transformations
+
+```python
+@dataclass
+class MyEnvActionPostprocessor(ProcessorStep):
+    """Convert policy actions to environment-specific format."""
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        action = transition["action"]
+
+        # Example: Convert from Cartesian to joint space
+        if self.action_space == "joint":
+            action = self.ik_solver(action)
+
+        # Example: Apply environment-specific safety limits
+        action = torch.clamp(action, self.min_action, self.max_action)
+
+        transition["action"] = action
+        return transition
+```
+
+### Coordinate System Conversions
+
+```python
+@dataclass
+class CoordinateTransformPostprocessor(ProcessorStep):
+    """Transform actions between coordinate systems."""
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        action = transition["action"]
+
+        # Example: Policy outputs in world frame, env expects base frame
+        action = self.world_to_base_transform(action)
+
+        transition["action"] = action
+        return transition
+```
+
+## Best Practices
+
+1. **Keep environment processors simple**: They should only handle environment-specific data format issues, not complex learning-related transformations.
+
+2. **Use policy processors for model requirements**: Normalization, batching, device placement, and tokenization belong in policy processors.
+
+3. **Expose all data from environments**: Let processors decide what to use rather than hardcoding choices in the environment.
+
+4. **Document conventions**: Clearly document any coordinate system conventions, camera orientations, or data formats that your processor handles.
+
+5. **Test independently**: Environment processors should be testable without loading full policies or environments.
+
+## Summary
+
+Environment processors provide a **clean separation** between environment-specific data transformations and policy-specific model requirements. This architecture:
+
+- ✅ Enables easy experimentation with different state representations
+- ✅ Allows policies to work seamlessly across different environments
+- ✅ Keeps environment code focused on simulation/hardware interface
+- ✅ Makes processor pipelines more maintainable and debuggable
+- ✅ Follows the single responsibility principle
+
+The key insight: **Environments define data formats, processors standardize them, policies consume standardized data.** Each layer has a clear, focused responsibility.

docs/source/envhub.mdx

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

docs/source/envhub_leisaac.mdx

@@ -0,0 +1,301 @@
+# LeIsaac × LeRobot EnvHub
+
+LeRobot EnvHub now supports **imitation learning in simulation** with LeIsaac.
+Spin up everyday manipulation tasks, teleoperate the robot, collect demos, push them to the Hub, and train policies in LeRobot — all in one loop.
+
+[LeIsaac](https://github.com/LightwheelAI/leisaac) integrates with IsaacLab and the SO101 Leader/Follower setup to provide:
+
+- 🕹️ **Teleoperation-first workflows** for data collection
+- 📦 **Built-in data conversion** ready for LeRobot training
+- 🤖 **Everyday skills** like picking oranges, lifting cubes, cleaning tables, and folding cloth
+- ☁️ **Ongoing upgrades** from [LightWheel](https://lightwheel.ai/): cloud simulation, EnvHub support, Sim2Real tooling, and more
+
+Below you’ll find the currently supported LeIsaac tasks exposed through LeRobot EnvHub.
+
+# Available Environments
+
+The following table lists all available tasks and environments in LeIsaac x LeRobot Envhub. You can also get the latest list of environments by running the following command:
+
+```bash
+python scripts/environments/list_envs.py
+```
+
+| Task                                                                                                                                                            | Environment ID                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                | Task Description                                                                                                           | Related Robot                                              |
+| :-------------------------------------------------------------------------------------------------------------------------------------------------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------------- | :--------------------------------------------------------- |
+| <video src="https://github.com/user-attachments/assets/466eddff-f720-4f99-94d5-5e123e4c302c" autoplay loop muted playsinline style="max-width: 300px;"></video> | [LeIsaac-SO101-PickOrange-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/pick_orange/pick_orange_env_cfg.py)<br /><br />[LeIsaac-SO101-PickOrange-Direct-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/pick_orange/direct/pick_orange_env.py)                                                                                                                                                                                                                        | Pick three oranges and put them into the plate, then reset the arm to rest state.                                          | Single-Arm SO101 Follower                                  |
+| <video src="https://github.com/user-attachments/assets/1e4eb83a-0b38-40fb-a0b2-ddb0fe201e6d" autoplay loop muted playsinline style="max-width: 300px;"></video> | [LeIsaac-SO101-LiftCube-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/lift_cube/lift_cube_env_cfg.py)<br /><br />[LeIsaac-SO101-LiftCube-Direct-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/lift_cube/direct/lift_cube_env.py)                                                                                                                                                                                                                                    | Lift the red cube up.                                                                                                      | Single-Arm SO101 Follower                                  |
+| <video src="https://github.com/user-attachments/assets/e49d8f1c-dcc9-412b-a88f-100680d8a45b" autoplay loop muted playsinline style="max-width: 300px;"></video> | [LeIsaac-SO101-CleanToyTable-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/clean_toy_table/clean_toy_table_env_cfg.py)<br /><br />[LeIsaac-SO101-CleanToyTable-BiArm-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/clean_toy_table/clean_toy_table_bi_arm_env_cfg.py)<br /><br />[LeIsaac-SO101-CleanToyTable-BiArm-Direct-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/clean_toy_table/direct/clean_toy_table_bi_arm_env.py) | Pick two letter e objects into the box, and reset the arm to rest state.                                                   | Single-Arm SO101 Follower<br /><br />Bi-Arm SO101 Follower |
+| <video src="https://github.com/user-attachments/assets/e29a0f8a-9286-4ce6-b45d-342c3d3ba754" autoplay loop muted playsinline style="max-width: 300px;"></video> | [LeIsaac-SO101-FoldCloth-BiArm-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/fold_cloth/fold_cloth_bi_arm_env_cfg.py)<br /><br />[LeIsaac-SO101-FoldCloth-BiArm-Direct-v0](https://github.com/LightwheelAI/leisaac/blob/main/source/leisaac/leisaac/tasks/fold_cloth/direct/fold_cloth_bi_arm_env.py)                                                                                                                                                                                                    | Fold the cloth, and reset the arm to rest state.<br /><br />_Note: Only the DirectEnv support check_success in this task._ | Bi-Arm SO101 Follower                                      |
+
+# Load LeIsaac directly in LeRobot with one line of code
+
+> EnvHub: Share LeIsaac environments through HuggingFace
+
+[EnvHub](https://huggingface.co/docs/lerobot/envhub) is our reproducible environment hub, spin up a packaged simulation with one line, experiment immediately, and publish your own tasks for the community.
+
+LeIsaac offers EnvHub support so you can consume or share tasks with only a few commands.
+
+<video
+  controls
+  src="https://github.com/user-attachments/assets/687666f5-ebe0-421d-84a0-eb86116ac5f8"
+  style={{ width: "100%", maxWidth: "960px", borderRadius: "8px" }}
+/>
+
+## How to get started, environment Setup
+
+Run the following commands to setup your code environments:
+
+```bash
+# Refer to Getting Started/Installation to install leisaac firstly
+conda create -n leisaac_envhub python=3.11
+conda activate leisaac_envhub
+
+conda install -c "nvidia/label/cuda-12.8.1" cuda-toolkit
+pip install -U torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu128
+pip install 'leisaac[isaaclab] @ git+https://github.com/LightwheelAI/leisaac.git#subdirectory=source/leisaac' --extra-index-url https://pypi.nvidia.com
+
+# Install lerobot
+pip install lerobot==0.4.1
+
+# Fix numpy version
+pip install numpy==1.26.0
+```
+
+## Usage Example
+
+EnvHub exposes every LeIsaac-supported task in a uniform interface. The examples below load `so101_pick_orange` and demonstrate a random-action rollout and an interactive teleoperation.
+
+### Random Action
+
+<details>
+<summary>Click to expand code example</summary>
+
+```python
+# envhub_random_action.py
+
+import torch
+from lerobot.envs.factory import make_env
+
+# Load from the hub
+envs_dict = make_env("LightwheelAI/leisaac_env:envs/so101_pick_orange.py", n_envs=1, trust_remote_code=True)
+
+# Access the environment
+suite_name = next(iter(envs_dict))
+sync_vector_env = envs_dict[suite_name][0]
+# retrieve the isaac environment from the sync vector env
+env = sync_vector_env.envs[0].unwrapped
+
+# Use it like any gym environment
+obs, info = env.reset()
+
+while True:
+    action = torch.tensor(env.action_space.sample())
+    obs, reward, terminated, truncated, info = env.step(action)
+    if terminated or truncated:
+        obs, info = env.reset()
+
+env.close()
+```
+
+</details>
+
+```bash
+python envhub_random_action.py
+```
+
+You should see the SO101 arm swinging under purely random commands.
+
+### Teleoperation
+
+LeRobot’s teleoperation stack can drive the simulated arm.
+
+Connect the SO101 Leader controller, run the calibration command below.
+
+```bash
+lerobot-calibrate \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/ttyACM0 \
+    --teleop.id=leader
+```
+
+And then launch the teleop script.
+
+<details>
+<summary>Click to expand code example</summary>
+
+```python
+# envhub_teleop_example.py
+
+import logging
+import time
+import gymnasium as gym
+
+from dataclasses import asdict, dataclass
+from pprint import pformat
+
+from lerobot.teleoperators import (  # noqa: F401
+    Teleoperator,
+    TeleoperatorConfig,
+    make_teleoperator_from_config,
+    so101_leader,
+)
+from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.utils import init_logging
+from lerobot.envs.factory import make_env
+
+
+@dataclass
+class TeleoperateConfig:
+    teleop: TeleoperatorConfig
+    env_name: str = "so101_pick_orange"
+    fps: int = 60
+
+
+@dataclass
+class EnvWrap:
+    env: gym.Env
+
+
+def make_env_from_leisaac(env_name: str = "so101_pick_orange"):
+    envs_dict = make_env(
+        f'LightwheelAI/leisaac_env:envs/{env_name}.py',
+        n_envs=1,
+        trust_remote_code=True
+    )
+    suite_name = next(iter(envs_dict))
+    sync_vector_env = envs_dict[suite_name][0]
+    env = sync_vector_env.envs[0].unwrapped
+
+    return env
+
+
+def teleop_loop(teleop: Teleoperator, env: gym.Env, fps: int):
+    from leisaac.devices.action_process import preprocess_device_action
+    from leisaac.assets.robots.lerobot import SO101_FOLLOWER_MOTOR_LIMITS
+    from leisaac.utils.env_utils import dynamic_reset_gripper_effort_limit_sim
+
+    env_wrap = EnvWrap(env=env)
+
+    obs, info = env.reset()
+    while True:
+        loop_start = time.perf_counter()
+        if env.cfg.dynamic_reset_gripper_effort_limit:
+            dynamic_reset_gripper_effort_limit_sim(env, 'so101leader')
+
+        raw_action = teleop.get_action()
+        processed_action = preprocess_device_action(
+            dict(
+                so101_leader=True,
+                joint_state={
+                    k.removesuffix(".pos"): v for k, v in raw_action.items()},
+                motor_limits=SO101_FOLLOWER_MOTOR_LIMITS),
+            env_wrap
+        )
+        obs, reward, terminated, truncated, info = env.step(processed_action)
+        if terminated or truncated:
+            obs, info = env.reset()
+
+        dt_s = time.perf_counter() - loop_start
+        precise_sleep(1 / fps - dt_s)
+        loop_s = time.perf_counter() - loop_start
+        print(f"\ntime: {loop_s * 1e3:.2f}ms ({1 / loop_s:.0f} Hz)")
+
+
+def teleoperate(cfg: TeleoperateConfig):
+    init_logging()
+    logging.info(pformat(asdict(cfg)))
+
+    teleop = make_teleoperator_from_config(cfg.teleop)
+    env = make_env_from_leisaac(cfg.env_name)
+
+    teleop.connect()
+    if hasattr(env, 'initialize'):
+        env.initialize()
+    try:
+        teleop_loop(teleop=teleop, env=env, fps=cfg.fps)
+    except KeyboardInterrupt:
+        pass
+    finally:
+        teleop.disconnect()
+        env.close()
+
+
+def main():
+    teleoperate(TeleoperateConfig(
+        teleop=so101_leader.SO101LeaderConfig(
+            port="/dev/ttyACM0",
+            id='leader',
+            use_degrees=False,
+        ),
+        env_name="so101_pick_orange",
+        fps=60,
+    ))
+
+
+if __name__ == "__main__":
+    main()
+
+```
+
+</details>
+
+```bash
+python envhub_teleop_example.py
+```
+
+Running the script lets you operate the simulated arm using the physical Leader device.
+
+## ☁️ Cloud Simulation (No GPU Required)
+
+Don’t have a local GPU or the right drivers? No problem! You can run LeIsaac entirely in the cloud with zero setup.
+LeIsaac works out-of-the-box on **NVIDIA Brev**, giving you a fully configured environment directly in your browser.
+
+👉 **Start here:** [https://lightwheelai.github.io/leisaac/docs/cloud_simulation/nvidia_brev](https://lightwheelai.github.io/leisaac/docs/cloud_simulation/nvidia_brev)
+
+Once your instance is deployed, simply open the link for **port 80 (HTTP)** to launch **Visual Studio Code Server** (default password: `password`). From there, you can run simulations, edit code, and visualize IsaacLab environments — all from your web browser.
+
+**No GPU, no drivers, no local installation. Just click and run.**
+
+## Additional Notes
+
+We keep EnvHub coverage aligned with the LeIsaac task. Currently supported:
+
+- `so101_pick_orange`
+- `so101_lift_cube`
+- `so101_clean_toytable`
+- `bi_so101_fold_cloth`
+
+Switch tasks by targeting a different script when calling `make_env`, for example:
+
+```python
+envs_dict_pick_orange = make_env("LightwheelAI/leisaac_env:envs/so101_pick_orange.py", n_envs=1, trust_remote_code=True)
+envs_dict_lift_cube = make_env("LightwheelAI/leisaac_env:envs/so101_lift_cube.py", n_envs=1, trust_remote_code=True)
+envs_dict_clean_toytable = make_env("LightwheelAI/leisaac_env:envs/so101_clean_toytable.py", n_envs=1, trust_remote_code=True)
+envs_dict_fold_cloth = make_env("LightwheelAI/leisaac_env:envs/bi_so101_fold_cloth.py", n_envs=1, trust_remote_code=True)
+```
+
+Note: when working with `bi_so101_fold_cloth`, call `initialize()` immediately after retrieving the env before performing any other operations:
+
+<details>
+<summary>Click to expand code example</summary>
+
+```python
+import torch
+from lerobot.envs.factory import make_env
+
+# Load from the hub
+envs_dict = make_env("LightwheelAI/leisaac_env:envs/bi_so101_fold_cloth.py", n_envs=1, trust_remote_code=True)
+
+# Access the environment
+suite_name = next(iter(envs_dict))
+sync_vector_env = envs_dict[suite_name][0]
+# retrieve the isaac environment from the sync vector env
+env = sync_vector_env.envs[0].unwrapped
+
+# NOTE: initialize() first
+env.initialize()
+
+# other operation with env...
+```
+
+</details>

docs/source/groot.mdx

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

docs/source/il_robots.mdx

@@ -165,7 +165,7 @@ huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
+HF_USER=$(hf auth whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -393,7 +393,7 @@ import time
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 episode_idx = 0
@@ -415,7 +415,7 @@ for idx in range(dataset.num_frames):
     }
     robot.send_action(action)
 
-    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+    precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
 
 robot.disconnect()
 ```
@@ -513,13 +513,14 @@ 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.policies.factory import make_processor
+
 
 NUM_EPISODES = 5
 FPS = 30
@@ -562,7 +563,7 @@ init_rerun(session_name="recording")
 # Connect the robot
 robot.connect()
 
-preprocessor, postprocessor = make_processor(
+preprocessor, postprocessor = make_pre_post_processors(
     policy_cfg=policy,
     pretrained_path=HF_MODEL_ID,
     dataset_stats=dataset.meta.stats,

docs/source/il_sim.mdx

@@ -1,220 +0,0 @@
-# Imitation Learning in Sim
-
-This tutorial will explain how to train a neural network to control a robot in simulation with imitation learning.
-
-**You'll learn:**
-
-1. How to record a dataset in simulation with [gym-hil](https://github.com/huggingface/gym-hil) and visualize the dataset.
-2. How to train a policy using your data.
-3. How to evaluate your policy in simulation and visualize the results.
-
-For the simulation environment we use the same [repo](https://github.com/huggingface/gym-hil) that is also being used by the Human-In-the-Loop (HIL) reinforcement learning algorithm.
-This environment is based on [MuJoCo](https://mujoco.org) and allows you to record datasets in LeRobotDataset format.
-Teleoperation is easiest with a controller like the Logitech F710, but you can also use your keyboard if you are up for the challenge.
-
-## Installation
-
-First, install the `gym_hil` package within the LeRobot environment, go to your LeRobot folder and run this command:
-
-```bash
-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/lerobot/config_examples/resolve/main/sim_il/env_config.json).
-
-To teleoperate and collect a dataset, we need to modify this config file. Here's an example configuration for imitation learning data collection:
-
-```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"
-}
-```
-
-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:
-
-<hfoptions id="teleop_sim">
-<hfoption id="Linux">
-
-```bash
-python -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
-```
-
-</hfoption>
-<hfoption id="MacOS">
-
-```bash
-mjpython -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
-```
-
-</hfoption>
-</hfoptions>
-
-Once rendered you can teleoperate the robot with the gamepad or keyboard, below you can find the gamepad/keyboard controls.
-
-Note that to teleoperate the robot you have to hold the "Human Take Over Pause Policy" Button `RB` to enable control!
-
-**Gamepad Controls**
-
-<p align="center">
-  <img
-    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/gamepad_guide.jpg?raw=true"
-    alt="Figure shows the control mappings on a Logitech gamepad."
-    title="Gamepad Control Mapping"
-    width="100%"
-  ></img>
-</p>
-<p align="center">
-  <i>Gamepad button mapping for robot control and episode management</i>
-</p>
-
-**Keyboard controls**
-
-For keyboard controls use the `spacebar` to enable control and the following keys to move the robot:
-
-```bash
-  Arrow keys: Move in X-Y plane
-  Shift and Shift_R: Move in Z axis
-  Right Ctrl and Left Ctrl: Open and close gripper
-  ESC: Exit
-```
-
-## Visualize a dataset
-
-If you uploaded your dataset to the hub you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id.
-
-<p align="center">
-  <img
-    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/dataset_visualizer_sim.png"
-    alt="Figure shows the dataset visualizer"
-    title="Dataset visualization"
-    width="100%"
-  ></img>
-</p>
-<p align="center">
-  <i>Dataset visualizer</i>
-</p>
-
-## Train a policy
-
-To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=${HF_USER}/il_gym \
-  --policy.type=act \
-  --output_dir=outputs/train/il_sim_test \
-  --job_name=il_sim_test \
-  --policy.device=cuda \
-  --wandb.enable=true
-```
-
-Let's explain the command:
-
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/il_gym`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-3. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
-4. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-
-Training should take several hours, 100k steps (which is the default) will take about 1h on Nvidia A100. You will find checkpoints in `outputs/train/il_sim_test/checkpoints`.
-
-#### Train using Collab
-
-If your local computer doesn't have a powerful GPU you could utilize Google Collab to train your model by following the [ACT training notebook](./notebooks#training-act).
-
-#### Upload policy checkpoints
-
-Once training is done, upload the latest checkpoint with:
-
-```bash
-huggingface-cli upload ${HF_USER}/il_sim_test \
-  outputs/train/il_sim_test/checkpoints/last/pretrained_model
-```
-
-You can also upload intermediate checkpoints with:
-
-```bash
-CKPT=010000
-huggingface-cli upload ${HF_USER}/il_sim_test${CKPT} \
-  outputs/train/il_sim_test/checkpoints/${CKPT}/pretrained_model
-```
-
-## Evaluate your policy in Sim
-
-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).
-
-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
-
-<hfoptions id="eval_policy">
-<hfoption id="Linux">
-
-```bash
-python -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
-```
-
-</hfoption>
-<hfoption id="MacOS">
-
-```bash
-mjpython -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
-```
-
-</hfoption>
-</hfoptions>
-
-> [!WARNING]
-> While the main workflow of training ACT in simulation is straightforward, there is significant room for exploring how to set up the task, define the initial state of the environment, and determine the type of data required during collection to learn the most effective policy. If your trained policy doesn't perform well, investigate the quality of the dataset it was trained on using our visualizers, as well as the action values and various hyperparameters related to ACT and the simulation.
-
-Congrats 🎉, you have finished this tutorial. If you want to continue with using LeRobot in simulation follow this [Tutorial on reinforcement learning in sim with HIL-SERL](https://huggingface.co/docs/lerobot/hilserl_sim)
-
-> [!TIP]
-> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).

docs/source/installation.mdx

@@ -1,8 +1,15 @@
 # Installation
 
+## Install [`miniforge`](https://conda-forge.org/download/)
+
+```bash
+wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
+bash Miniforge3-$(uname)-$(uname -m).sh
+```
+
 ## Environment Setup
 
-Create a virtual environment with Python 3.10, using [`Miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install)
+Create a virtual environment with Python 3.10, using conda:
 
 ```bash
 conda create -y -n lerobot python=3.10
@@ -14,7 +21,7 @@ Then activate your conda environment, you have to do this each time you open a s
 conda activate lerobot
 ```
 
-When using `miniconda`, install `ffmpeg` in your environment:
+When using `conda`, install `ffmpeg` in your environment:
 
 ```bash
 conda install ffmpeg -c conda-forge
@@ -74,6 +81,9 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
+> [!NOTE]
+> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
+
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
@@ -91,7 +101,7 @@ 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.
 
@@ -208,34 +208,36 @@ LeRobot supports saving and loading calibration data automatically. This is usef
 
 <!-- prettier-ignore-start -->
 ```python
-> @property
-> def is_calibrated(self) -> bool:
->    return True
->
-> def calibrate(self) -> None:
->    pass
-> ```
+@property
+def is_calibrated(self) -> bool:
+    return True
+
+def calibrate(self) -> None:
+    pass
+```
+<!-- prettier-ignore-end -->
 
 ### `is_calibrated`
 
 This should reflect whether your robot has the required calibration loaded.
 
-```
-<!-- prettier-ignore-end -->python
+<!-- prettier-ignore-start -->
+```python
 @property
 def is_calibrated(self) -> bool:
     return self.bus.is_calibrated
 ```
+<!-- prettier-ignore-end -->
 
 ### `calibrate()`
 
 The goal of the calibration is twofold:
-    - Know the physical range of motion of each motors in order to only send commands within this range.
-    - Normalize raw motors positions to sensible continuous values (e.g. percentages, degrees) instead of arbitrary discrete value dependant on the specific motor used that will not replicate elsewhere.
+
+- Know the physical range of motion of each motors in order to only send commands within this range.
+- Normalize raw motors positions to sensible continuous values (e.g. percentages, degrees) instead of arbitrary discrete value dependant on the specific motor used that will not replicate elsewhere.
 
 It should implement the logic for calibration (if relevant) and update the `self.calibration` dictionary. If you are using Feetech or Dynamixel motors, our bus interfaces already include methods to help with this.
 
-
 <!-- prettier-ignore-start -->
 ```python
 def calibrate(self) -> None:
@@ -335,6 +337,134 @@ 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

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

docs/source/lerobot-dataset-v3.mdx

@@ -279,3 +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

@@ -125,3 +125,42 @@ lerobot-train \
 LeRobot uses MuJoCo for simulation. You need to set the rendering backend before training or evaluation:
 
 - `export MUJOCO_GL=egl` → for headless servers (e.g. HPC, cloud)
+
+## Reproducing π₀.₅ results
+
+We reproduce the results of π₀.₅ on the LIBERO benchmark using the LeRobot implementation. We take the Physical Intelligence LIBERO base model (`pi05_libero`) and finetune for an additional 6k steps in bfloat16, with batch size of 256 on 8 H100 GPUs using the [HuggingFace LIBERO dataset](https://huggingface.co/datasets/HuggingFaceVLA/libero).
+
+The finetuned model can be found here:
+
+- **π₀.₅ LIBERO**: [lerobot/pi05_libero_finetuned](https://huggingface.co/lerobot/pi05_libero_finetuned)
+
+We then evaluate the finetuned model using the LeRobot LIBERO implementation, by running the following command:
+
+```bash
+lerobot-eval \
+  --output_dir=/logs/ \
+  --env.type=libero \
+  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
+  --eval.batch_size=1 \
+  --eval.n_episodes=10 \
+  --policy.path=pi05_libero_finetuned \
+  --policy.n_action_steps=10 \
+  --output_dir=./eval_logs/ \
+  --env.max_parallel_tasks=1
+```
+
+**Note:** We set `n_action_steps=10`, similar to the original OpenPI implementation.
+
+### Results
+
+We obtain the following results on the LIBERO benchmark:
+
+| Model    | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average  |
+| -------- | -------------- | ------------- | ----------- | --------- | -------- |
+| **π₀.₅** | 97.0           | 99.0          | 98.0        | 96.0      | **97.5** |
+
+These results are consistent with the original [results](https://github.com/Physical-Intelligence/openpi/tree/main/examples/libero#results) reported by Physical Intelligence:
+
+| Model    | LIBERO Spatial | LIBERO Object | LIBERO Goal | LIBERO 10 | Average   |
+| -------- | -------------- | ------------- | ----------- | --------- | --------- |
+| **π₀.₅** | 98.8           | 98.2          | 98.0        | 92.4      | **96.85** |

docs/source/metaworld.mdx

@@ -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/multi_gpu_training.mdx

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

docs/source/phone_teleop.mdx

@@ -79,7 +79,7 @@ After running the example:
 - Android: after starting the script, open the printed local URL on your phone, tap Start, then press and hold Move.
 - iOS: open HEBI Mobile I/O first; B1 enables motion. A3 controls the gripper.
 
-Additionally you can customize mapping or safety limits by editing the processor steps shown in the examples. You can also remap inputs (e.g., use a different analog input) or adapt the pipeline to other robots (e.g., LeKiwi) by modifying the input and kinematics steps. More about this in the [Processors for Robots and Teleoperators](./processors_robots_teleop.mdx) guide.
+Additionally you can customize mapping or safety limits by editing the processor steps shown in the examples. You can also remap inputs (e.g., use a different analog input) or adapt the pipeline to other robots (e.g., LeKiwi) by modifying the input and kinematics steps. More about this in the [Processors for Robots and Teleoperators](./processors_robots_teleop) guide.
 
 - Run this example to record a dataset, which saves absolute end effector observations and actions:
 

docs/source/pi0.mdx

@@ -0,0 +1,84 @@
+# 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]"
+   ```
+
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
+## 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 +1,112 @@
+# 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]"
+   ```
+
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
+## 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/policy_groot_README.md

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

docs/source/rtc.mdx

@@ -0,0 +1,188 @@
+# Real-Time Chunking (RTC)
+
+Real-Time Chunking (RTC) is an inference-time method that allows large, flow-matching based robotic policies, such as [Pi0](./pi0), [Pi0.5](./pi05), and [SmolVLA](./smolvla), to produce smooth, continuous, and reactive motion despite having high inference latency.
+
+These policies generate chunks of future actions (e.g., 50 steps at a time) instead of single actions.
+Because the models are large, producing each chunk takes longer than the time it takes the robot to execute it.
+Naively executing chunks leads to problems such as pauses, jerky transitions, or sudden changes in strategy whenever the next chunk arrives late or disagrees with the previously executed actions.
+
+RTC solves this by asynchronously generating the next chunk while the robot continues executing the current one, and by guiding the new chunk so it aligns smoothly with the portion of the previous chunk that has already been executed.
+
+## How RTC Works (simplified)
+
+RTC lets the robot think ahead while it’s still moving. When the robot is carrying out one chunk of actions, RTC starts creating the next chunk early.
+But since the robot has already moved a bit by the time the new chunk is ready, RTC has to make sure the new chunk still lines up smoothly with what the robot is currently doing.
+
+To do this, RTC treats the beginning of the new chunk like an inpainting or “fill-in-the-gaps” problem:
+it gently adjusts the first part of the new chunk so it blends naturally with the robot’s ongoing motion. The result is no pauses, no sudden jumps.
+
+In technical terms, RTC adds a guidance term to the flow-matching denoising process that forces the overlapping timesteps of the new chunk to stay close to the executed portion of the previous chunk, typically using a soft transition mask.
+
+## Quick Start
+
+### Installation
+
+RTC is built into LeRobot. Just install the policy dependencies you need:
+
+```bash
+# For Pi0 or Pi0.5
+pip install -e ".[pi]"
+
+# For SmolVLA
+pip install -e ".[smolvla]"
+```
+
+### Using RTC with Pi0
+
+You can find a complete reference implementation in [eval_with_real_robot.py](examples/rtc/eval_with_real_robot.py).
+The snippet below provides a simplified pseudo-example of how RTC operates with Pi0 in your pipeline:
+
+```python
+from lerobot.policies.pi0 import PI0Policy, PI0Config
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.action_queue import ActionQueue
+
+# Load Pi0 with RTC enabled
+policy_cfg = PI0Config()
+
+# Enable RTC
+policy_cfg.rtc_config = RTCConfig(
+    enabled=True,
+    execution_horizon=10,  # How many steps to blend with previous chunk
+    max_guidance_weight=10.0,  # How strongly to enforce consistency
+    prefix_attention_schedule=RTCAttentionSchedule.EXP,  # Exponential blend
+)
+
+# Load the policy
+policy = PI0Policy.from_pretrained("lerobot/pi0_base", policy_cfg=policy_cfg, device="cuda")
+
+# Now use predict_action_chunk with RTC parameters
+inference_delay = 4  # How many steps of inference latency, this values should be calculated based on the inference latency of the policy
+
+# Initialize the action queue
+action_queue = ActionQueue(policy_cfg.rtc_config)
+
+# Start in a separate thread with the following function
+def get_actions():
+  while True:
+    if should_get_actions:
+
+      prev_actions = action_queue.get_left_over()
+      obs = get_robot_observations(robot)
+
+      # Generate actions WITH RTC
+      actions = policy.predict_action_chunk(
+          obs,
+          inference_delay=inference_delay,
+          prev_chunk_left_over=prev_actions,
+      )
+
+      action_queue.merge(
+          actions, actions, inference_delay
+      )
+
+for step in range(num_steps):
+    action = action_queue.get()
+
+    # Execute the first N actions
+    execute_actions(action)
+```
+
+## Key Parameters
+
+`RTCConfig` has the following parameters to tune:
+
+**`execution_horizon`**: How many timesteps from the previous chunk to maintain consistency with. Higher values mean smoother transitions but potentially less reactivity.
+
+Typical values: 8-12 steps
+
+```python
+RTCConfig(execution_horizon=10)
+```
+
+**`max_guidance_weight`**: How strongly to enforce consistency with the previous chunk. This is a hyperparameter that can be tuned to balance the smoothness of the transitions and the reactivity of the policy. For 10 steps flow matching (SmolVLA, Pi0, Pi0.5), a value of 10.0 is a optimal value.
+
+**`prefix_attention_schedule`**: How to weight consistency across the overlap region.
+
+- `LINEAR`: Linear decay from inference_delay to execution_horizon
+- `EXP`: Exponential decay (recommended for getting started)
+- `ONES`: Full weight across entire execution_horizon
+- `ZEROS`: Binary (full weight up to inference_delay, then zero)
+
+**`inference_delay`**: How many timesteps of inference latency your system has. This is passed to `predict_action_chunk()` rather than the config, since it may vary at runtime.
+
+## Testing RTC Offline
+
+Before running on a real robot, test RTC with dataset samples to visualize how it works:
+
+```bash
+python examples/rtc/eval_dataset.py \
+    --policy.path=lerobot/pi0_libero_finetuned \
+    --dataset.repo_id=HuggingFaceVLA/libero \
+    --rtc.execution_horizon=10 \
+    --rtc.max_guidance_weight=10.0 \
+    --device=cuda
+```
+
+The script generates a visualization of the denoising process, comparing standard generation (left) with RTC (right). In the RTC plots, you can see how the first few steps (blue/purple lines) are guided to match the red ground truth trajectory (previous chunk's tail), ensuring a smooth transition between chunks.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/flow_matching.png"
+    alt="Denoising steps with and without RTC"
+    width="100%"
+  />
+</p>
+
+## Testing RTC with a Real Robot
+
+```bash
+python examples/rtc/eval_with_real_robot.py \
+    --policy.path=${HF_USERNAME}/policy_repo_id \
+    --robot.type=so100_follower \
+    --robot.port=/dev/tty.usbmodem58FA0834591 \
+    --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --task="Move green small object into the purple platform" \
+    --duration=120 \
+    --device=cuda
+```
+
+## How It Differs from the Async Inference in LeRobot
+
+Both RTC and [async inference](./async) improve real-time robot control, but they solve different problems.
+
+| Aspect        | Async Inference                                                            | RTC                                                 |
+| ------------- | -------------------------------------------------------------------------- | --------------------------------------------------- |
+| **Problem**   | Idle frames while waiting for inference                                    | Discontinuities between action chunks               |
+| **Solution**  | Decouple prediction from execution                                         | Guide new chunks to continue smoothly from previous |
+| **Benefit**   | No waiting, continuous action                                              | Smooth transitions, natural motion                  |
+| **Best Used** | Async inference is best used with large models with high inference latency | Flow-matching based policies                        |
+
+**Use both together** for maximum smoothness and reactivity!
+
+## Advanced: Debug Tracking
+
+RTC includes built-in debug tracking to help you understand what's happening during inference:
+
+```python
+# Enable debug tracking
+policy_cfg.rtc_config.debug = True
+policy_cfg.rtc_config.debug_maxlen = 100
+
+# After inference, access debug data
+debug_data = policy.rtc_processor.get_debug_data()
+
+# Visualize denoising steps, corrections, etc.
+from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer
+visualizer = RTCDebugVisualizer()
+# ... create plots
+```
+
+See `examples/rtc/eval_dataset.py` for a complete example of visualization.
+
+## References
+
+- [Smooth-As-Butter Robot Policies](https://alexander-soare.github.io/robotics/2025/08/05/smooth-as-butter-robot-policies.html) - Excellent technical explanation with real robot results
+- [Physical Intelligence - Real-Time Chunking](https://www.physicalintelligence.company/research/real_time_chunking) - Original paper and research
+- [Kinetix RTC Implementation](https://github.com/Physical-Intelligence/real-time-chunking-kinetix) - Reference implementation from Physical Intelligence

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!
 

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/backward_compatibility/replay.py

@@ -45,7 +45,7 @@ from lerobot.robots import (  # noqa: F401
     so101_follower,
 )
 from lerobot.utils.constants import ACTION
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import (
     init_logging,
     log_say,
@@ -97,7 +97,7 @@ def replay(cfg: ReplayConfig):
         robot.send_action(action)
 
         dt_s = time.perf_counter() - start_episode_t
-        busy_wait(1 / dataset.fps - dt_s)
+        precise_sleep(1 / dataset.fps - dt_s)
 
     robot.disconnect()
 

examples/dataset/load_lerobot_dataset.py

@@ -34,115 +34,118 @@ from huggingface_hub import HfApi
 import lerobot
 from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 
-# We ported a number of existing datasets ourselves, use this to see the list:
-print("List of available datasets:")
-pprint(lerobot.available_datasets)
 
-# You can also browse through the datasets created/ported by the community on the hub using the hub api:
-hub_api = HfApi()
-repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
-pprint(repo_ids)
+def main():
+    # We ported a number of existing datasets ourselves, use this to see the list:
+    print("List of available datasets:")
+    pprint(lerobot.available_datasets)
 
-# Or simply explore them in your web browser directly at:
-# https://huggingface.co/datasets?other=LeRobot
+    # You can also browse through the datasets created/ported by the community on the hub using the hub api:
+    hub_api = HfApi()
+    repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
+    pprint(repo_ids)
 
-# Let's take this one for this example
-repo_id = "lerobot/aloha_mobile_cabinet"
-# We can have a look and fetch its metadata to know more about it:
-ds_meta = LeRobotDatasetMetadata(repo_id)
+    # Or simply explore them in your web browser directly at:
+    # https://huggingface.co/datasets?other=LeRobot
 
-# By instantiating just this class, you can quickly access useful information about the content and the
-# structure of the dataset without downloading the actual data yet (only metadata files — which are
-# lightweight).
-print(f"Total number of episodes: {ds_meta.total_episodes}")
-print(f"Average number of frames per episode: {ds_meta.total_frames / ds_meta.total_episodes:.3f}")
-print(f"Frames per second used during data collection: {ds_meta.fps}")
-print(f"Robot type: {ds_meta.robot_type}")
-print(f"keys to access images from cameras: {ds_meta.camera_keys=}\n")
+    # Let's take this one for this example
+    repo_id = "lerobot/aloha_mobile_cabinet"
+    # We can have a look and fetch its metadata to know more about it:
+    ds_meta = LeRobotDatasetMetadata(repo_id)
 
-print("Tasks:")
-print(ds_meta.tasks)
-print("Features:")
-pprint(ds_meta.features)
+    # By instantiating just this class, you can quickly access useful information about the content and the
+    # structure of the dataset without downloading the actual data yet (only metadata files — which are
+    # lightweight).
+    print(f"Total number of episodes: {ds_meta.total_episodes}")
+    print(f"Average number of frames per episode: {ds_meta.total_frames / ds_meta.total_episodes:.3f}")
+    print(f"Frames per second used during data collection: {ds_meta.fps}")
+    print(f"Robot type: {ds_meta.robot_type}")
+    print(f"keys to access images from cameras: {ds_meta.camera_keys=}\n")
 
-# You can also get a short summary by simply printing the object:
-print(ds_meta)
+    print("Tasks:")
+    print(ds_meta.tasks)
+    print("Features:")
+    pprint(ds_meta.features)
 
-# You can then load the actual dataset from the hub.
-# Either load any subset of episodes:
-dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
+    # You can also get a short summary by simply printing the object:
+    print(ds_meta)
 
-# And see how many frames you have:
-print(f"Selected episodes: {dataset.episodes}")
-print(f"Number of episodes selected: {dataset.num_episodes}")
-print(f"Number of frames selected: {dataset.num_frames}")
+    # You can then load the actual dataset from the hub.
+    # Either load any subset of episodes:
+    dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
 
-# Or simply load the entire dataset:
-dataset = LeRobotDataset(repo_id)
-print(f"Number of episodes selected: {dataset.num_episodes}")
-print(f"Number of frames selected: {dataset.num_frames}")
+    # And see how many frames you have:
+    print(f"Selected episodes: {dataset.episodes}")
+    print(f"Number of episodes selected: {dataset.num_episodes}")
+    print(f"Number of frames selected: {dataset.num_frames}")
 
-# The previous metadata class is contained in the 'meta' attribute of the dataset:
-print(dataset.meta)
+    # Or simply load the entire dataset:
+    dataset = LeRobotDataset(repo_id)
+    print(f"Number of episodes selected: {dataset.num_episodes}")
+    print(f"Number of frames selected: {dataset.num_frames}")
 
-# LeRobotDataset actually wraps an underlying Hugging Face dataset
-# (see https://huggingface.co/docs/datasets for more information).
-print(dataset.hf_dataset)
+    # The previous metadata class is contained in the 'meta' attribute of the dataset:
+    print(dataset.meta)
 
-# LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
-# with the latter, like iterating through the dataset.
-# The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
-# episodes, you can access the frame indices of any episode using dataset.meta.episodes. Here, we access
-# frame indices associated to the first episode:
-episode_index = 0
-from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
-to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
+    # LeRobotDataset actually wraps an underlying Hugging Face dataset
+    # (see https://huggingface.co/docs/datasets for more information).
+    print(dataset.hf_dataset)
 
-# Then we grab all the image frames from the first camera:
-camera_key = dataset.meta.camera_keys[0]
-frames = [dataset[idx][camera_key] for idx in range(from_idx, to_idx)]
+    # LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
+    # with the latter, like iterating through the dataset.
+    # The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
+    # episodes, you can access the frame indices of any episode using dataset.meta.episodes. Here, we access
+    # frame indices associated to the first episode:
+    episode_index = 0
+    from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
+    to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
 
-# The objects returned by the dataset are all torch.Tensors
-print(type(frames[0]))
-print(frames[0].shape)
+    # Then we grab all the image frames from the first camera:
+    camera_key = dataset.meta.camera_keys[0]
+    frames = [dataset[idx][camera_key] for idx in range(from_idx, to_idx)]
 
-# Since we're using pytorch, the shape is in pytorch, channel-first convention (c, h, w).
-# We can compare this shape with the information available for that feature
-pprint(dataset.features[camera_key])
-# In particular:
-print(dataset.features[camera_key]["shape"])
-# The shape is in (h, w, c) which is a more universal format.
+    # The objects returned by the dataset are all torch.Tensors
+    print(type(frames[0]))
+    print(frames[0].shape)
 
-# For many machine learning applications we need to load the history of past observations or trajectories of
-# future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
-# differences with the current loaded frame. For instance:
-delta_timestamps = {
-    # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
-    camera_key: [-1, -0.5, -0.20, 0],
-    # loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
-    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
-    # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
-    "action": [t / dataset.fps for t in range(64)],
-}
-# Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
-# timestamp, you still get a valid timestamp.
+    # Since we're using pytorch, the shape is in pytorch, channel-first convention (c, h, w).
+    # We can compare this shape with the information available for that feature
+    pprint(dataset.features[camera_key])
+    # In particular:
+    print(dataset.features[camera_key]["shape"])
+    # The shape is in (h, w, c) which is a more universal format.
 
-dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
-print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
-print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
-print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
+    # For many machine learning applications we need to load the history of past observations or trajectories of
+    # future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
+    # differences with the current loaded frame. For instance:
+    delta_timestamps = {
+        # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
+        camera_key: [-1, -0.5, -0.20, 0],
+        # loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
+        "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
+        # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
+        "action": [t / dataset.fps for t in range(64)],
+    }
+    # Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
+    # timestamp, you still get a valid timestamp.
 
-# Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
-# PyTorch datasets.
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    num_workers=4,
-    batch_size=32,
-    shuffle=True,
-)
+    dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
+    print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
+    print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
+    print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
 
-for batch in dataloader:
-    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
-    print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
-    print(f"{batch['action'].shape=}")  # (32, 64, c)
-    break
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=4,
+        batch_size=32,
+        shuffle=True,
+    )
+    for batch in dataloader:
+        print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
+        print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
+        print(f"{batch['action'].shape=}")  # (32, 64, c)
+        break
+
+
+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

@@ -33,83 +33,68 @@ 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 & robot
-robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 
-robot = LeKiwiClient(robot_config)
+def main():
+    # Create the robot configuration & robot
+    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 
-# Create policy
-policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+    robot = LeKiwiClient(robot_config)
 
-# Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, ACTION)
-obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
-dataset_features = {**action_features, **obs_features}
+    # Create policy
+    policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_DATASET_ID,
-    fps=FPS,
-    features=dataset_features,
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
+    # Configure the dataset features
+    action_features = hw_to_dataset_features(robot.action_features, ACTION)
+    obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
+    dataset_features = {**action_features, **obs_features}
 
-# 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()
-
-# 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}")
-
-    # Main record loop
-    record_loop(
-        robot=robot,
-        events=events,
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_DATASET_ID,
         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,
+        features=dataset_features,
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and (
-        (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-    ):
-        log_say("Reset the environment")
+    # 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()
+
+    # 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}")
+
+        # 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,
@@ -118,19 +103,42 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             robot_observation_processor=robot_observation_processor,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-record episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (
+            (recorded_episodes < 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=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    recorded_episodes += 1
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-listener.stop()
-dataset.push_to_hub()
+        # Save episode
+        dataset.save_episode()
+        recorded_episodes += 1
+
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/record.py

@@ -34,78 +34,62 @@ 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)
+def main():
+    # 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()
 
-# TODO(Steven): Update this example to use pipelines
-teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+    # Initialize the robot and teleoperator
+    robot = LeKiwiClient(robot_config)
+    leader_arm = SO100Leader(leader_arm_config)
+    keyboard = KeyboardTeleop(keyboard_config)
 
-# Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, ACTION)
-obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
-dataset_features = {**action_features, **obs_features}
+    # TODO(Steven): Update this example to use pipelines
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
 
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_REPO_ID,
-    fps=FPS,
-    features=dataset_features,
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
+    # Configure the dataset features
+    action_features = hw_to_dataset_features(robot.action_features, ACTION)
+    obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
+    dataset_features = {**action_features, **obs_features}
 
-# 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()
-
-# 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 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}")
-
-    # Main record loop
-    record_loop(
-        robot=robot,
-        events=events,
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_REPO_ID,
         fps=FPS,
-        dataset=dataset,
-        teleop=[leader_arm, keyboard],
-        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,
+        features=dataset_features,
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and (
-        (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-    ):
-        log_say("Reset the environment")
+    # 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()
+
+    # 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 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}")
+
+        # Main record loop
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
+            dataset=dataset,
             teleop=[leader_arm, keyboard],
-            control_time_s=RESET_TIME_SEC,
+            control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=teleop_action_processor,
@@ -113,21 +97,45 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             robot_observation_processor=robot_observation_processor,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-record episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (
+            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+        ):
+            log_say("Reset the environment")
+            record_loop(
+                robot=robot,
+                events=events,
+                fps=FPS,
+                teleop=[leader_arm, keyboard],
+                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,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    recorded_episodes += 1
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-leader_arm.disconnect()
-keyboard.disconnect()
-listener.stop()
-dataset.push_to_hub()
+        # Save episode
+        dataset.save_episode()
+        recorded_episodes += 1
+
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    leader_arm.disconnect()
+    keyboard.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/replay.py

@@ -20,42 +20,48 @@ 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.robot_utils import precise_sleep
 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)
+def main():
+    # Initialize the robot config
+    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 
-# Fetch the dataset to replay
-dataset = LeRobotDataset("<hf_username>/<dataset_repo_id>", episodes=[EPISODE_IDX])
-# Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
-episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
-actions = episode_frames.select_columns(ACTION)
+    # Initialize the robot
+    robot = LeKiwiClient(robot_config)
 
-# Connect to the robot
-robot.connect()
+    # Fetch the dataset to replay
+    dataset = LeRobotDataset("<hf_username>/<dataset_repo_id>", episodes=[EPISODE_IDX])
+    # Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
+    episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
+    actions = episode_frames.select_columns(ACTION)
 
-if not robot.is_connected:
-    raise ValueError("Robot is not connected!")
+    # Connect to the robot
+    robot.connect()
 
-print("Starting replay loop...")
-log_say(f"Replaying episode {EPISODE_IDX}")
-for idx in range(len(episode_frames)):
-    t0 = time.perf_counter()
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
 
-    # Get recorded action from dataset
-    action = {
-        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-    }
+    print("Starting replay loop...")
+    log_say(f"Replaying episode {EPISODE_IDX}")
+    for idx in range(len(episode_frames)):
+        t0 = time.perf_counter()
 
-    # Send action to robot
-    _ = robot.send_action(action)
+        # Get recorded action from dataset
+        action = {
+            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        }
 
-    busy_wait(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+        # Send action to robot
+        _ = robot.send_action(action)
 
-robot.disconnect()
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+
+    robot.disconnect()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/teleoperate.py

@@ -19,54 +19,60 @@ 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.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 FPS = 30
 
-# Create the robot and teleoperator configurations
-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)
+def main():
+    # Create the robot and teleoperator configurations
+    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")
 
-# 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()
+    # Initialize the robot and teleoperator
+    robot = LeKiwiClient(robot_config)
+    leader_arm = SO100Leader(teleop_arm_config)
+    keyboard = KeyboardTeleop(keyboard_config)
 
-# Init rerun viewer
-init_rerun(session_name="lekiwi_teleop")
+    # 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()
 
-if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-    raise ValueError("Robot or teleop is not connected!")
+    # Init rerun viewer
+    init_rerun(session_name="lekiwi_teleop")
 
-print("Starting teleop loop...")
-while True:
-    t0 = time.perf_counter()
+    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
+        raise ValueError("Robot or teleop is not connected!")
 
-    # Get robot observation
-    observation = robot.get_observation()
+    print("Starting teleop loop...")
+    while True:
+        t0 = time.perf_counter()
 
-    # 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)
+        # Get robot observation
+        observation = robot.get_observation()
 
-    action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
+        # 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)
 
-    # Send action to robot
-    _ = robot.send_action(action)
+        action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
 
-    # Visualize
-    log_rerun_data(observation=observation, action=action)
+        # Send action to robot
+        _ = robot.send_action(action)
 
-    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        # Visualize
+        log_rerun_data(observation=observation, action=action)
+
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/evaluate.py

@@ -52,125 +52,114 @@ 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,
+def main():
+    # 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,
     )
 
-    # 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")
+    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,
@@ -179,19 +168,40 @@ for episode_idx in range(NUM_EPISODES):
             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
+        # 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,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-listener.stop()
-dataset.push_to_hub()
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
+
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/record.py

@@ -50,133 +50,122 @@ 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)
+def main():
+    # 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
 
-# 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()),
-)
+    # Initialize the robot and teleoperator
+    robot = SO100Follower(robot_config)
+    phone = Phone(teleop_config)
 
-# 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,
+    # 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()),
     )
 
-    # 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")
+    # 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,
-            control_time_s=RESET_TIME_SEC,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=phone_to_robot_ee_pose_processor,
@@ -184,20 +173,42 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
             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
+        # 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,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-recording episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-phone.disconnect()
-listener.stop()
-dataset.push_to_hub()
+        # 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()
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/replay.py

@@ -29,72 +29,78 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 )
 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.robot_utils import precise_sleep
 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)
+def main():
+    # Initialize the robot config
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+    )
 
-# 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()),
-)
+    # Initialize the robot
+    robot = SO100Follower(robot_config)
 
-# 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,
-)
+    # 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()),
+    )
 
-# 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)
+    # 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,
+    )
 
-# Connect to the robot
-robot.connect()
+    # 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)
 
-if not robot.is_connected:
-    raise ValueError("Robot is not connected!")
+    # Connect to the robot
+    robot.connect()
 
-print("Starting replay loop...")
-log_say(f"Replaying episode {EPISODE_IDX}")
-for idx in range(len(episode_frames)):
-    t0 = time.perf_counter()
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
 
-    # Get recorded action from dataset
-    ee_action = {
-        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-    }
+    print("Starting replay loop...")
+    log_say(f"Replaying episode {EPISODE_IDX}")
+    for idx in range(len(episode_frames)):
+        t0 = time.perf_counter()
 
-    # Get robot observation
-    robot_obs = robot.get_observation()
+        # Get recorded action from dataset
+        ee_action = {
+            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        }
 
-    # Dataset EE -> robot joints
-    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+        # Get robot observation
+        robot_obs = robot.get_observation()
 
-    # Send action to robot
-    _ = robot.send_action(joint_action)
+        # Dataset EE -> robot joints
+        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+        # Send action to robot
+        _ = robot.send_action(joint_action)
 
-# Clean up
-robot.disconnect()
+        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+    # Clean up
+    robot.disconnect()
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/teleoperate.py

@@ -32,82 +32,90 @@ 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.robot_utils import precise_sleep
 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)
+def main():
+    # 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
 
-# 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()),
-)
+    # Initialize the robot and teleoperator
+    robot = SO100Follower(robot_config)
+    teleop_device = Phone(teleop_config)
 
-# 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,
-)
+    # 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()),
+    )
 
-# Connect to the robot and teleoperator
-robot.connect()
-teleop_device.connect()
+    # 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,
+    )
 
-# Init rerun viewer
-init_rerun(session_name="phone_so100_teleop")
+    # Connect to the robot and teleoperator
+    robot.connect()
+    teleop_device.connect()
 
-if not robot.is_connected or not teleop_device.is_connected:
-    raise ValueError("Robot or teleop is not connected!")
+    # Init rerun viewer
+    init_rerun(session_name="phone_so100_teleop")
 
-print("Starting teleop loop. Move your phone to teleoperate the robot...")
-while True:
-    t0 = time.perf_counter()
+    if not robot.is_connected or not teleop_device.is_connected:
+        raise ValueError("Robot or teleop is not connected!")
 
-    # Get robot observation
-    robot_obs = robot.get_observation()
+    print("Starting teleop loop. Move your phone to teleoperate the robot...")
+    while True:
+        t0 = time.perf_counter()
 
-    # Get teleop action
-    phone_obs = teleop_device.get_action()
+        # Get robot observation
+        robot_obs = robot.get_observation()
 
-    # Phone -> EE pose -> Joints transition
-    joint_action = phone_to_robot_joints_processor((phone_obs, robot_obs))
+        # Get teleop action
+        phone_obs = teleop_device.get_action()
 
-    # Send action to robot
-    _ = robot.send_action(joint_action)
+        # Phone -> EE pose -> Joints transition
+        joint_action = phone_to_robot_joints_processor((phone_obs, robot_obs))
 
-    # Visualize
-    log_rerun_data(observation=phone_obs, action=joint_action)
+        # Send action to robot
+        _ = robot.send_action(joint_action)
 
-    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        # Visualize
+        log_rerun_data(observation=phone_obs, action=joint_action)
+
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+
+
+if __name__ == "__main__":
+    main()

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/port_datasets/slurm_aggregate_shards.py

@@ -15,16 +15,12 @@
 # limitations under the License.
 
 import argparse
-import logging
 from pathlib import Path
 
 from datatrove.executor import LocalPipelineExecutor
 from datatrove.executor.slurm import SlurmPipelineExecutor
 from datatrove.pipeline.base import PipelineStep
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
-
-from lerobot.datasets.aggregate import aggregate_datasets
-from lerobot.utils.utils import init_logging
+from port_droid import DROID_SHARDS
 
 
 class AggregateDatasets(PipelineStep):
@@ -38,6 +34,11 @@ class AggregateDatasets(PipelineStep):
         self.aggr_repo_id = aggregated_repo_id
 
     def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+
+        from lerobot.datasets.aggregate import aggregate_datasets
+        from lerobot.utils.utils import init_logging
+
         init_logging()
 
         # Since aggregate_datasets already handles parallel processing internally,

examples/port_datasets/slurm_port_shards.py

@@ -20,7 +20,7 @@ from pathlib import Path
 from datatrove.executor import LocalPipelineExecutor
 from datatrove.executor.slurm import SlurmPipelineExecutor
 from datatrove.pipeline.base import PipelineStep
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
+from port_droid import DROID_SHARDS
 
 
 class PortDroidShards(PipelineStep):
@@ -35,7 +35,7 @@ class PortDroidShards(PipelineStep):
 
     def run(self, data=None, rank: int = 0, world_size: int = 1):
         from datasets.utils.tqdm import disable_progress_bars
-        from port_datasets.droid_rlds.port_droid import port_droid, validate_dataset
+        from port_droid import port_droid, validate_dataset
 
         from lerobot.utils.utils import init_logging
 

examples/port_datasets/slurm_upload.py

@@ -24,7 +24,7 @@ from datatrove.executor.slurm import SlurmPipelineExecutor
 from datatrove.pipeline.base import PipelineStep
 from huggingface_hub import HfApi
 from huggingface_hub.constants import REPOCARD_NAME
-from port_datasets.droid_rlds.port_droid import DROID_SHARDS
+from port_droid import DROID_SHARDS
 
 from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.utils import create_lerobot_dataset_card
@@ -185,11 +185,11 @@ class UploadDataset(PipelineStep):
 
 
 def make_upload_executor(
-    repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
+    repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, private=False, slurm=True
 ):
     kwargs = {
         "pipeline": [
-            UploadDataset(repo_id),
+            UploadDataset(repo_id, private=private),
         ],
         "logging_dir": str(logs_dir / job_name),
     }
@@ -267,6 +267,12 @@ def main():
         default="1950M",
         help="Memory per cpu that each worker will use.",
     )
+    parser.add_argument(
+        "--private",
+        action="store_true",
+        default=False,
+        help="Whether to create a private repository.",
+    )
 
     init_logging()
 

examples/rtc/eval_dataset.py

@@ -0,0 +1,951 @@
+#!/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.
+
+"""
+Evaluate Real-Time Chunking (RTC) performance on dataset samples.
+
+This script takes two random samples from a dataset:
+- Uses actions from the first sample as previous chunk
+- Generates new actions for the second sample with and without RTC
+
+It compares action predictions with and without RTC on dataset samples,
+measuring consistency and ground truth alignment.
+
+Usage:
+    # Basic usage with smolvla policy
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=mps \
+        --rtc.max_guidance_weight=10.0 \
+        --rtc.prefix_attention_schedule=EXP \
+        --seed=10
+
+    # Basic usage with pi0.5 policy
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi05_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=10 \
+        --device=mps
+        --seed=10
+
+    # Basic usage with pi0.5 policy with cuda device
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi05_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    # Basic usage with pi0 policy with cuda device
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi0_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lipsop/reuben_pi0 \
+        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    # With torch.compile for faster inference (PyTorch 2.0+)
+    # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=mps \
+        --use_torch_compile=true \
+        --torch_compile_mode=max-autotune
+
+    # With torch.compile on CUDA (CUDA graphs disabled by default)
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=cuda \
+        --use_torch_compile=true \
+        --torch_compile_mode=reduce-overhead
+
+    # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --use_torch_compile=true \
+        --torch_compile_backend=inductor \
+        --torch_compile_mode=max-autotune \
+        --torch_compile_disable_cudagraphs=false
+"""
+
+import gc
+import logging
+import os
+import random
+from dataclasses import dataclass, field
+
+import numpy as np
+import torch
+
+try:
+    import matplotlib.pyplot as plt
+
+    MATPLOTLIB_AVAILABLE = True
+except ImportError:
+    MATPLOTLIB_AVAILABLE = False
+    plt = None
+
+from lerobot.configs import parser
+from lerobot.configs.default import DatasetConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.factory import resolve_delta_timestamps
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging
+
+
+def set_seed(seed: int):
+    """Set random seed for reproducibility."""
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+        torch.cuda.manual_seed_all(seed)
+    if torch.backends.mps.is_available():
+        torch.mps.manual_seed(seed)
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+
+
+def _check_matplotlib_available():
+    """Check if matplotlib is available, raise helpful error if not."""
+    if not MATPLOTLIB_AVAILABLE:
+        raise ImportError(
+            "matplotlib is required for RTC debug visualizations. "
+            "Please install it by running:\n"
+            "  uv pip install matplotlib"
+        )
+
+
+@dataclass
+class RTCEvalConfig(HubMixin):
+    """Configuration for RTC evaluation."""
+
+    # Policy configuration
+    policy: PreTrainedConfig | None = None
+
+    # Dataset configuration
+    dataset: DatasetConfig = field(default_factory=DatasetConfig)
+
+    # RTC configuration
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            enabled=True,
+            execution_horizon=20,
+            max_guidance_weight=10.0,
+            prefix_attention_schedule=RTCAttentionSchedule.EXP,
+            debug=True,
+            debug_maxlen=1000,
+        )
+    )
+
+    # Device configuration
+    device: str | None = field(
+        default=None,
+        metadata={"help": "Device to run on (cuda, cpu, mps, auto)"},
+    )
+
+    # Output configuration
+    output_dir: str = field(
+        default="rtc_debug_output",
+        metadata={"help": "Directory to save debug visualizations"},
+    )
+
+    # Seed configuration
+    seed: int = field(
+        default=42,
+        metadata={"help": "Random seed for reproducibility"},
+    )
+
+    inference_delay: int = field(
+        default=4,
+        metadata={"help": "Inference delay for RTC"},
+    )
+
+    # Torch compile configuration
+    use_torch_compile: bool = field(
+        default=False,
+        metadata={"help": "Use torch.compile for faster inference (PyTorch 2.0+)"},
+    )
+
+    torch_compile_backend: str = field(
+        default="inductor",
+        metadata={"help": "Backend for torch.compile (inductor, aot_eager, cudagraphs)"},
+    )
+
+    torch_compile_mode: str = field(
+        default="default",
+        metadata={"help": "Compilation mode (default, reduce-overhead, max-autotune)"},
+    )
+
+    torch_compile_disable_cudagraphs: bool = field(
+        default=True,
+        metadata={
+            "help": "Disable CUDA graphs in torch.compile. Required due to in-place tensor "
+            "operations in denoising loop (x_t += dt * v_t) which cause tensor aliasing issues."
+        },
+    )
+
+    def __post_init__(self):
+        # Parse policy path
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        else:
+            raise ValueError("Policy path is required (--policy.path)")
+
+        # Auto-detect device if not specified
+        if self.device is None or self.device == "auto":
+            if torch.cuda.is_available():
+                self.device = "cuda"
+            elif torch.backends.mps.is_available():
+                self.device = "mps"
+            else:
+                self.device = "cpu"
+            logging.info(f"Auto-detected device: {self.device}")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
+        return ["policy"]
+
+
+class RTCEvaluator:
+    """Evaluator for RTC on dataset samples."""
+
+    def __init__(self, cfg: RTCEvalConfig):
+        self.cfg = cfg
+        self.device = cfg.device
+
+        # Load dataset with proper delta_timestamps based on policy configuration
+        # Calculate delta_timestamps using the same logic as make_dataset factory
+        logging.info(f"Loading dataset: {cfg.dataset.repo_id}")
+
+        # Get dataset metadata to extract FPS
+        ds_meta = LeRobotDatasetMetadata(cfg.dataset.repo_id)
+
+        # Calculate delta_timestamps from policy's delta_indices
+        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
+
+        # Create dataset with calculated delta_timestamps
+        self.dataset = LeRobotDataset(
+            cfg.dataset.repo_id,
+            delta_timestamps=delta_timestamps,
+        )
+        logging.info(f"Dataset loaded: {len(self.dataset)} samples, {self.dataset.num_episodes} episodes")
+
+        # Create preprocessor/postprocessor
+        self.preprocessor, self.postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            preprocessor_overrides={
+                "device_processor": {"device": self.device},
+            },
+        )
+
+        logging.info("=" * 80)
+        logging.info("Ready to run evaluation with sequential policy loading:")
+        logging.info("  1. policy_prev_chunk - Generate reference chunk, then destroy")
+        logging.info("  2. policy_no_rtc - Generate without RTC, then destroy")
+        logging.info("  3. policy_rtc - Generate with RTC, then destroy")
+        logging.info("  Note: Only one policy in memory at a time for efficient memory usage")
+        logging.info("=" * 80)
+
+    def _init_policy(self, name: str, rtc_enabled: bool, rtc_debug: bool):
+        """Initialize a single policy instance with specified RTC configuration.
+
+        Args:
+            name: Name identifier for logging purposes
+            rtc_enabled: Whether to enable RTC for this policy
+            rtc_debug: Whether to enable debug tracking for this policy
+
+        Returns:
+            Configured policy instance with optional torch.compile applied
+        """
+        logging.info(f"Initializing {name}...")
+
+        # Load policy from pretrained
+        policy_class = get_policy_class(self.cfg.policy.type)
+
+        config = PreTrainedConfig.from_pretrained(self.cfg.policy.pretrained_path)
+
+        if self.cfg.policy.type == "pi05" or self.cfg.policy.type == "pi0":
+            config.compile_model = self.cfg.use_torch_compile
+
+        policy = policy_class.from_pretrained(self.cfg.policy.pretrained_path, config=config)
+        policy = policy.to(self.device)
+        policy.eval()
+
+        # Configure RTC
+        rtc_config = RTCConfig(
+            enabled=rtc_enabled,
+            execution_horizon=self.cfg.rtc.execution_horizon,
+            max_guidance_weight=self.cfg.rtc.max_guidance_weight,
+            prefix_attention_schedule=self.cfg.rtc.prefix_attention_schedule,
+            debug=rtc_debug,
+            debug_maxlen=self.cfg.rtc.debug_maxlen,
+        )
+        policy.config.rtc_config = rtc_config
+        policy.init_rtc_processor()
+
+        logging.info(f"  RTC enabled: {rtc_enabled}")
+        logging.info(f"  RTC debug: {rtc_debug}")
+        logging.info(f"  Policy config: {config}")
+
+        # Apply torch.compile to predict_action_chunk method if enabled
+        if self.cfg.use_torch_compile:
+            policy = self._apply_torch_compile(policy, name)
+
+        logging.info(f"✓ {name} initialized successfully")
+        return policy
+
+    def _apply_torch_compile(self, policy, policy_name: str):
+        """Apply torch.compile to the policy's predict_action_chunk method.
+
+        Args:
+            policy: Policy instance to compile
+            policy_name: Name for logging purposes
+
+        Returns:
+            Policy with compiled predict_action_chunk method
+        """
+
+        # PI models handle their own compilation
+        if policy.type == "pi05" or policy.type == "pi0":
+            return policy
+
+        try:
+            # Check if torch.compile is available (PyTorch 2.0+)
+            if not hasattr(torch, "compile"):
+                logging.warning(
+                    f"  [{policy_name}] torch.compile is not available. Requires PyTorch 2.0+. "
+                    f"Current version: {torch.__version__}. Skipping compilation."
+                )
+                return policy
+
+            logging.info(f"  [{policy_name}] Applying torch.compile to predict_action_chunk...")
+            logging.info(f"    Backend: {self.cfg.torch_compile_backend}")
+            logging.info(f"    Mode: {self.cfg.torch_compile_mode}")
+            logging.info(f"    Disable CUDA graphs: {self.cfg.torch_compile_disable_cudagraphs}")
+            logging.info("    Note: Debug tracker excluded from compilation via @torch._dynamo.disable")
+
+            # Compile the predict_action_chunk method
+            # - Debug tracker is excluded from compilation via @torch._dynamo.disable
+            # - CUDA graphs disabled to prevent tensor aliasing from in-place ops (x_t += dt * v_t)
+            compile_kwargs = {
+                "backend": self.cfg.torch_compile_backend,
+                "mode": self.cfg.torch_compile_mode,
+            }
+
+            # Disable CUDA graphs if requested (prevents tensor aliasing issues)
+            if self.cfg.torch_compile_disable_cudagraphs:
+                compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+            original_method = policy.predict_action_chunk
+            compiled_method = torch.compile(original_method, **compile_kwargs)
+            policy.predict_action_chunk = compiled_method
+            logging.info(f"  ✓ [{policy_name}] Successfully compiled predict_action_chunk")
+
+        except Exception as e:
+            logging.error(f"  [{policy_name}] Failed to apply torch.compile: {e}")
+            logging.warning(f"  [{policy_name}] Continuing without torch.compile")
+
+        return policy
+
+    def _destroy_policy(self, policy, policy_name: str):
+        """Explicitly destroy a policy and free all associated memory.
+
+        This method performs aggressive cleanup to ensure maximum memory is freed,
+        which is critical for large models (e.g., VLAs with billions of parameters).
+
+        Args:
+            policy: Policy instance to destroy
+            policy_name: Name for logging purposes
+        """
+        logging.info(f"  Destroying {policy_name} and freeing memory...")
+
+        try:
+            # Step 1: Move policy to CPU to free GPU/MPS memory
+            policy.cpu()
+
+            # Step 2: Delete the policy object
+            del policy
+
+            # Step 3: Force garbage collection to reclaim memory immediately
+            gc.collect()
+
+            # Step 4: Clear device-specific caches
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+                torch.cuda.synchronize()  # Ensure all operations complete
+
+            if torch.backends.mps.is_available():
+                torch.mps.empty_cache()
+
+            logging.info(f"  ✓ {policy_name} destroyed and memory freed")
+
+        except Exception as e:
+            logging.warning(f"  Warning: Error during {policy_name} cleanup: {e}")
+
+    def run_evaluation(self):
+        """Run evaluation on two random dataset samples using three separate policies.
+
+        Note: Policies are deinitalized after each step to free memory. Large models
+        (e.g., VLA models with billions of parameters) cannot fit three instances in
+        memory simultaneously. By deleting and garbage collecting after each step,
+        we ensure only one policy is loaded at a time.
+        """
+        # Create output directory
+        os.makedirs(self.cfg.output_dir, exist_ok=True)
+        logging.info(f"Output directory: {self.cfg.output_dir}")
+
+        logging.info("=" * 80)
+        logging.info("Starting RTC evaluation")
+        logging.info(f"Inference delay: {self.cfg.inference_delay}")
+        logging.info("=" * 80)
+
+        # Load two random samples from dataset
+        data_loader = torch.utils.data.DataLoader(self.dataset, batch_size=1, shuffle=True)
+        loader_iter = iter(data_loader)
+        first_sample = next(loader_iter)
+        second_sample = next(loader_iter)
+
+        preprocessed_first_sample = self.preprocessor(first_sample)
+        preprocessed_second_sample = self.preprocessor(second_sample)
+
+        # ============================================================================
+        # Step 1: Generate previous chunk using policy_prev_chunk
+        # ============================================================================
+        # This policy is only used to generate the reference chunk and then freed
+        logging.info("=" * 80)
+        logging.info("Step 1: Generating previous chunk with policy_prev_chunk")
+        logging.info("=" * 80)
+
+        # Initialize policy 1
+        policy_prev_chunk_policy = self._init_policy(
+            name="policy_prev_chunk",
+            rtc_enabled=False,
+            rtc_debug=False,
+        )
+        with torch.no_grad():
+            prev_chunk_left_over = policy_prev_chunk_policy.predict_action_chunk(
+                preprocessed_first_sample,
+            )[:, :25, :].squeeze(0)
+        logging.info(f"  Generated prev_chunk shape: {prev_chunk_left_over.shape}")
+
+        # Destroy policy_prev_chunk to free memory for large models
+        self._destroy_policy(policy_prev_chunk_policy, "policy_prev_chunk")
+
+        # ============================================================================
+        # Step 2: Generate actions WITHOUT RTC using policy_no_rtc
+        # ============================================================================
+        logging.info("=" * 80)
+        logging.info("Step 2: Generating actions WITHOUT RTC with policy_no_rtc")
+        logging.info("=" * 80)
+
+        set_seed(self.cfg.seed)
+
+        # Initialize policy 2
+        policy_no_rtc_policy = self._init_policy(
+            name="policy_no_rtc",
+            rtc_enabled=False,
+            rtc_debug=True,
+        )
+
+        # Sample noise (use same noise for both RTC and non-RTC for fair comparison)
+        noise_size = (1, policy_no_rtc_policy.config.chunk_size, policy_no_rtc_policy.config.max_action_dim)
+        noise = policy_no_rtc_policy.model.sample_noise(noise_size, self.device)
+        noise_clone = noise.clone()
+        policy_no_rtc_policy.rtc_processor.reset_tracker()
+        with torch.no_grad():
+            no_rtc_actions = policy_no_rtc_policy.predict_action_chunk(
+                preprocessed_second_sample,
+                noise=noise,
+            )
+        no_rtc_tracked_steps = policy_no_rtc_policy.rtc_processor.tracker.get_all_steps()
+        logging.info(f"  Tracked {len(no_rtc_tracked_steps)} steps without RTC")
+        logging.info(f"  Generated no_rtc_actions shape: {no_rtc_actions.shape}")
+
+        # Destroy policy_no_rtc to free memory before loading policy_rtc
+        self._destroy_policy(policy_no_rtc_policy, "policy_no_rtc")
+
+        # ============================================================================
+        # Step 3: Generate actions WITH RTC using policy_rtc
+        # ============================================================================
+        logging.info("=" * 80)
+        logging.info("Step 3: Generating actions WITH RTC with policy_rtc")
+        logging.info("=" * 80)
+
+        set_seed(self.cfg.seed)
+
+        # Initialize policy 3
+        policy_rtc_policy = self._init_policy(
+            name="policy_rtc",
+            rtc_enabled=True,
+            rtc_debug=True,
+        )
+        policy_rtc_policy.rtc_processor.reset_tracker()
+        with torch.no_grad():
+            rtc_actions = policy_rtc_policy.predict_action_chunk(
+                preprocessed_second_sample,
+                noise=noise_clone,
+                inference_delay=self.cfg.inference_delay,
+                prev_chunk_left_over=prev_chunk_left_over,
+                execution_horizon=self.cfg.rtc.execution_horizon,
+            )
+        rtc_tracked_steps = policy_rtc_policy.rtc_processor.get_all_debug_steps()
+        logging.info(f"  Tracked {len(rtc_tracked_steps)} steps with RTC")
+        logging.info(f"  Generated rtc_actions shape: {rtc_actions.shape}")
+
+        # Save num_steps before destroying policy (needed for plotting)
+        try:
+            num_steps = policy_rtc_policy.config.num_steps
+        except Exception as e:
+            logging.error(f"  Error getting num_steps: {e}")
+            num_steps = policy_rtc_policy.config.num_inference_steps
+            logging.warning(f"  Using num_inference_steps: {num_steps} instead of num_steps")
+
+        # Destroy policy_rtc after final use
+        self._destroy_policy(policy_rtc_policy, "policy_rtc")
+
+        # Plot and save results
+        logging.info("=" * 80)
+        logging.info("Plotting results...")
+        self.plot_tracked_data(rtc_tracked_steps, no_rtc_tracked_steps, prev_chunk_left_over, num_steps)
+
+        # Plot final actions comparison
+        logging.info("=" * 80)
+        logging.info("Plotting final actions comparison...")
+        self.plot_final_actions_comparison(rtc_actions, no_rtc_actions, prev_chunk_left_over)
+
+        logging.info("=" * 80)
+        logging.info("Evaluation completed successfully")
+
+    def plot_final_actions_comparison(self, rtc_actions, no_rtc_actions, prev_chunk_left_over):
+        """Plot final action predictions comparison on a single chart.
+
+        Args:
+            rtc_actions: Final actions from RTC policy
+            no_rtc_actions: Final actions from non-RTC policy
+            prev_chunk_left_over: Previous chunk used as ground truth
+        """
+        _check_matplotlib_available()
+
+        # Remove batch dimension if present
+        rtc_actions_plot = rtc_actions.squeeze(0).cpu() if len(rtc_actions.shape) == 3 else rtc_actions.cpu()
+        no_rtc_actions_plot = (
+            no_rtc_actions.squeeze(0).cpu() if len(no_rtc_actions.shape) == 3 else no_rtc_actions.cpu()
+        )
+        prev_chunk_plot = prev_chunk_left_over.cpu()
+
+        # Create figure with 6 subplots (one per action dimension)
+        fig, axes = plt.subplots(6, 1, figsize=(16, 12))
+        fig.suptitle("Final Action Predictions Comparison (Raw)", fontsize=16)
+
+        # Plot each action dimension
+        for dim_idx, ax in enumerate(axes):
+            # Plot previous chunk (ground truth) in red
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                prev_chunk_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="red",
+                label="Previous Chunk (Ground Truth)",
+                linewidth=2.5,
+                alpha=0.8,
+            )
+
+            # Plot no-RTC actions in blue
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                no_rtc_actions_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="blue",
+                label="No RTC",
+                linewidth=2,
+                alpha=0.7,
+            )
+
+            # Plot RTC actions in green
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                rtc_actions_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="green",
+                label="RTC",
+                linewidth=2,
+                alpha=0.7,
+            )
+
+            # Add vertical lines for inference delay and execution horizon
+            inference_delay = self.cfg.inference_delay
+            execution_horizon = self.cfg.rtc.execution_horizon
+
+            if inference_delay > 0:
+                ax.axvline(
+                    x=inference_delay - 1,
+                    color="orange",
+                    linestyle="--",
+                    alpha=0.5,
+                    label=f"Inference Delay ({inference_delay})",
+                )
+
+            if execution_horizon > 0:
+                ax.axvline(
+                    x=execution_horizon,
+                    color="purple",
+                    linestyle="--",
+                    alpha=0.5,
+                    label=f"Execution Horizon ({execution_horizon})",
+                )
+
+            ax.set_ylabel(f"Dim {dim_idx}", fontsize=10)
+            ax.grid(True, alpha=0.3)
+
+            # Set x-axis ticks to show all integer values
+            max_len = max(rtc_actions_plot.shape[0], no_rtc_actions_plot.shape[0], prev_chunk_plot.shape[0])
+            ax.set_xticks(range(0, max_len, max(1, max_len // 20)))  # Show ~20 ticks
+            ax.set_xlim(-0.5, max_len - 0.5)
+
+        axes[-1].set_xlabel("Step", fontsize=10)
+
+        # Collect legend handles and labels from first subplot
+        handles, labels = axes[0].get_legend_handles_labels()
+        # Remove duplicates while preserving order
+        seen = set()
+        unique_handles = []
+        unique_labels = []
+        for handle, label in zip(handles, labels, strict=True):
+            if label not in seen:
+                seen.add(label)
+                unique_handles.append(handle)
+                unique_labels.append(label)
+
+        # Add legend outside the plot area (to the right)
+        fig.legend(
+            unique_handles,
+            unique_labels,
+            loc="center right",
+            fontsize=9,
+            bbox_to_anchor=(1.0, 0.5),
+            framealpha=0.9,
+        )
+
+        # Save figure
+        output_path = os.path.join(self.cfg.output_dir, "final_actions_comparison.png")
+        fig.tight_layout(rect=[0, 0, 0.85, 1])  # Leave space for legend on right
+        fig.savefig(output_path, dpi=150, bbox_inches="tight")
+        logging.info(f"Saved final actions comparison to {output_path}")
+        plt.close(fig)
+
+    def plot_tracked_data(self, rtc_tracked_steps, no_rtc_tracked_steps, prev_chunk_left_over, num_steps):
+        _check_matplotlib_available()
+
+        # Create side-by-side figures for denoising visualization
+        fig_xt, axs_xt = self._create_figure("x_t Denoising: No RTC (left) vs RTC (right)")
+        fig_vt, axs_vt = self._create_figure("v_t Denoising: No RTC (left) vs RTC (right)")
+        fig_corr, axs_corr = self._create_figure("Correction: No RTC (left) vs RTC (right)")
+        fig_x1t, axs_x1t = self._create_figure(
+            "x1_t Predicted State & Error: No RTC (left - empty) vs RTC (right)"
+        )
+        self._plot_denoising_steps_from_tracker(
+            rtc_tracked_steps,
+            axs_xt[:, 1],  # Right column for x_t
+            axs_vt[:, 1],  # Right column for v_t
+            axs_corr[:, 1],  # Right column for correction
+            axs_x1t[:, 1],  # Right column for x1_t
+            num_steps,
+            add_labels=True,  # Add labels for RTC (right column)
+        )
+
+        self._plot_denoising_steps_from_tracker(
+            no_rtc_tracked_steps,
+            axs_xt[:, 0],  # Left column for x_t
+            axs_vt[:, 0],  # Left column for v_t
+            axs_corr[:, 0],  # Left column for correction
+            axs_x1t[:, 0],  # Left column for x1_t
+            num_steps,
+            add_labels=False,  # No labels for No RTC (left column)
+        )
+
+        # Plot no-RTC x_t data on right chart as orange dashed line for comparison
+        self._plot_no_rtc_xt_reference(no_rtc_tracked_steps, axs_xt[:, 1], num_steps)
+
+        # Plot ground truth on x_t axes
+        RTCDebugVisualizer.plot_waypoints(
+            axs_xt[:, 1], prev_chunk_left_over, start_from=0, color="red", label="Ground truth"
+        )
+
+        # Plot ground truth on x1_t axes
+        RTCDebugVisualizer.plot_waypoints(
+            axs_x1t[:, 1], prev_chunk_left_over, start_from=0, color="red", label="Ground truth"
+        )
+
+        # Plot ground truth on x_t axes (no labels for left column)
+        RTCDebugVisualizer.plot_waypoints(
+            axs_xt[:, 0], prev_chunk_left_over, start_from=0, color="red", label=None
+        )
+
+        RTCDebugVisualizer.plot_waypoints(
+            axs_x1t[:, 0], prev_chunk_left_over, start_from=0, color="red", label=None
+        )
+
+        # Add legends outside the plot area for each figure
+        self._add_figure_legend(fig_xt, axs_xt)
+        self._add_figure_legend(fig_vt, axs_vt)
+        self._add_figure_legend(fig_corr, axs_corr)
+        self._add_figure_legend(fig_x1t, axs_x1t)
+
+        # Save denoising plots
+        self._save_figure(fig_xt, os.path.join(self.cfg.output_dir, "denoising_xt_comparison.png"))
+        self._save_figure(fig_vt, os.path.join(self.cfg.output_dir, "denoising_vt_comparison.png"))
+        self._save_figure(fig_corr, os.path.join(self.cfg.output_dir, "denoising_correction_comparison.png"))
+        self._save_figure(fig_x1t, os.path.join(self.cfg.output_dir, "denoising_x1t_comparison.png"))
+
+    def _create_figure(self, title):
+        fig, axs = plt.subplots(6, 2, figsize=(24, 12))
+        fig.suptitle(title, fontsize=16)
+
+        for ax in axs[:, 0]:
+            ax.set_title("No RTC (N/A)" if ax == axs[0, 0] else "", fontsize=12)
+        for ax in axs[:, 1]:
+            ax.set_title("RTC" if ax == axs[0, 1] else "", fontsize=12)
+
+        return fig, axs
+
+    def _add_figure_legend(self, fig, axs):
+        """Add a legend outside the plot area on the right side.
+
+        Args:
+            fig: Matplotlib figure to add legend to
+            axs: Array of axes to collect legend handles from
+        """
+        # Collect all handles and labels from the first row of axes (right column)
+        handles, labels = axs[0, 1].get_legend_handles_labels()
+
+        # Remove duplicates while preserving order
+        seen = set()
+        unique_handles = []
+        unique_labels = []
+        for handle, label in zip(handles, labels, strict=True):
+            if label not in seen:
+                seen.add(label)
+                unique_handles.append(handle)
+                unique_labels.append(label)
+
+        # Add legend outside the plot area (to the right, close to charts)
+        if unique_handles:
+            fig.legend(
+                unique_handles,
+                unique_labels,
+                loc="center left",
+                fontsize=8,
+                bbox_to_anchor=(0.87, 0.5),
+                framealpha=0.9,
+                ncol=1,
+            )
+
+    def _save_figure(self, fig, path):
+        fig.tight_layout(rect=[0, 0, 0.85, 1])  # Leave space for legend/colorbar on right
+        fig.savefig(path, dpi=150, bbox_inches="tight")
+        logging.info(f"Saved figure to {path}")
+        plt.close(fig)
+
+    def _plot_denoising_steps_from_tracker(
+        self, tracked_steps, xt_axs, vt_axs, corr_axs, x1t_axs, num_steps, add_labels=True
+    ):
+        """Plot denoising steps from tracker data.
+
+        Args:
+            tracked_steps: List of DebugStep objects containing debug steps
+            xt_axs: Matplotlib axes for x_t plots (array of 6 axes)
+            vt_axs: Matplotlib axes for v_t plots (array of 6 axes)
+            corr_axs: Matplotlib axes for correction plots (array of 6 axes)
+            x1t_axs: Matplotlib axes for x1_t plots (array of 6 axes)
+            num_steps: Total number of denoising steps for colormap
+            add_labels: Whether to add legend labels for the plots
+        """
+
+        logging.info("=" * 80)
+        logging.info(f"Plotting {len(tracked_steps)} steps")
+
+        debug_steps = tracked_steps
+        if not debug_steps:
+            return
+
+        # Define colors for different denoise steps (using a colormap)
+        colors = plt.cm.viridis(np.linspace(0, 1, num_steps))
+
+        for step_idx, debug_step in enumerate(debug_steps):
+            color = colors[step_idx % len(colors)]
+            label = f"Step {step_idx}" if add_labels else None
+
+            # Plot x_t
+            if debug_step.x_t is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    xt_axs, debug_step.x_t, start_from=0, color=color, label=label
+                )
+
+            # Plot v_t
+            if debug_step.v_t is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    vt_axs, debug_step.v_t, start_from=0, color=color, label=label
+                )
+
+            # Plot correction on separate axes
+            if debug_step.correction is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    corr_axs,
+                    debug_step.correction,
+                    start_from=0,
+                    color=color,
+                    label=label,
+                )
+
+            # Plot x1_t (predicted state)
+            if x1t_axs is not None and debug_step.x1_t is not None:
+                x1t_label = f"x1_t Step {step_idx}" if add_labels else None
+                RTCDebugVisualizer.plot_waypoints(
+                    x1t_axs,
+                    debug_step.x1_t,
+                    start_from=0,
+                    color=color,
+                    label=x1t_label,
+                )
+
+            # Plot error in orange dashed
+            if x1t_axs is not None and debug_step.err is not None:
+                error_chunk = (
+                    debug_step.err[0].cpu().numpy()
+                    if len(debug_step.err.shape) == 3
+                    else debug_step.err.cpu().numpy()
+                )
+
+                num_dims = min(error_chunk.shape[-1], 6)
+                error_label = f"error Step {step_idx}" if add_labels else None
+                for j in range(num_dims):
+                    x1t_axs[j].plot(
+                        np.arange(0, error_chunk.shape[0]),
+                        error_chunk[:, j],
+                        color="orange",
+                        linestyle="--",
+                        alpha=0.7,
+                        label=error_label,
+                    )
+
+        # Recalculate axis limits after plotting to ensure proper scaling
+        self._rescale_axes(xt_axs)
+        self._rescale_axes(vt_axs)
+        self._rescale_axes(corr_axs)
+        self._rescale_axes(x1t_axs)
+
+    def _plot_no_rtc_xt_reference(self, no_rtc_tracked_steps, xt_axs, num_steps):
+        """Plot final no-RTC x_t data as orange dashed line on the RTC chart for comparison.
+
+        Args:
+            no_rtc_tracked_steps: List of DebugStep objects containing no-RTC debug steps
+            xt_axs: Matplotlib axes for x_t plots (array of 6 axes, right column)
+            num_steps: Total number of denoising steps for colormap
+        """
+        debug_steps = no_rtc_tracked_steps
+        if not debug_steps:
+            return
+
+        # Plot only the final x_t step as orange dashed line
+        final_step = debug_steps[-1]
+        logging.info("Plotting final no-RTC x_t step as orange dashed reference")
+
+        if final_step.x_t is not None:
+            x_t_chunk = (
+                final_step.x_t[0].cpu().numpy()
+                if len(final_step.x_t.shape) == 3
+                else final_step.x_t.cpu().numpy()
+            )
+
+            num_dims = min(x_t_chunk.shape[-1], 6)
+            for j in range(num_dims):
+                xt_axs[j].plot(
+                    np.arange(0, x_t_chunk.shape[0]),
+                    x_t_chunk[:, j],
+                    color="orange",
+                    linestyle="--",
+                    alpha=0.7,
+                    linewidth=2,
+                    label="No RTC (final)" if j == 0 else "",
+                )
+
+    def _rescale_axes(self, axes):
+        """Rescale axes to show all data with proper margins.
+
+        Args:
+            axes: Array of matplotlib axes to rescale
+        """
+        for ax in axes:
+            ax.relim()
+            ax.autoscale_view()
+
+            # Add 10% margin to y-axis for better visualization
+            ylim = ax.get_ylim()
+            y_range = ylim[1] - ylim[0]
+            if y_range > 0:  # Avoid division by zero
+                margin = y_range * 0.1
+                ax.set_ylim(ylim[0] - margin, ylim[1] + margin)
+
+            # Set x-axis ticks to show all integer values
+            xlim = ax.get_xlim()
+            max_len = int(xlim[1]) + 1
+            if max_len > 0:
+                ax.set_xticks(range(0, max_len, max(1, max_len // 20)))  # Show ~20 ticks
+                ax.set_xlim(-0.5, max_len - 0.5)
+
+
+@parser.wrap()
+def main(cfg: RTCEvalConfig):
+    """Main entry point for RTC evaluation."""
+    # Set random seed for reproducibility
+    set_seed(cfg.seed)
+
+    init_logging()
+
+    logging.info("=" * 80)
+    logging.info("RTC Dataset Evaluation")
+    logging.info(f"Config: {cfg}")
+    logging.info("=" * 80)
+
+    evaluator = RTCEvaluator(cfg)
+    evaluator.run_evaluation()
+
+
+if __name__ == "__main__":
+    main()

examples/rtc/eval_with_real_robot.py

@@ -0,0 +1,549 @@
+#!/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.
+
+"""
+Demo script showing how to use Real-Time Chunking (RTC) with action chunking policies on real robots.
+
+This script demonstrates:
+1. Creating a robot and policy (SmolVLA, Pi0, etc.) with RTC
+2. Consuming actions from the policy while the robot executes
+3. Periodically requesting new action chunks in the background using threads
+4. Managing action buffers and timing for real-time operation
+
+For simulation environments, see eval_with_simulation.py
+
+Usage:
+    # Run RTC with Real robot with RTC
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.device=mps \
+        --rtc.enabled=true \
+        --rtc.execution_horizon=20 \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+
+    # Run RTC with Real robot without RTC
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.device=mps \
+        --rtc.enabled=false \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+
+    # Run RTC with Real robot with pi0.5 policy
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/pi05_check_rtc \
+        --policy.device=mps \
+        --rtc.enabled=true \
+        --rtc.execution_horizon=20 \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+"""
+
+import logging
+import math
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+from threading import Event, Lock, Thread
+
+import torch
+from torch import Tensor
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.processor.factory import (
+    make_default_robot_action_processor,
+    make_default_robot_observation_processor,
+)
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.robots import (  # noqa: F401
+    Robot,
+    RobotConfig,
+    koch_follower,
+    so100_follower,
+    so101_follower,
+)
+from lerobot.robots.utils import make_robot_from_config
+from lerobot.utils.constants import OBS_IMAGES
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+class RobotWrapper:
+    def __init__(self, robot: Robot):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: Tensor):
+        with self.lock:
+            self.robot.send_action(action)
+
+    def observation_features(self) -> list[str]:
+        with self.lock:
+            return self.robot.observation_features
+
+    def action_features(self) -> list[str]:
+        with self.lock:
+            return self.robot.action_features
+
+
+@dataclass
+class RTCDemoConfig(HubMixin):
+    """Configuration for RTC demo with action chunking policies and real robots."""
+
+    # Policy configuration
+    policy: PreTrainedConfig | None = None
+
+    # Robot configuration
+    robot: RobotConfig | None = None
+
+    # RTC configuration
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            execution_horizon=10,
+            max_guidance_weight=1.0,
+            prefix_attention_schedule=RTCAttentionSchedule.EXP,
+        )
+    )
+
+    # Demo parameters
+    duration: float = 30.0  # Duration to run the demo (seconds)
+    fps: float = 10.0  # Action execution frequency (Hz)
+
+    # Compute device
+    device: str | None = None  # Device to run on (cuda, cpu, auto)
+
+    # Get new actions horizon. The amount of executed steps after which will be requested new actions.
+    # It should be higher than inference delay + execution horizon.
+    action_queue_size_to_get_new_actions: int = 30
+
+    # Task to execute
+    task: str = field(default="", metadata={"help": "Task to execute"})
+
+    # Torch compile configuration
+    use_torch_compile: bool = field(
+        default=False,
+        metadata={"help": "Use torch.compile for faster inference (PyTorch 2.0+)"},
+    )
+
+    torch_compile_backend: str = field(
+        default="inductor",
+        metadata={"help": "Backend for torch.compile (inductor, aot_eager, cudagraphs)"},
+    )
+
+    torch_compile_mode: str = field(
+        default="default",
+        metadata={"help": "Compilation mode (default, reduce-overhead, max-autotune)"},
+    )
+
+    torch_compile_disable_cudagraphs: bool = field(
+        default=True,
+        metadata={
+            "help": "Disable CUDA graphs in torch.compile. Required due to in-place tensor "
+            "operations in denoising loop (x_t += dt * v_t) which cause tensor aliasing issues."
+        },
+    )
+
+    def __post_init__(self):
+        # HACK: We parse again the cli args here to get the pretrained path if there was one.
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        else:
+            raise ValueError("Policy path is required")
+
+        # Validate that robot configuration is provided
+        if self.robot is None:
+            raise ValueError("Robot configuration must be provided")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
+        return ["policy"]
+
+
+def is_image_key(k: str) -> bool:
+    return k.startswith(OBS_IMAGES)
+
+
+def get_actions(
+    policy,
+    robot: RobotWrapper,
+    robot_observation_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: RTCDemoConfig,
+):
+    """Thread function to request action chunks from the policy.
+
+    Args:
+        policy: The policy instance (SmolVLA, Pi0, etc.)
+        robot: The robot instance for getting observations
+        robot_observation_processor: Processor for raw robot observations
+        action_queue: Queue to put new action chunks
+        shutdown_event: Event to signal shutdown
+        cfg: Demo configuration
+    """
+    try:
+        logger.info("[GET_ACTIONS] Starting get actions thread")
+
+        latency_tracker = LatencyTracker()  # Track latency of action chunks
+        fps = cfg.fps
+        time_per_chunk = 1.0 / fps
+
+        dataset_features = hw_to_dataset_features(robot.observation_features(), "observation")
+        policy_device = policy.config.device
+
+        # Load preprocessor and postprocessor from pretrained files
+        # The stats are embedded in the processor .safetensors files
+        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
+
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=None,  # Will load from pretrained processor files
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.policy.device},
+            },
+        )
+
+        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully with embedded stats")
+
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+
+        while not shutdown_event.is_set():
+            if action_queue.qsize() <= get_actions_threshold:
+                current_time = time.perf_counter()
+                action_index_before_inference = action_queue.get_action_index()
+                prev_actions = action_queue.get_left_over()
+
+                inference_latency = latency_tracker.max()
+                inference_delay = math.ceil(inference_latency / time_per_chunk)
+
+                obs = robot.get_observation()
+
+                # Apply robot observation processor
+                obs_processed = robot_observation_processor(obs)
+
+                obs_with_policy_features = build_dataset_frame(
+                    dataset_features, obs_processed, prefix="observation"
+                )
+
+                for name in obs_with_policy_features:
+                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                    if "image" in name:
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].type(torch.float32) / 255
+                        )
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        )
+                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
+                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
+
+                obs_with_policy_features["task"] = [cfg.task]  # Task should be a list, not a string!
+                obs_with_policy_features["robot_type"] = (
+                    robot.robot.name if hasattr(robot.robot, "name") else ""
+                )
+
+                preproceseded_obs = preprocessor(obs_with_policy_features)
+
+                # Generate actions WITH RTC
+                actions = policy.predict_action_chunk(
+                    preproceseded_obs,
+                    inference_delay=inference_delay,
+                    prev_chunk_left_over=prev_actions,
+                )
+
+                # Store original actions (before postprocessing) for RTC
+                original_actions = actions.squeeze(0).clone()
+
+                postprocessed_actions = postprocessor(actions)
+
+                postprocessed_actions = postprocessed_actions.squeeze(0)
+
+                new_latency = time.perf_counter() - current_time
+                new_delay = math.ceil(new_latency / time_per_chunk)
+                latency_tracker.add(new_latency)
+
+                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
+                    logger.warning(
+                        "[GET_ACTIONS] cfg.action_queue_size_to_get_new_actions Too small, It should be higher than inference delay + execution horizon."
+                    )
+
+                action_queue.merge(
+                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
+                )
+            else:
+                # Small sleep to prevent busy waiting
+                time.sleep(0.1)
+
+        logger.info("[GET_ACTIONS] get actions thread shutting down")
+    except Exception as e:
+        logger.error(f"[GET_ACTIONS] Fatal exception in get_actions thread: {e}")
+        logger.error(traceback.format_exc())
+        sys.exit(1)
+
+
+def actor_control(
+    robot: RobotWrapper,
+    robot_action_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: RTCDemoConfig,
+):
+    """Thread function to execute actions on the robot.
+
+    Args:
+        robot: The robot instance
+        action_queue: Queue to get actions from
+        shutdown_event: Event to signal shutdown
+        cfg: Demo configuration
+    """
+    try:
+        logger.info("[ACTOR] Starting actor thread")
+
+        action_count = 0
+        action_interval = 1.0 / cfg.fps
+
+        while not shutdown_event.is_set():
+            start_time = time.perf_counter()
+
+            # Try to get an action from the queue with timeout
+            action = action_queue.get()
+
+            if action is not None:
+                action = action.cpu()
+                action_dict = {key: action[i].item() for i, key in enumerate(robot.action_features())}
+                action_processed = robot_action_processor((action_dict, None))
+                robot.send_action(action_processed)
+
+                action_count += 1
+
+            dt_s = time.perf_counter() - start_time
+            time.sleep(max(0, (action_interval - dt_s) - 0.001))
+
+        logger.info(f"[ACTOR] Actor thread shutting down. Total actions executed: {action_count}")
+    except Exception as e:
+        logger.error(f"[ACTOR] Fatal exception in actor_control thread: {e}")
+        logger.error(traceback.format_exc())
+        sys.exit(1)
+
+
+def _apply_torch_compile(policy, cfg: RTCDemoConfig):
+    """Apply torch.compile to the policy's predict_action_chunk method.
+
+    Args:
+        policy: Policy instance to compile
+        cfg: Configuration containing torch compile settings
+
+    Returns:
+        Policy with compiled predict_action_chunk method
+    """
+
+    # PI models handle their own compilation
+    if policy.type == "pi05" or policy.type == "pi0":
+        return policy
+
+    try:
+        # Check if torch.compile is available (PyTorch 2.0+)
+        if not hasattr(torch, "compile"):
+            logger.warning(
+                f"torch.compile is not available. Requires PyTorch 2.0+. "
+                f"Current version: {torch.__version__}. Skipping compilation."
+            )
+            return policy
+
+        logger.info("Applying torch.compile to predict_action_chunk...")
+        logger.info(f"  Backend: {cfg.torch_compile_backend}")
+        logger.info(f"  Mode: {cfg.torch_compile_mode}")
+        logger.info(f"  Disable CUDA graphs: {cfg.torch_compile_disable_cudagraphs}")
+
+        # Compile the predict_action_chunk method
+        # - CUDA graphs disabled to prevent tensor aliasing from in-place ops (x_t += dt * v_t)
+        compile_kwargs = {
+            "backend": cfg.torch_compile_backend,
+            "mode": cfg.torch_compile_mode,
+        }
+
+        # Disable CUDA graphs if requested (prevents tensor aliasing issues)
+        if cfg.torch_compile_disable_cudagraphs:
+            compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+        original_method = policy.predict_action_chunk
+        compiled_method = torch.compile(original_method, **compile_kwargs)
+        policy.predict_action_chunk = compiled_method
+        logger.info("✓ Successfully compiled predict_action_chunk")
+
+    except Exception as e:
+        logger.error(f"Failed to apply torch.compile: {e}")
+        logger.warning("Continuing without torch.compile")
+
+    return policy
+
+
+@parser.wrap()
+def demo_cli(cfg: RTCDemoConfig):
+    """Main entry point for RTC demo with draccus configuration."""
+
+    # Initialize logging
+    init_logging()
+
+    logger.info(f"Using device: {cfg.device}")
+
+    # Setup signal handler for graceful shutdown
+    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
+    shutdown_event = signal_handler.shutdown_event
+
+    policy = None
+    robot = None
+    get_actions_thread = None
+    actor_thread = None
+
+    policy_class = get_policy_class(cfg.policy.type)
+
+    # Load config and set compile_model for pi0/pi05 models
+    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+
+    if cfg.policy.type == "pi05" or cfg.policy.type == "pi0":
+        config.compile_model = cfg.use_torch_compile
+
+    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+
+    # Turn on RTC
+    policy.config.rtc_config = cfg.rtc
+
+    # Init RTC processort, as by default if RTC disabled in the config
+    # The processor won't be created
+    policy.init_rtc_processor()
+
+    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
+
+    policy = policy.to(cfg.device)
+    policy.eval()
+
+    # Apply torch.compile to predict_action_chunk method if enabled
+    if cfg.use_torch_compile:
+        policy = _apply_torch_compile(policy, cfg)
+
+    # Create robot
+    logger.info(f"Initializing robot: {cfg.robot.type}")
+    robot = make_robot_from_config(cfg.robot)
+    robot.connect()
+    robot_wrapper = RobotWrapper(robot)
+
+    # Create robot observation processor
+    robot_observation_processor = make_default_robot_observation_processor()
+    robot_action_processor = make_default_robot_action_processor()
+
+    # Create action queue for communication between threads
+    action_queue = ActionQueue(cfg.rtc)
+
+    # Start chunk requester thread
+    get_actions_thread = Thread(
+        target=get_actions,
+        args=(policy, robot_wrapper, robot_observation_processor, action_queue, shutdown_event, cfg),
+        daemon=True,
+        name="GetActions",
+    )
+    get_actions_thread.start()
+    logger.info("Started get actions thread")
+
+    # Start action executor thread
+    actor_thread = Thread(
+        target=actor_control,
+        args=(robot_wrapper, robot_action_processor, action_queue, shutdown_event, cfg),
+        daemon=True,
+        name="Actor",
+    )
+    actor_thread.start()
+    logger.info("Started actor thread")
+
+    logger.info("Started stop by duration thread")
+
+    # Main thread monitors for duration or shutdown
+    logger.info(f"Running demo for {cfg.duration} seconds...")
+    start_time = time.time()
+
+    while not shutdown_event.is_set() and (time.time() - start_time) < cfg.duration:
+        time.sleep(10)
+
+        # Log queue status periodically
+        if int(time.time() - start_time) % 5 == 0:
+            logger.info(f"[MAIN] Action queue size: {action_queue.qsize()}")
+
+        if time.time() - start_time > cfg.duration:
+            break
+
+    logger.info("Demo duration reached or shutdown requested")
+
+    # Signal shutdown
+    shutdown_event.set()
+
+    # Wait for threads to finish
+    if get_actions_thread and get_actions_thread.is_alive():
+        logger.info("Waiting for chunk requester thread to finish...")
+        get_actions_thread.join()
+
+    if actor_thread and actor_thread.is_alive():
+        logger.info("Waiting for action executor thread to finish...")
+        actor_thread.join()
+
+    # Cleanup robot
+    if robot:
+        robot.disconnect()
+        logger.info("Robot disconnected")
+
+    logger.info("Cleanup completed")
+
+
+if __name__ == "__main__":
+    demo_cli()
+    logging.info("RTC demo finished")

examples/so100_to_so100_EE/evaluate.py

@@ -52,126 +52,114 @@ 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,
+def main():
+    # 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,
     )
 
-    # 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")
+    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,
@@ -180,19 +168,40 @@ for episode_idx in range(NUM_EPISODES):
             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
+        # 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,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-listener.stop()
-dataset.push_to_hub()
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
+
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/so100_to_so100_EE/record.py

@@ -48,134 +48,122 @@ 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)
+def main():
+    # 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")
 
-# 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()),
-)
+    # 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
-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,
+    # 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()),
     )
 
-    # 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")
+    # 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,
-            control_time_s=RESET_TIME_SEC,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=leader_joints_to_ee,
@@ -183,20 +171,42 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
             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
+        # 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,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-recording episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-leader.disconnect()
-follower.disconnect()
-listener.stop()
-dataset.push_to_hub()
+        # 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()
+
+
+if __name__ == "__main__":
+    main()

examples/so100_to_so100_EE/replay.py

@@ -30,72 +30,78 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 )
 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.robot_utils import precise_sleep
 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)
+def main():
+    # Initialize the robot config
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+    )
 
-# 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()),
-)
+    # Initialize the robot
+    robot = SO100Follower(robot_config)
 
-# 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,
-)
+    # 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()),
+    )
 
-# 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)
+    # 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,
+    )
 
-# Connect to the robot
-robot.connect()
+    # 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)
 
-if not robot.is_connected:
-    raise ValueError("Robot is not connected!")
+    # Connect to the robot
+    robot.connect()
 
-print("Starting replay loop...")
-log_say(f"Replaying episode {EPISODE_IDX}")
-for idx in range(len(episode_frames)):
-    t0 = time.perf_counter()
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
 
-    # Get recorded action from dataset
-    ee_action = {
-        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-    }
+    print("Starting replay loop...")
+    log_say(f"Replaying episode {EPISODE_IDX}")
+    for idx in range(len(episode_frames)):
+        t0 = time.perf_counter()
 
-    # Get robot observation
-    robot_obs = robot.get_observation()
+        # Get recorded action from dataset
+        ee_action = {
+            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        }
 
-    # Dataset EE -> robot joints
-    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+        # Get robot observation
+        robot_obs = robot.get_observation()
 
-    # Send action to robot
-    _ = robot.send_action(joint_action)
+        # Dataset EE -> robot joints
+        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+        # Send action to robot
+        _ = robot.send_action(joint_action)
 
-# Clean up
-robot.disconnect()
+        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+    # Clean up
+    robot.disconnect()
+
+
+if __name__ == "__main__":
+    main()

examples/so100_to_so100_EE/teleoperate.py

@@ -32,90 +32,96 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 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.robot_utils import precise_sleep
 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)
+def main():
+    # 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")
 
-# 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()),
-)
+    # 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
-leader_kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(leader.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
+    follower_kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(follower.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,
-)
+    # 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 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,
-)
+    # 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,
+    )
 
-# Connect to the robot and teleoperator
-follower.connect()
-leader.connect()
+    # 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,
+    )
 
-# Init rerun viewer
-init_rerun(session_name="so100_so100_EE_teleop")
+    # Connect to the robot and teleoperator
+    follower.connect()
+    leader.connect()
 
-print("Starting teleop loop...")
-while True:
-    t0 = time.perf_counter()
+    # Init rerun viewer
+    init_rerun(session_name="so100_so100_EE_teleop")
 
-    # Get robot observation
-    robot_obs = follower.get_observation()
+    print("Starting teleop loop...")
+    while True:
+        t0 = time.perf_counter()
 
-    # Get teleop observation
-    leader_joints_obs = leader.get_action()
+        # Get robot observation
+        robot_obs = follower.get_observation()
 
-    # teleop joints -> teleop EE action
-    leader_ee_act = leader_to_ee(leader_joints_obs)
+        # Get teleop observation
+        leader_joints_obs = leader.get_action()
 
-    # teleop EE -> robot joints
-    follower_joints_act = ee_to_follower_joints((leader_ee_act, robot_obs))
+        # teleop joints -> teleop EE action
+        leader_ee_act = leader_to_ee(leader_joints_obs)
 
-    # Send action to robot
-    _ = follower.send_action(follower_joints_act)
+        # teleop EE -> robot joints
+        follower_joints_act = ee_to_follower_joints((leader_ee_act, robot_obs))
 
-    # Visualize
-    log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
+        # Send action to robot
+        _ = follower.send_action(follower_joints_act)
 
-    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        # Visualize
+        log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
+
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/act/act_training_example.py

@@ -0,0 +1,104 @@
+"""This script demonstrates how to train ACT Policy on a real-world dataset."""
+
+from pathlib import Path
+
+import torch
+
+from lerobot.configs.types import FeatureType
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import dataset_to_policy_features
+from lerobot.policies.act.configuration_act import ACTConfig
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+
+
+def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:
+    if delta_indices is None:
+        return [0]
+
+    return [i / fps for i in delta_indices]
+
+
+def main():
+    output_directory = Path("outputs/robot_learning_tutorial/act")
+    output_directory.mkdir(parents=True, exist_ok=True)
+
+    # Select your device
+    device = torch.device("mps")  # or "cuda" or "cpu"
+
+    dataset_id = "lerobot/svla_so101_pickplace"
+
+    # This specifies the inputs the model will be expecting and the outputs it will produce
+    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+    features = dataset_to_policy_features(dataset_metadata.features)
+
+    output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    input_features = {key: ft for key, ft in features.items() if key not in output_features}
+
+    cfg = ACTConfig(input_features=input_features, output_features=output_features)
+    policy = ACTPolicy(cfg)
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
+
+    policy.train()
+    policy.to(device)
+
+    # To perform action chunking, ACT expects a given number of actions as targets
+    delta_timestamps = {
+        "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
+    }
+
+    # add image features if they are present
+    delta_timestamps |= {
+        k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps)
+        for k in cfg.image_features
+    }
+
+    # Instantiate the dataset
+    dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
+
+    # Create the optimizer and dataloader for offline training
+    optimizer = cfg.get_optimizer_preset().build(policy.parameters())
+    batch_size = 32
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=True,
+        pin_memory=device.type != "cpu",
+        drop_last=True,
+    )
+
+    # Number of training steps and logging frequency
+    training_steps = 1
+    log_freq = 1
+
+    # Run training loop
+    step = 0
+    done = False
+    while not done:
+        for batch in dataloader:
+            batch = preprocessor(batch)
+            loss, _ = policy.forward(batch)
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+
+            if step % log_freq == 0:
+                print(f"step: {step} loss: {loss.item():.3f}")
+            step += 1
+            if step >= training_steps:
+                done = True
+                break
+
+    # Save the policy checkpoint, alongside the pre/post processors
+    policy.save_pretrained(output_directory)
+    preprocessor.save_pretrained(output_directory)
+    postprocessor.save_pretrained(output_directory)
+
+    # Save all assets to the Hub
+    policy.push_to_hub("<user>/robot_learning_tutorial_act")
+    preprocessor.push_to_hub("<user>/robot_learning_tutorial_act")
+    postprocessor.push_to_hub("<user>/robot_learning_tutorial_act")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/act/act_using_example.py

@@ -0,0 +1,63 @@
+import torch
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.policies.act.modeling_act import ACTPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.utils import build_inference_frame, make_robot_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+
+def main():
+    device = torch.device("mps")  # or "cuda" or "cpu"
+    model_id = "<user>/robot_learning_tutorial_act"
+    model = ACTPolicy.from_pretrained(model_id)
+
+    dataset_id = "lerobot/svla_so101_pickplace"
+    # This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
+    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+    preprocess, postprocess = make_pre_post_processors(model.config, dataset_stats=dataset_metadata.stats)
+
+    # # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+    # # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
+
+    # Robot and environment configuration
+    # Camera keys must match the name and resolutions of the ones used for training!
+    # You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+    camera_config = {
+        "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    }
+
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+    robot = SO100Follower(robot_cfg)
+    robot.connect()
+
+    for _ in range(MAX_EPISODES):
+        for _ in range(MAX_STEPS_PER_EPISODE):
+            obs = robot.get_observation()
+            obs_frame = build_inference_frame(
+                observation=obs, ds_features=dataset_metadata.features, device=device
+            )
+
+            obs = preprocess(obs_frame)
+
+            action = model.select_action(obs)
+            action = postprocess(action)
+
+            action = make_robot_action(action, dataset_metadata.features)
+
+            robot.send_action(action)
+
+        print("Episode finished! Starting new episode...")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/async-inf/policy_server.py

@@ -0,0 +1,17 @@
+from lerobot.async_inference.configs import PolicyServerConfig
+from lerobot.async_inference.policy_server import serve
+
+
+def main():
+    host = ...  # something like "127.0.0.1" if you're exposing to localhost
+    port = ...  # something like 8080
+
+    config = PolicyServerConfig(
+        host=host,
+        port=port,
+    )
+    serve(config)
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/async-inf/robot_client.py

@@ -0,0 +1,61 @@
+import threading
+
+from lerobot.async_inference.configs import RobotClientConfig
+from lerobot.async_inference.helpers import visualize_action_queue_size
+from lerobot.async_inference.robot_client import RobotClient
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.robots.so100_follower import SO100FollowerConfig
+
+
+def main():
+    # these cameras must match the ones expected by the policy - find your cameras with lerobot-find-cameras
+    # check the config.json on the Hub for the policy you are using to see the expected camera specs
+    camera_cfg = {
+        "up": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "side": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    }
+
+    # # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+    # # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
+
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_cfg)
+
+    server_address = ...  # something like "127.0.0.1:8080" if using localhost
+
+    # 3. Create client configuration
+    client_cfg = RobotClientConfig(
+        robot=robot_cfg,
+        server_address=server_address,
+        policy_device="mps",
+        policy_type="act",
+        pretrained_name_or_path="<user>/robot_learning_tutorial_act",
+        chunk_size_threshold=0.5,  # g
+        actions_per_chunk=50,  # make sure this is less than the max actions of the policy
+    )
+
+    # 4. Create and start client
+    client = RobotClient(client_cfg)
+
+    # 5. Provide a textual description of the task
+    task = ...
+
+    if client.start():
+        # Start action receiver thread
+        action_receiver_thread = threading.Thread(target=client.receive_actions, daemon=True)
+        action_receiver_thread.start()
+
+        try:
+            # Run the control loop
+            client.control_loop(task)
+        except KeyboardInterrupt:
+            client.stop()
+            action_receiver_thread.join()
+            # (Optionally) plot the action queue size
+            visualize_action_queue_size(client.action_queue_size)
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/diffusion/diffusion_training_example.py

@@ -0,0 +1,105 @@
+"""This script demonstrates how to train Diffusion Policy on a real-world dataset."""
+
+from pathlib import Path
+
+import torch
+
+from lerobot.configs.types import FeatureType
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.datasets.utils import dataset_to_policy_features
+from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
+from lerobot.policies.factory import make_pre_post_processors
+
+
+def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:
+    if delta_indices is None:
+        return [0]
+
+    return [i / fps for i in delta_indices]
+
+
+def main():
+    output_directory = Path("outputs/robot_learning_tutorial/diffusion")
+    output_directory.mkdir(parents=True, exist_ok=True)
+
+    # Select your device
+    device = torch.device("mps")  # or "cuda" or "cpu"
+
+    dataset_id = "lerobot/svla_so101_pickplace"
+
+    # This specifies the inputs the model will be expecting and the outputs it will produce
+    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+    features = dataset_to_policy_features(dataset_metadata.features)
+
+    output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    input_features = {key: ft for key, ft in features.items() if key not in output_features}
+
+    cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
+    policy = DiffusionPolicy(cfg)
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
+
+    policy.train()
+    policy.to(device)
+
+    # To perform action chunking, ACT expects a given number of actions as targets
+    delta_timestamps = {
+        "observation.state": make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps),
+        "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
+    }
+
+    # add image features if they are present
+    delta_timestamps |= {
+        k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps)
+        for k in cfg.image_features
+    }
+
+    # Instantiate the dataset
+    dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
+
+    # Create the optimizer and dataloader for offline training
+    optimizer = cfg.get_optimizer_preset().build(policy.parameters())
+    batch_size = 32
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=True,
+        pin_memory=device.type != "cpu",
+        drop_last=True,
+    )
+
+    # Number of training steps and logging frequency
+    training_steps = 1
+    log_freq = 1
+
+    # Run training loop
+    step = 0
+    done = False
+    while not done:
+        for batch in dataloader:
+            batch = preprocessor(batch)
+            loss, _ = policy.forward(batch)
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+
+            if step % log_freq == 0:
+                print(f"step: {step} loss: {loss.item():.3f}")
+            step += 1
+            if step >= training_steps:
+                done = True
+                break
+
+    # Save the policy checkpoint, alongside the pre/post processors
+    policy.save_pretrained(output_directory)
+    preprocessor.save_pretrained(output_directory)
+    postprocessor.save_pretrained(output_directory)
+
+    # Save all assets to the Hub
+    policy.push_to_hub("<user>/robot_learning_tutorial_diffusion")
+    preprocessor.push_to_hub("<user>/robot_learning_tutorial_diffusion")
+    postprocessor.push_to_hub("<user>/robot_learning_tutorial_diffusion")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/diffusion/diffusion_using_example.py

@@ -0,0 +1,64 @@
+import torch
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.utils import build_inference_frame, make_robot_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+
+def main():
+    device = torch.device("mps")  # or "cuda" or "cpu"
+    model_id = "<user>/robot_learning_tutorial_diffusion"
+
+    model = DiffusionPolicy.from_pretrained(model_id)
+
+    dataset_id = "lerobot/svla_so101_pickplace"
+    # This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
+    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+    preprocess, postprocess = make_pre_post_processors(
+        model.config, model_id, dataset_stats=dataset_metadata.stats
+    )
+
+    # # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+    # # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
+
+    # Robot and environment configuration
+    # Camera keys must match the name and resolutions of the ones used for training!
+    # You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+    camera_config = {
+        "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    }
+
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+    robot = SO100Follower(robot_cfg)
+    robot.connect()
+
+    for _ in range(MAX_EPISODES):
+        for _ in range(MAX_STEPS_PER_EPISODE):
+            obs = robot.get_observation()
+            obs_frame = build_inference_frame(
+                observation=obs, ds_features=dataset_metadata.features, device=device
+            )
+
+            obs = preprocess(obs_frame)
+
+            action = model.select_action(obs)
+            action = postprocess(action)
+            action = make_robot_action(action, dataset_metadata.features)
+            robot.send_action(action)
+
+        print("Episode finished! Starting new episode...")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/pi0/using_pi0_example.py

@@ -0,0 +1,73 @@
+import torch
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.pi0.modeling_pi0 import PI0Policy
+from lerobot.policies.utils import build_inference_frame, make_robot_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+
+def main():
+    device = torch.device("mps")  # or "cuda" or "cpu"
+    model_id = "lerobot/pi0_base"
+
+    model = PI0Policy.from_pretrained(model_id)
+
+    preprocess, postprocess = make_pre_post_processors(
+        model.config,
+        model_id,
+        # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
+        preprocessor_overrides={"device_processor": {"device": str(device)}},
+    )
+
+    # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+    # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
+
+    # Robot and environment configuration
+    # Camera keys must match the name and resolutions of the ones used for training!
+    # You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+    camera_config = {
+        "base_0_rgb": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "left_wrist_0_rgb": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+        "right_wrist_0_rgb": OpenCVCameraConfig(index_or_path=2, width=640, height=480, fps=30),
+    }
+
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+    robot = SO100Follower(robot_cfg)
+    robot.connect()
+
+    task = ""  # something like "pick the red block"
+    robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
+
+    # This is used to match the raw observation keys to the keys expected by the policy
+    action_features = hw_to_dataset_features(robot.action_features, "action")
+    obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+    dataset_features = {**action_features, **obs_features}
+
+    for _ in range(MAX_EPISODES):
+        for _ in range(MAX_STEPS_PER_EPISODE):
+            obs = robot.get_observation()
+            obs_frame = build_inference_frame(
+                observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
+            )
+
+            obs = preprocess(obs_frame)
+
+            action = model.select_action(obs)
+            action = postprocess(action)
+            action = make_robot_action(action, dataset_features)
+            robot.send_action(action)
+
+        print("Episode finished! Starting new episode...")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/rl/hilserl_example.py

@@ -0,0 +1,347 @@
+import multiprocessing as mp
+import signal
+from pathlib import Path
+from queue import Empty, Full
+
+import torch
+import torch.optim as optim
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
+from lerobot.policies.sac.configuration_sac import SACConfig
+from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies.sac.reward_model.modeling_classifier import Classifier
+from lerobot.rl.buffer import ReplayBuffer
+from lerobot.rl.gym_manipulator import make_robot_env
+from lerobot.robots.so100_follower import SO100FollowerConfig
+from lerobot.teleoperators.so100_leader import SO100LeaderConfig
+from lerobot.teleoperators.utils import TeleopEvents
+
+LOG_EVERY = 10
+SEND_EVERY = 10
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+
+def run_learner(
+    transitions_queue: mp.Queue,
+    parameters_queue: mp.Queue,
+    shutdown_event: mp.Event,
+    policy_learner: SACPolicy,
+    online_buffer: ReplayBuffer,
+    offline_buffer: ReplayBuffer,
+    lr: float = 3e-4,
+    batch_size: int = 32,
+    device: torch.device = "mps",
+):
+    """The learner process - trains SAC policy on transitions streamed from the actor, updating parameters
+    for the actor to adopt."""
+    policy_learner.train()
+    policy_learner.to(device)
+
+    # Create Adam optimizer from scratch - simple and clean
+    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
+
+    print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
+    print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
+
+    training_step = 0
+
+    while not shutdown_event.is_set():
+        # retrieve incoming transitions from the actor process
+        try:
+            transitions = transitions_queue.get(timeout=0.1)
+            for transition in transitions:
+                # HIL-SERL: Add ALL transitions to online buffer
+                online_buffer.add(**transition)
+
+                # HIL-SERL: Add ONLY human intervention transitions to offline buffer
+                is_intervention = transition.get("complementary_info", {}).get("is_intervention", False)
+                if is_intervention:
+                    offline_buffer.add(**transition)
+                    print(
+                        f"[LEARNER] Human intervention detected! Added to offline buffer (now {len(offline_buffer)} transitions)"
+                    )
+
+        except Empty:
+            pass  # No transitions available, continue
+
+        # Train if we have enough data
+        if len(online_buffer) >= policy_learner.config.online_step_before_learning:
+            # Sample from online buffer (autonomous + human data)
+            online_batch = online_buffer.sample(batch_size // 2)
+
+            # Sample from offline buffer (human demonstrations only, either precollected or at runtime)
+            offline_batch = offline_buffer.sample(batch_size // 2)
+
+            # Combine batches - this is the key HIL-SERL mechanism!
+            batch = {}
+            for key in online_batch:
+                if key in offline_batch:
+                    batch[key] = torch.cat([online_batch[key], offline_batch[key]], dim=0)
+                else:
+                    batch[key] = online_batch[key]
+
+            loss, _ = policy_learner.forward(batch)
+
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+            training_step += 1
+
+            if training_step % LOG_EVERY == 0:
+                print(
+                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
+                    f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
+                )
+
+            # Send updated parameters to actor every 10 training steps
+            if training_step % SEND_EVERY == 0:
+                try:
+                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
+                    parameters_queue.put_nowait(state_dict)
+                    print("[LEARNER] Sent updated parameters to actor")
+                except Full:
+                    # Missing write due to queue not being consumed (should happen rarely)
+                    pass
+
+    print("[LEARNER] Learner process finished")
+
+
+def run_actor(
+    transitions_queue: mp.Queue,
+    parameters_queue: mp.Queue,
+    shutdown_event: mp.Event,
+    policy_actor: SACPolicy,
+    reward_classifier: Classifier,
+    env_cfg: HILSerlRobotEnvConfig,
+    device: torch.device = "mps",
+    output_directory: Path | None = None,
+):
+    """The actor process - interacts with environment and collects data.
+    The policy is frozen and only the parameters are updated, popping the most recent ones from a queue."""
+    policy_actor.eval()
+    policy_actor.to(device)
+
+    reward_classifier.eval()
+    reward_classifier.to(device)
+
+    # Create robot environment inside the actor process
+    env, teleop_device = make_robot_env(env_cfg)
+
+    try:
+        for episode in range(MAX_EPISODES):
+            if shutdown_event.is_set():
+                break
+
+            obs, _info = env.reset()
+            episode_reward = 0.0
+            step = 0
+            episode_transitions = []
+
+            print(f"[ACTOR] Starting episode {episode + 1}")
+
+            while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
+                try:
+                    new_params = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_params)
+                    print("[ACTOR] Updated policy parameters from learner")
+                except Empty:  # No new updated parameters available from learner, waiting
+                    pass
+
+                # Get action from policy
+                policy_obs = make_policy_obs(obs, device=device)
+                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
+                action = action_tensor.squeeze(0).cpu().numpy()
+
+                # Step environment
+                next_obs, _env_reward, terminated, truncated, _info = env.step(action)
+                done = terminated or truncated
+
+                # Predict reward
+                policy_next_obs = make_policy_obs(next_obs, device=device)
+                reward = reward_classifier.predict_reward(policy_next_obs)
+
+                if reward >= 1.0 and not done:  # success detected! halt episode
+                    terminated = True
+                    done = True
+
+                # In HIL-SERL, human interventions come from the teleop device
+                is_intervention = False
+                if hasattr(teleop_device, "get_teleop_events"):
+                    # Real intervention detection from teleop device
+                    teleop_events = teleop_device.get_teleop_events()
+                    is_intervention = teleop_events.get(TeleopEvents.IS_INTERVENTION, False)
+
+                # Store transition with intervention metadata
+                transition = {
+                    "state": policy_obs,
+                    "action": action,
+                    "reward": float(reward) if hasattr(reward, "item") else reward,
+                    "next_state": policy_next_obs,
+                    "done": done,
+                    "truncated": truncated,
+                    "complementary_info": {
+                        "is_intervention": is_intervention,
+                    },
+                }
+
+                episode_transitions.append(transition)
+
+                episode_reward += reward
+                step += 1
+
+                obs = next_obs
+
+                if done:
+                    break
+
+            # Send episode transitions to learner
+            transitions_queue.put_nowait(episode_transitions)
+
+    except KeyboardInterrupt:
+        print("[ACTOR] Interrupted by user")
+    finally:
+        # Clean up
+        if hasattr(env, "robot") and env.robot.is_connected:
+            env.robot.disconnect()
+        if teleop_device and hasattr(teleop_device, "disconnect"):
+            teleop_device.disconnect()
+        if output_directory is not None:
+            policy_actor.save_pretrained(output_directory)
+            print(f"[ACTOR] Latest actor policy saved at: {output_directory}")
+
+        print("[ACTOR] Actor process finished")
+
+
+def make_policy_obs(obs, device: torch.device = "cpu"):
+    return {
+        "observation.state": torch.from_numpy(obs["agent_pos"]).float().unsqueeze(0).to(device),
+        **{
+            f"observation.image.{k}": torch.from_numpy(obs["pixels"][k]).float().unsqueeze(0).to(device)
+            for k in obs["pixels"]
+        },
+    }
+
+
+def main():
+    """Main function - coordinates actor and learner processes."""
+
+    device = "mps"  # or "cuda" or "cpu"
+    output_directory = Path("outputs/robot_learning_tutorial/hil_serl")
+    output_directory.mkdir(parents=True, exist_ok=True)
+
+    # find ports using lerobot-find-port
+    follower_port = ...
+    leader_port = ...
+
+    # the robot ids are used the load the right calibration files
+    follower_id = ...
+    leader_id = ...
+
+    # A pretrained model (to be used in-distribution!)
+    reward_classifier_id = "<user>/reward_classifier_hil_serl_example"
+    reward_classifier = Classifier.from_pretrained(reward_classifier_id)
+
+    reward_classifier.to(device)
+    reward_classifier.eval()
+
+    # Robot and environment configuration
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id)
+    teleop_cfg = SO100LeaderConfig(port=leader_port, id=leader_id)
+    processor_cfg = HILSerlProcessorConfig(control_mode="leader")
+
+    env_cfg = HILSerlRobotEnvConfig(robot=robot_cfg, teleop=teleop_cfg, processor=processor_cfg)
+
+    # Create robot environment
+    env, teleop_device = make_robot_env(env_cfg)
+
+    obs_features = hw_to_dataset_features(env.robot.observation_features, "observation")
+    action_features = hw_to_dataset_features(env.robot.action_features, "action")
+
+    # Create SAC policy for action selection
+    policy_cfg = SACConfig(
+        device=device,
+        input_features=obs_features,
+        output_features=action_features,
+    )
+
+    policy_actor = SACPolicy(policy_cfg)
+    policy_learner = SACPolicy(policy_cfg)
+
+    demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
+    offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)
+
+    # Online buffer: initialized from scratch
+    online_replay_buffer = ReplayBuffer(device=device, state_keys=list(obs_features.keys()))
+    # Offline buffer: Created from dataset (pre-populated it with demonstrations)
+    offline_replay_buffer = ReplayBuffer.from_lerobot_dataset(
+        lerobot_dataset=offline_dataset, device=device, state_keys=list(obs_features.keys())
+    )
+
+    # Create communication channels between learner and actor processes
+    transitions_queue = mp.Queue(maxsize=10)
+    parameters_queue = mp.Queue(maxsize=2)
+    shutdown_event = mp.Event()
+
+    # Signal handler for graceful shutdown
+    def signal_handler(sig):
+        print(f"\nSignal {sig} received, shutting down...")
+        shutdown_event.set()
+
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTERM, signal_handler)
+
+    # Create processes
+    learner_process = mp.Process(
+        target=run_learner,
+        args=(
+            transitions_queue,
+            parameters_queue,
+            shutdown_event,
+            policy_learner,
+            online_replay_buffer,
+            offline_replay_buffer,
+        ),
+        kwargs={"device": device},  # can run on accelerated hardware for training
+    )
+
+    actor_process = mp.Process(
+        target=run_actor,
+        args=(
+            transitions_queue,
+            parameters_queue,
+            shutdown_event,
+            policy_actor,
+            reward_classifier,
+            env_cfg,
+            output_directory,
+        ),
+        kwargs={"device": "cpu"},  # actor is frozen, can run on CPU or accelerate for inference
+    )
+
+    learner_process.start()
+    actor_process.start()
+
+    try:
+        # Wait for actor to finish (it controls the episode loop)
+        actor_process.join()
+        shutdown_event.set()
+        learner_process.join(timeout=10)
+
+    except KeyboardInterrupt:
+        print("Main process interrupted")
+        shutdown_event.set()
+        actor_process.join(timeout=5)
+        learner_process.join(timeout=10)
+
+    finally:
+        if learner_process.is_alive():
+            learner_process.terminate()
+        if actor_process.is_alive():
+            actor_process.terminate()
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/rl/reward_classifier_example.py

@@ -0,0 +1,67 @@
+import torch
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.policies.factory import make_policy, make_pre_post_processors
+from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
+
+
+def main():
+    # Device to use for training
+    device = "mps"  # or "cuda", or "cpu"
+
+    # Load the dataset used for training
+    repo_id = "lerobot/example_hil_serl_dataset"
+    dataset = LeRobotDataset(repo_id)
+
+    # Configure the policy to extract features from the image frames
+    camera_keys = dataset.meta.camera_keys
+
+    config = RewardClassifierConfig(
+        num_cameras=len(camera_keys),
+        device=device,
+        # backbone model to extract features from the image frames
+        model_name="microsoft/resnet-18",
+    )
+
+    # Make policy, preprocessor, and optimizer
+    policy = make_policy(config, ds_meta=dataset.meta)
+    optimizer = config.get_optimizer_preset().build(policy.parameters())
+    preprocessor, _ = make_pre_post_processors(policy_cfg=config, dataset_stats=dataset.meta.stats)
+
+    classifier_id = "<user>/reward_classifier_hil_serl_example"
+
+    # Instantiate a dataloader
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=16, shuffle=True)
+
+    # Training loop
+    num_epochs = 5
+    for epoch in range(num_epochs):
+        total_loss = 0
+        total_accuracy = 0
+        for batch in dataloader:
+            # Preprocess the batch and move it to the correct device.
+            batch = preprocessor(batch)
+
+            # Forward pass
+            loss, output_dict = policy.forward(batch)
+
+            # Backward pass and optimization
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+            total_loss += loss.item()
+            total_accuracy += output_dict["accuracy"]
+
+        avg_loss = total_loss / len(dataloader)
+        avg_accuracy = total_accuracy / len(dataloader)
+        print(f"Epoch {epoch + 1}/{num_epochs}, Loss: {avg_loss:.4f}, Accuracy: {avg_accuracy:.2f}%")
+
+    print("Training finished!")
+
+    # You can now save the trained policy.
+    policy.push_to_hub(classifier_id)
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/smolvla/using_smolvla_example.py

@@ -0,0 +1,72 @@
+import torch
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.datasets.utils import hw_to_dataset_features
+from lerobot.policies.factory import make_pre_post_processors
+from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
+from lerobot.policies.utils import build_inference_frame, make_robot_action
+from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
+from lerobot.robots.so100_follower.so100_follower import SO100Follower
+
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
+
+
+def main():
+    device = torch.device("mps")  # or "cuda" or "cpu"
+    model_id = "lerobot/smolvla_base"
+
+    model = SmolVLAPolicy.from_pretrained(model_id)
+
+    preprocess, postprocess = make_pre_post_processors(
+        model.config,
+        model_id,
+        # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
+        preprocessor_overrides={"device_processor": {"device": str(device)}},
+    )
+
+    # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+    # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
+
+    # Robot and environment configuration
+    # Camera keys must match the name and resolutions of the ones used for training!
+    # You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+    camera_config = {
+        "camera1": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "camera2": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    }
+
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+    robot = SO100Follower(robot_cfg)
+    robot.connect()
+
+    task = ""  # something like "pick the red block"
+    robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
+
+    # This is used to match the raw observation keys to the keys expected by the policy
+    action_features = hw_to_dataset_features(robot.action_features, "action")
+    obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+    dataset_features = {**action_features, **obs_features}
+
+    for _ in range(MAX_EPISODES):
+        for _ in range(MAX_STEPS_PER_EPISODE):
+            obs = robot.get_observation()
+            obs_frame = build_inference_frame(
+                observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
+            )
+
+            obs = preprocess(obs_frame)
+
+            action = model.select_action(obs)
+            action = postprocess(action)
+            action = make_robot_action(action, dataset_features)
+            robot.send_action(action)
+
+        print("Episode finished! Starting new episode...")
+
+
+if __name__ == "__main__":
+    main()

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.3.4"
+version = "0.4.2"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -59,28 +59,30 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0",
-    "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",
+    "accelerate>=1.10.0,<2.0.0",
 
     # 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,<0.22.0", # TODO: Bumb dependency (compatible with protobuf)
 
     "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
-    "rerun-sdk>=0.21.0,<0.23.0", # TODO: Bumb dependency
+    "gymnasium>=1.1.1,<2.0.0",
+    "rerun-sdk>=0.24.0,<0.27.0",
 
     # Support dependencies
     "deepdiff>=7.0.1,<9.0.0",
@@ -92,51 +94,56 @@ dependencies = [
 [project.optional-dependencies]
 
 # Common
-pygame-dep = ["pygame>=2.5.1"]
-placo-dep = ["placo>=0.9.6"]
-transformers-dep = ["transformers>=4.52.0"]
-grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"]
+pygame-dep = ["pygame>=2.5.1,<2.7.0"]
+placo-dep = ["placo>=0.9.6,<0.10.0"]
+transformers-dep = ["transformers>=4.53.0,<5.0.0"]
+grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"] # TODO: Bumb dependency (compatible with wandb)
 
 # 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", "teleop>=0.1.0"]
-# stretch = [
-#     "hello-robot-stretch-body>=0.7.27 ; sys_platform == 'linux'",
-#     "pyrender @ git+https://github.com/mmatl/pyrender.git ; sys_platform == 'linux'",
-#     "pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'"
-# ] # TODO: Currently not supported
+phone = ["hebi-py>=2.8.0,<2.12.0", "teleop>=0.1.0,<0.2.0", "fastapi<1.0"]
 
 # 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>=0.1.11", "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"]
+groot = [
+    "lerobot[transformers-dep]",
+    "peft>=0.13.0,<1.0.0",
+    "dm-tree>=0.1.8,<1.0.0",
+    "timm>=1.0.0,<1.1.0",
+    "safetensors>=0.4.3,<1.0.0",
+    "Pillow>=10.0.0,<13.0.0",
+    "decord>=0.6.0,<1.0.0; (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
+    "ninja>=1.11.1,<2.0.0",
+    "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
+]
+hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # 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"]
-libero = ["lerobot[transformers-dep]", "libero @ git+https://github.com/huggingface/lerobot-libero.git@main#egg=libero"]
-
+aloha = ["gym-aloha>=0.1.2,<0.2.0"]
+pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
+metaworld = ["metaworld==3.0.0"]
 
 # All
 all = [
@@ -147,8 +154,9 @@ all = [
     "lerobot[reachy2]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
-    "lerobot[pi0]",
+    "lerobot[pi]",
     "lerobot[smolvla]",
+    # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
     "lerobot[hilserl]",
     "lerobot[async]",
     "lerobot[dev]",
@@ -156,9 +164,9 @@ all = [
     "lerobot[video_benchmark]",
     "lerobot[aloha]",
     "lerobot[pusht]",
-    "lerobot[xarm]",
     "lerobot[phone]",
     "lerobot[libero]",
+    "lerobot[metaworld]",
 ]
 
 [project.scripts]
@@ -175,6 +183,7 @@ 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]
@@ -232,9 +241,6 @@ exclude_dirs = [
     "tests",
     "benchmarks",
     "src/lerobot/datasets/push_dataset_to_hub",
-    "src/lerobot/datasets/v2/convert_dataset_v1_to_v2",
-    "src/lerobot/policies/pi0/conversion_scripts",
-    "src/lerobot/scripts/push_dataset_to_hub.py",
 ]
 skips = ["B101", "B311", "B404", "B603", "B615"]
 
@@ -249,6 +255,8 @@ default.extend-ignore-identifiers-re = [
     "pn",
     "ser",
     "ein",
+    "thw",
+    "inpt",
 ]
 
 # TODO: Uncomment when ready to use
@@ -270,80 +278,88 @@ default.extend-ignore-identifiers-re = [
 # 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"
+[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.*"
+ignore_errors = false
+
+
 # [[tool.mypy.overrides]]
 # module = "lerobot.utils.*"
-# # include = "src/lerobot/utils/**/*.py"
+# ignore_errors = false
+
+[[tool.mypy.overrides]]
+module = "lerobot.configs.*"
+ignore_errors = false
+
+# extra strictness for configs
+disallow_untyped_defs = true
+disallow_incomplete_defs = true
+check_untyped_defs = true
 
 # [[tool.mypy.overrides]]
-# module = "lerobot.configs.*"
-# # include = "src/lerobot/configs/**/*.py"
+# module = "lerobot.optim.*"
+# ignore_errors = false
+
+[[tool.mypy.overrides]]
+module = "lerobot.model.*"
+ignore_errors = false
 
-# # Data processing modules
 # [[tool.mypy.overrides]]
 # module = "lerobot.processor.*"
-# # include = "src/lerobot/processor/**/*.py"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.datasets.*"
-# # include = "src/lerobot/datasets/**/*.py"
+# ignore_errors = false
 
-# # Core machine learning modules
-# [[tool.mypy.overrides]]
-# module = "lerobot.optim.*"
-# # include = "src/lerobot/optim/**/*.py"
-
-# [[tool.mypy.overrides]]
-# module = "lerobot.model.*"
-# # include = "src/lerobot/model/**/*.py"
-
-# # Hardware interfaces
-# [[tool.mypy.overrides]]
-# module = "lerobot.cameras.*"
-# # include = "src/lerobot/cameras/**/*.py"
+[[tool.mypy.overrides]]
+module = "lerobot.cameras.*"
+ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.motors.*"
-# # include = "src/lerobot/motors/**/*.py"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"
-# # include = "src/lerobot/robots/**/*.py"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.teleoperators.*"
-# # include = "src/lerobot/teleoperators/**/*.py"
+# ignore_errors = false
 
-# # Complex modules (enable these last)
 # [[tool.mypy.overrides]]
 # module = "lerobot.policies.*"
-# # include = "src/lerobot/policies/**/*.py"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.rl.*"
-# # include = "src/lerobot/rl/**/*.py"
+# ignore_errors = false
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.envs.*"
-# # include = "src/lerobot/envs/**/*.py"
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.async_inference.*"
-# # include = "src/lerobot/async_inference/**/*.py"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.transport.*"
-# # include = "src/lerobot/transport/**/*.py"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"
-# # include = "src/lerobot/scripts/**/*.py"
+# ignore_errors = false

requirements-macos.txt

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

requirements-ubuntu.txt

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

requirements.in

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

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

@@ -142,11 +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

@@ -16,7 +16,7 @@ import logging
 import logging.handlers
 import os
 import time
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 
 import torch
@@ -25,7 +25,14 @@ from lerobot.configs.types import PolicyFeature
 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
@@ -55,15 +62,6 @@ 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, OBS_STR, use_video=False)
 
@@ -85,11 +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 = {}
 
@@ -98,9 +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
 
@@ -272,6 +268,7 @@ class RemotePolicyConfig:
     lerobot_features: dict[str, PolicyFeature]
     actions_per_chunk: int
     device: str = "cpu"
+    rename_map: dict[str, str] = field(default_factory=dict)
 
 
 def _compare_observation_states(obs1_state: torch.Tensor, obs2_state: torch.Tensor, atol: float) -> bool:

src/lerobot/async_inference/policy_server.py

@@ -15,7 +15,7 @@
 """
 Example:
 ```shell
-python src/lerobot/async_inference/policy_server.py \
+python -m lerobot.async_inference.policy_server \
      --host=127.0.0.1 \
      --port=8080 \
      --fps=30 \
@@ -32,12 +32,17 @@ 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.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
@@ -82,6 +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):
@@ -146,6 +153,19 @@ 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,
+                "rename_observations_processor": {"rename_map": policy_specs.rename_map},
+            },
+            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")
@@ -173,7 +193,7 @@ 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} | "
@@ -189,7 +209,7 @@ 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()
 
@@ -301,23 +321,6 @@ 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)
@@ -327,44 +330,76 @@ 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

@@ -48,10 +48,10 @@ 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,
@@ -75,7 +75,6 @@ from .helpers import (
     TimedObservation,
     get_logger,
     map_robot_keys_to_lerobot_features,
-    validate_robot_cameras_for_policy,
     visualize_action_queue_size,
 )
 
@@ -97,14 +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
 
@@ -214,7 +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
 
@@ -467,7 +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/camera.py

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

src/lerobot/cameras/configs.py

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

src/lerobot/cameras/opencv/__init__.py

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

src/lerobot/cameras/opencv/camera_opencv.py

@@ -25,11 +25,12 @@ from pathlib import Path
 from threading import Event, Lock, Thread
 from typing import Any
 
+from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
+
 # Fix MSMF hardware transform compatibility for Windows before importing cv2
 if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS" not in os.environ:
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
-import cv2
-import numpy as np
+import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
 from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
@@ -121,7 +122,7 @@ class OpenCVCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: np.ndarray | None = None
+        self.latest_frame: NDArray[Any] | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -140,7 +141,7 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
-    def connect(self, warmup: bool = True):
+    def connect(self, warmup: bool = True) -> None:
         """
         Connects to the OpenCV camera specified in the configuration.
 
@@ -180,12 +181,14 @@ class OpenCVCamera(Camera):
 
     def _configure_capture_settings(self) -> None:
         """
-        Applies the specified FPS, width, and height settings to the connected camera.
+        Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
 
         This method attempts to set the camera properties via OpenCV. It checks if
         the camera successfully applied the settings and raises an error if not.
+        FOURCC is set first (if specified) as it can affect the available FPS and resolution options.
 
         Args:
+            fourcc: The desired FOURCC code (e.g., "MJPG", "YUYV"). If None, auto-detect.
             fps: The desired frames per second. If None, the setting is skipped.
             width: The desired capture width. If None, the setting is skipped.
             height: The desired capture height. If None, the setting is skipped.
@@ -199,10 +202,11 @@ class OpenCVCamera(Camera):
         if not self.is_connected:
             raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
-        if self.fps is None:
-            self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
-        else:
-            self._validate_fps()
+        # Set FOURCC first (if specified) as it can affect available FPS/resolution options
+        if self.config.fourcc is not None:
+            self._validate_fourcc()
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
 
         default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
         default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
@@ -216,18 +220,56 @@ class OpenCVCamera(Camera):
         else:
             self._validate_width_and_height()
 
+        if self.fps is None:
+            self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
+        else:
+            self._validate_fps()
+
     def _validate_fps(self) -> None:
         """Validates and sets the camera's frames per second (FPS)."""
 
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
+        if self.fps is None:
+            raise ValueError(f"{self} FPS is not set")
+
         success = self.videocapture.set(cv2.CAP_PROP_FPS, float(self.fps))
         actual_fps = self.videocapture.get(cv2.CAP_PROP_FPS)
         # Use math.isclose for robust float comparison
         if not success or not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
             raise RuntimeError(f"{self} failed to set fps={self.fps} ({actual_fps=}).")
 
+    def _validate_fourcc(self) -> None:
+        """Validates and sets the camera's FOURCC code."""
+
+        fourcc_code = cv2.VideoWriter_fourcc(*self.config.fourcc)
+
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
+        success = self.videocapture.set(cv2.CAP_PROP_FOURCC, fourcc_code)
+        actual_fourcc_code = self.videocapture.get(cv2.CAP_PROP_FOURCC)
+
+        # Convert actual FOURCC code back to string for comparison
+        actual_fourcc_code_int = int(actual_fourcc_code)
+        actual_fourcc = "".join([chr((actual_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
+
+        if not success or actual_fourcc != self.config.fourcc:
+            logger.warning(
+                f"{self} failed to set fourcc={self.config.fourcc} (actual={actual_fourcc}, success={success}). "
+                f"Continuing with default format."
+            )
+
     def _validate_width_and_height(self) -> None:
         """Validates and sets the camera's frame capture width and height."""
 
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
+        if self.capture_width is None or self.capture_height is None:
+            raise ValueError(f"{self} capture_width or capture_height is not set")
+
         width_success = self.videocapture.set(cv2.CAP_PROP_FRAME_WIDTH, float(self.capture_width))
         height_success = self.videocapture.set(cv2.CAP_PROP_FRAME_HEIGHT, float(self.capture_height))
 
@@ -258,11 +300,12 @@ class OpenCVCamera(Camera):
         """
         found_cameras_info = []
 
+        targets_to_scan: list[str | int]
         if platform.system() == "Linux":
             possible_paths = sorted(Path("/dev").glob("video*"), key=lambda p: p.name)
             targets_to_scan = [str(p) for p in possible_paths]
         else:
-            targets_to_scan = list(range(MAX_OPENCV_INDEX))
+            targets_to_scan = [int(i) for i in range(MAX_OPENCV_INDEX)]
 
         for target in targets_to_scan:
             camera = cv2.VideoCapture(target)
@@ -271,6 +314,12 @@ class OpenCVCamera(Camera):
                 default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
                 default_fps = camera.get(cv2.CAP_PROP_FPS)
                 default_format = camera.get(cv2.CAP_PROP_FORMAT)
+
+                # Get FOURCC code and convert to string
+                default_fourcc_code = camera.get(cv2.CAP_PROP_FOURCC)
+                default_fourcc_code_int = int(default_fourcc_code)
+                default_fourcc = "".join([chr((default_fourcc_code_int >> 8 * i) & 0xFF) for i in range(4)])
+
                 camera_info = {
                     "name": f"OpenCV Camera @ {target}",
                     "type": "OpenCV",
@@ -278,6 +327,7 @@ class OpenCVCamera(Camera):
                     "backend_api": camera.getBackendName(),
                     "default_stream_profile": {
                         "format": default_format,
+                        "fourcc": default_fourcc,
                         "width": default_width,
                         "height": default_height,
                         "fps": default_fps,
@@ -289,7 +339,7 @@ class OpenCVCamera(Camera):
 
         return found_cameras_info
 
-    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
 
@@ -317,6 +367,9 @@ class OpenCVCamera(Camera):
 
         start_time = time.perf_counter()
 
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
+
         ret, frame = self.videocapture.read()
 
         if not ret or frame is None:
@@ -329,7 +382,7 @@ class OpenCVCamera(Camera):
 
         return processed_frame
 
-    def _postprocess_image(self, image: np.ndarray, color_mode: ColorMode | None = None) -> np.ndarray:
+    def _postprocess_image(self, image: NDArray[Any], color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw frame.
 
@@ -372,7 +425,7 @@ class OpenCVCamera(Camera):
 
         return processed_image
 
-    def _read_loop(self):
+    def _read_loop(self) -> None:
         """
         Internal loop run by the background thread for asynchronous reading.
 
@@ -383,6 +436,9 @@ class OpenCVCamera(Camera):
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
+        if self.stop_event is None:
+            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
+
         while not self.stop_event.is_set():
             try:
                 color_image = self.read()
@@ -419,7 +475,7 @@ class OpenCVCamera(Camera):
         self.thread = None
         self.stop_event = None
 
-    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
+    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
 
@@ -462,7 +518,7 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    def disconnect(self):
+    def disconnect(self) -> None:
         """
         Disconnects from the camera and cleans up resources.
 

src/lerobot/cameras/opencv/configuration_opencv.py

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

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

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

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

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

src/lerobot/cameras/realsense/camera_realsense.py

@@ -21,11 +21,12 @@ import time
 from threading import Event, Lock, Thread
 from typing import Any
 
-import cv2
-import numpy as np
+import cv2  # type: ignore  # TODO: add type stubs for OpenCV
+import numpy as np  # type: ignore  # TODO: add type stubs for numpy
+from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
 try:
-    import pyrealsense2 as rs
+    import pyrealsense2 as rs  # type: ignore  # TODO: add type stubs for pyrealsense2
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
@@ -132,7 +133,7 @@ class RealSenseCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_frame: np.ndarray | None = None
+        self.latest_frame: NDArray[Any] | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -150,7 +151,7 @@ class RealSenseCamera(Camera):
         """Checks if the camera pipeline is started and streams are active."""
         return self.rs_pipeline is not None and self.rs_profile is not None
 
-    def connect(self, warmup: bool = True):
+    def connect(self, warmup: bool = True) -> None:
         """
         Connects to the RealSense camera specified in the configuration.
 
@@ -264,7 +265,7 @@ class RealSenseCamera(Camera):
         serial_number = str(found_devices[0]["serial_number"])
         return serial_number
 
-    def _configure_rs_pipeline_config(self, rs_config):
+    def _configure_rs_pipeline_config(self, rs_config: Any) -> None:
         """Creates and configures the RealSense pipeline configuration object."""
         rs.config.enable_device(rs_config, self.serial_number)
 
@@ -293,6 +294,9 @@ class RealSenseCamera(Camera):
         if not self.is_connected:
             raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
 
+        if self.rs_profile is None:
+            raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
+
         stream = self.rs_profile.get_stream(rs.stream.color).as_video_stream_profile()
 
         if self.fps is None:
@@ -308,7 +312,7 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
-    def read_depth(self, timeout_ms: int = 200) -> np.ndarray:
+    def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (depth) synchronously from the camera.
 
@@ -336,6 +340,9 @@ class RealSenseCamera(Camera):
 
         start_time = time.perf_counter()
 
+        if self.rs_pipeline is None:
+            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
+
         ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
         if not ret or frame is None:
@@ -351,7 +358,7 @@ class RealSenseCamera(Camera):
 
         return depth_map_processed
 
-    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> np.ndarray:
+    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (color) synchronously from the camera.
 
@@ -376,6 +383,9 @@ class RealSenseCamera(Camera):
 
         start_time = time.perf_counter()
 
+        if self.rs_pipeline is None:
+            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
+
         ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
         if not ret or frame is None:
@@ -392,8 +402,8 @@ class RealSenseCamera(Camera):
         return color_image_processed
 
     def _postprocess_image(
-        self, image: np.ndarray, color_mode: ColorMode | None = None, depth_frame: bool = False
-    ) -> np.ndarray:
+        self, image: NDArray[Any], color_mode: ColorMode | None = None, depth_frame: bool = False
+    ) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw color frame.
 
@@ -438,7 +448,7 @@ class RealSenseCamera(Camera):
 
         return processed_image
 
-    def _read_loop(self):
+    def _read_loop(self) -> None:
         """
         Internal loop run by the background thread for asynchronous reading.
 
@@ -449,6 +459,9 @@ class RealSenseCamera(Camera):
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
         """
+        if self.stop_event is None:
+            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
+
         while not self.stop_event.is_set():
             try:
                 color_image = self.read(timeout_ms=500)
@@ -474,7 +487,7 @@ class RealSenseCamera(Camera):
         self.thread.daemon = True
         self.thread.start()
 
-    def _stop_read_thread(self):
+    def _stop_read_thread(self) -> None:
         """Signals the background read thread to stop and waits for it to join."""
         if self.stop_event is not None:
             self.stop_event.set()
@@ -486,7 +499,7 @@ class RealSenseCamera(Camera):
         self.stop_event = None
 
     # NOTE(Steven): Missing implementation for depth for now
-    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
+    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame data (color) asynchronously.
 
@@ -529,7 +542,7 @@ class RealSenseCamera(Camera):
 
         return frame
 
-    def disconnect(self):
+    def disconnect(self) -> None:
         """
         Disconnects from the camera, stops the pipeline, and cleans up resources.
 

src/lerobot/cameras/realsense/configuration_realsense.py

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

src/lerobot/cameras/utils.py

@@ -15,15 +15,19 @@
 # limitations under the License.
 
 import platform
+from typing import cast
+
+from lerobot.utils.import_utils import make_device_from_device_class
 
 from .camera import Camera
 from .configs import CameraConfig, Cv2Rotation
 
 
 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
 
@@ -40,20 +44,23 @@ 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
 
 
 def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
-    import cv2
+    import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
     if rotation == Cv2Rotation.ROTATE_90:
-        return cv2.ROTATE_90_CLOCKWISE
+        return int(cv2.ROTATE_90_CLOCKWISE)
     elif rotation == Cv2Rotation.ROTATE_180:
-        return cv2.ROTATE_180
+        return int(cv2.ROTATE_180)
     elif rotation == Cv2Rotation.ROTATE_270:
-        return cv2.ROTATE_90_COUNTERCLOCKWISE
+        return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
     else:
         return None
 
@@ -62,8 +69,8 @@ def get_cv2_backend() -> int:
     import cv2
 
     if platform.system() == "Windows":
-        return cv2.CAP_MSMF  # Use MSMF for Windows instead of AVFOUNDATION
+        return int(cv2.CAP_MSMF)  # Use MSMF for Windows instead of AVFOUNDATION
     # elif platform.system() == "Darwin":  # macOS
     #     return cv2.CAP_AVFOUNDATION
     else:  # Linux and others
-        return cv2.CAP_ANY
+        return int(cv2.CAP_ANY)

src/lerobot/configs/default.py

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

src/lerobot/configs/eval.py

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

src/lerobot/configs/parser.py

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

src/lerobot/configs/policies.py

@@ -14,12 +14,12 @@
 import abc
 import builtins
 import json
-import logging
 import os
 import tempfile
 from dataclasses import dataclass, field
+from logging import getLogger
 from pathlib import Path
-from typing import TypeVar
+from typing import Any, TypeVar
 
 import draccus
 from huggingface_hub import hf_hub_download
@@ -34,10 +34,11 @@ from lerobot.utils.hub import HubMixin
 from lerobot.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
 
 T = TypeVar("T", bound="PreTrainedConfig")
+logger = getLogger(__name__)
 
 
 @dataclass
-class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
+class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: ignore[misc,name-defined] #TODO: draccus issue
     """
     Base configuration class for policy models.
 
@@ -57,12 +58,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     input_features: dict[str, PolicyFeature] = field(default_factory=dict)
     output_features: dict[str, PolicyFeature] = field(default_factory=dict)
 
-    device: str | None = None  # cuda | cpu | mp
+    device: str | None = None  # e.g. "cuda", "cuda:0", "cpu", or "mps"
     # `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
     # automatic gradient scaling is used.
     use_amp: bool = False
 
-    push_to_hub: bool = True
+    push_to_hub: bool = True  # type: ignore[assignment] # TODO: use a different name to avoid override
     repo_id: str | None = None
 
     # Upload on private repository on the Hugging Face hub.
@@ -71,38 +72,43 @@ 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: Path | None = None
 
-    def __post_init__(self):
-        self.pretrained_path = None
+    def __post_init__(self) -> 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}'.")
+            logger.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
             self.device = auto_device.type
 
         # Automatically deactivate AMP if necessary
         if self.use_amp and not is_amp_available(self.device):
-            logging.warning(
+            logger.warning(
                 f"Automatic Mixed Precision (amp) is not available on device '{self.device}'. Deactivating AMP."
             )
             self.use_amp = False
 
     @property
     def type(self) -> str:
-        return self.get_choice_name(self.__class__)
+        choice_name = self.get_choice_name(self.__class__)
+        if not isinstance(choice_name, str):
+            raise TypeError(f"Expected string from get_choice_name, got {type(choice_name)}")
+        return choice_name
 
     @property
     @abc.abstractmethod
-    def observation_delta_indices(self) -> list | None:
+    def observation_delta_indices(self) -> list | None:  # type: ignore[type-arg] #TODO: No implementation
         raise NotImplementedError
 
     @property
     @abc.abstractmethod
-    def action_delta_indices(self) -> list | None:
+    def action_delta_indices(self) -> list | None:  # type: ignore[type-arg]    #TODO: No implementation
         raise NotImplementedError
 
     @property
     @abc.abstractmethod
-    def reward_delta_indices(self) -> list | None:
+    def reward_delta_indices(self) -> list | None:  # type: ignore[type-arg]    #TODO: No implementation
         raise NotImplementedError
 
     @abc.abstractmethod
@@ -152,13 +158,13 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
         pretrained_name_or_path: str | Path,
         *,
         force_download: bool = False,
-        resume_download: bool = None,
-        proxies: dict | None = None,
+        resume_download: bool | None = None,
+        proxies: dict[Any, Any] | None = None,
         token: str | bool | None = None,
         cache_dir: str | Path | None = None,
         local_files_only: bool = False,
         revision: str | None = None,
-        **policy_kwargs,
+        **policy_kwargs: Any,
     ) -> T:
         model_id = str(pretrained_name_or_path)
         config_file: str | None = None
@@ -166,7 +172,7 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
             if CONFIG_NAME in os.listdir(model_id):
                 config_file = os.path.join(model_id, CONFIG_NAME)
             else:
-                print(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
+                logger.error(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
         else:
             try:
                 config_file = hf_hub_download(
@@ -192,6 +198,9 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
         with draccus.config_type("json"):
             orig_config = draccus.parse(cls, config_file, args=[])
 
+        if config_file is None:
+            raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
+
         with open(config_file) as f:
             config = json.load(f)
 

src/lerobot/configs/train.py

@@ -16,6 +16,7 @@ import datetime as dt
 import os
 from dataclasses import dataclass, field
 from pathlib import Path
+from typing import Any
 
 import draccus
 from huggingface_hub import hf_hub_download
@@ -63,18 +64,18 @@ class TrainPipelineConfig(HubMixin):
     scheduler: LRSchedulerConfig | None = None
     eval: EvalConfig = field(default_factory=EvalConfig)
     wandb: WandBConfig = field(default_factory=WandBConfig)
+    checkpoint_path: Path | None = field(init=False, default=None)
+    # Rename map for the observation to override the image and state keys
+    rename_map: dict[str, str] = field(default_factory=dict)
 
-    def __post_init__(self):
-        self.checkpoint_path = None
-
-    def validate(self):
+    def validate(self) -> None:
         # HACK: We parse again the cli args here to get the pretrained paths if there was some.
         policy_path = parser.get_path_arg("policy")
         if policy_path:
             # Only load the policy config
             cli_overrides = parser.get_cli_overrides("policy")
             self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
-            self.policy.pretrained_path = policy_path
+            self.policy.pretrained_path = Path(policy_path)
         elif self.resume:
             # The entire train config is already loaded, we just need to get the checkpoint dir
             config_path = parser.parse_arg("config_path")
@@ -82,14 +83,22 @@ class TrainPipelineConfig(HubMixin):
                 raise ValueError(
                     f"A config_path is expected when resuming a run. Please specify path to {TRAIN_CONFIG_NAME}"
                 )
+
             if not Path(config_path).resolve().exists():
                 raise NotADirectoryError(
                     f"{config_path=} is expected to be a local path. "
                     "Resuming from the hub is not supported for now."
                 )
-            policy_path = Path(config_path).parent
-            self.policy.pretrained_path = policy_path
-            self.checkpoint_path = policy_path.parent
+
+            policy_dir = Path(config_path).parent
+            if self.policy is not None:
+                self.policy.pretrained_path = policy_dir
+            self.checkpoint_path = policy_dir.parent
+
+        if self.policy is None:
+            raise ValueError(
+                "Policy is not configured. Please specify a pretrained policy with `--policy.path`."
+            )
 
         if not self.job_name:
             if self.env is None:
@@ -126,8 +135,8 @@ class TrainPipelineConfig(HubMixin):
         """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
         return ["policy"]
 
-    def to_dict(self) -> dict:
-        return draccus.encode(self)
+    def to_dict(self) -> dict[str, Any]:
+        return draccus.encode(self)  # type: ignore[no-any-return]  # because of the third-party library draccus uses Any as the return type
 
     def _save_pretrained(self, save_directory: Path) -> None:
         with open(save_directory / TRAIN_CONFIG_NAME, "w") as f, draccus.config_type("json"):
@@ -139,13 +148,13 @@ class TrainPipelineConfig(HubMixin):
         pretrained_name_or_path: str | Path,
         *,
         force_download: bool = False,
-        resume_download: bool = None,
-        proxies: dict | None = None,
+        resume_download: bool | None = None,
+        proxies: dict[Any, Any] | None = None,
         token: str | bool | None = None,
         cache_dir: str | Path | None = None,
         local_files_only: bool = False,
         revision: str | None = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> "TrainPipelineConfig":
         model_id = str(pretrained_name_or_path)
         config_file: str | None = None
@@ -181,4 +190,6 @@ class TrainPipelineConfig(HubMixin):
 
 @dataclass(kw_only=True)
 class TrainRLServerPipelineConfig(TrainPipelineConfig):
-    dataset: DatasetConfig | None = None  # NOTE: In RL, we don't need an offline dataset
+    # NOTE: In RL, we don't need an offline dataset
+    # TODO: Make `TrainPipelineConfig.dataset` optional
+    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional

src/lerobot/configs/types.py

@@ -35,9 +35,18 @@ class NormalizationMode(str, Enum):
     MIN_MAX = "MIN_MAX"
     MEAN_STD = "MEAN_STD"
     IDENTITY = "IDENTITY"
+    QUANTILES = "QUANTILES"
+    QUANTILE10 = "QUANTILE10"
 
 
 @dataclass
 class PolicyFeature:
     type: FeatureType
-    shape: tuple
+    shape: tuple[int, ...]
+
+
+class RTCAttentionSchedule(str, Enum):
+    ZEROS = "ZEROS"
+    ONES = "ONES"
+    LINEAR = "LINEAR"
+    EXP = "EXP"

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]):
@@ -130,10 +130,34 @@ def update_meta_data(
     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():
-        df[f"videos/{key}/chunk_index"] = df[f"videos/{key}/chunk_index"] + video_idx["chunk"]
-        df[f"videos/{key}/file_index"] = df[f"videos/{key}/file_index"] + video_idx["file"]
-        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"]
+        # Store original video file indices before updating
+        orig_chunk_col = f"videos/{key}/chunk_index"
+        orig_file_col = f"videos/{key}/file_index"
+        df["_orig_chunk"] = df[orig_chunk_col].copy()
+        df["_orig_file"] = df[orig_file_col].copy()
+
+        # Update chunk and file indices to point to destination
+        df[orig_chunk_col] = video_idx["chunk"]
+        df[orig_file_col] = video_idx["file"]
+
+        # Apply per-source-file timestamp offsets
+        src_to_offset = video_idx.get("src_to_offset", {})
+        if src_to_offset:
+            # Apply offset based on original source file
+            for idx in df.index:
+                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
+                offset = src_to_offset.get(src_key, 0)
+                df.at[idx, f"videos/{key}/from_timestamp"] += offset
+                df.at[idx, f"videos/{key}/to_timestamp"] += offset
+        else:
+            # Fallback to simple offset (for backward compatibility)
+            df[f"videos/{key}/from_timestamp"] = (
+                df[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
+            )
+            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"]
@@ -193,6 +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")
@@ -236,6 +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)
@@ -249,6 +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(
@@ -263,21 +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,
@@ -286,25 +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
 
 
@@ -389,9 +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,
@@ -403,6 +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/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}")
 
@@ -87,8 +110,8 @@ def worker_thread_loop(queue: queue.Queue):
         if item is None:
             queue.task_done()
             break
-        image_array, fpath = item
-        write_image(image_array, fpath)
+        image_array, fpath, compress_level = item
+        write_image(image_array, fpath, compress_level)
         queue.task_done()
 
 
@@ -146,11 +169,13 @@ class AsyncImageWriter:
                 p.start()
                 self.processes.append(p)
 
-    def save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path):
+    def save_image(
+        self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1
+    ):
         if isinstance(image, torch.Tensor):
             # Convert tensor to numpy array to minimize main process time
             image = image.cpu().numpy()
-        self.queue.put((image, fpath))
+        self.queue.put((image, fpath, compress_level))
 
     def wait_until_done(self):
         self.queue.join()

src/lerobot/datasets/lerobot_dataset.py

@@ -13,8 +13,8 @@
 # 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 concurrent.futures
 import contextlib
-import gc
 import logging
 import shutil
 import tempfile
@@ -26,6 +26,8 @@ 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
@@ -46,13 +48,9 @@ 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,
@@ -60,7 +58,6 @@ 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,
@@ -90,10 +87,15 @@ 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 +109,54 @@ 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 +188,12 @@ 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 +201,12 @@ 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 +322,75 @@ 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,
@@ -366,9 +431,7 @@ class LeRobotDatasetMetadata:
         video_keys = [video_key] if video_key is not None else self.video_keys
         for key in video_keys:
             if not self.features[key].get("info", None):
-                video_path = self.root / self.video_path.format(
-                    video_key=video_key, chunk_index=0, file_index=0
-                )
+                video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
                 self.info["features"][key]["info"] = get_video_info(video_path)
 
     def update_chunk_settings(
@@ -438,6 +501,10 @@ 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)
@@ -452,14 +519,36 @@ 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
 
 
+def _encode_video_worker(video_key: str, episode_index: int, root: Path, fps: int) -> Path:
+    temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
+    fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
+    img_dir = (root / fpath).parent
+    encode_video_frames(img_dir, temp_path, fps, overwrite=True)
+    shutil.rmtree(img_dir)
+    return temp_path
+
+
 class LeRobotDataset(torch.utils.data.Dataset):
     def __init__(
         self,
@@ -603,6 +692,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # Unused attributes
         self.image_writer = None
         self.episode_buffer = None
+        self.writer = None
+        self.latest_episode = None
+        self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -611,6 +703,11 @@ 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:
@@ -620,15 +717,38 @@ class LeRobotDataset(torch.utils.data.Dataset):
             if not self._check_cached_episodes_sufficient():
                 raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
         except (AssertionError, FileNotFoundError, NotADirectoryError):
-            self.revision = get_safe_version(self.repo_id, self.revision)
+            if is_valid_version(self.revision):
+                self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
             self.hf_dataset = self.load_hf_dataset()
 
+        # Create mapping from absolute indices to relative indices when only a subset of the episodes are loaded
+        # Build a mapping: absolute_index -> relative_index_in_filtered_dataset
+        self._absolute_to_relative_idx = None
+        if self.episodes is not None:
+            self._absolute_to_relative_idx = {
+                abs_idx.item() if isinstance(abs_idx, torch.Tensor) else abs_idx: rel_idx
+                for rel_idx, abs_idx in enumerate(self.hf_dataset["index"])
+            }
+
         # Setup delta_indices
         if self.delta_timestamps is not None:
             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,
@@ -729,31 +849,40 @@ class LeRobotDataset(torch.utils.data.Dataset):
     def load_hf_dataset(self) -> datasets.Dataset:
         """hf_dataset contains all the observations, states, actions, rewards, etc."""
         features = get_hf_features_from_features(self.features)
-        hf_dataset = load_nested_dataset(self.root / "data", features=features)
+        hf_dataset = load_nested_dataset(self.root / "data", features=features, episodes=self.episodes)
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
     def _check_cached_episodes_sufficient(self) -> bool:
-        """Check if the cached dataset contains all requested episodes."""
+        """Check if the cached dataset contains all requested episodes and their video files."""
         if self.hf_dataset is None or len(self.hf_dataset) == 0:
             return False
 
         # Get available episode indices from cached dataset
         available_episodes = {
             ep_idx.item() if isinstance(ep_idx, torch.Tensor) else ep_idx
-            for ep_idx in self.hf_dataset["episode_index"]
+            for ep_idx in self.hf_dataset.unique("episode_index")
         }
 
         # Determine requested episodes
         if self.episodes is None:
-            # Requesting all episodes - check if we have all episodes from metadata
             requested_episodes = set(range(self.meta.total_episodes))
         else:
-            # Requesting specific episodes
             requested_episodes = set(self.episodes)
 
         # Check if all requested episodes are available in cached data
-        return requested_episodes.issubset(available_episodes)
+        if not requested_episodes.issubset(available_episodes):
+            return False
+
+        # Check if all required video files exist
+        if len(self.meta.video_keys) > 0:
+            for ep_idx in requested_episodes:
+                for vid_key in self.meta.video_keys:
+                    video_path = self.root / self.meta.get_video_file_path(ep_idx, vid_key)
+                    if not video_path.exists():
+                        return False
+
+        return True
 
     def create_hf_dataset(self) -> datasets.Dataset:
         features = get_hf_features_from_features(self.features)
@@ -769,8 +898,15 @@ 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:
@@ -813,7 +949,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         query_timestamps = {}
         for key in self.meta.video_keys:
             if query_indices is not None and key in query_indices:
-                timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
+                if self._absolute_to_relative_idx is not None:
+                    relative_indices = [self._absolute_to_relative_idx[idx] for idx in query_indices[key]]
+                    timestamps = self.hf_dataset[relative_indices]["timestamp"]
+                else:
+                    timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
                 query_timestamps[key] = torch.stack(timestamps).tolist()
             else:
                 query_timestamps[key] = [current_ts]
@@ -821,11 +961,32 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return query_timestamps
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
-        return {
-            key: torch.stack(self.hf_dataset[q_idx][key])
-            for key, q_idx in query_indices.items()
-            if key not in self.meta.video_keys
-        }
+        """
+        Query dataset for indices across keys, skipping video keys.
+
+        Tries column-first [key][indices] for speed, falls back to row-first.
+
+        Args:
+            query_indices: Dict mapping keys to index lists to retrieve
+
+        Returns:
+            Dict with stacked tensors of queried data (video keys excluded)
+        """
+        result: dict = {}
+        for key, q_idx in query_indices.items():
+            if key in self.meta.video_keys:
+                continue
+            # Map absolute indices to relative indices if needed
+            relative_indices = (
+                q_idx
+                if self._absolute_to_relative_idx is None
+                else [self._absolute_to_relative_idx[idx] for idx in q_idx]
+            )
+            try:
+                result[key] = torch.stack(self.hf_dataset[key][relative_indices])
+            except (KeyError, TypeError, IndexError):
+                result[key] = torch.stack(self.hf_dataset[relative_indices][key])
+        return result
 
     def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict[str, torch.Tensor]:
         """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
@@ -848,10 +1009,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         return item
 
+    def _ensure_hf_dataset_loaded(self):
+        """Lazy load the HF dataset only when needed for reading."""
+        if self._lazy_loading or self.hf_dataset is None:
+            # Close the writer before loading to ensure parquet file is properly finalized
+            if self.writer is not None:
+                self._close_writer()
+                self._writer_closed_for_reading = True
+            self.hf_dataset = self.load_hf_dataset()
+            self._lazy_loading = False
+
     def __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()
 
@@ -890,6 +1063,14 @@ 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 = {}
@@ -900,6 +1081,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             ep_buffer[key] = current_ep_idx if key == "episode_index" else []
         return ep_buffer
 
+    # TODO(Steven): consider move this to utils
     def _get_image_file_path(self, episode_index: int, image_key: str, frame_index: int) -> Path:
         fpath = DEFAULT_IMAGE_PATH.format(
             image_key=image_key, episode_index=episode_index, frame_index=frame_index
@@ -909,13 +1091,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
     def _get_image_file_dir(self, episode_index: int, image_key: str) -> Path:
         return self._get_image_file_path(episode_index, image_key, frame_index=0).parent
 
-    def _save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path) -> None:
+    def _save_image(
+        self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1
+    ) -> None:
         if self.image_writer is None:
             if isinstance(image, torch.Tensor):
                 image = image.cpu().numpy()
-            write_image(image, fpath)
+            write_image(image, fpath, compress_level=compress_level)
         else:
-            self.image_writer.save_image(image=image, fpath=fpath)
+            self.image_writer.save_image(image=image, fpath=fpath, compress_level=compress_level)
 
     def add_frame(self, frame: dict) -> None:
         """
@@ -953,14 +1137,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 )
                 if frame_index == 0:
                     img_path.parent.mkdir(parents=True, exist_ok=True)
-                self._save_image(frame[key], img_path)
+                compress_level = 1 if self.features[key]["dtype"] == "video" else 6
+                self._save_image(frame[key], img_path, compress_level)
                 self.episode_buffer[key].append(str(img_path))
             else:
                 self.episode_buffer[key].append(frame[key])
 
         self.episode_buffer["size"] += 1
 
-    def save_episode(self, episode_data: dict | None = None) -> None:
+    def save_episode(
+        self,
+        episode_data: dict | None = None,
+        parallel_encoding: bool = True,
+    ) -> None:
         """
         This will save to disk the current episode in self.episode_buffer.
 
@@ -972,6 +1161,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
             episode_data (dict | None, optional): Dict containing the episode data to save. If None, this will
                 save the current episode in self.episode_buffer, which is filled with 'add_frame'. Defaults to
                 None.
+            parallel_encoding (bool, optional): If True, encode videos in parallel using ProcessPoolExecutor.
+                Defaults to True on Linux, False on macOS as it tends to use all the CPU available already.
         """
         episode_buffer = episode_data if episode_data is not None else self.episode_buffer
 
@@ -1008,8 +1199,40 @@ class LeRobotDataset(torch.utils.data.Dataset):
         use_batched_encoding = self.batch_encoding_size > 1
 
         if has_video_keys and not use_batched_encoding:
-            for video_key in self.meta.video_keys:
-                ep_metadata.update(self._save_episode_video(video_key, episode_index))
+            num_cameras = len(self.meta.video_keys)
+            if parallel_encoding and num_cameras > 1:
+                # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
+                # num_cameras * num_threads = (total_cpu -1)
+                with concurrent.futures.ProcessPoolExecutor(max_workers=num_cameras) as executor:
+                    future_to_key = {
+                        executor.submit(
+                            _encode_video_worker,
+                            video_key,
+                            episode_index,
+                            self.root,
+                            self.fps,
+                        ): video_key
+                        for video_key in self.meta.video_keys
+                    }
+
+                    results = {}
+                    for future in concurrent.futures.as_completed(future_to_key):
+                        video_key = future_to_key[future]
+                        try:
+                            temp_path = future.result()
+                            results[video_key] = temp_path
+                        except Exception as exc:
+                            logging.error(f"Video encoding failed for {video_key}: {exc}")
+                            raise exc
+
+                for video_key in self.meta.video_keys:
+                    temp_path = results[video_key]
+                    ep_metadata.update(
+                        self._save_episode_video(video_key, episode_index, temp_path=temp_path)
+                    )
+            else:
+                for video_key in self.meta.video_keys:
+                    ep_metadata.update(self._save_episode_video(video_key, episode_index))
 
         # `meta.save_episode` need to be executed after encoding the videos
         self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats, ep_metadata)
@@ -1097,74 +1320,113 @@ 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
+            self._current_file_start_frame = 0
+            # However, if the episodes already exists
+            # It means we are resuming recording, so we need to load the latest episode
+            # Update the indices to avoid overwriting the latest episode
+            if self.meta.episodes is not None and len(self.meta.episodes) > 0:
+                latest_ep = self.meta.episodes[-1]
+                global_frame_index = latest_ep["dataset_to_index"]
+                chunk_idx = latest_ep["data/chunk_index"]
+                file_idx = latest_ep["data/file_index"]
+
+                # When resuming, move to the next file
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
+                self._current_file_start_frame = global_frame_index
         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 - self._current_file_start_frame
+            av_size_per_frame = (
+                latest_size_in_mb / frames_in_current_file if frames_in_current_file > 0 else 0
+            )
 
             # 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
+                self._current_file_start_frame = global_frame_index
 
-                # 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) -> dict:
+    def _save_episode_video(
+        self,
+        video_key: str,
+        episode_index: int,
+        temp_path: Path | None = None,
+    ) -> 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)
+        if temp_path is None:
+            ep_path = self._encode_temporary_episode_video(video_key, episode_index)
+        else:
+            ep_path = temp_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
@@ -1172,16 +1434,16 @@ 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
@@ -1264,11 +1526,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        temp_path = Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4"
-        img_dir = self._get_image_file_dir(episode_index, video_key)
-        encode_video_frames(img_dir, temp_path, self.fps, overwrite=True)
-        shutil.rmtree(img_dir)
-        return temp_path
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps)
 
     @classmethod
     def create(
@@ -1314,7 +1572,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_transforms = None
         obj.delta_timestamps = None
         obj.delta_indices = None
+        obj._absolute_to_relative_idx = None
         obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
+        obj.writer = None
+        obj.latest_episode = None
+        obj._current_file_start_frame = None
+        # Initialize tracking for incremental recording
+        obj._lazy_loading = False
+        obj._recorded_frames = 0
+        obj._writer_closed_for_reading = False
         return obj