feat(anyskin): adding support for anyskin sensors

.github/workflows/documentation.yml

@@ -18,11 +18,6 @@ name: Documentation
 on:
   # Allows running this workflow manually from the Actions tab
   workflow_dispatch:
-    inputs:
-      version:
-        description: 'Version tag (e.g. v0.1.2) - Leave empty for standard main build'
-        required: false
-        type: string
 
   # Triggers the workflow on push events to main for the docs folder
   push:
@@ -59,13 +54,7 @@ jobs:
     with:
       commit_sha: ${{ github.sha }}
       package: lerobot
-      additional_args: >-
-        --not_python_module
-        ${{
-          (github.event_name == 'release' && format('--version {0}', github.event.release.tag_name)) ||
-          (inputs.version != '' && format('--version {0}', inputs.version)) ||
-          ''
-        }}
+      additional_args: --not_python_module ${{ github.event_name == 'release' && format('--version {0}', github.event.release.tag_name) || '' }}
     secrets:
       token: ${{ secrets.HUGGINGFACE_PUSH }}
       hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/full_tests.yml

@@ -101,11 +101,9 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     if: |
-      github.repository == 'huggingface/lerobot' && (
-        (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
-        github.event_name == 'push' ||
-        github.event_name == 'workflow_dispatch'
-      )
+      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+      github.event_name == 'push' ||
+      github.event_name == 'workflow_dispatch'
     outputs:
       image_tag: ${{ steps.set_tag.outputs.image_tag }}
     env:
@@ -173,8 +171,6 @@ jobs:
         shell: bash
         working-directory: /lerobot
     steps:
-      - name: Fix ptxas permissions
-        run: chmod +x /lerobot/.venv/lib/python3.10/site-packages/triton/backends/nvidia/bin/ptxas
       - name: Run pytest on GPU
         run: pytest tests -vv --maxfail=10
       - name: Run end-to-end tests

.github/workflows/nightly.yml

@@ -188,7 +188,7 @@ jobs:
       - name: Verify GPU availability
         run: |
           nvidia-smi
-          python -c "import torch; print(f'PyTorch version: {torch.__version__}'); print(f'PyTorch CUDA available: {torch.cuda.is_available()}'); print(f'Number of GPUs: {torch.cuda.device_count()}')"
+          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

.github/workflows/unbound_deps_tests.yml

@@ -91,7 +91,6 @@ jobs:
     name: Build and Push Docker
     runs-on:
       group: aws-general-8-plus
-    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME }}
     env:

AI_POLICY.md

@@ -1,25 +0,0 @@
-# AI Usage Policy
-
-The LeRobot project welcomes contributions from everyone, and we have a few guidelines regarding AI usage to ensure high code quality, clear communication, and a healthy open-source ecosystem:
-
-- **Please disclose significant AI assistance.** If you used AI tools (e.g., Copilot, Claude, Cursor, ChatGPT) to generate a substantial portion of your code or text, let us know in your PR description. Transparency helps us review your changes more effectively.
-- **Own your code (The Human-in-the-Loop).** You must fully understand all the changes you are proposing. If you cannot explain what your AI-assisted code does or how it interacts with LeRobot's broader architecture, please take the time to learn and test it before submitting.
-- **Keep issues and discussions focused.** You are welcome to use AI to help draft issues or PR descriptions, but please review and edit them carefully before posting. AI can often be overly verbose; trimming the noise and getting straight to the point helps our maintainers address your needs faster.
-
-Our core maintainers also use AI tools to aid their workflows, but they do so while bringing deep contextual knowledge of the LeRobot codebase to validate the output. We ask all contributors to apply that same level of rigor.
-
-## Remember the Human Maintainers
-
-Please remember that LeRobot is maintained by a dedicated team of humans.
-
-Every discussion, issue, and pull request is read and reviewed by real people. While AI tools can generate thousands of lines of code in seconds, reviewing that code still takes human time and energy. Submitting unverified or low-effort AI output puts an unfair burden on our maintainers.
-
-Today, the quality of the AI output still heavily depends on the developer driving the tool. We ask that you respect our maintainers' time by thoroughly vetting, testing, and refining your submissions.
-
-## AI is Welcome Here
-
-LeRobot operates at the cutting edge of AI and robotics, and many of our maintainers actively embrace AI coding assistants as valuable productivity tools. We are a pro-AI project!
-
-Our reason for having an AI policy is not an anti-AI stance. Rather, it exists to ensure that AI is used to enhance human contributions, not replace them with unverified noise. It's about how the tools are used, not the tools themselves.
-
-We value the unique human insight you bring to the LeRobot community. Let AI empower your workflow, but always let your own judgment take the wheel.

CONTRIBUTING.md

@@ -2,7 +2,7 @@
 
 Everyone is welcome to contribute, and we value everybody's contribution. Code is not the only way to help the community. Answering questions, helping others, reaching out, and improving the documentation are immensely valuable.
 
-Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md) and our [AI policy](./AI_POLICY.md).
+Whichever way you choose to contribute, please be mindful to respect our [code of conduct](./CODE_OF_CONDUCT.md).
 
 ## Ways to Contribute
 
@@ -14,7 +14,7 @@ You can contribute in many ways:
 - **Documentation:** Improve examples, guides, and docstrings.
 - **Feedback:** Submit tickets related to bugs or desired new features.
 
-If you are unsure where to start, join our [Discord Channel](https://discord.gg/q8Dzzpym3f).
+If you are unsure where to start, join our [Discord Channel](https://discord.gg/JkrYNdmw).
 
 ## Development Setup
 

MANIFEST.in

@@ -1,3 +1,2 @@
 include src/lerobot/templates/lerobot_modelcard_template.md
 include src/lerobot/datasets/card_template.md
-include src/lerobot/envs/metaworld_config.json

README.md

@@ -128,7 +128,6 @@ Learn how to implement your own simulation environment or benchmark and distribu
 ## Resources
 
 - **[Documentation](https://huggingface.co/docs/lerobot/index):** The complete guide to tutorials & API.
-- **[Chinese Tutorials: LeRobot+SO-ARM101中文教程-同济子豪兄](https://zihao-ai.feishu.cn/wiki/space/7589642043471924447)** Detailed doc for assembling, teleoperate, dataset, train, deploy. Verified by Seed Studio and 5 global hackathon players.
 - **[Discord](https://discord.gg/q8Dzzpym3f):** Join the `LeRobot` server to discuss with the community.
 - **[X](https://x.com/LeRobotHF):** Follow us on X to stay up-to-date with the latest developments.
 - **[Robot Learning Tutorial](https://huggingface.co/spaces/lerobot/robot-learning-tutorial):** A free, hands-on course to learn robot learning using LeRobot.

benchmarks/audio/run_microphone_benchmark.py

@@ -0,0 +1,219 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+from soundfile import read
+
+from lerobot.microphones.configs import MicrophoneConfig
+from lerobot.microphones.portaudio import PortAudioMicrophone, PortAudioMicrophoneConfig
+from lerobot.microphones.utils import (
+    async_microphones_start_recording,
+    async_microphones_stop_recording,
+    make_microphones_from_configs,
+)
+from lerobot.utils.robot_utils import (
+    precise_sleep,
+)
+
+
+def main(
+    microphones_configs: dict[str, MicrophoneConfig],
+    audio_chunks_number: int,
+    audio_chunks_duration: float,
+    repetitions: int,
+    multiprocessing: bool = False,
+):
+    recording_dir = Path("outputs/audio_benchmark")
+    recording_dir.mkdir(parents=True, exist_ok=True)
+
+    # Create microphones
+    microphones = make_microphones_from_configs(microphones_configs)
+
+    # Connect microphones
+    for microphone in microphones.values():
+        microphone.connect()
+
+    all_audio_chunks = []
+    for i in range(repetitions):
+        print(f"Repetition {i + 1}/{repetitions}...")
+
+        # Create audio chunks
+        audio_chunks = {}
+        for microphone_key in microphones:
+            audio_chunks.update({microphone_key: []})
+
+        # Start recording
+        async_microphones_start_recording(
+            microphones,
+            output_files=[
+                recording_dir / f"{microphone_key}_recording_{i}.wav" for microphone_key in microphones
+            ],
+            multiprocessing=multiprocessing,
+        )
+
+        # Record audio chunks
+        for j in range(audio_chunks_number):
+            precise_sleep(audio_chunks_duration)
+
+            for microphone_key, microphone in microphones.items():
+                audio_chunk = microphone.read()
+                print(f"{microphone_key} - repetition {i} - chunk {j} - samples {audio_chunk.shape[0]}")
+                audio_chunks[microphone_key].append(audio_chunk)
+
+        # Stop recording
+        async_microphones_stop_recording(microphones)
+
+        for microphone_key in microphones:
+            audio_chunks[microphone_key] = np.concatenate(audio_chunks[microphone_key], axis=0)
+
+        all_audio_chunks.append(audio_chunks)
+
+    # Disconnect microphones
+    for microphone in microphones.values():
+        microphone.disconnect()
+
+    # Compute statistics
+    cmap = plt.get_cmap("tab10")
+    _, ax = plt.subplots(nrows=repetitions, ncols=len(microphones))
+    chunk_length = np.zeros((repetitions, len(microphones)))
+    record_length = np.zeros((repetitions, len(microphones)))
+    for i in range(repetitions):
+        for j, (microphone_key, microphone) in enumerate(microphones.items()):
+            # Get recorded audio chunks
+            recorded_audio_chunks = all_audio_chunks[i][microphone_key]
+
+            # Load recorded file
+            recorded_data, _ = read(recording_dir / f"{microphone_key}_recording_{i}.wav")
+            if recorded_data.ndim == 1:
+                recorded_data = np.expand_dims(recorded_data, axis=1)
+
+            record_length[i, j] = recorded_data.shape[0]
+            chunk_length[i, j] = recorded_audio_chunks.shape[0]
+
+            for k, (chunk_data, record_data) in enumerate(
+                zip(recorded_audio_chunks.T, recorded_data.T, strict=False)
+            ):
+                # Plot audio chunks and recorded data
+                ax[i, j].plot(
+                    np.arange(0, len(chunk_data)) / microphone.sample_rate,
+                    chunk_data,
+                    label=f"audio chunks - channel {k}",
+                    color=cmap(2 * k),
+                )
+                ax[i, j].plot(
+                    np.arange(0, len(record_data)) / microphone.sample_rate,
+                    record_data,
+                    label=f"recorded data - channel {k}",
+                    linestyle="dashed",
+                    color=cmap(2 * k + 1),
+                )
+
+                # Plot absolute difference (errors should be located at the end of the recordings)
+                if recorded_data.shape[0] - recorded_audio_chunks.shape[0] > 0:
+                    chunk_data = np.append(
+                        chunk_data, np.zeros(int(recorded_data.shape[0] - recorded_audio_chunks.shape[0]))
+                    )
+                else:
+                    record_data = np.append(
+                        record_data, np.zeros(int(-recorded_data.shape[0] + recorded_audio_chunks.shape[0]))
+                    )
+                ax[i, j].plot(
+                    np.arange(0, len(record_data)) / microphone.sample_rate,
+                    np.abs(chunk_data - record_data),
+                    label=f"differences - channel {k}",
+                    color="red",
+                    linestyle="dotted",
+                )
+                ax[i, j].set_title(f"{microphone_key} - repetition {i}")
+            ax[i, j].legend()
+
+    plt.show()
+
+    # Print statistics
+    differences = record_length - chunk_length
+    for i, (microphone_key, microphone) in enumerate(microphones.items()):
+        print(
+            f"Average recorded duration for {microphone_key} : {np.mean(record_length[:, i]) / microphone.sample_rate:.3f} seconds"
+        )
+        print(
+            f"Average chunk duration for {microphone_key} : {np.mean(chunk_length[:, i]) / microphone.sample_rate:.3f} seconds"
+        )
+        print(f"Average difference for {microphone_key} : {np.mean(differences[:, i]):.3f} samples")
+        print(
+            f"Average difference for {microphone_key} : {np.mean(differences[:, i]) / microphone.sample_rate:.3f} seconds"
+        )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--microphones_indices",
+        type=int,
+        nargs="+",
+        default=[microphone["index"] for microphone in PortAudioMicrophone.find_microphones()],
+    )
+    parser.add_argument(
+        "--microphones_sample_rate",
+        type=float,
+        nargs="+",
+        default=[None] * len(PortAudioMicrophone.find_microphones()),
+    )
+    parser.add_argument(
+        "--microphones_channels",
+        type=int,
+        nargs="+",
+        default=[None] * len(PortAudioMicrophone.find_microphones()),
+    )
+    parser.add_argument("--audio_chunks_number", type=int, default=2)
+    parser.add_argument(
+        "--audio_chunks_duration",
+        type=float,
+        default=1.0,
+    )
+    parser.add_argument(
+        "--repetitions",
+        type=int,
+        default=2,
+    )
+    parser.add_argument(
+        "--multiprocessing",
+        action="store_true",
+    )
+
+    args = vars(parser.parse_args())
+
+    args["microphones_configs"] = {}
+    for index, sample_rate, channels in zip(
+        args["microphones_indices"],
+        args["microphones_sample_rate"],
+        args["microphones_channels"],
+        strict=False,
+    ):
+        microphone_config = PortAudioMicrophoneConfig(
+            microphone_index=index,
+            sample_rate=sample_rate,
+            channels=channels,
+        )
+        args["microphones_configs"].update({f"microphone_{index}": microphone_config})
+    args.pop("microphones_indices")
+    args.pop("microphones_sample_rate")
+    args.pop("microphones_channels")
+
+    main(**args)

benchmarks/audio/run_tactile_benchmark.py

@@ -0,0 +1,136 @@
+#!/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 argparse
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+
+from lerobot.microphones.anyskin import AnyskinSensorConfig
+from lerobot.microphones.configs import MicrophoneConfig
+from lerobot.microphones.utils import (
+    async_microphones_start_recording,
+    async_microphones_stop_recording,
+    make_microphones_from_configs,
+)
+from lerobot.utils.robot_utils import (
+    precise_sleep,
+)
+
+
+def main(
+    sensors_configs: dict[str, MicrophoneConfig],
+    multiprocessing: bool = False,
+):
+    recording_dir = Path("outputs/tactile_benchmark")
+    recording_dir.mkdir(parents=True, exist_ok=True)
+
+    # Create microphones
+    sensors = make_microphones_from_configs(sensors_configs)
+
+    # Connect microphones
+    for sensor in sensors.values():
+        sensor.connect()
+
+    # Create audio chunks
+    data_chunks = {}
+    for sensor_key in sensors:
+        data_chunks.update({sensor_key: []})
+
+    # Start recording
+    async_microphones_start_recording(
+        sensors,
+        output_files=[recording_dir / f"{sensor_key}_recording.wav" for sensor_key in sensors],
+        multiprocessing=multiprocessing,
+    )
+
+    # Record audio chunks
+    precise_sleep(10.0)
+
+    for sensor_key, sensor in sensors.items():
+        data_chunk = sensor.read()
+        print(f"{sensor_key} - samples {data_chunk.shape[0]}")
+        data_chunks[sensor_key].append(data_chunk)
+
+    # Stop recording
+    async_microphones_stop_recording(sensors)
+
+    for sensor_key in sensors:
+        data_chunks[sensor_key] = np.concatenate(data_chunks[sensor_key], axis=0)
+
+    # Disconnect microphones
+    for sensor in sensors.values():
+        sensor.disconnect()
+
+    for sensor_key in sensors:
+        data, sample_rate = sf.read(recording_dir / f"{sensor_key}_recording.wav")
+        print(f"{sensor_key} - samples {data.shape[0]}")
+        print(f"{sensor_key} - sample rate {sample_rate}")
+        print(f"{sensor_key} - data {data}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--sensors_ports",
+        type=str,
+        nargs="+",
+    )
+    parser.add_argument(
+        "--sensors_baud_rate",
+        type=int,
+        nargs="+",
+    )
+    parser.add_argument(
+        "--sensors_sample_rate",
+        type=int,
+        nargs="+",
+    )
+    parser.add_argument(
+        "--sensors_channels",
+        type=int,
+        nargs="+",
+    )
+    parser.add_argument(
+        "--multiprocessing",
+        action="store_true",
+    )
+
+    args = vars(parser.parse_args())
+
+    args["sensors_configs"] = {}
+    for port, baud_rate, sample_rate, channels in zip(
+        args["sensors_ports"],
+        args["sensors_baud_rate"],
+        args["sensors_sample_rate"],
+        args["sensors_channels"],
+        strict=False,
+    ):
+        channels = [1, 2, 3, 4, 5]
+        sensor_config = AnyskinSensorConfig(
+            sensor_port=port,
+            baud_rate=baud_rate,
+            sample_rate=sample_rate,
+            channels=channels,
+        )
+        args["sensors_configs"].update({f"sensor_{port}": sensor_config})
+    args.pop("sensors_ports")
+    args.pop("sensors_baud_rate")
+    args.pop("sensors_sample_rate")
+    args.pop("sensors_channels")
+
+    main(**args)

benchmarks/video/README.md

@@ -28,9 +28,9 @@ We don't expect the same optimal settings for a dataset of images from a simulat
 For these reasons, we run this benchmark on four representative datasets:
 
 - `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
-- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
-- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
-- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
 
 Note: The datasets used for this benchmark need to be image datasets, not video datasets.
 
@@ -179,7 +179,7 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
+        aliberts/aloha_mobile_shrimp_image \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 2 20 None \
@@ -203,9 +203,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/kitchen \
     --vcodec libx264 libx265 \
     --pix-fmt yuv444p yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -221,9 +221,9 @@ python benchmark/video/run_video_benchmark.py \
     --output-dir outputs/video_benchmark \
     --repo-ids \
         lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/kitchen \
     --vcodec libsvtav1 \
     --pix-fmt yuv420p \
     --g 1 2 3 4 5 6 10 15 20 40 None \
@@ -252,37 +252,37 @@ Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_read
 
 These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
 
-| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
-| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                   | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
+| ---------------------------------- | ---------- | ------- | --------- | --------- | --------- |
+|                                    | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
 
-| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
-| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                   | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
+| ---------------------------------- | ------- | ------- | -------- | ------- | --------- |
+|                                    | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
 
-|                                   |          | vcodec   | pix_fmt      |          |           |              |
-| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |
+|                                    |          | vcodec   | pix_fmt      |          |           |              |
+| ---------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
+|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

docker/Dockerfile.internal

@@ -85,8 +85,6 @@ RUN if [ "$UNBOUND_DEPS" = "true" ]; then \
 
 RUN uv pip install --no-cache ".[all]"
 
-RUN chmod +x /lerobot/.venv/lib/python${PYTHON_VERSION}/site-packages/triton/backends/nvidia/bin/ptxas
-
 # Copy the rest of the application source code
 # Make sure to have the git-LFS files for testing
 COPY --chown=user_lerobot:user_lerobot . .

docs/source/_toctree.yml

@@ -7,6 +7,8 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
+  - local: cameras
+    title: Cameras
   - local: bring_your_own_policies
     title: Bring Your Own Policies
   - local: integrate_hardware
@@ -19,8 +21,6 @@
     title: Multi GPU training
   - local: peft_training
     title: Training with PEFT (e.g., LoRA)
-  - local: rename_map
-    title: Using Rename Map and Empty Cameras
   title: "Tutorials"
 - sections:
   - local: lerobot-dataset-v3
@@ -29,10 +29,6 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     title: Using the Dataset Tools
-  - local: dataset_subtask
-    title: Using Subtasks in the Dataset
-  - local: streaming_video_encoding
-    title: Streaming Video Encoding
   title: "Datasets"
 - sections:
   - local: act
@@ -103,19 +99,11 @@
     title: Unitree G1
   - local: earthrover_mini_plus
     title: Earth Rover Mini
-  - local: omx
-    title: OMX
-  - local: openarm
-    title: OpenArm
   title: "Robots"
 - sections:
   - local: phone_teleop
     title: Phone
   title: "Teleoperators"
-- sections:
-  - local: cameras
-    title: Cameras
-  title: "Sensors"
 - sections:
   - local: torch_accelerators
     title: PyTorch accelerators
@@ -125,8 +113,6 @@
     title: Notebooks
   - local: feetech
     title: Updating Feetech Firmware
-  - local: damiao
-    title: Damiao Motors and CAN Bus
   title: "Resources"
 - sections:
   - local: contributing

docs/source/act.mdx

@@ -88,8 +88,5 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_act_your_dataset \
   --dataset.num_episodes=10 \
   --dataset.single_task="Your task description" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   --policy.path=${HF_USER}/act_policy
 ```

docs/source/async.mdx

@@ -195,7 +195,6 @@ client_cfg = RobotClientConfig(
     robot=robot_cfg,
     server_address="localhost:8080",
     policy_device="mps",
-    client_device="cpu",
     policy_type="smolvla",
     pretrained_name_or_path="<user>/smolvla_async",
     chunk_size_threshold=0.5,

docs/source/cameras.mdx

@@ -1,22 +1,12 @@
 # Cameras
 
-LeRobot offers multiple options for video capture:
+LeRobot offers multiple options for video capture, including phone cameras, built-in laptop cameras, external webcams, and Intel RealSense cameras. To efficiently record frames from most cameras, you can use either the `OpenCVCamera` or `RealSenseCamera` class. For additional compatibility details on the `OpenCVCamera` class, refer to the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
 
-| Class             | Supported Cameras                   |
-| ----------------- | ----------------------------------- |
-| `OpenCVCamera`    | Phone, built-in laptop, USB webcams |
-| `ZMQCamera`       | Network-connected cameras           |
-| `RealSenseCamera` | Intel RealSense (with depth)        |
-| `Reachy2Camera`   | Reachy 2 robot cameras              |
+### Finding your camera
 
-> [!TIP]
-> For `OpenCVCamera` compatibility details, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
+To instantiate a camera, you need a camera identifier. This identifier might change if you reboot your computer or re-plug your camera, a behavior mostly dependant on your operating system.
 
-### Find your camera
-
-Every camera requires a unique identifier to be instantiated, allowing you to distinguish between multiple connected devices.
-
-`OpenCVCamera` and `RealSenseCamera` support auto-discovery. Run the command below to list available devices and their identifiers. Note that these identifiers may change after rebooting your computer or re-plugging the camera, depending on your operating system.
+To find the camera indices of the cameras plugged into your system, run the following script:
 
 ```bash
 lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
@@ -24,7 +14,7 @@ lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
 
 The output will look something like this if you have two cameras connected:
 
-```bash
+```
 --- Detected Cameras ---
 Camera #0:
   Name: OpenCV Camera @ 0
@@ -43,37 +33,13 @@ Camera #0:
 > [!WARNING]
 > When using Intel RealSense cameras in `macOS`, you could get this [error](https://github.com/IntelRealSense/librealsense/issues/12307): `Error finding RealSense cameras: failed to set power state`, this can be solved by running the same command with `sudo` permissions. Note that using RealSense cameras in `macOS` is unstable.
 
-`ZMQCamera` and `Reachy2Camera` do not support auto-discovery. They must be configured manually by providing their network address and port or robot SDK settings.
+## Use Cameras
 
-## Use cameras
+Below are two examples, demonstrating how to work with the API.
 
-### Frame access modes
-
-All camera classes implement three access modes for capturing frames:
-
-| Method                    | Behavior                                                                                                                                                   | Blocks?        | Best For                                 |
-| ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------- | ---------------------------------------- |
-| `read()`                  | Waits for the camera hardware to return a frame. May block for a long time depending on the camera and SDK.                                                | Yes            | Simple scripts, sequential capture       |
-| `async_read(timeout_ms)`  | Returns the latest unconsumed frame from background thread. Blocks only if buffer is empty, up to `timeout_ms`. Raises `TimeoutError` if no frame arrives. | With a timeout | Control loops synchronized to camera FPS |
-| `read_latest(max_age_ms)` | Peeks at the most recent frame in buffer (may be stale). Raises `TimeoutError` if frame is older than `max_age_ms`.                                        | No             | UI visualization, logging, monitoring    |
-
-### Usage examples
-
-The following examples show how to use the camera API to configure and capture frames from different camera types.
-
-- **Blocking and non-blocking frame capture** using an OpenCV-based camera
+- **Asynchronous frame capture** using an OpenCV-based camera
 - **Color and depth capture** using an Intel RealSense camera
 
-> [!WARNING]
-> Failing to cleanly disconnect cameras can cause resource leaks. Use the context manager protocol to ensure automatic cleanup:
->
-> ```python
-> with OpenCVCamera(config) as camera:
->     ...
-> ```
->
-> You can also call `connect()` and `disconnect()` manually, but always use a `finally` block for the latter.
-
 <hfoptions id="shell_restart">
 <hfoption id="Open CV Camera">
 
@@ -94,30 +60,16 @@ config = OpenCVCameraConfig(
 )
 
 # Instantiate and connect an `OpenCVCamera`, performing a warm-up read (default).
-with OpenCVCamera(config) as camera:
-
-    # Read a frame synchronously — blocks until hardware delivers a new frame
-    frame = camera.read()
-    print(f"read() call returned frame with shape:", frame.shape)
-
-    # Read a frame asynchronously with a timeout — returns the latest unconsumed frame or waits up to timeout_ms for a new one
-    try:
-        for i in range(10):
-            frame = camera.async_read(timeout_ms=200)
-            print(f"async_read call returned frame {i} with shape:", frame.shape)
-    except TimeoutError as e:
-        print(f"No frame received within timeout: {e}")
-
-    # Instantly return a frame - returns the most recent frame captured by the camera
-    try:
-        initial_frame = camera.read_latest(max_age_ms=1000)
-        for i in range(10):
-            frame = camera.read_latest(max_age_ms=1000)
-            print(f"read_latest call returned frame {i} with shape:", frame.shape)
-            print(f"Was a new frame received by the camera? {not (initial_frame == frame).any()}")
-    except TimeoutError as e:
-        print(f"Frame too old: {e}")
+camera = OpenCVCamera(config)
+camera.connect()
 
+# Read frames asynchronously in a loop via `async_read(timeout_ms)`
+try:
+    for i in range(10):
+        frame = camera.async_read(timeout_ms=200)
+        print(f"Async frame {i} shape:", frame.shape)
+finally:
+    camera.disconnect()
 ```
 <!-- prettier-ignore-end -->
 
@@ -159,10 +111,10 @@ finally:
 </hfoption>
 </hfoptions>
 
-## Use your phone's camera
+## Use your phone
 
 <hfoptions id="use phone">
-<hfoption id="iPhone & macOS">
+<hfoption id="Mac">
 
 To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
 
@@ -172,49 +124,83 @@ To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
 
 For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).
 
+Your iPhone should be detected automatically when running the camera setup script in the next section.
+
 </hfoption>
-<hfoption id="OBS virtual camera">
+<hfoption id="Linux">
 
-If you want to use your phone as a camera using OBS, follow these steps to set up a virtual camera.
+If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
 
-1. _(Linux only) Install `v4l2loopback-dkms` and `v4l-utils`_. These packages create virtual camera devices and verify their settings. Install with:
+1. _Install `v4l2loopback-dkms` and `v4l-utils`_. Those packages are required to create virtual camera devices (`v4l2loopback`) and verify their settings with the `v4l2-ctl` utility from `v4l-utils`. Install them using:
 
-```bash
+<!-- prettier-ignore-start -->
+```python
 sudo apt install v4l2loopback-dkms v4l-utils
 ```
+<!-- prettier-ignore-end -->
 
-2. _Install the [DroidCam app](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
-3. _Download and install [OBS Studio](https://obsproject.com)_.
-4. _Download and install the [DroidCam OBS plugin](https://droidcam.app/obs)_.
-5. _Start OBS Studio_.
+2. _Install [DroidCam](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
+3. _Install [OBS Studio](https://obsproject.com)_. This software will help you manage the camera feed. Install it using [Flatpak](https://flatpak.org):
 
-6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480` to avoid the watermarks.
-7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video` or `OBS > Preferences... > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it.
+<!-- prettier-ignore-start -->
+```python
+flatpak install flathub com.obsproject.Studio
+```
+<!-- prettier-ignore-end -->
+
+4. _Install the DroidCam OBS plugin_. This plugin integrates DroidCam with OBS Studio. Install it with:
+
+<!-- prettier-ignore-start -->
+```python
+flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
+```
+<!-- prettier-ignore-end -->
+
+5. _Start OBS Studio_. Launch with:
+
+<!-- prettier-ignore-start -->
+```python
+flatpak run com.obsproject.Studio
+```
+<!-- prettier-ignore-end -->
+
+6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480`.
+7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it in.
 8. _Start virtual camera_. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
-9. _Verify the virtual camera setup and resolution_.
-   - **Linux**: Use `v4l2-ctl` to list devices and check resolution:
-     ```bash
-     v4l2-ctl --list-devices  # find VirtualCam and note its /dev/videoX path
-     v4l2-ctl -d /dev/videoX --get-fmt-video  # replace with your VirtualCam path
-     ```
-     You should see `VirtualCam` listed and resolution `640x480`.
-   - **macOS**: Open Photo Booth or FaceTime and select "OBS Virtual Camera" as the input.
-   - **Windows**: The native Camera app doesn't support virtual cameras. Use a video conferencing app (Zoom, Teams) or run `lerobot-find-cameras opencv` directly to verify.
+9. _Verify the virtual camera setup_. Use `v4l2-ctl` to list the devices:
 
-<details>
-<summary><strong>Troubleshooting</strong></summary>
+<!-- prettier-ignore-start -->
+```python
+v4l2-ctl --list-devices
+```
+<!-- prettier-ignore-end -->
 
-> The virtual camera resolution is incorrect.
+You should see an entry like:
 
-Delete the virtual camera source and recreate it. The resolution cannot be changed after creation.
+```
+VirtualCam (platform:v4l2loopback-000):
+/dev/video1
+```
 
-> Error reading frame in background thread for OpenCVCamera(X): OpenCVCamera(X) frame width=640 or height=480 do not match configured width=1920 or height=1080.
+10. _Check the camera resolution_. Use `v4l2-ctl` to ensure that the virtual camera output resolution is `640x480`. Change `/dev/video1` to the port of your virtual camera from the output of `v4l2-ctl --list-devices`.
 
-This error is caused by OBS Virtual Camera advertising a `1920x1080` resolution despite rescaling. The only fix for now is to comment out the width and height check in `_postprocess_image()`.
+<!-- prettier-ignore-start -->
+```python
+v4l2-ctl -d /dev/video1 --get-fmt-video
+```
+<!-- prettier-ignore-end -->
 
-</details>
+You should see an entry like:
+
+```
+>>> Format Video Capture:
+>>>	Width/Height      : 640/480
+>>>	Pixel Format      : 'YUYV' (YUYV 4:2:2)
+```
+
+Troubleshooting: If the resolution is not correct you will have to delete the Virtual Camera port and try again as it cannot be changed.
+
+If everything is set up correctly, you can proceed with the rest of the tutorial.
 
 </hfoption>
 </hfoptions>
-
-If everything is set up correctly, your phone will appear as a standard OpenCV camera and can be used with `OpenCVCamera`.

docs/source/damiao.mdx

@@ -1,165 +0,0 @@
-# Damiao Motors and CAN Bus
-
-This guide covers setup and usage of Damiao motors with LeRobot via CAN bus communication.
-
-Currently, only Linux is supported, as the OpenArms CAN adapter only has drivers for Linux.
-
-## Linux CAN Setup
-
-Before using Damiao motors, you need to set up the CAN interface on your Linux system.
-
-### Install CAN Utilities
-
-```bash
-sudo apt-get install can-utils
-```
-
-### Configure CAN Interface (Manual)
-
-For standard CAN FD (recommended for OpenArms):
-
-```bash
-sudo ip link set can0 down
-sudo ip link set can0 type can bitrate 1000000 dbitrate 5000000 fd on
-sudo ip link set can0 up
-```
-
-For standard CAN (without FD):
-
-```bash
-sudo ip link set can0 down
-sudo ip link set can0 type can bitrate 1000000
-sudo ip link set can0 up
-```
-
-### Configure CAN Interface (Using LeRobot)
-
-LeRobot provides a utility script to setup and test CAN interfaces:
-
-```bash
-# Setup multiple interfaces (e.g., OpenArms Followers with 2 CAN buses)
-lerobot-setup-can --mode=setup --interfaces=can0,can1
-```
-
-## Debugging CAN Communication
-
-Use the built-in debug tools to test motor communication:
-
-```bash
-# Test motors on all interfaces
-lerobot-setup-can --mode=test --interfaces=can0,can1
-
-# Run speed/latency test
-lerobot-setup-can --mode=speed --interfaces=can0
-```
-
-The test mode will scan for motors (IDs 0x01-0x08) and report which ones respond. Example output:
-
-```
-can0: UP (CAN FD)
-  Motor 0x01 (joint_1): ✓ FOUND
-    → Response 0x11 [FD]: 00112233...
-  Motor 0x02 (joint_2): ✓ FOUND
-  Motor 0x03 (joint_3): ✗ No response
-  ...
-  Summary: 2/8 motors found
-```
-
-## Usage
-
-### Basic Setup
-
-```python
-from lerobot.motors import Motor
-from lerobot.motors.damiao import DamiaoMotorsBus
-
-# Define your motors with send/receive CAN IDs
-motors = {
-    "joint_1": Motor(id=0x01, motor_type_str="dm8009", recv_id=0x11),
-    "joint_2": Motor(id=0x02, motor_type_str="dm4340", recv_id=0x12),
-    "joint_3": Motor(id=0x03, motor_type_str="dm4310", recv_id=0x13),
-}
-
-# Create the bus
-bus = DamiaoMotorsBus(
-    port="can0",  # Linux socketcan interface
-    motors=motors,
-)
-
-# Connect
-bus.connect()
-```
-
-### Reading Motor States
-
-```python
-# Read single motor position (degrees)
-position = bus.read("Present_Position", "joint_1")
-
-# Read from multiple motors
-positions = bus.sync_read("Present_Position")  # All motors
-positions = bus.sync_read("Present_Position", ["joint_1", "joint_2"])
-
-# Read all states at once (position, velocity, torque)
-states = bus.sync_read_all_states()
-# Returns: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
-```
-
-### Writing Motor Commands
-
-```python
-# Enable torque
-bus.enable_torque()
-
-# Set goal position (degrees)
-bus.write("Goal_Position", "joint_1", 45.0)
-
-# Set positions for multiple motors
-bus.sync_write("Goal_Position", {
-    "joint_1": 45.0,
-    "joint_2": -30.0,
-    "joint_3": 90.0,
-})
-
-# Disable torque
-bus.disable_torque()
-```
-
-## Configuration Options
-
-| Parameter      | Default   | Description                                                 |
-| -------------- | --------- | ----------------------------------------------------------- |
-| `port`         | -         | CAN interface (`can0`) or serial port (`/dev/cu.usbmodem*`) |
-| `use_can_fd`   | `True`    | Enable CAN FD for higher data rates                         |
-| `bitrate`      | `1000000` | Nominal bitrate (1 Mbps)                                    |
-| `data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)                                |
-
-## Motor Configuration
-
-Each motor requires:
-
-- `id`: CAN ID for sending commands
-- `motor_type`: One of the supported motor types (e.g., `"dm8009"`, `"dm4340"`)
-- `recv_id`: CAN ID for receiving responses
-
-OpenArms default IDs follow the pattern: send ID `0x0N`, receive ID `0x1N` where N is the joint number.
-
-## Troubleshooting
-
-### No Response from Motors
-
-1. **Check power**
-2. **Verify CAN wiring**: Check CAN-H, CAN-L, and GND connections
-3. **Check motor IDs**: Use Damiao Debugging Tools to verify/configure IDs
-4. **Test CAN interface**: Run `candump can0` to see if messages are being received
-5. **Run diagnostics**: `lerobot-setup-can --mode=test --interfaces=can0`
-
-### Motor Timeout Parameter
-
-If motors were configured with timeout=0, they won't respond to commands. Use Damiao Debugging Tools to set a non-zero timeout value.
-
-### Verify CAN FD Status
-
-```bash
-ip -d link show can0 | grep fd
-```

docs/source/dataset_subtask.mdx

@@ -1,278 +0,0 @@
-# Using Subtasks in LeRobot Datasets
-
-Subtask support in robotics datasets has proven effective in improving robot reasoning and understanding. Subtasks are particularly useful for:
-
-- **Hierarchical policies**: Building policies that include subtask predictions to visualize robot reasoning in real time
-- **Reward modeling**: Helping reward models understand task progression (e.g., SARM-style stage-aware reward models)
-- **Task decomposition**: Breaking down complex manipulation tasks into atomic, interpretable steps
-
-LeRobotDataset now supports subtasks as part of its dataset structure, alongside tasks.
-
-## What are Subtasks?
-
-While a **task** describes the overall goal (e.g., "Pick up the apple and place it in the basket"), **subtasks** break down the execution into finer-grained steps:
-
-1. "Approach the apple"
-2. "Grasp the apple"
-3. "Lift the apple"
-4. "Move to basket"
-5. "Release the apple"
-
-Each frame in the dataset can be annotated with its corresponding subtask, enabling models to learn and predict these intermediate stages.
-
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/subtask-asset.png"
-  alt="An overview of subtask annotation showing how frames are labeled with intermediate subtask stages"
-  width="80%"
-/>
-
-<p>
-  <em>Figure: Overview of subtask annotation.</em>
-</p>
-
-**Reference:** _Subtask-learning based for robot self-assembly in flexible collaborative assembly in manufacturing_, Original Article, Published: 19 April 2022.
-
-## Dataset Structure
-
-Subtask information is stored in the dataset metadata:
-
-```
-my-dataset/
-├── data/
-│   └── ...
-├── meta/
-│   ├── info.json
-│   ├── stats.json
-│   ├── tasks.parquet
-│   ├── subtasks.parquet      # Subtask index → subtask string mapping
-│   └── episodes/
-│       └── ...
-└── videos/
-    └── ...
-```
-
-### Subtasks Parquet File
-
-The `meta/subtasks.parquet` file maps subtask indices to their natural language descriptions:
-
-| subtask_index | subtask (index column) |
-| ------------- | ---------------------- |
-| 0             | "Approach the apple"   |
-| 1             | "Grasp the apple"      |
-| 2             | "Lift the apple"       |
-| ...           | ...                    |
-
-### Frame-Level Annotations
-
-Each frame in the dataset can include a `subtask_index` field that references the subtasks parquet file:
-
-```python
-# Example frame data in the parquet file
-{
-    "index": 42,
-    "timestamp": 1.4,
-    "episode_index": 0,
-    "task_index": 0,
-    "subtask_index": 2,  # References "Lift the apple"
-    "observation.state": [...],
-    "action": [...],
-}
-```
-
-## Annotating Datasets with Subtasks
-
-We provide a HuggingFace Space for easily annotating any LeRobotDataset with subtasks:
-
-**[https://huggingface.co/spaces/lerobot/annotate](https://huggingface.co/spaces/lerobot/annotate)**
-
-After completing your annotation:
-
-1. Click "Push to Hub" to upload your annotated dataset
-2. You can also run the annotation space locally by following the instructions at [github.com/huggingface/lerobot-annotate](https://github.com/huggingface/lerobot-annotate)
-
-## Loading Datasets with Subtasks
-
-When you load a dataset with subtask annotations, the subtask information is automatically available:
-
-```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-# Load a dataset with subtask annotations
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-# Access a sample
-sample = dataset[100]
-
-# The sample includes both task and subtask information
-print(sample["task"])        # "Collect the fruit"
-print(sample["subtask"])     # "Grasp the apple"
-print(sample["task_index"])  # tensor(0)
-print(sample["subtask_index"])  # tensor(2)
-```
-
-### Checking for Subtask Support
-
-You can check if a dataset has subtask annotations:
-
-```python
-# Check if subtasks are available
-has_subtasks = (
-    "subtask_index" in dataset.features
-    and dataset.meta.subtasks is not None
-)
-
-if has_subtasks:
-    print(f"Dataset has {len(dataset.meta.subtasks)} unique subtasks")
-    print("Subtasks:", list(dataset.meta.subtasks.index))
-```
-
-## Using Subtasks for Training
-
-### With the Tokenizer Processor
-
-The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
-
-```python
-from lerobot.processor.tokenizer_processor import TokenizerProcessor
-from lerobot.processor.pipeline import ProcessorPipeline
-
-# Create a tokenizer processor
-tokenizer_processor = TokenizerProcessor(
-    tokenizer_name_or_path="google/paligemma-3b-pt-224",
-    padding="max_length",
-    max_length=64,
-)
-
-# The processor will automatically tokenize subtasks if present in the batch
-# and add them to the observation under:
-# - "observation.subtask.tokens"
-# - "observation.subtask.attention_mask"
-```
-
-When subtasks are available in the batch, the tokenizer processor adds:
-
-- `observation.subtask.tokens`: Tokenized subtask text
-- `observation.subtask.attention_mask`: Attention mask for the subtask tokens
-
-### DataLoader with Subtasks
-
-```python
-import torch
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    batch_size=16,
-    shuffle=True,
-)
-
-for batch in dataloader:
-    # Access subtask information in the batch
-    subtasks = batch["subtask"]  # List of subtask strings
-    subtask_indices = batch["subtask_index"]  # Tensor of subtask indices
-
-    # Use for training hierarchical policies or reward models
-    print(f"Batch subtasks: {set(subtasks)}")
-```
-
-## Example Datasets with Subtask Annotations
-
-Try loading a dataset with subtask annotations:
-
-```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-# Example dataset with subtask annotations
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-# Explore the subtasks
-print("Available subtasks:")
-for subtask_name in dataset.meta.subtasks.index:
-    print(f"  - {subtask_name}")
-
-# Get subtask distribution
-subtask_counts = {}
-for i in range(len(dataset)):
-    sample = dataset[i]
-    subtask = sample["subtask"]
-    subtask_counts[subtask] = subtask_counts.get(subtask, 0) + 1
-
-print("\nSubtask distribution:")
-for subtask, count in sorted(subtask_counts.items(), key=lambda x: -x[1]):
-    print(f"  {subtask}: {count} frames")
-```
-
-## Use Cases
-
-### 1. Hierarchical Policy Training
-
-Train policies that predict both actions and current subtask:
-
-```python
-class HierarchicalPolicy(nn.Module):
-    def __init__(self, num_subtasks):
-        super().__init__()
-        self.action_head = nn.Linear(hidden_dim, action_dim)
-        self.subtask_head = nn.Linear(hidden_dim, num_subtasks)
-
-    def forward(self, observations):
-        features = self.encoder(observations)
-        actions = self.action_head(features)
-        subtask_logits = self.subtask_head(features)
-        return actions, subtask_logits
-```
-
-### 2. Stage-Aware Reward Modeling (SARM)
-
-Build reward models that understand task progression:
-
-```python
-# SARM predicts:
-# - Stage: Which subtask is being executed (discrete)
-# - Progress: How far along the subtask (continuous 0-1)
-
-class SARMRewardModel(nn.Module):
-    def forward(self, observations):
-        features = self.encoder(observations)
-        stage_logits = self.stage_classifier(features)
-        progress = self.progress_regressor(features)
-        return stage_logits, progress
-```
-
-### 3. Progress Visualization
-
-Monitor robot execution by tracking subtask progression:
-
-```python
-def visualize_execution(model, observations):
-    for t, obs in enumerate(observations):
-        action, subtask_logits = model(obs)
-        predicted_subtask = subtask_names[subtask_logits.argmax()]
-        print(f"t={t}: Executing '{predicted_subtask}'")
-```
-
-## API Reference
-
-### LeRobotDataset Properties
-
-| Property                    | Type                   | Description                                |
-| --------------------------- | ---------------------- | ------------------------------------------ |
-| `meta.subtasks`             | `pd.DataFrame \| None` | DataFrame mapping subtask names to indices |
-| `features["subtask_index"]` | `dict`                 | Feature spec for subtask_index if present  |
-
-### Sample Keys
-
-When subtasks are available, each sample includes:
-
-| Key             | Type           | Description                          |
-| --------------- | -------------- | ------------------------------------ |
-| `subtask_index` | `torch.Tensor` | Integer index of the current subtask |
-| `subtask`       | `str`          | Natural language subtask description |
-
-## Related Resources
-
-- [SARM Paper](https://arxiv.org/pdf/2509.25358) - Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
-- [LeRobot Annotate Space](https://huggingface.co/spaces/lerobot/annotate) - Interactive annotation tool
-- [LeRobotDataset v3.0](./lerobot-dataset-v3) - Dataset format documentation

docs/source/earthrover_mini_plus.mdx

@@ -1,11 +1,5 @@
 # EarthRover Mini Plus
 
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Earth_Rover_Mini_5_240c9adc-4f9e-44b7-982f-5d1dc24af1d8.png.webp"
-  alt="EarthRover Mini Plus"
-  width="70%"
-/>
-
 The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.
 
 ## What You Need
@@ -185,16 +179,13 @@ echo $HF_USER
 Use the standard recording command:
 
 ```bash
-lerobot-record \
+python src/lerobot/scripts/lerobot_record.py \
     --robot.type=earthrover_mini_plus \
     --teleop.type=keyboard_rover \
     --dataset.repo_id=your_username/dataset_name \
     --dataset.num_episodes=2 \
     --dataset.fps=10 \
     --dataset.single_task="Navigate around obstacles" \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/groot.mdx

@@ -120,12 +120,9 @@ lerobot-record \
   --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" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
-  --policy.path=<user>/groot-bimanual \ # your trained model
-  --dataset.episode_time_s=30 \
+  --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
 ```
 

docs/source/hope_jr.mdx

@@ -224,15 +224,12 @@ lerobot-record \
     --teleop.port=/dev/tty.usbmodem1201 \
     --teleop.id=right \
     --teleop.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+    --dataset.repo_id=nepyope/hand_record_test_with_video_data \
     --dataset.single_task="Hand recording test with video data" \
     --dataset.num_episodes=1 \
     --dataset.episode_time_s=5 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -244,7 +241,7 @@ lerobot-replay \
     --robot.port=/dev/tty.usbmodem58760432281 \
     --robot.id=right \
     --robot.side=right \
-    --dataset.repo_id=<USER>/hand_record_test_with_camera \
+    --dataset.repo_id=nepyope/hand_record_test_with_camera \
     --dataset.episode=0
 ```
 
@@ -252,13 +249,13 @@ lerobot-replay \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/hand_record_test_with_video_data \
+  --dataset.repo_id=nepyope/hand_record_test_with_video_data \
   --policy.type=act \
   --output_dir=outputs/train/hopejr_hand \
   --job_name=hopejr \
   --policy.device=mps \
   --wandb.enable=true \
-  --policy.repo_id=<USER>/hand_test_policy
+  --policy.repo_id=nepyope/hand_test_policy
 ```
 
 ### Evaluate
@@ -273,11 +270,8 @@ lerobot-record \
   --robot.side=right \
   --robot.cameras='{"main": {"type": "opencv", "index_or_path": 0, "width": 640, "height": 480, "fps": 30}}' \
   --display_data=false \
-  --dataset.repo_id=<USER>/eval_hopejr \
+  --dataset.repo_id=nepyope/eval_hopejr \
   --dataset.single_task="Evaluate hopejr hand policy" \
   --dataset.num_episodes=10 \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

docs/source/il_robots.mdx

@@ -165,7 +165,7 @@ huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 Then store your Hugging Face repository name in a variable:
 
 ```bash
-HF_USER=$(hf auth whoami | awk -F': *' 'NR==1 {print $2}')
+HF_USER=$(hf auth whoami | head -n 1)
 echo $HF_USER
 ```
 
@@ -185,10 +185,7 @@ lerobot-record \
     --display_data=true \
     --dataset.repo_id=${HF_USER}/record-test \
     --dataset.num_episodes=5 \
-    --dataset.single_task="Grab the black cube" \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
+    --dataset.single_task="Grab the black cube"
 ```
 </hfoption>
 <hfoption id="API example">
@@ -518,9 +515,6 @@ lerobot-record  \
   --display_data=false \
   --dataset.repo_id=${HF_USER}/eval_so100 \
   --dataset.single_task="Put lego brick into the transparent box" \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/installation.mdx

@@ -1,15 +1,13 @@
 # Installation
 
-This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.10 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).
-
-## Step 1: Install [`miniforge`](https://conda-forge.org/download/)
+## 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
 ```
 
-## Step 2: Environment Setup
+## Environment Setup
 
 Create a virtual environment with Python 3.10, using conda:
 
@@ -40,14 +38,7 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
-> [!NOTE]
-> When installing LeRobot inside WSL (Windows Subsystem for Linux), make sure to install `evdev` with the following command:
->
-> ```bash
-> conda install evdev -c conda-forge
-> ```
-
-## Step 3: Install LeRobot 🤗
+## Install LeRobot 🤗
 
 ### From Source
 

docs/source/lekiwi.mdx

@@ -1,11 +1,5 @@
 # LeKiwi
 
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/1740517739083.jpeg"
-  alt="LeKiwi"
-  width="70%"
-/>
-
 In the steps below, we explain how to assemble the LeKiwi mobile robot.
 
 ## Source the parts

docs/source/lerobot-dataset-v3.mdx

@@ -41,10 +41,7 @@ lerobot-record \
   --display_data=true \
   --dataset.repo_id=${HF_USER}/record-test \
   --dataset.num_episodes=5 \
-  --dataset.single_task="Grab the black cube" \
-  --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
-  --dataset.encoder_threads=2
+  --dataset.single_task="Grab the black cube"
 ```
 
 See the [recording guide](./il_robots#record-a-dataset) for more details.

docs/source/libero.mdx

@@ -42,7 +42,6 @@ lerobot-eval \
 ```
 
 - `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
-- `--env.task_ids` picks task ids to run (`[0]`, `[1,2,3]`, etc.). Omit this flag (or set it to `null`) to run all tasks in the suite.
 - `--eval.batch_size` controls how many environments run in parallel.
 - `--eval.n_episodes` sets how many episodes to run in total.
 

docs/source/omx.mdx

@@ -1,197 +0,0 @@
-## Order and Assemble the parts
-
-First, assemble the OMX hardware following the official assembly guide.
-
-OMX Assembly Guide: https://ai.robotis.com/omx/assembly_guide_omx.html
-
-OMX robots are shipped preconfigured from the factory. Motor IDs, communication parameters, and joint offsets are already set, so no additional motor setup or calibration is required before using LeRobot.
-
-## Install LeRobot 🤗
-
-To install LeRobot, follow our [Installation Guide](./installation)
-
-In addition to these instructions, you need to install the Dynamixel SDK:
-
-```bash
-pip install -e ".[dynamixel]"
-```
-
-## Connect the robot
-
-To find the port for each bus servo adapter, run this script:
-
-```bash
-lerobot-find-port
-```
-
-This command runs and when prompted, disconnect the USB cable from either the leader or follower arm and press Enter. The output will show 'The port of this MotorsBus is [port]'. This identifies the port for the disconnected arm. Repeat for the other arm to identify both ports.
-
-<hfoptions id="find_port">
-<hfoption id="Mac">
-
-Example output on macOS:
-
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the USB cable from your MotorsBus and press Enter when done.
-
-[...Disconnect corresponding leader or follower arm and press Enter...]
-
-The port of this MotorsBus is /dev/tty.usbmodem575E0032081
-Reconnect the USB cable.
-```
-
-Where the found port is: `/dev/tty.usbmodem575E0032081` corresponding to your leader or follower arm.
-
-</hfoption>
-<hfoption id="Linux">
-
-On Linux, we strongly recommend using udev rules to assign persistent and human-readable device names to the OMX leader and follower arms. This avoids issues where device names such as ttyACM0 and ttyACM1 change when the robot is unplugged, replugged, or when the system is rebooted.
-
-#### 1. Find your device serial numbers
-
-You should have obtained the port numbers like ../../ttyACM? for the leader and follower using `lerobot-find-port`. You can match those results with the serial numbers using the `ls -l /dev/serial/by-id/` command.
-To create udev rules, you need the unique serial number for each OMX device. The easiest way is to list devices under:
-
-```bash
-ls -l /dev/serial/by-id/
-```
-
-You will see output similar to:
-
-```bash
-usb-ROBOTIS_OpenRB-150_228BDD7B503059384C2E3120FF0A2B19-if00 -> ../../ttyACM0
-usb-ROBOTIS_OpenRB-150_67E1ED68503059384C2E3120FF092234-if00 -> ../../ttyACM1
-```
-
-In each line, the serial number is the long string after `usb-ROBOTIS_OpenRB-150_` and before `-if00`.
-
-Follower serial: `228BDD7B503059384C2E3120FF0A2B19`
-
-Leader serial: `67E1ED68503059384C2E3120FF092234`
-
-#### 2. Create the udev rule
-
-Create a new udev rule file:
-
-```bash
-sudo nano /etc/udev/rules.d/99-omx.rules
-```
-
-Paste the following lines, replacing the serial numbers with the values you found above:
-
-```bash
-SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="228BDD7B503059384C2E3120FF0A2B19", SYMLINK+="omx_follower"
-SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{serial}=="67E1ED68503059384C2E3120FF092234", SYMLINK+="omx_leader"
-```
-
-Save the file and reload udev rules:
-
-```bash
-sudo udevadm control --reload-rules
-sudo udevadm trigger
-```
-
-Now unplug and replug both devices once.
-
-#### 3. Verify the symlinks
-
-Check that the persistent device names exist:
-
-```bash
-ls -l /dev/omx_follower /dev/omx_leader
-```
-
-You should see them pointing to ttyACM\* devices:
-
-```bash
-/dev/omx_follower -> ttyACM*
-/dev/omx_leader   -> ttyACM*
-```
-
-These names remain stable across reboots and reconnections.
-
-</hfoption>
-</hfoptions>
-
-## Teleoperate
-
-After identifying the correct ports, you can directly teleoperate the follower arm using the leader arm.
-
-<hfoptions id="teleoperate">
-<hfoption id="Mac">
-
-### Teleoperate without camera
-
-```bash
-lerobot-teleoperate \
-  --robot.type=omx_follower \
-  --robot.port=<your_follower_port> \
-  --robot.id=omx_follower_arm \
-  --teleop.type=omx_leader \
-  --teleop.port=<your_leader_port> \
-  --teleop.id=omx_leader_arm
-```
-
-During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
-
-### Teleoperate with camera
-
-You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
-
-```bash
-lerobot-teleoperate \
-  --robot.type=omx_follower \
-  --robot.port=<your_follower_port> \
-  --robot.id=omx_follower_arm \
-  --robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
-  --teleop.type=omx_leader \
-  --teleop.port=<your_leader_port> \
-  --teleop.id=omx_leader_arm \
-  --display_data=true
-```
-
-When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
-
-</hfoption>
-<hfoption id="Linux">
-
-### Teleoperate without camera
-
-```bash
-lerobot-teleoperate \
-  --robot.type=omx_follower \
-  --robot.port=/dev/omx_follower \
-  --robot.id=omx_follower_arm \
-  --teleop.type=omx_leader \
-  --teleop.port=/dev/omx_leader \
-  --teleop.id=omx_leader_arm
-```
-
-During teleoperation, motions of the leader arm are mirrored in real time by the follower arm. OMX is already preconfigured, teleoperation can begin immediately without any calibration steps.
-
-### Teleoperate with camera
-
-You can also enable camera input during teleoperation by providing a camera configuration for the follower arm.
-
-```bash
-lerobot-teleoperate \
-  --robot.type=omx_follower \
-  --robot.port=/dev/omx_follower \
-  --robot.id=omx_follower_arm \
-  --robot.cameras="{front: {type: opencv, index_or_path: '/dev/video0', width: 640, height: 480, fps: 30}}" \
-  --teleop.type=omx_leader \
-  --teleop.port=/dev/omx_leader \
-  --teleop.id=omx_leader_arm \
-  --display_data=true
-```
-
-When the camera is enabled, the camera stream is displayed in real time and synchronized with the robot state. This setup is useful for visual monitoring and can be reused later for demonstration recording and imitation learning.
-
-</hfoption>
-</hfoptions>
-
-Congrats 🎉, your robot is all set to learn a task on its own.
-
-> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/robotis).

docs/source/openarm.mdx

@@ -1,276 +0,0 @@
-# OpenArm
-
-[OpenArm](https://openarm.dev) is an open-source 7DOF humanoid arm designed for physical AI research and deployment.
-
-To get your OpenArm, assembled or DIY, and join the global community, browse verified and certified manufacturers worldwide at [openarm.dev](https://openarm.dev).
-
-## What's Unique?
-
-- **Human-Scale Design**: OpenArm is designed with human-like proportions, scaled for a person around 160-165cm tall. This provides an optimal balance between practical reach and manageable inertia for safe, responsive operation.
-
-- **Safety-First Architecture**: Built with QDD backdrivable motors and high compliance, OpenArm prioritizes safe human-robot interaction while maintaining practical payload capabilities (6.0kg peak / 4.1kg nominal) for real-world tasks.
-
-- **Built for Durability**: Critical structural components use aluminum and stainless steel construction, ensuring robust performance for repetitive data collection and continuous research use.
-
-- **Fully Accessible & Buildable**: Every component, from CNC parts and 3D-printed casings to electrical wiring is designed to be purchasable and buildable by individual researchers and labs, with complete fabrication data provided.
-
-- **Practical & Affordable**: At $6,500 USD for a complete bimanual system, OpenArm delivers research-grade capabilities at a fraction of traditional humanoid robot costs.
-
-## Platform Requirements
-
-<Tip warning={true}>
-  **Linux Only**: OpenArm currently only works on Linux. The CAN bus USB adapter
-  does not have macOS drivers and has not been tested on Windows.
-</Tip>
-
-## Safety Guide
-
-Before operating OpenArm, please read the [official safety guide](https://docs.openarm.dev/getting-started/safety-guide). Key points:
-
-- **Secure installation**: Fasten the arm to a flat, stable surface with screws or clamps
-- **Safe distance**: Keep body parts and objects outside the range of motion during operation
-- **Protective equipment**: Always wear safety goggles; use additional PPE as needed
-- **Payload limits**: Do not exceed specified payload limits (6.0kg peak / 4.1kg nominal per arm)
-- **Emergency stop**: Know the location and operation of the emergency stop device
-- **Regular inspection**: Check for loose screws, damaged mechanical limits, unusual noises, and wiring damage
-
-## Hardware Setup
-
-Follow the official [OpenArm hardware documentation](https://docs.openarm.dev) for:
-
-- Bill of materials and sourcing
-- 3D printing instructions
-- Mechanical assembly
-- Electrical wiring
-
-The hardware repositories are available at [github.com/enactic/openarm](https://github.com/enactic/openarm).
-
-## CAN Bus Setup
-
-OpenArm uses CAN bus communication with Damiao motors. Once you have the CAN bus USB adapter plugged into your Linux PC, follow the [Damiao Motors and CAN Bus guide](./damiao) to configure the interface.
-
-Quick setup:
-
-```bash
-# Setup CAN interfaces
-lerobot-setup-can --mode=setup --interfaces=can0,can1
-
-# Test motor communication
-lerobot-setup-can --mode=test --interfaces=can0,can1
-```
-
-## Install LeRobot 🤗
-
-Follow our [Installation Guide](./installation), then install the Damiao motor support:
-
-```bash
-pip install -e ".[damiao]"
-```
-
-## Usage
-
-### Follower Arm (Robot)
-
-<hfoptions id="follower">
-<hfoption id="Command">
-
-```bash
-lerobot-calibrate \
-    --robot.type=openarm_follower \
-    --robot.port=can0 \
-    --robot.side=right \
-    --robot.id=my_openarm_follower
-```
-
-</hfoption>
-<hfoption id="API example">
-
-```python
-from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
-
-config = OpenArmFollowerConfig(
-    port="can0",
-    side="right",  # or "left" for left arm
-    id="my_openarm_follower",
-)
-
-follower = OpenArmFollower(config)
-follower.connect()
-
-# Read current state
-obs = follower.get_observation()
-print(obs)
-
-# Send action (position in degrees)
-action = {
-    "joint_1.pos": 0.0,
-    "joint_2.pos": 0.0,
-    "joint_3.pos": 0.0,
-    "joint_4.pos": 45.0,
-    "joint_5.pos": 0.0,
-    "joint_6.pos": 0.0,
-    "joint_7.pos": 0.0,
-    "gripper.pos": 0.0,
-}
-follower.send_action(action)
-
-follower.disconnect()
-```
-
-</hfoption>
-</hfoptions>
-
-### Leader Arm (Teleoperator)
-
-The leader arm is used for teleoperation - manually moving it to control the follower arm.
-
-<hfoptions id="leader">
-<hfoption id="Command">
-
-```bash
-lerobot-calibrate \
-    --teleop.type=openarm_leader \
-    --teleop.port=can1 \
-    --teleop.id=my_openarm_leader
-```
-
-</hfoption>
-<hfoption id="API example">
-
-```python
-from lerobot.teleoperators.openarm_leader import OpenArmLeader, OpenArmLeaderConfig
-
-config = OpenArmLeaderConfig(
-    port="can1",
-    id="my_openarm_leader",
-    manual_control=True,  # Disable torque for manual movement
-)
-
-leader = OpenArmLeader(config)
-leader.connect()
-
-# Read current position (as action to send to follower)
-action = leader.get_action()
-print(action)
-
-leader.disconnect()
-```
-
-</hfoption>
-</hfoptions>
-
-### Teleoperation
-
-To teleoperate OpenArm with leader-follower control:
-
-```bash
-lerobot-teleoperate \
-    --robot.type=openarm_follower \
-    --robot.port=can0 \
-    --robot.side=right \
-    --robot.id=my_follower \
-    --teleop.type=openarm_leader \
-    --teleop.port=can1 \
-    --teleop.id=my_leader
-```
-
-### Bimanual Teleoperation
-
-To teleoperate a bimanual OpenArm setup with two leader and two follower arms:
-
-```bash
-lerobot-teleoperate \
-    --robot.type=bi_openarm_follower \
-    --robot.left_arm_config.port=can0 \
-    --robot.left_arm_config.side=left \
-    --robot.right_arm_config.port=can1 \
-    --robot.right_arm_config.side=right \
-    --robot.id=my_bimanual_follower \
-    --teleop.type=bi_openarm_leader \
-    --teleop.left_arm_config.port=can2 \
-    --teleop.right_arm_config.port=can3 \
-    --teleop.id=my_bimanual_leader
-```
-
-### Recording Data
-
-To record a dataset during teleoperation:
-
-```bash
-lerobot-record \
-    --robot.type=openarm_follower \
-    --robot.port=can0 \
-    --robot.side=right \
-    --robot.id=my_follower \
-    --teleop.type=openarm_leader \
-    --teleop.port=can1 \
-    --teleop.id=my_leader \
-    --repo-id=my_hf_username/my_openarm_dataset \
-    --fps=30 \
-    --num-episodes=10
-```
-
-## Configuration Options
-
-### Follower Configuration
-
-| Parameter             | Default   | Description                                                |
-| --------------------- | --------- | ---------------------------------------------------------- |
-| `port`                | -         | CAN interface (e.g., `can0`)                               |
-| `side`                | `None`    | Arm side: `"left"`, `"right"`, or `None` for custom limits |
-| `use_can_fd`          | `True`    | Enable CAN FD for higher data rates                        |
-| `can_bitrate`         | `1000000` | Nominal bitrate (1 Mbps)                                   |
-| `can_data_bitrate`    | `5000000` | CAN FD data bitrate (5 Mbps)                               |
-| `max_relative_target` | `None`    | Safety limit for relative target positions                 |
-| `position_kp`         | Per-joint | Position control proportional gains                        |
-| `position_kd`         | Per-joint | Position control derivative gains                          |
-
-### Leader Configuration
-
-| Parameter          | Default   | Description                         |
-| ------------------ | --------- | ----------------------------------- |
-| `port`             | -         | CAN interface (e.g., `can1`)        |
-| `manual_control`   | `True`    | Disable torque for manual movement  |
-| `use_can_fd`       | `True`    | Enable CAN FD for higher data rates |
-| `can_bitrate`      | `1000000` | Nominal bitrate (1 Mbps)            |
-| `can_data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)        |
-
-## Motor Configuration
-
-OpenArm uses Damiao motors with the following default configuration:
-
-| Joint                       | Motor Type | Send ID | Recv ID |
-| --------------------------- | ---------- | ------- | ------- |
-| joint_1 (Shoulder pan)      | DM8009     | 0x01    | 0x11    |
-| joint_2 (Shoulder lift)     | DM8009     | 0x02    | 0x12    |
-| joint_3 (Shoulder rotation) | DM4340     | 0x03    | 0x13    |
-| joint_4 (Elbow flex)        | DM4340     | 0x04    | 0x14    |
-| joint_5 (Wrist roll)        | DM4310     | 0x05    | 0x15    |
-| joint_6 (Wrist pitch)       | DM4310     | 0x06    | 0x16    |
-| joint_7 (Wrist rotation)    | DM4310     | 0x07    | 0x17    |
-| gripper                     | DM4310     | 0x08    | 0x18    |
-
-## Troubleshooting
-
-### No Response from Motors
-
-1. Check power supply connections
-2. Verify CAN wiring (CAN-H, CAN-L, GND)
-3. Run diagnostics: `lerobot-setup-can --mode=test --interfaces=can0`
-4. See the [Damiao troubleshooting guide](./damiao#troubleshooting) for more details
-
-### CAN Interface Not Found
-
-Ensure the CAN interface is configured:
-
-```bash
-ip link show can0
-```
-
-## Resources
-
-- [OpenArm Website](https://openarm.dev)
-- [OpenArm Documentation](https://docs.openarm.dev)
-- [OpenArm GitHub](https://github.com/enactic/openarm)
-- [Safety Guide](https://docs.openarm.dev/getting-started/safety-guide)
-- [Damiao Motors and CAN Bus](./damiao)

docs/source/phone_teleop.mdx

@@ -66,13 +66,12 @@ Run on of the examples scripts to teleoperate, record a dataset, replay a datase
 
 All scripts assume you configured your robot (e.g., SO-100 follower) and set the correct serial port.
 
-Additionally you need to **copy the URDF of the robot into the examples folder**. For the examples in this tutorial (using SO100/SO101), copy the `SO101` folder from the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101) into the `examples/phone_to_so100/` directory, so that the URDF file path becomes `examples/phone_to_so100/SO101/so101_new_calib.urdf`.
+Additionally you need to **copy the urdf of the robot to the examples folder**. For the examples in this tutorial (Using SO100/SO101) it is highly recommended to use the urdf in the [SO-ARM100 repo](https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf)
 
 - Run this example to teleoperate:
 
   ```bash
-  cd examples/phone_to_so100
-  python teleoperate.py
+  python examples/phone_to_so100/teleoperate.py
   ```
 
 After running the example:
@@ -85,22 +84,19 @@ Additionally you can customize mapping or safety limits by editing the processor
 - Run this example to record a dataset, which saves absolute end effector observations and actions:
 
   ```bash
-  cd examples/phone_to_so100
-  python record.py
+  python examples/phone_to_so100/record.py
   ```
 
 - Run this example to replay recorded episodes:
 
   ```bash
-  cd examples/phone_to_so100
-  python replay.py
+  python examples/phone_to_so100/replay.py
   ```
 
 - Run this example to evaluate a pretrained policy:
 
   ```bash
-  cd examples/phone_to_so100
-  python evaluate.py
+  python examples/phone_to_so100/evaluate.py
   ```
 
 ### Important pipeline steps and options

docs/source/pi0.mdx

@@ -60,7 +60,7 @@ policy.type=pi0
 For training π₀, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=pi0 \
     --output_dir=./outputs/pi0_training \

docs/source/pi05.mdx

@@ -56,7 +56,7 @@ policy.type=pi05
 Here's a complete training command for finetuning the base π₀.₅ model on your own dataset:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py\
     --dataset.repo_id=your_dataset \
     --policy.type=pi05 \
     --output_dir=./outputs/pi05_training \

docs/source/reachy2.mdx

@@ -159,9 +159,6 @@ lerobot-record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 
@@ -201,9 +198,6 @@ lerobot-record \
     --dataset.fps=15 \
     --dataset.push_to_hub=true \
     --dataset.private=true \
-    --dataset.streaming_encoding=true \
-    --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
     --display_data=true
 ```
 

docs/source/rename_map.mdx

@@ -1,145 +0,0 @@
-# Understanding the Rename Map and Empty Cameras
-
-When you train or evaluate a robot policy, your **dataset** or **environment** hands you observations under one set of keys (e.g. `observation.images.front`, `observation.images.eagle`), while your **policy** was built to expect another (e.g. `observation.images.image`, `observation.images.image2`). The rename map is how you bridge that gap without changing the policy or the data source.
-
-This guide explains why it exists, how to use it in training and evaluation, and when to use **empty cameras** so you can fine-tune multi-camera policies on datasets that have fewer views.
-
----
-
-## Why observation keys don’t always match
-
-Policies have a fixed set of **input feature names** (often coming from a pretrained config). For example:
-
-- **XVLA-base** expects three image keys: `observation.images.image`, `observation.images.image2`, `observation.images.image3`.
-- **pi0-fast-libero** might expect `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb`.
-
-Your dataset or sim might use completely different names: `observation.images.front`, `observation.images.eagle`, `observation.images.glove` (e.g. [svla_so100_sorting](https://huggingface.co/datasets/lerobot/svla_so100_sorting)). Or your eval env (e.g. LIBERO) might return `observation.images.image` and `observation.images.image2`.
-
-Rather than renaming columns in the dataset or editing the policy code, LeRobot lets you pass a **rename map**: a dictionary that says “when you see this key in the data, treat it as this key for the policy.” Renaming is applied in the preprocessing pipeline so the policy always receives the keys it expects.
-
----
-
-## How the rename map works
-
-The rename map is a dictionary:
-
-- **Keys** = observation keys as produced by your **dataset** (training) or **environment** (evaluation).
-- **Values** = the observation keys your **policy** expects.
-
-Only keys listed in the map are renamed; everything else is left as-is. Under the hood, the [RenameObservationsProcessorStep](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/rename_processor.py) runs in the preprocessor and rewrites observation keys (and keeps normalization stats aligned) so the batch matches the policy’s `input_features`.
-
-You can use the same idea for **training** (dataset → policy) and **evaluation** (env → policy).
-
-<p align="center">
-  <img
-    src="https://huggingface.co/datasets/jadechoghari/images/resolve/main/rename-map.png"
-    alt="Rename map: mapping dataset or environment observation keys to policy input keys"
-    style="max-width: 100%; height: auto;"
-  />
-</p>
-
----
-
-## Option 1: Use a rename map (recommended)
-
-You pass the mapping on the command line so dataset/env keys are renamed to what the policy expects. No need to change the policy repo or the data.
-
-### Training example: XVLA on a dataset with different camera names
-
-Suppose you fine-tune [lerobot/xvla-base](https://huggingface.co/lerobot/xvla-base) on a dataset whose images are stored under `observation.images.front`, `observation.images.eagle`, and `observation.images.glove`. XVLA expects `observation.images.image`, `observation.images.image2`, and `observation.images.image3`. Map the dataset keys to the policy keys:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=YOUR_DATASET \
-  --output_dir=./outputs/xvla_training \
-  --job_name=xvla_training \
-  --policy.path="lerobot/xvla-base" \
-  --policy.repo_id="HF_USER/xvla-your-robot" \
-  --policy.dtype=bfloat16 \
-  --policy.action_mode=auto \
-  --steps=20000 \
-  --policy.device=cuda \
-  --policy.freeze_vision_encoder=false \
-  --policy.freeze_language_encoder=false \
-  --policy.train_policy_transformer=true \
-  --policy.train_soft_prompts=true \
-  --rename_map='{"observation.images.front": "observation.images.image", "observation.images.eagle": "observation.images.image2", "observation.images.glove": "observation.images.image3"}'
-```
-
-Order of entries in the map doesn’t matter; each dataset key is renamed to the corresponding policy key.
-
-### Evaluation example: Policy trained on different camera names than the env
-
-You trained (or downloaded) a policy that expects `observation.images.base_0_rgb` and `observation.images.left_wrist_0_rgb` (e.g. [pi0fast-libero](https://huggingface.co/lerobot/pi0fast-libero)), but your evaluation environment (e.g. LIBERO) returns `observation.images.image` and `observation.images.image2`. Tell the eval script how to rename env keys to policy keys:
-
-```bash
-lerobot-eval \
-  --policy.path=lerobot/pi0fast-libero \
-  --env.type=libero \
-  ... \
-  --rename_map='{"observation.images.image": "observation.images.base_0_rgb", "observation.images.image2": "observation.images.left_wrist_0_rgb"}'
-```
-
-So: **key = what the env gives, value = what the policy expects.** Same convention as in training.
-
----
-
-## Option 2: Change the policy config (no rename map)
-
-If you prefer not to pass a rename map every time, you can **edit the policy’s `config.json`** so that its expected observation keys match your dataset or environment. For example, change the policy’s visual input keys to `observation.images.front`, `observation.images.eagle`, `observation.images.glove` to match your dataset, or to `observation.images.image` / `observation.images.image2` to match LIBERO.
-
-- **Training:** If the dataset’s camera keys match the (modified) policy config, you don’t need a rename map.
-- **Evaluation:** If the env’s keys match the (modified) policy config, you don’t need a rename map for eval either.
-
-The tradeoff: you’re changing the policy repo or your local checkpoint. That’s fine if you’re only ever using that one dataset or env; a rename map keeps the same policy usable across multiple data sources without touching the config.
-
----
-
-## When you have fewer cameras than the policy expects: empty cameras
-
-Some policies (e.g. XVLA) are built for a fixed number of image inputs (e.g. three). Your dataset might only have **two** cameras. You still want to fine-tune without changing the model architecture.
-
-LeRobot supports this with **empty cameras**: the config declares extra “slots” that the policy expects, but the dataset (or env) does not provide. Those slots are filled with placeholder keys and typically zero or masked inputs so the policy can run with fewer real views.
-
-<p align="center">
-  <img
-    src="https://huggingface.co/datasets/jadechoghari/images/resolve/main/empty_cam.png"
-    alt="Empty cameras: using placeholder slots when the dataset has fewer views than the policy expects"
-    style="max-width: 100%; height: auto;"
-  />
-</p>
-
-- In the policy config (e.g. [xvla-base config.json](https://huggingface.co/lerobot/xvla-base/blob/main/config.json)), `empty_cameras` is the number of these extra slots (default `0`).
-- For each slot, the config adds an observation key of the form:
-  `observation.images.empty_camera_0`, `observation.images.empty_camera_1`, …
-
-Example: XVLA-base has three visual inputs and `empty_cameras=0`. Your dataset has only two images. Set **`empty_cameras=1`**. Then:
-
-1. The config gains a third visual key: `observation.images.empty_camera_0`.
-2. You still use the rename map (or matching config keys) for the two real cameras.
-3. The third view is treated as “empty” (no corresponding dataset key); the policy ignores or masks it as needed.
-
-So you fine-tune on two observations only, and the third visual input is effectively unused. You do **not** need to add a fake third image to your dataset.
-
----
-
-## Where the rename map is used in the codebase
-
-- **Training** ([`lerobot_train.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_train.py)): `rename_map` is passed into `make_policy(..., rename_map=cfg.rename_map)` and into the preprocessor as `rename_observations_processor: {"rename_map": cfg.rename_map}`. Batches from the dataset are renamed before being fed to the policy.
-- **Evaluation** ([`lerobot_eval.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_eval.py)): Same idea—`rename_map` is passed to `make_policy` and to the preprocessor so env observations are renamed before the policy sees them.
-- **Processor** ([`rename_processor.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/processor/rename_processor.py)): `RenameObservationsProcessorStep` does the actual key renaming and updates feature metadata so normalization stats stay consistent with the renamed keys.
-
-If you see a feature mismatch error (“Missing features” / “Extra features”), the error message suggests using `--rename_map` with a mapping from your data’s keys to the policy’s expected keys.
-
----
-
-## Quick reference
-
-| Goal                                  | What to do                                                                                                 |
-| ------------------------------------- | ---------------------------------------------------------------------------------------------------------- |
-| Dataset keys ≠ policy keys (training) | `--rename_map='{"dataset_key": "policy_key", ...}'`                                                        |
-| Env keys ≠ policy keys (eval)         | `--rename_map='{"env_key": "policy_key", ...}'`                                                            |
-| Fewer cameras than policy expects     | Set `empty_cameras` in the policy config (e.g. `1` when you have 2 real cameras and the policy expects 3). |
-| Avoid passing a rename map            | Edit the policy’s `config.json` so its observation keys match your dataset or env.                         |
-
-The rename map keeps your pipeline flexible: one policy, many data sources, no code changes—just a small dictionary on the command line or in your config.

docs/source/sarm.mdx

@@ -269,7 +269,7 @@ This generates visualizations showing video frames with subtask boundaries overl
 Train with **no annotations** - uses linear progress from 0 to 1:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=single_stage \
@@ -288,7 +288,7 @@ lerobot-train \
 Train with **dense annotations only** (sparse auto-generated):
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dense_only \
@@ -307,7 +307,7 @@ lerobot-train \
 Train with **both sparse and dense annotations**:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=sarm \
   --policy.annotation_mode=dual \
@@ -468,7 +468,7 @@ This script:
 Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
   --use_rabc=true \

docs/source/smolvla.mdx

@@ -106,9 +106,6 @@ lerobot-record \
   --dataset.repo_id=${HF_USER}/eval_DATASET_NAME_test \  # <- This will be the dataset name on HF Hub
   --dataset.episode_time_s=50 \
   --dataset.num_episodes=10 \
-  --dataset.streaming_encoding=true \
-  --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/so101.mdx

@@ -1,18 +1,5 @@
 # SO-101
 
-<div style="display: flex; align-items: center; gap: 10px;">
-  <img
-    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Follower.webp"
-    alt="SO-101"
-    width="60%"
-  />
-  <img
-    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Leader.webp"
-    alt="SO-101"
-    width="60%"
-  />
-</div>
-
 In the steps below, we explain how to assemble our flagship robot, the SO-101.
 
 ## Source the parts

docs/source/streaming_video_encoding.mdx

@@ -1,155 +0,0 @@
-# Streaming Video Encoding Guide
-
-## 1. Overview
-
-Streaming video encoding eliminates the traditional PNG round-trip during video dataset recording. Instead of:
-
-1. Capture frame -> write PNG to disk -> (at episode end) read PNG's -> encode to MP4 -> delete PNG's
-
-Frames can be encoded in real-time during capture:
-
-1. Capture frame -> queue to encoder thread -> encode to MP4 directly
-
-This makes `save_episode()` near-instant (the video is already encoded by the time the episode ends) and removes the blocking wait that previously occurred between episodes, especially with multiple cameras in long episodes.
-
-## 2. Tuning Parameters
-
-| Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
-| ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
-| `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
-| `vcodec`                | `--dataset.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
-| `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
-| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |
-
-## 3. Performance Considerations
-
-Streaming encoding means the CPU is encoding video **during** the capture loop, not after. This creates a CPU budget that must be shared between:
-
-- **Control loop** (reading cameras, control the robot, writing non-video data)
-- **Encoder threads** (one pool per camera)
-- **Rerun visualization** (if enabled)
-- **OS and other processes**
-
-### Resolution & Number of Cameras Impact
-
-| Setup                     | Throughput (px/sec) | CPU Encoding Load | Notes                          |
-| ------------------------- | ------------------- | ----------------- | ------------------------------ |
-| 2camsx 640x480x3 @30fps   | 55M                 | Low               | Works on most systems          |
-| 2camsx 1280x720x3 @30fps  | 165M                | Moderate          | Comfortable on modern systems  |
-| 2camsx 1920x1080x3 @30fps | 373M                | High              | Requires powerful high-end CPU |
-
-### `encoder_threads` Tuning
-
-This parameter controls how many threads each encoder instance uses internally:
-
-- **Higher values** (e.g., 4-5): Faster encoding, but uses more CPU cores per camera. Good for high-end systems with many cores.
-- **Lower values** (e.g., 1-2): Less CPU per camera, freeing cores for capture and visualization. Good for low-res images and capable CPUs.
-- **`None` (default)**: Lets the codec decide. Information available in the codec logs.
-
-### Backpressure and Frame Dropping
-
-Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:
-
-1. The queue fills up (consuming RAM)
-2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
-3. A warning is logged: `"Encoder queue full for {camera}, dropped N frame(s)"`
-4. At episode end, total dropped frames per camera are reported
-
-### Symptoms of Encoder Falling Behind
-
-- **System feels laggy and freezes**: all CPUs are at 100%
-- **Dropped frame warnings** in the log or lower frames/FPS than expected in the recorded dataset
-- **Choppy robot movement**: If CPU is severely overloaded, even the capture loop may be affected
-- **Accumulated rerun lag**: Visualization falls behind real-time
-
-## 4. Hardware-Accelerated Encoding
-
-### When to Use
-
-Use HW encoding when:
-
-- CPU is the bottleneck (dropped frames, choppy robot, rerun lag)
-- You have compatible hardware (GPU or dedicated encoder)
-- You're recording at high throughput (high resolution or with many cameras)
-
-### Choosing a Codec
-
-| Codec                 | CPU Usage | File Size      | Quality | Notes                                                            |
-| --------------------- | --------- | -------------- | ------- | ---------------------------------------------------------------- |
-| `libsvtav1` (default) | High      | Smallest       | Best    | Default. Best compression but most CPU-intensive                 |
-| `h264`                | Medium    | ~30-50% larger | Good    | Software H.264. Lower CPU                                        |
-| HW encoders           | Very Low  | Largest        | Good    | Offloads to dedicated hardware. Best for CPU-constrained systems |
-
-### Available HW Encoders
-
-| Encoder             | Platform      | Hardware                                                                                         | CLI Value                            |
-| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
-| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=h264_videotoolbox` |
-| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=hevc_videotoolbox` |
-| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=h264_nvenc`        |
-| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=hevc_nvenc`        |
-| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.vcodec=h264_vaapi`        |
-| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.vcodec=h264_qsv`          |
-| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto`              |
-
-> [!NOTE]
-> In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
-
-> [!NOTE]
-> `libsvtav1` is the default because it provides the best training performance; other vcodecs can reduce CPU usage and be faster, but they typically produce larger files and may affect training time.
-
-## 5. Troubleshooting
-
-| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                  |
-| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
-| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`).                                                                                                                                                    |
-| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                     |
-| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                              |
-| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                              |
-| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.vcodec=auto`                                                                                                                                                                                                                             |
-| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                   |
-
-## 6. Recommended Configurations
-
-These estimates are conservative; we recommend testing them on your setup—start with a low load and increase it gradually.
-
-### High-End Systems: modern 12+ cores (24+ threads)
-
-A throughput between ~250-500M px/sec should be comfortable in CPU. For even better results try HW encoding if available.
-
-```bash
-# 3camsx 1280x720x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
-# 2camsx 1920x1080x3 @30fps: Defaults work well. Optionally increase encoder parallelism.
-lerobot-record --dataset.encoder_threads=5 ...
-
-# 3camsx 1920x1080x3 @30fps: Might require some tuning.
-```
-
-### Mid-Range Systems: modern 8+ cores (16+ threads) or Apple Silicon
-
-A throughput between ~80-300M px/sec should be possible in CPU.
-
-```bash
-# 3camsx 640x480x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
-# 2camsx 1280x720x3 @30fps: Defaults work well. Optionally decrease encoder parallelism.
-lerobot-record --dataset.encoder_threads=2 ...
-
-# 2camsx 1920x1080x3 @30fps: Might require some tuning.
-```
-
-### Low-Resource Systems: modern 4+ cores (8+ threads) or Raspberry Pi 5
-
-On very constrained systems, streaming encoding may compete too heavily with the capture loop. Disabling it falls back to the PNG-based approach where encoding happens between episodes (blocking, but doesn't interfere with capture). Alternatively, record at a lower throughput to reduce both capture and encoding load. Consider also changing codec to `h264` and using batch encoding.
-
-```bash
-# 2camsx 640x480x3 @30fps: Requires some tuning.
-
-# Use H.264, disable streaming, consider batching encoding
-lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
-```
-
-## 7. Closing note
-
-Performance ultimately depends on your exact setup — frames-per-second, resolution, CPU cores and load, available memory, episode length, and the encoder you choose. Always test with your target workload, be mindful about your CPU & system capabilities and tune `encoder_threads`, `encoder_queue_maxsize`, and
-`vcodec` reasonably. That said, a common practical configuration (for many applications) is three cameras at 640×480x3 @30fps; this usually runs fine with the default streaming video encoding settings in modern systems. Always verify your recorded dataset is healthy by comparing the video duration to the CLI episode duration and confirming the row count equals FPS × CLI duration.

docs/source/unitree_g1.mdx

@@ -188,111 +188,7 @@ Press `Ctrl+C` to stop the policy.
 
 ## Running in Simulation Mode (MuJoCo)
 
-You can test policies before deploying on the physical robot using MuJoCo simulation. Set `is_simulation=True` in config or pass `--robot.is_simulation=true` via CLI.
-
-### Calibrate Exoskeleton Teleoperator
-
-```bash
-lerobot-calibrate \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo
-```
-
-### Teleoperate in Simulation
-
-```bash
-lerobot-teleoperate \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=true \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --fps=100
-```
-
-### Record Dataset in Simulation
-
-```bash
-lerobot-record \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=true \
-    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --dataset.repo_id=your-username/dataset-name \
-    --dataset.single_task="Test" \
-    --dataset.num_episodes=2 \
-    --dataset.episode_time_s=5 \
-    --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
-```
-
-Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
-
----
-
-## Running on Real Robot
-
-Once the robot server is running on the G1 (see Part 3), you can teleoperate and record on the real robot.
-
-### Start the Camera Server
-
-On the robot, start the ZMQ image server:
-
-```bash
-python src/lerobot/cameras/zmq/image_server.py
-```
-
-Keep this running in a separate terminal for camera streaming during recording.
-
-### Teleoperate Real Robot
-
-```bash
-lerobot-teleoperate \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=false \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --fps=100
-```
-
-### Record Dataset on Real Robot
-
-```bash
-lerobot-record \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=false \
-    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --dataset.repo_id=your-username/dataset-name \
-    --dataset.single_task="Test" \
-    --dataset.num_episodes=2 \
-    --dataset.episode_time_s=5 \
-    --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true \
-    --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
-    --dataset.encoder_threads=2
-```
-
-**Note**: Update `server_address` to match your robot's camera server IP.
-
-Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/datasets/nepyope/teleop_test_real)
-
----
+You can now test policies before unleashing them on the physical robot using MuJoCo. To do so simply set `is_simulation=True` in config.
 
 ## Additional Resources
 

docs/source/using_dataset_tools.mdx

@@ -12,7 +12,6 @@ LeRobot provides several utilities for manipulating datasets:
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
 6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
-7. **Show the Info of Datasets** - Show the summary of datasets information such as number of episode etc.
 
 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`.
@@ -157,30 +156,6 @@ lerobot-edit-dataset \
 
 **Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.
 
-### Show the information of datasets
-
-Show the information of datasets such as number of episode, number of frame, File size and so on.
-No change will be made to the dataset
-
-```bash
-
-# Show dataset information without feature details
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_image \
-    --operation.type info \
-
-# Show dataset information with feature details
-lerobot-edit-dataset \
-    --repo_id lerobot/pusht_image \
-    --operation.type info \
-    --operation.show_features true
-
-```
-
-**Parameters:**
-
-- `parameters`: The flag to control show or no show dataset information with feature details.(default=false)
-
 ### Push to Hub
 
 Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:

docs/source/walloss.mdx

@@ -45,7 +45,7 @@ policy.type=wall_x
 For training WallX, you can use the standard LeRobot training script with the appropriate configuration:
 
 ```bash
-lerobot-train \
+python src/lerobot/scripts/lerobot_train.py \
     --dataset.repo_id=your_dataset \
     --policy.type=wall_x \
     --output_dir=./outputs/wallx_training \

docs/source/xvla.mdx

@@ -154,7 +154,7 @@ lerobot-train \
 
 ```bash
 lerobot-train \
-  --dataset.repo_id=<USER>/bimanual-so100-handover-cube \
+  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
   --output_dir=./outputs/xvla_bimanual \
   --job_name=xvla_so101_training \
   --policy.path="lerobot/xvla-base" \

examples/backward_compatibility/replay.py

@@ -22,7 +22,7 @@ lerobot-replay \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431541 \
     --robot.id=black \
-    --dataset.repo_id=<USER>/record-test \
+    --dataset.repo_id=aliberts/record-test \
     --dataset.episode=2
 ```
 """
@@ -57,7 +57,7 @@ class DatasetReplayConfig:
     repo_id: str
     # Episode to replay.
     episode: int
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | Path | None = None
     # Limit the frames per second. By default, uses the policy fps.
     fps: int = 30
@@ -81,25 +81,24 @@ def replay(cfg: ReplayConfig):
     actions = dataset.hf_dataset.select_columns(ACTION)
     robot.connect()
 
-    try:
-        log_say("Replaying episode", cfg.play_sounds, blocking=True)
-        for idx in range(dataset.num_frames):
-            start_episode_t = time.perf_counter()
+    log_say("Replaying episode", cfg.play_sounds, blocking=True)
+    for idx in range(dataset.num_frames):
+        start_episode_t = time.perf_counter()
 
-            action_array = actions[idx][ACTION]
-            action = {}
-            for i, name in enumerate(dataset.features[ACTION]["names"]):
-                key = f"{name.removeprefix('main_')}.pos"
-                action[key] = action_array[i].item()
+        action_array = actions[idx][ACTION]
+        action = {}
+        for i, name in enumerate(dataset.features[ACTION]["names"]):
+            key = f"{name.removeprefix('main_')}.pos"
+            action[key] = action_array[i].item()
 
-            action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
-            action["elbow_flex.pos"] -= 90
-            robot.send_action(action)
+        action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
+        action["elbow_flex.pos"] -= 90
+        robot.send_action(action)
 
-            dt_s = time.perf_counter() - start_episode_t
-            precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
-    finally:
-        robot.disconnect()
+        dt_s = time.perf_counter() - start_episode_t
+        precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
+
+    robot.disconnect()
 
 
 if __name__ == "__main__":

examples/lekiwi/evaluate.py

@@ -78,24 +78,40 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="lekiwi_evaluate")
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    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}")
+    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
+        # Main record loop
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            policy=policy,
+            preprocessor=preprocessor,  # Pass the pre and post policy processors
+            postprocessor=postprocessor,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            robot_observation_processor=robot_observation_processor,
+        )
+
+        # 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,
-                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,
@@ -104,42 +120,24 @@ def main():
                 robot_observation_processor=robot_observation_processor,
             )
 
-            # 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,
-                )
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-            if events["rerecord_episode"]:
-                log_say("Re-record episode")
-                events["rerecord_episode"] = False
-                events["exit_early"] = False
-                dataset.clear_episode_buffer()
-                continue
+        # Save episode
+        dataset.save_episode()
+        recorded_episodes += 1
 
-            # Save episode
-            dataset.save_episode()
-            recorded_episodes += 1
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
 
-    finally:
-        # Clean up
-        log_say("Stop recording")
-        robot.disconnect()
-        listener.stop()
-
-        dataset.finalize()
-        dataset.push_to_hub()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/lekiwi/record.py

@@ -74,23 +74,40 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="lekiwi_record")
 
-    try:
-        if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-            raise ValueError("Robot or teleop is not connected!")
+    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}")
+    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
+        # Main record loop
+        record_loop(
+            robot=robot,
+            events=events,
+            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,
+        )
+
+        # 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,
-                dataset=dataset,
                 teleop=[leader_arm, keyboard],
-                control_time_s=EPISODE_TIME_SEC,
+                control_time_s=RESET_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=teleop_action_processor,
@@ -98,44 +115,26 @@ def main():
                 robot_observation_processor=robot_observation_processor,
             )
 
-            # 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,
-                )
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-            if events["rerecord_episode"]:
-                log_say("Re-record episode")
-                events["rerecord_episode"] = False
-                events["exit_early"] = False
-                dataset.clear_episode_buffer()
-                continue
+        # Save episode
+        dataset.save_episode()
+        recorded_episodes += 1
 
-            # Save episode
-            dataset.save_episode()
-            recorded_episodes += 1
-    finally:
-        # Clean up
-        log_say("Stop recording")
-        robot.disconnect()
-        leader_arm.disconnect()
-        keyboard.disconnect()
-        listener.stop()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    leader_arm.disconnect()
+    keyboard.disconnect()
+    listener.stop()
 
-        dataset.finalize()
-        dataset.push_to_hub()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/lekiwi/replay.py

@@ -42,27 +42,25 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
 
-        print("Starting replay loop...")
-        log_say(f"Replaying episode {EPISODE_IDX}")
-        for idx in range(len(episode_frames)):
-            t0 = time.perf_counter()
+    print("Starting replay loop...")
+    log_say(f"Replaying episode {EPISODE_IDX}")
+    for idx in range(len(episode_frames)):
+        t0 = time.perf_counter()
 
-            # Get recorded action from dataset
-            action = {
-                name: float(actions[idx][ACTION][i])
-                for i, name in enumerate(dataset.features[ACTION]["names"])
-            }
+        # Get recorded action from dataset
+        action = {
+            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        }
 
-            # Send action to robot
-            _ = robot.send_action(action)
+        # Send action to robot
+        _ = robot.send_action(action)
 
-            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-    finally:
-        robot.disconnect()
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+
+    robot.disconnect()
 
 
 if __name__ == "__main__":

examples/lekiwi/teleoperate.py

@@ -43,12 +43,13 @@ def main():
     keyboard.connect()
 
     # Init rerun viewer
-    init_rerun(session_name="lekiwi_teleop")
+    init_rerun(session_name="lekiwi_teleop", robot=robot, reset_time=True)
 
     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 teleop loop...")
+    start = time.perf_counter()
     while True:
         t0 = time.perf_counter()
 
@@ -69,7 +70,7 @@ def main():
         _ = robot.send_action(action)
 
         # Visualize
-        log_rerun_data(observation=observation, action=action)
+        log_rerun_data(observation=observation, action=action, log_time=time.perf_counter() - start)
 
         precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 

examples/phone_to_so100/evaluate.py

@@ -142,24 +142,38 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="phone_so100_evaluate")
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    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}")
+    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
+        # Main record loop
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            policy=policy,
+            preprocessor=preprocessor,  # Pass the pre and post policy processors
+            postprocessor=postprocessor,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=make_default_teleop_action_processor(),
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose_processor,
+        )
+
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+            log_say("Reset the environment")
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
-                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,
@@ -168,41 +182,24 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose_processor,
             )
 
-            # Reset the environment if not stopping or re-recording
-            if not events["stop_recording"] and (
-                (episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]
-            ):
-                log_say("Reset the environment")
-                record_loop(
-                    robot=robot,
-                    events=events,
-                    fps=FPS,
-                    control_time_s=EPISODE_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=make_default_teleop_action_processor(),
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose_processor,
-                )
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-            if events["rerecord_episode"]:
-                log_say("Re-record episode")
-                events["rerecord_episode"] = False
-                events["exit_early"] = False
-                dataset.clear_episode_buffer()
-                continue
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
 
-            # Save episode
-            dataset.save_episode()
-            episode_idx += 1
-    finally:
-        # Clean up
-        log_say("Stop recording")
-        robot.disconnect()
-        listener.stop()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
 
-        dataset.finalize()
-        dataset.push_to_hub()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/record.py

@@ -149,23 +149,38 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="phone_so100_record")
 
-    try:
-        if not robot.is_connected or not phone.is_connected:
-            raise ValueError("Robot or teleop is not connected!")
+    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}")
+    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
+        # Main record loop
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            teleop=phone,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=phone_to_robot_ee_pose_processor,
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose,
+        )
+
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+            log_say("Reset the environment")
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
                 teleop=phone,
-                dataset=dataset,
-                control_time_s=EPISODE_TIME_SEC,
+                control_time_s=RESET_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=phone_to_robot_ee_pose_processor,
@@ -173,43 +188,25 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose,
             )
 
-            # Reset the environment if not stopping or re-recording
-            if not events["stop_recording"] and (
-                episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]
-            ):
-                log_say("Reset the environment")
-                record_loop(
-                    robot=robot,
-                    events=events,
-                    fps=FPS,
-                    teleop=phone,
-                    control_time_s=RESET_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=phone_to_robot_ee_pose_processor,
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose,
-                )
+        if events["rerecord_episode"]:
+            log_say("Re-recording episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-            if events["rerecord_episode"]:
-                log_say("Re-recording episode")
-                events["rerecord_episode"] = False
-                events["exit_early"] = False
-                dataset.clear_episode_buffer()
-                continue
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
 
-            # Save episode
-            dataset.save_episode()
-            episode_idx += 1
-    finally:
-        # Clean up
-        log_say("Stop recording")
-        robot.disconnect()
-        phone.disconnect()
-        listener.stop()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    phone.disconnect()
+    listener.stop()
 
-        dataset.finalize()
-        dataset.push_to_hub()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/replay.py

@@ -73,34 +73,32 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
 
-        print("Starting replay loop...")
-        log_say(f"Replaying episode {EPISODE_IDX}")
-        for idx in range(len(episode_frames)):
-            t0 = time.perf_counter()
+    print("Starting replay loop...")
+    log_say(f"Replaying episode {EPISODE_IDX}")
+    for idx in range(len(episode_frames)):
+        t0 = time.perf_counter()
 
-            # Get recorded action from dataset
-            ee_action = {
-                name: float(actions[idx][ACTION][i])
-                for i, name in enumerate(dataset.features[ACTION]["names"])
-            }
+        # Get recorded action from dataset
+        ee_action = {
+            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        }
 
-            # Get robot observation
-            robot_obs = robot.get_observation()
+        # Get robot observation
+        robot_obs = robot.get_observation()
 
-            # Dataset EE -> robot joints
-            joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+        # Dataset EE -> robot joints
+        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-            # Send action to robot
-            _ = robot.send_action(joint_action)
+        # Send action to robot
+        _ = robot.send_action(joint_action)
 
-            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-    finally:
-        # Clean up
-        robot.disconnect()
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+
+    # Clean up
+    robot.disconnect()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/teleoperate.py

@@ -89,12 +89,13 @@ def main():
     teleop_device.connect()
 
     # Init rerun viewer
-    init_rerun(session_name="phone_so100_teleop")
+    init_rerun(session_name="phone_so100_teleop", robot=robot, reset_time=True)
 
     if not robot.is_connected or not teleop_device.is_connected:
         raise ValueError("Robot or teleop is not connected!")
 
     print("Starting teleop loop. Move your phone to teleoperate the robot...")
+    start = time.perf_counter()
     while True:
         t0 = time.perf_counter()
 
@@ -111,7 +112,7 @@ def main():
         _ = robot.send_action(joint_action)
 
         # Visualize
-        log_rerun_data(observation=phone_obs, action=joint_action)
+        log_rerun_data(observation=phone_obs, action=joint_action, log_time=time.perf_counter() - start)
 
         precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 

examples/rtc/eval_dataset.py

@@ -27,8 +27,8 @@ measuring consistency and ground truth alignment.
 Usage:
     # Basic usage with smolvla policy
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --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 \
@@ -58,16 +58,16 @@ Usage:
         --device=cuda
 
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/reuben_pi0 \
-        --dataset.repo_id=<USER>/so101_cube_in_cup \
+        --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=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=mps \
         --use_torch_compile=true \
@@ -75,8 +75,8 @@ Usage:
 
     # With torch.compile on CUDA (CUDA graphs disabled by default)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
         --rtc.execution_horizon=8 \
         --device=cuda \
         --use_torch_compile=true \
@@ -84,8 +84,8 @@ Usage:
 
     # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
     uv run python examples/rtc/eval_dataset.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
-        --dataset.repo_id=<USER>/check_rtc \
+        --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 \

examples/rtc/eval_with_real_robot.py

@@ -28,7 +28,7 @@ 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=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \
@@ -41,7 +41,7 @@ Usage:
 
     # Run RTC with Real robot without RTC
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/smolvla_check_rtc_last3 \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
         --policy.device=mps \
         --rtc.enabled=false \
         --robot.type=so100_follower \
@@ -53,7 +53,7 @@ Usage:
 
     # Run RTC with Real robot with pi0.5 policy
     uv run examples/rtc/eval_with_real_robot.py \
-        --policy.path=<USER>/pi05_check_rtc \
+        --policy.path=helper2424/pi05_check_rtc \
         --policy.device=mps \
         --rtc.enabled=true \
         --rtc.execution_horizon=20 \

examples/so100_to_so100_EE/evaluate.py

@@ -142,24 +142,38 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="so100_so100_evaluate")
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    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}")
+    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
+        # Main record loop
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            policy=policy,
+            preprocessor=preprocessor,  # Pass the pre and post policy processors
+            postprocessor=postprocessor,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=make_default_teleop_action_processor(),
+            robot_action_processor=robot_ee_to_joints_processor,
+            robot_observation_processor=robot_joints_to_ee_pose_processor,
+        )
+
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+            log_say("Reset the environment")
             record_loop(
                 robot=robot,
                 events=events,
                 fps=FPS,
-                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,
@@ -168,41 +182,24 @@ def main():
                 robot_observation_processor=robot_joints_to_ee_pose_processor,
             )
 
-            # Reset the environment if not stopping or re-recording
-            if not events["stop_recording"] and (
-                (episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]
-            ):
-                log_say("Reset the environment")
-                record_loop(
-                    robot=robot,
-                    events=events,
-                    fps=FPS,
-                    control_time_s=EPISODE_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=make_default_teleop_action_processor(),
-                    robot_action_processor=robot_ee_to_joints_processor,
-                    robot_observation_processor=robot_joints_to_ee_pose_processor,
-                )
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-            if events["rerecord_episode"]:
-                log_say("Re-record episode")
-                events["rerecord_episode"] = False
-                events["exit_early"] = False
-                dataset.clear_episode_buffer()
-                continue
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
 
-            # Save episode
-            dataset.save_episode()
-            episode_idx += 1
-    finally:
-        # Clean up
-        log_say("Stop recording")
-        robot.disconnect()
-        listener.stop()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
 
-        dataset.finalize()
-        dataset.push_to_hub()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/record.py

@@ -146,23 +146,38 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="recording_phone")
 
-    try:
-        if not leader.is_connected or not follower.is_connected:
-            raise ValueError("Robot or teleop is not connected!")
+    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}")
+    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
+        # Main record loop
+        record_loop(
+            robot=follower,
+            events=events,
+            fps=FPS,
+            teleop=leader,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=leader_joints_to_ee,
+            robot_action_processor=ee_to_follower_joints,
+            robot_observation_processor=follower_joints_to_ee,
+        )
+
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+            log_say("Reset the environment")
             record_loop(
                 robot=follower,
                 events=events,
                 fps=FPS,
                 teleop=leader,
-                dataset=dataset,
-                control_time_s=EPISODE_TIME_SEC,
+                control_time_s=RESET_TIME_SEC,
                 single_task=TASK_DESCRIPTION,
                 display_data=True,
                 teleop_action_processor=leader_joints_to_ee,
@@ -170,44 +185,25 @@ def main():
                 robot_observation_processor=follower_joints_to_ee,
             )
 
-            # Reset the environment if not stopping or re-recording
-            if not events["stop_recording"] and (
-                episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]
-            ):
-                log_say("Reset the environment")
-                record_loop(
-                    robot=follower,
-                    events=events,
-                    fps=FPS,
-                    teleop=leader,
-                    control_time_s=RESET_TIME_SEC,
-                    single_task=TASK_DESCRIPTION,
-                    display_data=True,
-                    teleop_action_processor=leader_joints_to_ee,
-                    robot_action_processor=ee_to_follower_joints,
-                    robot_observation_processor=follower_joints_to_ee,
-                )
+        if events["rerecord_episode"]:
+            log_say("Re-recording episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-            if events["rerecord_episode"]:
-                log_say("Re-recording episode")
-                events["rerecord_episode"] = False
-                events["exit_early"] = False
-                dataset.clear_episode_buffer()
-                continue
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
 
-            # Save episode
-            dataset.save_episode()
-            episode_idx += 1
+    # Clean up
+    log_say("Stop recording")
+    leader.disconnect()
+    follower.disconnect()
+    listener.stop()
 
-    finally:
-        # Clean up
-        log_say("Stop recording")
-        leader.disconnect()
-        follower.disconnect()
-        listener.stop()
-
-        dataset.finalize()
-        dataset.push_to_hub()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/replay.py

@@ -74,35 +74,32 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
 
-        print("Starting replay loop...")
-        log_say(f"Replaying episode {EPISODE_IDX}")
-        for idx in range(len(episode_frames)):
-            t0 = time.perf_counter()
+    print("Starting replay loop...")
+    log_say(f"Replaying episode {EPISODE_IDX}")
+    for idx in range(len(episode_frames)):
+        t0 = time.perf_counter()
 
-            # Get recorded action from dataset
-            ee_action = {
-                name: float(actions[idx][ACTION][i])
-                for i, name in enumerate(dataset.features[ACTION]["names"])
-            }
+        # Get recorded action from dataset
+        ee_action = {
+            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        }
 
-            # Get robot observation
-            robot_obs = robot.get_observation()
+        # Get robot observation
+        robot_obs = robot.get_observation()
 
-            # Dataset EE -> robot joints
-            joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+        # Dataset EE -> robot joints
+        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-            # Send action to robot
-            _ = robot.send_action(joint_action)
+        # Send action to robot
+        _ = robot.send_action(joint_action)
 
-            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
-    finally:
-        # Clean up
-        robot.disconnect()
+    # Clean up
+    robot.disconnect()
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/teleoperate.py

@@ -94,9 +94,10 @@ def main():
     leader.connect()
 
     # Init rerun viewer
-    init_rerun(session_name="so100_so100_EE_teleop")
+    init_rerun(session_name="so100_so100_EE_teleop", robot=follower, reset_time=True)
 
     print("Starting teleop loop...")
+    start = time.perf_counter()
     while True:
         t0 = time.perf_counter()
 
@@ -116,7 +117,9 @@ def main():
         _ = follower.send_action(follower_joints_act)
 
         # Visualize
-        log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
+        log_rerun_data(
+            observation=leader_ee_act, action=follower_joints_act, log_time=time.perf_counter() - start
+        )
 
         precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 

examples/tutorial/async-inf/robot_client.py

@@ -30,7 +30,6 @@ def main():
         robot=robot_cfg,
         server_address=server_address,
         policy_device="mps",
-        client_device="cpu",
         policy_type="act",
         pretrained_name_or_path="<user>/robot_learning_tutorial_act",
         chunk_size_threshold=0.5,  # g

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.5"
+version = "0.4.3"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 dynamic = ["readme"]
 license = { text = "Apache-2.0" }
@@ -59,7 +59,7 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 dependencies = [
 
     # Hugging Face dependencies
-    "datasets>=4.0.0,<5.0.0",
+    "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",
@@ -76,9 +76,9 @@ dependencies = [
     "pyserial>=3.5,<4.0",
     "wandb>=0.24.0,<0.25.0",
 
-    "torch==2.10.0",
-    "torchcodec>=0.2.1,<0.11.0; 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: Bump dependency
-    "torchvision>=0.21.0,<0.26.0", # TODO: Bump dependency
+    "torch>=2.2.1,<2.8.0", # TODO: Bumb dependency
+    "torchcodec>=0.2.1,<0.6.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # TODO: Bumb dependency
+    "torchvision>=0.21.0,<0.23.0", # TODO: Bumb dependency
 
     "draccus==0.10.0", # TODO: Remove ==
     "gymnasium>=1.1.1,<2.0.0",
@@ -98,26 +98,18 @@ pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
 transformers-dep = ["transformers>=4.57.1,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
-can-dep = ["python-can>=4.2.0,<5.0.0"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
-damiao = ["lerobot[can-dep]"]
-robstride = ["lerobot[can-dep]"]
 
 # Robots
-openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
 unitree_g1 = [
     "pyzmq>=26.2.1,<28.0.0",
-    "onnxruntime>=1.16.0,<2.0.0",
-    "pin>=3.0.0,<4.0.0",
-    "meshcat>=0.3.0,<0.4.0",
-    "matplotlib>=3.9.0,<4.0.0",
-    "casadi>=3.6.0,<4.0.0",
+    "onnxruntime>=1.16.0,<2.0.0"
 ]
 reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
@@ -155,6 +147,7 @@ hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpci
 # Features
 async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
 peft = ["lerobot[transformers-dep]", "peft>=0.18.0,<1.0.0"]
+audio = ["sounddevice>=0.5.1,<0.6.0", "soundfile>=0.13.1,<0.14.0", "librosa>=0.11.0,<0.12.0", "torchaudio>=2.6.0,<2.10.0"]
 
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1", "mypy>=1.19.1"]
@@ -183,6 +176,7 @@ all = [
     "lerobot[xvla]",
     "lerobot[hilserl]",
     "lerobot[async]",
+    "lerobot[audio]",
     "lerobot[dev]",
     "lerobot[test]",
     "lerobot[video_benchmark]",
@@ -211,12 +205,8 @@ lerobot-info="lerobot.scripts.lerobot_info:main"
 lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
-lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 
 # ---------------- Tool Configurations ----------------
-[tool.setuptools.package-data]
-lerobot = ["envs/*.json"]
-
 [tool.setuptools.packages.find]
 where = ["src"]
 
@@ -290,7 +280,6 @@ default.extend-ignore-identifiers-re = [
     "thw",
     "inpt",
     "ROBOTIS",
-    "OT_VALUE"
 ]
 
 # TODO: Uncomment when ready to use
@@ -365,9 +354,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false
 
-[[tool.mypy.overrides]]
-module = "lerobot.motors.*"
-ignore_errors = false
+# [[tool.mypy.overrides]]
+# module = "lerobot.motors.*"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"

src/lerobot/__init__.py

@@ -29,6 +29,7 @@ Example:
         print(lerobot.available_policies_per_env)
         print(lerobot.available_robots)
         print(lerobot.available_cameras)
+        print(lerobot.available_microphones)
         print(lerobot.available_motors)
     ```
 
@@ -174,6 +175,13 @@ available_cameras = [
     "intelrealsense",
 ]
 
+# lists all available microphones from `lerobot/microphones`
+available_microphones = [
+    "portaudio",
+    "touchlab",
+    "anyskin",
+]
+
 # lists all available motors from `lerobot/motors`
 available_motors = [
     "dynamixel",

src/lerobot/async_inference/configs.py

@@ -126,12 +126,6 @@ class RobotClientConfig:
 
     # Device configuration
     policy_device: str = field(default="cpu", metadata={"help": "Device for policy inference"})
-    client_device: str = field(
-        default="cpu",
-        metadata={
-            "help": "Device to move actions to after receiving from server (e.g., for downstream planners)"
-        },
-    )
 
     # Control behavior configuration
     chunk_size_threshold: float = field(default=0.5, metadata={"help": "Threshold for chunk size control"})
@@ -167,9 +161,6 @@ class RobotClientConfig:
         if not self.policy_device:
             raise ValueError("policy_device cannot be empty")
 
-        if not self.client_device:
-            raise ValueError("client_device cannot be empty")
-
         if self.chunk_size_threshold < 0 or self.chunk_size_threshold > 1:
             raise ValueError(f"chunk_size_threshold must be between 0 and 1, got {self.chunk_size_threshold}")
 
@@ -193,7 +184,6 @@ class RobotClientConfig:
             "policy_type": self.policy_type,
             "pretrained_name_or_path": self.pretrained_name_or_path,
             "policy_device": self.policy_device,
-            "client_device": self.client_device,
             "chunk_size_threshold": self.chunk_size_threshold,
             "fps": self.fps,
             "actions_per_chunk": self.actions_per_chunk,

src/lerobot/async_inference/helpers.py

@@ -18,7 +18,6 @@ import os
 import time
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any
 
 import torch
 
@@ -40,8 +39,8 @@ from lerobot.utils.utils import init_logging
 
 Action = torch.Tensor
 
-# observation as received from the robot (can be numpy arrays, floats, etc.)
-RawObservation = dict[str, Any]
+# observation as received from the robot
+RawObservation = dict[str, torch.Tensor]
 
 # observation as those recorded in LeRobot dataset (keys are different)
 LeRobotObservation = dict[str, torch.Tensor]

src/lerobot/async_inference/policy_server.py

@@ -381,8 +381,6 @@ class PolicyServer(services_pb2_grpc.AsyncInferenceServicer):
         action_tensor = torch.stack(processed_actions, dim=1).squeeze(0)
         self.logger.debug(f"Postprocessed action shape: {action_tensor.shape}")
 
-        action_tensor = action_tensor.detach().cpu()
-
         """5. Convert to TimedAction list"""
         action_chunk = self._time_action_chunk(
             observation_t.get_timestamp(), list(action_tensor), observation_t.get_timestep()

src/lerobot/async_inference/robot_client.py

@@ -25,7 +25,6 @@ python src/lerobot/async_inference/robot_client.py \
     --policy_type=act \
     --pretrained_name_or_path=user/model \
     --policy_device=mps \
-    --client_device=cpu \
     --actions_per_chunk=50 \
     --chunk_size_threshold=0.5 \
     --aggregate_fn_name=weighted_average \
@@ -41,7 +40,6 @@ from collections.abc import Callable
 from dataclasses import asdict
 from pprint import pformat
 from queue import Queue
-from typing import Any
 
 import draccus
 import grpc
@@ -49,18 +47,27 @@ import torch
 
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
 from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.robots import (
-    RobotConfig,  # noqa: F401
+from lerobot.microphones.anyskin.configuration_anyskin import AnyskinSensorConfig  # noqa: F401
+from lerobot.microphones.portaudio.configuration_portaudio import PortAudioMicrophoneConfig  # noqa: F401
+from lerobot.microphones.touchlab.configuration_touchlab import TouchLabSensorConfig  # noqa: F401
+from lerobot.processor import RobotAction
+from lerobot.robots import (  # noqa: F401
+    Robot,
+    RobotConfig,
+    bi_so_follower,
+    koch_follower,
     make_robot_from_config,
+    omx_follower,
+    so_follower,
 )
 from lerobot.transport import (
     services_pb2,  # type: ignore
     services_pb2_grpc,  # type: ignore
 )
 from lerobot.transport.utils import grpc_channel_options, send_bytes_in_chunks
-from lerobot.utils.import_utils import register_third_party_plugins
 
 from .configs import RobotClientConfig
+from .constants import SUPPORTED_ROBOTS
 from .helpers import (
     Action,
     FPSTracker,
@@ -281,21 +288,6 @@ class RobotClient:
                 timed_actions = pickle.loads(actions_chunk.data)  # nosec
                 deserialize_time = time.perf_counter() - deserialize_start
 
-                # Log device type of received actions
-                if len(timed_actions) > 0:
-                    received_device = timed_actions[0].get_action().device.type
-                    self.logger.debug(f"Received actions on device: {received_device}")
-
-                # Move actions to client_device (e.g., for downstream planners that need GPU)
-                client_device = self.config.client_device
-                if client_device != "cpu":
-                    for timed_action in timed_actions:
-                        if timed_action.get_action().device.type != client_device:
-                            timed_action.action = timed_action.get_action().to(client_device)
-                    self.logger.debug(f"Converted actions to device: {client_device}")
-                else:
-                    self.logger.debug(f"Actions kept on device: {client_device}")
-
                 self.action_chunk_size = max(self.action_chunk_size, len(timed_actions))
 
                 # Calculate network latency if we have matching observations
@@ -362,7 +354,7 @@ class RobotClient:
         action = {key: action_tensor[i].item() for i, key in enumerate(self.robot.action_features)}
         return action
 
-    def control_loop_action(self, verbose: bool = False) -> dict[str, Any]:
+    def control_loop_action(self, verbose: bool = False) -> RobotAction:
         """Reading and performing actions in local queue"""
 
         # Lock only for queue operations
@@ -480,9 +472,8 @@ class RobotClient:
 def async_client(cfg: RobotClientConfig):
     logging.info(pformat(asdict(cfg)))
 
-    # TODO: Assert if checking robot support is still needed with the plugin system
-    # if cfg.robot.type not in SUPPORTED_ROBOTS:
-    #     raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
+    if cfg.robot.type not in SUPPORTED_ROBOTS:
+        raise ValueError(f"Robot {cfg.robot.type} not yet supported!")
 
     client = RobotClient(cfg)
 
@@ -508,5 +499,4 @@ def async_client(cfg: RobotClientConfig):
 
 
 if __name__ == "__main__":
-    register_third_party_plugins()
     async_client()  # run the client

src/lerobot/cameras/__init__.py

@@ -13,5 +13,5 @@
 # limitations under the License.
 
 from .camera import Camera
-from .configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
+from .configs import CameraConfig, ColorMode, Cv2Rotation
 from .utils import make_cameras_from_configs

src/lerobot/cameras/camera.py

@@ -15,12 +15,11 @@
 # limitations under the License.
 
 import abc
-import warnings
 from typing import Any
 
 from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
-from .configs import CameraConfig
+from .configs import CameraConfig, ColorMode
 
 
 class Camera(abc.ABC):
@@ -31,12 +30,20 @@ class Camera(abc.ABC):
 
     Manages basic camera properties (FPS, resolution) and core operations:
     - Connection/disconnection
-    - Frame capture (sync/async/latest)
+    - Frame capture (sync/async)
 
     Attributes:
         fps (int | None): Configured frames per second
         width (int | None): Frame width in pixels
         height (int | None): Frame height in pixels
+
+    Example:
+        class MyCamera(Camera):
+            def __init__(self, config): ...
+            @property
+            def is_connected(self) -> bool: ...
+            def connect(self, warmup=True): ...
+            # Plus other required methods
     """
 
     def __init__(self, config: CameraConfig):
@@ -49,32 +56,6 @@ class Camera(abc.ABC):
         self.width: int | None = config.width
         self.height: int | None = config.height
 
-    def __enter__(self):
-        """
-        Context manager entry.
-        Automatically connects to the camera.
-        """
-        self.connect()
-        return self
-
-    def __exit__(self, exc_type, exc_value, traceback) -> None:
-        """
-        Context manager exit.
-        Automatically disconnects, ensuring resources are released even on error.
-        """
-        self.disconnect()
-
-    def __del__(self) -> None:
-        """
-        Destructor safety net.
-        Attempts to disconnect if the object is garbage collected without cleanup.
-        """
-        try:
-            if self.is_connected:
-                self.disconnect()
-        except Exception:  # nosec B110
-            pass
-
     @property
     @abc.abstractmethod
     def is_connected(self) -> bool:
@@ -108,10 +89,12 @@ class Camera(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def read(self) -> NDArray[Any]:
-        """Capture and return a single frame from the camera synchronously.
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+        """Capture and return a single frame from the camera.
 
-        This is a blocking call that will wait for the hardware and its SDK.
+        Args:
+            color_mode: Desired color mode for the output frame. If None,
+                        uses the camera's default color mode.
 
         Returns:
             np.ndarray: Captured frame as a numpy array.
@@ -120,64 +103,17 @@ class Camera(abc.ABC):
 
     @abc.abstractmethod
     def async_read(self, timeout_ms: float = ...) -> NDArray[Any]:
-        """Return the most recent new frame.
-
-        This method retrieves the latest frame captured by the background thread.
-        If a new frame is already available in the buffer (captured since the last call),
-        it returns it immediately.
-
-        It blocks up to `timeout_ms` only if the buffer is empty or if the latest frame
-        was already consumed by a previous `async_read` call.
-
-        Essentially, this method return the latest unconsumed frame, waiting if necessary
-        for a new one to arrive within the specified timeout.
-
-        Usage:
-            - Ideal for control loops where you want to ensure every processed frame
-            is fresh, effectively synchronizing your loop to the camera's FPS.
-            - Causes of a timeout usually include: very low camera FPS, heavy processing load,
-            or if the camera is disconnected.
+        """Asynchronously capture and return a single frame from the camera.
 
         Args:
-            timeout_ms: Maximum time to wait for a new frame in milliseconds.
-                        Defaults to 200ms (0.2s).
+            timeout_ms: Maximum time to wait for a frame in milliseconds.
+                        Defaults to implementation-specific timeout.
 
         Returns:
             np.ndarray: Captured frame as a numpy array.
-
-        Raises:
-            TimeoutError: If no new frame arrives within `timeout_ms`.
         """
         pass
 
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
-        """Return the most recent frame captured immediately (Peeking).
-
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
-
-        Usage:
-            Ideal for scenarios requiring zero latency or decoupled frequencies & when
-            we want a guaranteed frame, such as UI visualization, logging, or
-            non-critical monitoring.
-
-        Returns:
-            NDArray[Any]: The frame image (numpy array).
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            NotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-        warnings.warn(
-            f"{self.__class__.__name__}.read_latest() is not implemented. "
-            "Please override read_latest(); it will be required in future releases.",
-            FutureWarning,
-            stacklevel=2,
-        )
-        return self.async_read()
-
     @abc.abstractmethod
     def disconnect(self) -> None:
         """Disconnect from the camera and release resources."""

src/lerobot/cameras/configs.py

@@ -25,10 +25,6 @@ class ColorMode(str, Enum):
     RGB = "rgb"
     BGR = "bgr"
 
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`color_mode` is expected to be in {list(cls)}, but {value} is provided.")
-
 
 class Cv2Rotation(int, Enum):
     NO_ROTATION = 0
@@ -36,25 +32,6 @@ class Cv2Rotation(int, Enum):
     ROTATE_180 = 180
     ROTATE_270 = -90
 
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`rotation` is expected to be in {list(cls)}, but {value} is provided.")
-
-
-# Subset from https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html
-class Cv2Backends(int, Enum):
-    ANY = 0
-    V4L2 = 200
-    DSHOW = 700
-    PVAPI = 800
-    ANDROID = 1000
-    AVFOUNDATION = 1200
-    MSMF = 1400
-
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`backend` is expected to be in {list(cls)}, but {value} is provided.")
-
 
 @dataclass(kw_only=True)
 class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus

src/lerobot/cameras/opencv/camera_opencv.py

@@ -32,11 +32,10 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
-from ..utils import get_cv2_rotation
+from ..utils import get_cv2_backend, get_cv2_rotation
 from .configuration_opencv import ColorMode, OpenCVCameraConfig
 
 # NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
@@ -71,24 +70,34 @@ class OpenCVCamera(Camera):
     Example:
         ```python
         from lerobot.cameras.opencv import OpenCVCamera
-        from lerobot.cameras.configuration_opencv import OpenCVCameraConfig
+        from lerobot.cameras.configuration_opencv import OpenCVCameraConfig, ColorMode, Cv2Rotation
 
         # Basic usage with camera index 0
         config = OpenCVCameraConfig(index_or_path=0)
         camera = OpenCVCamera(config)
         camera.connect()
 
-        # Read 1 frame synchronously (blocking)
+        # Read 1 frame synchronously
         color_image = camera.read()
+        print(color_image.shape)
 
-        # Read 1 frame asynchronously (waits for new frame with a timeout)
+        # Read 1 frame asynchronously
         async_image = camera.async_read()
 
-        # Get the latest frame immediately (no wait, returns timestamp)
-        latest_image, timestamp = camera.read_latest()
-
         # When done, properly disconnect the camera using
         camera.disconnect()
+
+        # Example with custom settings
+        custom_config = OpenCVCameraConfig(
+            index_or_path='/dev/video0', # Or use an index
+            fps=30,
+            width=1280,
+            height=720,
+            color_mode=ColorMode.RGB,
+            rotation=Cv2Rotation.ROTATE_90
+        )
+        custom_camera = OpenCVCamera(custom_config)
+        # ... connect, read, disconnect ...
         ```
     """
 
@@ -114,11 +123,10 @@ class OpenCVCamera(Camera):
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
         self.latest_frame: NDArray[Any] | None = None
-        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = config.backend
+        self.backend: int = get_cv2_backend()
 
         if self.height and self.width:
             self.capture_width, self.capture_height = self.width, self.height
@@ -133,23 +141,20 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the OpenCV camera specified in the configuration.
 
         Initializes the OpenCV VideoCapture object, sets desired camera properties
-        (FPS, width, height), starts the background reading thread and performs initial checks.
-
-        Args:
-            warmup (bool): If True, waits at connect() time until at least one valid frame
-                           has been captured by the background thread. Defaults to True.
+        (FPS, width, height), and performs initial checks.
 
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
-            ConnectionError: If the specified camera index/path is not found or fails to open.
-            RuntimeError: If the camera opens but fails to apply requested settings.
+            ConnectionError: If the specified camera index/path is not found or the camera is found but fails to open.
+            RuntimeError: If the camera opens but fails to apply requested FPS/resolution settings.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         # Use 1 thread for OpenCV operations to avoid potential conflicts or
         # blocking in multi-threaded applications, especially during data collection.
@@ -165,20 +170,15 @@ class OpenCVCamera(Camera):
             )
 
         self._configure_capture_settings()
-        self._start_read_thread()
 
-        if warmup and self.warmup_s > 0:
+        if warmup:
             start_time = time.time()
             while time.time() - start_time < self.warmup_s:
-                self.async_read(timeout_ms=self.warmup_s * 1000)
+                self.read()
                 time.sleep(0.1)
-            with self.frame_lock:
-                if self.latest_frame is None:
-                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
 
         logger.info(f"{self} connected.")
 
-    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """
         Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
@@ -196,8 +196,11 @@ class OpenCVCamera(Camera):
         Raises:
             RuntimeError: If the camera fails to set any of the specified properties
                           to the requested value.
-            DeviceNotConnectedError: If the camera is not connected.
+            DeviceNotConnectedError: If the camera is not connected when attempting
+                                     to configure settings.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
         # Set FOURCC first (if specified) as it can affect available FPS/resolution options
         if self.config.fourcc is not None:
@@ -336,18 +339,6 @@ class OpenCVCamera(Camera):
 
         return found_cameras_info
 
-    def _read_from_hardware(self) -> NDArray[Any]:
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
-
-        ret, frame = self.videocapture.read()
-
-        if not ret:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
-
-        return frame
-
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -355,6 +346,11 @@ class OpenCVCamera(Camera):
         This is a blocking call. It waits for the next available frame from the
         camera hardware via OpenCV.
 
+        Args:
+            color_mode (Optional[ColorMode]): If specified, overrides the default
+                color mode (`self.color_mode`) for this read operation (e.g.,
+                request RGB even if default is BGR).
+
         Returns:
             np.ndarray: The captured frame as a NumPy array in the format
                        (height, width, channels), using the specified or default
@@ -366,31 +362,34 @@ class OpenCVCamera(Camera):
                           received frame dimensions don't match expectations before rotation.
             ValueError: If an invalid `color_mode` is requested.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         start_time = time.perf_counter()
 
-        if color_mode is not None:
-            logger.warning(
-                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
-            )
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        ret, frame = self.videocapture.read()
 
-        self.new_frame_event.clear()
-        frame = self.async_read(timeout_ms=10000)
+        if not ret or frame is None:
+            raise RuntimeError(f"{self} read failed (status={ret}).")
+
+        processed_frame = self._postprocess_image(frame, color_mode)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
-        return frame
+        return processed_frame
 
-    def _postprocess_image(self, image: NDArray[Any]) -> NDArray[Any]:
+    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.
 
         Args:
             image (np.ndarray): The raw image frame (expected BGR format from OpenCV).
+            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
+                                             uses the instance's default `self.color_mode`.
 
         Returns:
             np.ndarray: The processed image frame.
@@ -400,10 +399,11 @@ class OpenCVCamera(Camera):
             RuntimeError: If the raw frame dimensions do not match the configured
                           `width` and `height`.
         """
+        requested_color_mode = self.color_mode if color_mode is None else color_mode
 
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+        if requested_color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
-                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+                f"Invalid color mode '{requested_color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
 
         h, w, c = image.shape
@@ -417,7 +417,7 @@ class OpenCVCamera(Camera):
             raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
 
         processed_image = image
-        if self.color_mode == ColorMode.RGB:
+        if requested_color_mode == ColorMode.RGB:
             processed_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
 
         if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE, cv2.ROTATE_180]:
@@ -431,7 +431,7 @@ class OpenCVCamera(Camera):
 
         On each iteration:
         1. Reads a color frame
-        2. Stores result in latest_frame and updates timestamp (thread-safe)
+        2. Stores result in latest_frame (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -439,37 +439,30 @@ class OpenCVCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
-        failure_count = 0
         while not self.stop_event.is_set():
             try:
-                raw_frame = self._read_from_hardware()
-                processed_frame = self._postprocess_image(raw_frame)
-                capture_time = time.perf_counter()
+                color_image = self.read()
 
                 with self.frame_lock:
-                    self.latest_frame = processed_frame
-                    self.latest_timestamp = capture_time
+                    self.latest_frame = color_image
                 self.new_frame_event.set()
-                failure_count = 0
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
-                if failure_count <= 10:
-                    failure_count += 1
-                    logger.warning(f"Error reading frame in background thread for {self}: {e}")
-                else:
-                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
+                logger.warning(f"Error reading frame in background thread for {self}: {e}")
 
     def _start_read_thread(self) -> None:
         """Starts or restarts the background read thread if it's not running."""
-        self._stop_read_thread()
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=0.1)
+        if self.stop_event is not None:
+            self.stop_event.set()
 
         self.stop_event = Event()
         self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
         self.thread.daemon = True
         self.thread.start()
-        time.sleep(0.1)
 
     def _stop_read_thread(self) -> None:
         """Signals the background read thread to stop and waits for it to join."""
@@ -482,12 +475,6 @@ class OpenCVCamera(Camera):
         self.thread = None
         self.stop_event = None
 
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -495,7 +482,6 @@ class OpenCVCamera(Camera):
         This method retrieves the most recent frame captured by the background
         read thread. It does not block waiting for the camera hardware directly,
         but may wait up to timeout_ms for the background thread to provide a frame.
-        It is “best effort” under high FPS.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -510,14 +496,17 @@ class OpenCVCamera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+            self._start_read_thread()
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            thread_alive = self.thread is not None and self.thread.is_alive()
             raise TimeoutError(
                 f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {self.thread.is_alive()}."
+                f"Read thread alive: {thread_alive}."
             )
 
         with self.frame_lock:
@@ -529,41 +518,6 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
-        """Return the most recent frame captured immediately (Peeking).
-
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
-
-        Returns:
-            NDArray[Any]: The frame image (numpy array).
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        with self.frame_lock:
-            frame = self.latest_frame
-            timestamp = self.latest_timestamp
-
-        if frame is None or timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
-
-        age_ms = (time.perf_counter() - timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
-
-        return frame
-
     def disconnect(self) -> None:
         """
         Disconnects from the camera and cleans up resources.
@@ -584,9 +538,4 @@ class OpenCVCamera(Camera):
             self.videocapture.release()
             self.videocapture = None
 
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/opencv/configuration_opencv.py

@@ -15,9 +15,9 @@
 from dataclasses import dataclass
 from pathlib import Path
 
-from ..configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
+from ..configs import CameraConfig, ColorMode, Cv2Rotation
 
-__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation", "Cv2Backends"]
+__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
 
 
 @CameraConfig.register_subclass("opencv")
@@ -50,7 +50,6 @@ class OpenCVCameraConfig(CameraConfig):
         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).
-        backend: OpenCV backend identifier (https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html). Defaults to ANY.
 
     Note:
         - Only 3-channel color output (RGB/BGR) is currently supported.
@@ -63,12 +62,22 @@ class OpenCVCameraConfig(CameraConfig):
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
     fourcc: str | None = None
-    backend: Cv2Backends = Cv2Backends.ANY
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
-        self.rotation = Cv2Rotation(self.rotation)
-        self.backend = Cv2Backends(self.backend)
+        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."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            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(

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

@@ -74,4 +74,7 @@ class Reachy2CameraConfig(CameraConfig):
                 f"`image_type` is expected to be 'left' or 'right' for teleop camera, and 'rgb' or 'depth' for depth camera, but {self.image_type} is provided."
             )
 
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -32,7 +32,6 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 
-from lerobot.utils.decorators import check_if_not_connected
 from lerobot.utils.import_utils import _reachy2_sdk_available
 
 if TYPE_CHECKING or _reachy2_sdk_available:
@@ -81,8 +80,6 @@ class Reachy2Camera(Camera):
         self.config = config
 
         self.color_mode = config.color_mode
-        self.latest_frame: NDArray[Any] | None = None
-        self.latest_timestamp: float | None = None
 
         self.cam_manager: CameraManager | None = None
 
@@ -124,12 +121,16 @@ class Reachy2Camera(Camera):
         """
         raise NotImplementedError("Camera detection is not implemented for Reachy2 cameras.")
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
 
-        This method retrieves the most recent frame available in Reachy 2's low-level software.
+        This is a blocking call.
+
+        Args:
+            color_mode (Optional[ColorMode]): If specified, overrides the default
+                color mode (`self.color_mode`) for this read operation (e.g.,
+                request RGB even if default is BGR).
 
         Returns:
             np.ndarray: The captured frame as a NumPy array in the format
@@ -138,13 +139,11 @@ class Reachy2Camera(Camera):
         """
         start_time = time.perf_counter()
 
-        if self.cam_manager is None:
+        if not self.is_connected:
             raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        if color_mode is not None:
-            logger.warning(
-                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
-            )
+        if self.cam_manager is None:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         frame: NDArray[Any] = np.empty((0, 0, 3), dtype=np.uint8)
 
@@ -166,27 +165,25 @@ class Reachy2Camera(Camera):
             raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")
 
         if frame is None:
-            raise RuntimeError(f"Internal error: No frame available for {self}.")
+            return np.empty((0, 0, 3), dtype=np.uint8)
 
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
-            )
-        if self.color_mode == ColorMode.RGB:
+        if self.config.color_mode == "rgb":
             frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
 
-        self.latest_frame = frame
-        self.latest_timestamp = time.perf_counter()
-
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
         return frame
 
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
-        Same as read()
+        Reads the latest available frame.
+
+        This method retrieves the most recent frame available in Reachy 2's low-level software.
+
+        Args:
+            timeout_ms (float): Maximum time in milliseconds to wait for a frame
+                to become available. Defaults to 200ms (0.2 seconds).
 
         Returns:
             np.ndarray: The latest captured frame as a NumPy array in the format
@@ -197,40 +194,16 @@ class Reachy2Camera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        return self.read()
+        frame = self.read()
 
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
-        """Return the most recent frame captured immediately (Peeking).
+        if frame is None:
+            raise RuntimeError(f"Internal error: No frame available for {self}.")
 
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+        return frame
 
-        Returns:
-            tuple[NDArray, float]:
-                - The frame image (numpy array).
-                - The timestamp (time.perf_counter) when this frame was captured.
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-
-        if self.latest_frame is None or self.latest_timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
-
-        age_ms = (time.perf_counter() - self.latest_timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
-
-        return self.latest_frame
-
-    @check_if_not_connected
     def disconnect(self) -> None:
         """
         Stops the background read thread (if running).
@@ -238,6 +211,8 @@ class Reachy2Camera(Camera):
         Raises:
             DeviceNotConnectedError: If the camera is already disconnected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} not connected.")
 
         if self.cam_manager is not None:
             self.cam_manager.disconnect()

src/lerobot/cameras/realsense/camera_realsense.py

@@ -30,8 +30,7 @@ try:
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -73,14 +72,15 @@ class RealSenseCamera(Camera):
         camera = RealSenseCamera(config)
         camera.connect()
 
-        # Read 1 frame synchronously (blocking)
+        # Read 1 frame synchronously
         color_image = camera.read()
+        print(color_image.shape)
 
-        # Read 1 frame asynchronously (waits for new frame with a timeout)
+        # Read 1 frame asynchronously
         async_image = camera.async_read()
 
-        # Get the latest frame immediately (no wait, returns timestamp)
-        latest_image, timestamp = camera.read_latest()
+        # When done, properly disconnect the camera using
+        camera.disconnect()
 
         # Example with depth capture and custom settings
         custom_config = RealSenseCameraConfig(
@@ -133,9 +133,7 @@ class RealSenseCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_color_frame: NDArray[Any] | None = None
-        self.latest_depth_frame: NDArray[Any] | None = None
-        self.latest_timestamp: float | None = None
+        self.latest_frame: NDArray[Any] | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -153,7 +151,6 @@ 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
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the RealSense camera specified in the configuration.
@@ -161,16 +158,14 @@ class RealSenseCamera(Camera):
         Initializes the RealSense pipeline, configures the required streams (color
         and optionally depth), starts the pipeline, and validates the actual stream settings.
 
-        Args:
-            warmup (bool): If True, waits at connect() time until at least one valid frame
-                           has been captured by the background thread. Defaults to True.
-
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
             ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
             ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
             RuntimeError: If the pipeline starts but fails to apply requested settings.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         self.rs_pipeline = rs.pipeline()
         rs_config = rs.config()
@@ -186,18 +181,15 @@ class RealSenseCamera(Camera):
             ) from e
 
         self._configure_capture_settings()
-        self._start_read_thread()
 
-        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
-        self.warmup_s = max(self.warmup_s, 1)
-
-        start_time = time.time()
-        while time.time() - start_time < self.warmup_s:
-            self.async_read(timeout_ms=self.warmup_s * 1000)
-            time.sleep(0.1)
-        with self.frame_lock:
-            if self.latest_color_frame is None or self.use_depth and self.latest_depth_frame is None:
-                raise ConnectionError(f"{self} failed to capture frames during warmup.")
+        if warmup:
+            time.sleep(
+                1
+            )  # NOTE(Steven): RS cameras need a bit of time to warm up before the first read. If we don't wait, the first read from the warmup will raise.
+            start_time = time.time()
+            while time.time() - start_time < self.warmup_s:
+                self.read()
+                time.sleep(0.1)
 
         logger.info(f"{self} connected.")
 
@@ -290,7 +282,6 @@ class RealSenseCamera(Camera):
             if self.use_depth:
                 rs_config.enable_stream(rs.stream.depth)
 
-    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """Sets fps, width, and height from device stream if not already configured.
 
@@ -300,6 +291,8 @@ class RealSenseCamera(Camera):
         Raises:
             DeviceNotConnectedError: If device is not connected.
         """
+        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.")
@@ -319,7 +312,6 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
-    @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (depth) synchronously from the camera.
@@ -327,6 +319,9 @@ class RealSenseCamera(Camera):
         This is a blocking call. It waits for a coherent set of frames (depth)
         from the camera hardware via the RealSense pipeline.
 
+        Args:
+            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
+
         Returns:
             np.ndarray: The depth map as a NumPy array (height, width)
                   of type `np.uint16` (raw depth values in millimeters) and rotation.
@@ -335,50 +330,44 @@ class RealSenseCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If reading frames from the pipeline fails or frames are invalid.
         """
-        if timeout_ms:
-            logger.warning(
-                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
-            )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
         if not self.use_depth:
             raise RuntimeError(
                 f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
             )
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        start_time = time.perf_counter()
 
-        self.new_frame_event.clear()
-
-        _ = self.async_read(timeout_ms=10000)
-
-        with self.frame_lock:
-            depth_map = self.latest_depth_frame
-
-        if depth_map is None:
-            raise RuntimeError("No depth frame available. Ensure camera is streaming.")
-
-        return depth_map
-
-    def _read_from_hardware(self):
         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=10000)
+        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
         if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
+            raise RuntimeError(f"{self} read_depth failed (status={ret}).")
 
-        return frame
+        depth_frame = frame.get_depth_frame()
+        depth_map = np.asanyarray(depth_frame.get_data())
 
-    @check_if_not_connected
-    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 0) -> NDArray[Any]:
+        depth_map_processed = self._postprocess_image(depth_map, depth_frame=True)
+
+        read_duration_ms = (time.perf_counter() - start_time) * 1e3
+        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
+
+        return depth_map_processed
+
+    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (color) synchronously from the camera.
 
         This is a blocking call. It waits for a coherent set of frames (color)
         from the camera hardware via the RealSense pipeline.
 
+        Args:
+            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
+
         Returns:
             np.ndarray: The captured color frame as a NumPy array
               (height, width, channels), processed according to `color_mode` and rotation.
@@ -389,36 +378,39 @@ class RealSenseCamera(Camera):
             ValueError: If an invalid `color_mode` is requested.
         """
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         start_time = time.perf_counter()
 
-        if color_mode is not None:
-            logger.warning(
-                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
-            )
+        if self.rs_pipeline is None:
+            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
 
-        if timeout_ms:
-            logger.warning(
-                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
-            )
+        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        if not ret or frame is None:
+            raise RuntimeError(f"{self} read failed (status={ret}).")
 
-        self.new_frame_event.clear()
+        color_frame = frame.get_color_frame()
+        color_image_raw = np.asanyarray(color_frame.get_data())
 
-        frame = self.async_read(timeout_ms=10000)
+        color_image_processed = self._postprocess_image(color_image_raw, color_mode)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
-        return frame
+        return color_image_processed
 
-    def _postprocess_image(self, image: NDArray[Any], depth_frame: bool = False) -> NDArray[Any]:
+    def _postprocess_image(
+        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.
 
         Args:
             image (np.ndarray): The raw image frame (expected RGB format from RealSense).
+            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
+                                             uses the instance's default `self.color_mode`.
 
         Returns:
             np.ndarray: The processed image frame according to `self.color_mode` and `self.rotation`.
@@ -429,9 +421,9 @@ class RealSenseCamera(Camera):
                           `width` and `height`.
         """
 
-        if self.color_mode and self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+        if color_mode and color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
-                f"Invalid requested color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+                f"Invalid requested color mode '{color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
 
         if depth_frame:
@@ -462,7 +454,7 @@ class RealSenseCamera(Camera):
 
         On each iteration:
         1. Reads a color frame with 500ms timeout
-        2. Stores result in latest_frame and updates timestamp (thread-safe)
+        2. Stores result in latest_frame (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -470,41 +462,25 @@ class RealSenseCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
-        failure_count = 0
         while not self.stop_event.is_set():
             try:
-                frame = self._read_from_hardware()
-                color_frame_raw = frame.get_color_frame()
-                color_frame = np.asanyarray(color_frame_raw.get_data())
-                processed_color_frame = self._postprocess_image(color_frame)
-
-                if self.use_depth:
-                    depth_frame_raw = frame.get_depth_frame()
-                    depth_frame = np.asanyarray(depth_frame_raw.get_data())
-                    processed_depth_frame = self._postprocess_image(depth_frame, depth_frame=True)
-
-                capture_time = time.perf_counter()
+                color_image = self.read(timeout_ms=500)
 
                 with self.frame_lock:
-                    self.latest_color_frame = processed_color_frame
-                    if self.use_depth:
-                        self.latest_depth_frame = processed_depth_frame
-                    self.latest_timestamp = capture_time
+                    self.latest_frame = color_image
                 self.new_frame_event.set()
-                failure_count = 0
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
-                if failure_count <= 10:
-                    failure_count += 1
-                    logger.warning(f"Error reading frame in background thread for {self}: {e}")
-                else:
-                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
+                logger.warning(f"Error reading frame in background thread for {self}: {e}")
 
     def _start_read_thread(self) -> None:
         """Starts or restarts the background read thread if it's not running."""
-        self._stop_read_thread()
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=0.1)
+        if self.stop_event is not None:
+            self.stop_event.set()
 
         self.stop_event = Event()
         self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
@@ -522,14 +498,7 @@ class RealSenseCamera(Camera):
         self.thread = None
         self.stop_event = None
 
-        with self.frame_lock:
-            self.latest_color_frame = None
-            self.latest_depth_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
     # NOTE(Steven): Missing implementation for depth for now
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame data (color) asynchronously.
@@ -537,7 +506,6 @@ class RealSenseCamera(Camera):
         This method retrieves the most recent color frame captured by the background
         read thread. It does not block waiting for the camera hardware directly,
         but may wait up to timeout_ms for the background thread to provide a frame.
-        It is “best effort” under high FPS.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -552,18 +520,21 @@ class RealSenseCamera(Camera):
             TimeoutError: If no frame data becomes available within the specified timeout.
             RuntimeError: If the background thread died unexpectedly or another error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+            self._start_read_thread()
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            thread_alive = self.thread is not None and self.thread.is_alive()
             raise TimeoutError(
                 f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {self.thread.is_alive()}."
+                f"Read thread alive: {thread_alive}."
             )
 
         with self.frame_lock:
-            frame = self.latest_color_frame
+            frame = self.latest_frame
             self.new_frame_event.clear()
 
         if frame is None:
@@ -571,42 +542,6 @@ class RealSenseCamera(Camera):
 
         return frame
 
-    # NOTE(Steven): Missing implementation for depth for now
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 500) -> NDArray[Any]:
-        """Return the most recent (color) frame captured immediately (Peeking).
-
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
-
-        Returns:
-            NDArray[Any]: The frame image (numpy array).
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        with self.frame_lock:
-            frame = self.latest_color_frame
-            timestamp = self.latest_timestamp
-
-        if frame is None or timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
-
-        age_ms = (time.perf_counter() - timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
-
-        return frame
-
     def disconnect(self) -> None:
         """
         Disconnects from the camera, stops the pipeline, and cleans up resources.
@@ -630,10 +565,4 @@ class RealSenseCamera(Camera):
             self.rs_pipeline = None
             self.rs_profile = None
 
-        with self.frame_lock:
-            self.latest_color_frame = None
-            self.latest_depth_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -60,8 +60,20 @@ class RealSenseCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
-        self.rotation = Cv2Rotation(self.rotation)
+        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."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            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."
+            )
 
         values = (self.fps, self.width, self.height)
         if any(v is not None for v in values) and any(v is None for v in values):

src/lerobot/cameras/utils.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import platform
 from typing import cast
 
 from lerobot.utils.import_utils import make_device_from_device_class
@@ -67,3 +68,14 @@ def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
         return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
     else:
         return None
+
+
+def get_cv2_backend() -> int:
+    import cv2
+
+    if platform.system() == "Windows":
+        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 int(cv2.CAP_ANY)

src/lerobot/cameras/zmq/camera_zmq.py

@@ -34,8 +34,7 @@ import cv2
 import numpy as np
 from numpy.typing import NDArray
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -46,12 +45,6 @@ logger = logging.getLogger(__name__)
 
 class ZMQCamera(Camera):
     """
-    Manages camera interactions via ZeroMQ for receiving frames from a remote server.
-
-    This class connects to a ZMQ Publisher, subscribes to frame topics, and decodes
-    incoming JSON messages containing Base64 encoded images. It supports both
-    synchronous and asynchronous frame reading patterns.
-
     Example usage:
         ```python
         from lerobot.cameras.zmq import ZMQCamera, ZMQCameraConfig
@@ -59,16 +52,7 @@ class ZMQCamera(Camera):
         config = ZMQCameraConfig(server_address="192.168.123.164", port=5555, camera_name="head_camera")
         camera = ZMQCamera(config)
         camera.connect()
-
-        # Read 1 frame synchronously (blocking)
-        color_image = camera.read()
-
-        # Read 1 frame asynchronously (waits for new frame with a timeout)
-        async_image = camera.async_read()
-
-        # Get the latest frame immediately (no wait, returns timestamp)
-        latest_image, timestamp = camera.read_latest()
-
+        frame = camera.read()
         camera.disconnect()
         ```
     """
@@ -84,17 +68,14 @@ class ZMQCamera(Camera):
         self.color_mode = config.color_mode
         self.timeout_ms = config.timeout_ms
 
-        # ZMQ Context and Socket
         self.context: zmq.Context | None = None
         self.socket: zmq.Socket | None = None
         self._connected = False
 
-        # Threading resources
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
         self.latest_frame: NDArray[Any] | None = None
-        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
     def __str__(self) -> str:
@@ -102,17 +83,12 @@ class ZMQCamera(Camera):
 
     @property
     def is_connected(self) -> bool:
-        """Checks if the ZMQ socket is initialized and connected."""
         return self._connected and self.context is not None and self.socket is not None
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
-        """Connect to ZMQ camera server.
-
-        Args:
-            warmup (bool): If True, waits for the camera to provide at least one
-                           valid frame before returning. Defaults to True.
-        """
+        """Connect to ZMQ camera server."""
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         logger.info(f"Connecting to {self}...")
 
@@ -127,28 +103,17 @@ class ZMQCamera(Camera):
             self.socket.connect(f"tcp://{self.server_address}:{self.port}")
             self._connected = True
 
-            # Auto-detect resolution if not provided
+            # Auto-detect resolution
             if self.width is None or self.height is None:
-                # Read directly from hardware because the thread isn't running yet
-                temp_frame = self._read_from_hardware()
-                h, w = temp_frame.shape[:2]
+                h, w = self.read().shape[:2]
                 self.height = h
                 self.width = w
-                logger.info(f"{self} resolution detected: {w}x{h}")
+                logger.info(f"{self} resolution: {w}x{h}")
 
-            self._start_read_thread()
             logger.info(f"{self} connected.")
 
             if warmup:
-                # Ensure we have captured at least one frame via the thread
-                start_time = time.time()
-                while time.time() - start_time < (self.config.warmup_s):  # Wait a bit more than timeout
-                    self.async_read(timeout_ms=self.config.warmup_s * 1000)
-                    time.sleep(0.1)
-
-                with self.frame_lock:
-                    if self.latest_frame is None:
-                        raise ConnectionError(f"{self} failed to capture frames during warmup.")
+                time.sleep(0.1)
 
         except Exception as e:
             self._cleanup()
@@ -166,14 +131,15 @@ class ZMQCamera(Camera):
 
     @staticmethod
     def find_cameras() -> list[dict[str, Any]]:
-        """
-        Detection not implemented for ZMQ cameras. These cameras require manual configuration (server address/port).
-        """
-        raise NotImplementedError("Camera detection is not implemented for ZMQ cameras.")
+        """ZMQ cameras require manual configuration (server address/port)."""
+        return []
 
-    def _read_from_hardware(self) -> NDArray[Any]:
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
-        Reads a single frame directly from the ZMQ socket.
+        Read a single frame from the ZMQ camera.
+
+        Returns:
+            np.ndarray: Decoded frame (height, width, 3)
         """
         if not self.is_connected or self.socket is None:
             raise DeviceNotConnectedError(f"{self} is not connected.")
@@ -181,7 +147,6 @@ class ZMQCamera(Camera):
         try:
             message = self.socket.recv_string()
         except Exception as e:
-            # Check for ZMQ timeout (EAGAIN/Again) without requiring global zmq import
             if type(e).__name__ == "Again":
                 raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
             raise
@@ -211,114 +176,42 @@ class ZMQCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
-    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
-        """
-        Reads a single frame synchronously from the camera.
-
-        This is a blocking call. It waits for the next available frame from the
-        camera background thread.
-
-        Returns:
-            np.ndarray: Decoded frame (height, width, 3)
-        """
-        start_time = time.perf_counter()
-
-        if color_mode is not None:
-            logger.warning(
-                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
-            )
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        self.new_frame_event.clear()
-        frame = self.async_read(timeout_ms=10000)
-
-        read_duration_ms = (time.perf_counter() - start_time) * 1e3
-        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
-
-        return frame
-
     def _read_loop(self) -> None:
-        """
-        Internal loop run by the background thread for asynchronous reading.
-        """
-        if self.stop_event is None:
-            raise RuntimeError(f"{self}: stop_event is not initialized.")
-
-        failure_count = 0
-        while not self.stop_event.is_set():
+        while self.stop_event and not self.stop_event.is_set():
             try:
-                frame = self._read_from_hardware()
-                capture_time = time.perf_counter()
-
+                frame = self.read()
                 with self.frame_lock:
                     self.latest_frame = frame
-                    self.latest_timestamp = capture_time
                 self.new_frame_event.set()
-                failure_count = 0
-
             except DeviceNotConnectedError:
                 break
-            except (TimeoutError, Exception) as e:
-                if failure_count <= 10:
-                    failure_count += 1
-                    logger.warning(f"Read error: {e}")
-                else:
-                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
+            except TimeoutError:
+                pass
+            except Exception as e:
+                logger.warning(f"Read error: {e}")
 
     def _start_read_thread(self) -> None:
-        if self.stop_event is not None:
-            self.stop_event.set()
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=2.0)
-
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
+        if self.thread and self.thread.is_alive():
+            return
         self.stop_event = Event()
-        self.thread = Thread(target=self._read_loop, daemon=True, name=f"{self}_read_loop")
+        self.thread = Thread(target=self._read_loop, daemon=True)
         self.thread.start()
-        time.sleep(0.1)
 
     def _stop_read_thread(self) -> None:
-        if self.stop_event is not None:
+        if self.stop_event:
             self.stop_event.set()
-
-        if self.thread is not None and self.thread.is_alive():
+        if self.thread and self.thread.is_alive():
             self.thread.join(timeout=2.0)
-
         self.thread = None
         self.stop_event = None
 
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
+    def async_read(self, timeout_ms: float = 10000) -> NDArray[Any]:
+        """Read latest frame asynchronously (non-blocking)."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
-    @check_if_not_connected
-    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
-        """
-        Reads the latest available frame asynchronously.
-
-        Args:
-            timeout_ms (float): Maximum time in milliseconds to wait for a frame
-                to become available. Defaults to 200ms.
-
-        Returns:
-            np.ndarray: The latest captured frame.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is not connected.
-            TimeoutError: If no frame data becomes available within the specified timeout.
-            RuntimeError: If the background thread is not running.
-        """
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        if not self.thread or not self.thread.is_alive():
+            self._start_read_thread()
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
             raise TimeoutError(f"{self} async_read timeout after {timeout_ms}ms")
@@ -332,54 +225,11 @@ class ZMQCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-        """Return the most recent frame captured immediately (Peeking).
-
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
-
-        Returns:
-            NDArray[Any]: The frame image (numpy array).
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        with self.frame_lock:
-            frame = self.latest_frame
-            timestamp = self.latest_timestamp
-
-        if frame is None or timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
-
-        age_ms = (time.perf_counter() - timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
-
-        return frame
-
     def disconnect(self) -> None:
         """Disconnect from ZMQ camera."""
-        if not self.is_connected and self.thread is None:
+        if not self.is_connected and not self.thread:
             raise DeviceNotConnectedError(f"{self} not connected.")
 
-        if self.thread is not None:
-            self._stop_read_thread()
-
+        self._stop_read_thread()
         self._cleanup()
-
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/zmq/configuration_zmq.py

@@ -29,10 +29,12 @@ class ZMQCameraConfig(CameraConfig):
     camera_name: str = "zmq_camera"
     color_mode: ColorMode = ColorMode.RGB
     timeout_ms: int = 5000
-    warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
+        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."
+            )
 
         if self.timeout_ms <= 0:
             raise ValueError(f"`timeout_ms` must be positive, but {self.timeout_ms} is provided.")

src/lerobot/configs/default.py

@@ -27,7 +27,7 @@ class DatasetConfig:
     # "dataset_index" into the returned item. The index mapping is made according to the order in which the
     # datasets are provided.
     repo_id: str
-    # Root directory where the dataset will be stored (e.g. 'dataset/path'). If None, defaults to $HF_LEROBOT_HOME/repo_id.
+    # Root directory where the dataset will be stored (e.g. 'dataset/path').
     root: str | None = None
     episodes: list[int] | None = None
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)

src/lerobot/configs/policies.py

@@ -45,12 +45,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy.
+        output_shapes: A dictionary defining the shapes of the output data for the policy.
+        input_normalization_modes: A dictionary with key representing the modality and the value specifies the
+            normalization mode to apply.
+        output_normalization_modes: Similar dictionary as `input_normalization_modes`, but to unnormalize to
+            the original scale.
     """
 
     n_obs_steps: int = 1
@@ -151,6 +151,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
             return {}
         return {key: ft for key, ft in self.input_features.items() if ft.type is FeatureType.VISUAL}
 
+    @property
+    def audio_features(self) -> dict[str, PolicyFeature]:
+        if not self.input_features:
+            return {}
+        return {key: ft for key, ft in self.input_features.items() if ft.type is FeatureType.AUDIO}
+
     @property
     def action_feature(self) -> PolicyFeature | None:
         if not self.output_features:

src/lerobot/configs/types.py

@@ -20,6 +20,7 @@ from enum import Enum
 class FeatureType(str, Enum):
     STATE = "STATE"
     VISUAL = "VISUAL"
+    AUDIO = "AUDIO"
     ENV = "ENV"
     ACTION = "ACTION"
     REWARD = "REWARD"

src/lerobot/datasets/aggregate.py

@@ -26,6 +26,8 @@ import tqdm
 from lerobot.datasets.compute_stats import aggregate_stats
 from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
+    DEFAULT_AUDIO_FILE_SIZE_IN_MB,
+    DEFAULT_AUDIO_PATH,
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
@@ -41,7 +43,7 @@ from lerobot.datasets.utils import (
     write_stats,
     write_tasks,
 )
-from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
+from lerobot.datasets.video_utils import concatenate_media_files, get_media_duration_in_s
 
 
 def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
@@ -110,72 +112,28 @@ def update_meta_data(
     meta_idx,
     data_idx,
     videos_idx,
+    audios_idx,
 ):
     """Updates metadata DataFrame with new chunk, file, and timestamp indices.
 
     Adjusts all indices and timestamps to account for previously aggregated
     data and videos in the destination dataset.
 
-    For data file indices, uses the 'src_to_dst' mapping from aggregate_data()
-    to correctly map source file indices to their destination locations.
-
     Args:
         df: DataFrame containing the metadata to be updated.
         dst_meta: Destination dataset metadata.
         meta_idx: Dictionary containing current metadata chunk and file indices.
         data_idx: Dictionary containing current data chunk and file indices.
         videos_idx: Dictionary containing current video indices and timestamps.
-
+        audios_idx: Dictionary containing current audio indices and timestamps.
     Returns:
         pd.DataFrame: Updated DataFrame with adjusted indices and timestamps.
     """
 
     df["meta/episodes/chunk_index"] = df["meta/episodes/chunk_index"] + meta_idx["chunk"]
     df["meta/episodes/file_index"] = df["meta/episodes/file_index"] + meta_idx["file"]
-
-    # Update data file indices using source-to-destination mapping
-    # This is critical for handling datasets that are already results of a merge
-    data_src_to_dst = data_idx.get("src_to_dst", {})
-    if data_src_to_dst:
-        # Store original indices for lookup
-        df["_orig_data_chunk"] = df["data/chunk_index"].copy()
-        df["_orig_data_file"] = df["data/file_index"].copy()
-
-        # Vectorized mapping from (src_chunk, src_file) to (dst_chunk, dst_file)
-        # This is much faster than per-row iteration for large metadata tables
-        mapping_index = pd.MultiIndex.from_tuples(
-            list(data_src_to_dst.keys()),
-            names=["chunk_index", "file_index"],
-        )
-        mapping_values = list(data_src_to_dst.values())
-        mapping_df = pd.DataFrame(
-            mapping_values,
-            index=mapping_index,
-            columns=["dst_chunk", "dst_file"],
-        )
-
-        # Construct a MultiIndex for each row based on original data indices
-        row_index = pd.MultiIndex.from_arrays(
-            [df["_orig_data_chunk"], df["_orig_data_file"]],
-            names=["chunk_index", "file_index"],
-        )
-
-        # Align mapping to rows; missing keys fall back to the default destination
-        reindexed = mapping_df.reindex(row_index)
-        reindexed[["dst_chunk", "dst_file"]] = reindexed[["dst_chunk", "dst_file"]].fillna(
-            {"dst_chunk": data_idx["chunk"], "dst_file": data_idx["file"]}
-        )
-
-        # Assign mapped destination indices back to the DataFrame
-        df["data/chunk_index"] = reindexed["dst_chunk"].to_numpy()
-        df["data/file_index"] = reindexed["dst_file"].to_numpy()
-
-        # Clean up temporary columns
-        df = df.drop(columns=["_orig_data_chunk", "_orig_data_file"])
-    else:
-        # Fallback to simple offset (backward compatibility for single-file sources)
-        df["data/chunk_index"] = df["data/chunk_index"] + data_idx["chunk"]
-        df["data/file_index"] = df["data/file_index"] + data_idx["file"]
+    df["data/chunk_index"] = df["data/chunk_index"] + data_idx["chunk"]
+    df["data/file_index"] = df["data/file_index"] + data_idx["file"]
     for key, video_idx in videos_idx.items():
         # Store original video file indices before updating
         orig_chunk_col = f"videos/{key}/chunk_index"
@@ -191,7 +149,8 @@ def update_meta_data(
         if src_to_dst:
             # Map each episode to its correct destination file and apply offset
             for idx in df.index:
-                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
+                # Convert to Python int to avoid numpy type mismatch in dict lookup
+                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
 
                 # Get destination chunk/file for this source file
                 dst_chunk, dst_file = src_to_dst.get(src_key, (video_idx["chunk"], video_idx["file"]))
@@ -207,7 +166,8 @@ def update_meta_data(
             df[orig_chunk_col] = video_idx["chunk"]
             df[orig_file_col] = video_idx["file"]
             for idx in df.index:
-                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
+                # Convert to Python int to avoid numpy type mismatch in dict lookup
+                src_key = (int(df.at[idx, "_orig_chunk"]), int(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
@@ -223,6 +183,36 @@ def update_meta_data(
         # Clean up temporary columns
         df = df.drop(columns=["_orig_chunk", "_orig_file"])
 
+    for key, audio_idx in audios_idx.items():
+        # Store original audio file indices before updating
+        orig_chunk_col = f"audio/{key}/chunk_index"
+        orig_file_col = f"audio/{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] = audio_idx["chunk"]
+        df[orig_file_col] = audio_idx["file"]
+
+        # Apply per-source-file timestamp offsets
+        src_to_offset = audio_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"audio/{key}/from_timestamp"] += offset
+                df.at[idx, f"audio/{key}/to_timestamp"] += offset
+        else:
+            # Fallback to simple offset (for backward compatibility)
+            df[f"audio/{key}/from_timestamp"] = (
+                df[f"audio/{key}/from_timestamp"] + audio_idx["latest_duration"]
+            )
+            df[f"audio/{key}/to_timestamp"] = df[f"audio/{key}/to_timestamp"] + audio_idx["latest_duration"]
+
+        # Clean up temporary columns
+        df = df.drop(columns=["_orig_chunk", "_orig_file"])
+
     df["dataset_from_index"] = df["dataset_from_index"] + dst_meta.info["total_frames"]
     df["dataset_to_index"] = df["dataset_to_index"] + dst_meta.info["total_frames"]
     df["episode_index"] = df["episode_index"] + dst_meta.info["total_episodes"]
@@ -237,6 +227,7 @@ def aggregate_datasets(
     aggr_root: Path | None = None,
     data_files_size_in_mb: float | None = None,
     video_files_size_in_mb: float | None = None,
+    audio_files_size_in_mb: float | None = None,
     chunk_size: int | None = None,
 ):
     """Aggregates multiple LeRobot datasets into a single unified dataset.
@@ -254,6 +245,7 @@ def aggregate_datasets(
         aggr_root: Optional root path for the aggregated dataset.
         data_files_size_in_mb: Maximum size for data files in MB (defaults to DEFAULT_DATA_FILE_SIZE_IN_MB)
         video_files_size_in_mb: Maximum size for video files in MB (defaults to DEFAULT_VIDEO_FILE_SIZE_IN_MB)
+        audio_files_size_in_mb: Maximum size for audio files in MB (defaults to DEFAULT_AUDIO_FILE_SIZE_IN_MB)
         chunk_size: Maximum number of files per chunk (defaults to DEFAULT_CHUNK_SIZE)
     """
     logging.info("Start aggregate_datasets")
@@ -262,6 +254,8 @@ def aggregate_datasets(
         data_files_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
     if video_files_size_in_mb is None:
         video_files_size_in_mb = DEFAULT_VIDEO_FILE_SIZE_IN_MB
+    if audio_files_size_in_mb is None:
+        audio_files_size_in_mb = DEFAULT_AUDIO_FILE_SIZE_IN_MB
     if chunk_size is None:
         chunk_size = DEFAULT_CHUNK_SIZE
 
@@ -274,6 +268,7 @@ def aggregate_datasets(
     )
     fps, robot_type, features = validate_all_metadata(all_metadata)
     video_keys = [key for key in features if features[key]["dtype"] == "video"]
+    audio_keys = [key for key in features if features[key]["dtype"] == "audio"]
 
     dst_meta = LeRobotDatasetMetadata.create(
         repo_id=aggr_repo_id,
@@ -285,6 +280,7 @@ def aggregate_datasets(
         chunks_size=chunk_size,
         data_files_size_in_mb=data_files_size_in_mb,
         video_files_size_in_mb=video_files_size_in_mb,
+        audio_files_size_in_mb=audio_files_size_in_mb,
     )
 
     logging.info("Find all tasks")
@@ -296,18 +292,18 @@ def aggregate_datasets(
     videos_idx = {
         key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in video_keys
     }
+    audios_idx = {
+        key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in audio_keys
+    }
 
     dst_meta.episodes = {}
 
     for src_meta in tqdm.tqdm(all_metadata, desc="Copy data and videos"):
         videos_idx = aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chunk_size)
+        audios_idx = aggregate_audio(src_meta, dst_meta, audios_idx, audio_files_size_in_mb, chunk_size)
         data_idx = aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_size)
 
-        meta_idx = aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx)
-
-        # Clear the src_to_dst mapping after processing each source dataset
-        # to avoid interference between different source datasets
-        data_idx.pop("src_to_dst", None)
+        meta_idx = aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx, audios_idx)
 
         dst_meta.info["total_episodes"] += src_meta.total_episodes
         dst_meta.info["total_frames"] += src_meta.total_frames
@@ -359,6 +355,10 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
         dst_file_durations = video_idx["dst_file_durations"]
 
         for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
+            # Convert to Python int to ensure consistent dict keys
+            src_chunk_idx = int(src_chunk_idx)
+            src_file_idx = int(src_file_idx)
+
             src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
                 video_key=key,
                 chunk_index=src_chunk_idx,
@@ -371,7 +371,7 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                 file_index=file_idx,
             )
 
-            src_duration = get_video_duration_in_s(src_path)
+            src_duration = get_media_duration_in_s(src_path, media_type="video")
             dst_key = (chunk_idx, file_idx)
 
             if not dst_path.exists():
@@ -410,7 +410,7 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                 current_dst_duration = dst_file_durations.get(dst_key, 0)
                 videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_dst_duration
                 videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
-                concatenate_video_files(
+                concatenate_media_files(
                     [dst_path, src_path],
                     dst_path,
                 )
@@ -425,22 +425,111 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
     return videos_idx
 
 
+def aggregate_audio(src_meta, dst_meta, audios_idx, audio_files_size_in_mb, chunk_size):
+    """Aggregates audio files from a source dataset into the destination dataset.
+
+    Handles audio file concatenation and rotation based on file size limits.
+    Creates new audio files when size limits are exceeded.
+
+    Args:
+        src_meta: Source dataset metadata.
+        dst_meta: Destination dataset metadata.
+        audio_idx: Dictionary tracking audio chunk and file indices.
+        audio_files_size_in_mb: Maximum size for audio files in MB (defaults to DEFAULT_AUDIO_FILE_SIZE_IN_MB)
+        chunk_size: Maximum number of files per chunk (defaults to DEFAULT_CHUNK_SIZE)
+
+    Returns:
+        dict: Updated audio_idx with current chunk and file indices.
+    """
+    for key in audios_idx:
+        audios_idx[key]["episode_duration"] = 0
+        # Track offset for each source (chunk, file) pair
+        audios_idx[key]["src_to_offset"] = {}
+
+    for key, audio_idx in audios_idx.items():
+        unique_chunk_file_pairs = {
+            (chunk, file)
+            for chunk, file in zip(
+                src_meta.episodes[f"audio/{key}/chunk_index"],
+                src_meta.episodes[f"audio/{key}/file_index"],
+                strict=False,
+            )
+        }
+        unique_chunk_file_pairs = sorted(unique_chunk_file_pairs)
+
+        chunk_idx = audio_idx["chunk"]
+        file_idx = audio_idx["file"]
+        current_offset = audio_idx["latest_duration"]
+
+        for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
+            src_path = src_meta.root / DEFAULT_AUDIO_PATH.format(
+                audio_key=key,
+                chunk_index=src_chunk_idx,
+                file_index=src_file_idx,
+            )
+
+            dst_path = dst_meta.root / DEFAULT_AUDIO_PATH.format(
+                audio_key=key,
+                chunk_index=chunk_idx,
+                file_index=file_idx,
+            )
+
+            src_duration = get_media_duration_in_s(src_path, media_type="audio")
+
+            if not dst_path.exists():
+                # Store offset before incrementing
+                audios_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))
+                audios_idx[key]["episode_duration"] += src_duration
+                current_offset += src_duration
+                continue
+
+            # Check file sizes before appending
+            src_size = get_file_size_in_mb(src_path)
+            dst_size = get_file_size_in_mb(dst_path)
+
+            if dst_size + src_size >= audio_files_size_in_mb:
+                # Rotate to a new file, this source becomes start of new destination
+                # So its offset should be 0
+                audios_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_AUDIO_PATH.format(
+                    audio_key=key,
+                    chunk_index=chunk_idx,
+                    file_index=file_idx,
+                )
+                dst_path.parent.mkdir(parents=True, exist_ok=True)
+                shutil.copy(str(src_path), str(dst_path))
+                # Reset offset for next file
+                current_offset = src_duration
+            else:
+                # Append to existing video file - use current accumulated offset
+                audios_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
+                concatenate_media_files(
+                    [dst_path, src_path],
+                    dst_path,
+                )
+                current_offset += src_duration
+
+            audios_idx[key]["episode_duration"] += src_duration
+
+        audios_idx[key]["chunk"] = chunk_idx
+        audios_idx[key]["file"] = file_idx
+
+    return audios_idx
+
+
 def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_size):
     """Aggregates data chunks from a source dataset into the destination dataset.
 
     Reads source data files, updates indices to match the aggregated dataset,
     and writes them to the destination with proper file rotation.
 
-    Tracks a `src_to_dst` mapping from source (chunk, file) to destination (chunk, file)
-    which is critical for correctly updating episode metadata when source datasets
-    have multiple data files (e.g., from a previous merge operation).
-
     Args:
         src_meta: Source dataset metadata.
         dst_meta: Destination dataset metadata.
         data_idx: Dictionary tracking data chunk and file indices.
-        data_files_size_in_mb: Maximum size for data files in MB.
-        chunk_size: Maximum number of files per chunk.
 
     Returns:
         dict: Updated data_idx with current chunk and file indices.
@@ -458,10 +547,6 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
     # retrieve features schema for proper image typing in parquet
     hf_features = get_hf_features_from_features(dst_meta.features) if contains_images else None
 
-    # Track source to destination file mapping for metadata update
-    # This is critical for handling datasets that are already results of a merge
-    src_to_dst: dict[tuple[int, int], tuple[int, int]] = {}
-
     for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
         src_path = src_meta.root / DEFAULT_DATA_PATH.format(
             chunk_index=src_chunk_idx, file_index=src_file_idx
@@ -474,9 +559,7 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             df = pd.read_parquet(src_path)
         df = update_data_df(df, src_meta, dst_meta)
 
-        # Write data and get the actual destination file it was written to
-        # This avoids duplicating the rotation logic here
-        data_idx, (dst_chunk, dst_file) = append_or_create_parquet_file(
+        data_idx = append_or_create_parquet_file(
             df,
             src_path,
             data_idx,
@@ -488,16 +571,10 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             hf_features=hf_features,
         )
 
-        # Record the mapping from source to actual destination
-        src_to_dst[(src_chunk_idx, src_file_idx)] = (dst_chunk, dst_file)
-
-    # Add the mapping to data_idx for use in metadata update
-    data_idx["src_to_dst"] = src_to_dst
-
     return data_idx
 
 
-def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
+def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx, audios_idx):
     """Aggregates metadata from a source dataset into the destination dataset.
 
     Reads source metadata files, updates all indices and timestamps,
@@ -509,6 +586,7 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
         meta_idx: Dictionary tracking metadata chunk and file indices.
         data_idx: Dictionary tracking data chunk and file indices.
         videos_idx: Dictionary tracking video indices and timestamps.
+        audios_idx: Dictionary tracking audio indices and timestamps.
 
     Returns:
         dict: Updated meta_idx with current chunk and file indices.
@@ -532,9 +610,10 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
             meta_idx,
             data_idx,
             videos_idx,
+            audios_idx,
         )
 
-        meta_idx, _ = append_or_create_parquet_file(
+        meta_idx = append_or_create_parquet_file(
             df,
             src_path,
             meta_idx,
@@ -548,7 +627,8 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
     # 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"]
-
+    for k in audios_idx:
+        audios_idx[k]["latest_duration"] += audios_idx[k]["episode_duration"]
     return meta_idx
 
 
@@ -562,7 +642,7 @@ def append_or_create_parquet_file(
     contains_images: bool = False,
     aggr_root: Path = None,
     hf_features: datasets.Features | None = None,
-) -> tuple[dict[str, int], tuple[int, int]]:
+):
     """Appends data to an existing parquet file or creates a new one based on size constraints.
 
     Manages file rotation when size limits are exceeded to prevent individual files
@@ -580,11 +660,9 @@ def append_or_create_parquet_file(
         hf_features: Optional HuggingFace Features schema for proper image typing.
 
     Returns:
-        tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
-               and (dst_chunk, dst_file) is the actual destination file the data was written to.
+        dict: Updated index dictionary with current chunk and file indices.
     """
-    dst_chunk, dst_file = idx["chunk"], idx["file"]
-    dst_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
+    dst_path = aggr_root / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
 
     if not dst_path.exists():
         dst_path.parent.mkdir(parents=True, exist_ok=True)
@@ -592,15 +670,14 @@ def append_or_create_parquet_file(
             to_parquet_with_hf_images(df, dst_path, features=hf_features)
         else:
             df.to_parquet(dst_path)
-        return idx, (dst_chunk, dst_file)
+        return idx
 
     src_size = get_parquet_file_size_in_mb(src_path)
     dst_size = get_parquet_file_size_in_mb(dst_path)
 
     if dst_size + src_size >= max_mb:
         idx["chunk"], idx["file"] = update_chunk_file_indices(idx["chunk"], idx["file"], chunk_size)
-        dst_chunk, dst_file = idx["chunk"], idx["file"]
-        new_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
+        new_path = aggr_root / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
         new_path.parent.mkdir(parents=True, exist_ok=True)
         final_df = df
         target_path = new_path
@@ -619,7 +696,7 @@ def append_or_create_parquet_file(
     else:
         final_df.to_parquet(target_path)
 
-    return idx, (dst_chunk, dst_file)
+    return idx
 
 
 def finalize_aggregation(aggr_meta, all_metadata):

src/lerobot/datasets/audio_utils.py

@@ -0,0 +1,275 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+import av
+import torch
+import torchaudio
+import torchcodec
+from numpy import ceil
+
+CHANNELS_LAYOUTS_MAPPING = {
+    1: "mono",
+    2: "stereo",
+    3: "2.1",
+    4: "3.1",
+    5: "4.1",
+    6: "5.1",
+    7: "6.1",
+    8: "7.1",
+    16: "hexadecagonal",
+    24: "22.2",
+}
+
+
+def decode_audio(
+    audio_path: Path | str,
+    timestamps: list[float],
+    duration: float,
+    start_time_s: float | None = 0.0,
+    backend: str | None = "torchcodec",
+) -> torch.Tensor:
+    """
+    Decodes audio using the specified backend.
+    Args:
+        audio_path (Path): Path to the audio file.
+        timestamps (list[float]): List of (starting) timestamps to extract audio chunks.
+        duration (float): Duration of the audio chunks in seconds.
+        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec".
+
+    Returns:
+        torch.Tensor: Decoded audio chunks.
+
+    Currently supports torchaudio.
+    """
+    if backend == "torchcodec":
+        return decode_audio_torchcodec(audio_path, timestamps, duration, start_time_s)
+    elif backend == "torchaudio":
+        return decode_audio_torchaudio(audio_path, timestamps, duration, start_time_s)
+    else:
+        raise ValueError(f"Unsupported video backend: {backend}")
+
+
+def decode_audio_torchcodec(
+    audio_path: Path | str,
+    timestamps: list[float],
+    duration: float,
+    start_time_s: float | None = 0.0,
+    log_loaded_timestamps: bool = False,
+) -> torch.Tensor:
+    # TODO(CarolinePascal) : add channels selection
+    audio_decoder = torchcodec.decoders.AudioDecoder(audio_path)
+    audio_sample_rate = audio_decoder.metadata.sample_rate
+    audio_channels = audio_decoder.metadata.num_channels
+    # TODO(CarolinePascal) : assert ts < total record duration
+
+    audio_chunks = []
+    timestamps = [
+        timestamp + start_time_s for timestamp in timestamps
+    ]  # Add an offset of start_time_s to each timestamp
+    for ts in timestamps:
+        current_audio_chunk = audio_decoder.get_samples_played_in_range(
+            start_seconds=max(0.0, ts - duration), stop_seconds=ts
+        )
+
+        current_audio_chunk_data = current_audio_chunk.data
+
+        # Case where the requested audio chunk starts before the beginning of the audio stream
+        if ts - duration < 0:
+            # No useful audio sample has been recorded
+            if ts < 1 / audio_sample_rate:
+                # TODO(CarolinePascal) : add low level white noise instead of zeros ?
+                current_audio_chunk_data = torch.zeros(
+                    (audio_channels, int(ceil(duration * audio_sample_rate)))
+                )
+            # At least one useful audio sample has been recorded
+            else:
+                # Pad the beginning of the audio chunk with zeros
+                # TODO(CarolinePascal) : add low level white noise instead of zeros ?
+                current_audio_chunk_data = torch.nn.functional.pad(
+                    current_audio_chunk_data,
+                    (int(ceil((duration - ts) * audio_sample_rate)), 0, 0, 0),  # left, right, top, bottom
+                )
+
+        if log_loaded_timestamps:
+            logging.info(
+                f"audio chunk loaded at timestamp={current_audio_chunk.pts_seconds:.4f} with duration={current_audio_chunk.duration_seconds:.4f}"
+            )
+
+        audio_chunks.append(current_audio_chunk_data)
+
+    audio_chunks = torch.stack(audio_chunks)
+
+    assert len(timestamps) == len(audio_chunks)
+    return audio_chunks
+
+
+def decode_audio_torchaudio(
+    audio_path: Path | str,
+    timestamps: list[float],
+    duration: float,
+    start_time_s: float | None = 0.0,
+    log_loaded_timestamps: bool = False,
+) -> torch.Tensor:
+    # TODO(CarolinePascal) : add channels selection
+    audio_path = str(audio_path)
+
+    reader = torchaudio.io.StreamReader(src=audio_path)
+    audio_sample_rate = reader.get_src_stream_info(reader.default_audio_stream).sample_rate
+    audio_channels = reader.get_src_stream_info(reader.default_audio_stream).num_channels
+    # TODO(CarolinePascal) : assert ts < total record duration
+
+    # TODO(CarolinePascal) : sort timestamps ?
+
+    reader.add_basic_audio_stream(
+        frames_per_chunk=int(ceil(duration * audio_sample_rate)),  # Too much is better than not enough
+        buffer_chunk_size=-1,  # No dropping frames
+        format="fltp",  # Format as float32
+    )
+
+    audio_chunks = []
+    timestamps = [
+        timestamp + start_time_s for timestamp in timestamps
+    ]  # Add an offset of start_time_s to each timestamp
+    for ts in timestamps:
+        reader.seek(max(0.0, ts - duration))  # Default to closest audio sample. Needs to be non-negative !
+        status = reader.fill_buffer()
+        if status != 0:
+            # Should not happen, but just in case
+            logging.warning("Audio stream reached end of recording before decoding desired timestamps.")
+
+        current_audio_chunk = reader.pop_chunks()[0]
+        current_audio_chunk_data = current_audio_chunk.t()  # Channel first format
+
+        # Case where the requested audio chunk starts before the beginning of the audio stream
+        if ts - duration < 0:
+            # No useful audio sample has been recorded
+            if ts < 1 / audio_sample_rate:
+                current_audio_chunk_data = torch.zeros(
+                    (audio_channels, int(ceil(duration * audio_sample_rate)))
+                )
+            # At least one useful audio sample has been recorded
+            else:
+                # Remove the superfluous last samples of the audio chunk
+                current_audio_chunk_data = current_audio_chunk_data[:, : int(ceil(ts * audio_sample_rate))]
+                # Pad the beginning of the audio chunk with zeros
+                # TODO(CarolinePascal) : add low level white noise instead of zeros ?
+                current_audio_chunk_data = torch.nn.functional.pad(
+                    current_audio_chunk_data,
+                    (int(ceil((duration - ts) * audio_sample_rate)), 0, 0, 0),  # left, right, top, bottom
+                )
+
+        if log_loaded_timestamps:
+            logging.info(
+                f"audio chunk loaded at starting timestamp={current_audio_chunk['pts']:.4f} with duration={len(current_audio_chunk) / audio_sample_rate:.4f}"
+            )
+
+        audio_chunks.append(current_audio_chunk_data)
+
+    audio_chunks = torch.stack(audio_chunks)
+
+    assert len(timestamps) == len(audio_chunks)
+    return audio_chunks
+
+
+def encode_audio(
+    input_path: Path | str,
+    output_path: Path | str,
+    codec: str = "aac",  # TODO(CarolinePascal) : investigate Fraunhofer FDK AAC (libfdk_aac) codec and and constant (file size control) /variable (quality control) bitrate options
+    bit_rate: int | None = None,
+    sample_rate: int | None = None,
+    log_level: int | None = av.logging.ERROR,
+    overwrite: bool = False,
+) -> None:
+    """Encodes an audio file using ffmpeg."""
+    output_path = Path(output_path)
+    output_path.parent.mkdir(parents=True, exist_ok=overwrite)
+
+    # Set logging level
+    if log_level is not None:
+        # "While less efficient, it is generally preferable to modify logging with Python’s logging"
+        logging.getLogger("libav").setLevel(log_level)
+
+    # Open input file
+    with av.open(str(input_path), "r") as input:
+        input_stream = input.streams.audio[0]  # Assuming the first stream is the audio stream to be encoded
+
+        # Define sub-sampling options
+        if sample_rate is None:
+            sample_rate = input_stream.rate
+
+        # Create and open output file (overwrite by default)
+        with av.open(str(output_path), "w") as output:
+            output_stream = output.add_stream(
+                codec, rate=sample_rate, layout=CHANNELS_LAYOUTS_MAPPING[input_stream.channels]
+            )
+
+            if bit_rate is not None:
+                output_stream.bit_rate = bit_rate
+
+            # Loop through input WAV packets and encode them
+            for input_frame in input.decode(
+                input_stream
+            ):  # This step handles both demuxing and decoding under the hood
+                packet = output_stream.encode(input_frame)
+                if packet:
+                    output.mux(packet)
+
+            # Flush the encoder
+            packet = output_stream.encode()
+            if packet:
+                output.mux(packet)
+
+    # Reset logging level
+    if log_level is not None:
+        av.logging.restore_default_callback()
+
+    if not output_path.exists():
+        raise OSError(f"Audio encoding did not work. File not found: {output_path}.")
+
+
+def get_audio_info(video_path: Path | str) -> dict:
+    # Set logging level
+    logging.getLogger("libav").setLevel(av.logging.ERROR)
+
+    # Getting audio stream information
+    audio_info = {}
+    with av.open(str(video_path), "r") as audio_file:
+        try:
+            audio_stream = audio_file.streams.audio[0]
+        except IndexError:
+            # Reset logging level
+            av.logging.restore_default_callback()
+            return {"has_audio": False}
+
+        audio_info["audio.channels"] = audio_stream.channels
+        audio_info["audio.codec"] = audio_stream.codec.canonical_name
+        # In an ideal loseless case : bit depth x sample rate x channels = bit rate.
+        # In an actual compressed case, the bit rate is set according to the compression level : the lower the bit rate, the more compression is applied.
+        audio_info["audio.bit_rate"] = audio_stream.bit_rate
+        audio_info["audio.sample_rate"] = audio_stream.sample_rate  # Number of samples per second
+        # In an ideal loseless case : fixed number of bits per sample.
+        # In an actual compressed case : variable number of bits per sample (often reduced to match a given depth rate).
+        audio_info["audio.bit_depth"] = audio_stream.format.bits
+        audio_info["audio.channel_layout"] = audio_stream.layout.name
+        audio_info["has_audio"] = True
+
+    # Reset logging level
+    av.logging.restore_default_callback()
+
+    return audio_info

src/lerobot/datasets/card_template.md

@@ -7,13 +7,6 @@
 
 This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
 
-{% if repo_id is defined and repo_id %}
-<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path={{ repo_id }}">
-<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
-<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
-</a>
-{% endif %}
-
 ## Dataset Description
 
 {{ dataset_description | default("", true) }}

src/lerobot/datasets/compute_stats.py

@@ -15,7 +15,7 @@
 # limitations under the License.
 import numpy as np
 
-from lerobot.datasets.utils import load_image_as_numpy
+from lerobot.datasets.utils import load_audio_from_path, load_image_as_numpy
 
 DEFAULT_QUANTILES = [0.01, 0.10, 0.50, 0.90, 0.99]
 
@@ -245,6 +245,20 @@ def sample_images(image_paths: list[str]) -> np.ndarray:
     return images
 
 
+def sample_audio_from_path(audio_path: str) -> np.ndarray:
+    """Samples audio data from an audio recording stored in a WAV file."""
+    data = load_audio_from_path(audio_path)
+    sampled_indices = sample_indices(len(data))
+
+    return data[sampled_indices]
+
+
+def sample_audio_from_data(data: np.ndarray) -> np.ndarray:
+    """Samples audio data from an audio recording stored in a numpy array."""
+    sampled_indices = sample_indices(len(data))
+    return data[sampled_indices]
+
+
 def _reshape_stats_by_axis(
     stats: dict[str, np.ndarray],
     axis: int | tuple[int, ...] | None,
@@ -512,6 +526,13 @@ def compute_episode_stats(
             ep_ft_array = sample_images(data)
             axes_to_reduce = (0, 2, 3)
             keepdims = True
+        elif features[key]["dtype"] == "audio":
+            try:
+                ep_ft_array = sample_audio_from_path(data[0])
+            except TypeError:  # Should only be triggered for LeKiwi robot, for which audio is stored chunk by chunk in a visual frame-like manner
+                ep_ft_array = sample_audio_from_data(data)
+            axes_to_reduce = 0
+            keepdims = True
         else:
             ep_ft_array = data
             axes_to_reduce = 0

src/lerobot/datasets/dataset_tools.py

@@ -567,22 +567,20 @@ def _copy_and_reindex_data(
 def _keep_episodes_from_video_with_av(
     input_path: Path,
     output_path: Path,
-    episodes_to_keep: list[tuple[int, int]],
+    episodes_to_keep: list[tuple[float, float]],
     fps: float,
     vcodec: str = "libsvtav1",
     pix_fmt: str = "yuv420p",
 ) -> None:
     """Keep only specified episodes from a video file using PyAV.
 
-    This function decodes frames from specified frame ranges and re-encodes them with
+    This function decodes frames from specified time ranges and re-encodes them with
     properly reset timestamps to ensure monotonic progression.
 
     Args:
         input_path: Source video file path.
         output_path: Destination video file path.
-        episodes_to_keep: List of (start_frame, end_frame) tuples for episodes to keep.
-            Ranges are half-open intervals: [start_frame, end_frame), where start_frame
-            is inclusive and end_frame is exclusive.
+        episodes_to_keep: List of (start_time, end_time) tuples for episodes to keep.
         fps: Frame rate of the video.
         vcodec: Video codec to use for encoding.
         pix_fmt: Pixel format for output video.
@@ -624,10 +622,9 @@ def _keep_episodes_from_video_with_av(
 
     # Create set of (start, end) ranges for fast lookup.
     # Convert to a sorted list for efficient checking.
-    frame_ranges = sorted(episodes_to_keep)
+    time_ranges = sorted(episodes_to_keep)
 
     # Track frame index for setting PTS and current range being processed.
-    src_frame_count = 0
     frame_count = 0
     range_idx = 0
 
@@ -637,20 +634,21 @@ def _keep_episodes_from_video_with_av(
             if frame is None:
                 continue
 
-            # Check if frame is in any of our desired frame ranges.
+            # Get frame timestamp.
+            frame_time = float(frame.pts * frame.time_base) if frame.pts is not None else 0.0
+
+            # Check if frame is in any of our desired time ranges.
             # Skip ranges that have already passed.
-            while range_idx < len(frame_ranges) and src_frame_count >= frame_ranges[range_idx][1]:
+            while range_idx < len(time_ranges) and frame_time >= time_ranges[range_idx][1]:
                 range_idx += 1
 
             # If we've passed all ranges, stop processing.
-            if range_idx >= len(frame_ranges):
+            if range_idx >= len(time_ranges):
                 break
 
             # Check if frame is in current range.
-            start_frame = frame_ranges[range_idx][0]
-
-            if src_frame_count < start_frame:
-                src_frame_count += 1
+            start_ts, end_ts = time_ranges[range_idx]
+            if frame_time < start_ts:
                 continue
 
             # Frame is in range - create a new frame with reset timestamps.
@@ -663,7 +661,6 @@ def _keep_episodes_from_video_with_av(
             for pkt in v_out.encode(new_frame):
                 out.mux(pkt)
 
-            src_frame_count += 1
             frame_count += 1
 
     # Flush encoder.
@@ -752,17 +749,15 @@ def _copy_and_reindex_videos(
                         f"videos/{video_key}/to_timestamp"
                     ]
             else:
-                # Build list of frame ranges to keep, in sorted order.
+                # Build list of time ranges to keep, in sorted order.
                 sorted_keep_episodes = sorted(episodes_in_file, key=lambda x: episode_mapping[x])
-                episodes_to_keep_ranges: list[tuple[int, int]] = []
+                episodes_to_keep_ranges: list[tuple[float, float]] = []
+
                 for old_idx in sorted_keep_episodes:
                     src_ep = src_dataset.meta.episodes[old_idx]
-                    from_frame = round(src_ep[f"videos/{video_key}/from_timestamp"] * src_dataset.meta.fps)
-                    to_frame = round(src_ep[f"videos/{video_key}/to_timestamp"] * src_dataset.meta.fps)
-                    assert src_ep["length"] == to_frame - from_frame, (
-                        f"Episode length mismatch: {src_ep['length']} vs {to_frame - from_frame}"
-                    )
-                    episodes_to_keep_ranges.append((from_frame, to_frame))
+                    from_ts = src_ep[f"videos/{video_key}/from_timestamp"]
+                    to_ts = src_ep[f"videos/{video_key}/to_timestamp"]
+                    episodes_to_keep_ranges.append((from_ts, to_ts))
 
                 # Use PyAV filters to efficiently re-encode only the desired segments.
                 assert src_dataset.meta.video_path is not None
@@ -1401,132 +1396,6 @@ BYTES_PER_KIB = 1024
 BYTES_PER_MIB = BYTES_PER_KIB * BYTES_PER_KIB
 
 
-def modify_tasks(
-    dataset: LeRobotDataset,
-    new_task: str | None = None,
-    episode_tasks: dict[int, str] | None = None,
-) -> LeRobotDataset:
-    """Modify tasks in a LeRobotDataset.
-
-    This function allows you to either:
-    1. Set a single task for the entire dataset (using `new_task`)
-    2. Set specific tasks for specific episodes (using `episode_tasks`)
-
-    You can combine both: `new_task` sets the default, and `episode_tasks` overrides
-    specific episodes.
-
-    The dataset is modified in-place, updating only the task-related files:
-    - meta/tasks.parquet
-    - data/**/*.parquet (task_index column)
-    - meta/episodes/**/*.parquet (tasks column)
-    - meta/info.json (total_tasks)
-
-    Args:
-        dataset: The source LeRobotDataset to modify.
-        new_task: A single task string to apply to all episodes. If None and episode_tasks
-            is also None, raises an error.
-        episode_tasks: Optional dict mapping episode indices to their task strings.
-            Overrides `new_task` for specific episodes.
-
-
-    Examples:
-        Set a single task for all episodes:
-            dataset = modify_tasks(dataset, new_task="Pick up the cube")
-
-        Set different tasks for specific episodes:
-            dataset = modify_tasks(
-                dataset,
-                episode_tasks={0: "Task A", 1: "Task B", 2: "Task A"}
-            )
-
-        Set a default task with overrides:
-            dataset = modify_tasks(
-                dataset,
-                new_task="Default task",
-                episode_tasks={5: "Special task for episode 5"}
-            )
-    """
-    if new_task is None and episode_tasks is None:
-        raise ValueError("Must specify at least one of new_task or episode_tasks")
-
-    if episode_tasks is not None:
-        valid_indices = set(range(dataset.meta.total_episodes))
-        invalid = set(episode_tasks.keys()) - valid_indices
-        if invalid:
-            raise ValueError(f"Invalid episode indices: {invalid}")
-
-    # Ensure episodes metadata is loaded
-    if dataset.meta.episodes is None:
-        dataset.meta.episodes = load_episodes(dataset.root)
-
-    # Build the mapping from episode index to task string
-    episode_to_task: dict[int, str] = {}
-    for ep_idx in range(dataset.meta.total_episodes):
-        if episode_tasks and ep_idx in episode_tasks:
-            episode_to_task[ep_idx] = episode_tasks[ep_idx]
-        elif new_task is not None:
-            episode_to_task[ep_idx] = new_task
-        else:
-            # Keep original task if not overridden and no default provided
-            original_tasks = dataset.meta.episodes[ep_idx]["tasks"]
-            if not original_tasks:
-                raise ValueError(f"Episode {ep_idx} has no tasks and no default task was provided")
-            episode_to_task[ep_idx] = original_tasks[0]
-
-    # Collect all unique tasks and create new task mapping
-    unique_tasks = sorted(set(episode_to_task.values()))
-    new_task_df = pd.DataFrame({"task_index": list(range(len(unique_tasks)))}, index=unique_tasks)
-    task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}
-
-    logging.info(f"Modifying tasks in {dataset.repo_id}")
-    logging.info(f"New tasks: {unique_tasks}")
-
-    root = dataset.root
-
-    # Update data files - modify task_index column
-    logging.info("Updating data files...")
-    data_dir = root / DATA_DIR
-
-    for parquet_path in tqdm(sorted(data_dir.rglob("*.parquet")), desc="Updating data"):
-        df = pd.read_parquet(parquet_path)
-
-        # Build a mapping from episode_index to new task_index for rows in this file
-        episode_indices_in_file = df["episode_index"].unique()
-        ep_to_new_task_idx = {
-            ep_idx: task_to_index[episode_to_task[ep_idx]] for ep_idx in episode_indices_in_file
-        }
-
-        # Update task_index column
-        df["task_index"] = df["episode_index"].map(ep_to_new_task_idx)
-        df.to_parquet(parquet_path, index=False)
-
-    # Update episodes metadata - modify tasks column
-    logging.info("Updating episodes metadata...")
-    episodes_dir = root / "meta" / "episodes"
-
-    for parquet_path in tqdm(sorted(episodes_dir.rglob("*.parquet")), desc="Updating episodes"):
-        df = pd.read_parquet(parquet_path)
-
-        # Update tasks column
-        df["tasks"] = df["episode_index"].apply(lambda ep_idx: [episode_to_task[ep_idx]])
-        df.to_parquet(parquet_path, index=False)
-
-    # Write new tasks.parquet
-    write_tasks(new_task_df, root)
-
-    # Update info.json
-    dataset.meta.info["total_tasks"] = len(unique_tasks)
-    write_info(dataset.meta.info, root)
-
-    # Reload metadata to reflect changes
-    dataset.meta.tasks = new_task_df
-    dataset.meta.episodes = load_episodes(root)
-
-    logging.info(f"Tasks: {unique_tasks}")
-
-    return dataset
-
-
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
     output_dir: Path,

src/lerobot/datasets/lerobot_dataset.py

@@ -33,12 +33,16 @@ import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
+from lerobot.datasets.audio_utils import decode_audio, encode_audio, get_audio_info
 from lerobot.datasets.compute_stats import aggregate_stats, compute_episode_stats
 from lerobot.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.datasets.utils import (
+    DEFAULT_AUDIO_CHUNK_DURATION,
     DEFAULT_EPISODES_PATH,
     DEFAULT_FEATURES,
     DEFAULT_IMAGE_PATH,
+    DEFAULT_INITIAL_AUDIO_BUFFER_DURATION,
+    DEFAULT_RAW_AUDIO_PATH,
     INFO_PATH,
     _validate_feature_names,
     check_delta_timestamps,
@@ -57,7 +61,6 @@ from lerobot.datasets.utils import (
     load_info,
     load_nested_dataset,
     load_stats,
-    load_subtasks,
     load_tasks,
     update_chunk_file_indices,
     validate_episode_buffer,
@@ -68,19 +71,20 @@ from lerobot.datasets.utils import (
     write_tasks,
 )
 from lerobot.datasets.video_utils import (
-    StreamingVideoEncoder,
     VideoFrame,
-    concatenate_video_files,
+    concatenate_media_files,
     decode_video_frames,
     encode_video_frames,
+    get_media_duration_in_s,
     get_safe_default_codec,
-    get_video_duration_in_s,
     get_video_info,
-    resolve_vcodec,
 )
+from lerobot.microphones import Microphone
+from lerobot.microphones.utils import async_microphones_start_recording
 from lerobot.utils.constants import HF_LEROBOT_HOME
 
 CODEBASE_VERSION = "v3.0"
+VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1"}
 
 
 class LeRobotDatasetMetadata:
@@ -164,7 +168,6 @@ class LeRobotDatasetMetadata:
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
         self.tasks = load_tasks(self.root)
-        self.subtasks = load_subtasks(self.root)
         self.episodes = load_episodes(self.root)
         self.stats = load_stats(self.root)
 
@@ -217,6 +220,19 @@ class LeRobotDatasetMetadata:
         fpath = self.video_path.format(video_key=vid_key, chunk_index=chunk_idx, file_index=file_idx)
         return Path(fpath)
 
+    def get_audio_file_path(self, ep_index: int, audio_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"audio/{audio_key}/chunk_index"]
+        file_idx = ep[f"audio/{audio_key}/file_index"]
+        fpath = self.audio_path.format(audio_key=audio_key, chunk_index=chunk_idx, file_index=file_idx)
+        return Path(fpath)
+
     @property
     def data_path(self) -> str:
         """Formattable string for the parquet files."""
@@ -227,6 +243,11 @@ class LeRobotDatasetMetadata:
         """Formattable string for the video files."""
         return self.info["video_path"]
 
+    @property
+    def audio_path(self) -> str | None:
+        """Formattable string for the audio files."""
+        return self.info["audio_path"]
+
     @property
     def robot_type(self) -> str | None:
         """Robot type used in recording this dataset."""
@@ -257,6 +278,11 @@ class LeRobotDatasetMetadata:
         """Keys to access visual modalities (regardless of their storage method)."""
         return [key for key, ft in self.features.items() if ft["dtype"] in ["video", "image"]]
 
+    @property
+    def audio_keys(self) -> list[str]:
+        """Keys to access audio modalities."""
+        return [key for key, ft in self.features.items() if ft["dtype"] == "audio"]
+
     @property
     def names(self) -> dict[str, list | dict]:
         """Names of the various dimensions of vector modalities."""
@@ -297,6 +323,11 @@ class LeRobotDatasetMetadata:
         """Max size of video file in mega bytes."""
         return self.info["video_files_size_in_mb"]
 
+    @property
+    def audio_files_size_in_mb(self) -> int:
+        """Max size of audio file in mega bytes."""
+        return self.info["audio_files_size_in_mb"]
+
     def get_task_index(self, task: str) -> int | None:
         """
         Given a task in natural language, returns its task_index if the task already exists in the dataset,
@@ -438,11 +469,27 @@ class LeRobotDatasetMetadata:
                 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_audio_info(self, audio_key: str | None = None) -> None:
+        """
+        Warning: this function writes info from first episode audio, implicitly assuming that all audio have
+        been encoded the same way. Also, this means it assumes the first episode exists.
+        """
+        if audio_key is not None and audio_key not in self.audio_keys:
+            raise ValueError(f"Audio key {audio_key} not found in dataset")
+
+        audio_keys = [audio_key] if audio_key is not None else self.audio_keys
+        for key in audio_keys:
+            if not self.features[key].get("info", None):
+                audio_path = self.root / self.audio_path.format(audio_key=key, chunk_index=0, file_index=0)
+                self.info["features"][key]["info"] = get_audio_info(audio_path)
+                self.info["features"][key]["info"]["start_time_s"] = DEFAULT_INITIAL_AUDIO_BUFFER_DURATION
+
     def update_chunk_settings(
         self,
         chunks_size: int | None = None,
         data_files_size_in_mb: int | None = None,
         video_files_size_in_mb: int | None = None,
+        audio_files_size_in_mb: int | None = None,
     ) -> None:
         """Update chunk and file size settings after dataset creation.
 
@@ -454,6 +501,7 @@ class LeRobotDatasetMetadata:
             chunks_size: Maximum number of files per chunk directory. If None, keeps current value.
             data_files_size_in_mb: Maximum size for data parquet files in MB. If None, keeps current value.
             video_files_size_in_mb: Maximum size for video files in MB. If None, keeps current value.
+            audio_files_size_in_mb: Maximum size for audio files in MB. If None, keeps current value.
         """
         if chunks_size is not None:
             if chunks_size <= 0:
@@ -470,6 +518,11 @@ class LeRobotDatasetMetadata:
                 raise ValueError(f"video_files_size_in_mb must be positive, got {video_files_size_in_mb}")
             self.info["video_files_size_in_mb"] = video_files_size_in_mb
 
+        if audio_files_size_in_mb is not None:
+            if audio_files_size_in_mb <= 0:
+                raise ValueError(f"audio_files_size_in_mb must be positive, got {audio_files_size_in_mb}")
+            self.info["audio_files_size_in_mb"] = audio_files_size_in_mb
+
         # Update the info file on disk
         write_info(self.info, self.root)
 
@@ -477,12 +530,13 @@ class LeRobotDatasetMetadata:
         """Get current chunk and file size settings.
 
         Returns:
-            Dict containing chunks_size, data_files_size_in_mb, and video_files_size_in_mb.
+            Dict containing chunks_size, data_files_size_in_mb, video_files_size_in_mb, and audio_files_size_in_mb.
         """
         return {
             "chunks_size": self.chunks_size,
             "data_files_size_in_mb": self.data_files_size_in_mb,
             "video_files_size_in_mb": self.video_files_size_in_mb,
+            "audio_files_size_in_mb": self.audio_files_size_in_mb,
         }
 
     def __repr__(self):
@@ -509,6 +563,7 @@ class LeRobotDatasetMetadata:
         chunks_size: int | None = None,
         data_files_size_in_mb: int | None = None,
         video_files_size_in_mb: int | None = None,
+        audio_files_size_in_mb: int | None = None,
     ) -> "LeRobotDatasetMetadata":
         """Creates metadata for a LeRobotDataset."""
         obj = cls.__new__(cls)
@@ -521,7 +576,6 @@ class LeRobotDatasetMetadata:
         _validate_feature_names(features)
 
         obj.tasks = None
-        obj.subtasks = None
         obj.episodes = None
         obj.stats = None
         obj.info = create_empty_dataset_info(
@@ -533,6 +587,7 @@ class LeRobotDatasetMetadata:
             chunks_size,
             data_files_size_in_mb,
             video_files_size_in_mb,
+            audio_files_size_in_mb,
         )
         if len(obj.video_keys) > 0 and not use_videos:
             raise ValueError()
@@ -546,19 +601,12 @@ class LeRobotDatasetMetadata:
 
 
 def _encode_video_worker(
-    video_key: str,
-    episode_index: int,
-    root: Path,
-    fps: int,
-    vcodec: str = "libsvtav1",
-    encoder_threads: int | None = None,
+    video_key: str, episode_index: int, root: Path, fps: int, vcodec: str = "libsvtav1"
 ) -> 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, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
-    )
+    encode_video_frames(img_dir, temp_path, fps, vcodec=vcodec, overwrite=True)
     shutil.rmtree(img_dir)
     return temp_path
 
@@ -575,12 +623,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         revision: str | None = None,
         force_cache_sync: bool = False,
         download_videos: bool = True,
+        download_audio: bool = True,
         video_backend: str | None = None,
+        audio_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
-        streaming_encoding: bool = False,
-        encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -612,6 +659,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                   task-conditioned training.
             - hf_dataset (from datasets.Dataset), which will read any values from parquet files.
             - videos (optional) from which frames are loaded to be synchronous with data from parquet files.
+            - audio (optional) from which audio is loaded to be synchronous with data from parquet files.
 
         A typical LeRobotDataset looks like this from its root path:
         .
@@ -637,19 +685,37 @@ class LeRobotDataset(torch.utils.data.Dataset):
         │   ├── info.json
         │   ├── stats.json
         │   └── tasks.parquet
-        └── videos
-            ├── observation.images.laptop
+        ├── videos
+        │   ├── observation.images.laptop
+        │   │   ├── chunk-000
+        │   │   │   ├── file-000.mp4
+        │   │   │   ├── file-001.mp4
+        │   │   │   └── ...
+        │   │   ├── chunk-001
+        │   │   │   └── ...
+        │   │   └── ...
+        │   ├── observation.images.phone
+        │   │   ├── chunk-000
+        │   │   │   ├── file-000.mp4
+        │   │   │   ├── file-001.mp4
+        │   │   │   └── ...
+        │   │   ├── chunk-001
+        │   │   │   └── ...
+        │   │   └── ...
+        │   └── ...
+        └── audio
+            ├── observation.audio.laptop
             │   ├── chunk-000
-            │   │   ├── file-000.mp4
-            │   │   ├── file-001.mp4
+            │   │   ├── file-000.m4a
+            │   │   ├── file-001.m4a
             │   │   └── ...
             │   ├── chunk-001
             │   │   └── ...
             │   └── ...
-            ├── observation.images.phone
+            ├── observation.audio.phone
             │   ├── chunk-000
-            │   │   ├── file-000.mp4
-            │   │   ├── file-001.mp4
+            │   │   ├── file-000.m4a
+            │   │   ├── file-001.m4a
             │   │   └── ...
             │   ├── chunk-001
             │   │   └── ...
@@ -664,11 +730,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         for the README).
 
         Args:
-            repo_id (str): This is the repo id that will be used to fetch the dataset.
-            root (Path | None, optional): Local directory where the dataset will be downloaded and
-                stored. If set, all dataset files will be stored directly under this path. If not set, the
-                dataset files will be stored under $HF_LEROBOT_HOME/repo_id (configurable via the
-                HF_LEROBOT_HOME environment variable).
+            repo_id (str): This is the repo id that will be used to fetch the dataset. Locally, the dataset
+                will be stored under root/repo_id.
+            root (Path | None, optional): Local directory to use for downloading/writing files. You can also
+                set the LEROBOT_HOME environment variable to point to a different location. Defaults to
+                '~/.cache/huggingface/lerobot'.
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
             image_transforms (Callable | None, optional): You can pass standard v2 image transforms from
@@ -689,22 +755,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
             download_videos (bool, optional): Flag to download the videos. Note that when set to True but the
                 video files are already present on local disk, they won't be downloaded again. Defaults to
                 True.
+            download_audio (bool, optional): Flag to download the audio. Defaults to True.
             video_backend (str | None, optional): Video backend to use for decoding videos. Defaults to torchcodec when available int the platform; otherwise, defaults to 'pyav'.
                 You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
+            audio_backend (str | None, optional): Audio backend to use for decoding audio. Defaults to 'torchcodec'.
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
             vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
-                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
-            streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
-                instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
-            encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
-                streaming encoding. Defaults to 30 (~1s at 30fps).
-            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
-                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
-                libsvtav1 and 'threads' for h264/hevc.
+                'libsvtav1'. Defaults to 'libsvtav1'. Use 'h264' for faster encoding on systems where AV1
+                encoding is CPU-heavy.
         """
         super().__init__()
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         self.repo_id = repo_id
         self.root = Path(root) if root else HF_LEROBOT_HOME / repo_id
         self.image_transforms = image_transforms
@@ -713,11 +776,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.tolerance_s = tolerance_s
         self.revision = revision if revision else CODEBASE_VERSION
         self.video_backend = video_backend if video_backend else get_safe_default_codec()
+        self.audio_backend = (
+            audio_backend if audio_backend else "torchcodec"
+        )  # Waiting for torchcodec release #TODO(CarolinePascal)
         self.delta_indices = None
         self.batch_encoding_size = batch_encoding_size
         self.episodes_since_last_encoding = 0
-        self.vcodec = resolve_vcodec(vcodec)
-        self._encoder_threads = encoder_threads
+        self.vcodec = vcodec
 
         # Unused attributes
         self.image_writer = None
@@ -725,7 +790,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.writer = None
         self.latest_episode = None
         self._current_file_start_frame = None  # Track the starting frame index of the current parquet file
-        self._streaming_encoder = None
 
         self.root.mkdir(exist_ok=True, parents=True)
 
@@ -747,7 +811,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # Check if cached dataset contains all requested episodes
             if not self._check_cached_episodes_sufficient():
                 raise FileNotFoundError("Cached dataset doesn't contain all requested episodes")
-        except (FileNotFoundError, NotADirectoryError):
+        except (AssertionError, FileNotFoundError, NotADirectoryError):
             if is_valid_version(self.revision):
                 self.revision = get_safe_version(self.repo_id, self.revision)
             self.download(download_videos)
@@ -767,19 +831,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
-        # Initialize streaming encoder for resumed recording
-        if streaming_encoding and len(self.meta.video_keys) > 0:
-            self._streaming_encoder = StreamingVideoEncoder(
-                fps=self.meta.fps,
-                vcodec=self.vcodec,
-                pix_fmt="yuv420p",
-                g=2,
-                crf=30,
-                preset=None,
-                queue_maxsize=encoder_queue_maxsize,
-                encoder_threads=encoder_threads,
-            )
-
     def _close_writer(self) -> None:
         """Close and cleanup the parquet writer if it exists."""
         writer = getattr(self, "writer", None)
@@ -800,6 +851,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         license: str | None = "apache-2.0",
         tag_version: bool = True,
         push_videos: bool = True,
+        push_audio: bool = True,
         private: bool = False,
         allow_patterns: list[str] | str | None = None,
         upload_large_folder: bool = False,
@@ -808,6 +860,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         ignore_patterns = ["images/"]
         if not push_videos:
             ignore_patterns.append("videos/")
+        if not push_audio:
+            ignore_patterns.append("audio/")
 
         hub_api = HfApi()
         hub_api.create_repo(
@@ -839,7 +893,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             hub_api.upload_folder(**upload_kwargs)
 
         card = create_lerobot_dataset_card(
-            tags=tags, dataset_info=self.meta.info, license=license, repo_id=self.repo_id, **card_kwargs
+            tags=tags, dataset_info=self.meta.info, license=license, **card_kwargs
         )
         card.push_to_hub(repo_id=self.repo_id, repo_type="dataset", revision=branch)
 
@@ -862,7 +916,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             ignore_patterns=ignore_patterns,
         )
 
-    def download(self, download_videos: bool = True) -> None:
+    def download(self, download_videos: bool = True, download_audio: bool = True) -> None:
         """Downloads the dataset from the given 'repo_id' at the provided version. If 'episodes' is given, this
         will only download those episodes (selected by their episode_index). If 'episodes' is None, the whole
         dataset will be downloaded. Thanks to the behavior of snapshot_download, if the files are already present
@@ -870,8 +924,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         # TODO(rcadene, aliberts): implement faster transfer
         # https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads
-        ignore_patterns = None if download_videos else "videos/"
         files = None
+        ignore_patterns = []
+        if not download_videos:
+            ignore_patterns.append("videos/")
+        if not download_audio:
+            ignore_patterns.append("audio/")
         if self.episodes is not None:
             files = self.get_episodes_file_paths()
         self.pull_from_repo(allow_patterns=files, ignore_patterns=ignore_patterns)
@@ -886,6 +944,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 for ep_idx in episodes
             ]
             fpaths += video_files
+
+        if len(self.meta.audio_keys) > 0:
+            audio_files = [
+                str(self.meta.get_compressed_audio_file_path(ep_idx, audio_key))
+                for audio_key in self.meta.audio_keys
+                for ep_idx in episodes
+            ]
+            fpaths += audio_files
+
         # episodes are stored in the same files, so we return unique paths only
         fpaths = list(set(fpaths))
         return fpaths
@@ -898,7 +965,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return hf_dataset
 
     def _check_cached_episodes_sufficient(self) -> bool:
-        """Check if the cached dataset contains all requested episodes and their video files."""
+        """Check if the cached dataset contains all requested episodes and their video and audio files."""
         if self.hf_dataset is None or len(self.hf_dataset) == 0:
             return False
 
@@ -926,6 +993,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
                     if not video_path.exists():
                         return False
 
+        # Check if all required audio files exist
+        if len(self.meta.audio_keys) > 0:
+            for ep_idx in requested_episodes:
+                for audio_key in self.meta.audio_keys:
+                    audio_path = self.root / self.meta.get_audio_file_path(ep_idx, audio_key)
+                    if not audio_path.exists():
+                        return False
+
         return True
 
     def create_hf_dataset(self) -> datasets.Dataset:
@@ -1004,7 +1079,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         query_indices: dict[str, list[int]] | None = None,
     ) -> dict[str, list[float]]:
         query_timestamps = {}
-        for key in self.meta.video_keys:
+        for key in self.meta.video_keys + self.meta.audio_keys:
             if query_indices is not None and key in query_indices:
                 if self._absolute_to_relative_idx is not None:
                     relative_indices = [self._absolute_to_relative_idx[idx] for idx in query_indices[key]]
@@ -1019,7 +1094,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
         """
-        Query dataset for indices across keys, skipping video keys.
+        Query dataset for indices across keys, skipping video keys and audio keys.
 
         Tries column-first [key][indices] for speed, falls back to row-first.
 
@@ -1031,7 +1106,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         result: dict = {}
         for key, q_idx in query_indices.items():
-            if key in self.meta.video_keys:
+            if key in self.meta.video_keys or key in self.meta.audio_keys:
                 continue
             # Map absolute indices to relative indices if needed
             relative_indices = (
@@ -1066,6 +1141,28 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         return item
 
+    # TODO(CarolinePascal): add variable query durations
+    def _query_audio(
+        self, query_timestamps: dict[str, list[float]], query_duration: float, ep_idx: int
+    ) -> dict[str, torch.Tensor]:
+        ep = self.meta.episodes[ep_idx]
+        item = {}
+        for audio_key, query_ts in query_timestamps.items():
+            # Episodes are stored sequentially on a single mp4 to reduce the number of files.
+            # Thus we load the start timestamp of the episode on this mp4 and,
+            # shift the query timestamp accordingly.
+            from_timestamp = ep[f"audio/{audio_key}/from_timestamp"]
+            shifted_query_ts = [from_timestamp + ts for ts in query_ts]
+
+            audio_path = self.root / self.meta.get_audio_file_path(ep_idx, audio_key)
+            start_time_s = self.meta.features[audio_key]["info"].get("start_time_s", 0.0)
+            audio_chunk = decode_audio(
+                audio_path, shifted_query_ts, query_duration, start_time_s, self.audio_backend
+            )
+            item[audio_key] = audio_chunk.squeeze(0)
+
+        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:
@@ -1095,11 +1192,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
             for key, val in query_result.items():
                 item[key] = val
 
-        if len(self.meta.video_keys) > 0:
+        if len(self.meta.video_keys) > 0 or len(self.meta.audio_keys) > 0:
             current_ts = item["timestamp"].item()
             query_timestamps = self._get_query_timestamps(current_ts, query_indices)
             video_frames = self._query_videos(query_timestamps, ep_idx)
-            item = {**video_frames, **item}
+            audio_chunks = self._query_audio(query_timestamps, DEFAULT_AUDIO_CHUNK_DURATION, ep_idx)
+            item = {**item, **video_frames, **audio_chunks}
 
         if self.image_transforms is not None:
             image_keys = self.meta.camera_keys
@@ -1109,12 +1207,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # Add task as a string
         task_idx = item["task_index"].item()
         item["task"] = self.meta.tasks.iloc[task_idx].name
-
-        # add subtask information if available
-        if "subtask_index" in self.features and self.meta.subtasks is not None:
-            subtask_idx = item["subtask_index"].item()
-            item["subtask"] = self.meta.subtasks.iloc[subtask_idx].name
-
         return item
 
     def __repr__(self):
@@ -1135,8 +1227,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         self._close_writer()
         self.meta._close_writer()
-        if self._streaming_encoder is not None:
-            self._streaming_encoder.close()
 
     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
@@ -1155,6 +1245,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         )
         return self.root / fpath
 
+    def _get_raw_audio_file_path(self, episode_index: int, audio_key: str) -> Path:
+        fpath = DEFAULT_RAW_AUDIO_PATH.format(audio_key=audio_key, episode_index=episode_index)
+        return self.root / fpath
+
     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
 
@@ -1191,13 +1285,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.episode_buffer["timestamp"].append(timestamp)
         self.episode_buffer["task"].append(frame.pop("task"))  # Remove task from frame after processing
 
-        # Start streaming encoder on first frame of episode (once, before iterating keys)
-        if frame_index == 0 and self._streaming_encoder is not None:
-            self._streaming_encoder.start_episode(
-                video_keys=list(self.meta.video_keys),
-                temp_dir=self.root,
-            )
-
         # Add frame features to episode_buffer
         for key in frame:
             if key not in self.features:
@@ -1205,10 +1292,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                     f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
                 )
 
-            if self.features[key]["dtype"] == "video" and self._streaming_encoder is not None:
-                self._streaming_encoder.feed_frame(key, frame[key])
-                self.episode_buffer[key].append(None)  # Placeholder (video keys are skipped in parquet)
-            elif self.features[key]["dtype"] in ["image", "video"]:
+            if self.features[key]["dtype"] in ["image", "video"]:
                 img_path = self._get_image_file_path(
                     episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
                 )
@@ -1217,11 +1301,43 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 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))
+            elif self.features[key]["dtype"] == "audio":
+                if (
+                    self.meta.robot_type == "lekiwi"
+                ):  # Raw data storage should only be triggered for LeKiwi robot, for which audio is stored chunk by chunk in a visual frame-like manner
+                    self.episode_buffer[key].append(frame[key])
+                else:  # Otherwise, only the audio file path is stored in the episode buffer
+                    if frame_index == 0:
+                        audio_path = self._get_raw_audio_file_path(
+                            episode_index=self.episode_buffer["episode_index"], audio_key=key
+                        )
+                        self.episode_buffer[key].append(str(audio_path))
             else:
                 self.episode_buffer[key].append(frame[key])
 
         self.episode_buffer["size"] += 1
 
+    def add_microphone_recording(self, microphone_key: str, microphone: Microphone) -> None:
+        """
+        Starts recording audio data provided by the microphone and directly writes it in a .wav file.
+        """
+
+        audio_file = self._get_raw_audio_file_path(self.num_episodes, "observation.audio." + microphone_key)
+        microphone.start_recording(output_file=audio_file)
+
+    def add_microphones_recordings(self, microphones: dict[str, Microphone]) -> None:
+        """
+        Starts recording audio data provided by multiple microphones and directly writes it in appropriate .wav files.
+        """
+
+        output_files = []
+        for microphone_key in microphones:
+            output_files.append(
+                self._get_raw_audio_file_path(self.num_episodes, "observation.audio." + microphone_key)
+            )
+
+        async_microphones_start_recording(microphones, output_files)
+
     def save_episode(
         self,
         episode_data: dict | None = None,
@@ -1265,42 +1381,24 @@ class LeRobotDataset(torch.utils.data.Dataset):
             # are processed separately by storing image path and frame info as meta data
             if key in ["index", "episode_index", "task_index"] or ft["dtype"] in ["image", "video"]:
                 continue
+            elif ft["dtype"] == "audio":
+                if (
+                    self.meta.robot_type == "lekiwi"
+                ):  # Raw data storage should only be triggered for LeKiwi robot, for which audio is stored chunk by chunk in a visual frame-like manner
+                    episode_buffer[key] = np.concatenate(episode_buffer[key], axis=0)
+                continue
             episode_buffer[key] = np.stack(episode_buffer[key])
 
         # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
-
-        has_video_keys = len(self.meta.video_keys) > 0
-        use_streaming = self._streaming_encoder is not None and has_video_keys
-        use_batched_encoding = self.batch_encoding_size > 1
-
-        if use_streaming:
-            # Compute stats for non-video features only (video stats come from encoder)
-            non_video_buffer = {
-                k: v
-                for k, v in episode_buffer.items()
-                if self.features.get(k, {}).get("dtype") not in ("video",)
-            }
-            non_video_features = {k: v for k, v in self.features.items() if v["dtype"] != "video"}
-            ep_stats = compute_episode_stats(non_video_buffer, non_video_features)
-        else:
-            ep_stats = compute_episode_stats(episode_buffer, self.features)
+        ep_stats = compute_episode_stats(episode_buffer, self.features)
 
         ep_metadata = self._save_episode_data(episode_buffer)
+        has_video_keys = len(self.meta.video_keys) > 0
+        has_audio_keys = len(self.meta.audio_keys) > 0
+        use_batched_encoding = self.batch_encoding_size > 1
 
-        if use_streaming:
-            # Finish streaming encoding and collect results
-            streaming_results = self._streaming_encoder.finish_episode()
-            for video_key in self.meta.video_keys:
-                temp_path, video_stats = streaming_results[video_key]
-                if video_stats is not None:
-                    # Format stats same as compute_episode_stats: normalize to [0,1], reshape to (C,1,1)
-                    ep_stats[video_key] = {
-                        k: v if k == "count" else np.squeeze(v.reshape(1, -1, 1, 1) / 255.0, axis=0)
-                        for k, v in video_stats.items()
-                    }
-                ep_metadata.update(self._save_episode_video(video_key, episode_index, temp_path=temp_path))
-        elif has_video_keys and not use_batched_encoding:
+        if (has_video_keys or has_audio_keys) and not use_batched_encoding:
             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:
@@ -1314,7 +1412,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
                             self.root,
                             self.fps,
                             self.vcodec,
-                            self._encoder_threads,
                         ): video_key
                         for video_key in self.meta.video_keys
                     }
@@ -1338,21 +1435,30 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 for video_key in self.meta.video_keys:
                     ep_metadata.update(self._save_episode_video(video_key, episode_index))
 
+            # TODO(Caroline): add parallel encoding for audio as well
+            for audio_key in self.meta.audio_keys:
+                ep_metadata.update(self._save_episode_audio(audio_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)
 
-        if has_video_keys and use_batched_encoding:
+        if (has_video_keys or has_audio_keys) and use_batched_encoding:
             # Check if we should trigger batch encoding
             self.episodes_since_last_encoding += 1
             if self.episodes_since_last_encoding == self.batch_encoding_size:
                 start_ep = self.num_episodes - self.batch_encoding_size
                 end_ep = self.num_episodes
-                self._batch_save_episode_video(start_ep, end_ep)
+                if has_video_keys:
+                    self._batch_save_episode_video(start_ep, end_ep)
+                if has_audio_keys:
+                    self._batch_save_episode_audio(start_ep, end_ep)
                 self.episodes_since_last_encoding = 0
 
         if not episode_data:
             # Reset episode buffer and clean up temporary images (if not already deleted during video encoding)
-            self.clear_episode_buffer(delete_images=len(self.meta.image_keys) > 0)
+            self.clear_episode_buffer(
+                delete_images=len(self.meta.image_keys) > 0, delete_audio=len(self.meta.audio_keys) > 0
+            )
 
     def _batch_save_episode_video(self, start_episode: int, end_episode: int | None = None) -> None:
         """
@@ -1403,7 +1509,70 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 dtype_backend="pyarrow"
             )  # allows NaN values along with integers
 
+            # Save the current episode's audio metadata to the dataframe
+            audio_ep_metadata = {}
+            for audio_key in self.meta.audio_keys:
+                audio_ep_metadata.update(self._save_episode_audio(audio_key, ep_idx))
+            audio_ep_metadata.pop("episode_index")
+            audio_ep_df = pd.DataFrame(audio_ep_metadata, index=[ep_idx]).convert_dtypes(
+                dtype_backend="pyarrow"
+            )  # allows NaN values along with integers
+
             episode_df = episode_df.combine_first(video_ep_df)
+            episode_df = episode_df.combine_first(audio_ep_df)
+            episode_df.to_parquet(episode_df_path)
+            self.meta.episodes = load_episodes(self.root)
+
+    def _batch_save_episode_audio(self, start_episode: int, end_episode: int | None = None) -> None:
+        """
+        Batch save audio for multiple episodes.
+
+        Args:
+            start_episode: Starting episode index (inclusive)
+            end_episode: Ending episode index (exclusive). If None, encodes all episodes from start_episode to the current episode.
+        """
+        if end_episode is None:
+            end_episode = self.num_episodes
+
+        logging.info(
+            f"Batch encoding {self.batch_encoding_size} audio for episodes {start_episode} to {end_episode - 1}"
+        )
+
+        chunk_idx = self.meta.episodes[start_episode]["data/chunk_index"]
+        file_idx = self.meta.episodes[start_episode]["data/file_index"]
+        episode_df_path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+        episode_df = pd.read_parquet(episode_df_path)
+
+        for ep_idx in range(start_episode, end_episode):
+            logging.info(f"Encoding audio for episode {ep_idx}")
+
+            if (
+                self.meta.episodes[ep_idx]["data/chunk_index"] != chunk_idx
+                or self.meta.episodes[ep_idx]["data/file_index"] != file_idx
+            ):
+                # The current episode is in a new chunk or file.
+                # Save previous episode dataframe and update the Hugging Face dataset by reloading it.
+                episode_df.to_parquet(episode_df_path)
+                self.meta.episodes = load_episodes(self.root)
+
+                # Load new episode dataframe
+                chunk_idx = self.meta.episodes[ep_idx]["data/chunk_index"]
+                file_idx = self.meta.episodes[ep_idx]["data/file_index"]
+                episode_df_path = self.root / DEFAULT_EPISODES_PATH.format(
+                    chunk_index=chunk_idx, file_index=file_idx
+                )
+                episode_df = pd.read_parquet(episode_df_path)
+
+            # Save the current episode's video metadata to the dataframe
+            audio_ep_metadata = {}
+            for audio_key in self.meta.audio_keys:
+                audio_ep_metadata.update(self._save_episode_audio(audio_key, ep_idx))
+            audio_ep_metadata.pop("episode_index")
+            audio_ep_df = pd.DataFrame(audio_ep_metadata, index=[ep_idx]).convert_dtypes(
+                dtype_backend="pyarrow"
+            )  # allows NaN values along with integers
+
+            episode_df = episode_df.combine_first(audio_ep_df)
             episode_df.to_parquet(episode_df_path)
             self.meta.episodes = load_episodes(self.root)
 
@@ -1514,7 +1683,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             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)
+        ep_duration_in_s = get_media_duration_in_s(ep_path, media_type="video")
 
         if (
             episode_index == 0
@@ -1560,7 +1729,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 latest_duration_in_s = 0.0
             else:
                 # Update latest video file
-                concatenate_video_files(
+                concatenate_media_files(
                     [latest_path, ep_path],
                     latest_path,
                 )
@@ -1582,11 +1751,79 @@ class LeRobotDataset(torch.utils.data.Dataset):
         }
         return metadata
 
-    def clear_episode_buffer(self, delete_images: bool = True) -> None:
-        # Cancel streaming encoder if active
-        if self._streaming_encoder is not None:
-            self._streaming_encoder.cancel_episode()
+    def _save_episode_audio(self, audio_key: str, episode_index: int) -> dict:
+        # Encode episode audio into a temporary audio file
+        ep_path = self._encode_temporary_episode_audio(audio_key, episode_index)
+        ep_size_in_mb = get_file_size_in_mb(ep_path)
+        ep_duration_in_s = get_media_duration_in_s(ep_path, media_type="audio")
 
+        if (
+            episode_index == 0
+            or self.meta.latest_episode is None
+            or f"audio/{audio_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"audio/{audio_key}/chunk_index"]
+                old_file_idx = self.meta.episodes[-1][f"audio/{audio_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.audio_path.format(
+                audio_key=audio_key, chunk_index=chunk_idx, file_index=file_idx
+            )
+            new_path.parent.mkdir(parents=True, exist_ok=True)
+            shutil.move(str(ep_path), str(new_path))
+        else:
+            # Retrieve information from the latest updated audio file using latest_episode
+            latest_ep = self.meta.latest_episode
+            chunk_idx = latest_ep[f"audio/{audio_key}/chunk_index"][0]
+            file_idx = latest_ep[f"audio/{audio_key}/file_index"][0]
+
+            latest_path = self.root / self.meta.audio_path.format(
+                audio_key=audio_key, chunk_index=chunk_idx, file_index=file_idx
+            )
+            latest_size_in_mb = get_file_size_in_mb(latest_path)
+            latest_duration_in_s = latest_ep[f"audio/{audio_key}/to_timestamp"][0]
+
+            if latest_size_in_mb + ep_size_in_mb >= self.meta.audio_files_size_in_mb:
+                # Move temporary episode audio to a new audio file in the dataset
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
+                new_path = self.root / self.meta.audio_path.format(
+                    audio_key=audio_key, chunk_index=chunk_idx, file_index=file_idx
+                )
+                new_path.parent.mkdir(parents=True, exist_ok=True)
+                shutil.move(str(ep_path), str(new_path))
+                latest_duration_in_s = 0.0
+            else:
+                # Update latest audio file
+                concatenate_media_files(
+                    [latest_path, ep_path],
+                    latest_path,
+                )
+
+        # Remove temporary directory
+        shutil.rmtree(str(ep_path.parent))
+
+        # Update audio info (only needed when first episode is encoded since it reads from episode 0)
+        if episode_index == 0:
+            self.meta.update_audio_info(audio_key)
+            write_info(self.meta.info, self.meta.root)  # ensure audio info always written properly
+
+        metadata = {
+            "episode_index": episode_index,
+            f"audio/{audio_key}/chunk_index": chunk_idx,
+            f"audio/{audio_key}/file_index": file_idx,
+            f"audio/{audio_key}/from_timestamp": latest_duration_in_s,
+            f"audio/{audio_key}/to_timestamp": latest_duration_in_s + ep_duration_in_s,
+        }
+        return metadata
+
+    def clear_episode_buffer(self, delete_images: bool = True, delete_audio: bool = True) -> None:
         # Clean up image files for the current episode buffer
         if delete_images:
             # Wait for the async image writer to finish
@@ -1600,6 +1837,16 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 if img_dir.is_dir():
                     shutil.rmtree(img_dir)
 
+        # Clean up audio files for the current episode buffer
+        if delete_audio:
+            episode_index = self.episode_buffer["episode_index"]
+            if isinstance(episode_index, np.ndarray):
+                episode_index = episode_index.item() if episode_index.size == 1 else episode_index[0]
+            for audio_key in self.meta.audio_keys:
+                audio_file = self._get_raw_audio_file_path(episode_index, audio_key)
+                if audio_file.is_file():
+                    audio_file.unlink()
+
         # Reset the buffer
         self.episode_buffer = self.create_episode_buffer()
 
@@ -1634,9 +1881,19 @@ 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.
         """
-        return _encode_video_worker(
-            video_key, episode_index, self.root, self.fps, self.vcodec, self._encoder_threads
-        )
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps, self.vcodec)
+
+    def _encode_temporary_episode_audio(self, audio_key: str, episode_index: int) -> Path:
+        """
+        Use ffmpeg to convert raw audio files into m4a audio files.
+        Note: `encode_episode_audio` is a blocking call. Making it asynchronous shouldn't speedup encoding,
+        since audio encoding with ffmpeg is already using multithreading.
+        """
+        temp_path = Path(tempfile.mkdtemp(dir=self.root)) / f"{audio_key}_{episode_index:03d}.m4a"
+        raw_audio_file = self._get_raw_audio_file_path(episode_index, audio_key)
+        encode_audio(raw_audio_file, temp_path, overwrite=True)
+        raw_audio_file.unlink()
+        return temp_path
 
     @classmethod
     def create(
@@ -1651,15 +1908,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
         image_writer_processes: int = 0,
         image_writer_threads: int = 0,
         video_backend: str | None = None,
+        audio_backend: str | None = None,
         batch_encoding_size: int = 1,
         vcodec: str = "libsvtav1",
-        metadata_buffer_size: int = 10,
-        streaming_encoding: bool = False,
-        encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
     ) -> "LeRobotDataset":
         """Create a LeRobot Dataset from scratch in order to record data."""
-        vcodec = resolve_vcodec(vcodec)
+        if vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -1668,7 +1923,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
             features=features,
             root=root,
             use_videos=use_videos,
-            metadata_buffer_size=metadata_buffer_size,
         )
         obj.repo_id = obj.meta.repo_id
         obj.root = obj.meta.root
@@ -1678,7 +1932,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.batch_encoding_size = batch_encoding_size
         obj.episodes_since_last_encoding = 0
         obj.vcodec = vcodec
-        obj._encoder_threads = encoder_threads
 
         if image_writer_processes or image_writer_threads:
             obj.start_image_writer(image_writer_processes, image_writer_threads)
@@ -1700,22 +1953,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj._lazy_loading = False
         obj._recorded_frames = 0
         obj._writer_closed_for_reading = False
-
-        # Initialize streaming encoder
-        if streaming_encoding and len(obj.meta.video_keys) > 0:
-            obj._streaming_encoder = StreamingVideoEncoder(
-                fps=fps,
-                vcodec=vcodec,
-                pix_fmt="yuv420p",
-                g=2,
-                crf=30,
-                preset=None,
-                queue_maxsize=encoder_queue_maxsize,
-                encoder_threads=encoder_threads,
-            )
-        else:
-            obj._streaming_encoder = None
-
+        obj.audio_backend = (
+            audio_backend if audio_backend is not None else "torchcodec"
+        )  # Waiting for torchcodec release #TODO(CarolinePascal)
         return obj
 
 
@@ -1736,6 +1976,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
         tolerances_s: dict | None = None,
         download_videos: bool = True,
         video_backend: str | None = None,
+        audio_backend: str | None = None,
     ):
         super().__init__()
         self.repo_ids = repo_ids
@@ -1753,6 +1994,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
                 tolerance_s=self.tolerances_s[repo_id],
                 download_videos=download_videos,
                 video_backend=video_backend,
+                audio_backend=audio_backend,
             )
             for repo_id in repo_ids
         ]
@@ -1771,12 +2013,11 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
             )
         for repo_id, ds in zip(self.repo_ids, self._datasets, strict=True):
             extra_keys = set(ds.features).difference(intersection_features)
-            if extra_keys:
-                logging.warning(
-                    f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
-                    "other datasets."
-                )
-                self.disabled_features.update(extra_keys)
+            logging.warning(
+                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                "other datasets."
+            )
+            self.disabled_features.update(extra_keys)
 
         self.image_transforms = image_transforms
         self.delta_timestamps = delta_timestamps

src/lerobot/datasets/transforms.py

@@ -216,17 +216,16 @@ class ImageTransformsConfig:
 
 
 def make_transform_from_config(cfg: ImageTransformConfig):
-    if cfg.type == "SharpnessJitter":
+    if cfg.type == "Identity":
+        return v2.Identity(**cfg.kwargs)
+    elif cfg.type == "ColorJitter":
+        return v2.ColorJitter(**cfg.kwargs)
+    elif cfg.type == "SharpnessJitter":
         return SharpnessJitter(**cfg.kwargs)
-
-    transform_cls = getattr(v2, cfg.type, None)
-    if isinstance(transform_cls, type) and issubclass(transform_cls, Transform):
-        return transform_cls(**cfg.kwargs)
-
-    raise ValueError(
-        f"Transform '{cfg.type}' is not valid. It must be a class in "
-        f"torchvision.transforms.v2 or 'SharpnessJitter'."
-    )
+    elif cfg.type == "RandomAffine":
+        return v2.RandomAffine(**cfg.kwargs)
+    else:
+        raise ValueError(f"Transform '{cfg.type}' is not valid.")
 
 
 class ImageTransforms(Transform):

src/lerobot/datasets/utils.py

@@ -36,6 +36,7 @@ from datasets.table import embed_table_storage
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
 from PIL import Image as PILImage
+from soundfile import read
 from torchvision import transforms
 
 from lerobot.configs.types import FeatureType, PolicyFeature
@@ -50,6 +51,7 @@ from lerobot.utils.utils import SuppressProgressBars, is_valid_numpy_dtype_strin
 DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
 DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
 DEFAULT_VIDEO_FILE_SIZE_IN_MB = 200  # Max size per file
+DEFAULT_AUDIO_FILE_SIZE_IN_MB = 100  # Max size per file
 
 INFO_PATH = "meta/info.json"
 STATS_PATH = "meta/stats.json"
@@ -57,14 +59,19 @@ STATS_PATH = "meta/stats.json"
 EPISODES_DIR = "meta/episodes"
 DATA_DIR = "data"
 VIDEO_DIR = "videos"
+AUDIO_DIR = "audio"
 
 CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
-DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
+DEFAULT_AUDIO_PATH = AUDIO_DIR + "/{audio_key}/" + CHUNK_FILE_PATTERN + ".m4a"
 DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{frame_index:06d}.png"
+DEFAULT_RAW_AUDIO_PATH = "raw_audio/{audio_key}/episode_{episode_index:06d}.wav"
+
+DEFAULT_AUDIO_CHUNK_DURATION = 0.5  # seconds
+DEFAULT_INITIAL_AUDIO_BUFFER_DURATION = 1.0  # seconds
 
 LEGACY_EPISODES_PATH = "meta/episodes.jsonl"
 LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
@@ -122,9 +129,19 @@ def load_nested_dataset(
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
     with SuppressProgressBars():
-        # We use .from_parquet() memory-mapped loading for efficiency
-        filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
-        return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)
+        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
+        # PyArrow loads the entire dataset into memory
+        if episodes is None:
+            return Dataset.from_parquet([str(path) for path in paths], features=features)
+
+        arrow_dataset = pa_ds.dataset(paths, format="parquet")
+        filter_expr = pa_ds.field("episode_index").isin(episodes)
+        table = arrow_dataset.to_table(filter=filter_expr)
+
+        if features is not None:
+            table = table.cast(features.arrow_schema)
+
+        return Dataset(table)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:
@@ -344,14 +361,6 @@ def load_tasks(local_dir: Path) -> pandas.DataFrame:
     return tasks
 
 
-def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
-    """Load subtasks from subtasks.parquet if it exists."""
-    subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
-    if subtasks_path.exists():
-        return pd.read_parquet(subtasks_path)
-    return None
-
-
 def write_episodes(episodes: Dataset, local_dir: Path) -> None:
     """Write episode metadata to a parquet file in the LeRobot v3.0 format.
     This function writes episode-level metadata to a single parquet file.
@@ -407,6 +416,16 @@ def load_image_as_numpy(
     return img_array
 
 
+def load_audio_from_path(fpath: str | Path) -> np.ndarray:
+    audio_data, _ = read(fpath, dtype="float32")
+
+    # Fill missing channel dimension when loading mono audio data
+    if audio_data.ndim == 1:
+        audio_data = np.expand_dims(audio_data, axis=1)
+
+    return audio_data
+
+
 def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[torch.Tensor | str]]:
     """Convert a batch from a Hugging Face dataset to torch tensors.
 
@@ -575,7 +594,7 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
     """
     hf_features = {}
     for key, ft in features.items():
-        if ft["dtype"] == "video":
+        if ft["dtype"] == "video" or ft["dtype"] == "audio":
             continue
         elif ft["dtype"] == "image":
             hf_features[key] = datasets.Image()
@@ -638,7 +657,12 @@ def hw_to_dataset_features(
         for key, ftype in hw_features.items()
         if ftype is float or (isinstance(ftype, PolicyFeature) and ftype.type != FeatureType.VISUAL)
     }
-    cam_fts = {key: shape for key, shape in hw_features.items() if isinstance(shape, tuple)}
+    cam_fts = {
+        key: shape for key, shape in hw_features.items() if isinstance(shape, tuple) and len(shape) == 3
+    }
+    mic_fts = {
+        key: shape for key, shape in hw_features.items() if isinstance(shape, tuple) and len(shape) == 2
+    }
 
     if joint_fts and prefix == ACTION:
         features[prefix] = {
@@ -661,6 +685,14 @@ def hw_to_dataset_features(
             "names": ["height", "width", "channels"],
         }
 
+    for key, parameters in mic_fts.items():
+        features[f"{prefix}.audio.{key}"] = {
+            "dtype": "audio",
+            "shape": (len(parameters[1]),),
+            "names": ["channels"],
+            "info": {"sample_rate": parameters[0]},
+        }
+
     _validate_feature_names(features)
     return features
 
@@ -690,6 +722,8 @@ def build_dataset_frame(
             frame[key] = np.array([values[name] for name in ft["names"]], dtype=np.float32)
         elif ft["dtype"] in ["image", "video"]:
             frame[key] = values[key.removeprefix(f"{prefix}.images.")]
+        elif ft["dtype"] == "audio":
+            frame[key] = values[key.removeprefix(f"{prefix}.audio.")]
 
     return frame
 
@@ -723,6 +757,10 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
             # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
             if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
                 shape = (shape[2], shape[0], shape[1])
+        elif ft["dtype"] == "audio":
+            type = FeatureType.AUDIO
+            if len(shape) != 2:
+                raise ValueError(f"Number of dimensions of {key} != 2 (shape={shape})")
         elif key == OBS_ENV_STATE:
             type = FeatureType.ENV
         elif key.startswith(OBS_STR):
@@ -801,6 +839,7 @@ def create_empty_dataset_info(
     chunks_size: int | None = None,
     data_files_size_in_mb: int | None = None,
     video_files_size_in_mb: int | None = None,
+    audio_files_size_in_mb: int | None = None,
 ) -> dict:
     """Create a template dictionary for a new dataset's `info.json`.
 
@@ -810,6 +849,10 @@ def create_empty_dataset_info(
         features (dict): The LeRobot features dictionary for the dataset.
         use_videos (bool): Whether the dataset will store videos.
         robot_type (str | None): The type of robot used, if any.
+        chunks_size (int | None): The maximum number of files per chunk directory.
+        data_files_size_in_mb (int | None): The maximum size for data files in MB.
+        video_files_size_in_mb (int | None): The maximum size for video files in MB.
+        audio_files_size_in_mb (int | None): The maximum size for audio files in MB.
 
     Returns:
         dict: A dictionary with the initial dataset metadata.
@@ -823,10 +866,12 @@ def create_empty_dataset_info(
         "chunks_size": chunks_size or DEFAULT_CHUNK_SIZE,
         "data_files_size_in_mb": data_files_size_in_mb or DEFAULT_DATA_FILE_SIZE_IN_MB,
         "video_files_size_in_mb": video_files_size_in_mb or DEFAULT_VIDEO_FILE_SIZE_IN_MB,
+        "audio_files_size_in_mb": audio_files_size_in_mb or DEFAULT_AUDIO_FILE_SIZE_IN_MB,
         "fps": fps,
         "splits": {},
         "data_path": DEFAULT_DATA_PATH,
         "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
+        "audio_path": DEFAULT_AUDIO_PATH,
         "features": features,
     }
 
@@ -1050,6 +1095,8 @@ def validate_feature_dtype_and_shape(
         return validate_feature_numpy_array(name, expected_dtype, expected_shape, value)
     elif expected_dtype in ["image", "video"]:
         return validate_feature_image_or_video(name, expected_shape, value)
+    elif expected_dtype == "audio":
+        return validate_feature_audio(name, expected_shape, value)
     elif expected_dtype == "string":
         return validate_feature_string(name, value)
     else:
@@ -1116,6 +1163,23 @@ def validate_feature_image_or_video(
     return error_message
 
 
+def validate_feature_audio(name: str, expected_shape: list[str], value: np.ndarray):
+    error_message = ""
+    if isinstance(value, np.ndarray):
+        actual_shape = value.shape
+        c = expected_shape
+        if (len(actual_shape) != 2 and len(actual_shape) != 1) or actual_shape[-1] != c[
+            -1
+        ]:  # The number of frames might be different
+            error_message += (
+                f"The feature '{name}' of shape '{actual_shape}' does not have the expected shape '{c}'.\n"
+            )
+    else:
+        error_message += f"The feature '{name}' is expected to be of type 'np.ndarray', but type '{type(value)}' provided instead.\n"
+
+    return error_message
+
+
 def validate_feature_string(name: str, value: str) -> str:
     """Validate a feature that is expected to be a string.
 

src/lerobot/datasets/v30/convert_dataset_v21_to_v30.py

@@ -59,6 +59,8 @@ from requests import HTTPError
 from lerobot.datasets.compute_stats import aggregate_stats
 from lerobot.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
 from lerobot.datasets.utils import (
+    DEFAULT_AUDIO_FILE_SIZE_IN_MB,
+    DEFAULT_AUDIO_PATH,
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
@@ -79,7 +81,7 @@ from lerobot.datasets.utils import (
     write_stats,
     write_tasks,
 )
-from lerobot.datasets.video_utils import concatenate_video_files, get_video_duration_in_s
+from lerobot.datasets.video_utils import concatenate_media_files, get_media_duration_in_s
 from lerobot.utils.constants import HF_LEROBOT_HOME
 from lerobot.utils.utils import init_logging
 
@@ -311,12 +313,12 @@ def convert_videos_of_camera(root: Path, new_root: Path, video_key: str, video_f
 
     for ep_path in tqdm.tqdm(ep_paths, desc=f"convert videos of {video_key}"):
         ep_size_in_mb = get_file_size_in_mb(ep_path)
-        ep_duration_in_s = get_video_duration_in_s(ep_path)
+        ep_duration_in_s = get_media_duration_in_s(ep_path, media_type="video")
 
         # Check if adding this episode would exceed the limit
         if size_in_mb + ep_size_in_mb >= video_file_size_in_mb and len(paths_to_cat) > 0:
             # Size limit would be exceeded, save current accumulation WITHOUT this episode
-            concatenate_video_files(
+            concatenate_media_files(
                 paths_to_cat,
                 new_root
                 / DEFAULT_VIDEO_PATH.format(video_key=video_key, chunk_index=chunk_idx, file_index=file_idx),
@@ -352,7 +354,7 @@ def convert_videos_of_camera(root: Path, new_root: Path, video_key: str, video_f
 
     # Write remaining videos if any
     if paths_to_cat:
-        concatenate_video_files(
+        concatenate_media_files(
             paths_to_cat,
             new_root
             / DEFAULT_VIDEO_PATH.format(video_key=video_key, chunk_index=chunk_idx, file_index=file_idx),
@@ -367,8 +369,124 @@ def convert_videos_of_camera(root: Path, new_root: Path, video_key: str, video_f
     return episodes_metadata
 
 
+def get_audio_keys(root):
+    info = load_info(root)
+    features = info["features"]
+    audio_keys = [key for key, ft in features.items() if ft["dtype"] == "audio"]
+    return audio_keys
+
+
+def convert_audios(root: Path, new_root: Path, audio_file_size_in_mb: int):
+    logging.info(f"Converting audios from {root} to {new_root}")
+
+    audio_keys = get_audio_keys(root)
+    if len(audio_keys) == 0:
+        return None
+
+    audio_keys = sorted(audio_keys)
+
+    eps_metadata_per_mic = []
+    for microphone in audio_keys:
+        eps_metadata = convert_audios_of_microphone(root, new_root, microphone, audio_file_size_in_mb)
+        eps_metadata_per_mic.append(eps_metadata)
+
+    num_eps_per_mic = [len(eps_mic_map) for eps_mic_map in eps_metadata_per_mic]
+    if len(set(num_eps_per_mic)) != 1:
+        raise ValueError(f"All microphones dont have same number of episodes ({num_eps_per_mic}).")
+
+    episodes_metadata = []
+    num_microphones = len(audio_keys)
+    num_episodes = num_eps_per_mic[0]
+    for ep_idx in tqdm.tqdm(range(num_episodes), desc="convert audios"):
+        # Sanity check
+        ep_ids = [
+            eps_metadata_per_mic[mic_idx][ep_idx]["episode_index"] for mic_idx in range(num_microphones)
+        ]
+        ep_ids += [ep_idx]
+        if len(set(ep_ids)) != 1:
+            raise ValueError(f"All episode indices need to match ({ep_ids}).")
+
+        ep_dict = {}
+        for mic_idx in range(num_microphones):
+            ep_dict.update(eps_metadata_per_mic[mic_idx][ep_idx])
+        episodes_metadata.append(ep_dict)
+
+    return episodes_metadata
+
+
+def convert_audios_of_microphone(root: Path, new_root: Path, audio_key: str, audio_file_size_in_mb: int):
+    # Access old paths to m4a
+    audios_dir = root / "audio"
+    ep_paths = sorted(audios_dir.glob(f"*/{audio_key}/*.m4a"))
+
+    ep_idx = 0
+    chunk_idx = 0
+    file_idx = 0
+    size_in_mb = 0
+    duration_in_s = 0.0
+    paths_to_cat = []
+    episodes_metadata = []
+
+    for ep_path in tqdm.tqdm(ep_paths, desc=f"convert audios of {audio_key}"):
+        ep_size_in_mb = get_file_size_in_mb(ep_path)
+        ep_duration_in_s = get_media_duration_in_s(ep_path, media_type="audio")
+
+        # Check if adding this episode would exceed the limit
+        if size_in_mb + ep_size_in_mb >= audio_file_size_in_mb and len(paths_to_cat) > 0:
+            # Size limit would be exceeded, save current accumulation WITHOUT this episode
+            concatenate_media_files(
+                paths_to_cat,
+                new_root
+                / DEFAULT_AUDIO_PATH.format(audio_key=audio_key, chunk_index=chunk_idx, file_index=file_idx),
+            )
+
+            # Update episodes metadata for the file we just saved
+            for i, _ in enumerate(paths_to_cat):
+                past_ep_idx = ep_idx - len(paths_to_cat) + i
+                episodes_metadata[past_ep_idx][f"audio/{audio_key}/chunk_index"] = chunk_idx
+                episodes_metadata[past_ep_idx][f"audio/{audio_key}/file_index"] = file_idx
+
+            # Move to next file and start fresh with current episode
+            chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
+            size_in_mb = 0
+            duration_in_s = 0.0
+            paths_to_cat = []
+
+        # Add current episode metadata
+        ep_metadata = {
+            "episode_index": ep_idx,
+            f"audio/{audio_key}/chunk_index": chunk_idx,  # Will be updated when file is saved
+            f"audio/{audio_key}/file_index": file_idx,  # Will be updated when file is saved
+            f"audio/{audio_key}/from_timestamp": duration_in_s,
+            f"audio/{audio_key}/to_timestamp": duration_in_s + ep_duration_in_s,
+        }
+        episodes_metadata.append(ep_metadata)
+
+        # Add current episode to accumulation
+        paths_to_cat.append(ep_path)
+        size_in_mb += ep_size_in_mb
+        duration_in_s += ep_duration_in_s
+        ep_idx += 1
+
+    # Write remaining videos if any
+    if paths_to_cat:
+        concatenate_media_files(
+            paths_to_cat,
+            new_root
+            / DEFAULT_AUDIO_PATH.format(audio_key=audio_key, chunk_index=chunk_idx, file_index=file_idx),
+        )
+
+        # Update episodes metadata for the final file
+        for i, _ in enumerate(paths_to_cat):
+            past_ep_idx = ep_idx - len(paths_to_cat) + i
+            episodes_metadata[past_ep_idx][f"audio/{audio_key}/chunk_index"] = chunk_idx
+            episodes_metadata[past_ep_idx][f"audio/{audio_key}/file_index"] = file_idx
+
+    return episodes_metadata
+
+
 def generate_episode_metadata_dict(
-    episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_videos=None
+    episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_videos=None, episodes_audios=None
 ):
     num_episodes = len(episodes_metadata)
     episodes_legacy_metadata_vals = list(episodes_legacy_metadata.values())
@@ -392,16 +510,30 @@ def generate_episode_metadata_dict(
             ep_video = episodes_videos[i]
             ep_ids_set.add(ep_video["episode_index"])
 
+        if episodes_audios is None:
+            ep_audio = {}
+        else:
+            ep_audio = episodes_audios[i]
+            ep_ids_set.add(ep_audio["episode_index"])
+
         if len(ep_ids_set) != 1:
             raise ValueError(f"Number of episodes is not the same ({ep_ids_set}).")
 
-        ep_dict = {**ep_metadata, **ep_video, **ep_legacy_metadata, **flatten_dict({"stats": ep_stats})}
+        ep_dict = {
+            **ep_metadata,
+            **ep_video,
+            **ep_audio,
+            **ep_legacy_metadata,
+            **flatten_dict({"stats": ep_stats}),
+        }
         ep_dict["meta/episodes/chunk_index"] = 0
         ep_dict["meta/episodes/file_index"] = 0
         yield ep_dict
 
 
-def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_metadata=None):
+def convert_episodes_metadata(
+    root, new_root, episodes_metadata, episodes_video_metadata=None, episodes_audio_metadata=None
+):
     logging.info(f"Converting episodes metadata from {root} to {new_root}")
 
     episodes_legacy_metadata = legacy_load_episodes(root)
@@ -410,13 +542,19 @@ def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_
     num_eps_set = {len(episodes_legacy_metadata), len(episodes_metadata)}
     if episodes_video_metadata is not None:
         num_eps_set.add(len(episodes_video_metadata))
+    if episodes_audio_metadata is not None:
+        num_eps_set.add(len(episodes_audio_metadata))
 
     if len(num_eps_set) != 1:
         raise ValueError(f"Number of episodes is not the same ({num_eps_set}).")
 
     ds_episodes = Dataset.from_generator(
         lambda: generate_episode_metadata_dict(
-            episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_video_metadata
+            episodes_legacy_metadata,
+            episodes_metadata,
+            episodes_stats,
+            episodes_video_metadata,
+            episodes_audio_metadata,
         )
     )
     write_episodes(ds_episodes, new_root)
@@ -425,20 +563,22 @@ def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_
     write_stats(stats, new_root)
 
 
-def convert_info(root, new_root, data_file_size_in_mb, video_file_size_in_mb):
+def convert_info(root, new_root, data_file_size_in_mb, video_file_size_in_mb, audio_file_size_in_mb):
     info = load_info(root)
     info["codebase_version"] = V30
     del info["total_chunks"]
     del info["total_videos"]
     info["data_files_size_in_mb"] = data_file_size_in_mb
     info["video_files_size_in_mb"] = video_file_size_in_mb
+    info["audio_files_size_in_mb"] = audio_file_size_in_mb
     info["data_path"] = DEFAULT_DATA_PATH
     info["video_path"] = DEFAULT_VIDEO_PATH if info["video_path"] is not None else None
+    info["audio_path"] = DEFAULT_AUDIO_PATH if info["audio_path"] is not None else None
     info["fps"] = int(info["fps"])
     logging.info(f"Converting info from {root} to {new_root}")
     for key in info["features"]:
-        if info["features"][key]["dtype"] == "video":
-            # already has fps in video_info
+        if info["features"][key]["dtype"] == "video" or info["features"][key]["dtype"] == "audio":
+            # already has fps in video_info or audio_info
             continue
         info["features"][key]["fps"] = info["fps"]
     write_info(info, new_root)
@@ -449,6 +589,7 @@ def convert_dataset(
     branch: str | None = None,
     data_file_size_in_mb: int | None = None,
     video_file_size_in_mb: int | None = None,
+    audio_file_size_in_mb: int | None = None,
     root: str | Path | None = None,
     push_to_hub: bool = True,
     force_conversion: bool = False,
@@ -457,6 +598,8 @@ def convert_dataset(
         data_file_size_in_mb = DEFAULT_DATA_FILE_SIZE_IN_MB
     if video_file_size_in_mb is None:
         video_file_size_in_mb = DEFAULT_VIDEO_FILE_SIZE_IN_MB
+    if audio_file_size_in_mb is None:
+        audio_file_size_in_mb = DEFAULT_AUDIO_FILE_SIZE_IN_MB
 
     # First check if the dataset already has a v3.0 version
     if root is None and not force_conversion:
@@ -498,7 +641,10 @@ def convert_dataset(
     convert_tasks(root, new_root)
     episodes_metadata = convert_data(root, new_root, data_file_size_in_mb)
     episodes_videos_metadata = convert_videos(root, new_root, video_file_size_in_mb)
-    convert_episodes_metadata(root, new_root, episodes_metadata, episodes_videos_metadata)
+    episodes_audios_metadata = convert_audios(root, new_root, audio_file_size_in_mb)
+    convert_episodes_metadata(
+        root, new_root, episodes_metadata, episodes_videos_metadata, episodes_audios_metadata
+    )
 
     shutil.move(str(root), str(old_root))
     shutil.move(str(new_root), str(root))
@@ -511,7 +657,7 @@ def convert_dataset(
             print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
             pass
         hub_api.delete_files(
-            delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*"],
+            delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*", "audio/chunk*"],
             repo_id=repo_id,
             revision=branch,
             repo_type="dataset",
@@ -529,7 +675,7 @@ if __name__ == "__main__":
         type=str,
         required=True,
         help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset "
-        "(e.g. `lerobot/pusht`, `<USER>/aloha_sim_insertion_human`).",
+        "(e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
     )
     parser.add_argument(
         "--branch",
@@ -549,6 +695,12 @@ if __name__ == "__main__":
         default=None,
         help="File size in MB. Defaults to 100 for data and 500 for videos.",
     )
+    parser.add_argument(
+        "--audio-file-size-in-mb",
+        type=int,
+        default=None,
+        help="File size in MB. Defaults to 100 for audio.",
+    )
     parser.add_argument(
         "--root",
         type=str,

src/lerobot/datasets/video_utils.py

@@ -13,106 +13,25 @@
 # 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 contextlib
 import glob
 import importlib
 import logging
-import queue
 import shutil
 import tempfile
-import threading
 import warnings
 from dataclasses import dataclass, field
-from fractions import Fraction
 from pathlib import Path
 from threading import Lock
 from typing import Any, ClassVar
 
 import av
 import fsspec
-import numpy as np
 import pyarrow as pa
 import torch
 import torchvision
 from datasets.features.features import register_feature
 from PIL import Image
 
-# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
-# Determines the order of preference for auto-selection when vcodec="auto" is used.
-HW_ENCODERS = [
-    "h264_videotoolbox",  # macOS
-    "hevc_videotoolbox",  # macOS
-    "h264_nvenc",  # NVIDIA GPU
-    "hevc_nvenc",  # NVIDIA GPU
-    "h264_vaapi",  # Linux Intel/AMD
-    "h264_qsv",  # Intel Quick Sync
-]
-
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
-
-
-def _get_codec_options(
-    vcodec: str,
-    g: int | None = 2,
-    crf: int | None = 30,
-    preset: int | None = None,
-) -> dict:
-    """Build codec-specific options dict for video encoding."""
-    options = {}
-
-    # GOP size (keyframe interval) - supported by VideoToolbox and software encoders
-    if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
-        options["g"] = str(g)
-
-    # Quality control (codec-specific parameter names)
-    if crf is not None:
-        if vcodec in ("h264", "hevc", "libsvtav1"):
-            options["crf"] = str(crf)
-        elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
-            quality = max(1, min(100, int(100 - crf * 2)))
-            options["q:v"] = str(quality)
-        elif vcodec in ("h264_nvenc", "hevc_nvenc"):
-            options["rc"] = "constqp"
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_vaapi",):
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_qsv",):
-            options["global_quality"] = str(crf)
-
-    # Preset (only for libsvtav1)
-    if vcodec == "libsvtav1":
-        options["preset"] = str(preset) if preset is not None else "12"
-
-    return options
-
-
-def detect_available_hw_encoders() -> list[str]:
-    """Probe PyAV/FFmpeg for available hardware video encoders."""
-    available = []
-    for codec_name in HW_ENCODERS:
-        try:
-            av.codec.Codec(codec_name, "w")
-            available.append(codec_name)
-        except Exception:  # nosec B110
-            pass  # nosec B110
-    return available
-
-
-def resolve_vcodec(vcodec: str) -> str:
-    """Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
-    if vcodec not in VALID_VIDEO_CODECS:
-        raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
-    if vcodec != "auto":
-        logging.info(f"Using video codec: {vcodec}")
-        return vcodec
-    available = detect_available_hw_encoders()
-    for encoder in HW_ENCODERS:
-        if encoder in available:
-            logging.info(f"Auto-selected video codec: {encoder}")
-            return encoder
-    logging.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
-    return "libsvtav1"
-
 
 def get_safe_default_codec():
     if importlib.util.find_spec("torchcodec"):
@@ -227,17 +146,16 @@ def decode_video_frames_torchvision(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    if not is_within_tol.all():
-        raise FrameTimestampError(
-            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-            " It means that the closest frame that can be loaded from the video is too far away in time."
-            " This might be due to synchronization issues with timestamps during data collection."
-            " To be safe, we advise to ignore this item during training."
-            f"\nqueried timestamps: {query_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
-            f"\nvideo: {video_path}"
-            f"\nbackend: {backend}"
-        )
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
+        f"\nbackend: {backend}"
+    )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -249,11 +167,7 @@ def decode_video_frames_torchvision(
     # convert to the pytorch format which is float32 in [0,1] range (and channel first)
     closest_frames = closest_frames.type(torch.float32) / 255
 
-    if len(timestamps) != len(closest_frames):
-        raise FrameTimestampError(
-            f"Number of retrieved frames ({len(closest_frames)}) does not match "
-            f"number of queried timestamps ({len(timestamps)})"
-        )
+    assert len(timestamps) == len(closest_frames)
     return closest_frames
 
 
@@ -358,16 +272,15 @@ def decode_video_frames_torchcodec(
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
-    if not is_within_tol.all():
-        raise FrameTimestampError(
-            f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
-            " It means that the closest frame that can be loaded from the video is too far away in time."
-            " This might be due to synchronization issues with timestamps during data collection."
-            " To be safe, we advise to ignore this item during training."
-            f"\nqueried timestamps: {query_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
-            f"\nvideo: {video_path}"
-        )
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
+    )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
@@ -396,13 +309,14 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: int | None = av.logging.WARNING,
+    log_level: int | None = av.logging.ERROR,
     overwrite: bool = False,
     preset: int | None = None,
-    encoder_threads: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    vcodec = resolve_vcodec(vcodec)
+    # Check encoder availability
+    if vcodec not in ["h264", "hevc", "libsvtav1"]:
+        raise ValueError(f"Unsupported video codec: {vcodec}. Supported codecs are: h264, hevc, libsvtav1.")
 
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
@@ -433,22 +347,21 @@ def encode_video_frames(
         width, height = dummy_image.size
 
     # Define video codec options
-    video_options = _get_codec_options(vcodec, g, crf, preset)
+    video_options = {}
+
+    if g is not None:
+        video_options["g"] = str(g)
+
+    if crf is not None:
+        video_options["crf"] = str(crf)
 
     if fast_decode:
         key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
         video_options[key] = value
 
-    if encoder_threads is not None:
-        if vcodec == "libsvtav1":
-            lp_param = f"lp={encoder_threads}"
-            if "svtav1-params" in video_options:
-                video_options["svtav1-params"] += f":{lp_param}"
-            else:
-                video_options["svtav1-params"] = lp_param
-        else:
-            video_options["threads"] = str(encoder_threads)
+    if vcodec == "libsvtav1":
+        video_options["preset"] = str(preset) if preset is not None else "12"
 
     # Set logging level
     if log_level is not None:
@@ -484,42 +397,42 @@ def encode_video_frames(
         raise OSError(f"Video encoding did not work. File not found: {video_path}.")
 
 
-def concatenate_video_files(
-    input_video_paths: list[Path | str], output_video_path: Path, overwrite: bool = True
+def concatenate_media_files(
+    input_media_paths: list[Path | str], output_media_path: Path, overwrite: bool = True
 ):
     """
-    Concatenate multiple video files into a single video file using pyav.
+    Concatenate multiple media files (video & audio) into a single media file using pyav.
 
-    This function takes a list of video input file paths and concatenates them into a single
-    output video file. It uses ffmpeg's concat demuxer with stream copy mode for fast
+    This function takes a list of input media file paths and concatenates them into a single
+    output media file. It uses ffmpeg's concat demuxer with stream copy mode for fast
     concatenation without re-encoding.
 
     Args:
-        input_video_paths: Ordered list of input video file paths to concatenate.
-        output_video_path: Path to the output video file.
-        overwrite: Whether to overwrite the output video file if it already exists. Default is True.
+        input_media_paths: Ordered list of input media file paths to concatenate.
+        output_media_path: Path to the output media file.
+        overwrite: Whether to overwrite the output media file if it already exists. Default is True.
 
     Note:
-        - Creates a temporary directory for intermediate files that is cleaned up after use.
-        - Uses ffmpeg's concat demuxer which requires all input videos to have the same
+        - Creates a temporary .ffconcat file and container audio/video file that are cleaned up after use.
+        - Uses ffmpeg's concat demuxer which requires all input media files to have the same
           codec, resolution, and frame rate for proper concatenation.
     """
 
-    output_video_path = Path(output_video_path)
+    output_media_path = Path(output_media_path)
 
-    if output_video_path.exists() and not overwrite:
-        logging.warning(f"Video file already exists: {output_video_path}. Skipping concatenation.")
+    if output_media_path.exists() and not overwrite:
+        logging.warning(f"Media file already exists: {output_media_path}. Skipping concatenation.")
         return
 
-    output_video_path.parent.mkdir(parents=True, exist_ok=True)
+    output_media_path.parent.mkdir(parents=True, exist_ok=True)
 
-    if len(input_video_paths) == 0:
-        raise FileNotFoundError("No input video paths provided.")
+    if len(input_media_paths) == 0:
+        raise FileNotFoundError("No input media paths provided.")
 
-    # Create a temporary .ffconcat file to list the input video paths
+    # Create a temporary .ffconcat file to list the input media paths
     with tempfile.NamedTemporaryFile(mode="w", suffix=".ffconcat", delete=False) as tmp_concatenate_file:
         tmp_concatenate_file.write("ffconcat version 1.0\n")
-        for input_path in input_video_paths:
+        for input_path in input_media_paths:
             tmp_concatenate_file.write(f"file '{str(input_path.resolve())}'\n")
         tmp_concatenate_file.flush()
         tmp_concatenate_path = tmp_concatenate_file.name
@@ -529,11 +442,12 @@ def concatenate_video_files(
         tmp_concatenate_path, mode="r", format="concat", options={"safe": "0"}
     )  # safe = 0 allows absolute paths as well as relative paths
 
-    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_named_file:
-        tmp_output_video_path = tmp_named_file.name
+    # Using an intermediate container to store the concatenated media file is necessary to avoid inplace concatenation read-write race conditions.
+    with tempfile.NamedTemporaryFile(suffix=output_media_path.suffix, delete=False) as tmp_named_file:
+        tmp_output_media_path = tmp_named_file.name
 
     output_container = av.open(
-        tmp_output_video_path, mode="w", options={"movflags": "faststart"}
+        tmp_output_media_path, mode="w", options={"movflags": "faststart"}
     )  # faststart is to move the metadata to the beginning of the file to speed up loading
 
     # Replicate input streams in output container
@@ -548,6 +462,7 @@ def concatenate_video_files(
             stream_map[input_stream.index].time_base = input_stream.time_base
 
     # Demux + remux packets (no re-encode)
+    last_dts = None
     for packet in input_container.demux():
         # Skip packets from un-mapped streams
         if packet.stream.index not in stream_map:
@@ -556,6 +471,16 @@ def concatenate_video_files(
         # Skip demux flushing packets
         if packet.dts is None:
             continue
+        else:
+            # Enforce strictly increasing decoding timestamps (DTS)
+            if last_dts is not None and packet.dts <= last_dts:
+                shift = last_dts - packet.dts + 1
+                packet.dts += shift
+                packet.pts += shift  # Presenting timestamps (PTS) are the same as DTS here
+                logging.warning(
+                    f"Non-monotonic DTS; previous: {last_dts}, current: {packet.dts - shift}; changing to {packet.dts}. This may result in incorrect timestamps in the output file."
+                )
+            last_dts = packet.dts
 
         output_stream = stream_map[packet.stream.index]
         packet.stream = output_stream
@@ -563,352 +488,10 @@ def concatenate_video_files(
 
     input_container.close()
     output_container.close()
-    shutil.move(tmp_output_video_path, output_video_path)
+    shutil.move(tmp_output_media_path, output_media_path)
     Path(tmp_concatenate_path).unlink()
 
 
-class _CameraEncoderThread(threading.Thread):
-    """A thread that encodes video frames streamed via a queue into an MP4 file.
-
-    One instance is created per camera per episode. Frames are received as numpy arrays
-    from the main thread, encoded in real-time using PyAV (which releases the GIL during
-    encoding), and written to disk. Stats are computed incrementally using
-    RunningQuantileStats and returned via result_queue.
-    """
-
-    def __init__(
-        self,
-        video_path: Path,
-        fps: int,
-        vcodec: str,
-        pix_fmt: str,
-        g: int | None,
-        crf: int | None,
-        preset: int | None,
-        frame_queue: queue.Queue,
-        result_queue: queue.Queue,
-        stop_event: threading.Event,
-        encoder_threads: int | None = None,
-    ):
-        super().__init__(daemon=True)
-        self.video_path = video_path
-        self.fps = fps
-        self.vcodec = vcodec
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
-        self.frame_queue = frame_queue
-        self.result_queue = result_queue
-        self.stop_event = stop_event
-        self.encoder_threads = encoder_threads
-
-    def run(self) -> None:
-        from lerobot.datasets.compute_stats import RunningQuantileStats, auto_downsample_height_width
-
-        container = None
-        output_stream = None
-        stats_tracker = RunningQuantileStats()
-        frame_count = 0
-
-        try:
-            logging.getLogger("libav").setLevel(av.logging.WARNING)
-
-            while True:
-                try:
-                    frame_data = self.frame_queue.get(timeout=1)
-                except queue.Empty:
-                    if self.stop_event.is_set():
-                        break
-                    continue
-
-                if frame_data is None:
-                    # Sentinel: flush and close
-                    break
-
-                # Ensure HWC uint8 numpy array
-                if isinstance(frame_data, np.ndarray):
-                    if frame_data.ndim == 3 and frame_data.shape[0] == 3:
-                        # CHW -> HWC
-                        frame_data = frame_data.transpose(1, 2, 0)
-                    if frame_data.dtype != np.uint8:
-                        frame_data = (frame_data * 255).astype(np.uint8)
-
-                # Open container on first frame (to get width/height)
-                if container is None:
-                    height, width = frame_data.shape[:2]
-                    video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
-                    if self.encoder_threads is not None:
-                        if self.vcodec == "libsvtav1":
-                            lp_param = f"lp={self.encoder_threads}"
-                            if "svtav1-params" in video_options:
-                                video_options["svtav1-params"] += f":{lp_param}"
-                            else:
-                                video_options["svtav1-params"] = lp_param
-                        else:
-                            video_options["threads"] = str(self.encoder_threads)
-                    Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
-                    container = av.open(str(self.video_path), "w")
-                    output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
-                    output_stream.pix_fmt = self.pix_fmt
-                    output_stream.width = width
-                    output_stream.height = height
-                    output_stream.time_base = Fraction(1, self.fps)
-
-                # Encode frame with explicit timestamps
-                pil_img = Image.fromarray(frame_data)
-                video_frame = av.VideoFrame.from_image(pil_img)
-                video_frame.pts = frame_count
-                video_frame.time_base = Fraction(1, self.fps)
-                packet = output_stream.encode(video_frame)
-                if packet:
-                    container.mux(packet)
-
-                # Update stats with downsampled frame (per-channel stats like compute_episode_stats)
-                img_chw = frame_data.transpose(2, 0, 1)  # HWC -> CHW
-                img_downsampled = auto_downsample_height_width(img_chw)
-                # Reshape CHW to (H*W, C) for per-channel stats
-                channels = img_downsampled.shape[0]
-                img_for_stats = img_downsampled.transpose(1, 2, 0).reshape(-1, channels)
-                stats_tracker.update(img_for_stats)
-
-                frame_count += 1
-
-            # Flush encoder
-            if output_stream is not None:
-                packet = output_stream.encode()
-                if packet:
-                    container.mux(packet)
-
-            if container is not None:
-                container.close()
-
-            av.logging.restore_default_callback()
-
-            # Get stats and put on result queue
-            if frame_count >= 2:
-                stats = stats_tracker.get_statistics()
-                self.result_queue.put(("ok", stats))
-            else:
-                self.result_queue.put(("ok", None))
-
-        except Exception as e:
-            logging.error(f"Encoder thread error: {e}")
-            if container is not None:
-                with contextlib.suppress(Exception):
-                    container.close()
-            self.result_queue.put(("error", str(e)))
-
-
-class StreamingVideoEncoder:
-    """Manages per-camera encoder threads for real-time video encoding during recording.
-
-    Instead of writing frames as PNG images and then encoding to MP4 at episode end,
-    this class streams frames directly to encoder threads, eliminating the
-    PNG round-trip and making save_episode() near-instant.
-
-    Uses threading instead of multiprocessing to avoid the overhead of pickling large
-    numpy arrays through multiprocessing.Queue. PyAV's encode() releases the GIL,
-    so encoding runs in parallel with the main recording loop.
-    """
-
-    def __init__(
-        self,
-        fps: int,
-        vcodec: str = "libsvtav1",
-        pix_fmt: str = "yuv420p",
-        g: int | None = 2,
-        crf: int | None = 30,
-        preset: int | None = None,
-        queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
-    ):
-        self.fps = fps
-        self.vcodec = resolve_vcodec(vcodec)
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
-        self.queue_maxsize = queue_maxsize
-        self.encoder_threads = encoder_threads
-
-        self._frame_queues: dict[str, queue.Queue] = {}
-        self._result_queues: dict[str, queue.Queue] = {}
-        self._threads: dict[str, _CameraEncoderThread] = {}
-        self._stop_events: dict[str, threading.Event] = {}
-        self._video_paths: dict[str, Path] = {}
-        self._dropped_frames: dict[str, int] = {}
-        self._episode_active = False
-
-    def start_episode(self, video_keys: list[str], temp_dir: Path) -> None:
-        """Start encoder threads for a new episode.
-
-        Args:
-            video_keys: List of video feature keys (e.g. ["observation.images.laptop"])
-            temp_dir: Base directory for temporary MP4 files
-        """
-        if self._episode_active:
-            self.cancel_episode()
-
-        self._dropped_frames.clear()
-
-        for video_key in video_keys:
-            frame_queue: queue.Queue = queue.Queue(maxsize=self.queue_maxsize)
-            result_queue: queue.Queue = queue.Queue(maxsize=1)
-            stop_event = threading.Event()
-
-            temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
-            video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"
-
-            encoder_thread = _CameraEncoderThread(
-                video_path=video_path,
-                fps=self.fps,
-                vcodec=self.vcodec,
-                pix_fmt=self.pix_fmt,
-                g=self.g,
-                crf=self.crf,
-                preset=self.preset,
-                frame_queue=frame_queue,
-                result_queue=result_queue,
-                stop_event=stop_event,
-                encoder_threads=self.encoder_threads,
-            )
-            encoder_thread.start()
-
-            self._frame_queues[video_key] = frame_queue
-            self._result_queues[video_key] = result_queue
-            self._threads[video_key] = encoder_thread
-            self._stop_events[video_key] = stop_event
-            self._video_paths[video_key] = video_path
-
-        self._episode_active = True
-
-    def feed_frame(self, video_key: str, image: np.ndarray) -> None:
-        """Feed a frame to the encoder for a specific camera.
-
-        A copy of the image is made before enqueueing to prevent race conditions
-        with camera drivers that may reuse buffers. If the encoder queue is full
-        (encoder can't keep up), the frame is dropped with a warning instead of
-        crashing the recording session.
-
-        Args:
-            video_key: The video feature key
-            image: numpy array in (H,W,C) or (C,H,W) format, uint8 or float
-
-        Raises:
-            RuntimeError: If the encoder thread has crashed
-        """
-        if not self._episode_active:
-            raise RuntimeError("No active episode. Call start_episode() first.")
-
-        thread = self._threads[video_key]
-        if not thread.is_alive():
-            # Check for error
-            try:
-                status, msg = self._result_queues[video_key].get_nowait()
-                if status == "error":
-                    raise RuntimeError(f"Encoder thread for {video_key} crashed: {msg}")
-            except queue.Empty:
-                pass
-            raise RuntimeError(f"Encoder thread for {video_key} is not alive")
-
-        try:
-            self._frame_queues[video_key].put(image.copy(), timeout=0.1)
-        except queue.Full:
-            self._dropped_frames[video_key] = self._dropped_frames.get(video_key, 0) + 1
-            count = self._dropped_frames[video_key]
-            # Log periodically to avoid spam (1st, then every 10th)
-            if count == 1 or count % 10 == 0:
-                logging.warning(
-                    f"Encoder queue full for {video_key}, dropped {count} frame(s). "
-                    f"Consider using vcodec='auto' for hardware encoding or increasing encoder_queue_maxsize."
-                )
-
-    def finish_episode(self) -> dict[str, tuple[Path, dict | None]]:
-        """Finish encoding the current episode.
-
-        Sends sentinel values, waits for encoder threads to complete,
-        and collects results.
-
-        Returns:
-            Dict mapping video_key to (mp4_path, stats_dict_or_None)
-        """
-        if not self._episode_active:
-            raise RuntimeError("No active episode to finish.")
-
-        results = {}
-
-        # Report dropped frames
-        for video_key, count in self._dropped_frames.items():
-            if count > 0:
-                logging.warning(f"Episode finished with {count} dropped frame(s) for {video_key}.")
-
-        # Send sentinel to all queues
-        for video_key in self._frame_queues:
-            self._frame_queues[video_key].put(None)
-
-        # Wait for all threads and collect results
-        for video_key in self._threads:
-            self._threads[video_key].join(timeout=120)
-            if self._threads[video_key].is_alive():
-                logging.error(f"Encoder thread for {video_key} did not finish in time")
-                self._stop_events[video_key].set()
-                self._threads[video_key].join(timeout=5)
-                results[video_key] = (self._video_paths[video_key], None)
-                continue
-
-            try:
-                status, data = self._result_queues[video_key].get(timeout=5)
-                if status == "error":
-                    raise RuntimeError(f"Encoder thread for {video_key} failed: {data}")
-                results[video_key] = (self._video_paths[video_key], data)
-            except queue.Empty:
-                logging.error(f"No result from encoder thread for {video_key}")
-                results[video_key] = (self._video_paths[video_key], None)
-
-        self._cleanup()
-        self._episode_active = False
-        return results
-
-    def cancel_episode(self) -> None:
-        """Cancel the current episode, stopping encoder threads and cleaning up."""
-        if not self._episode_active:
-            return
-
-        # Signal all threads to stop
-        for video_key in self._stop_events:
-            self._stop_events[video_key].set()
-
-        # Wait for threads to finish
-        for video_key in self._threads:
-            self._threads[video_key].join(timeout=5)
-
-            # Clean up temp MP4 files
-            video_path = self._video_paths.get(video_key)
-            if video_path is not None and video_path.exists():
-                shutil.rmtree(str(video_path.parent), ignore_errors=True)
-
-        self._cleanup()
-        self._episode_active = False
-
-    def close(self) -> None:
-        """Close the encoder, canceling any in-progress episode."""
-        if self._episode_active:
-            self.cancel_episode()
-
-    def _cleanup(self) -> None:
-        """Clean up queues and thread tracking dicts."""
-        for q in self._frame_queues.values():
-            with contextlib.suppress(Exception):
-                while not q.empty():
-                    q.get_nowait()
-        self._frame_queues.clear()
-        self._result_queues.clear()
-        self._threads.clear()
-        self._stop_events.clear()
-        self._video_paths.clear()
-
-
 @dataclass
 class VideoFrame:
     # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
@@ -941,41 +524,9 @@ with warnings.catch_warnings():
     register_feature(VideoFrame, "VideoFrame")
 
 
-def get_audio_info(video_path: Path | str) -> dict:
-    # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
-
-    # Getting audio stream information
-    audio_info = {}
-    with av.open(str(video_path), "r") as audio_file:
-        try:
-            audio_stream = audio_file.streams.audio[0]
-        except IndexError:
-            # Reset logging level
-            av.logging.restore_default_callback()
-            return {"has_audio": False}
-
-        audio_info["audio.channels"] = audio_stream.channels
-        audio_info["audio.codec"] = audio_stream.codec.canonical_name
-        # In an ideal loseless case : bit depth x sample rate x channels = bit rate.
-        # In an actual compressed case, the bit rate is set according to the compression level : the lower the bit rate, the more compression is applied.
-        audio_info["audio.bit_rate"] = audio_stream.bit_rate
-        audio_info["audio.sample_rate"] = audio_stream.sample_rate  # Number of samples per second
-        # In an ideal loseless case : fixed number of bits per sample.
-        # In an actual compressed case : variable number of bits per sample (often reduced to match a given depth rate).
-        audio_info["audio.bit_depth"] = audio_stream.format.bits
-        audio_info["audio.channel_layout"] = audio_stream.layout.name
-        audio_info["has_audio"] = True
-
-    # Reset logging level
-    av.logging.restore_default_callback()
-
-    return audio_info
-
-
 def get_video_info(video_path: Path | str) -> dict:
     # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.WARNING)
+    logging.getLogger("libav").setLevel(av.logging.ERROR)
 
     # Getting video stream information
     video_info = {}
@@ -1002,9 +553,6 @@ def get_video_info(video_path: Path | str) -> dict:
     # Reset logging level
     av.logging.restore_default_callback()
 
-    # Adding audio stream information
-    video_info.update(**get_audio_info(video_path))
-
     return video_info
 
 
@@ -1019,22 +567,22 @@ def get_video_pixel_channels(pix_fmt: str) -> int:
         raise ValueError("Unknown format")
 
 
-def get_video_duration_in_s(video_path: Path | str) -> float:
+def get_media_duration_in_s(media_path: Path | str, media_type: str = "video") -> float:
     """
-    Get the duration of a video file in seconds using PyAV.
+    Get the duration of a media file (video & audio) in seconds using PyAV.
 
     Args:
-        video_path: Path to the video file.
+        media_path: Path to the media file.
 
     Returns:
-        Duration of the video in seconds.
+        Duration of the media file in seconds.
     """
-    with av.open(str(video_path)) as container:
-        # Get the first video stream
-        video_stream = container.streams.video[0]
+    with av.open(str(media_path)) as container:
+        # Get the first stream
+        stream = container.streams.video[0] if media_type == "video" else container.streams.audio[0]
         # Calculate duration: stream.duration * stream.time_base gives duration in seconds
-        if video_stream.duration is not None:
-            duration = float(video_stream.duration * video_stream.time_base)
+        if stream.duration is not None:
+            duration = float(stream.duration * stream.time_base)
         else:
             # Fallback to container duration if stream duration is not available
             duration = float(container.duration / av.time_base)
@@ -1043,12 +591,12 @@ def get_video_duration_in_s(video_path: Path | str) -> float:
 
 class VideoEncodingManager:
     """
-    Context manager that ensures proper video encoding and data cleanup even if exceptions occur.
+    Context manager that ensures proper video and audio encoding and data cleanup even if exceptions occur.
 
     This manager handles:
     - Batch encoding for any remaining episodes when recording interrupted
-    - Cleaning up temporary image files from interrupted episodes
-    - Removing empty image directories
+    - Cleaning up temporary image and audio files from interrupted episodes
+    - Removing empty image and audio directories
 
     Args:
         dataset: The LeRobotDataset instance
@@ -1061,15 +609,8 @@ class VideoEncodingManager:
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        streaming_encoder = getattr(self.dataset, "_streaming_encoder", None)
-
-        if streaming_encoder is not None:
-            # Handle streaming encoder cleanup
-            if exc_type is not None:
-                streaming_encoder.cancel_episode()
-            streaming_encoder.close()
-        elif self.dataset.episodes_since_last_encoding > 0:
-            # Handle any remaining episodes that haven't been batch encoded
+        # Handle any remaining episodes that haven't been batch encoded
+        if self.dataset.episodes_since_last_encoding > 0:
             if exc_type is not None:
                 logging.info("Exception occurred. Encoding remaining episodes before exit...")
             else:
@@ -1082,12 +623,13 @@ class VideoEncodingManager:
                 f"from episode {start_ep} to {end_ep - 1}"
             )
             self.dataset._batch_save_episode_video(start_ep, end_ep)
+            self.dataset._batch_save_episode_audio(start_ep, end_ep)
 
         # Finalize the dataset to properly close all writers
         self.dataset.finalize()
 
-        # Clean up episode images if recording was interrupted (only for non-streaming mode)
-        if exc_type is not None and streaming_encoder is None:
+        # Clean up episode images if recording was interrupted
+        if exc_type is not None:
             interrupted_episode_index = self.dataset.num_episodes
             for key in self.dataset.meta.video_keys:
                 img_dir = self.dataset._get_image_file_path(
@@ -1098,15 +640,38 @@ class VideoEncodingManager:
                         f"Cleaning up interrupted episode images for episode {interrupted_episode_index}, camera {key}"
                     )
                     shutil.rmtree(img_dir)
+            for key in self.dataset.meta.audio_keys:
+                audio_file = self.dataset._get_raw_audio_file_path(
+                    episode_index=interrupted_episode_index, audio_key=key
+                )
+                if audio_file.exists():
+                    logging.debug(
+                        f"Cleaning up interrupted episode audio for episode {interrupted_episode_index}, microphone {key}"
+                    )
+                    audio_file.unlink()
 
         # Clean up any remaining images directory if it's empty
         img_dir = self.dataset.root / "images"
-        if img_dir.exists():
-            png_files = list(img_dir.rglob("*.png"))
-            if len(png_files) == 0:
+        # Check for any remaining PNG files
+        png_files = list(img_dir.rglob("*.png"))
+        if len(png_files) == 0:
+            # Only remove the images directory if no PNG files remain
+            if img_dir.exists():
                 shutil.rmtree(img_dir)
                 logging.debug("Cleaned up empty images directory")
-            else:
-                logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
+        else:
+            logging.debug(f"Images directory is not empty, containing {len(png_files)} PNG files")
+
+        # Clean up any remaining audio directory if it's empty
+        audio_dir = self.dataset.root / "raw_audio"
+        # Check for any remaining WAV files
+        wav_files = list(audio_dir.rglob("*.wav"))
+        if len(wav_files) == 0:
+            # Only remove the raw_audio directory if no WAV files remain
+            if audio_dir.exists():
+                shutil.rmtree(audio_dir)
+                logging.debug("Cleaned up empty audio directory")
+        else:
+            logging.debug(f"Audio directory is not empty, containing {len(wav_files)} WAV files")
 
         return False  # Don't suppress the original exception

src/lerobot/envs/configs.py

@@ -205,7 +205,6 @@ class ObservationConfig:
 
     add_joint_velocity_to_observation: bool = False
     add_current_to_observation: bool = False
-    add_ee_pose_to_observation: bool = False
     display_cameras: bool = False
 
 
@@ -261,7 +260,6 @@ class HILSerlRobotEnvConfig(EnvConfig):
 @dataclass
 class LiberoEnv(EnvConfig):
     task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
-    task_ids: list[int] | None = None
     fps: int = 30
     episode_length: int | None = None
     obs_type: str = "pixels_agent_pos"
@@ -340,10 +338,10 @@ class LiberoEnv(EnvConfig):
 
     @property
     def gym_kwargs(self) -> dict:
-        kwargs: dict[str, Any] = {"obs_type": self.obs_type, "render_mode": self.render_mode}
-        if self.task_ids is not None:
-            kwargs["task_ids"] = self.task_ids
-        return kwargs
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+        }
 
 
 @EnvConfig.register_subclass("metaworld")

src/lerobot/envs/libero.py

@@ -112,7 +112,6 @@ class LiberoEnv(gym.Env):
         visualization_height: int = 480,
         init_states: bool = True,
         episode_index: int = 0,
-        n_envs: int = 1,
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
         control_mode: str = "relative",
@@ -146,9 +145,7 @@ class LiberoEnv(gym.Env):
         self.episode_length = episode_length
         # Load once and keep
         self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
-        self._reset_stride = n_envs  # when performing a reset, append `_reset_stride` to `init_state_id`.
-
-        self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state
+        self._init_state_id = self.episode_index  # tie each sub-env to a fixed init state
 
         self._env = self._make_envs_task(task_suite, self.task_id)
         default_steps = 500
@@ -298,8 +295,7 @@ class LiberoEnv(gym.Env):
         self._env.seed(seed)
         raw_obs = self._env.reset()
         if self.init_states and self._init_states is not None:
-            raw_obs = self._env.set_init_state(self._init_states[self.init_state_id % len(self._init_states)])
-            self.init_state_id += self._reset_stride  # Change init_state_id when reset
+            raw_obs = self._env.set_init_state(self._init_states[self._init_state_id])
 
         # After reset, objects may be unstable (slightly floating, intersecting, etc.).
         # Step the simulator with a no-op action for a few frames so everything settles.
@@ -377,7 +373,6 @@ def _make_env_fns(
             init_states=init_states,
             episode_length=episode_length,
             episode_index=episode_index,
-            n_envs=n_envs,
             control_mode=control_mode,
             **local_kwargs,
         )

src/lerobot/microphones/__init__.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .config_openarm_mini import OpenArmMiniConfig
-from .openarm_mini import OpenArmMini
-
-__all__ = ["OpenArmMini", "OpenArmMiniConfig"]
+from .configs import MicrophoneConfig
+from .microphone import Microphone
+from .utils import make_microphones_from_configs

src/lerobot/microphones/anyskin/__init__.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,5 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .damiao import DamiaoMotorsBus
-from .tables import *
+from .configuration_anyskin import AnyskinSensorConfig
+from .sensor_anyskin import AnyskinSensor

src/lerobot/microphones/anyskin/configuration_anyskin.py

@@ -0,0 +1,45 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from ..configs import MicrophoneConfig
+
+
+@MicrophoneConfig.register_subclass("anyskin")
+@dataclass
+class AnyskinSensorConfig(MicrophoneConfig):
+    """Configuration class for Anyskin tactile sensors (technically not a microphone, but behaves like one acquisition-wise).
+
+    This class provides configuration options for Anyskin tactile sensors, including serial port, sample rate and channels.
+
+    Example configurations:
+    ```python
+    # Basic configurations
+    AnyskinSensorConfig("/dev/ttyACM0", 16000)  # Serial port /dev/ttyACM0, 16000Hz
+    AnyskinSensorConfig("/dev/ttyACM1", 44100)  # Serial port /dev/ttyACM1, 44100Hz
+    ```
+
+    Attributes:
+        sensor_port: Serial port of the tactile sensor.
+        baud_rate: Baud rate of the tactile sensor.
+        sample_rate: Sample rate in Hz for the tactile sensor.
+        channels: List of channel numbers to use for the tactile sensor.
+    """
+
+    sensor_port: str
+    baud_rate: int = 115_200
+    sensor_id: int = 0
+    burst_mode: bool = True
+    temp_filtered: bool = False

src/lerobot/microphones/anyskin/sensor_anyskin.py

@@ -0,0 +1,473 @@
+# 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.
+
+"""
+Provides the AnyskinSensor class for capturing tactile data from Anyskin tactile sensors.
+"""
+
+from doctest import master
+import logging
+import time
+from multiprocessing import (
+    Event as process_Event,
+    JoinableQueue as process_Queue,
+    Process,
+)
+from pathlib import Path
+from queue import Empty
+from threading import Barrier, Event, Event as thread_Event, Thread
+from typing import Any
+
+from lerobot.utils.hub import T
+import numpy as np
+from serial import Serial, serialutil
+from soundfile import SoundFile
+
+from lerobot.utils.errors import (
+    DeviceAlreadyConnectedError,
+    DeviceAlreadyRecordingError,
+    DeviceNotConnectedError,
+    DeviceNotRecordingError,
+)
+from lerobot.utils.shared_array import SharedArray
+
+from ..microphone import Microphone
+from .configuration_anyskin import AnyskinSensorConfig
+
+from anyskin import AnySkinBase, AnySkinDummy 
+
+logger = logging.getLogger(__name__)
+
+MAX_MAGNETS_CHANNELS = 5
+
+class AnyskinSensor(Microphone):
+    """
+    The AnyskinSensor class handles all Anyskin tactile sensors.
+
+    A AnyskinSensor instance requires the serial port of the tactile sensor, which may be obtained using `python -m lerobot.find_port`. It also requires the recording sample rate as well as the list of recorded channels.
+
+    Example of usage:
+    ```python
+    from lerobot.common.robot_devices.microphones.configs import AnyskinSensorConfig
+
+    config = AnyskinSensorConfig(sensor_port="/dev/ttyACM0", baud_rate=115200, sample_rate=115, channels=[1])
+    microphone = AnyskinSensor(config)
+
+    microphone.connect()
+    microphone.start_recording("some/output/file.wav")
+    ...
+    audio_readings = microphone.read()  # Gets all recorded audio data since the last read or since the beginning of the recording. The longer the period the longer the reading time !
+    ...
+    microphone.stop_recording()
+    microphone.disconnect()
+    ```
+    """
+
+    def __init__(self, config: AnyskinSensorConfig):
+        """ "
+        Initializes the AnyskinSensor instance.
+
+        Args:
+            config: The configuration settings for the sensor.
+        """
+        super().__init__(config)
+
+        # Sensor port
+        self.sensor_port = config.sensor_port
+
+        # Baud rate
+        self.baud_rate = config.baud_rate
+
+        # Input audio recording process and events
+        self.record_process = None
+        self.record_stop_event = process_Event()
+        self.record_start_event = process_Event()
+        self.record_close_event = process_Event()
+        self.record_is_started_event = process_Event()
+        self.audio_callback_start_event = process_Event()
+
+        # Process-safe concurrent queue to send audio from the recording process to the writing process/thread
+        self.write_queue = process_Queue()
+
+        # SharedArray to store audio from the recording process.
+        self.read_shared_array = None
+        self.local_read_shared_array = None
+        # Thread/Process to handle data writing in a separate thread/process (safely)
+        self.write_thread = None
+        self.write_stop_event = None
+        self.write_is_started_event = None
+
+        self.logs = {}
+
+    def __str__(self) -> str:
+        return f"{self.__class__.__name__}({self.sensor_port})"
+
+    @property
+    def is_connected(self) -> bool:
+        """Check if the sensor is currently connected.
+
+        Returns:
+            bool: True if the sensor is connected and ready to start recording,
+                  False otherwise.
+        """
+        return self.record_process is not None and self.record_process.is_alive()
+
+    @property
+    def is_recording(self) -> bool:
+        """Check if the sensor is currently recording.
+
+        Returns:
+            bool: True if the sensor is recording, False otherwise.
+        """
+        return self.record_is_started_event.is_set()
+
+    @property
+    def is_writing(self) -> bool:
+        """Check if the sensor is currently writing to a file.
+
+        Returns:
+            bool: True if the sensor is writing to a file, False otherwise.
+        """
+        return self.write_thread is not None and self.write_is_started_event.is_set()
+
+    @staticmethod
+    def find_microphones() -> list[dict[str, Any]]:
+        """Detects available sensors connected to the system.
+
+        Returns:
+            List[Dict[str, Any]]: A list of dictionaries,
+            where each dictionary contains information about a detected sensor.
+        """
+        pass
+
+    def connect(self) -> None:
+        """
+        Establish connection to the sensor.
+        """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"Sensor connected to {self.sensor_port} is already connected.")
+
+        # Create or reset queue and shared array
+        self.read_shared_array = SharedArray(
+            shape=(self.sample_rate * 10, len(self.channels)),
+            dtype=np.dtype("int16"),
+        )
+        self.local_read_shared_array = self.read_shared_array.get_local_array()
+        self.write_queue = process_Queue()
+
+        # Reset events
+        self.record_start_event.clear()
+        self.record_stop_event.clear()
+        self.record_close_event.clear()
+        self.record_is_started_event.clear()
+        self.audio_callback_start_event.clear()
+
+        # Create and start an audio input stream with a recording callback
+        # Remark: this is done in a separate process so that audio recording is not impacted by the main thread CPU usage, especially the busy_wait function.
+        process_init_event = process_Event()
+        self.record_process = Process(
+            target=self._record_process,
+            args=(
+                self.sensor_port,
+                self.baud_rate,
+                self.channels,
+                process_init_event,
+                self.record_start_event,
+                self.record_stop_event,
+                self.record_close_event,
+                self.record_is_started_event,
+                self.audio_callback_start_event,
+                self.write_queue,
+                self.read_shared_array,
+            ),
+        )
+        self.record_process.daemon = True
+        self.record_process.start()
+
+        is_init = process_init_event.wait(
+            timeout=5.0
+        )  # Wait for the recording process to be started, and to potentially raise an error on failure.
+        if not self.is_connected or not is_init:
+            raise RuntimeError(f"Error connecting sensor connected to {self.sensor_port}.")
+
+        logger.info(f"{self} connected.")
+
+    @staticmethod
+    def _record_process(
+        sensor_port,
+        baud_rate,
+        channels,
+        process_init_event,
+        record_start_event,
+        record_stop_event,
+        record_close_event,
+        record_is_started_event,
+        audio_callback_start_event,
+        write_queue,
+        read_shared_array,
+    ) -> None:
+        channels_index = np.array(channels) - 1
+        local_read_shared_array = read_shared_array.get_local_array()
+
+        def tactile_callback(tactile_sensor: AnySkinBase):
+            """
+            Parse the tactile data from the raw input data.
+            """
+            if audio_callback_start_event.is_set():
+                timestamp, indata = tactile_sensor.get_sample()
+                indata = indata.reshape(-1, MAX_MAGNETS_CHANNELS)
+                write_queue.put_nowait(indata[:, channels_index])
+                read_shared_array.write(local_read_shared_array, indata[:, channels_index])
+
+        try:
+            tactile_sensor = AnySkinBase(
+                num_mags=MAX_MAGNETS_CHANNELS,
+                port=sensor_port,
+                baudrate=baud_rate,
+                burst_mode=True,
+                device_id=0,    #TODO(CarolinePascal): create an abstract increasing id for each sensor
+                temp_filtered=False,
+            )   #TODO(CarolinePascal): add timeout on serial connection ?
+        except (serialutil.SerialException, AttributeError) as e:
+            raise RuntimeError(f"Error connecting sensor connected to {sensor_port}: {e}")
+
+        process_init_event.set()
+
+        while True:
+            start_flag = record_start_event.wait(timeout=0.1)
+            if record_close_event.is_set():
+                break
+            elif not start_flag:
+                continue
+            record_is_started_event.set()
+            while not record_stop_event.is_set():
+                tactile_callback(tactile_sensor)    # Initial flush is already done in the constructor.
+            record_is_started_event.clear()
+        tactile_sensor.close()  # Closes the inherited serial connection.
+
+    def disconnect(self) -> None:
+        """
+        Disconnect the sensor and release any resources.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Sensor connected to {self.sensor_port} is not connected.")
+
+        if self.is_recording:
+            self.stop_recording()
+
+        self.record_close_event.set()
+        self.read_shared_array.delete()
+        self.write_queue.close()
+        self.record_process.join()
+
+        if self.is_connected:
+            raise RuntimeError(f"Error disconnecting sensor connected to {self.sensor_port}.")
+
+        logger.info(f"{self} disconnected.")
+
+    def start_recording(
+        self,
+        output_file: str | Path | None = None,
+        multiprocessing: bool | None = False,
+        overwrite: bool | None = True,
+        barrier: Barrier | None = None,
+    ) -> None:
+        """
+        Start recording tactile data from the sensor.
+
+        Args:
+            output_file: Optional path to save the recorded tactile data.
+            multiprocessing: If True, enables multiprocessing for recording. Defaults to multithreading otherwise.
+            overwrite: If True, overwrites existing files at output_file path.
+            barrier: If not None, ensures that multiple sensors start recording at the same time.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Sensor connected to {self.sensor_port} is not connected.")
+        if self.is_recording:
+            raise DeviceAlreadyRecordingError(f"Sensor connected to {self.sensor_port} is already recording.")
+
+        # Reset queue and shared memory
+        self.read_shared_array.reset()
+        self._clear_queue(self.write_queue)
+
+        # Reset stop event
+        self.record_stop_event.clear()
+
+        # Write recordings into a file if output_file is provided
+        if output_file is not None:
+            output_file = Path(output_file)
+            output_file.parent.mkdir(parents=True, exist_ok=True)
+
+            if output_file.exists():
+                if overwrite:
+                    output_file.unlink()
+                else:
+                    raise FileExistsError(
+                        f"Output file {output_file} already exists. Set overwrite to True to overwrite it."
+                    )
+
+            if multiprocessing:
+                self.write_stop_event = process_Event()
+                self.write_is_started_event = process_Event()
+                self.write_thread = Process(
+                    target=AnyskinSensor._write_loop,
+                    args=(
+                        self.write_queue,
+                        self.write_stop_event,
+                        self.write_is_started_event,
+                        self.sample_rate,
+                        self.channels,
+                        output_file,
+                    ),
+                )
+            else:
+                self.write_stop_event = thread_Event()
+                self.write_is_started_event = thread_Event()
+                self.write_thread = Thread(
+                    target=AnyskinSensor._write_loop,
+                    args=(
+                        self.write_queue,
+                        self.write_stop_event,
+                        self.write_is_started_event,
+                        self.sample_rate,
+                        self.channels,
+                        output_file,
+                    ),
+                )
+            self.write_thread.daemon = True
+            self.write_thread.start()
+            self.write_is_started_event.wait()  # Wait for the writing thread/process to be started.
+
+        self.record_start_event.set()  # Start the input audio stream process
+        self.record_is_started_event.wait()  # Wait for the input audio stream process to be actually started
+
+        if barrier is not None:
+            barrier.wait()  # Wait for multiple input audio streams to be started at the same time
+
+        self.audio_callback_start_event.set()
+
+        if not self.is_recording:
+            raise RuntimeError(f"Error starting recording for sensor connected to {self.sensor_port}.")
+        if output_file is not None and not self.is_writing:
+            raise RuntimeError(f"Error starting writing for sensor connected to {self.sensor_port}.")
+
+    def _read(self) -> np.ndarray:
+        """
+        Thread/Process-safe callback to read available audio data
+        """
+        return self.read_shared_array.read(self.local_read_shared_array, flush=True)
+
+    def read(self) -> np.ndarray:
+        """Capture and return a single audio chunk from the sensor.
+
+        Returns:
+            np.ndarray: Captured audio chunk as a numpy array.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Sensor connected to {self.sensor_port} is not connected.")
+        if not self.is_recording:
+            raise RuntimeError(f"Sensor connected to {self.sensor_port} is not recording.")
+
+        start_time = time.perf_counter()
+
+        tactile_readings = self._read()
+
+        # log the number of seconds it took to read the audio chunk
+        self.logs["delta_timestamp_s"] = time.perf_counter() - start_time
+
+        # log the utc time at which the audio chunk was received
+        self.logs["timestamp_utc"] = time.perf_counter()
+
+        return tactile_readings
+
+    def _read_loop(self) -> None:
+        """Internal loop run by the background thread for asynchronous reading."""
+
+    def stop_recording(self) -> None:
+        """Stop recording audio from the sensor."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Sensor connected to {self.sensor_port} is not connected.")
+        if not self.is_recording:
+            raise DeviceNotRecordingError(f"Sensor connected to {self.sensor_port} is not recording.")
+
+        self.audio_callback_start_event.clear()
+        self.record_start_event.clear()  # Ensures the audio stream is not started again !
+        self.record_stop_event.set()
+
+        self.read_shared_array.reset()
+        self._clear_queue(self.write_queue, join_queue=True)
+
+        if self.is_writing:
+            self.write_stop_event.set()
+            self.write_thread.join()
+
+        timeout = 1.0
+        while self.is_recording and timeout > 0:
+            time.sleep(0.01)
+            timeout -= 0.01
+
+        if self.is_recording:
+            raise RuntimeError(f"Error stopping recording for sensor connected to {self.sensor_port}.")
+        if self.is_writing:
+            raise RuntimeError(f"Error stopping writing for sensor connected to {self.sensor_port}.")
+
+    def __del__(self) -> None:
+        if self.is_connected:
+            self.disconnect()
+
+    @staticmethod
+    def _clear_queue(queue, join_queue: bool = False):
+        """
+        Clears the queue by getting all items until it is empty. The longer the queue, the longer it takes to clear it.
+        """
+        try:
+            while True:
+                queue.get_nowait()
+                queue.task_done()
+        except Empty:
+            if join_queue:
+                queue.join()
+            return
+
+    @staticmethod
+    def _write_loop(
+        queue,
+        write_stop_event: Event,
+        write_is_started_event: Event,
+        sample_rate: int,
+        channels: list[int],
+        output_file: Path,
+    ) -> None:
+        """
+        Thread/Process-safe loop to write audio data into a file.
+        """
+        # Can only be run on a single process/thread for file writing safety
+        with SoundFile(
+            output_file,
+            mode="w",
+            samplerate=sample_rate,
+            channels=len(channels),
+            format="WAV",
+            subtype="FLOAT",  # Subtype for float32 values
+        ) as file:
+            write_is_started_event.set()
+            while not write_stop_event.is_set():
+                try:
+                    file.write(
+                        queue.get(timeout=0.005)
+                    )  # Timeout set as the usual sounddevice buffer size. get_nowait is not possible here as it saturates the thread.
+                    queue.task_done()
+                except Empty:
+                    continue
+        write_is_started_event.clear()

src/lerobot/microphones/configs.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,7 +12,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .config_openarm_follower import OpenArmFollowerConfig, OpenArmFollowerConfigBase
-from .openarm_follower import OpenArmFollower
+import abc
+from dataclasses import dataclass
 
-__all__ = ["OpenArmFollower", "OpenArmFollowerConfig", "OpenArmFollowerConfigBase"]
+import draccus
+
+
+@dataclass(kw_only=True)
+class MicrophoneConfig(draccus.ChoiceRegistry, abc.ABC):
+    sample_rate: int | None = None
+    channels: list[int] | None = None
+
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)

src/lerobot/microphones/microphone.py

@@ -0,0 +1,140 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import abc
+from pathlib import Path
+from threading import Barrier
+from typing import Any
+
+import numpy as np
+
+from .configs import MicrophoneConfig
+
+
+class Microphone(abc.ABC):
+    """Base class for microphone implementations.
+
+    Defines a standard interface for microphone operations across different backends.
+    Subclasses must implement all abstract methods.
+
+    Manages basic microphone properties (sample rate, channels) and core operations:
+    - Connection/disconnection
+    - Start/stop recording
+    - Audio chunk reading
+
+    Attributes:
+        sample_rate (int | None): Configured sample rate in Hz
+        channels (list[int] | None): List of channel numbers to record
+
+    Example:
+        class MyMicrophone(Microphone):
+            def __init__(self, config): ...
+            @property
+            def is_connected(self) -> bool: ...
+            def connect(self): ...
+            # Plus other required methods
+    """
+
+    def __init__(self, config: MicrophoneConfig):
+        """Initialize the microphone with the given configuration.
+
+        Args:
+            config: Microphone configuration containing sample rate and channels.
+        """
+        self.sample_rate: int | None = config.sample_rate
+        self.channels: list[int] | None = config.channels
+
+    @property
+    @abc.abstractmethod
+    def is_connected(self) -> bool:
+        """Check if the microphone is currently connected.
+
+        Returns:
+            bool: True if the microphone is connected and ready to start recording,
+                  False otherwise.
+        """
+        pass
+
+    @property
+    @abc.abstractmethod
+    def is_recording(self) -> bool:
+        """Check if the microphone is currently recording.
+
+        Returns:
+            bool: True if the microphone is recording, False otherwise.
+        """
+        pass
+
+    @property
+    @abc.abstractmethod
+    def is_writing(self) -> bool:
+        """Check if the microphone is currently writing to a file.
+
+        Returns:
+            bool: True if the microphone is writing to a file, False otherwise.
+        """
+        pass
+
+    @staticmethod
+    @abc.abstractmethod
+    def find_microphones() -> list[dict[str, Any]]:
+        """Detects available microphones connected to the system.
+
+        Returns:
+            List[Dict[str, Any]]: A list of dictionaries,
+            where each dictionary contains information about a detected microphone.
+        """
+        pass
+
+    @abc.abstractmethod
+    def connect(self) -> None:
+        """Establish connection to the microphone."""
+        pass
+
+    @abc.abstractmethod
+    def start_recording(
+        self,
+        output_file: str | Path | None = None,
+        multiprocessing: bool | None = False,
+        overwrite: bool | None = True,
+        barrier: Barrier | None = None,
+    ) -> None:
+        """Start recording audio from the microphone.
+
+        Args:
+            output_file: Optional path to save the recorded audio.
+            multiprocessing: If True, enables multiprocessing for recording. Defaults to multithreading otherwise.
+            overwrite: If True, overwrites existing files at output_file path.
+            barrier: If not None, ensures that multiple microphones start recording at the same time.
+        """
+        pass
+
+    @abc.abstractmethod
+    def read(self) -> np.ndarray:
+        """Capture and return a single audio chunk from the microphone.
+
+        Returns:
+            np.ndarray: Captured audio chunk as a numpy array.
+        """
+        pass
+
+    @abc.abstractmethod
+    def stop_recording(self) -> None:
+        """Stop recording audio from the microphone."""
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self) -> None:
+        """Disconnect the microphone and release any resources."""
+        pass

src/lerobot/microphones/portaudio/__init__.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,7 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .config_openarm_leader import OpenArmLeaderConfig, OpenArmLeaderConfigBase
-from .openarm_leader import OpenArmLeader
-
-__all__ = ["OpenArmLeader", "OpenArmLeaderConfig", "OpenArmLeaderConfigBase"]
+from .configuration_portaudio import PortAudioMicrophoneConfig
+from .microphone_portaudio import PortAudioMicrophone

src/lerobot/microphones/portaudio/configuration_portaudio.py

@@ -0,0 +1,41 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+
+from ..configs import MicrophoneConfig
+
+
+@MicrophoneConfig.register_subclass("portaudio")
+@dataclass
+class PortAudioMicrophoneConfig(MicrophoneConfig):
+    """Configuration class for PortAudio-based microphone devices.
+
+    This class provides configuration options for microphones accessed through PortAudio with the sounddevice Python package.
+    including device index, sample rate and channels.
+
+    Example configurations:
+    ```python
+    # Basic configurations
+    PortAudioMicrophoneConfig(0, 16000, [1])  # Device index 0, 16000Hz, mono
+    PortAudioMicrophoneConfig(1, 44100, [1, 2])  # Device index 1, 44100Hz, stereo
+    ```
+
+    Attributes:
+        microphone_index: Device index for the microphone.
+        sample_rate: Sample rate in Hz for the microphone.
+        channels: List of channel numbers to use for the microphone.
+    """
+
+    microphone_index: int

src/lerobot/microphones/portaudio/interface_sounddevice_sdk.py

@@ -0,0 +1,394 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import abc
+import time
+from collections.abc import Callable
+from threading import Event, Thread
+from typing import Any
+
+import numpy as np
+from sounddevice import PortAudioError
+
+from lerobot.utils.robot_utils import precise_sleep
+
+
+# --- Interface definitions for InputStream ---
+class IInputStream(abc.ABC):
+    @abc.abstractmethod
+    def __init__(
+        self,
+        samplerate: float | None = None,
+        blocksize: int | None = None,
+        device: int | str | None = None,
+        channels: int | None = None,
+        dtype: str | np.dtype | None = None,
+        latency: float | str | None = None,
+        callback: Callable[[Any, int, Any, Any], None] | None = None,
+    ):
+        pass
+
+    @abc.abstractmethod
+    def start(self) -> None:
+        pass
+
+    @abc.abstractmethod
+    def stop(self) -> None:
+        pass
+
+    @abc.abstractmethod
+    def close(self) -> None:
+        pass
+
+
+class ISounddeviceSDK(abc.ABC):
+    """Interface defining the contract for the Sounddevice SDK."""
+
+    InputStream: type[IInputStream]
+
+    @abc.abstractmethod
+    def query_devices(self, device: int | str | None = None, kind: str | None = None) -> list[dict[str, Any]]:
+        pass
+
+
+# --- Real SDK Adapter ---
+
+
+class SounddeviceSDKAdapter(ISounddeviceSDK):
+    """Adapts the real sounddevice library to the ISounddeviceSDK interface."""
+
+    _sounddevice = None
+
+    def __init__(self):
+        try:
+            import sounddevice
+
+            SounddeviceSDKAdapter._sounddevice = sounddevice
+        except ImportError as e:
+            raise ImportError("sounddevice library not found") from e
+
+    # --- Inner Class Implementation ---
+    class RealInputStream(IInputStream):
+        def __init__(
+            self,
+            samplerate: int | None = None,
+            blocksize: int | None = None,
+            device: int | None = None,
+            channels: int | None = None,
+            dtype: str | np.dtype | None = None,
+            latency: float | str | None = None,
+            callback: Callable[[Any, int, Any, Any], None] | None = None,
+        ):
+            import sounddevice
+
+            self._input_stream = sounddevice.InputStream(
+                samplerate=samplerate,
+                blocksize=blocksize,
+                device=device,
+                channels=channels,
+                dtype=dtype,
+                latency=latency,
+                callback=callback,
+            )
+
+        def start(self) -> None:
+            self._input_stream.start()
+
+        def stop(self) -> None:
+            self._input_stream.stop()
+
+        def close(self) -> None:
+            self._input_stream.close()
+
+        def __del__(self):
+            self._input_stream.stop()
+            self._input_stream.close()
+
+        @property
+        def active(self) -> bool:
+            return self._input_stream.active
+
+        @property
+        def stopped(self) -> bool:
+            return self._input_stream.stopped
+
+        @property
+        def closed(self) -> bool:
+            return self._input_stream.closed
+
+    InputStream = RealInputStream
+
+    def query_devices(self, device: int | str | None = None, kind: str | None = None) -> list[dict[str, Any]]:
+        return SounddeviceSDKAdapter._sounddevice.query_devices(device, kind)
+
+
+# Emulates a 48kHz stereo microphone
+VALID_DTYPE = {
+    "float32",
+    "int32",
+    "int16",
+    "int8",
+    "uint8",
+    np.float32,
+    np.int32,
+    np.int16,
+    np.int8,
+    np.uint8,
+}
+VALID_LATENCY = {"low", "high"}
+
+VALID_DEVICES = [
+    {
+        "index": 0,
+        "name": "Built-in Microphone",
+        "hostapi": 0,
+        "max_input_channels": 2,
+        "max_output_channels": 0,
+        "default_low_input_latency": 0.01,
+        "default_low_output_latency": 0.001,
+        "default_high_input_latency": 0.1,
+        "default_high_output_latency": 0.01,
+        "default_samplerate": 48000.0,
+    },
+    {
+        "index": 1,
+        "name": "Built-in Output",
+        "hostapi": 0,
+        "max_input_channels": 0,
+        "max_output_channels": 2,
+        "default_low_input_latency": 0.04,
+        "default_low_output_latency": 0.04,
+        "default_high_input_latency": 0.12,
+        "default_high_output_latency": 0.12,
+        "default_samplerate": 48000.0,
+    },
+    {
+        "index": 2,
+        "name": "USB Audio Device",
+        "hostapi": 0,
+        "max_input_channels": 1,
+        "max_output_channels": 0,
+        "default_low_input_latency": 0.03,
+        "default_low_output_latency": 0.01,
+        "default_high_input_latency": 0.04,
+        "default_high_output_latency": 0.03,
+        "default_samplerate": 16000.0,
+    },
+]
+
+# -- Fake SDK Adapter ---
+
+
+class FakeSounddeviceSDKAdapter(ISounddeviceSDK):
+    """Implements the ISounddeviceSDK interface with fake behaviour for testing."""
+
+    # --- Inner Class Implementation ---
+    class FakeInputStream(IInputStream):
+        def __init__(
+            self,
+            samplerate: float | None = None,
+            blocksize: int | None = None,
+            device: int | str | None = None,
+            channels: int | None = None,
+            dtype: str | None = None,
+            latency: str | None = None,
+            callback: Callable[[Any, int, Any, Any], None] | None = None,
+        ):
+            self.samplerate = samplerate
+            self.blocksize = blocksize
+            self.device = device
+            self.channels = channels
+            self.dtype = dtype
+            self.latency = latency
+            self.callback = callback
+
+            self._validate_settings()
+
+            self._active = False
+            self._closed = False
+
+            if self.callback is not None:
+                self._streaming_thread = Thread(target=self._streaming_loop, daemon=True)
+                self._streaming_thread_stop_event = Event()
+
+        @property
+        def active(self) -> bool:
+            """True when the stream is active, False otherwise."""
+            return self._active
+
+        @property
+        def stopped(self) -> bool:
+            """True when the stream is stopped, False otherwise."""
+            return not self._active
+
+        @property
+        def closed(self) -> bool:
+            """True after a call to close(), False otherwise."""
+            return self._closed
+
+        def _get_device_info(self):
+            """Returns the device info for the device."""
+            for device in VALID_DEVICES:
+                if (isinstance(self.device, int) and device["index"] == self.device) or (
+                    isinstance(self.device, str) and device["name"] == self.device
+                ):
+                    return device
+            raise PortAudioError(f"No input device matching {self.device}")
+
+        def _validate_device(self):
+            """Validates the device against the valid devices."""
+            valid_device_indices = [device["index"] for device in VALID_DEVICES]
+            valid_device_names = [device["name"] for device in VALID_DEVICES]
+
+            if self.device is not None:
+                if isinstance(self.device, (int, str)):
+                    # Check if device index is valid
+                    if isinstance(self.device, int) and self.device not in valid_device_indices:
+                        raise PortAudioError(f"Error querying device {self.device}")
+
+                    # Check if device name is valid
+                    if isinstance(self.device, str) and self.device not in valid_device_names:
+                        raise PortAudioError(f"No input device matching {self.device}")
+                else:
+                    raise PortAudioError(f"Device must be int or str, got {type(self.device)}")
+            else:
+                # Default to first input device
+                input_devices = [d for d in VALID_DEVICES if d["max_input_channels"] > 0]
+                if input_devices:
+                    self.device = input_devices[0]["index"]
+
+        def _validate_samplerate(self):
+            """Validates the samplerate against the device's maximum samplerate."""
+            device_info = self._get_device_info()
+            if self.samplerate is None:
+                self.samplerate = device_info["default_samplerate"]
+            elif self.samplerate > device_info["default_samplerate"] or self.samplerate < 1000:
+                raise PortAudioError("Error opening InputStream: Invalid sample rate")
+
+        def _validate_channels(self):
+            """Validates the channels against the device's maximum channels."""
+            device_info = self._get_device_info()
+            if self.channels is None:
+                self.channels = device_info["max_input_channels"]
+            elif self.channels > device_info["max_input_channels"] or self.channels < 1:
+                raise PortAudioError("Error opening InputStream: Invalid number of channels")
+
+        def _validate_dtype(self):
+            """Validates the dtype against the valid dtypes."""
+            if self.dtype is not None:
+                if self.dtype not in VALID_DTYPE:
+                    raise PortAudioError("Invalid input sample format")
+            else:
+                self.dtype = "float32"  # Default dtype
+
+        def _validate_latency(self):
+            """Validates the latency against the valid latencies."""
+            if self.latency is not None:
+                if self.latency not in VALID_LATENCY:
+                    raise PortAudioError("Invalid latency")
+            else:
+                self.latency = "low"  # Default latency
+
+            if isinstance(self.latency, str):
+                device_info = self._get_device_info()
+                if self.latency == "low":
+                    self.latency = device_info["default_low_input_latency"]
+                elif self.latency == "high":
+                    self.latency = device_info["default_high_input_latency"]
+
+        def _validate_settings(self):
+            """Validates the input parameters against available devices and valid options."""
+            self._validate_device()
+            self._validate_samplerate()
+            self._validate_channels()
+            self._validate_dtype()
+            self._validate_latency()
+
+        def _simulated_audio_data(self) -> np.ndarray:
+            """Generates a simulated audio signal for testing purposes with proper value ranges."""
+            duration_samples = int(self.samplerate * self.latency)
+
+            # Generate output according to dtype
+            if self.dtype in {"float32", np.float32}:
+                # Generate values between -1 and 1 for float32
+                data = np.random.uniform(-1.0, 1.0, (duration_samples, self.channels)).astype(self.dtype)
+            else:
+                # Use np.iinfo to get proper range for integer types
+                info = np.iinfo(self.dtype)
+                data = np.random.randint(
+                    info.min, info.max + 1, (duration_samples, self.channels), dtype=self.dtype
+                )
+
+            return data
+
+        def _streaming_loop(self):
+            if self.callback is not None:
+                while not self._streaming_thread_stop_event.is_set():
+                    precise_sleep(self.latency)
+                    tmp_data = self._simulated_audio_data()
+                    self.callback(
+                        tmp_data,
+                        len(tmp_data),
+                        time.perf_counter(),
+                        None,
+                    )
+
+        def start(self) -> None:
+            """Start the fake input stream."""
+            if not self.active and self.callback is not None:
+                self._streaming_thread.start()
+            self._active = True
+
+        def stop(self) -> None:
+            """Stop the fake input stream."""
+            if self.callback is not None:
+                self._streaming_thread_stop_event.set()
+                self._streaming_thread.join()
+            self._active = False
+
+        def close(self) -> None:
+            """Close the fake input stream."""
+            if self.active and self.callback is not None:
+                self.stop()
+            self._active = False
+            self._closed = True
+
+        def __del__(self):
+            self.close()
+
+    InputStream = FakeInputStream
+
+    def query_devices(self, device: int | str | None = None, kind: str | None = None) -> list[dict[str, Any]]:
+        """Returns a realistic list of audio devices including speakers and microphones."""
+        if device is not None:
+            # Return specific device
+            for valid_device in VALID_DEVICES:
+                if (isinstance(device, int) and valid_device["index"] == device) or (
+                    isinstance(device, str) and valid_device["name"] == device
+                ):
+                    return valid_device
+            raise PortAudioError(f"Error querying device {device}")
+
+        elif kind is not None:
+            for valid_device in VALID_DEVICES:
+                if (
+                    valid_device["max_input_channels"] > 0
+                    and kind == "input"
+                    or valid_device["max_output_channels"] > 0
+                    and kind == "output"
+                ):
+                    return valid_device
+            raise PortAudioError(f"No {kind} device found")
+
+        return VALID_DEVICES

src/lerobot/microphones/portaudio/microphone_portaudio.py

@@ -0,0 +1,566 @@
+# 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.
+
+"""
+Provides the PortAudioMicrophone class for capturing audio from microphones using the PortAudio library through the sounddevice Python package.
+"""
+
+import logging
+import time
+from multiprocessing import (
+    Event as process_Event,
+    JoinableQueue as process_Queue,
+    Process,
+)
+from pathlib import Path
+from queue import Empty
+from threading import Barrier, Event, Event as thread_Event, Thread
+from typing import Any
+
+import numpy as np
+from soundfile import SoundFile
+
+from lerobot.microphones.portaudio.interface_sounddevice_sdk import ISounddeviceSDK, SounddeviceSDKAdapter
+from lerobot.utils.errors import (
+    DeviceAlreadyConnectedError,
+    DeviceAlreadyRecordingError,
+    DeviceNotConnectedError,
+    DeviceNotRecordingError,
+)
+from lerobot.utils.shared_array import SharedArray
+
+from ..microphone import Microphone
+from .configuration_portaudio import PortAudioMicrophoneConfig
+
+logger = logging.getLogger(__name__)
+
+
+class PortAudioMicrophone(Microphone):
+    """
+    The PortAudioMicrophone class handles all microphones compatible with sounddevice (and the underlying PortAudio library). Most microphones and sound cards are compatible, across all OS (Linux, Mac, Windows).
+
+    A PortAudioMicrophone instance requires the sounddevice index of the microphone, which may be obtained using `python -m sounddevice`. It also requires the recording sample rate as well as the list of recorded channels.
+
+    Example of usage:
+    ```python
+    from lerobot.common.robot_devices.microphones.configs import PortAudioMicrophoneConfig
+
+    config = PortAudioMicrophoneConfig(microphone_index=0, sample_rate=16000, channels=[1])
+    microphone = PortAudioMicrophone(config)
+
+    microphone.connect()
+    microphone.start_recording("some/output/file.wav")
+    ...
+    audio_readings = microphone.read()  # Gets all recorded audio data since the last read or since the beginning of the recording. The longer the period the longer the reading time !
+    ...
+    microphone.stop_recording()
+    microphone.disconnect()
+    ```
+    """
+
+    def __init__(self, config: PortAudioMicrophoneConfig, sounddevice_sdk: ISounddeviceSDK = None):
+        """
+        Initializes the PortAudioMicrophone instance.
+
+        Args:
+            config: The configuration settings for the microphone.
+        """
+        super().__init__(config)
+
+        if sounddevice_sdk is None:
+            self.sounddevice_sdk = SounddeviceSDKAdapter()
+        else:
+            self.sounddevice_sdk = sounddevice_sdk
+
+        # Microphone index
+        self.microphone_index = config.microphone_index
+
+        # Input audio recording process and events
+        self.record_process = None
+        self.record_stop_event = process_Event()
+        self.record_start_event = process_Event()
+        self.record_close_event = process_Event()
+        self.record_is_started_event = process_Event()
+        self.audio_callback_start_event = process_Event()
+
+        # Process-safe concurrent queue to send audio from the recording process to the writing process/thread
+        self.write_queue = process_Queue()
+
+        # SharedArray to store audio from the recording process.
+        self.read_shared_array = None
+        self.local_read_shared_array = None
+        # Thread/Process to handle data writing in a separate thread/process (safely)
+        self.write_thread = None
+        self.write_stop_event = None
+        self.write_is_started_event = None
+
+        self.logs = {}
+
+    def __str__(self) -> str:
+        return f"{self.__class__.__name__}({self.microphone_index})"
+
+    @property
+    def is_connected(self) -> bool:
+        return self.record_process is not None and self.record_process.is_alive()
+
+    @property
+    def is_recording(self) -> bool:
+        return self.record_is_started_event.is_set()
+
+    @property
+    def is_writing(self) -> bool:
+        return self.write_thread is not None and self.write_is_started_event.is_set()
+
+    @staticmethod
+    def find_microphones(
+        device: int | str | None = None, sounddevice_sdk: ISounddeviceSDK = None
+    ) -> list[dict[str, Any]] | dict[str, Any]:
+        """
+        Detects available microphones connected to the system.
+
+        Args:
+            device: The device to find microphones for. If None, all microphones are found.
+
+        Returns:
+            List[Dict[str, Any]]: A list of dictionaries,
+            where each dictionary contains information about a detected microphone : index, name, sample rate, channels.
+        """
+
+        if sounddevice_sdk is None:
+            sounddevice_sdk = SounddeviceSDKAdapter()
+
+        found_microphones_info = []
+
+        devices = sounddevice_sdk.query_devices()
+        for d in devices:
+            if d["max_input_channels"] > 0:
+                microphone_info = {
+                    "index": d["index"],
+                    "name": d["name"],
+                    "sample_rate": int(d["default_samplerate"]),
+                    "channels": np.arange(1, d["max_input_channels"] + 1),
+                }
+
+                if device is None or (
+                    (isinstance(device, int) and d["index"] == device)
+                    or (isinstance(device, str) and d["name"] == device)
+                ):
+                    found_microphones_info.append(microphone_info)
+
+        if device is not None:
+            if len(found_microphones_info) == 0:
+                raise RuntimeError(f"No microphone found for device {device}")
+            else:
+                return found_microphones_info[0]
+
+        if len(found_microphones_info) == 0:
+            logger.warning("No microphone found !")
+
+        return found_microphones_info
+
+    def _configure_capture_settings(self) -> None:
+        """
+        Validates the microphone index, sample rate and channels settings specified in the constructor's config to the un-connected microphone.
+
+        This method actually checks the specified settings and fills the sample rate and channels settings if not specified before attempting to start a PortAudio stream.
+
+        Raises:
+            RuntimeError: If one of the specified settings is not compatible with the microphone.
+            DeviceAlreadyConnectedError: If the microphone is connected when attempting to configure settings.
+        """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(
+                f"Cannot configure settings for {self} as it is already connected."
+            )
+
+        self._validate_microphone_index()
+        self._validate_sample_rate()
+        self._validate_channels()
+
+    def _validate_microphone_index(self) -> None:
+        """ "Validates the microphone index against available devices by checking if it has at least one input channel."""
+
+        try:
+            PortAudioMicrophone.find_microphones(self.microphone_index, self.sounddevice_sdk)
+        except RuntimeError as e:
+            raise RuntimeError(
+                f"{e}. Available microphones: {PortAudioMicrophone.find_microphones(sounddevice_sdk=self.sounddevice_sdk)}"
+            ) from e
+
+    def _validate_sample_rate(self) -> None:
+        """Validates the sample rate against the actual microphone's default sample rate."""
+
+        actual_sample_rate = PortAudioMicrophone.find_microphones(
+            self.microphone_index, self.sounddevice_sdk
+        )["sample_rate"]
+
+        if self.sample_rate is not None:
+            try:
+                self.sample_rate = int(self.sample_rate)
+            except ValueError as e:
+                raise RuntimeError(
+                    f"Cannot convert the provided sample rate ({self.sample_rate} Hz) to an integer."
+                ) from e
+
+            if self.sample_rate > actual_sample_rate or self.sample_rate < 1000:
+                raise RuntimeError(
+                    f"Provided sample rate {self.sample_rate} is either too low or too high compared to the sample rate of the microphone {actual_sample_rate}."
+                )
+            else:
+                if self.sample_rate < actual_sample_rate:
+                    logger.warning(
+                        "Provided sample rate is lower than the sample rate of the microphone. Performance may be impacted."
+                    )
+        else:
+            self.sample_rate = actual_sample_rate
+
+    def _validate_channels(self) -> None:
+        """Validates the channels against the actual microphone's maximum input channels."""
+
+        actual_channels = PortAudioMicrophone.find_microphones(self.microphone_index, self.sounddevice_sdk)[
+            "channels"
+        ]
+
+        if self.channels is not None and len(self.channels) > 0:
+            if not all(channel in actual_channels for channel in self.channels):
+                raise RuntimeError(
+                    f"Some of the provided channels {self.channels} are outside the possible channel range of the microphone {actual_channels}."
+                )
+        else:
+            self.channels = actual_channels
+
+        # Get channels index instead of number for slicing
+        self.channels_index = np.array(self.channels) - 1
+
+    def connect(self) -> None:
+        """
+        Connects the microphone and checks if the requested acquisition parameters are compatible with the microphone.
+        """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"Microphone {self.microphone_index} is already connected.")
+
+        self._configure_capture_settings()
+
+        # Create or reset queue and shared array
+        self.read_shared_array = SharedArray(
+            shape=(self.sample_rate * 10, len(self.channels)),
+            dtype=np.dtype("float32"),
+        )
+        self.local_read_shared_array = self.read_shared_array.get_local_array()
+        self.write_queue = process_Queue()
+
+        # Reset events
+        self.record_start_event.clear()
+        self.record_stop_event.clear()
+        self.record_close_event.clear()
+        self.record_is_started_event.clear()
+        self.audio_callback_start_event.clear()
+
+        # Create and start an audio input stream with a recording callback
+        # Remark: this is done in a separate process so that audio recording is not impacted by the main thread CPU usage, especially the precise_sleep function.
+        process_init_event = process_Event()
+        self.record_process = Process(
+            target=self._record_process,
+            args=(
+                self.microphone_index,
+                self.sample_rate,
+                self.channels,
+                process_init_event,
+                self.record_start_event,
+                self.record_stop_event,
+                self.record_close_event,
+                self.record_is_started_event,
+                self.audio_callback_start_event,
+                self.write_queue,
+                self.read_shared_array,
+                self.sounddevice_sdk,
+            ),
+        )
+        self.record_process.daemon = True
+        self.record_process.start()
+
+        is_init = process_init_event.wait(
+            timeout=5.0
+        )  # Wait for the recording process to be started, and to potentially raise an error on failure.
+        if not self.is_connected or not is_init:
+            raise RuntimeError(f"Error connecting microphone {self.microphone_index}.")
+
+        logger.info(f"{self} connected.")
+
+    def disconnect(self) -> None:
+        """
+        Disconnects the microphone and stops the recording.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Microphone {self.microphone_index} is not connected.")
+
+        if self.is_recording:
+            self.stop_recording()
+
+        self.record_close_event.set()
+        self.read_shared_array.delete()
+        self.write_queue.close()
+        self.record_process.join()
+
+        if self.is_connected:
+            raise RuntimeError(f"Error disconnecting microphone {self.microphone_index}.")
+
+        logger.info(f"{self} disconnected.")
+
+    def _read(self) -> np.ndarray:
+        """
+        Thread/Process-safe callback to read available audio data
+        """
+        return self.read_shared_array.read(self.local_read_shared_array, flush=True)
+
+    def read(self) -> np.ndarray:
+        """
+        Reads the last audio chunk recorded by the microphone, e.g. all samples recorded since the last read or since the beginning of the recording.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Microphone {self.microphone_index} is not connected.")
+        if not self.is_recording:
+            raise RuntimeError(f"Microphone {self.microphone_index} is not recording.")
+
+        start_time = time.perf_counter()
+
+        audio_readings = self._read()
+
+        # log the number of seconds it took to read the audio chunk
+        self.logs["delta_timestamp_s"] = time.perf_counter() - start_time
+
+        # log the utc time at which the audio chunk was received
+        self.logs["timestamp_utc"] = time.perf_counter()
+
+        return audio_readings
+
+    @staticmethod
+    def _record_process(
+        microphone_index,
+        sample_rate,
+        channels,
+        process_init_event,
+        record_start_event,
+        record_stop_event,
+        record_close_event,
+        record_is_started_event,
+        audio_callback_start_event,
+        write_queue,
+        read_shared_array,
+        sounddevice_sdk,
+    ) -> None:
+        """
+        Process callback used to create an unpickable sounddevice audio input stream with a recording callback and start, stop and close it based on multiprocessing events.
+        """
+
+        channels_index = np.array(channels) - 1
+        local_read_shared_array = read_shared_array.get_local_array()
+
+        def audio_callback(indata, frames, timestamp, status) -> None:
+            """
+            Low-level sounddevice callback.
+            """
+            if status:
+                logger.warning(status)
+            if audio_callback_start_event.is_set():
+                write_queue.put_nowait(indata[:, channels_index])
+                read_shared_array.write(local_read_shared_array, indata[:, channels_index])
+
+        # Create the audio stream
+        # InputStream must be instantiated in the process as it is not pickable.
+        stream = sounddevice_sdk.InputStream(
+            device=microphone_index,
+            samplerate=sample_rate,
+            channels=max(channels),
+            dtype="float32",
+            blocksize=0,  # Varying input buffer length, but no additional latency
+            latency="low",  # Low latency mode (not enabled by default !)
+            # never_drop_input=True, # Disabled as it generates an error for some devices
+            callback=audio_callback,
+        )
+        process_init_event.set()
+
+        while True:
+            start_flag = record_start_event.wait(timeout=0.1)
+            if record_close_event.is_set():
+                break
+            elif not start_flag:
+                continue
+            stream.start()
+            record_is_started_event.set()
+            record_stop_event.wait()
+            stream.stop()  # stream.stop() waits for all buffers to be processed, stream.abort() flushes the buffers !
+            record_is_started_event.clear()
+        stream.close()
+
+    def start_recording(
+        self,
+        output_file: str | None = None,
+        multiprocessing: bool | None = False,
+        overwrite: bool | None = True,
+        barrier: Barrier | None = None,
+    ) -> None:
+        """
+        Starts the recording of the microphone. If output_file is provided, the audio will be written to this file.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Microphone {self.microphone_index} is not connected.")
+        if self.is_recording:
+            raise DeviceAlreadyRecordingError(f"Microphone {self.microphone_index} is already recording.")
+
+        # Reset queue and shared memory
+        self.read_shared_array.reset()
+        self._clear_queue(self.write_queue)
+
+        # Reset stop event
+        self.record_stop_event.clear()
+
+        # Write recordings into a file if output_file is provided
+        if output_file is not None:
+            output_file = Path(output_file)
+            output_file.parent.mkdir(parents=True, exist_ok=True)
+
+            if output_file.exists():
+                if overwrite:
+                    output_file.unlink()
+                else:
+                    raise FileExistsError(
+                        f"Output file {output_file} already exists. Set overwrite to True to overwrite it."
+                    )
+
+            if multiprocessing:
+                self.write_stop_event = process_Event()
+                self.write_is_started_event = process_Event()
+                self.write_thread = Process(
+                    target=PortAudioMicrophone._write_loop,
+                    args=(
+                        self.write_queue,
+                        self.write_stop_event,
+                        self.write_is_started_event,
+                        self.sample_rate,
+                        self.channels,
+                        output_file,
+                    ),
+                )
+            else:
+                self.write_stop_event = thread_Event()
+                self.write_is_started_event = thread_Event()
+                self.write_thread = Thread(
+                    target=PortAudioMicrophone._write_loop,
+                    args=(
+                        self.write_queue,
+                        self.write_stop_event,
+                        self.write_is_started_event,
+                        self.sample_rate,
+                        self.channels,
+                        output_file,
+                    ),
+                )
+            self.write_thread.daemon = True
+            self.write_thread.start()
+            self.write_is_started_event.wait()  # Wait for the writing thread/process to be started.
+
+        self.record_start_event.set()  # Start the input audio stream process
+        self.record_is_started_event.wait()  # Wait for the input audio stream process to be actually started
+
+        if barrier is not None:
+            barrier.wait()  # Wait for multiple input audio streams to be started at the same time
+
+        self.audio_callback_start_event.set()
+
+        if not self.is_recording:
+            raise RuntimeError(f"Error starting recording for microphone {self.microphone_index}.")
+        if output_file is not None and not self.is_writing:
+            raise RuntimeError(f"Error starting writing for microphone {self.microphone_index}.")
+
+    def stop_recording(self) -> None:
+        """
+        Stops the recording of the microphones.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Microphone {self.microphone_index} is not connected.")
+        if not self.is_recording:
+            raise DeviceNotRecordingError(f"Microphone {self.microphone_index} is not recording.")
+
+        self.audio_callback_start_event.clear()
+        self.record_start_event.clear()  # Ensures the audio stream is not started again !
+        self.record_stop_event.set()
+
+        # Wait for the stream to be stopped (might lead to race condition if the stream is not properly stopped on array reset and queue clearing)
+        timeout = 1.0
+        while self.is_recording and timeout > 0:
+            time.sleep(0.01)
+            timeout -= 0.01
+
+        self.read_shared_array.reset()
+        self._clear_queue(self.write_queue, join_queue=True)
+
+        if self.is_writing:
+            self.write_stop_event.set()
+            self.write_thread.join()
+
+        if self.is_recording:
+            raise RuntimeError(f"Error stopping recording for microphone {self.microphone_index}.")
+        if self.is_writing:
+            raise RuntimeError(f"Error stopping writing for microphone {self.microphone_index}.")
+
+    @staticmethod
+    def _write_loop(
+        queue,
+        write_stop_event: Event,
+        write_is_started_event: Event,
+        sample_rate: int,
+        channels: list[int],
+        output_file: Path,
+    ) -> None:
+        """
+        Thread/Process-safe loop to write audio data into a file.
+        """
+        # Can only be run on a single process/thread for file writing safety
+        with SoundFile(
+            output_file,
+            mode="w",
+            samplerate=sample_rate,
+            channels=len(channels),
+            format="WAV",
+            subtype="FLOAT",  # By default, a much lower quality WAV file is created !
+        ) as file:
+            write_is_started_event.set()
+            while not write_stop_event.is_set():
+                try:
+                    file.write(
+                        queue.get(timeout=0.005)
+                    )  # Timeout set as the usual sounddevice buffer size. get_nowait is not possible here as it saturates the thread.
+                    queue.task_done()
+                except Empty:
+                    continue
+        write_is_started_event.clear()
+
+    def __del__(self) -> None:
+        if self.is_connected:
+            self.disconnect()
+
+    @staticmethod
+    def _clear_queue(queue, join_queue: bool = False):
+        """
+        Clears the queue by getting all items until it is empty. The longer the queue, the longer it takes to clear it.
+        """
+        try:
+            while True:
+                queue.get_nowait()
+                queue.task_done()
+        except Empty:
+            if join_queue:
+                queue.join()
+            return