Minor addition

Placed robot docs in robot folders and created symlink for docs
Check and improve spelling
2025-05-21 16:36:08 +02:00 · 2025-05-21 16:36:04 +02:00 · 2025-05-21 16:36:01 +02:00 · 2025-05-21 16:35:57 +02:00 · 2025-05-21 16:35:54 +02:00 · 2025-05-21 16:35:50 +02:00
319 changed files with 16339 additions and 23581 deletions
--- a/.cache/calibration/aloha_default/left_follower.json
+++ b/.cache/calibration/aloha_default/left_follower.json
@@ -1,68 +0,0 @@
-{
-    "homing_offset": [
-        2048,
-        3072,
-        3072,
-        -1024,
-        -1024,
-        2048,
-        -2048,
-        2048,
-        -2048
-    ],
-    "drive_mode": [
-        1,
-        1,
-        1,
-        0,
-        0,
-        1,
-        0,
-        1,
-        0
-    ],
-    "start_pos": [
-        2015,
-        3058,
-        3061,
-        1071,
-        1071,
-        2035,
-        2152,
-        2029,
-        2499
-    ],
-    "end_pos": [
-        -1008,
-        -1963,
-        -1966,
-        2141,
-        2143,
-        -971,
-        3043,
-        -1077,
-        3144
-    ],
-    "calib_mode": [
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "LINEAR"
-    ],
-    "motor_names": [
-        "waist",
-        "shoulder",
-        "shoulder_shadow",
-        "elbow",
-        "elbow_shadow",
-        "forearm_roll",
-        "wrist_angle",
-        "wrist_rotate",
-        "gripper"
-    ]
-}
--- a/.cache/calibration/aloha_default/left_leader.json
+++ b/.cache/calibration/aloha_default/left_leader.json
@@ -1,68 +0,0 @@
-{
-    "homing_offset": [
-        2048,
-        3072,
-        3072,
-        -1024,
-        -1024,
-        2048,
-        -2048,
-        2048,
-        -1024
-    ],
-    "drive_mode": [
-        1,
-        1,
-        1,
-        0,
-        0,
-        1,
-        0,
-        1,
-        0
-    ],
-    "start_pos": [
-        2035,
-        3024,
-        3019,
-        979,
-        981,
-        1982,
-        2166,
-        2124,
-        1968
-    ],
-    "end_pos": [
-        -990,
-        -2017,
-        -2015,
-        2078,
-        2076,
-        -1030,
-        3117,
-        -1016,
-        2556
-    ],
-    "calib_mode": [
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "LINEAR"
-    ],
-    "motor_names": [
-        "waist",
-        "shoulder",
-        "shoulder_shadow",
-        "elbow",
-        "elbow_shadow",
-        "forearm_roll",
-        "wrist_angle",
-        "wrist_rotate",
-        "gripper"
-    ]
-}
--- a/.cache/calibration/aloha_default/right_follower.json
+++ b/.cache/calibration/aloha_default/right_follower.json
@@ -1,68 +0,0 @@
-{
-    "homing_offset": [
-        2048,
-        3072,
-        3072,
-        -1024,
-        -1024,
-        2048,
-        -2048,
-        2048,
-        -2048
-    ],
-    "drive_mode": [
-        1,
-        1,
-        1,
-        0,
-        0,
-        1,
-        0,
-        1,
-        0
-    ],
-    "start_pos": [
-        2056,
-        2895,
-        2896,
-        1191,
-        1190,
-        2018,
-        2051,
-        2056,
-        2509
-    ],
-    "end_pos": [
-        -1040,
-        -2004,
-        -2006,
-        2126,
-        2127,
-        -1010,
-        3050,
-        -1117,
-        3143
-    ],
-    "calib_mode": [
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "LINEAR"
-    ],
-    "motor_names": [
-        "waist",
-        "shoulder",
-        "shoulder_shadow",
-        "elbow",
-        "elbow_shadow",
-        "forearm_roll",
-        "wrist_angle",
-        "wrist_rotate",
-        "gripper"
-    ]
-}
--- a/.cache/calibration/aloha_default/right_leader.json
+++ b/.cache/calibration/aloha_default/right_leader.json
@@ -1,68 +0,0 @@
-{
-    "homing_offset": [
-        2048,
-        3072,
-        3072,
-        -1024,
-        -1024,
-        2048,
-        -2048,
-        2048,
-        -2048
-    ],
-    "drive_mode": [
-        1,
-        1,
-        1,
-        0,
-        0,
-        1,
-        0,
-        1,
-        0
-    ],
-    "start_pos": [
-        2068,
-        3034,
-        3030,
-        1038,
-        1041,
-        1991,
-        1948,
-        2090,
-        1985
-    ],
-    "end_pos": [
-        -1025,
-        -2014,
-        -2015,
-        2058,
-        2060,
-        -955,
-        3091,
-        -940,
-        2576
-    ],
-    "calib_mode": [
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "DEGREE",
-        "LINEAR"
-    ],
-    "motor_names": [
-        "waist",
-        "shoulder",
-        "shoulder_shadow",
-        "elbow",
-        "elbow_shadow",
-        "forearm_roll",
-        "wrist_angle",
-        "wrist_rotate",
-        "gripper"
-    ]
-}
--- a/.gitattributes
+++ b/.gitattributes
@@ -11,10 +11,11 @@
 # 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.
-
 *.memmap filter=lfs diff=lfs merge=lfs -text
 *.stl filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.json !text !filter !merge !diff
+tests/artifacts/cameras/*.png filter=lfs diff=lfs merge=lfs -text
+*.bag filter=lfs diff=lfs merge=lfs -text
--- a/.github/workflows/build_documentation.yml
+++ b/.github/workflows/build_documentation.yml
@@ -0,0 +1,23 @@
+name: Build documentation
+
+on:
+  workflow_dispatch:
+  push:
+    paths:
+      - "docs/**"
+    branches:
+    - main
+    - doc-builder*
+    - v*-release
+
+
+jobs:
+  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
+    with:
+      commit_sha: ${{ github.sha }}
+      package: lerobot
+      additional_args: --not_python_module
+    secrets:
+      token: ${{ secrets.HUGGINGFACE_PUSH }}
+      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
--- a/.github/workflows/build_pr_documentation.yml
+++ b/.github/workflows/build_pr_documentation.yml
@@ -0,0 +1,19 @@
+name: Build PR Documentation
+
+on:
+  pull_request:
+    paths:
+      - "docs/**"
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
+    with:
+      commit_sha: ${{ github.event.pull_request.head.sha }}
+      pr_number: ${{ github.event.number }}
+      package: lerobot
+      additional_args: --not_python_module
--- a/.github/workflows/upload_pr_documentation.yml
+++ b/.github/workflows/upload_pr_documentation.yml
@@ -0,0 +1,16 @@
+name: Upload PR Documentation
+
+on: # zizmor: ignore[dangerous-triggers] We follow the same pattern as in Transformers
+  workflow_run:
+    workflows: [ "Build PR Documentation" ]
+    types:
+    - completed
+
+jobs:
+  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@main
+    with:
+      package_name: lerobot
+    secrets:
+      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
+      comment_bot_token: ${{ secrets.COMMENT_BOT_TOKEN }}
--- a/.gitignore
+++ b/.gitignore
@@ -12,6 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+# Dev scripts
+.dev
+
 # Logging
 logs
 tmp
@@ -26,7 +29,6 @@ outputs

 # VS Code
 .vscode
-.devcontainer

 # HPC
 nautilus/*.yaml
@@ -92,10 +94,8 @@ coverage.xml
 .hypothesis/
 .pytest_cache/

-# Ignore .cache except calibration
+# Ignore .cache
 .cache/*
-!.cache/calibration/
-!.cache/calibration/**

 # Translations
 *.mo
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -36,8 +36,8 @@ repos:
      - id: end-of-file-fixer
      - id: trailing-whitespace

-  - repo: https://github.com/crate-ci/typos
-    rev: v1.30.2
+  - repo: https://github.com/adhtruong/mirrors-typos
+    rev: v1.31.1
    hooks:
      - id: typos
        args: [--force-exclude]
@@ -46,9 +46,9 @@ repos:
    rev: v3.19.1
    hooks:
    -   id: pyupgrade
-        exclude: '^(.*_pb2_grpc\.py|.*_pb2\.py$)'
+
  - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.9.10
+    rev: v0.11.5
    hooks:
      - id: ruff
        args: [--fix]
@@ -57,12 +57,12 @@ repos:

  ##### Security #####
  - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.24.0
+    rev: v8.24.3
    hooks:
      - id: gitleaks

  - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.4.1
+    rev: v1.5.2
    hooks:
      - id: zizmor

--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -269,9 +269,6 @@ Follow these steps to start contributing:
   the PR as a draft PR. These are useful to avoid duplicated work, and to differentiate
   it from PRs ready to be merged;
 4. Make sure existing tests pass;
-<!-- 5. Add high-coverage tests. No quality testing = no merge.
-
-See an example of a good PR here: https://github.com/huggingface/lerobot/pull/ -->

 ### Tests

--- a/README.md
+++ b/README.md
@@ -23,21 +23,35 @@
 </div>

 <h2 align="center">
-    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
-        Build Your Own SO-100 Robot!</a></p>
+    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/12_use_so101.md">
+        Build Your Own SO-101 Robot!</a></p>
 </h2>

 <div align="center">
-  <img src="media/so100/leader_follower.webp?raw=true" alt="SO-100 leader and follower arms" title="SO-100 leader and follower arms" width="50%">
+  <div style="display: flex; gap: 1rem; justify-content: center; align-items: center;" >
+    <img
+      src="media/so101/so101.webp?raw=true"
+      alt="SO-101 follower arm"
+      title="SO-101 follower arm"
+      style="width: 40%;"
+    />
+    <img
+      src="media/so101/so101-leader.webp?raw=true"
+      alt="SO-101 leader arm"
+      title="SO-101 leader arm"
+      style="width: 40%;"
+    />
+  </div>

-  <p><strong>Meet the SO-100 – Just $110 per arm!</strong></p>
+
+  <p><strong>Meet the updated SO100, the SO-101 – Just €114 per arm!</strong></p>
  <p>Train it in minutes with a few simple moves on your laptop.</p>
  <p>Then sit back and watch your creation act autonomously! 🤯</p>

-  <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
-      Get the full SO-100 tutorial here.</a></p>
+  <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/12_use_so101.md">
+      See the full SO-101 tutorial here.</a></p>

-  <p>Want to take it to the next level? Make your SO-100 mobile by building LeKiwi!</p>
+  <p>Want to take it to the next level? Make your SO-101 mobile by building LeKiwi!</p>
  <p>Check out the <a href="https://github.com/huggingface/lerobot/blob/main/examples/11_use_lekiwi.md">LeKiwi tutorial</a> and bring your robot to life on wheels.</p>

  <img src="media/lekiwi/kiwi.webp?raw=true" alt="LeKiwi mobile robot" title="LeKiwi mobile robot" width="50%">
@@ -51,7 +65,6 @@

 ---

-
 🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.

 🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
@@ -98,18 +111,25 @@ conda create -y -n lerobot python=3.10
 conda activate lerobot
 ```

-When using `miniconda`, if you don't have `ffmpeg` in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
 ```bash
-conda install ffmpeg
+conda install ffmpeg -c conda-forge
 ```

+> **NOTE:** This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
+>  - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
+>  ```bash
+>  conda install ffmpeg=7.1.1 -c conda-forge
+>  ```
+>  - _[On Linux only]_ 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`.
+
 Install 🤗 LeRobot:
 ```bash
-pip install --no-binary=av -e .
+pip install -e .
 ```

 > **NOTE:** If you encounter build errors, you may need to install additional dependencies (`cmake`, `build-essential`, and `ffmpeg libs`). On Linux, run:
-`sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
+`sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config`. For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)

 For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
 - [aloha](https://github.com/huggingface/gym-aloha)
@@ -118,7 +138,7 @@ For simulations, 🤗 LeRobot comes with gymnasium environments that can be inst

 For instance, to install 🤗 LeRobot with aloha and pusht, use:
 ```bash
-pip install --no-binary=av -e ".[aloha, pusht]"
+pip install -e ".[aloha, pusht]"
 ```

 To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
@@ -201,7 +221,7 @@ dataset attributes:
  │  ├ episode_index (int64): index of the episode for this sample
  │  ├ frame_index (int64): index of the frame for this sample in the episode ; starts at 0 for each episode
  │  ├ timestamp (float32): timestamp in the episode
-  │  ├ next.done (bool): indicates the end of en episode ; True for the last frame in each episode
+  │  ├ next.done (bool): indicates the end of an episode ; True for the last frame in each episode
  │  └ index (int64): general index in the whole dataset
  ├ episode_data_index: contains 2 tensors with the start and end indices of each episode
  │  ├ from (1D int64 tensor): first frame index for each episode — shape (num episodes,) starts with 0
@@ -250,7 +270,7 @@ See `python lerobot/scripts/eval.py --help` for more instructions.

 ### Train your own policy

-Check out [example 3](./examples/3_train_policy.py) that illustrate how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
+Check out [example 3](./examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.

 To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding `--wandb.enable=true`.

@@ -301,7 +321,7 @@ Once you have trained a policy you may upload it to the Hugging Face hub using a
 You first need to find the checkpoint folder located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). Within that there is a `pretrained_model` directory which should contain:
 - `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
 - `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
- `train_config.json`: A consolidated configuration containing all parameter userd for training. The policy configuration should match `config.json` exactly. Thisis useful for anyone who wants to evaluate your policy or for reproducibility.
+- `train_config.json`: A consolidated configuration containing all parameters used for training. The policy configuration should match `config.json` exactly. This is useful for anyone who wants to evaluate your policy or for reproducibility.

 To upload these to the hub, run the following:
 ```bash
--- a/benchmarks/video/capture_camera_feed.py
+++ b/benchmarks/video/capture_camera_feed.py
@@ -17,12 +17,21 @@

 import argparse
 import datetime as dt
+import os
+import time
 from pathlib import Path

 import cv2
+import rerun as rr
+
+# see https://rerun.io/docs/howto/visualization/limit-ram
+RERUN_MEMORY_LIMIT = os.getenv("LEROBOT_RERUN_MEMORY_LIMIT", "5%")


-def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int):
+def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int, duration: int):
+    rr.init("lerobot_capture_camera_feed")
+    rr.spawn(memory_limit=RERUN_MEMORY_LIMIT)
+
    now = dt.datetime.now()
    capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
    if not capture_dir.exists():
@@ -39,24 +48,21 @@ def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height
    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)

    frame_index = 0
-    while True:
+    start_time = time.time()
+    while time.time() - start_time < duration:
        ret, frame = cap.read()

        if not ret:
            print("Error: Could not read frame.")
            break
-
-        cv2.imshow("Video Stream", frame)
+        rr.log("video/stream", rr.Image(frame.numpy()), static=True)
        cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
        frame_index += 1

-        # Break the loop on 'q' key press
-        if cv2.waitKey(1) & 0xFF == ord("q"):
-            break
-
-    # Release the capture and destroy all windows
+    # Release the capture
    cap.release()
-    cv2.destroyAllWindows()
+
+    # TODO(Steven): Add a graceful shutdown via a close() method for the Viewer context, though not currently supported in the Rerun API.


 if __name__ == "__main__":
@@ -86,5 +92,11 @@ if __name__ == "__main__":
        default=720,
        help="Height of the captured images.",
    )
+    parser.add_argument(
+        "--duration",
+        type=int,
+        default=20,
+        help="Duration in seconds for which the video stream should be captured.",
+    )
    args = parser.parse_args()
    display_and_save_video_stream(**vars(args))
--- a/benchmarks/video/run_video_benchmark.py
+++ b/benchmarks/video/run_video_benchmark.py
@@ -32,11 +32,7 @@ import numpy as np
 import pandas as pd
 import PIL
 import torch
-from skimage.metrics import (
-    mean_squared_error,
-    peak_signal_noise_ratio,
-    structural_similarity,
-)
+from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
 from tqdm import tqdm

 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
@@ -98,11 +94,7 @@ def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> t


 def save_decoded_frames(
-    imgs_dir: Path,
-    save_dir: Path,
-    frames: torch.Tensor,
-    timestamps: list[float],
-    fps: int,
+    imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
 ) -> None:
    if save_dir.exists() and len(list(save_dir.glob("frame_*.png"))) == len(timestamps):
        return
@@ -112,10 +104,7 @@ def save_decoded_frames(
        idx = int(ts * fps)
        frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame_{idx:06d}_decoded.png")
-        shutil.copyfile(
-            imgs_dir / f"frame_{idx:06d}.png",
-            save_dir / f"frame_{idx:06d}_original.png",
-        )
+        shutil.copyfile(imgs_dir / f"frame_{idx:06d}.png", save_dir / f"frame_{idx:06d}_original.png")


 def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
@@ -131,11 +120,7 @@ def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
    imgs_dataset = hf_dataset.select_columns(img_keys[0])

    for i, item in enumerate(
-        tqdm(
-            imgs_dataset,
-            desc=f"saving {dataset.repo_id} first episode images",
-            leave=False,
-        )
+        tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
    ):
        img = item[img_keys[0]]
        img.save(str(imgs_dir / f"frame_{i:06d}.png"), quality=100)
@@ -290,9 +275,7 @@ def benchmark_encoding_decoding(
    random.seed(seed)
    benchmark_table = []
    for timestamps_mode in tqdm(
-        decoding_cfg["timestamps_modes"],
-        desc="decodings (timestamps_modes)",
-        leave=False,
+        decoding_cfg["timestamps_modes"], desc="decodings (timestamps_modes)", leave=False
    ):
        for backend in tqdm(decoding_cfg["backends"], desc="decodings (backends)", leave=False):
            benchmark_row = benchmark_decoding(
@@ -433,7 +416,7 @@ if __name__ == "__main__":
        "--vcodec",
        type=str,
        nargs="*",
-        default=["libx264", "libx265", "libsvtav1"],
+        default=["libx264", "hevc", "libsvtav1"],
        help="Video codecs to be tested",
    )
    parser.add_argument(
@@ -463,7 +446,7 @@ if __name__ == "__main__":
    #     nargs="*",
    #     default=[0, 1],
    #     help="Use the fastdecode tuning option. 0 disables it. "
-    #         "For libx264 and libx265, only 1 is possible. "
+    #         "For libx264 and libx265/hevc, only 1 is possible. "
    #         "For libsvtav1, 1, 2 or 3 are possible values with a higher number meaning a faster decoding optimization",
    # )
    parser.add_argument(
--- a/docker/lerobot-gpu-dev/Dockerfile
+++ b/docker/lerobot-gpu-dev/Dockerfile
@@ -14,7 +14,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
    tcpdump sysstat screen tmux \
    libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
    speech-dispatcher portaudio19-dev libgeos-dev \
-    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
+    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv python${PYTHON_VERSION}-dev \
    && apt-get clean && rm -rf /var/lib/apt/lists/*

 # Install ffmpeg build dependencies. See:
--- a/docs/README.md
+++ b/docs/README.md
@@ -0,0 +1,137 @@
+<!---
+Copyright 2020 The HuggingFace 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.
+-->
+
+# Generating the documentation
+
+To generate the documentation, you first have to build it. Several packages are necessary to build the doc,
+you can install them with the following command, at the root of the code repository:
+
+```bash
+pip install -e ".[docs]"
+```
+
+You will also need `nodejs`. Please refer to their [installation page](https://nodejs.org/en/download)
+
+---
+**NOTE**
+
+You only need to generate the documentation to inspect it locally (if you're planning changes and want to
+check how they look before committing for instance). You don't have to `git commit` the built documentation.
+
+---
+
+## Building the documentation
+
+Once you have setup the `doc-builder` and additional packages, you can generate the documentation by
+typing the following command:
+
+```bash
+doc-builder build lerobot docs/source/ --build_dir ~/tmp/test-build
+```
+
+You can adapt the `--build_dir` to set any temporary folder that you prefer. This command will create it and generate
+the MDX files that will be rendered as the documentation on the main website. You can inspect them in your favorite
+Markdown editor.
+
+## Previewing the documentation
+
+To preview the docs, first install the `watchdog` module with:
+
+```bash
+pip install watchdog
+```
+
+Then run the following command:
+
+```bash
+doc-builder preview lerobot docs/source/
+```
+
+The docs will be viewable at [http://localhost:3000](http://localhost:3000). You can also preview the docs once you have opened a PR. You will see a bot add a comment to a link where the documentation with your changes lives.
+
+---
+**NOTE**
+
+The `preview` command only works with existing doc files. When you add a completely new file, you need to update `_toctree.yml` & restart `preview` command (`ctrl-c` to stop it & call `doc-builder preview ...` again).
+
+---
+
+## Adding a new element to the navigation bar
+
+Accepted files are Markdown (.md).
+
+Create a file with its extension and put it in the source directory. You can then link it to the toc-tree by putting
+the filename without the extension in the [`_toctree.yml`](https://github.com/huggingface/lerobot/blob/main/docs/source/_toctree.yml) file.
+
+## Renaming section headers and moving sections
+
+It helps to keep the old links working when renaming the section header and/or moving sections from one document to another. This is because the old links are likely to be used in Issues, Forums, and Social media and it'd make for a much more superior user experience if users reading those months later could still easily navigate to the originally intended information.
+
+Therefore, we simply keep a little map of moved sections at the end of the document where the original section was. The key is to preserve the original anchor.
+
+So if you renamed a section from: "Section A" to "Section B", then you can add at the end of the file:
+
+```
+Sections that were moved:
+
+[ <a href="#section-b">Section A</a><a id="section-a"></a> ]
+```
+and of course, if you moved it to another file, then:
+
+```
+Sections that were moved:
+
+[ <a href="../new-file#section-b">Section A</a><a id="section-a"></a> ]
+```
+
+Use the relative style to link to the new file so that the versioned docs continue to work.
+
+For an example of a rich moved sections set please see the very end of [the transformers Trainer doc](https://github.com/huggingface/transformers/blob/main/docs/source/en/main_classes/trainer.md).
+
+### Adding a new tutorial
+
+Adding a new tutorial or section is done in two steps:
+
+- Add a new file under `./source`. This file can either be ReStructuredText (.rst) or Markdown (.md).
+- Link that file in `./source/_toctree.yml` on the correct toc-tree.
+
+Make sure to put your new file under the proper section. If you have a doubt, feel free to ask in a Github Issue or PR.
+
+### Writing source documentation
+
+Values that should be put in `code` should either be surrounded by backticks: \`like so\`. Note that argument names
+and objects like True, None or any strings should usually be put in `code`.
+
+#### Writing a multi-line code block
+
+Multi-line code blocks can be useful for displaying examples. They are done between two lines of three backticks as usual in Markdown:
+
+
+````
+```
+# first line of code
+# second line
+# etc
+```
+````
+
+#### Adding an image
+
+Due to the rapidly growing repository, it is important to make sure that no files that would significantly weigh down the repository are added. This includes images, videos, and other non-text files. We prefer to leverage a hf.co hosted `dataset` like
+the ones hosted on [`hf-internal-testing`](https://huggingface.co/hf-internal-testing) in which to place these files and reference
+them by URL. We recommend putting them in the following dataset: [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images).
+If an external contribution, feel free to add the images to your PR and ask a Hugging Face member to migrate your images
+to this dataset.
--- a/docs/source/_toctree.yml
+++ b/docs/source/_toctree.yml
@@ -0,0 +1,26 @@
+- sections:
+  - local: index
+    title: LeRobot
+  - local: installation
+    title: Installation
+  title: Get started
+- sections:
+  - local: getting_started_real_world_robot
+    title: Getting Started with Real-World Robots
+  - local: cameras
+    title: Cameras
+  title: "Tutorials"
+- sections:
+  - local: so101
+    title: SO-101
+  - local: so100
+    title: SO-100
+  - local: koch
+    title: Koch v1.1
+  - local: lekiwi
+    title: LeKiwi
+  title: "Robots"
+- sections:
+  - local: contributing
+    title: Contribute to LeRobot
+  title: "Contribute"
--- a/docs/source/cameras.mdx
+++ b/docs/source/cameras.mdx
@@ -0,0 +1,190 @@
+# Cameras
+
+Here we describe how to set up and use a camera with LeRobot. We support different ways of capturing videos in LeRobot, such as using a phone camera, an integrated laptop camera, an external webcam, or an Intel realsense camera.
+## Set up Cameras
+
+There are three ways to connect and use a camera with LeRobot:
+1. Use [Camera Class](./setup_cameras?use+phone=Mac#use-opencvcamera), which allows you to use any camera: usb, realsense, laptop webcam
+2. Use [iPhone camera](./setup_cameras?use+phone=Mac#use-your-phone) with MacOS
+3. Use [Phone camera](./setup_cameras?use+phone=Linux#use-your-phone) on Linux
+
+### Use Camera Class
+
+In LeRobot, you can efficiently record frames from most cameras using either the OpenCVCamera class or the RealSenseCamera class. For more details on compatibility for the OpenCVCamera class, see [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 index. When you only have one camera like a webcam of a laptop, the camera index is usually `0` but it might differ, and the camera index might change if you reboot your computer or re-plug your camera. This behavior depends on your operating system.
+
+To find the camera indices, run the following script:
+```bash
+python lerobot/find_cameras.py list-cameras
+```
+
+The output will look something like this if you have two cameras connected:
+```
+--- Detected Cameras ---
+Camera #0:
+  Name: OpenCV Camera @ 0
+  Type: OpenCV
+  Id: 0
+  Backend api: AVFOUNDATION
+  Default stream profile:
+    Format: 16.0
+    Width: 1920
+    Height: 1080
+    Fps: 15.0
+--------------------
+Camera #1:
+  Name: OpenCV Camera @ 1
+  Type: OpenCV
+  Id: 1
+  Backend api: AVFOUNDATION
+  Default stream profile:
+    Format: 16.0
+    Width: 1920
+    Height: 1080
+    Fps: 1.0
+--------------------
+```
+
+> [!WARNING]
+> On , you could get this error: `Error finding RealSense cameras: failed to set power state`, this can be solved by running the same command with `sudo` permissions.
+
+### Use your phone
+<hfoptions id="use phone">
+<hfoption id="Mac">
+
+To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
+- Ensure your Mac is running macOS 13 or later, and your iPhone is on iOS 16 or later.
+- Sign in both devices with the same Apple ID.
+- Connect your devices with a USB cable or turn on Wi-Fi and Bluetooth for a wireless connection.
+
+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="Linux">
+
+If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
+
+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:
+```python
+sudo apt install v4l2loopback-dkms v4l-utils
+```
+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):
+```python
+flatpak install flathub com.obsproject.Studio
+```
+4. *Install the DroidCam OBS plugin*. This plugin integrates DroidCam with OBS Studio. Install it with:
+```python
+flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
+```
+5. *Start OBS Studio*. Launch with:
+```python
+flatpak run com.obsproject.Studio
+```
+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*. Use `v4l2-ctl` to list the devices:
+```python
+v4l2-ctl --list-devices
+```
+You should see an entry like:
+```
+VirtualCam (platform:v4l2loopback-000):
+/dev/video1
+```
+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`.
+```python
+v4l2-ctl -d /dev/video1 --get-fmt-video
+```
+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>
+
+## Use Cameras
+
+Below are two examples, demonstrating how to work with the API.
+
+- **Asynchronous frame capture** using an OpenCV-based camera
+- **Color and depth capture** using an Intel RealSense camera
+
+### Asynchronous OpenCV Camera
+
+This snippet shows how to:
+
+* Construct an `OpenCVCameraConfig` with your desired FPS, resolution, color mode, and rotation.
+* Instantiate and connect an `OpenCVCamera`, performing a warm-up read.
+* Read frames asynchronously in a loop via `async_read(timeout_ms)`.
+
+```python
+from lerobot.common.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.common.cameras.opencv.camera_opencv import OpenCVCamera
+from lerobot.common.cameras.configs import ColorMode, Cv2Rotation
+
+config = OpenCVCameraConfig(
+    index_or_path=0,
+    fps=15,
+    width=1920,
+    height=1080,
+    color_mode=ColorMode.RGB,
+    rotation=Cv2Rotation.NO_ROTATION
+)
+
+camera = OpenCVCamera(config)
+camera.connect(do_warmup_read=True)
+
+try:
+    for i in range(10):
+        frame = camera.async_read(timeout_ms=1000)
+        print(f"Async frame {i} shape:", frame.shape)
+finally:
+    camera.disconnect()
+```
+
+### Intel RealSense Camera (Color + Depth)
+
+This snippet shows how to:
+
+* Create a `RealSenseCameraConfig` specifying your camera’s serial number and enabling depth.
+* Instantiate and connect a `RealSenseCamera` with warm-up.
+* Capture a color frame via `read()` and a depth map via `read_depth()`.
+
+```python
+from lerobot.common.cameras.intel.configuration_realsense import RealSenseCameraConfig
+from lerobot.common.cameras.intel.camera_realsense import RealSenseCamera
+from lerobot.common.cameras.configs import ColorMode, Cv2Rotation
+
+config = RealSenseCameraConfig(
+    serial_number="233522074606",
+    fps=15,
+    width=640,
+    height=480,
+    color_mode=ColorMode.RGB,
+    use_depth=True,
+    rotation=Cv2Rotation.NO_ROTATION
+)
+
+camera = RealSenseCamera(config)
+camera.connect(do_warmup_read=True)
+
+try:
+    color_frame = camera.read()
+    depth_map = camera.read_depth()
+    print("Color frame shape:", color_frame.shape)
+    print("Depth map shape:", depth_map.shape)
+finally:
+    camera.disconnect()
+```
--- a/docs/source/contributing.md
+++ b/docs/source/contributing.md
@@ -0,0 +1 @@
+../../CONTRIBUTING.md
--- a/docs/source/getting_started_real_world_robot.mdx
+++ b/docs/source/getting_started_real_world_robot.mdx
@@ -0,0 +1,384 @@
+# Getting Started with Real-World Robots
+
+This tutorial will explain how to train a neural network to control a real robot autonomously.
+
+**You'll learn:**
+1. How to record and visualize your dataset.
+2. How to train a policy using your data and prepare it for evaluation.
+3. How to evaluate your policy and visualize the results.
+
+By following these steps, you'll be able to replicate tasks, such as picking up a Lego block and placing it in a bin with a high success rate, as shown in [this video](https://x.com/RemiCadene/status/1814680760592572934).
+
+This tutorial isn’t tied to a specific robot: we walk you through the commands and API snippets you can adapt for any supported platform.
+
+During data collection, you’ll use a “teloperation” device, such as a leader arm or keyboard to teleoperate the robot and record its motion trajectories.
+
+Once you’ve gathered enough trajectories, you’ll train a neural network to imitate these trajectories and deploy the trained model so your robot can perform the task autonomously.
+
+If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.
+
+## Set up and Calibrate
+
+If you haven't yet set up and calibrated your robot and teleop device, please do so by following the robot-specific tutorial.
+
+## Teleoperate
+
+In this example, we’ll demonstrate how to teleoperate the SO101 robot. For each command, we also provide a corresponding API example.
+
+<hfoptions id="teleoperate_so101">
+<hfoption id="Command">
+```bash
+python -m lerobot.teleoperate \
+    --robot.type=so101_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.cameras="{}" \
+    --robot.id=my_red_robot_arm \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=my_blue_leader_arm
+```
+</hfoption>
+<hfoption id="API example">
+```python
+from lerobot.common.teleoperators.so101_leader import SO101LeaderConfig, SO101Leader
+from lerobot.common.robots.so101_follower import SO101FollowerConfig, SO101Follower
+
+robot_config = SO101FollowerConfig(
+    port="/dev/tty.usbmodem58760431541",
+    id="my_red_robot_arm",
+)
+
+teleop_config = SO101LeaderConfig(
+    port="/dev/tty.usbmodem58760431551",
+    id="my_blue_leader_arm",
+)
+
+robot = SO101Follower(robot_config)
+teleop_device = SO101Leader(teleop_config)
+robot.connect()
+teleop_device.connect()
+
+while True:
+    action = teleop_device.get_action()
+    robot.send_action(action)
+```
+</hfoption>
+</hfoptions>
+
+The teleoperate command will automatically:
+1. Identify any missing calibrations and initiate the calibration procedure.
+2. Connect the robot and teleop device and start teleoperation.
+
+## Cameras
+
+To add cameras to your setup, follow this [Guide](./cameras#setup-cameras).
+
+## Teleoperate with cameras
+
+With `rerun`, you can teleoperate again while simultaneously visualizing the camera feeds and joint positions. In this example, we’re using the Koch arm.
+
+<hfoptions id="teleoperate_koch_camera">
+<hfoption id="Command">
+```bash
+python -m lerobot.teleoperate \
+    --robot.type=koch_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.id=my_koch_robot \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
+    --teleop.type=koch_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=my_koch_teleop \
+    --display_data=true
+```
+</hfoption>
+<hfoption id="API example">
+```python
+from lerobot.common.cameras.opencv.configuration_opencv import OpenCVCameraConfig
+from lerobot.common.teleoperators.koch_leader import KochLeaderConfig, KochLeader
+from lerobot.common.robots.koch_follower import KochFollowerConfig, KochFollower
+
+camera_config = {
+    "front": OpenCVCameraConfig(index_or_path=0, width=1920, height=1080, fps=30)
+}
+
+robot_config = KochFollowerConfig(
+    port="/dev/tty.usbmodem585A0076841",
+    id="my_red_robot_arm",
+    cameras=camera_config
+)
+
+teleop_config = KochLeaderConfig(
+    port="/dev/tty.usbmodem58760431551",
+    id="my_blue_leader_arm",
+)
+
+robot = KochFollower(robot_config)
+teleop_device = KochLeader(teleop_config)
+robot.connect()
+teleop_device.connect()
+
+while True:
+    observation = robot.get_observation()
+    action = teleop_device.get_action()
+    robot.send_action(action)
+```
+</hfoption>
+</hfoptions>
+
+## Teleoperate LeKiwi
+
+> [!TIP]
+> If you're using a Mac, you might need to give Terminal permission to access your keyboard for teleoperation. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
+
+TODO(pepijn): modify these commands with new API, and explain you can also map the arms to a keyboard with the new API now
+
+To teleoperate SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
+```bash
+python -m lerobot.common.robots.lekiwi.lekiwi_host
+```
+
+Then on your laptop, also run `conda activate lerobot` and this script:
+```bash
+python -m lerobot.teleoperate \
+    --robot.type=lekiwi \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.cameras="{}" \
+    --robot.id=my_lekiwi \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=my_blue_leader_arm
+```
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`. For the `--control.type=remote_robot` you will also need to set `--control.viewer_ip` and `--control.viewer_port`
+
+You should see on your laptop something like this: ```[INFO] Connected to remote robot at tcp://172.17.133.91:5555 and video stream at tcp://172.17.133.91:5556.``` Now you can move the leader arm and use the keyboard (w,a,s,d) to drive forward, left, backwards, right. And use (z,x) to turn left or turn right. You can use (r,f) to increase and decrease the speed of the mobile robot. There are three speed modes, see the table below:
+
+| Speed Mode | Linear Speed (m/s) | Rotation Speed (deg/s) |
+| ---------- | ------------------ | ---------------------- |
+| Fast       | 0.4                | 90                     |
+| Medium     | 0.25               | 60                     |
+| Slow       | 0.1                | 30                     |
+
+
+| Key | Action         |
+| --- | -------------- |
+| W   | Move forward   |
+| A   | Move left      |
+| S   | Move backward  |
+| D   | Move right     |
+| Z   | Turn left      |
+| X   | Turn right     |
+| R   | Increase speed |
+| F   | Decrease speed |
+
+> [!TIP]
+>  If you use a different keyboard, you can change the keys for each command in the [`LeKiwiConfig`](../lerobot/common/robot_devices/robots/configs.py).
+
+### Wired version
+If you have the **wired** LeKiwi version, please run all commands on your laptop.
+
+## Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset.
+
+We use the Hugging Face hub features for uploading your dataset. If you haven't previously used the Hub, make sure you can login via the cli using a write-access token, this token can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens).
+
+Add your token to the CLI by running this command:
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Then store your Hugging Face repository name in a variable:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Now you can record a dataset. To record 2 episodes and upload your dataset to the hub, execute this command tailored to the SO101.
+```bash
+python -m lerobot.record \
+    --robot.type=so101_follower \
+    --robot.port=/dev/tty.usbmodem585A0076841 \
+    --robot.id=my_red_robot_arm \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=my_blue_leader_arm \
+    --display_data=true \
+    --dataset.repo_id=aliberts/record-test \
+    --dataset.num_episodes=2 \
+    --dataset.single_task="Grab the black cube"
+```
+
+You will see a lot of lines appearing like this one:
+```
+INFO 2024-08-10 15:02:58 ol_robot.py:219 dt:33.34 (30.0hz) dtRlead: 5.06 (197.5hz) dtWfoll: 0.25 (3963.7hz) dtRfoll: 6.22 (160.7hz) dtRlaptop: 32.57 (30.7hz) dtRphone: 33.84 (29.5hz)
+```
+
+| Field | Meaning |
+|:---|:---|
+| `2024-08-10 15:02:58` | Timestamp when `print` was called. |
+| `ol_robot.py:219` | Source file and line number of the `print` call (`lerobot/scripts/control_robot.py` at line `219`). |
+| `dt: 33.34 (30.0 Hz)` | Delta time (ms) between teleop steps (target: 30.0 Hz, `--fps 30`). Yellow if step is too slow. |
+| `dtRlead: 5.06 (197.5 Hz)` | Delta time (ms) for reading present position from the **leader arm**. |
+| `dtWfoll: 0.25 (3963.7 Hz)` | Delta time (ms) for writing goal position to the **follower arm** (asynchronous). |
+| `dtRfoll: 6.22 (160.7 Hz)` | Delta time (ms) for reading present position from the **follower arm**. |
+| `dtRlaptop: 32.57 (30.7 Hz)` | Delta time (ms) for capturing an image from the **laptop camera** (async thread). |
+| `dtRphone: 33.84 (29.5 Hz)` | Delta time (ms) for capturing an image from the **phone camera** (async thread). |
+
+
+#### Dataset upload
+Locally, your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}` (e.g. `data/cadene/so101_test`). At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/so101_test) that you can obtain by running:
+```bash
+echo https://huggingface.co/datasets/${HF_USER}/so101_test
+```
+Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
+
+You can look for other LeRobot datasets on the hub by searching for `LeRobot` [tags](https://huggingface.co/datasets?other=LeRobot).
+
+#### Record function
+
+The `record` function provides a suite of tools for capturing and managing data during robot operation:
+
+##### 1. Data Storage
+- Data is stored using the `LeRobotDataset` format and is stored on disk during recording.
+- By default, the dataset is pushed to your Hugging Face page after recording.
+  - To disable uploading, use `--dataset.push_to_hub=False`.
+
+##### 2. Checkpointing and Resuming
+- Checkpoints are automatically created during recording.
+- If an issue occurs, you can resume by re-running the same command with `--control.resume=true`.
+- To start recording from scratch, **manually delete** the dataset directory.
+
+##### 3. Recording Parameters
+Set the flow of data recording using command-line arguments:
+- `--warmup_time_s=10`
+  Number of seconds before starting data collection (default: **10 seconds**).
+  Allows devices to warm up and synchronize.
+- `--dataset.episode_time_s=60`
+  Duration of each data recording episode (default: **60 seconds**).
+- `--dataset.reset_time_s=60`
+  Duration for resetting the environment after each episode (default: **60 seconds**).
+- `--dataset.num_episodes=50`
+  Total number of episodes to record (default: **50**).
+
+##### 4. Keyboard Controls During Recording
+Control the data recording flow using keyboard shortcuts:
+- Press **Right Arrow (`→`)**: Early stop the current episode or reset time and move to the next.
+- Press **Left Arrow (`←`)**: Cancel the current episode and re-record it.
+- Press **Escape (`ESC`)**: Immediately stop the session, encode videos, and upload the dataset.
+
+#### Tips for gathering data
+
+Once you're comfortable with data recording, you can create a larger dataset for training. A good starting task is grasping an object at different locations and placing it in a bin. We suggest recording at least 50 episodes, with 10 episodes per location. Keep the cameras fixed and maintain consistent grasping behavior throughout the recordings. Also make sure the object you are manipulating is visible on the camera's. A good rule of thumb is you should be able to do the task yourself by only looking at the camera images.
+
+In the following sections, you’ll train your neural network. After achieving reliable grasping performance, you can start introducing more variations during data collection, such as additional grasp locations, different grasping techniques, and altering camera positions.
+
+Avoid adding too much variation too quickly, as it may hinder your results.
+
+If you want to dive deeper into this important topic, you can check out the [blog post](https://huggingface.co/blog/lerobot-datasets#what-makes-a-good-dataset) we wrote on what makes a good dataset.
+
+
+#### Troubleshooting:
+- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
+
+## Visualize a dataset
+
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+```bash
+echo ${HF_USER}/so101_test
+```
+
+If you didn't upload with `--control.push_to_hub=false`, you can visualize it locally with (via a window in the browser `http://127.0.0.1:9090` with the visualization tool):
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/so101_test \
+  --local-files-only 1
+```
+
+This will launch a local web server that looks like this:
+<div style="text-align:center;">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/visualize_dataset_html.webp?raw=true" alt="Koch v1.1 leader and follower arms" title="Koch v1.1 leader and follower arms" width="100%"></img>
+</div>
+
+## Replay an episode
+
+A useful feature is the `replay` function, which allows you to replay any episode that you've recorded or episodes from any dataset out there. This function helps you test the repeatability of your robot's actions and assess transferability across robots of the same model.
+
+You can replay the first episode on your robot with:
+```bash
+python -m lerobot.replay \
+    --robot.type=so101_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.id=black \
+    --dataset.repo_id=aliberts/record-test \
+    --dataset.episode=2
+```
+
+Your robot should replicate movements similar to those you recorded. For example, check out [this video](https://x.com/RemiCadene/status/1793654950905680090) where we use `replay` on a Aloha robot from [Trossen Robotics](https://www.trossenrobotics.com).
+
+## Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+python lerobot/scripts/train.py \
+  --dataset.repo_id=${HF_USER}/so101_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_so101_test \
+  --job_name=act_so101_test \
+  --policy.device=cuda \
+  --wandb.enable=true
+```
+
+Let's explain the command:
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so101_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+
+Training should take several hours. You will find checkpoints in `outputs/train/act_so101_test/checkpoints`.
+
+To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:
+```bash
+python lerobot/scripts/train.py \
+  --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
+  --resume=true
+```
+
+#### Upload policy checkpoints
+
+Once training is done, upload the latest checkpoint with:
+```bash
+huggingface-cli upload ${HF_USER}/act_so101_test \
+  outputs/train/act_so101_test/checkpoints/last/pretrained_model
+```
+
+You can also upload intermediate checkpoints with:
+```bash
+CKPT=010000
+huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
+  outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
+```
+
+## Evaluate your policy
+
+TODO(pepijn): modify this command further
+
+You can use the `record` script from [`lerobot/record.py`](https://github.com/huggingface/lerobot/blob/main/lerobot/record.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
+```bash
+python -m lerobot.record \
+    --robot.type=so101_follower \
+    --robot.port=/dev/tty.usbmodem585A0076841 \
+    --robot.id=my_red_robot_arm \
+    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=my_blue_leader_arm \
+    --display_data=true \
+    --dataset.repo_id=aliberts/eval_act \
+    --policy.checkpoint_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model \
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_so101_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so101_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so101_test`).
--- a/docs/source/index.mdx
+++ b/docs/source/index.mdx
@@ -0,0 +1,19 @@
+<div class="flex justify-center">
+  <a target="_blank" href="https://huggingface.co/lerobot">
+      <img alt="HuggingFace Expert Acceleration Program" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-logo-thumbnail.png" style="width: 100%"></img>
+  </a>
+</div>
+
+# LeRobot
+
+**State-of-the-art machine learning for real-world robotics**
+
+🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier for entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
+
+🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
+
+🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulated environments so that everyone can get started.
+
+🤗 LeRobot hosts pretrained models and datasets on the LeRobot HuggingFace page.
+
+Join the LeRobot community on [Discord](https://discord.gg/s3KuuzsPFb)
--- a/docs/source/installation.mdx
+++ b/docs/source/installation.mdx
@@ -0,0 +1,100 @@
+# Installation
+
+## Install LeRobot
+
+Download our source code:
+```bash
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+```
+
+Create a virtual environment with Python 3.10, using [`Miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install)
+```bash
+conda create -y -n lerobot python=3.10
+```
+
+Now restart the shell by running:
+<hfoptions id="shell_restart">
+<hfoption id="Windows">
+
+```bash
+source ~/.bashrc
+```
+</hfoption>
+<hfoption id="Mac">
+
+```bash
+source ~/.bash_profile
+```
+</hfoption>
+<hfoption id="zshell">
+
+```bash
+source ~/.zshrc
+```
+</hfoption>
+</hfoptions>
+
+Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
+```bash
+conda activate lerobot
+```
+
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+> [!TIP]
+> This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
+>  - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
+>  ```bash
+>  conda install ffmpeg=7.1.1 -c conda-forge
+>  ```
+>  - _[On Linux only]_ 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`.
+
+Install 🤗 LeRobot:
+```bash
+cd lerobot && pip install
+```
+
+## Troubleshooting
+If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
+To install these for linux run:
+```bash
+sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config
+```
+For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
+
+## Sim
+For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
+- [aloha](https://github.com/huggingface/gym-aloha)
+- [xarm](https://github.com/huggingface/gym-xarm)
+- [pusht](https://github.com/huggingface/gym-pusht)
+
+For instance, to install 🤗 LeRobot with aloha and pusht, use:
+```bash
+pip install -e ".[aloha, pusht]"
+```
+
+## Motor SDK
+For Koch v1.1 install the Dynamixel SDK, for SO100/SO101/Moss install the Feetech SDK.
+
+<hfoptions id="install motors">
+<hfoption id="Feetech">
+```bash
+pip install -e ".[feetech]"
+```
+</hfoption>
+<hfoption id="Dynamixel">
+```bash
+pip install -e ".[dynamixel]"
+```
+</hfoption>
+</hfoptions>
+
+## W&B
+To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
+```bash
+wandb login
+```
--- a/docs/source/koch.md
+++ b/docs/source/koch.md
@@ -0,0 +1 @@
+../../lerobot/common/robots/koch_follower/koch.md
--- a/docs/source/lekiwi.md
+++ b/docs/source/lekiwi.md
@@ -0,0 +1 @@
+../../lerobot/common/robots/lekiwi/lekiwi.md
--- a/docs/source/so100.md
+++ b/docs/source/so100.md
@@ -0,0 +1 @@
+../../lerobot/common/robots/so100_follower/so100.md
--- a/docs/source/so101.md
+++ b/docs/source/so101.md
@@ -0,0 +1 @@
+../../lerobot/common/robots/so101_follower/so101.md
--- a/examples/10_use_so100.md
+++ b/examples/10_use_so100.md
@@ -1,615 +0,0 @@
-# Using the [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot
-
-## Table of Contents
-
-  - [A. Source the parts](#a-source-the-parts)
-  - [B. Install LeRobot](#b-install-lerobot)
-  - [C. Configure the Motors](#c-configure-the-motors)
-  - [D. Step-by-Step Assembly Instructions](#d-step-by-step-assembly-instructions)
-  - [E. Calibrate](#e-calibrate)
-  - [F. Teleoperate](#f-teleoperate)
-  - [G. Record a dataset](#g-record-a-dataset)
-  - [H. Visualize a dataset](#h-visualize-a-dataset)
-  - [I. Replay an episode](#i-replay-an-episode)
-  - [J. Train a policy](#j-train-a-policy)
-  - [K. Evaluate your policy](#k-evaluate-your-policy)
-  - [L. More Information](#l-more-information)
-
-## A. Source the parts
-
-Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts,
-and advice if it's your first time printing or if you don't own a 3D printer.
-
-Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
-
-## B. Install LeRobot
-
-> [!TIP]
-> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
-
-On your computer:
-
-#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
-
-#### 2. Restart shell
-Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
-
-#### 3. Create and activate a fresh conda environment for lerobot
-
-<details>
-<summary><strong>Video install instructions</strong></summary>
-
-<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
-
-</details>
-
-```bash
-conda create -y -n lerobot python=3.10
-```
-
-Then activate your conda environment (do this each time you open a shell to use lerobot!):
-```bash
-conda activate lerobot
-```
-
-#### 4. Clone LeRobot:
-```bash
-git clone https://github.com/huggingface/lerobot.git ~/lerobot
-```
-
-#### 5. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install --no-binary=av -e ".[feetech]"
-```
-
-Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms :robot:.
-Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
-
-## C. Configure the motors
-
-> [!NOTE]
-> Throughout this tutorial you will find videos on how to do the steps, the full video tutorial can be found here: [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I).
-
-### 1. Find the USB ports associated to each arm
-
-Designate one bus servo adapter and 6 motors for your leader arm, and similarly the other bus servo adapter and 6 motors for the follower arm. It's convenient to label them and write on each motor if it's for the follower `F` or for the leader `L` and it's ID from 1 to 6 (F1...F6 and L1...L6).
-
-#### a. Run the script to find port
-
-<details>
-<summary><strong>Video finding port</strong></summary>
-  <video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
-  <video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
-</details>
-
-To find the port for each bus servo adapter, run the utility script:
-```bash
-python lerobot/scripts/find_motors_bus_port.py
-```
-
-#### b. Example outputs
-
-Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
-```
-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 leader arm and press Enter...]
-
-The port of this MotorsBus is /dev/tty.usbmodem575E0031751
-Reconnect the usb cable.
-```
-Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
-```
-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 follower arm and press Enter...]
-
-The port of this MotorsBus is /dev/tty.usbmodem575E0032081
-Reconnect the usb cable.
-```
-
-#### c. Troubleshooting
-On Linux, you might need to give access to the USB ports by running:
-```bash
-sudo chmod 666 /dev/ttyACM0
-sudo chmod 666 /dev/ttyACM1
-```
-
-#### d. Update config file
-
-IMPORTANTLY: Now that you have your ports, update the **port** default values of [`SO100RobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
-```python
-@RobotConfig.register_subclass("so100")
-@dataclass
-class So100RobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/so100"
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-```
-
-### 2. Assembling the Base
-Let's begin with assembling the follower arm base
-
-#### a. Set IDs for all 12 motors
-
-<details>
-<summary><strong>Video configuring motor</strong></summary>
-  <video src="https://github.com/user-attachments/assets/ef9b3317-2e11-4858-b9d3-f0a02fb48ecf"></video>
-  <video src="https://github.com/user-attachments/assets/f36b5ed5-c803-4ebe-8947-b39278776a0d"></video>
-</details>
-
-Plug your first motor F1 and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate. Replace the text after --port to the corresponding follower control board port and run this command in cmd:
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 1
-```
-
-> [!NOTE]
-> These motors are currently limited. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
-
-Then unplug your motor and plug the second motor and set its ID to 2.
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 2
-```
-
-Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
-
-
-#### b. Remove the gears of the 6 leader motors
-
-<details>
-<summary><strong>Video removing gears</strong></summary>
-
-<video src="https://github.com/user-attachments/assets/0c95b88c-5b85-413d-ba19-aee2f864f2a7"></video>
-
-</details>
-
-
-Follow the video for removing gears. You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
-
-## D. Step-by-Step Assembly Instructions
-
-**Step 1: Clean Parts**
- Remove all support material from the 3D-printed parts.
---
-
-### Additional Guidance
-
-<details>
-<summary><strong>Video assembling arms</strong></summary>
-
-<video src="https://github.com/user-attachments/assets/488a39de-0189-4461-9de3-05b015f90cca"></video>
-
-</details>
-
-**Note:**
-This video provides visual guidance for assembling the arms, but it doesn't specify when or how to do the wiring. Inserting the cables beforehand is much easier than doing it afterward. The first arm may take a bit more than 1 hour to assemble, but once you get used to it, you can assemble the second arm in under 1 hour.
-
---
-
-### First Motor
-
-**Step 2: Insert Wires**
- Insert two wires into the first motor.
-
-  <img src="../media/tutorial/img1.jpg" style="height:300px;">
-
-**Step 3: Install in Base**
- Place the first motor into the base.
-
-  <img src="../media/tutorial/img2.jpg" style="height:300px;">
-
-**Step 4: Secure Motor**
- Fasten the motor with 4 screws. Two from the bottom and two from top.
-
-**Step 5: Attach Motor Holder**
- Slide over the first motor holder and fasten it using two screws (one on each side).
-
-  <img src="../media/tutorial/img4.jpg" style="height:300px;">
-
-**Step 6: Attach Motor Horns**
- Install both motor horns, securing the top horn with a screw. Try not to move the motor position when attaching the motor horn, especially for the leader arms, where we removed the gears.
-
-  <img src="../media/tutorial/img5.jpg" style="height:300px;">
-<details>
-  <summary><strong>Video adding motor horn</strong></summary>
-  <video src="https://github.com/user-attachments/assets/ef3391a4-ad05-4100-b2bd-1699bf86c969"></video>
-</details>
-
-**Step 7: Attach Shoulder Part**
- Route one wire to the back of the robot and the other to the left or in photo towards you (see photo).
- Attach the shoulder part.
-
-  <img src="../media/tutorial/img6.jpg" style="height:300px;">
-
-**Step 8: Secure Shoulder**
- Tighten the shoulder part with 4 screws on top and 4 on the bottom
-*(access bottom holes by turning the shoulder).*
-
---
-
-### Second Motor Assembly
-
-**Step 9: Install Motor 2**
- Slide the second motor in from the top and link the wire from motor 1 to motor 2.
-
-  <img src="../media/tutorial/img8.jpg" style="height:300px;">
-
-**Step 10: Attach Shoulder Holder**
- Add the shoulder motor holder.
- Ensure the wire from motor 1 to motor 2 goes behind the holder while the other wire is routed upward (see photo).
- This part can be tight to assemble, you can use a workbench like the image or a similar setup to push the part around the motor.
-
-  <div style="display: flex;">
-    <img src="../media/tutorial/img9.jpg" style="height:250px;">
-    <img src="../media/tutorial/img10.jpg" style="height:250px;">
-    <img src="../media/tutorial/img12.jpg" style="height:250px;">
-  </div>
-
-**Step 11: Secure Motor 2**
- Fasten the second motor with 4 screws.
-
-**Step 12: Attach Motor Horn**
- Attach both motor horns to motor 2, again use the horn screw.
-
-**Step 13: Attach Base**
- Install the base attachment using 2 screws.
-
-  <img src="../media/tutorial/img11.jpg" style="height:300px;">
-
-**Step 14: Attach Upper Arm**
- Attach the upper arm with 4 screws on each side.
-
-  <img src="../media/tutorial/img13.jpg" style="height:300px;">
-
---
-
-### Third Motor Assembly
-
-**Step 15: Install Motor 3**
- Route the motor cable from motor 2 through the cable holder to motor 3, then secure motor 3 with 4 screws.
-
-**Step 16: Attach Motor Horn**
- Attach both motor horns to motor 3 and secure one again with a horn screw.
-
-  <img src="../media/tutorial/img14.jpg" style="height:300px;">
-
-**Step 17: Attach Forearm**
- Connect the forearm to motor 3 using 4 screws on each side.
-
-  <img src="../media/tutorial/img15.jpg" style="height:300px;">
-
---
-
-### Fourth Motor Assembly
-
-**Step 18: Install Motor 4**
- Slide in motor 4, attach the cable from motor 3, and secure the cable in its holder with a screw.
-
-  <div style="display: flex;">
-    <img src="../media/tutorial/img16.jpg" style="height:300px;">
-    <img src="../media/tutorial/img19.jpg" style="height:300px;">
-  </div>
-
-**Step 19: Attach Motor Holder 4**
- Install the fourth motor holder (a tight fit). Ensure one wire is routed upward and the wire from motor 3 is routed downward (see photo).
-
-  <img src="../media/tutorial/img17.jpg" style="height:300px;">
-
-**Step 20: Secure Motor 4 & Attach Horn**
- Fasten motor 4 with 4 screws and attach its motor horns, use for one a horn screw.
-
-  <img src="../media/tutorial/img18.jpg" style="height:300px;">
-
---
-
-### Wrist Assembly
-
-**Step 21: Install Motor 5**
- Insert motor 5 into the wrist holder and secure it with 2 front screws.
-
-  <img src="../media/tutorial/img20.jpg" style="height:300px;">
-
-**Step 22: Attach Wrist**
- Connect the wire from motor 4 to motor 5. And already insert the other wire for the gripper.
- Secure the wrist to motor 4 using 4 screws on both sides.
-
-  <img src="../media/tutorial/img22.jpg" style="height:300px;">
-
-**Step 23: Attach Wrist Horn**
- Install only one motor horn on the wrist motor and secure it with a horn screw.
-
-  <img src="../media/tutorial/img23.jpg" style="height:300px;">
-
---
-
-### Follower Configuration
-
-**Step 24: Attach Gripper**
- Attach the gripper to motor 5.
-
-  <img src="../media/tutorial/img24.jpg" style="height:300px;">
-
-**Step 25: Install Gripper Motor**
- Insert the gripper motor, connect the motor wire from motor 5 to motor 6, and secure it with 3 screws on each side.
-
-  <img src="../media/tutorial/img25.jpg" style="height:300px;">
-
-**Step 26: Attach Gripper Horn & Claw**
- Attach the motor horns and again use a horn screw.
- Install the gripper claw and secure it with 4 screws on both sides.
-
-  <img src="../media/tutorial/img26.jpg" style="height:300px;">
-
-**Step 27: Mount Controller**
- Attach the motor controller on the back.
-
-  <div style="display: flex;">
-    <img src="../media/tutorial/img27.jpg" style="height:300px;">
-    <img src="../media/tutorial/img28.jpg" style="height:300px;">
-  </div>
-
-*Assembly complete – proceed to Leader arm assembly.*
-
---
-
-### Leader Configuration
-
-For the leader configuration, perform **Steps 1–23**. Make sure that you removed the motor gears from the motors.
-
-**Step 24: Attach Leader Holder**
- Mount the leader holder onto the wrist and secure it with a screw.
-
-  <img src="../media/tutorial/img29.jpg" style="height:300px;">
-
-**Step 25: Attach Handle**
- Attach the handle to motor 5 using 4 screws.
-
-  <img src="../media/tutorial/img30.jpg" style="height:300px;">
-
-**Step 26: Install Gripper Motor**
- Insert the gripper motor, secure it with 3 screws on each side, attach a motor horn using a horn screw, and connect the motor wire.
-
-  <img src="../media/tutorial/img31.jpg" style="height:300px;">
-
-**Step 27: Attach Trigger**
- Attach the follower trigger with 4 screws.
-
-  <img src="../media/tutorial/img32.jpg" style="height:300px;">
-
-**Step 28: Mount Controller**
- Attach the motor controller on the back.
-
-  <div style="display: flex;">
-    <img src="../media/tutorial/img27.jpg" style="height:300px;">
-    <img src="../media/tutorial/img28.jpg" style="height:300px;">
-  </div>
-
-*Assembly complete – proceed to calibration.*
-
-
-## E. Calibrate
-
-Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one SO-100 robot to work on another.
-
-#### a. Manual calibration of follower arm
-
-> [!IMPORTANT]
-> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
-
-You will need to move the follower arm to these positions sequentially:
-
-| 1. Zero position                                                                                                                                             | 2. Rotated position                                                                                                                                                   | 3. Rest position                                                                                                                                             |
-| ------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="../media/so100/follower_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/so100/follower_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/so100/follower_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
-
-Make sure both arms are connected and run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_follower"]'
-```
-
-#### b. Manual calibration of leader arm
-Follow step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
-
-| 1. Zero position                                                                                                                                       | 2. Rotated position                                                                                                                                             | 3. Rest position                                                                                                                                       |
-| ------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
-
-Run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_leader"]'
-```
-
-## F. Teleoperate
-
-**Simple teleop**
-Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --robot.cameras='{}' \
-  --control.type=teleoperate
-```
-
-
-#### a. Teleop with displaying cameras
-Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --control.type=teleoperate
-```
-
-## G. Record a dataset
-
-Once you're familiar with teleoperation, you can record your first dataset with SO-100.
-
-If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
-```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
-```
-
-Store your Hugging Face repository name in a variable to run these commands:
-```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
-echo $HF_USER
-```
-
-Record 2 episodes and upload your dataset to the hub:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/so100_test \
-  --control.tags='["so100","tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.push_to_hub=true
-```
-
-Note: You can resume recording by adding `--control.resume=true`.
-
-## H. Visualize a dataset
-
-If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
-```bash
-echo ${HF_USER}/so100_test
-```
-
-If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with (a window can be opened in the browser `http://127.0.0.1:9090` with the visualization tool):
-```bash
-python lerobot/scripts/visualize_dataset_html.py \
-  --repo-id ${HF_USER}/so100_test \
-  --local-files-only 1
-```
-
-## I. Replay an episode
-
-Now try to replay the first episode on your robot:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/so100_test \
-  --control.episode=0
-```
-
-## J. Train a policy
-
-To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-```bash
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/so100_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_so100_test \
-  --job_name=act_so100_test \
-  --policy.device=cuda \
-  --wandb.enable=true
-```
-
-Let's explain it:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so100_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
-5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-
-Training should take several hours. You will find checkpoints in `outputs/train/act_so100_test/checkpoints`.
-
-To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so100_test` policy:
-```bash
-python lerobot/scripts/train.py \
-  --config_path=outputs/train/act_so100_test/checkpoints/last/pretrained_model/train_config.json \
-  --resume=true
-```
-
-## K. Evaluate your policy
-
-You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=so100 \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/eval_act_so100_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=10 \
-  --control.push_to_hub=true \
-  --control.policy.path=outputs/train/act_so100_test/checkpoints/last/pretrained_model
-```
-
-As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so100_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so100_test`).
-
-## L. More Information
-
-Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
-
-> [!TIP]
->  If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb) in the channel [`#so100-arm`](https://discord.com/channels/1216765309076115607/1237741463832363039).
--- a/examples/11_use_lekiwi.md
+++ b/examples/11_use_lekiwi.md
@@ -1,579 +0,0 @@
-# Using the [LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi) Robot with LeRobot
-
-## Table of Contents
-
-  - [A. Source the parts](#a-source-the-parts)
-  - [B. Install software Pi](#b-install-software-on-pi)
-  - [C. Setup LeRobot laptop/pc](#c-install-lerobot-on-laptop)
-  - [D. Assemble the arms](#d-assembly)
-  - [E. Calibrate](#e-calibration)
-  - [F. Teleoperate](#f-teleoperate)
-  - [G. Record a dataset](#g-record-a-dataset)
-  - [H. Visualize a dataset](#h-visualize-a-dataset)
-  - [I. Replay an episode](#i-replay-an-episode)
-  - [J. Train a policy](#j-train-a-policy)
-  - [K. Evaluate your policy](#k-evaluate-your-policy)
-
-> [!TIP]
->  If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb) in the channel [`#mobile-so-100-arm`](https://discord.com/channels/1216765309076115607/1318390825528332371).
-
-## A. Source the parts
-
-Follow this [README](https://github.com/SIGRobotics-UIUC/LeKiwi). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts, and advice if it's your first time printing or if you don't own a 3D printer.
-
-Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
-
-### Wired version
-If you have the **wired** LeKiwi version you can skip the installation of the Raspberry Pi and setting up SSH. You can also run all commands directly on your PC for both the LeKiwi scripts and the leader arm scripts for teleoperating.
-
-## B. Install software on Pi
-Now we have to setup the remote PC that will run on the LeKiwi Robot. This is normally a Raspberry Pi, but can be any PC that can run on 5V and has enough usb ports (2 or more) for the cameras and motor control board.
-
-### Install OS
-For setting up the Raspberry Pi and its SD-card see: [Setup PI](https://www.raspberrypi.com/documentation/computers/getting-started.html). Here is explained how to download the [Imager](https://www.raspberrypi.com/software/) to install Raspberry Pi OS or Ubuntu.
-
-### Setup SSH
-After setting up your Pi, you should enable and setup [SSH](https://www.raspberrypi.com/news/coding-on-raspberry-pi-remotely-with-visual-studio-code/) (Secure Shell Protocol) so you can login into the Pi from your laptop without requiring a screen, keyboard and mouse in the Pi. A great tutorial on how to do this can be found [here](https://www.raspberrypi.com/documentation/computers/remote-access.html#ssh). Logging into your Pi can be done in your Command Prompt (cmd) or if you use VSCode you can use [this](https://marketplace.visualstudio.com/items?itemName=ms-vscode-remote.remote-ssh) extension.
-
-### Install LeRobot
-
-On your Raspberry Pi:
-
-#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
-
-#### 2. Restart shell
-Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
-
-#### 3. Create and activate a fresh conda environment for lerobot
-
-<details>
-<summary><strong>Video install instructions</strong></summary>
-
-<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
-
-</details>
-
-```bash
-conda create -y -n lerobot python=3.10
-```
-
-Then activate your conda environment (do this each time you open a shell to use lerobot!):
-```bash
-conda activate lerobot
-```
-
-#### 4. Clone LeRobot:
-```bash
-git clone https://github.com/huggingface/lerobot.git ~/lerobot
-```
-
-#### 5. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install --no-binary=av -e ".[feetech]"
-```
-
-## C. Install LeRobot on laptop
-If you already have install LeRobot on your laptop you can skip this step, otherwise please follow along as we do the same steps we did on the Pi.
-
-> [!TIP]
-> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
-
-On your computer:
-
-#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
-
-#### 2. Restart shell
-Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
-
-#### 3. Create and activate a fresh conda environment for lerobot
-
-<details>
-<summary><strong>Video install instructions</strong></summary>
-
-<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
-
-</details>
-
-```bash
-conda create -y -n lerobot python=3.10
-```
-
-Then activate your conda environment (do this each time you open a shell to use lerobot!):
-```bash
-conda activate lerobot
-```
-
-#### 4. Clone LeRobot:
-```bash
-git clone https://github.com/huggingface/lerobot.git ~/lerobot
-```
-
-#### 5. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install --no-binary=av -e ".[feetech]"
-```
-
-Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms and Mobile base :robot:.
-Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
-
-# D. Assembly
-
-First we will assemble the two SO100 arms. One to attach to the mobile base and one for teleoperation. Then we will assemble the mobile base.
-
-## SO100 Arms
-### Configure motors
-The instructions for configuring the motors can be found [Here](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#c-configure-the-motors) in step C of the SO100 tutorial. Besides the ID's for the arm motors we also need to set the motor ID's for the mobile base. These needs to be in a specific order to work. Below an image of the motor ID's and motor mounting positions for the mobile base. Note that we only use one Motor Control board on LeKiwi. This means the motor ID's for the wheels are 7, 8 and 9.
-
-<img src="../media/lekiwi/motor_ids.webp?raw=true" alt="Motor ID's for mobile robot" title="Motor ID's for mobile robot" width="60%">
-
-### Assemble arms
-[Assemble arms instruction](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#d-assemble-the-arms)
-
-## Mobile base (LeKiwi)
-[Assemble LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi)
-
-### Update config
-Both config files on the LeKiwi LeRobot and on the laptop should be the same. First we should find the Ip address of the Raspberry Pi of the mobile manipulator. This is the same Ip address used in SSH. We also need the usb port of the control board of the leader arm on the laptop and the port of the control board on LeKiwi. We can find these ports with the following script.
-
-#### a. Run the script to find port
-
-<details>
-<summary><strong>Video finding port</strong></summary>
-  <video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
-  <video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
-</details>
-
-To find the port for each bus servo adapter, run the utility script:
-```bash
-python lerobot/scripts/find_motors_bus_port.py
-```
-
-#### b. Example outputs
-
-Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
-
-[...Disconnect leader arm and press Enter...]
-
-The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
-Reconnect the usb cable.
-```
-Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
-
-[...Disconnect follower arm and press Enter...]
-
-The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
-Reconnect the usb cable.
-```
-
-#### c. Troubleshooting
-On Linux, you might need to give access to the USB ports by running:
-```bash
-sudo chmod 666 /dev/ttyACM0
-sudo chmod 666 /dev/ttyACM1
-```
-
-#### d. Update config file
-
-IMPORTANTLY: Now that you have your ports of leader and follower arm and ip address of the mobile-so100, update the **ip** in Network configuration, **port** in leader_arms and **port** in lekiwi. In the [`LeKiwiRobotConfig`](../lerobot/common/robot_devices/robots/configs.py) file. Where you will find something like:
-```python
-@RobotConfig.register_subclass("lekiwi")
-@dataclass
-class LeKiwiRobotConfig(RobotConfig):
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    # Network Configuration
-    ip: str = "172.17.133.91"
-    port: int = 5555
-    video_port: int = 5556
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "mobile": OpenCVCameraConfig(camera_index="/dev/video0", fps=30, width=640, height=480),
-            "mobile2": OpenCVCameraConfig(camera_index="/dev/video2", fps=30, width=640, height=480),
-        }
-    )
-
-    calibration_dir: str = ".cache/calibration/lekiwi"
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0077581",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/ttyACM0",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                    "left_wheel": (7, "sts3215"),
-                    "back_wheel": (8, "sts3215"),
-                    "right_wheel": (9, "sts3215"),
-                },
-            ),
-        }
-    )
-
-    teleop_keys: dict[str, str] = field(
-        default_factory=lambda: {
-            # Movement
-            "forward": "w",
-            "backward": "s",
-            "left": "a",
-            "right": "d",
-            "rotate_left": "z",
-            "rotate_right": "x",
-            # Speed control
-            "speed_up": "r",
-            "speed_down": "f",
-            # quit teleop
-            "quit": "q",
-        }
-    )
-
-    mock: bool = False
-```
-
-## Wired version
-
-For the wired LeKiwi version your configured IP address should refer to your own laptop (127.0.0.1), because leader arm and LeKiwi are in this case connected to own laptop. Below and example configuration for this wired setup:
-```python
-@RobotConfig.register_subclass("lekiwi")
-@dataclass
-class LeKiwiRobotConfig(RobotConfig):
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    # Network Configuration
-    ip: str = "127.0.0.1"
-    port: int = 5555
-    video_port: int = 5556
-
-    cameras: dict[str, CameraConfig] = field(
-        default_factory=lambda: {
-            "front": OpenCVCameraConfig(
-                camera_index=0, fps=30, width=640, height=480, rotation=90
-            ),
-            "wrist": OpenCVCameraConfig(
-                camera_index=1, fps=30, width=640, height=480, rotation=180
-            ),
-        }
-    )
-
-    calibration_dir: str = ".cache/calibration/lekiwi"
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0077581",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431061",
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                    "left_wheel": (7, "sts3215"),
-                    "back_wheel": (8, "sts3215"),
-                    "right_wheel": (9, "sts3215"),
-                },
-            ),
-        }
-    )
-
-    teleop_keys: dict[str, str] = field(
-        default_factory=lambda: {
-            # Movement
-            "forward": "w",
-            "backward": "s",
-            "left": "a",
-            "right": "d",
-            "rotate_left": "z",
-            "rotate_right": "x",
-            # Speed control
-            "speed_up": "r",
-            "speed_down": "f",
-            # quit teleop
-            "quit": "q",
-        }
-    )
-
-    mock: bool = False
-```
-
-# E. Calibration
-Now we have to calibrate the leader arm and the follower arm. The wheel motors don't have to be calibrated.
-
-
-### Calibrate follower arm (on mobile base)
-> [!IMPORTANT]
-> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
-
-You will need to move the follower arm to these positions sequentially:
-
-| 1. Zero position                                                                                                                                                  | 2. Rotated position                                                                                                                                                        | 3. Rest position                                                                                                                                                  |
-| ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| <img src="../media/lekiwi/mobile_calib_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/lekiwi/mobile_calib_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/lekiwi/mobile_calib_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
-
-Make sure the arm is connected to the Raspberry Pi and run this script (on the Raspberry Pi) to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=lekiwi \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_follower"]'
-```
-
-### Wired version
-If you have the **wired** LeKiwi version please run all commands including this calibration command on your laptop.
-
-### Calibrate leader arm
-Then to calibrate the leader arm (which is attached to the laptop/pc). You will need to move the leader arm to these positions sequentially:
-
-| 1. Zero position                                                                                                                                       | 2. Rotated position                                                                                                                                             | 3. Rest position                                                                                                                                       |
-| ------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
-
-Run this script (on your laptop/pc) to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=lekiwi \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_leader"]'
-```
-
-# F. Teleoperate
-To teleoperate SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=lekiwi \
-  --control.type=remote_robot
-```
-
-Then on your laptop, also run `conda activate lerobot` and this script:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=lekiwi \
-  --control.type=teleoperate \
-  --control.fps=30
-```
-
-You should see on your laptop something like this: ```[INFO] Connected to remote robot at tcp://172.17.133.91:5555 and video stream at tcp://172.17.133.91:5556.``` Now you can move the leader arm and use the keyboard (w,a,s,d) to drive forward, left, backwards, right. And use (z,x) to turn left or turn right. You can use (r,f) to increase and decrease the speed of the mobile robot. There are three speed modes, see the table below:
-| Speed Mode | Linear Speed (m/s) | Rotation Speed (deg/s) |
-| ---------- | ------------------ | ---------------------- |
-| Fast       | 0.4                | 90                     |
-| Medium     | 0.25               | 60                     |
-| Slow       | 0.1                | 30                     |
-
-
-| Key | Action         |
-| --- | -------------- |
-| W   | Move forward   |
-| A   | Move left      |
-| S   | Move backward  |
-| D   | Move right     |
-| Z   | Turn left      |
-| X   | Turn right     |
-| R   | Increase speed |
-| F   | Decrease speed |
-
-> [!TIP]
->  If you use a different keyboard you can change the keys for each command in the [`LeKiwiRobotConfig`](../lerobot/common/robot_devices/robots/configs.py).
-
-### Wired version
-If you have the **wired** LeKiwi version please run all commands including both these teleoperation commands on your laptop.
-
-## Troubleshoot communication
-
-If you are having trouble connecting to the Mobile SO100, follow these steps to diagnose and resolve the issue.
-
-### 1. Verify IP Address Configuration
-Make sure that the correct ip for the Pi is set in the configuration file. To check the Raspberry Pi's IP address, run (on the Pi command line):
-```bash
-hostname -I
-```
-
-### 2. Check if Pi is reachable from laptop/pc
-Try pinging the Raspberry Pi from your laptop:
-```bach
-ping <your_pi_ip_address>
-```
-
-If the ping fails:
- Ensure the Pi is powered on and connected to the same network.
- Check if SSH is enabled on the Pi.
-
-### 3. Try SSH connection
-If you can't SSH into the Pi, it might not be properly connected. Use:
-```bash
-ssh <your_pi_user_name>@<your_pi_ip_address>
-```
-If you get a connection error:
- Ensure SSH is enabled on the Pi by running:
-  ```bash
-  sudo raspi-config
-  ```
-  Then navigate to: **Interfacing Options -> SSH** and enable it.
-
-### 4. Same config file
-Make sure the configuration file on both your laptop/pc and the Raspberry Pi is the same.
-
-# G. Record a dataset
-Once you're familiar with teleoperation, you can record your first dataset with LeKiwi.
-
-To start the program on LeKiwi, SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=lekiwi \
-  --control.type=remote_robot
-```
-
-If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
-```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
-```
-
-Store your Hugging Face repository name in a variable to run these commands:
-```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
-echo $HF_USER
-```
-On your laptop then run this command to record 2 episodes and upload your dataset to the hub:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=lekiwi \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/lekiwi_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.push_to_hub=true
-```
-
-Note: You can resume recording by adding `--control.resume=true`.
-
-### Wired version
-If you have the **wired** LeKiwi version please run all commands including both these record dataset commands on your laptop.
-
-# H. Visualize a dataset
-
-If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
-```bash
-echo ${HF_USER}/lekiwi_test
-```
-
-If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with (a window can be opened in the browser `http://127.0.0.1:9090` with the visualization tool):
-```bash
-python lerobot/scripts/visualize_dataset_html.py \
-  --repo-id ${HF_USER}/lekiwi_test \
-  --local-files-only 1
-```
-
-# I. Replay an episode
-Now try to replay the first episode on your robot:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=lekiwi \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/lekiwi_test \
-  --control.episode=0
-```
-
-## J. Train a policy
-
-To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-```bash
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/lekiwi_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_lekiwi_test \
-  --job_name=act_lekiwi_test \
-  --policy.device=cuda \
-  --wandb.enable=true
-```
-
-Let's explain it:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/lekiwi_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
-5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-
-Training should take several hours. You will find checkpoints in `outputs/train/act_lekiwi_test/checkpoints`.
-
-## K. Evaluate your policy
-
-You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=lekiwi \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Drive to the red block and pick it up" \
-  --control.repo_id=${HF_USER}/eval_act_lekiwi_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=10 \
-  --control.push_to_hub=true \
-  --control.policy.path=outputs/train/act_lekiwi_test/checkpoints/last/pretrained_model
-```
-
-As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_lekiwi_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_lekiwi_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_lekiwi_test`).
--- a/examples/11_use_moss.md
+++ b/examples/11_use_moss.md
@@ -1,328 +0,0 @@
-This tutorial explains how to use [Moss v1](https://github.com/jess-moss/moss-robot-arms) with LeRobot.
-
-## Source the parts
-
-Follow this [README](https://github.com/jess-moss/moss-robot-arms). It contains the bill of materials with link to source the parts, as well as the instructions to 3D print the parts and advice if it's your first time printing or if you don't own a 3D printer already.
-
-**Important**: Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
-
-## Install LeRobot
-
-On your computer:
-
-1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
-```bash
-mkdir -p ~/miniconda3
-wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
-bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
-rm ~/miniconda3/miniconda.sh
-~/miniconda3/bin/conda init bash
-```
-
-2. Restart shell or `source ~/.bashrc`
-
-3. Create and activate a fresh conda environment for lerobot
-```bash
-conda create -y -n lerobot python=3.10 && conda activate lerobot
-```
-
-4. Clone LeRobot:
-```bash
-git clone https://github.com/huggingface/lerobot.git ~/lerobot
-```
-
-5. Install LeRobot with dependencies for the feetech motors:
-```bash
-cd ~/lerobot && pip install --no-binary=av -e ".[feetech]"
-```
-
-## Configure the motors
-
-Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the use of our scripts below.
-
-**Find USB ports associated to your arms**
-To find the correct ports for each arm, run the utility script twice:
-```bash
-python lerobot/scripts/find_motors_bus_port.py
-```
-
-Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
-
-[...Disconnect leader arm and press Enter...]
-
-The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
-Reconnect the usb cable.
-```
-
-Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
-```
-Finding all available ports for the MotorBus.
-['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
-Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
-
-[...Disconnect follower arm and press Enter...]
-
-The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
-Reconnect the usb cable.
-```
-
-Troubleshooting: On Linux, you might need to give access to the USB ports by running:
-```bash
-sudo chmod 666 /dev/ttyACM0
-sudo chmod 666 /dev/ttyACM1
-```
-
-#### Update config file
-
-IMPORTANTLY: Now that you have your ports, update the **port** default values of [`MossRobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
-```python
-@RobotConfig.register_subclass("moss")
-@dataclass
-class MossRobotConfig(ManipulatorRobotConfig):
-    calibration_dir: str = ".cache/calibration/moss"
-    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
-    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
-    # the number of motors in your follower arms.
-    max_relative_target: int | None = None
-
-    leader_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-
-    follower_arms: dict[str, MotorsBusConfig] = field(
-        default_factory=lambda: {
-            "main": FeetechMotorsBusConfig(
-                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
-                motors={
-                    # name: (index, model)
-                    "shoulder_pan": [1, "sts3215"],
-                    "shoulder_lift": [2, "sts3215"],
-                    "elbow_flex": [3, "sts3215"],
-                    "wrist_flex": [4, "sts3215"],
-                    "wrist_roll": [5, "sts3215"],
-                    "gripper": [6, "sts3215"],
-                },
-            ),
-        }
-    )
-```
-
-**Configure your motors**
-Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 1
-```
-
-Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
-
-Then unplug your motor and plug the second motor and set its ID to 2.
-```bash
-python lerobot/scripts/configure_motor.py \
-  --port /dev/tty.usbmodem58760432961 \
-  --brand feetech \
-  --model sts3215 \
-  --baudrate 1000000 \
-  --ID 2
-```
-
-Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
-
-**Remove the gears of the 6 leader motors**
-Follow step 2 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
-
-**Add motor horn to the motors**
-Follow step 3 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). For Moss v1, you need to align the holes on the motor horn to the motor spline to be approximately 3, 6, 9 and 12 o'clock.
-Try to avoid rotating the motor while doing so to keep position 2048 set during configuration. It is especially tricky for the leader motors as it is more sensible without the gears, but it's ok if it's a bit rotated.
-
-## Assemble the arms
-
-Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
-
-## Calibrate
-
-Next, you'll need to calibrate your Moss v1 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one Moss v1 robot to work on another.
-
-**Manual calibration of follower arm**
-/!\ Contrarily to step 6 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
-
-You will need to move the follower arm to these positions sequentially:
-
-| 1. Zero position                                                                                                                                              | 2. Rotated position                                                                                                                                                    | 3. Rest position                                                                                                                                              |
-| ------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| <img src="../media/moss/follower_zero.webp?raw=true" alt="Moss v1 follower arm zero position" title="Moss v1 follower arm zero position" style="width:100%;"> | <img src="../media/moss/follower_rotated.webp?raw=true" alt="Moss v1 follower arm rotated position" title="Moss v1 follower arm rotated position" style="width:100%;"> | <img src="../media/moss/follower_rest.webp?raw=true" alt="Moss v1 follower arm rest position" title="Moss v1 follower arm rest position" style="width:100%;"> |
-
-Make sure both arms are connected and run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_follower"]'
-```
-
-**Manual calibration of leader arm**
-Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
-
-| 1. Zero position                                                                                                                                        | 2. Rotated position                                                                                                                                              | 3. Rest position                                                                                                                                        |
-| ------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| <img src="../media/moss/leader_zero.webp?raw=true" alt="Moss v1 leader arm zero position" title="Moss v1 leader arm zero position" style="width:100%;"> | <img src="../media/moss/leader_rotated.webp?raw=true" alt="Moss v1 leader arm rotated position" title="Moss v1 leader arm rotated position" style="width:100%;"> | <img src="../media/moss/leader_rest.webp?raw=true" alt="Moss v1 leader arm rest position" title="Moss v1 leader arm rest position" style="width:100%;"> |
-
-Run this script to launch manual calibration:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --robot.cameras='{}' \
-  --control.type=calibrate \
-  --control.arms='["main_leader"]'
-```
-
-## Teleoperate
-
-**Simple teleop**
-Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --robot.cameras='{}' \
-  --control.type=teleoperate
-```
-
-
-**Teleop with displaying cameras**
-Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --control.type=teleoperate
-```
-
-## Record a dataset
-
-Once you're familiar with teleoperation, you can record your first dataset with Moss v1.
-
-If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
-```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
-```
-
-Store your Hugging Face repository name in a variable to run these commands:
-```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
-echo $HF_USER
-```
-
-Record 2 episodes and upload your dataset to the hub:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/moss_test \
-  --control.tags='["moss","tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=2 \
-  --control.push_to_hub=true
-```
-
-Note: You can resume recording by adding `--control.resume=true`.
-
-## Visualize a dataset
-
-If you uploaded your dataset to the hub with `--control.push_to_hub=true`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
-```bash
-echo ${HF_USER}/moss_test
-```
-
-If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with:
-```bash
-python lerobot/scripts/visualize_dataset_html.py \
-  --repo-id ${HF_USER}/moss_test \
-  --local-files-only 1
-```
-
-## Replay an episode
-
-Now try to replay the first episode on your robot:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --control.type=replay \
-  --control.fps=30 \
-  --control.repo_id=${HF_USER}/moss_test \
-  --control.episode=0
-```
-
-## Train a policy
-
-To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-```bash
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/moss_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_moss_test \
-  --job_name=act_moss_test \
-  --policy.device=cuda \
-  --wandb.enable=true
-```
-
-Let's explain it:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/moss_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
-5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-
-Training should take several hours. You will find checkpoints in `outputs/train/act_moss_test/checkpoints`.
-
-## Evaluate your policy
-
-You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
-```bash
-python lerobot/scripts/control_robot.py \
-  --robot.type=moss \
-  --control.type=record \
-  --control.fps=30 \
-  --control.single_task="Grasp a lego block and put it in the bin." \
-  --control.repo_id=${HF_USER}/eval_act_moss_test \
-  --control.tags='["tutorial"]' \
-  --control.warmup_time_s=5 \
-  --control.episode_time_s=30 \
-  --control.reset_time_s=30 \
-  --control.num_episodes=10 \
-  --control.push_to_hub=true \
-  --control.policy.path=outputs/train/act_moss_test/checkpoints/last/pretrained_model
-```
-
-As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
-1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_moss_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_moss_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_moss_test`).
-
-## More
-
-Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
-
-If you have any question or need help, please reach out on Discord in the channel [`#moss-arm`](https://discord.com/channels/1216765309076115607/1275374638985252925).
--- a/examples/12_train_hilserl_classifier.md
+++ b/examples/12_train_hilserl_classifier.md
@@ -1,94 +0,0 @@
-# Training a HIL-SERL Reward Classifier with LeRobot
-
-This tutorial provides step-by-step instructions for training a reward classifier using LeRobot.
-
---
-
-## Training Script Overview
-
-LeRobot includes a ready-to-use training script located at [`lerobot/scripts/train_hilserl_classifier.py`](../../lerobot/scripts/train_hilserl_classifier.py). Here's an outline of its workflow:
-
-1. **Configuration Loading**
-   The script uses Hydra to load a configuration file for subsequent steps. (Details on Hydra follow below.)
-
-2. **Dataset Initialization**
-   It loads a `LeRobotDataset` containing images and rewards. To optimize performance, a weighted random sampler is used to balance class sampling.
-
-3. **Classifier Initialization**
-   A lightweight classification head is built on top of a frozen, pretrained image encoder from HuggingFace. The classifier outputs either:
-   - A single probability (binary classification), or
-   - Logits (multi-class classification).
-
-4. **Training Loop Execution**
-   The script performs:
-   - Forward and backward passes,
-   - Optimization steps,
-   - Periodic logging, evaluation, and checkpoint saving.
-
---
-
-## Configuring with Hydra
-
-For detailed information about Hydra usage, refer to [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md). However, note that training the reward classifier differs slightly and requires a separate configuration file.
-
-### Config File Setup
-
-The default `default.yaml` cannot launch the reward classifier training directly. Instead, you need a configuration file like [`lerobot/configs/policy/hilserl_classifier.yaml`](../../lerobot/configs/policy/hilserl_classifier.yaml), with the following adjustment:
-
-Replace the `dataset_repo_id` field with the identifier for your dataset, which contains images and sparse rewards:
-
-```yaml
-# Example: lerobot/configs/policy/reward_classifier.yaml
-dataset_repo_id: "my_dataset_repo_id"
-## Typical logs and metrics
-```
-When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
-
-After that, you will see training log like this one:
-
-```
-[2024-11-29 18:26:36,999][root][INFO] -
-Epoch 5/5
-Training:  82%|██████████████████████████████████████████████████████████████████████████████▋                 | 91/111 [00:50<00:09,  2.04it/s, loss=0.2999, acc=69.99%]
-```
-
-or evaluation log like:
-
-```
-Validation: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:20<00:00,  1.37it/s]
-```
-
-### Metrics Tracking with Weights & Biases (WandB)
-
-If `wandb.enable` is set to `true`, the training and evaluation logs will also be saved in WandB. This allows you to track key metrics in real-time, including:
-
- **Training Metrics**:
-  - `train/accuracy`
-  - `train/loss`
-  - `train/dataloading_s`
- **Evaluation Metrics**:
-  - `eval/accuracy`
-  - `eval/loss`
-  - `eval/eval_s`
-
-#### Additional Features
-
-You can also log sample predictions during evaluation. Each logged sample will include:
-
- The **input image**.
- The **predicted label**.
- The **true label**.
- The **classifier's "confidence" (logits/probability)**.
-
-These logs can be useful for diagnosing and debugging performance issues.
-
-
-#### Generate protobuf files
-
-```bash
-python -m grpc_tools.protoc \
-    -I lerobot/scripts/server \
-    --python_out=lerobot/scripts/server \
-    --grpc_python_out=lerobot/scripts/server \
-    lerobot/scripts/server/hilserl.proto
-```
--- a/examples/1_load_lerobot_dataset.py
+++ b/examples/1_load_lerobot_dataset.py
@@ -32,10 +32,7 @@ import torch
 from huggingface_hub import HfApi

 import lerobot
-from lerobot.common.datasets.lerobot_dataset import (
-    LeRobotDataset,
-    LeRobotDatasetMetadata,
-)
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata

 # We ported a number of existing datasets ourselves, use this to see the list:
 print("List of available datasets:")
@@ -122,7 +119,7 @@ print(dataset.features[camera_key]["shape"])
 delta_timestamps = {
    # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
    camera_key: [-1, -0.5, -0.20, 0],
-    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
+    # loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
    # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
    "action": [t / dataset.fps for t in range(64)],
@@ -146,6 +143,6 @@ dataloader = torch.utils.data.DataLoader(

 for batch in dataloader:
    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
-    print(f"{batch['observation.state'].shape=}")  # (32, 5, c)
+    print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
    print(f"{batch['action'].shape=}")  # (32, 64, c)
    break
--- a/examples/2_evaluate_pretrained_policy.py
+++ b/examples/2_evaluate_pretrained_policy.py
@@ -13,12 +13,12 @@
 # limitations under the License.

 """
-This scripts demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
+This script demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
 training outputs directory. In the latter case, you might want to run examples/3_train_policy.py first.

 It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
 ```bash
-pip install --no-binary=av -e ".[pusht]"`
+pip install -e ".[pusht]"
 ```
 """

@@ -119,7 +119,7 @@ while not done:
    rewards.append(reward)
    frames.append(env.render())

-    # The rollout is considered done when the success state is reach (i.e. terminated is True),
+    # The rollout is considered done when the success state is reached (i.e. terminated is True),
    # or the maximum number of iterations is reached (i.e. truncated is True)
    done = terminated | truncated | done
    step += 1
--- a/examples/3_train_policy.py
+++ b/examples/3_train_policy.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-"""This scripts demonstrates how to train Diffusion Policy on the PushT environment.
+"""This script demonstrates how to train Diffusion Policy on the PushT environment.

 Once you have trained a model with this script, you can try to evaluate it on
 examples/2_evaluate_pretrained_policy.py
@@ -22,10 +22,7 @@ from pathlib import Path

 import torch

-from lerobot.common.datasets.lerobot_dataset import (
-    LeRobotDataset,
-    LeRobotDatasetMetadata,
-)
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.common.datasets.utils import dataset_to_policy_features
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
@@ -80,24 +77,7 @@ def main():
        # Load the previous action (-0.1), the next action to be executed (0.0),
        # and 14 future actions with a 0.1 seconds spacing. All these actions will be
        # used to supervise the policy.
-        "action": [
-            -0.1,
-            0.0,
-            0.1,
-            0.2,
-            0.3,
-            0.4,
-            0.5,
-            0.6,
-            0.7,
-            0.8,
-            0.9,
-            1.0,
-            1.1,
-            1.2,
-            1.3,
-            1.4,
-        ],
+        "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
    }

    # We can then instantiate the dataset with these delta_timestamps configuration.
--- a/examples/4_train_policy_with_script.md
+++ b/examples/4_train_policy_with_script.md
@@ -1,10 +1,10 @@
 This tutorial will explain the training script, how to use it, and particularly how to configure everything needed for the training run.
-> **Note:** The following assume you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
+> **Note:** The following assumes you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.


 ## The training script

-LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
+LeRobot offers a training script at [`lerobot/scripts/train.py`](../lerobot/scripts/train.py). At a high level it does the following:

 - Initialize/load a configuration for the following steps using.
 - Instantiates a dataset.
@@ -21,9 +21,9 @@ In the training script, the main function `train` expects a `TrainPipelineConfig
 def train(cfg: TrainPipelineConfig):
 ```

-You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
+You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)

-When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated for this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
+When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated to this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)

 Let's have a look at a simplified example. Amongst other attributes, the training config has the following attributes:
 ```python
@@ -43,14 +43,14 @@ class DatasetConfig:
 ```

 This creates a hierarchical relationship where, for example assuming we have a `cfg` instance of `TrainPipelineConfig`, we can access the `repo_id` value with `cfg.dataset.repo_id`.
-From the command line, we can specify this value with using a very similar syntax `--dataset.repo_id=repo/id`.
+From the command line, we can specify this value by using a very similar syntax `--dataset.repo_id=repo/id`.

 By default, every field takes its default value specified in the dataclass. If a field doesn't have a default value, it needs to be specified either from the command line or from a config file – which path is also given in the command line (more in this below). In the example above, the `dataset` field doesn't have a default value which means it must be specified.


 ## Specifying values from the CLI

-Let's say that we want to train [Diffusion Policy](../../lerobot/common/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
+Let's say that we want to train [Diffusion Policy](../lerobot/common/policies/diffusion) on the [pusht](https://huggingface.co/datasets/lerobot/pusht) dataset, using the [gym_pusht](https://github.com/huggingface/gym-pusht) environment for evaluation. The command to do so would look like this:
 ```bash
 python lerobot/scripts/train.py \
    --dataset.repo_id=lerobot/pusht \
@@ -60,10 +60,10 @@ python lerobot/scripts/train.py \

 Let's break this down:
 - To specify the dataset, we just need to specify its `repo_id` on the hub which is the only required argument in the `DatasetConfig`. The rest of the fields have default values and in this case we are fine with those so we can just add the option `--dataset.repo_id=lerobot/pusht`.
- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../../lerobot/common/policies)
- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../../lerobot/common/envs/configs.py)
+- To specify the policy, we can just select diffusion policy using `--policy` appended with `.type`. Here, `.type` is a special argument which allows us to select config classes inheriting from `draccus.ChoiceRegistry` and that have been decorated with the `register_subclass()` method. To have a better explanation of this feature, have a look at this [Draccus demo](https://github.com/dlwh/draccus?tab=readme-ov-file#more-flexible-configuration-with-choice-types). In our code, we use this mechanism mainly to select policies, environments, robots, and some other components like optimizers. The policies available to select are located in [lerobot/common/policies](../lerobot/common/policies)
+- Similarly, we select the environment with `--env.type=pusht`. The different environment configs are available in [`lerobot/common/envs/configs.py`](../lerobot/common/envs/configs.py)

-Let's see another example. Let's say you've been training [ACT](../../lerobot/common/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
+Let's see another example. Let's say you've been training [ACT](../lerobot/common/policies/act) on [lerobot/aloha_sim_insertion_human](https://huggingface.co/datasets/lerobot/aloha_sim_insertion_human) using the [gym-aloha](https://github.com/huggingface/gym-aloha) environment for evaluation with:
 ```bash
 python lerobot/scripts/train.py \
    --policy.type=act \
@@ -74,7 +74,7 @@ python lerobot/scripts/train.py \
 > Notice we added `--output_dir` to explicitly tell where to write outputs from this run (checkpoints, training state, configs etc.). This is not mandatory and if you don't specify it, a default directory will be created from the current date and time, env.type and policy.type. This will typically look like `outputs/train/2025-01-24/16-10-05_aloha_act`.

 We now want to train a different policy for aloha on another task. We'll change the dataset and use [lerobot/aloha_sim_transfer_cube_human](https://huggingface.co/datasets/lerobot/aloha_sim_transfer_cube_human) instead. Of course, we also need to change the task of the environment as well to match this other task.
-Looking at the [`AlohaEnv`](../../lerobot/common/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
+Looking at the [`AlohaEnv`](../lerobot/common/envs/configs.py) config, the task is `"AlohaInsertion-v0"` by default, which corresponds to the task we trained on in the command above. The [gym-aloha](https://github.com/huggingface/gym-aloha?tab=readme-ov-file#description) environment also has the `AlohaTransferCube-v0` task which corresponds to this other task we want to train on. Putting this together, we can train this new policy on this different task using:
 ```bash
 python lerobot/scripts/train.py \
    --policy.type=act \
@@ -135,7 +135,7 @@ will start a training run with the same configuration used for training [lerobot

 ## Resume training

-Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to that here.
+Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to do that here.

 Let's reuse the command from the previous run and add a few more options:
 ```bash
--- a/examples/7_get_started_with_real_robot.md
+++ b/examples/7_get_started_with_real_robot.md
@@ -33,7 +33,7 @@ First, install the additional dependencies required for robots built with dynami

 Using `pip`:
 ```bash
-pip install --no-binary=av -e ".[dynamixel]"
+pip install -e ".[dynamixel]"
 ```

 Using `poetry`:
@@ -55,6 +55,9 @@ Finally, connect both arms to your computer via USB. Note that the USB doesn't p
 Now you are ready to configure your motors for the first time, as detailed in the sections below. In the upcoming sections, you'll learn about our classes and functions by running some python code in an interactive session, or by copy-pasting it in a python file.

 If you have already configured your motors the first time, you can streamline the process by directly running the teleoperate script (which is detailed further in the tutorial):
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
+
 ```bash
 python lerobot/scripts/control_robot.py \
  --robot.type=koch \
@@ -80,7 +83,7 @@ python lerobot/scripts/configure_motor.py \
  --brand dynamixel \
  --model xl330-m288 \
  --baudrate 1000000 \
-  --ID 1
+  --id 1
 ```

 Then unplug your first motor and plug the second motor and set its ID to 2.
@@ -90,7 +93,7 @@ python lerobot/scripts/configure_motor.py \
  --brand dynamixel \
  --model xl330-m288 \
  --baudrate 1000000 \
-  --ID 2
+  --id 2
 ```

 Redo the process for all your motors until ID 6.
@@ -374,7 +377,7 @@ robot = ManipulatorRobot(robot_config)

 The `KochRobotConfig` is used to set the associated settings and calibration process. For instance, we activate the torque of the gripper of the leader Koch v1.1 arm and position it at a 40 degree angle to use it as a trigger.

-For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `AlohaRobotConfig` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected for Aloha.
+For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `AlohaRobotConfig` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected for Aloha.

 **Calibrate and Connect the ManipulatorRobot**

@@ -396,7 +399,7 @@ And here are the corresponding positions for the leader arm:

 You can watch a [video tutorial of the calibration procedure](https://youtu.be/8drnU9uRY24) for more details.

-During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask yo to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to measure if the values changed negatively or positively.
+During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask you to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to measure if the values changed negatively or positively.

 Finally, the rest position ensures that the follower and leader arms are roughly aligned after calibration, preventing sudden movements that could damage the motors when starting teleoperation.

@@ -619,7 +622,7 @@ camera_01_frame_000047.png

 Note: Some cameras may take a few seconds to warm up, and the first frame might be black or green.

-Finally, run this code to instantiate and connectyour camera:
+Finally, run this code to instantiate and connect your camera:
 ```python
 from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
 from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
@@ -827,11 +830,6 @@ It contains:
 - `dtRphone:33.84 (29.5hz)` which is the delta time of capturing an image from the phone camera in the thread running asynchronously.

 Troubleshooting:
- On Linux, if you encounter any issue during video encoding with `ffmpeg: unknown encoder libsvtav1`, you can:
-  - install with conda-forge by running `conda install -c conda-forge ffmpeg` (it should be compiled with `libsvtav1`),
-  - or, install [Homebrew](https://brew.sh) and run `brew install ffmpeg` (it should be compiled with `libsvtav1`),
-  - or, 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`.
 - On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).

 At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/koch_test) that you can obtain by running:
--- a/examples/advanced/2_calculate_validation_loss.py
+++ b/examples/advanced/2_calculate_validation_loss.py
@@ -26,10 +26,7 @@ import math

 import torch

-from lerobot.common.datasets.lerobot_dataset import (
-    LeRobotDataset,
-    LeRobotDatasetMetadata,
-)
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy


@@ -54,24 +51,7 @@ def main():
        # Load the previous action (-0.1), the next action to be executed (0.0),
        # and 14 future actions with a 0.1 seconds spacing. All these actions will be
        # used to calculate the loss.
-        "action": [
-            -0.1,
-            0.0,
-            0.1,
-            0.2,
-            0.3,
-            0.4,
-            0.5,
-            0.6,
-            0.7,
-            0.8,
-            0.9,
-            1.0,
-            1.1,
-            1.2,
-            1.3,
-            1.4,
-        ],
+        "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
    }

    # Load the last 10% of episodes of the dataset as a validation set.
@@ -86,7 +66,7 @@ def main():
    print(f"Number of episodes in full dataset: {total_episodes}")
    print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
    print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
-    # - Load train an val datasets
+    # - Load train and val datasets
    train_dataset = LeRobotDataset(
        "lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps
    )
--- a/examples/robots/lekiwi_client_app.py
+++ b/examples/robots/lekiwi_client_app.py
@@ -0,0 +1,98 @@
+# 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 logging
+import time
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
+from lerobot.common.robots.lekiwi.lekiwi_client import OBS_STATE, LeKiwiClient
+from lerobot.common.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
+from lerobot.common.teleoperators.so100 import SO100Leader, SO100LeaderConfig
+
+NB_CYCLES_CLIENT_CONNECTION = 250
+
+
+def main():
+    logging.info("Configuring Teleop Devices")
+    leader_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem58760434171")
+    leader_arm = SO100Leader(leader_arm_config)
+
+    keyboard_config = KeyboardTeleopConfig()
+    keyboard = KeyboardTeleop(keyboard_config)
+
+    logging.info("Configuring LeKiwi Client")
+    robot_config = LeKiwiClientConfig(remote_ip="192.0.2.42", id="lekiwi")
+    robot = LeKiwiClient(robot_config)
+
+    logging.info("Creating LeRobot Dataset")
+
+    # The observations that we get are expected to be in body frame (x,y,theta)
+    obs_dict = {f"{OBS_STATE}." + key: value for key, value in robot.state_feature.items()}
+    # The actions that we send are expected to be in wheel frame (motor encoders)
+    act_dict = {"action." + key: value for key, value in robot.action_feature.items()}
+
+    features_dict = {
+        **act_dict,
+        **obs_dict,
+        **robot.camera_features,
+    }
+    dataset = LeRobotDataset.create(
+        repo_id="user/lekiwi" + str(int(time.time())),
+        fps=10,
+        features=features_dict,
+    )
+
+    logging.info("Connecting Teleop Devices")
+    leader_arm.connect()
+    keyboard.connect()
+
+    logging.info("Connecting remote LeKiwi")
+    robot.connect()
+
+    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
+        logging.error("Failed to connect to all devices")
+        return
+
+    logging.info("Starting LeKiwi teleoperation")
+    i = 0
+    while i < NB_CYCLES_CLIENT_CONNECTION:
+        arm_action = leader_arm.get_action()
+        base_action = keyboard.get_action()
+        action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
+
+        action_sent = robot.send_action(action)
+        observation = robot.get_observation()
+
+        frame = {**action_sent, **observation}
+        frame.update({"task": "Dummy Example Task Dataset"})
+
+        logging.info("Saved a frame into the dataset")
+        dataset.add_frame(frame)
+        i += 1
+
+    logging.info("Disconnecting Teleop Devices and LeKiwi Client")
+    robot.disconnect()
+    leader_arm.disconnect()
+    keyboard.disconnect()
+
+    logging.info("Uploading dataset to the hub")
+    dataset.save_episode()
+    dataset.push_to_hub()
+
+    logging.info("Finished LeKiwi cleanly")
+
+
+if __name__ == "__main__":
+    main()
--- a/lerobot/init.py
+++ b/lerobot/init.py
@@ -181,7 +181,7 @@ available_robots = [
    "koch_bimanual",
    "aloha",
    "so100",
-    "moss",
+    "so101",
 ]

 # lists all available cameras from `lerobot/common/robot_devices/cameras`
--- a/lerobot/calibrate.py
+++ b/lerobot/calibrate.py
@@ -0,0 +1,84 @@
+# 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.
+
+"""
+Helper to recalibrate your device (robot or teleoperator).
+
+Example:
+
+```shell
+python -m lerobot.calibrate \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=blue
+```
+"""
+
+import logging
+from dataclasses import asdict, dataclass
+from pprint import pformat
+
+import draccus
+
+from lerobot.common.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.common.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.common.robots import (  # noqa: F401
+    Robot,
+    RobotConfig,
+    koch_follower,
+    lekiwi,
+    make_robot_from_config,
+    so100_follower,
+    so101_follower,
+)
+from lerobot.common.teleoperators import (  # noqa: F401
+    Teleoperator,
+    TeleoperatorConfig,
+    koch_leader,
+    make_teleoperator_from_config,
+    so100_leader,
+    so101_leader,
+)
+from lerobot.common.utils.utils import init_logging
+
+
+@dataclass
+class CalibrateConfig:
+    teleop: TeleoperatorConfig | None = None
+    robot: RobotConfig | None = None
+
+    def __post_init__(self):
+        if bool(self.teleop) == bool(self.robot):
+            raise ValueError("Choose either a teleop or a robot.")
+
+        self.device = self.robot if self.robot else self.teleop
+
+
+@draccus.wrap()
+def calibrate(cfg: CalibrateConfig):
+    init_logging()
+    logging.info(pformat(asdict(cfg)))
+
+    if isinstance(cfg.device, RobotConfig):
+        device = make_robot_from_config(cfg.device)
+    elif isinstance(cfg.device, TeleoperatorConfig):
+        device = make_teleoperator_from_config(cfg.device)
+
+    device.connect(calibrate=False)
+    device.calibrate()
+    device.disconnect()
+
+
+if __name__ == "__main__":
+    calibrate()
--- a/lerobot/common/cameras/init.py
+++ b/lerobot/common/cameras/init.py
@@ -0,0 +1,17 @@
+# 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.
+
+from .camera import Camera
+from .configs import CameraConfig, ColorMode, Cv2Rotation
+from .utils import make_cameras_from_configs
--- a/lerobot/common/cameras/camera.py
+++ b/lerobot/common/cameras/camera.py
@@ -0,0 +1,122 @@
+#!/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 abc
+from typing import Any, Dict, List
+
+import numpy as np
+
+from .configs import CameraConfig, ColorMode
+
+
+class Camera(abc.ABC):
+    """Base class for camera implementations.
+
+    Defines a standard interface for camera operations across different backends.
+    Subclasses must implement all abstract methods.
+
+    Manages basic camera properties (FPS, resolution) and core operations:
+    - Connection/disconnection
+    - 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
+        warmup_time (int | None): Time reading frames before returning from connect (in seconds)
+
+    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):
+        """Initialize the camera with the given configuration.
+
+        Args:
+            config: Camera configuration containing FPS and resolution.
+        """
+        self.fps: int | None = config.fps
+        self.width: int | None = config.width
+        self.height: int | None = config.height
+        self.warmup_time: int | None = config.warmup_time
+
+    @property
+    @abc.abstractmethod
+    def is_connected(self) -> bool:
+        """Check if the camera is currently connected.
+
+        Returns:
+            bool: True if the camera is connected and ready to capture frames,
+                  False otherwise.
+        """
+        pass
+
+    @staticmethod
+    @abc.abstractmethod
+    def find_cameras() -> List[Dict[str, Any]]:
+        """Detects available cameras connected to the system.
+        Returns:
+            List[Dict[str, Any]]: A list of dictionaries,
+            where each dictionary contains information about a detected camera.
+        """
+        pass
+
+    @abc.abstractmethod
+    def connect(self, warmup: bool = True) -> None:
+        """Establish connection to the camera.
+
+        Args:
+            warmup: If True (default), captures a warmup frame before returning. Useful
+                   for cameras that require time to adjust capture settings.
+                   If False, skips the warmup frame.
+        """
+        pass
+
+    @abc.abstractmethod
+    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
+        """Capture and return a single frame from the camera.
+
+        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.
+        """
+        pass
+
+    @abc.abstractmethod
+    def async_read(self, timeout_ms: float = ...) -> np.ndarray:
+        """Asynchronously capture and return a single frame from the camera.
+
+        Args:
+            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.
+        """
+        pass
+
+    @abc.abstractmethod
+    def disconnect(self) -> None:
+        """Disconnect from the camera and release resources."""
+        pass
--- a/lerobot/common/cameras/configs.py
+++ b/lerobot/common/cameras/configs.py
@@ -0,0 +1,45 @@
+#!/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 abc
+from dataclasses import dataclass
+from enum import Enum
+
+import draccus
+
+
+class ColorMode(str, Enum):
+    RGB = "rgb"
+    BGR = "bgr"
+
+
+class Cv2Rotation(int, Enum):
+    NO_ROTATION = 0
+    ROTATE_90 = 90
+    ROTATE_180 = 180
+    ROTATE_270 = -90
+
+
+@dataclass(kw_only=True)
+class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
+    fps: int | None = None
+    width: int | None = None
+    height: int | None = None
+    warmup_time: int | None = None
+
+    @property
+    def type(self) -> str:
+        return self.get_choice_name(self.__class__)
--- a/lerobot/common/policies/hilserl/configuration_hilserl.py
+++ b/lerobot/common/policies/hilserl/configuration_hilserl.py
@@ -1,7 +1,4 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team.
-# All rights reserved.
+# 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.
@@ -15,9 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-from dataclasses import dataclass
-
-
-@dataclass
-class HILSerlConfig:
-    pass
+from .camera_opencv import OpenCVCamera
+from .configuration_opencv import OpenCVCameraConfig
--- a/lerobot/common/cameras/opencv/camera_opencv.py
+++ b/lerobot/common/cameras/opencv/camera_opencv.py
@@ -0,0 +1,513 @@
+# 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.
+
+"""
+Provides the OpenCVCamera class for capturing frames from cameras using OpenCV.
+"""
+
+import contextlib
+import logging
+import math
+import platform
+import queue
+import time
+from pathlib import Path
+from threading import Event, Thread
+from typing import Any, Dict, List
+
+import cv2
+import numpy as np
+
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+
+from ..camera import Camera
+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,
+# if you have 3 cameras, they should be associated to index 0, 1, and 2. This is the case
+# on MacOS. However, on Ubuntu, the indices are different like 6, 16, 23.
+# When you change the USB port or reboot the computer, the operating system might
+# treat the same cameras as new devices. Thus we select a higher bound to search indices.
+MAX_OPENCV_INDEX = 60
+
+logger = logging.getLogger(__name__)
+
+
+class OpenCVCamera(Camera):
+    """
+    Manages camera interactions using OpenCV for efficient frame recording.
+
+    This class provides a high-level interface to connect to, configure, and read
+    frames from cameras compatible with OpenCV's VideoCapture. It supports both
+    synchronous and asynchronous frame reading.
+
+    An OpenCVCamera instance requires a camera index (e.g., 0) or a device path
+    (e.g., '/dev/video0' on Linux). Camera indices can be unstable across reboots
+    or port changes, especially on Linux. Use the provided utility script to find
+    available camera indices or paths:
+    ```bash
+    python -m lerobot.find_cameras opencv
+    ```
+
+    The camera's default settings (FPS, resolution, color mode) are used unless
+    overridden in the configuration.
+
+    Example:
+        ```python
+        from lerobot.common.cameras.opencv import OpenCVCamera
+        from lerobot.common.cameras.configuration_opencv import OpenCVCameraConfig, ColorMode, Cv2Rotation
+
+        # Basic usage with camera index 0
+        config = OpenCVCameraConfig(index_or_path=0)
+        camera = OpenCVCamera(config)
+        try:
+            camera.connect()
+            print(f"Connected to {camera}")
+            color_image = camera.read() # Synchronous read
+            print(f"Read frame shape: {color_image.shape}")
+            async_image = camera.async_read() # Asynchronous read
+            print(f"Async read frame shape: {async_image.shape}")
+        except Exception as e:
+            print(f"An error occurred: {e}")
+        finally:
+            camera.disconnect()
+            print(f"Disconnected from {camera}")
+
+        # 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 ...
+        ```
+    """
+
+    def __init__(self, config: OpenCVCameraConfig):
+        """
+        Initializes the OpenCVCamera instance.
+
+        Args:
+            config: The configuration settings for the camera.
+        """
+        super().__init__(config)
+
+        self.config = config
+        self.index_or_path = config.index_or_path
+
+        self.fps = config.fps
+        self.color_mode = config.color_mode
+        self.warmup_time = config.warmup_time
+
+        self.videocapture: cv2.VideoCapture | None = None
+
+        self.thread: Thread | None = None
+        self.stop_event: Event | None = None
+        self.frame_queue: queue.Queue = queue.Queue(maxsize=1)
+
+        self.rotation: int | None = get_cv2_rotation(config.rotation)
+        self.backend: int = get_cv2_backend()
+
+        if self.height and self.width:
+            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+                self.capture_width, self.capture_height = self.height, self.width
+            else:
+                self.capture_width, self.capture_height = self.width, self.height
+
+    def __str__(self) -> str:
+        return f"{self.__class__.__name__}({self.index_or_path})"
+
+    @property
+    def is_connected(self) -> bool:
+        """Checks if the camera is currently connected and opened."""
+        return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
+
+    def connect(self, warmup: bool = True):
+        """
+        Connects to the OpenCV camera specified in the configuration.
+
+        Initializes the OpenCV VideoCapture object, sets desired camera properties
+        (FPS, width, height), and performs initial checks.
+
+        Raises:
+            DeviceAlreadyConnectedError: If the camera is already connected.
+            ValueError: If the specified camera index/path is not found or accessible.
+            ConnectionError: If 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.
+        cv2.setNumThreads(1)
+
+        self.videocapture = cv2.VideoCapture(self.index_or_path, self.backend)
+
+        if not self.videocapture.isOpened():
+            self.videocapture.release()
+            self.videocapture = None
+            raise ConnectionError(
+                f"Failed to open OpenCV camera {self.index_or_path}."
+                f"Run 'python -m lerobot.find_cameras opencv' for details about the available cameras in your system."
+            )
+
+        self._configure_capture_settings()
+
+        if warmup:
+            if self.warmup_time is None:
+                raise ValueError(
+                    f"Warmup time is not set for {self}. Please set a warmup time in the configuration."
+                )
+            logger.debug(f"Reading a warm-up frames for {self} for {self.warmup_time} seconds...")
+            start_time = time.time()
+            while time.time() - start_time < self.warmup_time:
+                self.read()
+                time.sleep(0.1)
+
+        logger.debug(f"Camera {self.index_or_path} connected and configured successfully.")
+
+    def _configure_capture_settings(self) -> None:
+        """
+        Applies the specified FPS, width, and height settings to the connected camera.
+
+        This method attempts to set the camera properties via OpenCV. It checks if
+        the camera successfully applied the settings and raises an error if not.
+
+        Args:
+            fps: The desired frames per second. If None, the setting is skipped.
+            width: The desired capture width. If None, the setting is skipped.
+            height: The desired capture height. If None, the setting is skipped.
+
+        Raises:
+            RuntimeError: If the camera fails to set any of the specified properties
+                          to the requested value.
+            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.")
+
+        if self.fps is None:
+            self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
+            logger.info(f"FPS set to camera default: {self.fps}.")
+        else:
+            self._validate_fps()
+
+        default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
+        default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
+
+        if self.width is None or self.height is None:
+            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+                self.width, self.height = default_height, default_width
+                self.capture_width, self.capture_height = default_width, default_height
+            else:
+                self.width, self.height = default_width, default_height
+                self.capture_width, self.capture_height = default_width, default_height
+            logger.info(f"Capture width set to camera default: {self.width}.")
+            logger.info(f"Capture height set to camera default: {self.height}.")
+        else:
+            self._validate_width_and_height()
+
+    def _validate_fps(self) -> None:
+        """Validates and sets the camera's frames per second (FPS)."""
+
+        success = self.videocapture.set(cv2.CAP_PROP_FPS, float(self.fps))
+        actual_fps = self.videocapture.get(cv2.CAP_PROP_FPS)
+        # Use math.isclose for robust float comparison
+        if not success or not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
+            raise RuntimeError(
+                f"Failed to set requested FPS {self.fps} for {self}. Actual value reported: {actual_fps} set success: {success})."
+            )
+        logger.debug(f"FPS set to {actual_fps} for {self}.")
+
+    def _validate_width_and_height(self) -> None:
+        """Validates and sets the camera's frame capture width and height."""
+
+        success = self.videocapture.set(cv2.CAP_PROP_FRAME_WIDTH, float(self.capture_width))
+        actual_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
+        if not success or self.capture_width != actual_width:
+            raise RuntimeError(
+                f"Failed to set requested capture width {self.capture_width} for {self}. Actual value: {actual_width} (set success: {success})."
+            )
+        logger.debug(f"Capture width set to {actual_width} for {self}.")
+
+        success = self.videocapture.set(cv2.CAP_PROP_FRAME_HEIGHT, float(self.capture_height))
+        actual_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
+        if not success or self.capture_height != actual_height:
+            raise RuntimeError(
+                f"Failed to set requested capture height {self.capture_height} for {self}. Actual value: {actual_height} (set success: {success})."
+            )
+        logger.debug(f"Capture height set to {actual_height} for {self}.")
+
+    @staticmethod
+    def find_cameras() -> List[Dict[str, Any]]:
+        """
+        Detects available OpenCV cameras connected to the system.
+
+        On Linux, it scans '/dev/video*' paths. On other systems (like macOS, Windows),
+        it checks indices from 0 up to `MAX_OPENCV_INDEX`.
+
+        Returns:
+            List[Dict[str, Any]]: A list of dictionaries,
+            where each dictionary contains 'type', 'id' (port index or path),
+            and the default profile properties (width, height, fps, format).
+        """
+        found_cameras_info = []
+
+        if platform.system() == "Linux":
+            possible_paths = sorted(Path("/dev").glob("video*"), key=lambda p: p.name)
+            targets_to_scan = [str(p) for p in possible_paths]
+        else:
+            targets_to_scan = list(range(MAX_OPENCV_INDEX))
+
+        logger.debug(f"Found potential paths: {targets_to_scan}")
+        for target in targets_to_scan:
+            camera = cv2.VideoCapture(target)
+            if camera.isOpened():
+                default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
+                default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                default_fps = camera.get(cv2.CAP_PROP_FPS)
+                default_format = camera.get(cv2.CAP_PROP_FORMAT)
+                camera_info = {
+                    "name": f"OpenCV Camera @ {target}",
+                    "type": "OpenCV",
+                    "id": target,
+                    "backend_api": camera.getBackendName(),
+                    "default_stream_profile": {
+                        "format": default_format,
+                        "width": default_width,
+                        "height": default_height,
+                        "fps": default_fps,
+                    },
+                }
+
+                found_cameras_info.append(camera_info)
+                camera.release()
+
+        if not found_cameras_info:
+            logger.warning("No OpenCV devices detected.")
+
+        logger.info(f"Detected OpenCV cameras: {[cam['id'] for cam in found_cameras_info]}")
+        return found_cameras_info
+
+    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
+        """
+        Reads a single frame synchronously from the 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
+                       color mode and applying any configured rotation.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If reading the frame from the camera fails or if the
+                          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()
+
+        # NOTE(Steven): Are we okay with this blocking an undefined amount of time?
+        ret, frame = self.videocapture.read()
+
+        if not ret or frame is None:
+            raise RuntimeError(
+                f"Failed to capture frame from {self}. '.read()' returned status={ret} and frame is None."
+            )
+
+        # Post-process the frame (color conversion, dimension check, rotation)
+        processed_frame = self._postprocess_image(frame, color_mode)
+
+        read_duration_ms = (time.perf_counter() - start_time) * 1e3
+        logger.debug(f"{self} synchronous read took: {read_duration_ms:.1f}ms")
+
+        return processed_frame
+
+    def _postprocess_image(self, image: np.ndarray, color_mode: ColorMode | None = None) -> np.ndarray:
+        """
+        Applies color conversion, dimension validation, and rotation to a raw frame.
+
+        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.
+
+        Raises:
+            ValueError: If the requested `color_mode` is invalid.
+            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 requested_color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"Invalid requested color mode '{requested_color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+            )
+
+        h, w, c = image.shape
+
+        if h != self.capture_height or w != self.capture_width:
+            raise RuntimeError(
+                f"Captured frame dimensions ({h}x{w}) do not match configured capture dimensions ({self.capture_height}x{self.capture_width}) for {self}."
+            )
+
+        processed_image = image
+        if requested_color_mode == ColorMode.RGB:
+            processed_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+            logger.debug(f"Converted frame from BGR to RGB for {self}.")
+
+        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+            processed_image = cv2.rotate(processed_image, self.rotation)
+            logger.debug(f"Rotated frame by {self.config.rotation} degrees for {self}.")
+
+        return processed_image
+
+    def _read_loop(self):
+        """
+        Internal loop run by the background thread for asynchronous reading.
+
+        Continuously reads frames from the camera using the synchronous `read()`
+        method and places the latest frame into the `frame_queue`. It overwrites
+        any previous frame in the queue.
+        """
+        logger.debug(f"Starting read loop thread for {self}.")
+        while not self.stop_event.is_set():
+            try:
+                color_image = self.read()
+
+                with contextlib.suppress(queue.Empty):
+                    _ = self.frame_queue.get_nowait()
+                self.frame_queue.put(color_image)
+                logger.debug(f"Frame placed in queue for {self}.")
+
+            except DeviceNotConnectedError:
+                logger.error(f"Read loop for {self} stopped: Camera disconnected.")
+                break
+            except Exception as e:
+                logger.warning(f"Error reading frame in background thread for {self}: {e}")
+
+        logger.debug(f"Stopping read loop thread for {self}.")
+
+    def _start_read_thread(self) -> None:
+        """Starts or restarts the background read thread if it's not running."""
+        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"OpenCVCameraReadLoop-{self}-{self.index_or_path}"
+        )
+        self.thread.daemon = True
+        self.thread.start()
+        logger.debug(f"Read thread started for {self}.")
+
+    def _stop_read_thread(self) -> None:
+        """Signals the background read thread to stop and waits for it to join."""
+        if self.stop_event is not None:
+            self.stop_event.set()
+
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=2.0)
+            if self.thread.is_alive():
+                logger.warning(f"Read thread for {self} did not terminate gracefully after 2 seconds.")
+            else:
+                logger.debug(f"Read thread for {self} joined successfully.")
+
+        self.thread = None
+        self.stop_event = None
+        logger.debug(f"Read thread stopped for {self}.")
+
+    def async_read(self, timeout_ms: float = 2000) -> np.ndarray:
+        """
+        Reads the latest available frame asynchronously.
+
+        This method retrieves the most recent frame captured by the background
+        read thread. It does not block waiting for the camera hardware directly,
+        only waits for a frame to appear in the internal queue up to the specified
+        timeout.
+
+        Args:
+            timeout_ms (float): Maximum time in milliseconds to wait for a frame
+                to become available in the queue. Defaults to 2000ms (2 seconds).
+
+        Returns:
+            np.ndarray: The latest captured frame as a NumPy array in the format
+                       (height, width, channels), processed according to configuration.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            TimeoutError: If no frame becomes available within the specified timeout.
+            RuntimeError: If an unexpected error occurs while retrieving from the queue.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
+        if self.thread is None or not self.thread.is_alive():
+            self._start_read_thread()
+
+        try:
+            return self.frame_queue.get(timeout=timeout_ms / 1000.0)
+        except queue.Empty as e:
+            thread_alive = self.thread is not None and self.thread.is_alive()
+            raise TimeoutError(
+                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
+                f"Read thread alive: {thread_alive}."
+            ) from e
+        except Exception as e:
+            logger.exception(f"Unexpected error getting frame from queue for {self}: {e}")
+            raise RuntimeError(f"Error getting frame from queue for camera {self.index_or_path}: {e}") from e
+
+    def disconnect(self):
+        """
+        Disconnects from the camera and cleans up resources.
+
+        Stops the background read thread (if running) and releases the OpenCV
+        VideoCapture object.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is already disconnected.
+        """
+        if not self.is_connected and self.thread is None:
+            raise DeviceNotConnectedError(f"{self} not connected.")
+
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        if self.videocapture is not None:
+            self.videocapture.release()
+            self.videocapture = None
+
+        logger.info(f"{self} disconnected.")
--- a/lerobot/common/cameras/opencv/configuration_opencv.py
+++ b/lerobot/common/cameras/opencv/configuration_opencv.py
@@ -0,0 +1,71 @@
+# 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.
+
+from dataclasses import dataclass
+from pathlib import Path
+
+from ..configs import CameraConfig, ColorMode, Cv2Rotation
+
+
+@CameraConfig.register_subclass("opencv")
+@dataclass
+class OpenCVCameraConfig(CameraConfig):
+    """Configuration class for OpenCV-based camera devices or video files.
+
+    This class provides configuration options for cameras accessed through OpenCV,
+    supporting both physical camera devices and video files. It includes settings
+    for resolution, frame rate, color mode, and image rotation.
+
+    Example configurations:
+    ```python
+    # Basic configurations
+    OpenCVCameraConfig(0, 30, 1280, 720)   # 1280x720 @ 30FPS
+    OpenCVCameraConfig(/dev/video4, 60, 640, 480)   # 640x480 @ 60FPS
+
+    # Advanced configurations
+    OpenCVCameraConfig(128422271347, 30, 640, 480, rotation=Cv2Rotation.ROTATE_90)     # With 90° rotation
+    ```
+
+    Attributes:
+        index_or_path: Either an integer representing the camera device index,
+                      or a Path object pointing to a video file.
+        fps: Requested frames per second for the color stream.
+        width: Requested frame width in pixels for the color stream.
+        height: Requested frame height in pixels for the color stream.
+        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
+        rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
+
+    Note:
+        - Only 3-channel color output (RGB/BGR) is currently supported.
+    """
+
+    index_or_path: int | Path
+    color_mode: ColorMode = ColorMode.RGB
+    rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
+
+    def __post_init__(self):
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        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."
+            )
--- a/lerobot/common/cameras/realsense/init.py
+++ b/lerobot/common/cameras/realsense/init.py
@@ -0,0 +1,16 @@
+# 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.
+
+from .camera_realsense import RealSenseCamera
+from .configuration_realsense import RealSenseCameraConfig
--- a/lerobot/common/cameras/realsense/camera_realsense.py
+++ b/lerobot/common/cameras/realsense/camera_realsense.py
@@ -0,0 +1,612 @@
+# 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.
+
+"""
+Provides the RealSenseCamera class for capturing frames from Intel RealSense cameras.
+"""
+
+import contextlib
+import logging
+import queue
+import time
+from threading import Event, Thread
+from typing import Any, Dict, List
+
+import cv2
+import numpy as np
+
+try:
+    import pyrealsense2 as rs
+except Exception as e:
+    logging.info(f"Could not import realsense: {e}")
+
+from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
+
+from ..camera import Camera
+from ..configs import ColorMode
+from ..utils import get_cv2_rotation
+from .configuration_realsense import RealSenseCameraConfig
+
+logger = logging.getLogger(__name__)
+
+
+class RealSenseCamera(Camera):
+    """
+    Manages interactions with Intel RealSense cameras for frame and depth recording.
+
+    This class provides an interface similar to `OpenCVCamera` but tailored for
+    RealSense devices, leveraging the `pyrealsense2` library. It uses the camera's
+    unique serial number for identification, offering more stability than device
+    indices, especially on Linux. It also supports capturing depth maps alongside
+    color frames.
+
+    Use the provided utility script to find available camera indices and default profiles:
+    ```bash
+    python -m lerobot.find_cameras realsense
+    ```
+
+    A `RealSenseCamera` instance requires a configuration object specifying the
+    camera's serial number or a unique device name. If using the name, ensure only
+    one camera with that name is connected.
+
+    The camera's default settings (FPS, resolution, color mode) from the stream
+    profile are used unless overridden in the configuration.
+
+    Example:
+        ```python
+        from lerobot.common.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
+        from lerobot.common.cameras import ColorMode, Cv2Rotation
+
+        # Basic usage with serial number
+        config = RealSenseCameraConfig(serial_number_or_name="1234567890") # Replace with actual SN
+        camera = RealSenseCamera(config)
+        camera.connect()
+
+        # Read 1 frame synchronously
+        color_image = camera.read()
+        print(color_image.shape)
+
+        # Read 1 frame asynchronously
+        async_image = camera.async_read()
+
+        # When done, properly disconnect the camera using
+        camera.disconnect()
+
+        # Example with depth capture and custom settings
+        custom_config = RealSenseCameraConfig(
+            serial_number_or_name="1234567890", # Replace with actual SN
+            fps=30,
+            width=1280,
+            height=720,
+            color_mode=ColorMode.BGR, # Request BGR output
+            rotation=Cv2Rotation.NO_ROTATION,
+            use_depth=True
+        )
+        depth_camera = RealSenseCamera(custom_config)
+        try:
+            depth_camera.connect()
+            depth_map = depth_camera.read_depth()
+            print(f"Depth shape: {depth_map.shape}")
+        finally:
+            depth_camera.disconnect()
+
+        # Example using a unique camera name
+        name_config = RealSenseCameraConfig(serial_number_or_name="Intel RealSense D435") # If unique
+        name_camera = RealSenseCamera(name_config)
+        # ... connect, read, disconnect ...
+        ```
+    """
+
+    def __init__(self, config: RealSenseCameraConfig):
+        """
+        Initializes the RealSenseCamera instance.
+
+        Args:
+            config: The configuration settings for the camera.
+        """
+
+        super().__init__(config)
+
+        self.config = config
+
+        if isinstance(config.serial_number_or_name, int):
+            self.serial_number = str(config.serial_number_or_name)
+        else:
+            self.serial_number = self._find_serial_number_from_name(config.serial_number_or_name)
+
+        self.fps = config.fps
+        self.color_mode = config.color_mode
+        self.use_depth = config.use_depth
+        self.warmup_time = config.warmup_time
+
+        self.rs_pipeline: rs.pipeline | None = None
+        self.rs_profile: rs.pipeline_profile | None = None
+
+        self.thread: Thread | None = None
+        self.stop_event: Event | None = None
+        self.frame_queue: queue.Queue = queue.Queue(maxsize=1)
+
+        self.rotation: int | None = get_cv2_rotation(config.rotation)
+
+        if self.height and self.width:
+            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+                self.capture_width, self.capture_height = self.height, self.width
+            else:
+                self.capture_width, self.capture_height = self.width, self.height
+
+    def __str__(self) -> str:
+        return f"{self.__class__.__name__}({self.serial_number})"
+
+    @property
+    def is_connected(self) -> bool:
+        """Checks if the camera pipeline is started and streams are active."""
+        return self.rs_pipeline is not None and self.rs_profile is not None
+
+    def connect(self, warmup: bool = True):
+        """
+        Connects to the RealSense camera specified in the configuration.
+
+        Initializes the RealSense pipeline, configures the required streams (color
+        and optionally depth), starts the pipeline, and validates the actual stream settings.
+
+        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 = self._make_rs_pipeline_config()
+
+        try:
+            self.rs_profile = self.rs_pipeline.start(rs_config)
+            logger.debug(f"Successfully started pipeline for camera {self.serial_number}.")
+        except RuntimeError as e:
+            self.rs_profile = None
+            self.rs_pipeline = None
+            raise ConnectionError(
+                f"Failed to open {self} camera. Run 'python -m lerobot.find_cameras realsense' for details about the available cameras in your system."
+            ) from e
+
+        logger.debug(f"Validating stream configuration for {self}...")
+        self._validate_capture_settings()
+
+        if warmup:
+            if self.warmup_time is None:
+                raise ValueError(
+                    f"Warmup time is not set for {self}. Please set a warmup time in the configuration."
+                )
+            logger.debug(f"Reading a warm-up frames for {self} for {self.warmup_time} seconds...")
+            start_time = time.time()
+            while time.time() - start_time < self.warmup_time:
+                self.read()
+                time.sleep(0.1)
+
+        logger.info(f"{self} connected.")
+
+    @staticmethod
+    def find_cameras() -> List[Dict[str, Any]]:
+        """
+        Detects available Intel RealSense cameras connected to the system.
+
+        Returns:
+            List[Dict[str, Any]]: A list of dictionaries,
+            where each dictionary contains 'type', 'id' (serial number), 'name',
+            firmware version, USB type, and other available specs, and the default profile properties (width, height, fps, format).
+
+        Raises:
+            OSError: If pyrealsense2 is not installed.
+            ImportError: If pyrealsense2 is not installed.
+        """
+        found_cameras_info = []
+        context = rs.context()
+        devices = context.query_devices()
+
+        for device in devices:
+            camera_info = {
+                "name": device.get_info(rs.camera_info.name),
+                "type": "RealSense",
+                "id": device.get_info(rs.camera_info.serial_number),
+                "firmware_version": device.get_info(rs.camera_info.firmware_version),
+                "usb_type_descriptor": device.get_info(rs.camera_info.usb_type_descriptor),
+                "physical_port": device.get_info(rs.camera_info.physical_port),
+                "product_id": device.get_info(rs.camera_info.product_id),
+                "product_line": device.get_info(rs.camera_info.product_line),
+            }
+
+            # Get stream profiles for each sensor
+            sensors = device.query_sensors()
+            for sensor in sensors:
+                profiles = sensor.get_stream_profiles()
+
+                for profile in profiles:
+                    if profile.is_video_stream_profile() and profile.is_default():
+                        vprofile = profile.as_video_stream_profile()
+                        stream_info = {
+                            "stream_type": vprofile.stream_name(),
+                            "format": vprofile.format().name,
+                            "width": vprofile.width(),
+                            "height": vprofile.height(),
+                            "fps": vprofile.fps(),
+                        }
+                        camera_info["default_stream_profile"] = stream_info
+
+            found_cameras_info.append(camera_info)
+
+        logger.info(f"Detected RealSense cameras: {[cam['id'] for cam in found_cameras_info]}")
+        return found_cameras_info
+
+    def _find_serial_number_from_name(self, name: str) -> str:
+        """Finds the serial number for a given unique camera name."""
+        camera_infos = self.find_cameras()
+        found_devices = [cam for cam in camera_infos if str(cam["name"]) == name]
+
+        if not found_devices:
+            available_names = [cam["name"] for cam in camera_infos]
+            raise ValueError(
+                f"No RealSense camera found with name '{name}'. Available camera names: {available_names}"
+            )
+
+        if len(found_devices) > 1:
+            serial_numbers = [dev["serial_number"] for dev in found_devices]
+            raise ValueError(
+                f"Multiple RealSense cameras found with name '{name}'. "
+                f"Please use a unique serial number instead. Found SNs: {serial_numbers}"
+            )
+
+        serial_number = str(found_devices[0]["serial_number"])
+        logger.info(f"Found serial number '{serial_number}' for camera name '{name}'.")
+        return serial_number
+
+    def _make_rs_pipeline_config(self) -> rs.config:
+        """Creates and configures the RealSense pipeline configuration object."""
+        rs_config = rs.config()
+        rs.config.enable_device(rs_config, self.serial_number)
+
+        if self.width and self.height and self.fps:
+            logger.debug(
+                f"Requesting Color Stream: {self.capture_width}x{self.capture_height} @ {self.fps} FPS, Format: {rs.format.rgb8}"
+            )
+            rs_config.enable_stream(
+                rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps
+            )
+            if self.use_depth:
+                logger.debug(
+                    f"Requesting Depth Stream: {self.capture_width}x{self.capture_height} @ {self.fps} FPS, Format: {rs.format.z16}"
+                )
+                rs_config.enable_stream(
+                    rs.stream.depth, self.capture_width, self.capture_height, rs.format.z16, self.fps
+                )
+        else:
+            logger.debug(f"Requesting Color Stream: Default settings, Format: {rs.stream.color}")
+            rs_config.enable_stream(rs.stream.color)
+            if self.use_depth:
+                logger.debug(f"Requesting Depth Stream: Default settings, Format: {rs.stream.depth}")
+                rs_config.enable_stream(rs.stream.depth)
+
+        return rs_config
+
+    def _validate_capture_settings(self) -> None:
+        """
+        Validates if the actual stream settings match the requested configuration.
+
+        This method compares the requested FPS, width, and height against the
+        actual settings obtained from the active RealSense profile after the
+        pipeline has started.
+
+        Raises:
+            RuntimeError: If the actual camera settings significantly deviate
+                          from the requested ones.
+            DeviceNotConnectedError: If the camera is not connected when attempting
+                                     to validate settings.
+        """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
+
+        stream = self.rs_profile.get_stream(rs.stream.color).as_video_stream_profile()
+
+        if self.fps is None:
+            self.fps = stream.fps()
+
+        if self.width is None or self.height is None:
+            actual_width = int(round(stream.width()))
+            actual_height = int(round(stream.height()))
+            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+                self.width, self.height = actual_height, actual_width
+                self.capture_width, self.capture_height = actual_width, actual_height
+            else:
+                self.width, self.height = actual_width, actual_height
+                self.capture_width, self.capture_height = actual_width, actual_height
+            logger.info(f"Capture width set to camera default: {self.width}.")
+            logger.info(f"Capture height set to camera default: {self.height}.")
+        else:
+            self._validate_width_and_height(stream)
+
+        if self.use_depth:
+            stream = self.rs_profile.get_stream(rs.stream.depth).as_video_stream_profile()
+            self._validate_width_and_height(stream)
+
+    def _validate_width_and_height(self, stream) -> None:
+        """Validates and sets the internal capture width and height based on actual stream width."""
+        actual_width = int(round(stream.width()))
+        actual_height = int(round(stream.height()))
+
+        if self.capture_width != actual_width:
+            raise RuntimeError(
+                f"Failed to set requested capture width {self.capture_width} for {self}. Actual value: {actual_width}."
+            )
+        logger.debug(f"Capture width set to {actual_width} for {self}.")
+
+        if self.capture_height != actual_height:
+            raise RuntimeError(
+                f"Failed to set requested capture height {self.capture_height} for {self}. Actual value: {actual_height}."
+            )
+        logger.debug(f"Capture height set to {actual_height} for {self}.")
+
+    def read_depth(self, timeout_ms: int = 100) -> np.ndarray:
+        """
+        Reads a single frame (depth) synchronously from the 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 100ms.
+
+        Returns:
+            np.ndarray: The depth map as a NumPy array (height, width)
+                  of type `np.uint16` (raw depth values in millimeters) and rotation.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If reading frames from the pipeline fails or frames are invalid.
+        """
+
+        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}."
+            )
+
+        start_time = time.perf_counter()
+
+        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
+
+        if not ret or frame is None:
+            raise RuntimeError(f"{self} failed to capture frame. Returned status='{ret}'.")
+
+        depth_frame = frame.get_depth_frame()
+        depth_map = np.asanyarray(depth_frame.get_data())
+
+        depth_map_processed = self._postprocess_image(depth_map, depth_frame=True)
+
+        read_duration_ms = (time.perf_counter() - start_time) * 1e3
+        logger.debug(f"{self} synchronous read took: {read_duration_ms:.1f}ms")
+
+        return depth_map_processed
+
+    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 100) -> np.ndarray:
+        """
+        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 100ms.
+
+        Returns:
+            np.ndarray: The captured color frame as a NumPy array
+              (height, width, channels), processed according to `color_mode` and rotation.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            RuntimeError: If reading frames from the pipeline fails or frames are invalid.
+            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()
+
+        ret, frame = self.rs_pipeline.try_wait_for_frames(
+            timeout_ms=timeout_ms
+        )  # NOTE(Steven): This read has a timeout while opencv doesn't
+
+        if not ret or frame is None:
+            raise RuntimeError(
+                f"Failed to capture frame from {self}. '.read()' returned status={ret} and frame is None."
+            )
+
+        color_frame = frame.get_color_frame()
+        color_image_raw = np.asanyarray(color_frame.get_data())
+
+        color_image_processed = self._postprocess_image(color_image_raw, color_mode)
+
+        read_duration_ms = (time.perf_counter() - start_time) * 1e3
+        logger.debug(f"{self} synchronous read took: {read_duration_ms:.1f}ms")
+
+        return color_image_processed
+
+    def _postprocess_image(
+        self, image: np.ndarray, color_mode: ColorMode | None = None, depth_frame: bool = False
+    ) -> np.ndarray:
+        """
+        Applies color conversion, dimension validation, and rotation to a raw color frame.
+
+        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`.
+
+        Raises:
+            ValueError: If the requested `color_mode` is invalid.
+            RuntimeError: If the raw frame dimensions do not match the configured
+                          `width` and `height`.
+        """
+
+        if color_mode and color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"Invalid requested color mode '{color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+            )
+
+        if depth_frame:
+            h, w = image.shape
+        else:
+            h, w, _c = image.shape
+
+        if h != self.capture_height or w != self.capture_width:
+            raise RuntimeError(
+                f"Captured frame dimensions ({h}x{w}) do not match configured capture dimensions ({self.capture_height}x{self.capture_width}) for {self}."
+            )
+
+        processed_image = image
+        if self.color_mode == ColorMode.BGR:
+            processed_image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
+            logger.debug(f"Converted frame from RGB to BGR for {self}.")
+
+        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
+            processed_image = cv2.rotate(processed_image, self.rotation)
+            logger.debug(f"Rotated frame by {self.config.rotation} degrees for {self}.")
+
+        return processed_image
+
+    def _read_loop(self):
+        """
+        Internal loop run by the background thread for asynchronous reading.
+
+        Continuously reads frames (color and optional depth) using `read()`
+        and places the latest result (single image or tuple) into the `frame_queue`.
+        It overwrites any previous frame in the queue.
+        """
+        logger.debug(f"Starting read loop thread for {self}.")
+        while not self.stop_event.is_set():
+            try:
+                frame_data = self.read(timeout_ms=500)
+
+                with contextlib.suppress(queue.Empty):
+                    _ = self.frame_queue.get_nowait()
+                self.frame_queue.put(frame_data)
+                logger.debug(f"Frame data placed in queue for {self}.")
+
+            except DeviceNotConnectedError:
+                logger.error(f"Read loop for {self} stopped: Camera disconnected.")
+                break
+            except Exception as e:
+                logger.warning(f"Error reading frame in background thread for {self}: {e}")
+
+        logger.debug(f"Stopping read loop thread for {self}.")
+
+    def _start_read_thread(self) -> None:
+        """Starts or restarts the background read thread if it's not running."""
+        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()
+        logger.debug(f"Read thread started for {self}.")
+
+    def _stop_read_thread(self):
+        """Signals the background read thread to stop and waits for it to join."""
+        if self.stop_event is not None:
+            self.stop_event.set()
+
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=2.0)
+            if self.thread.is_alive():
+                logger.warning(f"Read thread for {self} did not terminate gracefully after 2 seconds.")
+            else:
+                logger.debug(f"Read thread for {self} joined successfully.")
+
+        self.thread = None
+        self.stop_event = None
+        logger.debug(f"Read thread stopped for {self}.")
+
+    # NOTE(Steven): Missing implementation for depth for now
+    def async_read(self, timeout_ms: float = 100) -> np.ndarray:
+        """
+        Reads the latest available frame data (color or color+depth) asynchronously.
+
+        This method retrieves the most recent frame captured by the background
+        read thread. It does not block waiting for the camera hardware directly,
+        only waits for a frame to appear in the internal queue up to the specified
+        timeout.
+
+        Args:
+            timeout_ms (float): Maximum time in milliseconds to wait for a frame
+                to become available in the queue. Defaults to 100ms (0.1 seconds).
+
+        Returns:
+            np.ndarray:
+            The latest captured frame data (color image), processed according to configuration.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is not connected.
+            TimeoutError: If no frame data becomes available within the specified timeout.
+            RuntimeError: If the background thread died unexpectedly or another queue 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():
+            self._start_read_thread()
+
+        try:
+            return self.frame_queue.get(timeout=timeout_ms / 1000.0)
+        except queue.Empty as e:
+            thread_alive = self.thread is not None and self.thread.is_alive()
+            raise TimeoutError(
+                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
+                f"Read thread alive: {thread_alive}."
+            ) from e
+        except Exception as e:
+            raise RuntimeError(f"Error getting frame data from queue for camera {self}: {e}") from e
+
+    def disconnect(self):
+        """
+        Disconnects from the camera, stops the pipeline, and cleans up resources.
+
+        Stops the background read thread (if running) and stops the RealSense pipeline.
+
+        Raises:
+            DeviceNotConnectedError: If the camera is already disconnected (pipeline not running).
+        """
+
+        if not self.is_connected and self.thread is None:
+            raise DeviceNotConnectedError(
+                f"Attempted to disconnect {self}, but it appears already disconnected."
+            )
+
+        if self.thread is not None:
+            self._stop_read_thread()
+
+        if self.rs_pipeline is not None:
+            logger.debug(f"Stopping RealSense pipeline object for {self}.")
+            self.rs_pipeline.stop()
+            self.rs_pipeline = None
+            self.rs_profile = None
+
+        logger.info(f"{self} disconnected.")
--- a/lerobot/common/cameras/realsense/configuration_realsense.py
+++ b/lerobot/common/cameras/realsense/configuration_realsense.py
@@ -0,0 +1,80 @@
+# 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.
+
+from dataclasses import dataclass
+
+from ..configs import CameraConfig, ColorMode, Cv2Rotation
+
+
+@CameraConfig.register_subclass("intelrealsense")
+@dataclass
+class RealSenseCameraConfig(CameraConfig):
+    """Configuration class for Intel RealSense cameras.
+
+    This class provides specialized configuration options for Intel RealSense cameras,
+    including support for depth sensing and device identification via serial number or name.
+
+    Example configurations for Intel RealSense D405:
+    ```python
+    # Basic configurations
+    RealSenseCameraConfig(128422271347, 30, 1280, 720)   # 1280x720 @ 30FPS
+    RealSenseCameraConfig(128422271347, 60, 640, 480)   # 640x480 @ 60FPS
+
+    # Advanced configurations
+    RealSenseCameraConfig(128422271347, 30, 640, 480, use_depth=True)  # With depth sensing
+    RealSenseCameraConfig(128422271347, 30, 640, 480, rotation=Cv2Rotation.ROTATE_90)     # With 90° rotation
+    ```
+
+    Attributes:
+        fps: Requested frames per second for the color stream.
+        width: Requested frame width in pixels for the color stream.
+        height: Requested frame height in pixels for the color stream.
+        serial_number_or_name: Unique serial number or human-readable name to identify the camera.
+        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
+        use_depth: Whether to enable depth stream. Defaults to False.
+        rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
+
+    Note:
+        - Either name or serial_number must be specified.
+        - Depth stream configuration (if enabled) will use the same FPS as the color stream.
+        - The actual resolution and FPS may be adjusted by the camera to the nearest supported mode.
+        - For `fps`, `width` and `height`, either all of them need to be set, or none of them.
+    """
+
+    serial_number_or_name: int | str
+    color_mode: ColorMode = ColorMode.RGB
+    use_depth: bool = False
+    rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION  # NOTE(Steven): Check if draccus can parse to an enum
+
+    def __post_init__(self):
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        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):
+            raise ValueError(
+                "For `fps`, `width` and `height`, either all of them need to be set, or none of them."
+            )
--- a/lerobot/common/cameras/utils.py
+++ b/lerobot/common/cameras/utils.py
@@ -0,0 +1,75 @@
+#!/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 platform
+from pathlib import Path
+from typing import TypeAlias
+
+import numpy as np
+from PIL import Image
+
+from .camera import Camera
+from .configs import CameraConfig, Cv2Rotation
+
+IndexOrPath: TypeAlias = int | Path
+
+
+def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[str, Camera]:
+    cameras = {}
+
+    for key, cfg in camera_configs.items():
+        if cfg.type == "opencv":
+            from .opencv import OpenCVCamera
+
+            cameras[key] = OpenCVCamera(cfg)
+
+        elif cfg.type == "intelrealsense":
+            from .realsense.camera_realsense import RealSenseCamera
+
+            cameras[key] = RealSenseCamera(cfg)
+        else:
+            raise ValueError(f"The motor type '{cfg.type}' is not valid.")
+
+    return cameras
+
+
+def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
+    import cv2
+
+    if rotation == Cv2Rotation.ROTATE_90:
+        return cv2.ROTATE_90_CLOCKWISE
+    elif rotation == Cv2Rotation.ROTATE_180:
+        return cv2.ROTATE_180
+    elif rotation == Cv2Rotation.ROTATE_270:
+        return cv2.ROTATE_90_COUNTERCLOCKWISE
+    else:
+        return None
+
+
+def get_cv2_backend() -> int:
+    import cv2
+
+    if platform.system() == "Windows":
+        return cv2.CAP_AVFOUNDATION
+    else:
+        return cv2.CAP_ANY
+
+
+def save_image(img_array: np.ndarray, camera_index: int, frame_index: int, images_dir: Path):
+    img = Image.fromarray(img_array)
+    path = images_dir / f"camera_{camera_index:02d}_frame_{frame_index:06d}.png"
+    path.parent.mkdir(parents=True, exist_ok=True)
+    img.save(str(path), quality=100)
--- a/lerobot/common/constants.py
+++ b/lerobot/common/constants.py
@@ -17,12 +17,15 @@ from pathlib import Path

 from huggingface_hub.constants import HF_HOME

-OBS_ENV = "observation.environment_state"
-OBS_ROBOT = "observation.state"
+OBS_ENV_STATE = "observation.environment_state"
+OBS_STATE = "observation.state"
 OBS_IMAGE = "observation.image"
 OBS_IMAGES = "observation.images"
 ACTION = "action"

+ROBOTS = "robots"
+TELEOPERATORS = "teleoperators"
+
 # files & directories
 CHECKPOINTS_DIR = "checkpoints"
 LAST_CHECKPOINT_LINK = "last"
@@ -34,12 +37,16 @@ OPTIMIZER_STATE = "optimizer_state.safetensors"
 OPTIMIZER_PARAM_GROUPS = "optimizer_param_groups.json"
 SCHEDULER_STATE = "scheduler_state.json"

-# cache dir
-default_cache_path = Path(HF_HOME) / "lerobot"
-HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
-
 if "LEROBOT_HOME" in os.environ:
    raise ValueError(
        f"You have a 'LEROBOT_HOME' environment variable set to '{os.getenv('LEROBOT_HOME')}'.\n"
        "'LEROBOT_HOME' is deprecated, please use 'HF_LEROBOT_HOME' instead."
    )
+
+# cache dir
+default_cache_path = Path(HF_HOME) / "lerobot"
+HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
+
+# calibration dir
+default_calibration_path = HF_LEROBOT_HOME / "calibration"
+HF_LEROBOT_CALIBRATION = Path(os.getenv("HF_LEROBOT_CALIBRATION", default_calibration_path)).expanduser()
--- a/lerobot/common/datasets/compute_stats.py
+++ b/lerobot/common/datasets/compute_stats.py
@@ -19,10 +19,7 @@ from lerobot.common.datasets.utils import load_image_as_numpy


 def estimate_num_samples(
-    dataset_len: int,
-    min_num_samples: int = 100,
-    max_num_samples: int = 10_000,
-    power: float = 0.75,
+    dataset_len: int, min_num_samples: int = 100, max_num_samples: int = 10_000, power: float = 0.75
 ) -> int:
    """Heuristic to estimate the number of samples based on dataset size.
    The power controls the sample growth relative to dataset size.
@@ -126,9 +123,7 @@ def _assert_type_and_shape(stats_list: list[dict[str, dict]]):
                    raise ValueError(f"Shape of '{k}' must be (3,1,1), but is {v.shape} instead.")


-def aggregate_feature_stats(
-    stats_ft_list: list[dict[str, dict]],
-) -> dict[str, dict[str, np.ndarray]]:
+def aggregate_feature_stats(stats_ft_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
    """Aggregates stats for a single feature."""
    means = np.stack([s["mean"] for s in stats_ft_list])
    variances = np.stack([s["std"] ** 2 for s in stats_ft_list])
@@ -157,9 +152,7 @@ def aggregate_feature_stats(
    }


-def aggregate_stats(
-    stats_list: list[dict[str, dict]],
-) -> dict[str, dict[str, np.ndarray]]:
+def aggregate_stats(stats_list: list[dict[str, dict]]) -> dict[str, dict[str, np.ndarray]]:
    """Aggregate stats from multiple compute_stats outputs into a single set of stats.

    The final stats will have the union of all data keys from each of the stats dicts.
--- a/lerobot/common/datasets/factory.py
+++ b/lerobot/common/datasets/factory.py
@@ -49,7 +49,7 @@ def resolve_delta_timestamps(
                "observation.state": [-0.04, -0.02, 0]
                "observation.action": [-0.02, 0, 0.02]
            }
-            returns `None` if the the resulting dict is empty.
+            returns `None` if the resulting dict is empty.
    """
    delta_timestamps = {}
    for key in ds_meta.features:
--- a/lerobot/common/datasets/lerobot_dataset.py
+++ b/lerobot/common/datasets/lerobot_dataset.py
@@ -48,7 +48,6 @@ from lerobot.common.datasets.utils import (
    embed_images,
    get_delta_indices,
    get_episode_data_index,
-    get_features_from_robot,
    get_hf_features_from_features,
    get_safe_version,
    hf_transform_to_torch,
@@ -72,7 +71,6 @@ from lerobot.common.datasets.video_utils import (
    get_safe_default_codec,
    get_video_info,
 )
-from lerobot.common.robot_devices.robots.utils import Robot

 CODEBASE_VERSION = "v2.1"

@@ -304,10 +302,9 @@ class LeRobotDatasetMetadata:
        cls,
        repo_id: str,
        fps: int,
-        root: str | Path | None = None,
-        robot: Robot | None = None,
+        features: dict,
        robot_type: str | None = None,
-        features: dict | None = None,
+        root: str | Path | None = None,
        use_videos: bool = True,
    ) -> "LeRobotDatasetMetadata":
        """Creates metadata for a LeRobotDataset."""
@@ -317,33 +314,27 @@ class LeRobotDatasetMetadata:

        obj.root.mkdir(parents=True, exist_ok=False)

-        if robot is not None:
-            features = {**(features or {}), **get_features_from_robot(robot)}
-            robot_type = robot.robot_type
-            if not all(cam.fps == fps for cam in robot.cameras.values()):
-                logging.warning(
-                    f"Some cameras in your {robot.robot_type} robot don't have an fps matching the fps of your dataset."
-                    "In this case, frames from lower fps cameras will be repeated to fill in the blanks."
-                )
-        elif features is None:
-            raise ValueError(
-                "Dataset features must either come from a Robot or explicitly passed upon creation."
-            )
-        else:
-            # TODO(aliberts, rcadene): implement sanity check for features
-            features = {**features, **DEFAULT_FEATURES}
+        # if robot is not None:
+        #     features = get_features_from_robot(robot, use_videos)
+        #     robot_type = robot.robot_type
+        #     if not all(cam.fps == fps for cam in robot.cameras.values()):
+        #         logging.warning(
+        #             f"Some cameras in your {robot.robot_type} robot don't have an fps matching the fps of your dataset."
+        #             "In this case, frames from lower fps cameras will be repeated to fill in the blanks."
+        #         )

-            # check if none of the features contains a "/" in their names,
-            # as this would break the dict flattening in the stats computation, which uses '/' as separator
-            for key in features:
-                if "/" in key:
-                    raise ValueError(f"Feature names should not contain '/'. Found '/' in feature '{key}'.")
+        # TODO(aliberts, rcadene): implement sanity check for features
+        features = {**features, **DEFAULT_FEATURES}

-            features = {**features, **DEFAULT_FEATURES}
+        # check if none of the features contains a "/" in their names,
+        # as this would break the dict flattening in the stats computation, which uses '/' as separator
+        for key in features:
+            if "/" in key:
+                raise ValueError(f"Feature names should not contain '/'. Found '/' in feature '{key}'.")

        obj.tasks, obj.task_to_task_index = {}, {}
        obj.episodes_stats, obj.stats, obj.episodes = {}, {}, {}
-        obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, robot_type, features, use_videos)
+        obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, features, use_videos, robot_type)
        if len(obj.video_keys) > 0 and not use_videos:
            raise ValueError()
        write_json(obj.info, obj.root / INFO_PATH)
@@ -785,7 +776,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
        else:
            self.image_writer.save_image(image=image, fpath=fpath)

-    def add_frame(self, frame: dict) -> None:
+    def add_frame(self, frame: dict, task: str, timestamp: float | None = None) -> None:
        """
        This function only adds the frame to the episode_buffer. Apart from images — which are written in a
        temporary directory — nothing is written to disk. To save those frames, the 'save_episode()' method
@@ -803,17 +794,14 @@ class LeRobotDataset(torch.utils.data.Dataset):

        # Automatically add frame_index and timestamp to episode buffer
        frame_index = self.episode_buffer["size"]
-        timestamp = frame.pop("timestamp") if "timestamp" in frame else frame_index / self.fps
+        if timestamp is None:
+            timestamp = frame_index / self.fps
        self.episode_buffer["frame_index"].append(frame_index)
        self.episode_buffer["timestamp"].append(timestamp)
+        self.episode_buffer["task"].append(task)

        # Add frame features to episode_buffer
        for key in frame:
-            if key == "task":
-                # Note: we associate the task in natural language to its task index during `save_episode`
-                self.episode_buffer["task"].append(frame["task"])
-                continue
-
            if key not in self.features:
                raise ValueError(
                    f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
@@ -821,9 +809,7 @@ class LeRobotDataset(torch.utils.data.Dataset):

            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,
+                    episode_index=self.episode_buffer["episode_index"], image_key=key, frame_index=frame_index
                )
                if frame_index == 0:
                    img_path.parent.mkdir(parents=True, exist_ok=True)
@@ -869,10 +855,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
        for key, ft in self.features.items():
            # index, episode_index, task_index are already processed above, and image and video
            # 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",
-            ]:
+            if key in ["index", "episode_index", "task_index"] or ft["dtype"] in ["image", "video"]:
                continue
            episode_buffer[key] = np.stack(episode_buffer[key])

@@ -994,10 +977,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
        cls,
        repo_id: str,
        fps: int,
+        features: dict,
        root: str | Path | None = None,
-        robot: Robot | None = None,
        robot_type: str | None = None,
-        features: dict | None = None,
        use_videos: bool = True,
        tolerance_s: float = 1e-4,
        image_writer_processes: int = 0,
@@ -1009,10 +991,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
        obj.meta = LeRobotDatasetMetadata.create(
            repo_id=repo_id,
            fps=fps,
-            root=root,
-            robot=robot,
            robot_type=robot_type,
            features=features,
+            root=root,
            use_videos=use_videos,
        )
        obj.repo_id = obj.meta.repo_id
@@ -1058,7 +1039,7 @@ class MultiLeRobotDataset(torch.utils.data.Dataset):
        super().__init__()
        self.repo_ids = repo_ids
        self.root = Path(root) if root else HF_LEROBOT_HOME
-        self.tolerances_s = tolerances_s if tolerances_s else {repo_id: 1e-4 for repo_id in repo_ids}
+        self.tolerances_s = tolerances_s if tolerances_s else dict.fromkeys(repo_ids, 0.0001)
        # Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
        # are handled by this class.
        self._datasets = [
--- a/lerobot/common/datasets/online_buffer.py
+++ b/lerobot/common/datasets/online_buffer.py
@@ -154,32 +154,14 @@ class OnlineBuffer(torch.utils.data.Dataset):
            OnlineBuffer.NEXT_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": ()},
            # Since the memmap is initialized with all-zeros, this keeps track of which indices are occupied
            # with real data rather than the dummy initialization.
-            OnlineBuffer.OCCUPANCY_MASK_KEY: {
-                "dtype": np.dtype("?"),
-                "shape": (buffer_capacity,),
-            },
-            OnlineBuffer.INDEX_KEY: {
-                "dtype": np.dtype("int64"),
-                "shape": (buffer_capacity,),
-            },
-            OnlineBuffer.FRAME_INDEX_KEY: {
-                "dtype": np.dtype("int64"),
-                "shape": (buffer_capacity,),
-            },
-            OnlineBuffer.EPISODE_INDEX_KEY: {
-                "dtype": np.dtype("int64"),
-                "shape": (buffer_capacity,),
-            },
-            OnlineBuffer.TIMESTAMP_KEY: {
-                "dtype": np.dtype("float64"),
-                "shape": (buffer_capacity,),
-            },
+            OnlineBuffer.OCCUPANCY_MASK_KEY: {"dtype": np.dtype("?"), "shape": (buffer_capacity,)},
+            OnlineBuffer.INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.FRAME_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.EPISODE_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.TIMESTAMP_KEY: {"dtype": np.dtype("float64"), "shape": (buffer_capacity,)},
        }
        for k, v in data_spec.items():
-            complete_data_spec[k] = {
-                "dtype": v["dtype"],
-                "shape": (buffer_capacity, *v["shape"]),
-            }
+            complete_data_spec[k] = {"dtype": v["dtype"], "shape": (buffer_capacity, *v["shape"])}
        return complete_data_spec

    def add_data(self, data: dict[str, np.ndarray]):
--- a/lerobot/common/datasets/push_dataset_to_hub/utils.py
+++ b/lerobot/common/datasets/push_dataset_to_hub/utils.py
@@ -77,9 +77,7 @@ def check_repo_id(repo_id: str) -> None:


 # TODO(aliberts): remove
-def calculate_episode_data_index(
-    hf_dataset: datasets.Dataset,
-) -> Dict[str, torch.Tensor]:
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
    """
    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.

--- a/lerobot/common/datasets/sampler.py
+++ b/lerobot/common/datasets/sampler.py
@@ -43,10 +43,7 @@ class EpisodeAwareSampler:
        ):
            if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
                indices.extend(
-                    range(
-                        start_index.item() + drop_n_first_frames,
-                        end_index.item() - drop_n_last_frames,
-                    )
+                    range(start_index.item() + drop_n_first_frames, end_index.item() - drop_n_last_frames)
                )

        self.indices = indices
--- a/lerobot/common/datasets/transforms.py
+++ b/lerobot/common/datasets/transforms.py
@@ -128,7 +128,7 @@ class SharpnessJitter(Transform):
            raise TypeError(f"{sharpness=} should be a single number or a sequence with length 2.")

        if not 0.0 <= sharpness[0] <= sharpness[1]:
-            raise ValueError(f"sharpnesss values should be between (0., inf), but got {sharpness}.")
+            raise ValueError(f"sharpness values should be between (0., inf), but got {sharpness}.")

        return float(sharpness[0]), float(sharpness[1])

--- a/lerobot/common/datasets/utils.py
+++ b/lerobot/common/datasets/utils.py
@@ -40,7 +40,7 @@ from lerobot.common.datasets.backward_compatibility import (
    BackwardCompatibilityError,
    ForwardCompatibilityError,
 )
-from lerobot.common.robot_devices.robots.utils import Robot
+from lerobot.common.robots import Robot
 from lerobot.common.utils.utils import is_valid_numpy_dtype_string
 from lerobot.configs.types import DictLike, FeatureType, PolicyFeature

@@ -225,10 +225,7 @@ def load_episodes(local_dir: Path) -> dict:
 def write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path):
    # We wrap episode_stats in a dictionary since `episode_stats["episode_index"]`
    # is a dictionary of stats and not an integer.
-    episode_stats = {
-        "episode_index": episode_index,
-        "stats": serialize_dict(episode_stats),
-    }
+    episode_stats = {"episode_index": episode_index, "stats": serialize_dict(episode_stats)}
    append_jsonlines(episode_stats, local_dir / EPISODES_STATS_PATH)


@@ -243,7 +240,7 @@ def load_episodes_stats(local_dir: Path) -> dict:
 def backward_compatible_episodes_stats(
    stats: dict[str, dict[str, np.ndarray]], episodes: list[int]
 ) -> dict[str, dict[str, np.ndarray]]:
-    return {ep_idx: stats for ep_idx in episodes}
+    return dict.fromkeys(episodes, stats)


 def load_image_as_numpy(
@@ -390,6 +387,52 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
    return datasets.Features(hf_features)


+def _validate_feature_names(features: dict[str, dict]) -> None:
+    invalid_features = {name: ft for name, ft in features.items() if "/" in name}
+    if invalid_features:
+        raise ValueError(f"Feature names should not contain '/'. Found '/' in '{invalid_features}'.")
+
+
+def hw_to_dataset_features(
+    hw_features: dict[str, type | tuple], prefix: str, use_video: bool = True
+) -> dict[str, dict]:
+    features = {}
+    joint_fts = {key: ftype for key, ftype in hw_features.items() if ftype is float}
+    cam_fts = {key: shape for key, shape in hw_features.items() if isinstance(shape, tuple)}
+
+    if joint_fts:
+        features[f"{prefix}.joints"] = {
+            "dtype": "float32",
+            "shape": (len(joint_fts),),
+            "names": list(joint_fts),
+        }
+
+    for key, shape in cam_fts.items():
+        features[f"{prefix}.cameras.{key}"] = {
+            "dtype": "video" if use_video else "image",
+            "shape": shape,
+            "names": ["height", "width", "channels"],
+        }
+
+    _validate_feature_names(features)
+    return features
+
+
+def build_dataset_frame(
+    ds_features: dict[str, dict], values: dict[str, Any], prefix: str
+) -> dict[str, np.ndarray]:
+    frame = {}
+    for key, ft in ds_features.items():
+        if key in DEFAULT_FEATURES or not key.startswith(prefix):
+            continue
+        elif ft["dtype"] == "float32" and len(ft["shape"]) == 1:
+            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}.cameras.")]
+
+    return frame
+
+
 def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
    camera_ft = {}
    if robot.cameras:
@@ -412,7 +455,7 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea

            names = ft["names"]
            # Backward compatibility for "channel" which is an error introduced in LeRobotDataset v2.0 for ported datasets.
-            if names is not None and names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
+            if names[2] in ["channel", "channels"]:  # (h, w, c) -> (c, h, w)
                shape = (shape[2], shape[0], shape[1])
        elif key == "observation.environment_state":
            type = FeatureType.ENV
@@ -434,9 +477,9 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
 def create_empty_dataset_info(
    codebase_version: str,
    fps: int,
-    robot_type: str,
    features: dict,
    use_videos: bool,
+    robot_type: str | None = None,
 ) -> dict:
    return {
        "codebase_version": codebase_version,
@@ -543,10 +586,7 @@ def check_timestamps_sync(


 def check_delta_timestamps(
-    delta_timestamps: dict[str, list[float]],
-    fps: int,
-    tolerance_s: float,
-    raise_value_error: bool = True,
+    delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
 ) -> bool:
    """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
    This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be
@@ -705,16 +745,12 @@ class IterableNamespace(SimpleNamespace):


 def validate_frame(frame: dict, features: dict):
-    optional_features = {"timestamp"}
-    expected_features = (set(features) - set(DEFAULT_FEATURES.keys())) | {"task"}
-    actual_features = set(frame.keys())
+    expected_features = set(features) - set(DEFAULT_FEATURES)
+    actual_features = set(frame)

-    error_message = validate_features_presence(actual_features, expected_features, optional_features)
+    error_message = validate_features_presence(actual_features, expected_features)

-    if "task" in frame:
-        error_message += validate_feature_string("task", frame["task"])
-
-    common_features = actual_features & (expected_features | optional_features)
+    common_features = actual_features & expected_features
    for name in common_features - {"task"}:
        error_message += validate_feature_dtype_and_shape(name, features[name], frame[name])

@@ -722,12 +758,10 @@ def validate_frame(frame: dict, features: dict):
        raise ValueError(error_message)


-def validate_features_presence(
-    actual_features: set[str], expected_features: set[str], optional_features: set[str]
-):
+def validate_features_presence(actual_features: set[str], expected_features: set[str]):
    error_message = ""
    missing_features = expected_features - actual_features
-    extra_features = actual_features - (expected_features | optional_features)
+    extra_features = actual_features - expected_features

    if missing_features or extra_features:
        error_message += "Feature mismatch in `frame` dictionary:\n"
--- a/lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py
+++ b/lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py
@@ -27,7 +27,7 @@ from textwrap import dedent

 from lerobot import available_datasets
 from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset
-from lerobot.common.robot_devices.robots.configs import AlohaRobotConfig
+from lerobot.common.robots.aloha.configuration_aloha import AlohaRobotConfig

 LOCAL_DIR = Path("data/")

@@ -118,10 +118,7 @@ DATASETS = {
        "single_task": "Place the battery into the slot of the remote controller.",
        **ALOHA_STATIC_INFO,
    },
-    "aloha_static_candy": {
-        "single_task": "Pick up the candy and unwrap it.",
-        **ALOHA_STATIC_INFO,
-    },
+    "aloha_static_candy": {"single_task": "Pick up the candy and unwrap it.", **ALOHA_STATIC_INFO},
    "aloha_static_coffee": {
        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray, then push the 'Hot Water' and 'Travel Mug' buttons.",
        **ALOHA_STATIC_INFO,
@@ -170,22 +167,13 @@ DATASETS = {
        "single_task": "Pick up the plastic cup with the left arm, then pop its lid open with the right arm.",
        **ALOHA_STATIC_INFO,
    },
-    "aloha_static_ziploc_slide": {
-        "single_task": "Slide open the ziploc bag.",
-        **ALOHA_STATIC_INFO,
-    },
-    "aloha_sim_insertion_scripted": {
-        "single_task": "Insert the peg into the socket.",
-        **ALOHA_STATIC_INFO,
-    },
+    "aloha_static_ziploc_slide": {"single_task": "Slide open the ziploc bag.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
    "aloha_sim_insertion_scripted_image": {
        "single_task": "Insert the peg into the socket.",
        **ALOHA_STATIC_INFO,
    },
-    "aloha_sim_insertion_human": {
-        "single_task": "Insert the peg into the socket.",
-        **ALOHA_STATIC_INFO,
-    },
+    "aloha_sim_insertion_human": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
    "aloha_sim_insertion_human_image": {
        "single_task": "Insert the peg into the socket.",
        **ALOHA_STATIC_INFO,
@@ -206,19 +194,10 @@ DATASETS = {
        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
        **ALOHA_STATIC_INFO,
    },
-    "pusht": {
-        "single_task": "Push the T-shaped block onto the T-shaped target.",
-        **PUSHT_INFO,
-    },
-    "pusht_image": {
-        "single_task": "Push the T-shaped block onto the T-shaped target.",
-        **PUSHT_INFO,
-    },
+    "pusht": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "pusht_image": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
    "unitreeh1_fold_clothes": {"single_task": "Fold the sweatshirt.", **UNITREEH_INFO},
-    "unitreeh1_rearrange_objects": {
-        "single_task": "Put the object into the bin.",
-        **UNITREEH_INFO,
-    },
+    "unitreeh1_rearrange_objects": {"single_task": "Put the object into the bin.", **UNITREEH_INFO},
    "unitreeh1_two_robot_greeting": {
        "single_task": "Greet the other robot with a high five.",
        **UNITREEH_INFO,
@@ -228,31 +207,13 @@ DATASETS = {
        **UNITREEH_INFO,
    },
    "xarm_lift_medium": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
-    "xarm_lift_medium_image": {
-        "single_task": "Pick up the cube and lift it.",
-        **XARM_INFO,
-    },
-    "xarm_lift_medium_replay": {
-        "single_task": "Pick up the cube and lift it.",
-        **XARM_INFO,
-    },
-    "xarm_lift_medium_replay_image": {
-        "single_task": "Pick up the cube and lift it.",
-        **XARM_INFO,
-    },
+    "xarm_lift_medium_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
    "xarm_push_medium": {"single_task": "Push the cube onto the target.", **XARM_INFO},
-    "xarm_push_medium_image": {
-        "single_task": "Push the cube onto the target.",
-        **XARM_INFO,
-    },
-    "xarm_push_medium_replay": {
-        "single_task": "Push the cube onto the target.",
-        **XARM_INFO,
-    },
-    "xarm_push_medium_replay_image": {
-        "single_task": "Push the cube onto the target.",
-        **XARM_INFO,
-    },
+    "xarm_push_medium_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
    "umi_cup_in_the_wild": {
        "single_task": "Put the cup on the plate.",
        "license": "apache-2.0",
--- a/lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py
+++ b/lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py
@@ -141,8 +141,8 @@ from lerobot.common.datasets.video_utils import (
    get_image_pixel_channels,
    get_video_info,
 )
-from lerobot.common.robot_devices.robots.configs import RobotConfig
-from lerobot.common.robot_devices.robots.utils import make_robot_config
+from lerobot.common.robots import RobotConfig
+from lerobot.common.robots.utils import make_robot_config

 V16 = "v1.6"
 V20 = "v2.0"
@@ -379,12 +379,7 @@ def fix_lfs_video_files_tracking(work_dir: Path, lfs_untracked_videos: list[str]
    for i in range(0, len(lfs_untracked_videos), 100):
        files = lfs_untracked_videos[i : i + 100]
        try:
-            subprocess.run(
-                ["git", "rm", "--cached", *files],
-                cwd=work_dir,
-                capture_output=True,
-                check=True,
-            )
+            subprocess.run(["git", "rm", "--cached", *files], cwd=work_dir, capture_output=True, check=True)
        except subprocess.CalledProcessError as e:
            print("git rm --cached ERROR:")
            print(e.stderr)
@@ -407,17 +402,7 @@ def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:
    repo_url = f"https://huggingface.co/datasets/{repo_id}"
    env = {"GIT_LFS_SKIP_SMUDGE": "1"}  # Prevent downloading LFS files
    subprocess.run(
-        [
-            "git",
-            "clone",
-            "--branch",
-            branch,
-            "--single-branch",
-            "--depth",
-            "1",
-            repo_url,
-            str(work_dir),
-        ],
+        ["git", "clone", "--branch", branch, "--single-branch", "--depth", "1", repo_url, str(work_dir)],
        check=True,
        env=env,
    )
@@ -425,11 +410,7 @@ def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:

 def _get_lfs_untracked_videos(work_dir: Path, video_files: list[str]) -> list[str]:
    lfs_tracked_files = subprocess.run(
-        ["git", "lfs", "ls-files", "-n"],
-        cwd=work_dir,
-        capture_output=True,
-        text=True,
-        check=True,
+        ["git", "lfs", "ls-files", "-n"], cwd=work_dir, capture_output=True, text=True, check=True
    )
    lfs_tracked_files = set(lfs_tracked_files.stdout.splitlines())
    return [f for f in video_files if f not in lfs_tracked_files]
@@ -443,11 +424,7 @@ def get_videos_info(repo_id: str, local_dir: Path, video_keys: list[str], branch
    ]
    hub_api = HfApi()
    hub_api.snapshot_download(
-        repo_id=repo_id,
-        repo_type="dataset",
-        local_dir=local_dir,
-        revision=branch,
-        allow_patterns=video_files,
+        repo_id=repo_id, repo_type="dataset", local_dir=local_dir, revision=branch, allow_patterns=video_files
    )
    videos_info_dict = {}
    for vid_key, vid_path in zip(video_keys, video_files, strict=True):
@@ -474,11 +451,7 @@ def convert_dataset(

    hub_api = HfApi()
    hub_api.snapshot_download(
-        repo_id=repo_id,
-        repo_type="dataset",
-        revision=v1,
-        local_dir=v1x_dir,
-        ignore_patterns="videos*/",
+        repo_id=repo_id, repo_type="dataset", revision=v1, local_dir=v1x_dir, ignore_patterns="videos*/"
    )
    branch = "main"
    if test_branch:
@@ -508,7 +481,7 @@ def convert_dataset(

    # Tasks
    if single_task:
-        tasks_by_episodes = {ep_idx: single_task for ep_idx in episode_indices}
+        tasks_by_episodes = dict.fromkeys(episode_indices, single_task)
        dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
        tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
    elif tasks_path:
@@ -536,21 +509,12 @@ def convert_dataset(
        dataset = dataset.remove_columns(video_keys)
        clean_gitattr = Path(
            hub_api.hf_hub_download(
-                repo_id=GITATTRIBUTES_REF,
-                repo_type="dataset",
-                local_dir=local_dir,
-                filename=".gitattributes",
+                repo_id=GITATTRIBUTES_REF, repo_type="dataset", local_dir=local_dir, filename=".gitattributes"
            )
        ).absolute()
        with tempfile.TemporaryDirectory() as tmp_video_dir:
            move_videos(
-                repo_id,
-                video_keys,
-                total_episodes,
-                total_chunks,
-                Path(tmp_video_dir),
-                clean_gitattr,
-                branch,
+                repo_id, video_keys, total_episodes, total_chunks, Path(tmp_video_dir), clean_gitattr, branch
            )
        videos_info = get_videos_info(repo_id, v1x_dir, video_keys=video_keys, branch=branch)
        for key in video_keys:
@@ -579,11 +543,7 @@ def convert_dataset(

    # Episodes
    episodes = [
-        {
-            "episode_index": ep_idx,
-            "tasks": tasks_by_episodes[ep_idx],
-            "length": episode_lengths[ep_idx],
-        }
+        {"episode_index": ep_idx, "tasks": tasks_by_episodes[ep_idx], "length": episode_lengths[ep_idx]}
        for ep_idx in episode_indices
    ]
    write_jsonlines(episodes, v20_dir / EPISODES_PATH)
@@ -612,12 +572,7 @@ def convert_dataset(
        hub_api.delete_folder(repo_id=repo_id, path_in_repo="data", repo_type="dataset", revision=branch)

    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
-        hub_api.delete_folder(
-            repo_id=repo_id,
-            path_in_repo="meta_data",
-            repo_type="dataset",
-            revision=branch,
-        )
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta_data", repo_type="dataset", revision=branch)

    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta", repo_type="dataset", revision=branch)
--- a/lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py
+++ b/lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py
@@ -37,16 +37,8 @@ import logging
 from huggingface_hub import HfApi

 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.common.datasets.utils import (
-    EPISODES_STATS_PATH,
-    STATS_PATH,
-    load_stats,
-    write_info,
-)
-from lerobot.common.datasets.v21.convert_stats import (
-    check_aggregate_stats,
-    convert_stats,
-)
+from lerobot.common.datasets.utils import EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
+from lerobot.common.datasets.v21.convert_stats import check_aggregate_stats, convert_stats

 V20 = "v2.0"
 V21 = "v2.1"
@@ -87,16 +79,10 @@ def convert_dataset(

    hub_api = HfApi()
    if hub_api.file_exists(
-        repo_id=dataset.repo_id,
-        filename=STATS_PATH,
-        revision=branch,
-        repo_type="dataset",
+        repo_id=dataset.repo_id, filename=STATS_PATH, revision=branch, repo_type="dataset"
    ):
        hub_api.delete_file(
-            path_in_repo=STATS_PATH,
-            repo_id=dataset.repo_id,
-            revision=branch,
-            repo_type="dataset",
+            path_in_repo=STATS_PATH, repo_id=dataset.repo_id, revision=branch, repo_type="dataset"
        )

    hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
--- a/lerobot/common/datasets/v21/convert_stats.py
+++ b/lerobot/common/datasets/v21/convert_stats.py
@@ -17,11 +17,7 @@ from concurrent.futures import ThreadPoolExecutor, as_completed
 import numpy as np
 from tqdm import tqdm

-from lerobot.common.datasets.compute_stats import (
-    aggregate_stats,
-    get_feature_stats,
-    sample_indices,
-)
+from lerobot.common.datasets.compute_stats import aggregate_stats, get_feature_stats, sample_indices
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.datasets.utils import write_episode_stats

@@ -99,9 +95,5 @@ def check_aggregate_stats(
            if key in reference_stats and stat in reference_stats[key]:
                err_msg = f"feature='{key}' stats='{stat}'"
                np.testing.assert_allclose(
-                    val,
-                    reference_stats[key][stat],
-                    rtol=rtol,
-                    atol=atol,
-                    err_msg=err_msg,
+                    val, reference_stats[key][stat], rtol=rtol, atol=atol, err_msg=err_msg
                )
--- a/lerobot/common/datasets/video_utils.py
+++ b/lerobot/common/datasets/video_utils.py
@@ -13,16 +13,15 @@
 # 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 glob
 import importlib
-import json
 import logging
-import subprocess
 import warnings
-from collections import OrderedDict
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, ClassVar

+import av
 import pyarrow as pa
 import torch
 import torchvision
@@ -252,51 +251,83 @@ def encode_video_frames(
    g: int | None = 2,
    crf: int | None = 30,
    fast_decode: int = 0,
-    log_level: str | None = "error",
+    log_level: int | None = av.logging.ERROR,
    overwrite: bool = False,
 ) -> None:
    """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
+    # 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)
-    video_path.parent.mkdir(parents=True, exist_ok=True)

-    ffmpeg_args = OrderedDict(
-        [
-            ("-f", "image2"),
-            ("-r", str(fps)),
-            ("-i", str(imgs_dir / "frame_%06d.png")),
-            ("-vcodec", vcodec),
-            ("-pix_fmt", pix_fmt),
-        ]
+    video_path.parent.mkdir(parents=True, exist_ok=overwrite)
+
+    # Encoders/pixel formats incompatibility check
+    if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
+        logging.warning(
+            f"Incompatible pixel format 'yuv444p' for codec {vcodec}, auto-selecting format 'yuv420p'"
+        )
+        pix_fmt = "yuv420p"
+
+    # Get input frames
+    template = "frame_" + ("[0-9]" * 6) + ".png"
+    input_list = sorted(
+        glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("_")[-1].split(".")[0])
    )

+    # Define video output frame size (assuming all input frames are the same size)
+    if len(input_list) == 0:
+        raise FileNotFoundError(f"No images found in {imgs_dir}.")
+    dummy_image = Image.open(input_list[0])
+    width, height = dummy_image.size
+
+    # Define video codec options
+    video_options = {}
+
    if g is not None:
-        ffmpeg_args["-g"] = str(g)
+        video_options["g"] = str(g)

    if crf is not None:
-        ffmpeg_args["-crf"] = str(crf)
+        video_options["crf"] = str(crf)

    if fast_decode:
-        key = "-svtav1-params" if vcodec == "libsvtav1" else "-tune"
+        key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
-        ffmpeg_args[key] = value
+        video_options[key] = value

+    # Set logging level
    if log_level is not None:
-        ffmpeg_args["-loglevel"] = str(log_level)
+        # "While less efficient, it is generally preferable to modify logging with Python’s logging"
+        logging.getLogger("libav").setLevel(log_level)

-    ffmpeg_args = [item for pair in ffmpeg_args.items() for item in pair]
-    if overwrite:
-        ffmpeg_args.append("-y")
+    # Create and open output file (overwrite by default)
+    with av.open(str(video_path), "w") as output:
+        output_stream = output.add_stream(vcodec, fps, options=video_options)
+        output_stream.pix_fmt = pix_fmt
+        output_stream.width = width
+        output_stream.height = height

-    ffmpeg_cmd = ["ffmpeg"] + ffmpeg_args + [str(video_path)]
-    # redirect stdin to subprocess.DEVNULL to prevent reading random keyboard inputs from terminal
-    subprocess.run(ffmpeg_cmd, check=True, stdin=subprocess.DEVNULL)
+        # Loop through input frames and encode them
+        for input_data in input_list:
+            input_image = Image.open(input_data).convert("RGB")
+            input_frame = av.VideoFrame.from_image(input_image)
+            packet = output_stream.encode(input_frame)
+            if packet:
+                output.mux(packet)
+
+        # Flush the encoder
+        packet = output_stream.encode()
+        if packet:
+            output.mux(packet)
+
+    # Reset logging level
+    if log_level is not None:
+        av.logging.restore_default_callback()

    if not video_path.exists():
-        raise OSError(
-            f"Video encoding did not work. File not found: {video_path}. "
-            f"Try running the command manually to debug: `{''.join(ffmpeg_cmd)}`"
-        )
+        raise OSError(f"Video encoding did not work. File not found: {video_path}.")


@dataclass
@@ -332,78 +363,68 @@ with warnings.catch_warnings():


 def get_audio_info(video_path: Path | str) -> dict:
-    ffprobe_audio_cmd = [
-        "ffprobe",
-        "-v",
-        "error",
-        "-select_streams",
-        "a:0",
-        "-show_entries",
-        "stream=channels,codec_name,bit_rate,sample_rate,bit_depth,channel_layout,duration",
-        "-of",
-        "json",
-        str(video_path),
-    ]
-    result = subprocess.run(ffprobe_audio_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
-    if result.returncode != 0:
-        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+    # Set logging level
+    logging.getLogger("libav").setLevel(av.logging.ERROR)

-    info = json.loads(result.stdout)
-    audio_stream_info = info["streams"][0] if info.get("streams") else None
-    if audio_stream_info is None:
-        return {"has_audio": False}
+    # 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}

-    # Return the information, defaulting to None if no audio stream is present
-    return {
-        "has_audio": True,
-        "audio.channels": audio_stream_info.get("channels", None),
-        "audio.codec": audio_stream_info.get("codec_name", None),
-        "audio.bit_rate": int(audio_stream_info["bit_rate"]) if audio_stream_info.get("bit_rate") else None,
-        "audio.sample_rate": int(audio_stream_info["sample_rate"])
-        if audio_stream_info.get("sample_rate")
-        else None,
-        "audio.bit_depth": audio_stream_info.get("bit_depth", None),
-        "audio.channel_layout": audio_stream_info.get("channel_layout", None),
-    }
+        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:
-    ffprobe_video_cmd = [
-        "ffprobe",
-        "-v",
-        "error",
-        "-select_streams",
-        "v:0",
-        "-show_entries",
-        "stream=r_frame_rate,width,height,codec_name,nb_frames,duration,pix_fmt",
-        "-of",
-        "json",
-        str(video_path),
-    ]
-    result = subprocess.run(ffprobe_video_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
-    if result.returncode != 0:
-        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+    # Set logging level
+    logging.getLogger("libav").setLevel(av.logging.ERROR)

-    info = json.loads(result.stdout)
-    video_stream_info = info["streams"][0]
+    # Getting video stream information
+    video_info = {}
+    with av.open(str(video_path), "r") as video_file:
+        try:
+            video_stream = video_file.streams.video[0]
+        except IndexError:
+            # Reset logging level
+            av.logging.restore_default_callback()
+            return {}

-    # Calculate fps from r_frame_rate
-    r_frame_rate = video_stream_info["r_frame_rate"]
-    num, denom = map(int, r_frame_rate.split("/"))
-    fps = num / denom
+        video_info["video.height"] = video_stream.height
+        video_info["video.width"] = video_stream.width
+        video_info["video.codec"] = video_stream.codec.canonical_name
+        video_info["video.pix_fmt"] = video_stream.pix_fmt
+        video_info["video.is_depth_map"] = False

-    pixel_channels = get_video_pixel_channels(video_stream_info["pix_fmt"])
+        # Calculate fps from r_frame_rate
+        video_info["video.fps"] = int(video_stream.base_rate)

-    video_info = {
-        "video.fps": fps,
-        "video.height": video_stream_info["height"],
-        "video.width": video_stream_info["width"],
-        "video.channels": pixel_channels,
-        "video.codec": video_stream_info["codec_name"],
-        "video.pix_fmt": video_stream_info["pix_fmt"],
-        "video.is_depth_map": False,
-        **get_audio_info(video_path),
-    }
+        pixel_channels = get_video_pixel_channels(video_stream.pix_fmt)
+        video_info["video.channels"] = pixel_channels
+
+    # Reset logging level
+    av.logging.restore_default_callback()
+
+    # Adding audio stream information
+    video_info.update(**get_audio_info(video_path))

    return video_info

--- a/lerobot/common/envs/configs.py
+++ b/lerobot/common/envs/configs.py
@@ -14,12 +14,10 @@

 import abc
 from dataclasses import dataclass, field
-from typing import Any, Dict, Optional, Tuple

 import draccus

-from lerobot.common.constants import ACTION, OBS_ENV, OBS_IMAGE, OBS_IMAGES, OBS_ROBOT
-from lerobot.common.robot_devices.robots.configs import RobotConfig
+from lerobot.common.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 from lerobot.configs.types import FeatureType, PolicyFeature


@@ -34,7 +32,8 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
    def type(self) -> str:
        return self.get_choice_name(self.__class__)

-    @abc.abstractproperty
+    @property
+    @abc.abstractmethod
    def gym_kwargs(self) -> dict:
        raise NotImplementedError()

@@ -55,7 +54,7 @@ class AlohaEnv(EnvConfig):
    features_map: dict[str, str] = field(
        default_factory=lambda: {
            "action": ACTION,
-            "agent_pos": OBS_ROBOT,
+            "agent_pos": OBS_STATE,
            "top": f"{OBS_IMAGE}.top",
            "pixels/top": f"{OBS_IMAGES}.top",
        }
@@ -96,8 +95,8 @@ class PushtEnv(EnvConfig):
    features_map: dict[str, str] = field(
        default_factory=lambda: {
            "action": ACTION,
-            "agent_pos": OBS_ROBOT,
-            "environment_state": OBS_ENV,
+            "agent_pos": OBS_STATE,
+            "environment_state": OBS_ENV_STATE,
            "pixels": OBS_IMAGE,
        }
    )
@@ -138,7 +137,7 @@ class XarmEnv(EnvConfig):
    features_map: dict[str, str] = field(
        default_factory=lambda: {
            "action": ACTION,
-            "agent_pos": OBS_ROBOT,
+            "agent_pos": OBS_STATE,
            "pixels": OBS_IMAGE,
        }
    )
@@ -156,135 +155,3 @@ class XarmEnv(EnvConfig):
            "visualization_height": self.visualization_height,
            "max_episode_steps": self.episode_length,
        }
-
-
-@dataclass
-class VideoRecordConfig:
-    """Configuration for video recording in ManiSkill environments."""
-
-    enabled: bool = False
-    record_dir: str = "videos"
-    trajectory_name: str = "trajectory"
-
-
-@dataclass
-class WrapperConfig:
-    """Configuration for environment wrappers."""
-
-    joint_masking_action_space: list[bool] | None = None
-
-
-@dataclass
-class EEActionSpaceConfig:
-    """Configuration parameters for end-effector action space."""
-
-    x_step_size: float
-    y_step_size: float
-    z_step_size: float
-    bounds: Dict[str, Any]  # Contains 'min' and 'max' keys with position bounds
-    use_gamepad: bool = False
-
-
-@dataclass
-class EnvWrapperConfig:
-    """Configuration for environment wrappers."""
-
-    display_cameras: bool = False
-    use_relative_joint_positions: bool = True
-    add_joint_velocity_to_observation: bool = False
-    add_ee_pose_to_observation: bool = False
-    crop_params_dict: Optional[Dict[str, Tuple[int, int, int, int]]] = None
-    resize_size: Optional[Tuple[int, int]] = None
-    control_time_s: float = 20.0
-    fixed_reset_joint_positions: Optional[Any] = None
-    reset_time_s: float = 5.0
-    joint_masking_action_space: Optional[Any] = None
-    ee_action_space_params: Optional[EEActionSpaceConfig] = None
-    use_gripper: bool = False
-    gripper_quantization_threshold: float | None = None
-    gripper_penalty: float  = 0.0
-    gripper_penalty_in_reward: bool = False
-    open_gripper_on_reset: bool = False
-
-
-@EnvConfig.register_subclass(name="gym_manipulator")
-@dataclass
-class HILSerlRobotEnvConfig(EnvConfig):
-    """Configuration for the HILSerlRobotEnv environment."""
-
-    robot: Optional[RobotConfig] = None
-    wrapper: Optional[EnvWrapperConfig] = None
-    fps: int = 10
-    name: str = "real_robot"
-    mode: str = None  # Either "record", "replay", None
-    repo_id: Optional[str] = None
-    dataset_root: Optional[str] = None
-    task: str = ""
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: Optional[str] = None
-    reward_classifier: dict[str, str | None] = field(
-        default_factory=lambda: {
-            "pretrained_path": None,
-            "config_path": None,
-        }
-    )
-
-    def gym_kwargs(self) -> dict:
-        return {}
-
-
-@EnvConfig.register_subclass("maniskill_push")
-@dataclass
-class ManiskillEnvConfig(EnvConfig):
-    """Configuration for the ManiSkill environment."""
-
-    name: str = "maniskill/pushcube"
-    task: str = "PushCube-v1"
-    image_size: int = 64
-    control_mode: str = "pd_ee_delta_pose"
-    state_dim: int = 25
-    action_dim: int = 7
-    fps: int = 200
-    episode_length: int = 50
-    obs_type: str = "rgb"
-    render_mode: str = "rgb_array"
-    render_size: int = 64
-    device: str = "cuda"
-    robot: str = "so100"  # This is a hack to make the robot config work
-    video_record: VideoRecordConfig = field(default_factory=VideoRecordConfig)
-    wrapper: WrapperConfig = field(default_factory=WrapperConfig)
-    mock_gripper: bool = False
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(7,)),
-            "observation.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 64, 64)),
-            "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(25,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": ACTION,
-            "observation.image": OBS_IMAGE,
-            "observation.state": OBS_ROBOT,
-        }
-    )
-    reward_classifier: dict[str, str | None] = field(
-        default_factory=lambda: {
-            "pretrained_path": None,
-            "config_path": None,
-        }
-    )
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "obs_type": self.obs_type,
-            "render_mode": self.render_mode,
-            "max_episode_steps": self.episode_length,
-            "control_mode": self.control_mode,
-            "sensor_configs": {"width": self.image_size, "height": self.image_size},
-            "num_envs": 1,
-        }
--- a/lerobot/common/envs/utils.py
+++ b/lerobot/common/envs/utils.py
@@ -13,7 +13,11 @@
 # 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 warnings
+from typing import Any
+
 import einops
+import gymnasium as gym
 import numpy as np
 import torch
 from torch import Tensor
@@ -33,35 +37,29 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
    """
    # map to expected inputs for the policy
    return_observations = {}
-    # TODO: You have to merge all tensors from agent key and extra key
-    # You don't keep sensor param key in the observation
-    # And you keep sensor data rgb
-    for key, img in observations.items():
-        if "images" not in key:
-            continue
+    if "pixels" in observations:
+        if isinstance(observations["pixels"], dict):
+            imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
+        else:
+            imgs = {"observation.image": observations["pixels"]}

-        # TODO(aliberts, rcadene): use transforms.ToTensor()?
-        if not torch.is_tensor(img):
+        for imgkey, img in imgs.items():
+            # TODO(aliberts, rcadene): use transforms.ToTensor()?
            img = torch.from_numpy(img)

-        if img.ndim == 3:
-            img = img.unsqueeze(0)
+            # sanity check that images are channel last
+            _, h, w, c = img.shape
+            assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"

-        # sanity check that images are channel last
-        _, h, w, c = img.shape
-        assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
+            # sanity check that images are uint8
+            assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"

-        # sanity check that images are uint8
-        assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
+            # convert to channel first of type float32 in range [0,1]
+            img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
+            img = img.type(torch.float32)
+            img /= 255

-        # convert to channel first of type float32 in range [0,1]
-        img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
-        img = img.type(torch.float32)
-        img /= 255
-
-        return_observations[key] = img
-        # obs state agent qpos and qvel
-        # image
+            return_observations[imgkey] = img

    if "environment_state" in observations:
        return_observations["observation.environment_state"] = torch.from_numpy(
@@ -70,8 +68,7 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten

    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
    # requirement for "agent_pos"
-    # return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
-    return_observations["observation.state"] = observations["observation.state"].float()
+    return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
    return return_observations


@@ -89,44 +86,42 @@ def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
        else:
            feature = ft

-        policy_key = env_cfg.features_map.get(key, key)
+        policy_key = env_cfg.features_map[key]
        policy_features[policy_key] = feature

    return policy_features


-def preprocess_maniskill_observation(
-    observations: dict[str, np.ndarray],
-) -> dict[str, Tensor]:
-    """Convert environment observation to LeRobot format observation.
-    Args:
-        observation: Dictionary of observation batches from a Gym vector environment.
-    Returns:
-        Dictionary of observation batches with keys renamed to LeRobot format and values as tensors.
-    """
-    # map to expected inputs for the policy
-    return_observations = {}
-    # TODO: You have to merge all tensors from agent key and extra key
-    # You don't keep sensor param key in the observation
-    # And you keep sensor data rgb
-    q_pos = observations["agent"]["qpos"]
-    q_vel = observations["agent"]["qvel"]
-    tcp_pos = observations["extra"]["tcp_pose"]
-    img = observations["sensor_data"]["base_camera"]["rgb"]
+def are_all_envs_same_type(env: gym.vector.VectorEnv) -> bool:
+    first_type = type(env.envs[0])  # Get type of first env
+    return all(type(e) is first_type for e in env.envs)  # Fast type check

-    _, h, w, c = img.shape
-    assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"

-    # sanity check that images are uint8
-    assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
+def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
+    with warnings.catch_warnings():
+        warnings.simplefilter("once", UserWarning)  # Apply filter only in this function

-    # convert to channel first of type float32 in range [0,1]
-    img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
-    img = img.type(torch.float32)
-    img /= 255
+        if not (hasattr(env.envs[0], "task_description") and hasattr(env.envs[0], "task")):
+            warnings.warn(
+                "The environment does not have 'task_description' and 'task'. Some policies require these features.",
+                UserWarning,
+                stacklevel=2,
+            )
+        if not are_all_envs_same_type(env):
+            warnings.warn(
+                "The environments have different types. Make sure you infer the right task from each environment. Empty task will be passed instead.",
+                UserWarning,
+                stacklevel=2,
+            )

-    state = torch.cat([q_pos, q_vel, tcp_pos], dim=-1)

-    return_observations["observation.image"] = img
-    return_observations["observation.state"] = state
-    return return_observations
+def add_envs_task(env: gym.vector.VectorEnv, observation: dict[str, Any]) -> dict[str, Any]:
+    """Adds task feature to the observation dict with respect to the first environment attribute."""
+    if hasattr(env.envs[0], "task_description"):
+        observation["task"] = env.call("task_description")
+    elif hasattr(env.envs[0], "task"):
+        observation["task"] = env.call("task")
+    else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
+        num_envs = observation[list(observation.keys())[0]].shape[0]
+        observation["task"] = ["" for _ in range(num_envs)]
+    return observation
--- a/lerobot/common/errors.py
+++ b/lerobot/common/errors.py
@@ -0,0 +1,43 @@
+# 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.
+
+
+class DeviceNotConnectedError(ConnectionError):
+    """Exception raised when the device is not connected."""
+
+    def __init__(self, message="This device is not connected. Try calling `connect()` first."):
+        self.message = message
+        super().__init__(self.message)
+
+
+class DeviceAlreadyConnectedError(ConnectionError):
+    """Exception raised when the device is already connected."""
+
+    def __init__(
+        self,
+        message="This device is already connected. Try not calling `connect()` twice.",
+    ):
+        self.message = message
+        super().__init__(self.message)
+
+
+class InvalidActionError(ValueError):
+    """Exception raised when an action is already invalid."""
+
+    def __init__(
+        self,
+        message="The action is invalid. Check the value follows what it is expected from the action space.",
+    ):
+        self.message = message
+        super().__init__(self.message)
--- a/lerobot/common/motors/init.py
+++ b/lerobot/common/motors/init.py
@@ -0,0 +1 @@
+from .motors_bus import Motor, MotorCalibration, MotorNormMode, MotorsBus
--- a/lerobot/common/robot_devices/motors/configs.py
+++ b/lerobot/common/robot_devices/motors/configs.py
--- a/lerobot/common/motors/dynamixel/init.py
+++ b/lerobot/common/motors/dynamixel/init.py
@@ -0,0 +1,2 @@
+from .dynamixel import DriveMode, DynamixelMotorsBus, OperatingMode, TorqueMode
+from .tables import *
--- a/lerobot/common/motors/dynamixel/dynamixel.py
+++ b/lerobot/common/motors/dynamixel/dynamixel.py
@@ -0,0 +1,259 @@
+# 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.
+
+# TODO(aliberts): Should we implement FastSyncRead/Write?
+# https://github.com/ROBOTIS-GIT/DynamixelSDK/pull/643
+# https://github.com/ROBOTIS-GIT/DynamixelSDK/releases/tag/3.8.2
+# https://emanual.robotis.com/docs/en/dxl/protocol2/#fast-sync-read-0x8a
+# -> Need to check compatibility across models
+
+import logging
+from copy import deepcopy
+from enum import Enum
+
+from lerobot.common.utils.encoding_utils import decode_twos_complement, encode_twos_complement
+
+from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from .tables import (
+    AVAILABLE_BAUDRATES,
+    MODEL_BAUDRATE_TABLE,
+    MODEL_CONTROL_TABLE,
+    MODEL_ENCODING_TABLE,
+    MODEL_NUMBER_TABLE,
+    MODEL_RESOLUTION,
+)
+
+PROTOCOL_VERSION = 2.0
+DEFAULT_BAUDRATE = 1_000_000
+DEFAULT_TIMEOUT_MS = 1000
+
+NORMALIZED_DATA = ["Goal_Position", "Present_Position"]
+CONVERT_UINT32_TO_INT32_REQUIRED = ["Goal_Position", "Present_Position"]
+
+logger = logging.getLogger(__name__)
+
+
+class OperatingMode(Enum):
+    # DYNAMIXEL only controls current(torque) regardless of speed and position. This mode is ideal for a
+    # gripper or a system that only uses current(torque) control or a system that has additional
+    # velocity/position controllers.
+    CURRENT = 0
+
+    # This mode controls velocity. This mode is identical to the Wheel Mode(endless) from existing DYNAMIXEL.
+    # This mode is ideal for wheel-type robots.
+    VELOCITY = 1
+
+    # This mode controls position. This mode is identical to the Joint Mode from existing DYNAMIXEL. Operating
+    # position range is limited by the Max Position Limit(48) and the Min Position Limit(52). This mode is
+    # ideal for articulated robots that each joint rotates less than 360 degrees.
+    POSITION = 3
+
+    # This mode controls position. This mode is identical to the Multi-turn Position Control from existing
+    # DYNAMIXEL. 512 turns are supported(-256[rev] ~ 256[rev]). This mode is ideal for multi-turn wrists or
+    # conveyer systems or a system that requires an additional reduction gear. Note that Max Position
+    # Limit(48), Min Position Limit(52) are not used on Extended Position Control Mode.
+    EXTENDED_POSITION = 4
+
+    # This mode controls both position and current(torque). Up to 512 turns are supported (-256[rev] ~
+    # 256[rev]). This mode is ideal for a system that requires both position and current control such as
+    # articulated robots or grippers.
+    CURRENT_POSITION = 5
+
+    # This mode directly controls PWM output. (Voltage Control Mode)
+    PWM = 16
+
+
+class DriveMode(Enum):
+    NON_INVERTED = 0
+    INVERTED = 1
+
+
+class TorqueMode(Enum):
+    ENABLED = 1
+    DISABLED = 0
+
+
+def _split_into_byte_chunks(value: int, length: int) -> list[int]:
+    import dynamixel_sdk as dxl
+
+    if length == 1:
+        data = [value]
+    elif length == 2:
+        data = [dxl.DXL_LOBYTE(value), dxl.DXL_HIBYTE(value)]
+    elif length == 4:
+        data = [
+            dxl.DXL_LOBYTE(dxl.DXL_LOWORD(value)),
+            dxl.DXL_HIBYTE(dxl.DXL_LOWORD(value)),
+            dxl.DXL_LOBYTE(dxl.DXL_HIWORD(value)),
+            dxl.DXL_HIBYTE(dxl.DXL_HIWORD(value)),
+        ]
+    return data
+
+
+class DynamixelMotorsBus(MotorsBus):
+    """
+    The Dynamixel implementation for a MotorsBus. It relies on the python dynamixel sdk to communicate with
+    the motors. For more info, see the Dynamixel SDK Documentation:
+    https://emanual.robotis.com/docs/en/software/dynamixel/dynamixel_sdk/sample_code/python_read_write_protocol_2_0/#python-read-write-protocol-20
+    """
+
+    available_baudrates = deepcopy(AVAILABLE_BAUDRATES)
+    default_baudrate = DEFAULT_BAUDRATE
+    default_timeout = DEFAULT_TIMEOUT_MS
+    model_baudrate_table = deepcopy(MODEL_BAUDRATE_TABLE)
+    model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
+    model_encoding_table = deepcopy(MODEL_ENCODING_TABLE)
+    model_number_table = deepcopy(MODEL_NUMBER_TABLE)
+    model_resolution_table = deepcopy(MODEL_RESOLUTION)
+    normalized_data = deepcopy(NORMALIZED_DATA)
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+    ):
+        super().__init__(port, motors, calibration)
+        import dynamixel_sdk as dxl
+
+        self.port_handler = dxl.PortHandler(self.port)
+        self.packet_handler = dxl.PacketHandler(PROTOCOL_VERSION)
+        self.sync_reader = dxl.GroupSyncRead(self.port_handler, self.packet_handler, 0, 0)
+        self.sync_writer = dxl.GroupSyncWrite(self.port_handler, self.packet_handler, 0, 0)
+        self._comm_success = dxl.COMM_SUCCESS
+        self._no_error = 0x00
+
+    def _assert_protocol_is_compatible(self, instruction_name: str) -> None:
+        pass
+
+    def _handshake(self) -> None:
+        self._assert_motors_exist()
+
+    def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
+        model = self.motors[motor].model
+        search_baudrates = (
+            [initial_baudrate] if initial_baudrate is not None else self.model_baudrate_table[model]
+        )
+
+        for baudrate in search_baudrates:
+            self.set_baudrate(baudrate)
+            id_model = self.broadcast_ping()
+            if id_model:
+                found_id, found_model = next(iter(id_model.items()))
+                expected_model_nb = self.model_number_table[model]
+                if found_model != expected_model_nb:
+                    raise RuntimeError(
+                        f"Found one motor on {baudrate=} with id={found_id} but it has a "
+                        f"model number '{found_model}' different than the one expected: '{expected_model_nb}'. "
+                        f"Make sure you are connected only connected to the '{motor}' motor (model '{model}')."
+                    )
+                return baudrate, found_id
+
+        raise RuntimeError(f"Motor '{motor}' (model '{model}') was not found. Make sure it is connected.")
+
+    def configure_motors(self) -> None:
+        # By default, Dynamixel motors have a 500µs delay response time (corresponding to a value of 250 on
+        # the 'Return_Delay_Time' address). We ensure this is reduced to the minimum of 2µs (value of 0).
+        for motor in self.motors:
+            self.write("Return_Delay_Time", motor, 0)
+
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        offsets = self.sync_read("Homing_Offset", normalize=False)
+        mins = self.sync_read("Min_Position_Limit", normalize=False)
+        maxes = self.sync_read("Max_Position_Limit", normalize=False)
+        drive_modes = self.sync_read("Drive_Mode", normalize=False)
+
+        calibration = {}
+        for motor, m in self.motors.items():
+            calibration[motor] = MotorCalibration(
+                id=m.id,
+                drive_mode=drive_modes[motor],
+                homing_offset=offsets[motor],
+                range_min=mins[motor],
+                range_max=maxes[motor],
+            )
+
+        return calibration
+
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration]) -> None:
+        for motor, calibration in calibration_dict.items():
+            self.write("Homing_Offset", motor, calibration.homing_offset)
+            self.write("Min_Position_Limit", motor, calibration.range_min)
+            self.write("Max_Position_Limit", motor, calibration.range_max)
+
+        self.calibration = calibration_dict
+
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        for motor in self._get_motors_list(motors):
+            self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
+
+    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
+        addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
+        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
+
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        for motor in self._get_motors_list(motors):
+            self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
+
+    def _encode_sign(self, data_name: str, ids_values: dict[int, int]) -> dict[int, int]:
+        for id_ in ids_values:
+            model = self._id_to_model(id_)
+            encoding_table = self.model_encoding_table.get(model)
+            if encoding_table and data_name in encoding_table:
+                n_bytes = encoding_table[data_name]
+                ids_values[id_] = encode_twos_complement(ids_values[id_], n_bytes)
+
+        return ids_values
+
+    def _decode_sign(self, data_name: str, ids_values: dict[int, int]) -> dict[int, int]:
+        for id_ in ids_values:
+            model = self._id_to_model(id_)
+            encoding_table = self.model_encoding_table.get(model)
+            if encoding_table and data_name in encoding_table:
+                n_bytes = encoding_table[data_name]
+                ids_values[id_] = decode_twos_complement(ids_values[id_], n_bytes)
+
+        return ids_values
+
+    def _get_half_turn_homings(self, positions: dict[NameOrID, Value]) -> dict[NameOrID, Value]:
+        """
+        On Dynamixel Motors:
+        Present_Position = Actual_Position + Homing_Offset
+        """
+        half_turn_homings = {}
+        for motor, pos in positions.items():
+            model = self._get_motor_model(motor)
+            max_res = self.model_resolution_table[model] - 1
+            half_turn_homings[motor] = int(max_res / 2) - pos
+
+        return half_turn_homings
+
+    def _split_into_byte_chunks(self, value: int, length: int) -> list[int]:
+        return _split_into_byte_chunks(value, length)
+
+    def broadcast_ping(self, num_retry: int = 0, raise_on_error: bool = False) -> dict[int, int] | None:
+        for n_try in range(1 + num_retry):
+            data_list, comm = self.packet_handler.broadcastPing(self.port_handler)
+            if self._is_comm_success(comm):
+                break
+            logger.debug(f"Broadcast ping failed on port '{self.port}' ({n_try=})")
+            logger.debug(self.packet_handler.getTxRxResult(comm))
+
+        if not self._is_comm_success(comm):
+            if raise_on_error:
+                raise ConnectionError(self.packet_handler.getTxRxResult(comm))
+
+            return
+
+        return {id_: data[0] for id_, data in data_list.items()}
--- a/lerobot/common/motors/dynamixel/tables.py
+++ b/lerobot/common/motors/dynamixel/tables.py
@@ -0,0 +1,197 @@
+# 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.
+
+# TODO(Steven): Consider doing the following:
+# from enum import Enum
+# class MyControlTableKey(Enum):
+#   ID = "ID"
+#   GOAL_SPEED = "Goal_Speed"
+#   ...
+#
+# MY_CONTROL_TABLE ={
+#   MyControlTableKey.ID.value: (5,1)
+#   MyControlTableKey.GOAL_SPEED.value: (46, 2)
+#   ...
+# }
+# This allows me do to:
+# bus.write(MyControlTableKey.GOAL_SPEED, ...)
+# Instead of:
+# bus.write("Goal_Speed", ...)
+# This is important for two reasons:
+# 1. The linter will tell me if I'm trying to use an invalid key, instead of me realizing when I get the RunTimeError
+# 2. We can change the value of the MyControlTableKey enums without impacting the client code
+
+
+# {data_name: (address, size_byte)}
+# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#control-table
+X_SERIES_CONTROL_TABLE = {
+    "Model_Number": (0, 2),
+    "Model_Information": (2, 4),
+    "Firmware_Version": (6, 1),
+    "ID": (7, 1),
+    "Baud_Rate": (8, 1),
+    "Return_Delay_Time": (9, 1),
+    "Drive_Mode": (10, 1),
+    "Operating_Mode": (11, 1),
+    "Secondary_ID": (12, 1),
+    "Protocol_Type": (13, 1),
+    "Homing_Offset": (20, 4),
+    "Moving_Threshold": (24, 4),
+    "Temperature_Limit": (31, 1),
+    "Max_Voltage_Limit": (32, 2),
+    "Min_Voltage_Limit": (34, 2),
+    "PWM_Limit": (36, 2),
+    "Current_Limit": (38, 2),
+    "Acceleration_Limit": (40, 4),
+    "Velocity_Limit": (44, 4),
+    "Max_Position_Limit": (48, 4),
+    "Min_Position_Limit": (52, 4),
+    "Shutdown": (63, 1),
+    "Torque_Enable": (64, 1),
+    "LED": (65, 1),
+    "Status_Return_Level": (68, 1),
+    "Registered_Instruction": (69, 1),
+    "Hardware_Error_Status": (70, 1),
+    "Velocity_I_Gain": (76, 2),
+    "Velocity_P_Gain": (78, 2),
+    "Position_D_Gain": (80, 2),
+    "Position_I_Gain": (82, 2),
+    "Position_P_Gain": (84, 2),
+    "Feedforward_2nd_Gain": (88, 2),
+    "Feedforward_1st_Gain": (90, 2),
+    "Bus_Watchdog": (98, 1),
+    "Goal_PWM": (100, 2),
+    "Goal_Current": (102, 2),
+    "Goal_Velocity": (104, 4),
+    "Profile_Acceleration": (108, 4),
+    "Profile_Velocity": (112, 4),
+    "Goal_Position": (116, 4),
+    "Realtime_Tick": (120, 2),
+    "Moving": (122, 1),
+    "Moving_Status": (123, 1),
+    "Present_PWM": (124, 2),
+    "Present_Current": (126, 2),
+    "Present_Velocity": (128, 4),
+    "Present_Position": (132, 4),
+    "Velocity_Trajectory": (136, 4),
+    "Position_Trajectory": (140, 4),
+    "Present_Input_Voltage": (144, 2),
+    "Present_Temperature": (146, 1),
+}
+
+# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#baud-rate8
+X_SERIES_BAUDRATE_TABLE = {
+    9_600: 0,
+    57_600: 1,
+    115_200: 2,
+    1_000_000: 3,
+    2_000_000: 4,
+    3_000_000: 5,
+    4_000_000: 6,
+}
+
+# {data_name: size_byte}
+X_SERIES_ENCODINGS_TABLE = {
+    "Homing_Offset": X_SERIES_CONTROL_TABLE["Homing_Offset"][1],
+    "Goal_PWM": X_SERIES_CONTROL_TABLE["Goal_PWM"][1],
+    "Goal_Current": X_SERIES_CONTROL_TABLE["Goal_Current"][1],
+    "Goal_Velocity": X_SERIES_CONTROL_TABLE["Goal_Velocity"][1],
+    "Present_PWM": X_SERIES_CONTROL_TABLE["Present_PWM"][1],
+    "Present_Current": X_SERIES_CONTROL_TABLE["Present_Current"][1],
+    "Present_Velocity": X_SERIES_CONTROL_TABLE["Present_Velocity"][1],
+}
+
+MODEL_ENCODING_TABLE = {
+    "x_series": X_SERIES_ENCODINGS_TABLE,
+    "xl330-m077": X_SERIES_ENCODINGS_TABLE,
+    "xl330-m288": X_SERIES_ENCODINGS_TABLE,
+    "xl430-w250": X_SERIES_ENCODINGS_TABLE,
+    "xm430-w350": X_SERIES_ENCODINGS_TABLE,
+    "xm540-w270": X_SERIES_ENCODINGS_TABLE,
+    "xc430-w150": X_SERIES_ENCODINGS_TABLE,
+}
+
+# {model: model_resolution}
+# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#specifications
+MODEL_RESOLUTION = {
+    "x_series": 4096,
+    "xl330-m077": 4096,
+    "xl330-m288": 4096,
+    "xl430-w250": 4096,
+    "xm430-w350": 4096,
+    "xm540-w270": 4096,
+    "xc430-w150": 4096,
+}
+
+# {model: model_number}
+# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#control-table-of-eeprom-area
+MODEL_NUMBER_TABLE = {
+    "xl330-m077": 1190,
+    "xl330-m288": 1200,
+    "xl430-w250": 1060,
+    "xm430-w350": 1020,
+    "xm540-w270": 1120,
+    "xc430-w150": 1070,
+}
+
+# {model: available_operating_modes}
+# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#operating-mode11
+MODEL_OPERATING_MODES = {
+    "xl330-m077": [0, 1, 3, 4, 5, 16],
+    "xl330-m288": [0, 1, 3, 4, 5, 16],
+    "xl430-w250": [1, 3, 4, 16],
+    "xm430-w350": [0, 1, 3, 4, 5, 16],
+    "xm540-w270": [0, 1, 3, 4, 5, 16],
+    "xc430-w150": [1, 3, 4, 16],
+}
+
+MODEL_CONTROL_TABLE = {
+    "x_series": X_SERIES_CONTROL_TABLE,
+    "xl330-m077": X_SERIES_CONTROL_TABLE,
+    "xl330-m288": X_SERIES_CONTROL_TABLE,
+    "xl430-w250": X_SERIES_CONTROL_TABLE,
+    "xm430-w350": X_SERIES_CONTROL_TABLE,
+    "xm540-w270": X_SERIES_CONTROL_TABLE,
+    "xc430-w150": X_SERIES_CONTROL_TABLE,
+}
+
+MODEL_BAUDRATE_TABLE = {
+    "x_series": X_SERIES_BAUDRATE_TABLE,
+    "xl330-m077": X_SERIES_BAUDRATE_TABLE,
+    "xl330-m288": X_SERIES_BAUDRATE_TABLE,
+    "xl430-w250": X_SERIES_BAUDRATE_TABLE,
+    "xm430-w350": X_SERIES_BAUDRATE_TABLE,
+    "xm540-w270": X_SERIES_BAUDRATE_TABLE,
+    "xc430-w150": X_SERIES_BAUDRATE_TABLE,
+}
+
+AVAILABLE_BAUDRATES = [
+    9_600,
+    19_200,
+    38_400,
+    57_600,
+    115_200,
+    230_400,
+    460_800,
+    500_000,
+    576_000,
+    921_600,
+    1_000_000,
+    1_152_000,
+    2_000_000,
+    2_500_000,
+    3_000_000,
+    3_500_000,
+    4_000_000,
+]
--- a/lerobot/common/motors/feetech/init.py
+++ b/lerobot/common/motors/feetech/init.py
@@ -0,0 +1,2 @@
+from .feetech import DriveMode, FeetechMotorsBus, OperatingMode, TorqueMode
+from .tables import *
--- a/lerobot/common/motors/feetech/feetech.py
+++ b/lerobot/common/motors/feetech/feetech.py
@@ -0,0 +1,441 @@
+# 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 logging
+from copy import deepcopy
+from enum import Enum
+from pprint import pformat
+
+from lerobot.common.utils.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
+
+from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
+from .tables import (
+    FIRMWARE_MAJOR_VERSION,
+    FIRMWARE_MINOR_VERSION,
+    MODEL_BAUDRATE_TABLE,
+    MODEL_CONTROL_TABLE,
+    MODEL_ENCODING_TABLE,
+    MODEL_NUMBER,
+    MODEL_NUMBER_TABLE,
+    MODEL_PROTOCOL,
+    MODEL_RESOLUTION,
+    SCAN_BAUDRATES,
+)
+
+DEFAULT_PROTOCOL_VERSION = 0
+DEFAULT_BAUDRATE = 1_000_000
+DEFAULT_TIMEOUT_MS = 1000
+
+NORMALIZED_DATA = ["Goal_Position", "Present_Position"]
+
+logger = logging.getLogger(__name__)
+
+
+class OperatingMode(Enum):
+    # position servo mode
+    POSITION = 0
+    # The motor is in constant speed mode, which is controlled by parameter 0x2e, and the highest bit 15 is
+    # the direction bit
+    VELOCITY = 1
+    # PWM open-loop speed regulation mode, with parameter 0x2c running time parameter control, bit11 as
+    # direction bit
+    PWM = 2
+    # In step servo mode, the number of step progress is represented by parameter 0x2a, and the highest bit 15
+    # is the direction bit
+    STEP = 3
+
+
+class DriveMode(Enum):
+    NON_INVERTED = 0
+    INVERTED = 1
+
+
+class TorqueMode(Enum):
+    ENABLED = 1
+    DISABLED = 0
+
+
+def _split_into_byte_chunks(value: int, length: int) -> list[int]:
+    import scservo_sdk as scs
+
+    if length == 1:
+        data = [value]
+    elif length == 2:
+        data = [scs.SCS_LOBYTE(value), scs.SCS_HIBYTE(value)]
+    elif length == 4:
+        data = [
+            scs.SCS_LOBYTE(scs.SCS_LOWORD(value)),
+            scs.SCS_HIBYTE(scs.SCS_LOWORD(value)),
+            scs.SCS_LOBYTE(scs.SCS_HIWORD(value)),
+            scs.SCS_HIBYTE(scs.SCS_HIWORD(value)),
+        ]
+    return data
+
+
+def patch_setPacketTimeout(self, packet_length):  # noqa: N802
+    """
+    HACK: This patches the PortHandler behavior to set the correct packet timeouts.
+
+    It fixes https://gitee.com/ftservo/SCServoSDK/issues/IBY2S6
+    The bug is fixed on the official Feetech SDK repo (https://gitee.com/ftservo/FTServo_Python)
+    but because that version is not published on PyPI, we rely on the (unofficial) on that is, which needs
+    patching.
+    """
+    self.packet_start_time = self.getCurrentTime()
+    self.packet_timeout = (self.tx_time_per_byte * packet_length) + (self.tx_time_per_byte * 3.0) + 50
+
+
+class FeetechMotorsBus(MotorsBus):
+    """
+    The FeetechMotorsBus class allows to efficiently read and write to the attached motors. It relies on the
+    python feetech sdk to communicate with the motors, which is itself based on the dynamixel sdk.
+    """
+
+    available_baudrates = deepcopy(SCAN_BAUDRATES)
+    default_baudrate = DEFAULT_BAUDRATE
+    default_timeout = DEFAULT_TIMEOUT_MS
+    model_baudrate_table = deepcopy(MODEL_BAUDRATE_TABLE)
+    model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
+    model_encoding_table = deepcopy(MODEL_ENCODING_TABLE)
+    model_number_table = deepcopy(MODEL_NUMBER_TABLE)
+    model_resolution_table = deepcopy(MODEL_RESOLUTION)
+    normalized_data = deepcopy(NORMALIZED_DATA)
+
+    def __init__(
+        self,
+        port: str,
+        motors: dict[str, Motor],
+        calibration: dict[str, MotorCalibration] | None = None,
+        protocol_version: int = DEFAULT_PROTOCOL_VERSION,
+    ):
+        super().__init__(port, motors, calibration)
+        self.protocol_version = protocol_version
+        self._assert_same_protocol()
+        import scservo_sdk as scs
+
+        self.port_handler = scs.PortHandler(self.port)
+        # HACK: monkeypatch
+        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(
+            self.port_handler, scs.PortHandler
+        )
+        self.packet_handler = scs.PacketHandler(protocol_version)
+        self.sync_reader = scs.GroupSyncRead(self.port_handler, self.packet_handler, 0, 0)
+        self.sync_writer = scs.GroupSyncWrite(self.port_handler, self.packet_handler, 0, 0)
+        self._comm_success = scs.COMM_SUCCESS
+        self._no_error = 0x00
+
+        if any(MODEL_PROTOCOL[model] != self.protocol_version for model in self.models):
+            raise ValueError(f"Some motors are incompatible with protocol_version={self.protocol_version}")
+
+    def _assert_same_protocol(self) -> None:
+        if any(MODEL_PROTOCOL[model] != self.protocol_version for model in self.models):
+            raise RuntimeError("Some motors use an incompatible protocol.")
+
+    def _assert_protocol_is_compatible(self, instruction_name: str) -> None:
+        if instruction_name == "sync_read" and self.protocol_version == 1:
+            raise NotImplementedError(
+                "'Sync Read' is not available with Feetech motors using Protocol 1. Use 'Read' sequentially instead."
+            )
+        if instruction_name == "broadcast_ping" and self.protocol_version == 1:
+            raise NotImplementedError(
+                "'Broadcast Ping' is not available with Feetech motors using Protocol 1. Use 'Ping' sequentially instead."
+            )
+
+    def _assert_same_firmware(self) -> None:
+        firmware_versions = self._read_firmware_version(self.ids)
+        if len(set(firmware_versions.values())) != 1:
+            raise RuntimeError(
+                "Some Motors use different firmware versions. Update their firmware first using Feetech's software. "
+                "Visit https://www.feetechrc.com/software."
+            )
+
+    def _handshake(self) -> None:
+        self._assert_motors_exist()
+        self._assert_same_firmware()
+
+    def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
+        if self.protocol_version == 0:
+            return self._find_single_motor_p0(motor, initial_baudrate)
+        else:
+            return self._find_single_motor_p1(motor, initial_baudrate)
+
+    def _find_single_motor_p0(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
+        model = self.motors[motor].model
+        search_baudrates = (
+            [initial_baudrate] if initial_baudrate is not None else self.model_baudrate_table[model]
+        )
+        expected_model_nb = self.model_number_table[model]
+
+        for baudrate in search_baudrates:
+            self.set_baudrate(baudrate)
+            id_model = self.broadcast_ping()
+            if id_model:
+                found_id, found_model = next(iter(id_model.items()))
+                if found_model != expected_model_nb:
+                    raise RuntimeError(
+                        f"Found one motor on {baudrate=} with id={found_id} but it has a "
+                        f"model number '{found_model}' different than the one expected: '{expected_model_nb}'. "
+                        f"Make sure you are connected only connected to the '{motor}' motor (model '{model}')."
+                    )
+                return baudrate, found_id
+
+        raise RuntimeError(f"Motor '{motor}' (model '{model}') was not found. Make sure it is connected.")
+
+    def _find_single_motor_p1(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
+        import scservo_sdk as scs
+
+        model = self.motors[motor].model
+        search_baudrates = (
+            [initial_baudrate] if initial_baudrate is not None else self.model_baudrate_table[model]
+        )
+        expected_model_nb = self.model_number_table[model]
+
+        for baudrate in search_baudrates:
+            self.set_baudrate(baudrate)
+            for id_ in range(scs.MAX_ID + 1):
+                found_model = self.ping(id_)
+                if found_model is not None:
+                    if found_model != expected_model_nb:
+                        raise RuntimeError(
+                            f"Found one motor on {baudrate=} with id={id_} but it has a "
+                            f"model number '{found_model}' different than the one expected: '{expected_model_nb}'. "
+                            f"Make sure you are connected only connected to the '{motor}' motor (model '{model}')."
+                        )
+                    return baudrate, id_
+
+        raise RuntimeError(f"Motor '{motor}' (model '{model}') was not found. Make sure it is connected.")
+
+    def configure_motors(self) -> None:
+        for motor in self.motors:
+            # By default, Feetech motors have a 500µs delay response time (corresponding to a value of 250 on
+            # the 'Return_Delay_Time' address). We ensure this is reduced to the minimum of 2µs (value of 0).
+            self.write("Return_Delay_Time", motor, 0)
+            # Set 'Maximum_Acceleration' to 254 to speedup acceleration and deceleration of the motors.
+            # Note: this address is not in the official STS3215 Memory Table
+            self.write("Maximum_Acceleration", motor, 254)
+            self.write("Acceleration", motor, 254)
+
+    def read_calibration(self) -> dict[str, MotorCalibration]:
+        if self.protocol_version == 0:
+            offsets = self.sync_read("Homing_Offset", normalize=False)
+            mins = self.sync_read("Min_Position_Limit", normalize=False)
+            maxes = self.sync_read("Max_Position_Limit", normalize=False)
+            drive_modes = dict.fromkeys(self.motors, 0)
+        else:
+            offsets, mins, maxes, drive_modes = {}, {}, {}, {}
+            for motor in self.motors:
+                offsets[motor] = 0
+                mins[motor] = self.read("Min_Position_Limit", motor, normalize=False)
+                maxes[motor] = self.read("Max_Position_Limit", motor, normalize=False)
+                drive_modes[motor] = 0
+
+        # TODO(aliberts): add set/get_drive_mode?
+
+        calibration = {}
+        for motor, m in self.motors.items():
+            calibration[motor] = MotorCalibration(
+                id=m.id,
+                drive_mode=drive_modes[motor],
+                homing_offset=offsets[motor],
+                range_min=mins[motor],
+                range_max=maxes[motor],
+            )
+
+        return calibration
+
+    def write_calibration(self, calibration_dict: dict[str, MotorCalibration]) -> None:
+        for motor, calibration in calibration_dict.items():
+            if self.protocol_version == 0:
+                self.write("Homing_Offset", motor, calibration.homing_offset)
+            self.write("Min_Position_Limit", motor, calibration.range_min)
+            self.write("Max_Position_Limit", motor, calibration.range_max)
+
+        self.calibration = calibration_dict
+
+    def _get_half_turn_homings(self, positions: dict[NameOrID, Value]) -> dict[NameOrID, Value]:
+        """
+        On Feetech Motors:
+        Present_Position = Actual_Position - Homing_Offset
+        """
+        half_turn_homings = {}
+        for motor, pos in positions.items():
+            model = self._get_motor_model(motor)
+            max_res = self.model_resolution_table[model] - 1
+            half_turn_homings[motor] = pos - int(max_res / 2)
+
+        return half_turn_homings
+
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        for motor in self._get_motors_list(motors):
+            self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
+            self.write("Lock", motor, 0, num_retry=num_retry)
+
+    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
+        addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
+        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
+        addr, length = get_address(self.model_ctrl_table, model, "Lock")
+        self._write(addr, length, motor_id, 0, num_retry=num_retry)
+
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+        for motor in self._get_motors_list(motors):
+            self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
+            self.write("Lock", motor, 1, num_retry=num_retry)
+
+    def _encode_sign(self, data_name: str, ids_values: dict[int, int]) -> dict[int, int]:
+        for id_ in ids_values:
+            model = self._id_to_model(id_)
+            encoding_table = self.model_encoding_table.get(model)
+            if encoding_table and data_name in encoding_table:
+                sign_bit = encoding_table[data_name]
+                ids_values[id_] = encode_sign_magnitude(ids_values[id_], sign_bit)
+
+        return ids_values
+
+    def _decode_sign(self, data_name: str, ids_values: dict[int, int]) -> dict[int, int]:
+        for id_ in ids_values:
+            model = self._id_to_model(id_)
+            encoding_table = self.model_encoding_table.get(model)
+            if encoding_table and data_name in encoding_table:
+                sign_bit = encoding_table[data_name]
+                ids_values[id_] = decode_sign_magnitude(ids_values[id_], sign_bit)
+
+        return ids_values
+
+    def _split_into_byte_chunks(self, value: int, length: int) -> list[int]:
+        return _split_into_byte_chunks(value, length)
+
+    def _broadcast_ping(self) -> tuple[dict[int, int], int]:
+        import scservo_sdk as scs
+
+        data_list = {}
+
+        status_length = 6
+
+        rx_length = 0
+        wait_length = status_length * scs.MAX_ID
+
+        txpacket = [0] * 6
+
+        tx_time_per_byte = (1000.0 / self.port_handler.getBaudRate()) * 10.0
+
+        txpacket[scs.PKT_ID] = scs.BROADCAST_ID
+        txpacket[scs.PKT_LENGTH] = 2
+        txpacket[scs.PKT_INSTRUCTION] = scs.INST_PING
+
+        result = self.packet_handler.txPacket(self.port_handler, txpacket)
+        if result != scs.COMM_SUCCESS:
+            self.port_handler.is_using = False
+            return data_list, result
+
+        # set rx timeout
+        self.port_handler.setPacketTimeoutMillis((wait_length * tx_time_per_byte) + (3.0 * scs.MAX_ID) + 16.0)
+
+        rxpacket = []
+        while True:
+            rxpacket += self.port_handler.readPort(wait_length - rx_length)
+            rx_length = len(rxpacket)
+
+            if self.port_handler.isPacketTimeout():  # or rx_length >= wait_length
+                break
+
+        self.port_handler.is_using = False
+
+        if rx_length == 0:
+            return data_list, scs.COMM_RX_TIMEOUT
+
+        while True:
+            if rx_length < status_length:
+                return data_list, scs.COMM_RX_CORRUPT
+
+            # find packet header
+            for idx in range(0, (rx_length - 1)):
+                if (rxpacket[idx] == 0xFF) and (rxpacket[idx + 1] == 0xFF):
+                    break
+
+            if idx == 0:  # found at the beginning of the packet
+                # calculate checksum
+                checksum = 0
+                for idx in range(2, status_length - 1):  # except header & checksum
+                    checksum += rxpacket[idx]
+
+                checksum = ~checksum & 0xFF
+                if rxpacket[status_length - 1] == checksum:
+                    result = scs.COMM_SUCCESS
+                    data_list[rxpacket[scs.PKT_ID]] = rxpacket[scs.PKT_ERROR]
+
+                    del rxpacket[0:status_length]
+                    rx_length = rx_length - status_length
+
+                    if rx_length == 0:
+                        return data_list, result
+                else:
+                    result = scs.COMM_RX_CORRUPT
+                    # remove header (0xFF 0xFF)
+                    del rxpacket[0:2]
+                    rx_length = rx_length - 2
+            else:
+                # remove unnecessary packets
+                del rxpacket[0:idx]
+                rx_length = rx_length - idx
+
+    def broadcast_ping(self, num_retry: int = 0, raise_on_error: bool = False) -> dict[int, int] | None:
+        self._assert_protocol_is_compatible("broadcast_ping")
+        for n_try in range(1 + num_retry):
+            ids_status, comm = self._broadcast_ping()
+            if self._is_comm_success(comm):
+                break
+            logger.debug(f"Broadcast ping failed on port '{self.port}' ({n_try=})")
+            logger.debug(self.packet_handler.getTxRxResult(comm))
+
+        if not self._is_comm_success(comm):
+            if raise_on_error:
+                raise ConnectionError(self.packet_handler.getTxRxResult(comm))
+            return
+
+        ids_errors = {id_: status for id_, status in ids_status.items() if self._is_error(status)}
+        if ids_errors:
+            display_dict = {id_: self.packet_handler.getRxPacketError(err) for id_, err in ids_errors.items()}
+            logger.error(f"Some motors found returned an error status:\n{pformat(display_dict, indent=4)}")
+
+        return self._read_model_number(list(ids_status), raise_on_error)
+
+    def _read_firmware_version(self, motor_ids: list[int], raise_on_error: bool = False) -> dict[int, str]:
+        firmware_versions = {}
+        for id_ in motor_ids:
+            firm_ver_major, comm, error = self._read(
+                *FIRMWARE_MAJOR_VERSION, id_, raise_on_error=raise_on_error
+            )
+            if not self._is_comm_success(comm) or self._is_error(error):
+                return
+
+            firm_ver_minor, comm, error = self._read(
+                *FIRMWARE_MINOR_VERSION, id_, raise_on_error=raise_on_error
+            )
+            if not self._is_comm_success(comm) or self._is_error(error):
+                return
+
+            firmware_versions[id_] = f"{firm_ver_major}.{firm_ver_minor}"
+
+        return firmware_versions
+
+    def _read_model_number(self, motor_ids: list[int], raise_on_error: bool = False) -> dict[int, int]:
+        model_numbers = {}
+        for id_ in motor_ids:
+            model_nb, comm, error = self._read(*MODEL_NUMBER, id_, raise_on_error=raise_on_error)
+            if not self._is_comm_success(comm) or self._is_error(error):
+                return
+
+            model_numbers[id_] = model_nb
+
+        return model_numbers
--- a/lerobot/common/motors/feetech/tables.py
+++ b/lerobot/common/motors/feetech/tables.py
@@ -0,0 +1,251 @@
+# 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.
+
+FIRMWARE_MAJOR_VERSION = (0, 1)
+FIRMWARE_MINOR_VERSION = (1, 1)
+MODEL_NUMBER = (3, 2)
+
+# TODO(Steven): Consider doing the following:
+# from enum import Enum
+# class MyControlTableKey(Enum):
+#   ID = "ID"
+#   GOAL_SPEED = "Goal_Speed"
+#   ...
+#
+# MY_CONTROL_TABLE ={
+#   MyControlTableKey.ID.value: (5,1)
+#   MyControlTableKey.GOAL_SPEED.value: (46, 2)
+#   ...
+# }
+# This allows me do to:
+# bus.write(MyControlTableKey.GOAL_SPEED, ...)
+# Instead of:
+# bus.write("Goal_Speed", ...)
+# This is important for two reasons:
+# 1. The linter will tell me if I'm trying to use an invalid key, instead of me realizing when I get the RunTimeError
+# 2. We can change the value of the MyControlTableKey enums without impacting the client code
+
+# data_name: (address, size_byte)
+# http://doc.feetech.cn/#/prodinfodownload?srcType=FT-SMS-STS-emanual-229f4476422d4059abfb1cb0
+STS_SMS_SERIES_CONTROL_TABLE = {
+    # EPROM
+    "Firmware_Major_Version": FIRMWARE_MAJOR_VERSION,  # read-only
+    "Firmware_Minor_Version": FIRMWARE_MINOR_VERSION,  # read-only
+    "Model_Number": MODEL_NUMBER,  # read-only
+    "ID": (5, 1),
+    "Baud_Rate": (6, 1),
+    "Return_Delay_Time": (7, 1),
+    "Response_Status_Level": (8, 1),
+    "Min_Position_Limit": (9, 2),
+    "Max_Position_Limit": (11, 2),
+    "Max_Temperature_Limit": (13, 1),
+    "Max_Voltage_Limit": (14, 1),
+    "Min_Voltage_Limit": (15, 1),
+    "Max_Torque_Limit": (16, 2),
+    "Phase": (18, 1),
+    "Unloading_Condition": (19, 1),
+    "LED_Alarm_Condition": (20, 1),
+    "P_Coefficient": (21, 1),
+    "D_Coefficient": (22, 1),
+    "I_Coefficient": (23, 1),
+    "Minimum_Startup_Force": (24, 2),
+    "CW_Dead_Zone": (26, 1),
+    "CCW_Dead_Zone": (27, 1),
+    "Protection_Current": (28, 2),
+    "Angular_Resolution": (30, 1),
+    "Homing_Offset": (31, 2),
+    "Operating_Mode": (33, 1),
+    "Protective_Torque": (34, 1),
+    "Protection_Time": (35, 1),
+    "Overload_Torque": (36, 1),
+    "Velocity_closed_loop_P_proportional_coefficient": (37, 1),
+    "Over_Current_Protection_Time": (38, 1),
+    "Velocity_closed_loop_I_integral_coefficient": (39, 1),
+    # SRAM
+    "Torque_Enable": (40, 1),
+    "Acceleration": (41, 1),
+    "Goal_Position": (42, 2),
+    "Goal_Time": (44, 2),
+    "Goal_Velocity": (46, 2),
+    "Torque_Limit": (48, 2),
+    "Lock": (55, 1),
+    "Present_Position": (56, 2),  # read-only
+    "Present_Velocity": (58, 2),  # read-only
+    "Present_Load": (60, 2),  # read-only
+    "Present_Voltage": (62, 1),  # read-only
+    "Present_Temperature": (63, 1),  # read-only
+    "Status": (65, 1),  # read-only
+    "Moving": (66, 1),  # read-only
+    "Present_Current": (69, 2),  # read-only
+    "Goal_Position_2": (71, 2),  # read-only
+    # Factory
+    "Moving_Velocity": (80, 1),
+    "Moving_Velocity_Threshold": (80, 1),
+    "DTs": (81, 1),  # (ms)
+    "Velocity_Unit_factor": (82, 1),
+    "Hts": (83, 1),  # (ns) valid for firmware >= 2.54, other versions keep 0
+    "Maximum_Velocity_Limit": (84, 1),
+    "Maximum_Acceleration": (85, 1),
+    "Acceleration_Multiplier ": (86, 1),  # Acceleration multiplier in effect when acceleration is 0
+}
+
+# http://doc.feetech.cn/#/prodinfodownload?srcType=FT-SCSCL-emanual-cbcc8ab2e3384282a01d4bf3
+SCS_SERIES_CONTROL_TABLE = {
+    # EPROM
+    "Firmware_Major_Version": FIRMWARE_MAJOR_VERSION,  # read-only
+    "Firmware_Minor_Version": FIRMWARE_MINOR_VERSION,  # read-only
+    "Model_Number": MODEL_NUMBER,  # read-only
+    "ID": (5, 1),
+    "Baud_Rate": (6, 1),
+    "Return_Delay_Time": (7, 1),
+    "Response_Status_Level": (8, 1),
+    "Min_Position_Limit": (9, 2),
+    "Max_Position_Limit": (11, 2),
+    "Max_Temperature_Limit": (13, 1),
+    "Max_Voltage_Limit": (14, 1),
+    "Min_Voltage_Limit": (15, 1),
+    "Max_Torque_Limit": (16, 2),
+    "Phase": (18, 1),
+    "Unloading_Condition": (19, 1),
+    "LED_Alarm_Condition": (20, 1),
+    "P_Coefficient": (21, 1),
+    "D_Coefficient": (22, 1),
+    "I_Coefficient": (23, 1),
+    "Minimum_Startup_Force": (24, 2),
+    "CW_Dead_Zone": (26, 1),
+    "CCW_Dead_Zone": (27, 1),
+    "Protective_Torque": (37, 1),
+    "Protection_Time": (38, 1),
+    # SRAM
+    "Torque_Enable": (40, 1),
+    "Acceleration": (41, 1),
+    "Goal_Position": (42, 2),
+    "Running_Time": (44, 2),
+    "Goal_Velocity": (46, 2),
+    "Lock": (48, 1),
+    "Present_Position": (56, 2),  # read-only
+    "Present_Velocity": (58, 2),  # read-only
+    "Present_Load": (60, 2),  # read-only
+    "Present_Voltage": (62, 1),  # read-only
+    "Present_Temperature": (63, 1),  # read-only
+    "Sync_Write_Flag": (64, 1),  # read-only
+    "Status": (65, 1),  # read-only
+    "Moving": (66, 1),  # read-only
+    # Factory
+    "PWM_Maximum_Step": (78, 1),
+    "Moving_Velocity_Threshold*50": (79, 1),
+    "DTs": (80, 1),  # (ms)
+    "Minimum_Velocity_Limit*50": (81, 1),
+    "Maximum_Velocity_Limit*50": (82, 1),
+    "Acceleration_2": (83, 1),  # don't know what that is
+}
+
+STS_SMS_SERIES_BAUDRATE_TABLE = {
+    1_000_000: 0,
+    500_000: 1,
+    250_000: 2,
+    128_000: 3,
+    115_200: 4,
+    57_600: 5,
+    38_400: 6,
+    19_200: 7,
+}
+
+SCS_SERIES_BAUDRATE_TABLE = {
+    1_000_000: 0,
+    500_000: 1,
+    250_000: 2,
+    128_000: 3,
+    115_200: 4,
+    57_600: 5,
+    38_400: 6,
+    19_200: 7,
+}
+
+MODEL_CONTROL_TABLE = {
+    "sts_series": STS_SMS_SERIES_CONTROL_TABLE,
+    "scs_series": SCS_SERIES_CONTROL_TABLE,
+    "sms_series": STS_SMS_SERIES_CONTROL_TABLE,
+    "sts3215": STS_SMS_SERIES_CONTROL_TABLE,
+    "sts3250": STS_SMS_SERIES_CONTROL_TABLE,
+    "scs0009": SCS_SERIES_CONTROL_TABLE,
+    "sm8512bl": STS_SMS_SERIES_CONTROL_TABLE,
+}
+
+MODEL_RESOLUTION = {
+    "sts_series": 4096,
+    "sms_series": 4096,
+    "scs_series": 1024,
+    "sts3215": 4096,
+    "sts3250": 4096,
+    "sm8512bl": 65536,
+    "scs0009": 1024,
+}
+
+MODEL_BAUDRATE_TABLE = {
+    "sts_series": STS_SMS_SERIES_BAUDRATE_TABLE,
+    "sms_series": STS_SMS_SERIES_BAUDRATE_TABLE,
+    "scs_series": SCS_SERIES_BAUDRATE_TABLE,
+    "sm8512bl": STS_SMS_SERIES_BAUDRATE_TABLE,
+    "sts3215": STS_SMS_SERIES_BAUDRATE_TABLE,
+    "sts3250": STS_SMS_SERIES_BAUDRATE_TABLE,
+    "scs0009": SCS_SERIES_BAUDRATE_TABLE,
+}
+
+# Sign-Magnitude encoding bits
+STS_SMS_SERIES_ENCODINGS_TABLE = {
+    "Homing_Offset": 11,
+    "Goal_Velocity": 15,
+}
+
+MODEL_ENCODING_TABLE = {
+    "sts_series": STS_SMS_SERIES_ENCODINGS_TABLE,
+    "sms_series": STS_SMS_SERIES_ENCODINGS_TABLE,
+    "scs_series": {},
+    "sts3215": STS_SMS_SERIES_ENCODINGS_TABLE,
+    "sts3250": STS_SMS_SERIES_ENCODINGS_TABLE,
+    "sm8512bl": STS_SMS_SERIES_ENCODINGS_TABLE,
+    "scs0009": {},
+}
+
+SCAN_BAUDRATES = [
+    4_800,
+    9_600,
+    14_400,
+    19_200,
+    38_400,
+    57_600,
+    115_200,
+    128_000,
+    250_000,
+    500_000,
+    1_000_000,
+]
+
+MODEL_NUMBER_TABLE = {
+    "sts3215": 777,
+    "sts3250": 2825,
+    "sm8512bl": 11272,
+    "scs0009": 1284,
+}
+
+MODEL_PROTOCOL = {
+    "sts_series": 0,
+    "sms_series": 0,
+    "scs_series": 1,
+    "sts3215": 0,
+    "sts3250": 0,
+    "sm8512bl": 0,
+    "scs0009": 1,
+}
--- a/lerobot/common/motors/motors_bus.py
+++ b/lerobot/common/motors/motors_bus.py
--- a/lerobot/common/robot_devices/motors/utils.py
+++ b/lerobot/common/robot_devices/motors/utils.py
@@ -12,37 +12,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-from typing import Protocol
-
-from lerobot.common.robot_devices.motors.configs import (
-    DynamixelMotorsBusConfig,
-    FeetechMotorsBusConfig,
-    MotorsBusConfig,
-)
+from .configs import MotorsBusConfig
+from .motors_bus import MotorsBus


-class MotorsBus(Protocol):
-    def motor_names(self): ...
-    def set_calibration(self): ...
-    def apply_calibration(self): ...
-    def revert_calibration(self): ...
-    def read(self): ...
-    def write(self): ...
-
-
-def make_motors_buses_from_configs(
-    motors_bus_configs: dict[str, MotorsBusConfig],
-) -> list[MotorsBus]:
+def make_motors_buses_from_configs(motors_bus_configs: dict[str, MotorsBusConfig]) -> list[MotorsBus]:
    motors_buses = {}

    for key, cfg in motors_bus_configs.items():
        if cfg.type == "dynamixel":
-            from lerobot.common.robot_devices.motors.dynamixel import DynamixelMotorsBus
+            from .dynamixel import DynamixelMotorsBus

            motors_buses[key] = DynamixelMotorsBus(cfg)

        elif cfg.type == "feetech":
-            from lerobot.common.robot_devices.motors.feetech import FeetechMotorsBus
+            from lerobot.common.motors.feetech.feetech import FeetechMotorsBus

            motors_buses[key] = FeetechMotorsBus(cfg)

@@ -54,13 +38,16 @@ def make_motors_buses_from_configs(

 def make_motors_bus(motor_type: str, **kwargs) -> MotorsBus:
    if motor_type == "dynamixel":
-        from lerobot.common.robot_devices.motors.dynamixel import DynamixelMotorsBus
+        from .configs import DynamixelMotorsBusConfig
+        from .dynamixel import DynamixelMotorsBus

        config = DynamixelMotorsBusConfig(**kwargs)
        return DynamixelMotorsBus(config)

    elif motor_type == "feetech":
-        from lerobot.common.robot_devices.motors.feetech import FeetechMotorsBus
+        from feetech import FeetechMotorsBus
+
+        from .configs import FeetechMotorsBusConfig

        config = FeetechMotorsBusConfig(**kwargs)
        return FeetechMotorsBus(config)
--- a/lerobot/common/optim/optimizers.py
+++ b/lerobot/common/optim/optimizers.py
@@ -14,9 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import abc
-from dataclasses import asdict, dataclass, field
+from dataclasses import asdict, dataclass
 from pathlib import Path
-from typing import Any

 import draccus
 import torch
@@ -45,7 +44,7 @@ class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
        return "adam"

    @abc.abstractmethod
-    def build(self) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
+    def build(self) -> torch.optim.Optimizer:
        raise NotImplementedError


@@ -95,76 +94,7 @@ class SGDConfig(OptimizerConfig):
        return torch.optim.SGD(params, **kwargs)


-@OptimizerConfig.register_subclass("multi_adam")
-@dataclass
-class MultiAdamConfig(OptimizerConfig):
-    """Configuration for multiple Adam optimizers with different parameter groups.
-
-    This creates a dictionary of Adam optimizers, each with its own hyperparameters.
-
-    Args:
-        lr: Default learning rate (used if not specified for a group)
-        weight_decay: Default weight decay (used if not specified for a group)
-        optimizer_groups: Dictionary mapping parameter group names to their hyperparameters
-        grad_clip_norm: Gradient clipping norm
-    """
-
-    lr: float = 1e-3
-    weight_decay: float = 0.0
-    grad_clip_norm: float = 10.0
-    optimizer_groups: dict[str, dict[str, Any]] = field(default_factory=dict)
-
-    def build(self, params_dict: dict[str, list]) -> dict[str, torch.optim.Optimizer]:
-        """Build multiple Adam optimizers.
-
-        Args:
-            params_dict: Dictionary mapping parameter group names to lists of parameters
-                         The keys should match the keys in optimizer_groups
-
-        Returns:
-            Dictionary mapping parameter group names to their optimizers
-        """
-        optimizers = {}
-
-        for name, params in params_dict.items():
-            # Get group-specific hyperparameters or use defaults
-            group_config = self.optimizer_groups.get(name, {})
-
-            # Create optimizer with merged parameters (defaults + group-specific)
-            optimizer_kwargs = {
-                "lr": group_config.get("lr", self.lr),
-                "betas": group_config.get("betas", (0.9, 0.999)),
-                "eps": group_config.get("eps", 1e-5),
-                "weight_decay": group_config.get("weight_decay", self.weight_decay),
-            }
-
-            optimizers[name] = torch.optim.Adam(params, **optimizer_kwargs)
-
-        return optimizers
-
-
-def save_optimizer_state(
-    optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer], save_dir: Path
-) -> None:
-    """Save optimizer state to disk.
-
-    Args:
-        optimizer: Either a single optimizer or a dictionary of optimizers.
-        save_dir: Directory to save the optimizer state.
-    """
-    if isinstance(optimizer, dict):
-        # Handle dictionary of optimizers
-        for name, opt in optimizer.items():
-            optimizer_dir = save_dir / name
-            optimizer_dir.mkdir(exist_ok=True, parents=True)
-            _save_single_optimizer_state(opt, optimizer_dir)
-    else:
-        # Handle single optimizer
-        _save_single_optimizer_state(optimizer, save_dir)
-
-
-def _save_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
-    """Save a single optimizer's state to disk."""
+def save_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
    state = optimizer.state_dict()
    param_groups = state.pop("param_groups")
    flat_state = flatten_dict(state)
@@ -172,44 +102,11 @@ def _save_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Pat
    write_json(param_groups, save_dir / OPTIMIZER_PARAM_GROUPS)


-def load_optimizer_state(
-    optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer], save_dir: Path
-) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
-    """Load optimizer state from disk.
-
-    Args:
-        optimizer: Either a single optimizer or a dictionary of optimizers.
-        save_dir: Directory to load the optimizer state from.
-
-    Returns:
-        The updated optimizer(s) with loaded state.
-    """
-    if isinstance(optimizer, dict):
-        # Handle dictionary of optimizers
-        loaded_optimizers = {}
-        for name, opt in optimizer.items():
-            optimizer_dir = save_dir / name
-            if optimizer_dir.exists():
-                loaded_optimizers[name] = _load_single_optimizer_state(opt, optimizer_dir)
-            else:
-                loaded_optimizers[name] = opt
-        return loaded_optimizers
-    else:
-        # Handle single optimizer
-        return _load_single_optimizer_state(optimizer, save_dir)
-
-
-def _load_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> torch.optim.Optimizer:
-    """Load a single optimizer's state from disk."""
+def load_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> torch.optim.Optimizer:
    current_state_dict = optimizer.state_dict()
    flat_state = load_file(save_dir / OPTIMIZER_STATE)
    state = unflatten_dict(flat_state)
-
-    # Handle case where 'state' key might not exist (for newly created optimizers)
-    if "state" in state:
-        loaded_state_dict = {"state": {int(k): v for k, v in state["state"].items()}}
-    else:
-        loaded_state_dict = {"state": {}}
+    loaded_state_dict = {"state": {int(k): v for k, v in state["state"].items()}}

    if "param_groups" in current_state_dict:
        param_groups = deserialize_json_into_object(
--- a/lerobot/common/optim/schedulers.py
+++ b/lerobot/common/optim/schedulers.py
@@ -49,11 +49,7 @@ class DiffuserSchedulerConfig(LRSchedulerConfig):
    def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
        from diffusers.optimization import get_scheduler

-        kwargs = {
-            **asdict(self),
-            "num_training_steps": num_training_steps,
-            "optimizer": optimizer,
-        }
+        kwargs = {**asdict(self), "num_training_steps": num_training_steps, "optimizer": optimizer}
        return get_scheduler(**kwargs)


@@ -75,10 +71,7 @@ class VQBeTSchedulerConfig(LRSchedulerConfig):
                progress = float(adjusted_step - self.num_warmup_steps) / float(
                    max(1, num_training_steps - self.num_warmup_steps)
                )
-                return max(
-                    0.0,
-                    0.5 * (1.0 + math.cos(math.pi * float(self.num_cycles) * 2.0 * progress)),
-                )
+                return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(self.num_cycles) * 2.0 * progress)))

        return LambdaLR(optimizer, lr_lambda, -1)

--- a/lerobot/common/policies/act/modeling_act.py
+++ b/lerobot/common/policies/act/modeling_act.py
@@ -241,9 +241,7 @@ class ACTTemporalEnsembler:
            # Note: The last dimension is unsqueeze to make sure we can broadcast properly for tensor
            # operations later.
            self.ensembled_actions_count = torch.ones(
-                (self.chunk_size, 1),
-                dtype=torch.long,
-                device=self.ensembled_actions.device,
+                (self.chunk_size, 1), dtype=torch.long, device=self.ensembled_actions.device
            )
        else:
            # self.ensembled_actions will have shape (batch_size, chunk_size - 1, action_dim). Compute
@@ -255,10 +253,7 @@ class ACTTemporalEnsembler:
            # The last action, which has no prior online average, needs to get concatenated onto the end.
            self.ensembled_actions = torch.cat([self.ensembled_actions, actions[:, -1:]], dim=1)
            self.ensembled_actions_count = torch.cat(
-                [
-                    self.ensembled_actions_count,
-                    torch.ones_like(self.ensembled_actions_count[-1:]),
-                ]
+                [self.ensembled_actions_count, torch.ones_like(self.ensembled_actions_count[-1:])]
            )
        # "Consume" the first action.
        action, self.ensembled_actions, self.ensembled_actions_count = (
@@ -338,11 +333,7 @@ class ACT(nn.Module):
        # Backbone for image feature extraction.
        if self.config.image_features:
            backbone_model = getattr(torchvision.models, config.vision_backbone)(
-                replace_stride_with_dilation=[
-                    False,
-                    False,
-                    config.replace_final_stride_with_dilation,
-                ],
+                replace_stride_with_dilation=[False, False, config.replace_final_stride_with_dilation],
                weights=config.pretrained_backbone_weights,
                norm_layer=FrozenBatchNorm2d,
            )
@@ -436,11 +427,7 @@ class ACT(nn.Module):
            action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)

            if self.config.robot_state_feature:
-                vae_encoder_input = [
-                    cls_embed,
-                    robot_state_embed,
-                    action_embed,
-                ]  # (B, S+2, D)
+                vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
            else:
                vae_encoder_input = [cls_embed, action_embed]
            vae_encoder_input = torch.cat(vae_encoder_input, axis=1)
@@ -553,10 +540,7 @@ class ACTEncoder(nn.Module):
        self.norm = nn.LayerNorm(config.dim_model) if config.pre_norm else nn.Identity()

    def forward(
-        self,
-        x: Tensor,
-        pos_embed: Tensor | None = None,
-        key_padding_mask: Tensor | None = None,
+        self, x: Tensor, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None
    ) -> Tensor:
        for layer in self.layers:
            x = layer(x, pos_embed=pos_embed, key_padding_mask=key_padding_mask)
@@ -619,10 +603,7 @@ class ACTDecoder(nn.Module):
    ) -> Tensor:
        for layer in self.layers:
            x = layer(
-                x,
-                encoder_out,
-                decoder_pos_embed=decoder_pos_embed,
-                encoder_pos_embed=encoder_pos_embed,
+                x, encoder_out, decoder_pos_embed=decoder_pos_embed, encoder_pos_embed=encoder_pos_embed
            )
        if self.norm is not None:
            x = self.norm(x)
--- a/lerobot/common/policies/diffusion/modeling_diffusion.py
+++ b/lerobot/common/policies/diffusion/modeling_diffusion.py
@@ -33,7 +33,7 @@ from diffusers.schedulers.scheduling_ddim import DDIMScheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 from torch import Tensor, nn

-from lerobot.common.constants import OBS_ENV, OBS_ROBOT
+from lerobot.common.constants import OBS_ENV_STATE, OBS_STATE
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pretrained import PreTrainedPolicy
@@ -209,10 +209,7 @@ class DiffusionModel(nn.Module):

    # ========= inference  ============
    def conditional_sample(
-        self,
-        batch_size: int,
-        global_cond: Tensor | None = None,
-        generator: torch.Generator | None = None,
+        self, batch_size: int, global_cond: Tensor | None = None, generator: torch.Generator | None = None
    ) -> Tensor:
        device = get_device_from_parameters(self)
        dtype = get_dtype_from_parameters(self)
@@ -241,8 +238,8 @@ class DiffusionModel(nn.Module):

    def _prepare_global_conditioning(self, batch: dict[str, Tensor]) -> Tensor:
        """Encode image features and concatenate them all together along with the state vector."""
-        batch_size, n_obs_steps = batch[OBS_ROBOT].shape[:2]
-        global_cond_feats = [batch[OBS_ROBOT]]
+        batch_size, n_obs_steps = batch[OBS_STATE].shape[:2]
+        global_cond_feats = [batch[OBS_STATE]]
        # Extract image features.
        if self.config.image_features:
            if self.config.use_separate_rgb_encoder_per_camera:
@@ -257,10 +254,7 @@ class DiffusionModel(nn.Module):
                # Separate batch and sequence dims back out. The camera index dim gets absorbed into the
                # feature dim (effectively concatenating the camera features).
                img_features = einops.rearrange(
-                    img_features_list,
-                    "(n b s) ... -> b s (n ...)",
-                    b=batch_size,
-                    s=n_obs_steps,
+                    img_features_list, "(n b s) ... -> b s (n ...)", b=batch_size, s=n_obs_steps
                )
            else:
                # Combine batch, sequence, and "which camera" dims before passing to shared encoder.
@@ -270,15 +264,12 @@ class DiffusionModel(nn.Module):
                # Separate batch dim and sequence dim back out. The camera index dim gets absorbed into the
                # feature dim (effectively concatenating the camera features).
                img_features = einops.rearrange(
-                    img_features,
-                    "(b s n) ... -> b s (n ...)",
-                    b=batch_size,
-                    s=n_obs_steps,
+                    img_features, "(b s n) ... -> b s (n ...)", b=batch_size, s=n_obs_steps
                )
            global_cond_feats.append(img_features)

        if self.config.env_state_feature:
-            global_cond_feats.append(batch[OBS_ENV])
+            global_cond_feats.append(batch[OBS_ENV_STATE])

        # Concatenate features then flatten to (B, global_cond_dim).
        return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)
@@ -524,9 +515,7 @@ class DiffusionRgbEncoder(nn.Module):


 def _replace_submodules(
-    root_module: nn.Module,
-    predicate: Callable[[nn.Module], bool],
-    func: Callable[[nn.Module], nn.Module],
+    root_module: nn.Module, predicate: Callable[[nn.Module], bool], func: Callable[[nn.Module], nn.Module]
 ) -> nn.Module:
    """
    Args:
@@ -644,14 +633,10 @@ class DiffusionConditionalUnet1d(nn.Module):
        self.mid_modules = nn.ModuleList(
            [
                DiffusionConditionalResidualBlock1d(
-                    config.down_dims[-1],
-                    config.down_dims[-1],
-                    **common_res_block_kwargs,
+                    config.down_dims[-1], config.down_dims[-1], **common_res_block_kwargs
                ),
                DiffusionConditionalResidualBlock1d(
-                    config.down_dims[-1],
-                    config.down_dims[-1],
-                    **common_res_block_kwargs,
+                    config.down_dims[-1], config.down_dims[-1], **common_res_block_kwargs
                ),
            ]
        )
--- a/lerobot/common/policies/factory.py
+++ b/lerobot/common/policies/factory.py
@@ -24,8 +24,8 @@ from lerobot.common.envs.configs import EnvConfig
 from lerobot.common.envs.utils import env_to_policy_features
 from lerobot.common.policies.act.configuration_act import ACTConfig
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
-from lerobot.common.policies.hilserl.classifier.configuration_classifier import ClassifierConfig
 from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
@@ -55,14 +55,10 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
        from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy

        return PI0Policy
-    elif name == "sac":
-        from lerobot.common.policies.sac.modeling_sac import SACPolicy
+    elif name == "pi0fast":
+        from lerobot.common.policies.pi0fast.modeling_pi0fast import PI0FASTPolicy

-        return SACPolicy
-    elif name == "hilserl_classifier":
-        from lerobot.common.policies.hilserl.classifier.modeling_classifier import Classifier
-
-        return Classifier
+        return PI0FASTPolicy
    else:
        raise NotImplementedError(f"Policy with name {name} is not implemented.")

@@ -78,8 +74,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
        return VQBeTConfig(**kwargs)
    elif policy_type == "pi0":
        return PI0Config(**kwargs)
-    elif policy_type == "hilserl_classifier":
-        return ClassifierConfig(**kwargs)
+    elif policy_type == "pi0fast":
+        return PI0FASTConfig(**kwargs)
    else:
        raise ValueError(f"Policy type '{policy_type}' is not available.")

--- a/lerobot/common/policies/hilserl/classifier/configuration_classifier.py
+++ b/lerobot/common/policies/hilserl/classifier/configuration_classifier.py
@@ -1,53 +0,0 @@
-from dataclasses import dataclass
-from typing import List
-
-from lerobot.common.optim.optimizers import AdamWConfig, OptimizerConfig
-from lerobot.common.optim.schedulers import LRSchedulerConfig
-from lerobot.configs.policies import PreTrainedConfig
-
-
-@PreTrainedConfig.register_subclass(name="hilserl_classifier")
-@dataclass
-class ClassifierConfig(PreTrainedConfig):
-    """Configuration for the Classifier model."""
-
-    name: str = "hilserl_classifier"
-    num_classes: int = 2
-    hidden_dim: int = 256
-    dropout_rate: float = 0.1
-    model_name: str = "helper2424/resnet10"
-    device: str = "cpu"
-    model_type: str = "cnn"  # "transformer" or "cnn"
-    num_cameras: int = 2
-    learning_rate: float = 1e-4
-    normalization_mode = None
-    # output_features: Dict[str, PolicyFeature] = field(
-    #     default_factory=lambda: {"next.reward": PolicyFeature(type=FeatureType.REWARD, shape=(1,))}
-    # )
-
-    @property
-    def observation_delta_indices(self) -> List | None:
-        return None
-
-    @property
-    def action_delta_indices(self) -> List | None:
-        return None
-
-    @property
-    def reward_delta_indices(self) -> List | None:
-        return None
-
-    def get_optimizer_preset(self) -> OptimizerConfig:
-        return AdamWConfig(
-            lr=self.learning_rate,
-            weight_decay=0.01,
-            grad_clip_norm=1.0,
-        )
-
-    def get_scheduler_preset(self) -> LRSchedulerConfig | None:
-        return None
-
-    def validate_features(self) -> None:
-        """Validate feature configurations."""
-        # Classifier doesn't need specific feature validation
-        pass
--- a/lerobot/common/policies/hilserl/classifier/modeling_classifier.py
+++ b/lerobot/common/policies/hilserl/classifier/modeling_classifier.py
@@ -1,237 +0,0 @@
-import logging
-from typing import Dict, Optional, Tuple
-
-import torch
-from torch import Tensor, nn
-
-from lerobot.common.constants import OBS_IMAGE
-from lerobot.common.policies.hilserl.classifier.configuration_classifier import (
-    ClassifierConfig,
-)
-from lerobot.common.policies.normalize import Normalize, Unnormalize
-from lerobot.common.policies.pretrained import PreTrainedPolicy
-
-logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s")
-logger = logging.getLogger(__name__)
-
-
-class ClassifierOutput:
-    """Wrapper for classifier outputs with additional metadata."""
-
-    def __init__(
-        self,
-        logits: Tensor,
-        probabilities: Optional[Tensor] = None,
-        hidden_states: Optional[Tensor] = None,
-    ):
-        self.logits = logits
-        self.probabilities = probabilities
-        self.hidden_states = hidden_states
-
-    def __repr__(self):
-        return (
-            f"ClassifierOutput(logits={self.logits}, "
-            f"probabilities={self.probabilities}, "
-            f"hidden_states={self.hidden_states})"
-        )
-
-
-class Classifier(PreTrainedPolicy):
-    """Image classifier built on top of a pre-trained encoder."""
-
-    name = "hilserl_classifier"
-    config_class = ClassifierConfig
-
-    def __init__(
-        self,
-        config: ClassifierConfig,
-        dataset_stats: Dict[str, Dict[str, Tensor]] | None = None,
-    ):
-        from transformers import AutoModel
-
-        super().__init__(config)
-        self.config = config
-
-        # Initialize normalization (standardized with the policy framework)
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
-        # Set up encoder
-        encoder = AutoModel.from_pretrained(self.config.model_name, trust_remote_code=True)
-        # Extract vision model if we're given a multimodal model
-        if hasattr(encoder, "vision_model"):
-            logging.info("Multimodal model detected - using vision encoder only")
-            self.encoder = encoder.vision_model
-            self.vision_config = encoder.config.vision_config
-        else:
-            self.encoder = encoder
-            self.vision_config = getattr(encoder, "config", None)
-
-        # Model type from config
-        self.is_cnn = self.config.model_type == "cnn"
-
-        # For CNNs, initialize backbone
-        if self.is_cnn:
-            self._setup_cnn_backbone()
-
-        self._freeze_encoder()
-        self._build_classifier_head()
-
-    def _setup_cnn_backbone(self):
-        """Set up CNN encoder"""
-        if hasattr(self.encoder, "fc"):
-            self.feature_dim = self.encoder.fc.in_features
-            self.encoder = nn.Sequential(*list(self.encoder.children())[:-1])
-        elif hasattr(self.encoder.config, "hidden_sizes"):
-            self.feature_dim = self.encoder.config.hidden_sizes[-1]  # Last channel dimension
-        else:
-            raise ValueError("Unsupported CNN architecture")
-
-    def _freeze_encoder(self) -> None:
-        """Freeze the encoder parameters."""
-        for param in self.encoder.parameters():
-            param.requires_grad = False
-
-    def _build_classifier_head(self) -> None:
-        """Initialize the classifier head architecture."""
-        # Get input dimension based on model type
-        if self.is_cnn:
-            input_dim = self.feature_dim
-        else:  # Transformer models
-            if hasattr(self.encoder.config, "hidden_size"):
-                input_dim = self.encoder.config.hidden_size
-            else:
-                raise ValueError("Unsupported transformer architecture since hidden_size is not found")
-
-        self.classifier_head = nn.Sequential(
-            nn.Linear(input_dim * self.config.num_cameras, self.config.hidden_dim),
-            nn.Dropout(self.config.dropout_rate),
-            nn.LayerNorm(self.config.hidden_dim),
-            nn.ReLU(),
-            nn.Linear(
-                self.config.hidden_dim,
-                1 if self.config.num_classes == 2 else self.config.num_classes,
-            ),
-        )
-
-    def _get_encoder_output(self, x: torch.Tensor) -> torch.Tensor:
-        """Extract the appropriate output from the encoder."""
-        with torch.no_grad():
-            if self.is_cnn:
-                # The HF ResNet applies pooling internally
-                outputs = self.encoder(x)
-                # Get pooled output directly
-                features = outputs.pooler_output
-
-                if features.dim() > 2:
-                    features = features.squeeze(-1).squeeze(-1)
-                return features
-            else:  # Transformer models
-                outputs = self.encoder(x)
-                if hasattr(outputs, "pooler_output") and outputs.pooler_output is not None:
-                    return outputs.pooler_output
-                return outputs.last_hidden_state[:, 0, :]
-
-    def extract_images_and_labels(self, batch: Dict[str, Tensor]) -> Tuple[list, Tensor]:
-        """Extract image tensors and label tensors from batch."""
-        # Find image keys in input features
-        image_keys = [key for key in self.config.input_features if key.startswith(OBS_IMAGE)]
-
-        # Extract the images and labels
-        images = [batch[key] for key in image_keys]
-        labels = batch["next.reward"]
-
-        return images, labels
-
-    def predict(self, xs: list) -> ClassifierOutput:
-        """Forward pass of the classifier for inference."""
-        encoder_outputs = torch.hstack([self._get_encoder_output(x) for x in xs])
-        logits = self.classifier_head(encoder_outputs)
-
-        if self.config.num_classes == 2:
-            logits = logits.squeeze(-1)
-            probabilities = torch.sigmoid(logits)
-        else:
-            probabilities = torch.softmax(logits, dim=-1)
-
-        return ClassifierOutput(logits=logits, probabilities=probabilities, hidden_states=encoder_outputs)
-
-    def forward(self, batch: Dict[str, Tensor]) -> Tuple[Tensor, Dict[str, Tensor]]:
-        """Standard forward pass for training compatible with train.py."""
-        # Normalize inputs if needed
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
-        # Extract images and labels
-        images, labels = self.extract_images_and_labels(batch)
-
-        # Get predictions
-        outputs = self.predict(images)
-
-        # Calculate loss
-        if self.config.num_classes == 2:
-            # Binary classification
-            loss = nn.functional.binary_cross_entropy_with_logits(outputs.logits, labels)
-            predictions = (torch.sigmoid(outputs.logits) > 0.5).float()
-        else:
-            # Multi-class classification
-            loss = nn.functional.cross_entropy(outputs.logits, labels.long())
-            predictions = torch.argmax(outputs.logits, dim=1)
-
-        # Calculate accuracy for logging
-        correct = (predictions == labels).sum().item()
-        total = labels.size(0)
-        accuracy = 100 * correct / total
-
-        # Return loss and metrics for logging
-        output_dict = {
-            "accuracy": accuracy,
-            "correct": correct,
-            "total": total,
-        }
-
-        return loss, output_dict
-
-    def predict_reward(self, batch, threshold=0.6):
-        """Legacy method for compatibility."""
-        images, _ = self.extract_images_and_labels(batch)
-        if self.config.num_classes == 2:
-            probs = self.predict(images).probabilities
-            logging.debug(f"Predicted reward images: {probs}")
-            return (probs > threshold).float()
-        else:
-            return torch.argmax(self.predict(images).probabilities, dim=1)
-
-    # Methods required by PreTrainedPolicy abstract class
-
-    def get_optim_params(self) -> dict:
-        """Return optimizer parameters for the policy."""
-        return {
-            "params": self.parameters(),
-            "lr": getattr(self.config, "learning_rate", 1e-4),
-            "weight_decay": getattr(self.config, "weight_decay", 0.01),
-        }
-
-    def reset(self):
-        """Reset any stateful components (required by PreTrainedPolicy)."""
-        # Classifier doesn't have stateful components that need resetting
-        pass
-
-    def select_action(self, batch: Dict[str, Tensor]) -> Tensor:
-        """Return action (class prediction) based on input observation."""
-        images, _ = self.extract_images_and_labels(batch)
-
-        with torch.no_grad():
-            outputs = self.predict(images)
-
-            if self.config.num_classes == 2:
-                # For binary classification return 0 or 1
-                return (outputs.probabilities > 0.5).float()
-            else:
-                # For multi-class return the predicted class
-                return torch.argmax(outputs.probabilities, dim=1)
--- a/lerobot/common/policies/normalize.py
+++ b/lerobot/common/policies/normalize.py
@@ -79,46 +79,28 @@ def create_stats_buffers(
            )

        # TODO(aliberts, rcadene): harmonize this to only use one framework (np or torch)
-        if stats and key in stats:
-            if norm_mode is NormalizationMode.MEAN_STD:
-                if "mean" not in stats[key] or "std" not in stats[key]:
-                    raise ValueError(
-                        f"Missing 'mean' or 'std' in stats for key {key} with MEAN_STD normalization"
-                    )
-
-                if isinstance(stats[key]["mean"], np.ndarray):
+        if stats:
+            if isinstance(stats[key]["mean"], np.ndarray):
+                if norm_mode is NormalizationMode.MEAN_STD:
                    buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(dtype=torch.float32)
                    buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(dtype=torch.float32)
-                elif isinstance(stats[key]["mean"], torch.Tensor):
-                    # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
-                    # tensors anywhere (for example, when we use the same stats for normalization and
-                    # unnormalization). See the logic here
-                    # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
-                    buffer["mean"].data = stats[key]["mean"].clone().to(dtype=torch.float32)
-                    buffer["std"].data = stats[key]["std"].clone().to(dtype=torch.float32)
-                else:
-                    type_ = type(stats[key]["mean"])
-                    raise ValueError(
-                        f"np.ndarray or torch.Tensor expected for 'mean', but type is '{type_}' instead."
-                    )
-
-            elif norm_mode is NormalizationMode.MIN_MAX:
-                if "min" not in stats[key] or "max" not in stats[key]:
-                    raise ValueError(
-                        f"Missing 'min' or 'max' in stats for key {key} with MIN_MAX normalization"
-                    )
-
-                if isinstance(stats[key]["min"], np.ndarray):
+                elif norm_mode is NormalizationMode.MIN_MAX:
                    buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=torch.float32)
                    buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=torch.float32)
-                elif isinstance(stats[key]["min"], torch.Tensor):
+            elif isinstance(stats[key]["mean"], torch.Tensor):
+                # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
+                # tensors anywhere (for example, when we use the same stats for normalization and
+                # unnormalization). See the logic here
+                # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
+                if norm_mode is NormalizationMode.MEAN_STD:
+                    buffer["mean"].data = stats[key]["mean"].clone().to(dtype=torch.float32)
+                    buffer["std"].data = stats[key]["std"].clone().to(dtype=torch.float32)
+                elif norm_mode is NormalizationMode.MIN_MAX:
                    buffer["min"].data = stats[key]["min"].clone().to(dtype=torch.float32)
                    buffer["max"].data = stats[key]["max"].clone().to(dtype=torch.float32)
-                else:
-                    type_ = type(stats[key]["min"])
-                    raise ValueError(
-                        f"np.ndarray or torch.Tensor expected for 'min', but type is '{type_}' instead."
-                    )
+            else:
+                type_ = type(stats[key]["mean"])
+                raise ValueError(f"np.ndarray or torch.Tensor expected, but type is '{type_}' instead.")

        stats_buffers[key] = buffer
    return stats_buffers
@@ -167,13 +149,12 @@ class Normalize(nn.Module):
            setattr(self, "buffer_" + key.replace(".", "_"), buffer)

    # TODO(rcadene): should we remove torch.no_grad?
-    # @torch.no_grad
+    @torch.no_grad
    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
        batch = dict(batch)  # shallow copy avoids mutating the input batch
        for key, ft in self.features.items():
            if key not in batch:
                # FIXME(aliberts, rcadene): This might lead to silent fail!
-                # NOTE: (azouitine) This continues help us for instantiation SACPolicy
                continue

            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
@@ -242,7 +223,7 @@ class Unnormalize(nn.Module):
            setattr(self, "buffer_" + key.replace(".", "_"), buffer)

    # TODO(rcadene): should we remove torch.no_grad?
-    # @torch.no_grad
+    @torch.no_grad
    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
        batch = dict(batch)  # shallow copy avoids mutating the input batch
        for key, ft in self.features.items():
--- a/lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py
+++ b/lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py
@@ -61,11 +61,7 @@ from lerobot.common.policies.pi0.conversion_scripts.conversion_utils import (
 )
 from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy

-PRECISIONS = {
-    "bfloat16": torch.bfloat16,
-    "float32": torch.float32,
-    "float16": torch.float16,
-}
+PRECISIONS = {"bfloat16": torch.bfloat16, "float32": torch.float32, "float16": torch.float16}


 def slice_paligemma_state_dict(state_dict, config):
--- a/lerobot/common/policies/pi0/flex_attention.py
+++ b/lerobot/common/policies/pi0/flex_attention.py
@@ -48,32 +48,18 @@ def flex_attention_forward(

    key_states = key_states[:, :, :, None, :]
    key_states = key_states.expand(
-        batch_size,
-        key_states.shape[1],
-        num_key_value_heads,
-        num_key_value_groups,
-        head_dim,
+        batch_size, key_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
    )
    key_states = key_states.reshape(
-        batch_size,
-        key_states.shape[1],
-        num_key_value_heads * num_key_value_groups,
-        head_dim,
+        batch_size, key_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
    )

    value_states = value_states[:, :, :, None, :]
    value_states = value_states.expand(
-        batch_size,
-        value_states.shape[1],
-        num_key_value_heads,
-        num_key_value_groups,
-        head_dim,
+        batch_size, value_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
    )
    value_states = value_states.reshape(
-        batch_size,
-        value_states.shape[1],
-        num_key_value_heads * num_key_value_groups,
-        head_dim,
+        batch_size, value_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
    )

    query_states = query_states.transpose(1, 2)
--- a/lerobot/common/policies/pi0/modeling_pi0.py
+++ b/lerobot/common/policies/pi0/modeling_pi0.py
@@ -24,7 +24,7 @@ Designed by Physical Intelligence. Ported from Jax by Hugging Face.

 Install pi0 extra dependencies:
 ```bash
-pip install --no-binary=av -e ".[pi0]"
+pip install -e ".[pi0]"
 ```

 Example of finetuning the pi0 pretrained model (`pi0_base` in `openpi`):
@@ -57,7 +57,7 @@ import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 from transformers import AutoTokenizer

-from lerobot.common.constants import ACTION, OBS_ROBOT
+from lerobot.common.constants import ACTION, OBS_STATE
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.common.policies.pi0.paligemma_with_expert import (
@@ -69,11 +69,7 @@ from lerobot.common.utils.utils import get_safe_dtype


 def create_sinusoidal_pos_embedding(
-    time: torch.tensor,
-    dimension: int,
-    min_period: float,
-    max_period: float,
-    device="cpu",
+    time: torch.tensor, dimension: int, min_period: float, max_period: float, device="cpu"
 ) -> Tensor:
    """Computes sine-cosine positional embedding vectors for scalar positions."""
    if dimension % 2 != 0:
@@ -275,7 +271,7 @@ class PI0Policy(PreTrainedPolicy):
        self.eval()

        if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+            batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])

        batch = self.normalize_inputs(batch)

@@ -307,7 +303,7 @@ class PI0Policy(PreTrainedPolicy):
    def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> tuple[Tensor, dict[str, Tensor]]:
        """Do a full training forward pass to compute the loss"""
        if self.config.adapt_to_pi_aloha:
-            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+            batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])

        batch = self.normalize_inputs(batch)
@@ -384,7 +380,7 @@ class PI0Policy(PreTrainedPolicy):

    def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
        """Tokenize the text input"""
-        device = batch[OBS_ROBOT].device
+        device = batch[OBS_STATE].device
        tasks = batch["task"]

        # PaliGemma prompt has to end with a new line
@@ -431,7 +427,7 @@ class PI0Policy(PreTrainedPolicy):

    def prepare_state(self, batch):
        """Pad state"""
-        state = pad_vector(batch[OBS_ROBOT], self.config.max_state_dim)
+        state = pad_vector(batch[OBS_STATE], self.config.max_state_dim)
        return state

    def prepare_action(self, batch):
@@ -581,11 +577,7 @@ class PI0FlowMatching(nn.Module):

        # Embed timestep using sine-cosine positional encoding with sensitivity in the range [0, 1]
        time_emb = create_sinusoidal_pos_embedding(
-            timestep,
-            self.config.proj_width,
-            min_period=4e-3,
-            max_period=4.0,
-            device=device,
+            timestep, self.config.proj_width, min_period=4e-3, max_period=4.0, device=device
        )
        time_emb = time_emb.type(dtype=dtype)

@@ -617,15 +609,7 @@ class PI0FlowMatching(nn.Module):
        return embs, pad_masks, att_masks

    def forward(
-        self,
-        images,
-        img_masks,
-        lang_tokens,
-        lang_masks,
-        state,
-        actions,
-        noise=None,
-        time=None,
+        self, images, img_masks, lang_tokens, lang_masks, state, actions, noise=None, time=None
    ) -> Tensor:
        """Do a full training forward pass and compute the loss (batch_size x num_steps x num_motors)"""
        if noise is None:
@@ -671,11 +655,7 @@ class PI0FlowMatching(nn.Module):
        device = state.device

        if noise is None:
-            actions_shape = (
-                bsize,
-                self.config.n_action_steps,
-                self.config.max_action_dim,
-            )
+            actions_shape = (bsize, self.config.n_action_steps, self.config.max_action_dim)
            noise = self.sample_noise(actions_shape, device)

        prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
--- a/lerobot/common/policies/pi0/paligemma_with_expert.py
+++ b/lerobot/common/policies/pi0/paligemma_with_expert.py
@@ -293,18 +293,12 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
                    # in `transformers`. (molbap)
                    key_states = torch.cat([past_key_values[layer_idx]["key_states"], key_states], dim=1)
                    value_states = torch.cat(
-                        [past_key_values[layer_idx]["value_states"], value_states],
-                        dim=1,
+                        [past_key_values[layer_idx]["value_states"], value_states], dim=1
                    )

            attention_interface = self.get_attention_interface()
            att_output = attention_interface(
-                attention_mask,
-                batch_size,
-                head_dim,
-                query_states,
-                key_states,
-                value_states,
+                attention_mask, batch_size, head_dim, query_states, key_states, value_states
            )
            att_output = att_output.to(dtype=torch.bfloat16)

@@ -364,24 +358,12 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
        return attention_interface

    def flash_attention_forward(
-        self,
-        attention_mask,
-        batch_size,
-        head_dim,
-        query_states,
-        key_states,
-        value_states,
+        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
    ):
        raise NotImplementedError("FA2 is not implemented (yet)")

    def eager_attention_forward(
-        self,
-        attention_mask,
-        batch_size,
-        head_dim,
-        query_states,
-        key_states,
-        value_states,
+        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
    ):
        num_att_heads = self.config.paligemma_config.text_config.num_attention_heads
        num_key_value_heads = self.config.paligemma_config.text_config.num_key_value_heads
@@ -393,31 +375,17 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
        sequence_length = key_states.shape[1]

        key_states = key_states[:, :, :, None, :].expand(
-            batch_size,
-            sequence_length,
-            num_key_value_heads,
-            num_key_value_groups,
-            head_dim,
+            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
        )
        key_states = key_states.reshape(
-            batch_size,
-            sequence_length,
-            num_key_value_heads * num_key_value_groups,
-            head_dim,
+            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
        )

        value_states = value_states[:, :, :, None, :].expand(
-            batch_size,
-            sequence_length,
-            num_key_value_heads,
-            num_key_value_groups,
-            head_dim,
+            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
        )
        value_states = value_states.reshape(
-            batch_size,
-            sequence_length,
-            num_key_value_heads * num_key_value_groups,
-            head_dim,
+            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
        )

        # Attention here is upcasted to float32 to match the original eager implementation.
--- a/lerobot/common/policies/pi0fast/configuration_pi0fast.py
+++ b/lerobot/common/policies/pi0fast/configuration_pi0fast.py
@@ -0,0 +1,136 @@
+from dataclasses import dataclass, field
+
+from lerobot.common.optim.optimizers import AdamWConfig
+from lerobot.common.optim.schedulers import (
+    CosineDecayWithWarmupSchedulerConfig,
+)
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+
+
+@PreTrainedConfig.register_subclass("pi0fast")
+@dataclass
+class PI0FASTConfig(PreTrainedConfig):
+    # Input / output structure.
+    n_obs_steps: int = 1
+    chunk_size: int = 10
+    n_action_steps: int = 5
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MEAN_STD,
+        }
+    )
+
+    # Shorter state and action vectors will be padded
+    max_state_dim: int = 32  # 32
+    max_action_dim: int = 32  # 32
+
+    # Image preprocessing
+    resize_imgs_with_padding: tuple[int, int] = (224, 224)
+    interpolate_like_pi: bool = False
+
+    # Add empty images. Used by pi0_aloha_sim which adds the empty
+    # left and right wrist cameras in addition to the top camera.
+    empty_cameras: int = 0
+
+    # Converts the joint and gripper values from the standard Aloha space to
+    # the space used by the pi internal runtime which was used to train the base model.
+    adapt_to_pi_aloha: bool = False
+
+    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
+    # Gripper dimensions will remain in absolute values.
+    use_delta_joint_actions_aloha: bool = False
+
+    # Tokenizer
+    tokenizer_max_length: int = 48
+
+    # Projector
+    proj_width: int = 1024
+
+    # Decoding
+    max_decoding_steps: int = 256
+    fast_skip_tokens: int = 128  # Skip last 128 tokens in PaliGemma vocab since they are special tokens
+    max_input_seq_len: int = 256  # 512
+
+    # Utils
+    use_cache: bool = True
+
+    # Frozen parameters
+    freeze_vision_encoder: bool = True
+    freeze_lm_head: bool = True
+
+    # Training presets
+    optimizer_lr: float = 1e-4
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-5
+
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    checkpoint_path: str = None
+
+    padding_side: str = "right"
+
+    precision: str = "bfloat16"
+    grad_clip_norm: float = 1
+
+    # Allows padding/truncation of generated action tokens during detokenization to ensure decoding.
+    # In the original version, tensors of 0s were generated if shapes didn't match for stable decoding.
+    relaxed_action_decoding: bool = True
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        """Input validation (not exhaustive)."""
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
+                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
+            )
+        if self.n_obs_steps != 1:
+            raise ValueError(
+                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
+            )
+
+    def validate_features(self) -> None:
+        for i in range(self.empty_cameras):
+            key = f"observation.images.empty_camera_{i}"
+            empty_camera = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(3, 480, 640),
+            )
+            self.input_features[key] = empty_camera
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None
--- a/lerobot/common/policies/pi0fast/modeling_pi0fast.py
+++ b/lerobot/common/policies/pi0fast/modeling_pi0fast.py
@@ -0,0 +1,973 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and 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.
+
+"""
+π0+FAST: Efficient Action Tokenization for Vision-Language-Action Models
+
+[Paper](https://arxiv.org/abs/2501.09747)
+[Jax code](https://github.com/Physical-Intelligence/openpi)
+
+Designed by Physical Intelligence. Ported from Jax by Hugging Face.
+
+Example of finetuning the pi0+FAST pretrained model (`pi0_fast_base` in `openpi`):
+```bash
+python lerobot/scripts/train.py \
+--policy.path=lerobot/pi0fast_base \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of training the pi0+FAST neural network with from scratch:
+```bash
+python lerobot/scripts/train.py \
+--policy.type=pi0fast \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of using the pi0 pretrained model outside LeRobot training framework:
+```python
+policy = PI0FASTPolicy.from_pretrained("lerobot/pi0fast_base")
+```
+
+"""
+
+from collections import deque
+from functools import partial
+
+import numpy as np
+import torch
+import torch.nn.functional as F  # noqa: N812
+from PIL import Image
+from scipy.fft import idct
+from torch import Tensor, nn
+from transformers import AutoProcessor, AutoTokenizer, PaliGemmaForConditionalGeneration
+from transformers.cache_utils import HybridCache, StaticCache
+from transformers.models.auto import CONFIG_MAPPING
+
+from lerobot.common.constants import ACTION, OBS_ROBOT
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+
+PRECISION = {
+    "float16": torch.float16,
+    "float32": torch.float32,
+    "bfloat16": torch.bfloat16,
+}
+
+
+def normalize(x, min_val, max_val):
+    return (x - min_val) / (max_val - min_val)
+
+
+def unnormalize(x, min_val, max_val):
+    return x * (max_val - min_val) + min_val
+
+
+def safe_arcsin(value):
+    # This ensures that the input stays within
+    # [−1,1] to avoid invalid values for arcsin
+    return torch.arcsin(torch.clamp(value, -1.0, 1.0))
+
+
+def aloha_gripper_to_angular(value):
+    # Aloha transforms the gripper positions into a linear space. The following code
+    # reverses this transformation to be consistent with pi0 which is pretrained in
+    # angular space.
+    #
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_POSITION_OPEN, PUPPET_GRIPPER_POSITION_CLOSED
+    value = unnormalize(value, min_val=0.01844, max_val=0.05800)
+
+    # This is the inverse of the angular to linear transformation inside the Interbotix code.
+    def linear_to_radian(linear_position, arm_length, horn_radius):
+        value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
+        return safe_arcsin(value)
+
+    # The constants are taken from the Interbotix code.
+    value = linear_to_radian(value, arm_length=0.036, horn_radius=0.022)
+
+    # Normalize to [0, 1].
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+def aloha_gripper_from_angular(value):
+    # Convert from the gripper position used by pi0 to the gripper position that is used by Aloha.
+    # Note that the units are still angular but the range is different.
+
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    value = unnormalize(value, min_val=0.4, max_val=1.5)
+
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_JOINT_OPEN, PUPPET_GRIPPER_JOINT_CLOSE
+    return normalize(value, min_val=-0.6213, max_val=1.4910)
+
+
+def aloha_gripper_from_angular_inv(value):
+    # Directly inverts the gripper_from_angular function.
+    value = unnormalize(value, min_val=-0.6213, max_val=1.4910)
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+class PI0FASTPolicy(PreTrainedPolicy):
+    """Wrapper class around PI0FAST tokenizer and model to train and run inference within LeRobot."""
+
+    config_class = PI0FASTConfig
+    name = "pi0fast"
+
+    def __init__(
+        self,
+        config: PI0FASTConfig,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                    the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_targets = Normalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+
+        self.language_tokenizer = AutoProcessor.from_pretrained("google/paligemma-3b-pt-224")
+        self.model = PI0FAST(config)
+
+        self.reset()
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    def get_optim_params(self) -> dict:
+        return self.parameters()
+
+    def _pi_aloha_decode_state(self, state):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            state[:, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            state[:, motor_idx] = aloha_gripper_to_angular(state[:, motor_idx])
+        return state
+
+    def _pi_aloha_encode_actions(self, actions):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
+        return actions
+
+    def _pi_aloha_encode_actions_inv(self, actions):
+        # Flip the joints again.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
+        return actions
+
+    @torch.no_grad
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method wraps `select_actions` in order to return one action at a time for execution in the
+        environment. It works by managing the actions in a queue and only calling `select_actions` when the
+        queue is empty.
+        """
+        self.eval()
+
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+
+        batch = self.normalize_inputs(batch)
+
+        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
+        # querying the policy.
+        if len(self._action_queue) == 0:
+            actions = self.model.generate_actions(batch)
+
+            actions = actions[:, : self.config.n_action_steps]
+
+            original_action_dim = self.config.action_feature.shape[
+                0
+            ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
+            actions = actions[:, :, :original_action_dim]
+
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.adapt_to_pi_aloha:
+                actions = self._pi_aloha_encode_actions(actions)
+
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
+        loss_dict = self.model.forward(batch)
+        return loss_dict["loss"], loss_dict
+
+
+def block_causal_update_causal_mask(
+    attention_mask,
+    token_type_ids=None,
+    past_key_values=None,
+    cache_position=None,
+    input_tensor=None,
+    attn_implementation: str = "eager",
+    dtype: torch.dtype = "float32",
+):
+    """
+    Update the causal mask during training and generation. It can be customized to different attention masks.
+    """
+    if attn_implementation == "flash_attention_2":
+        if attention_mask is not None and 0.0 in attention_mask:
+            return attention_mask
+        return None
+    using_static_cache = isinstance(past_key_values, StaticCache)
+    min_dtype = torch.finfo(dtype).min
+
+    if input_tensor is None:
+        input_tensor = attention_mask
+
+    inputs_lead_dim, sequence_length = input_tensor.shape[:2]
+
+    if using_static_cache or isinstance(past_key_values, HybridCache):
+        target_length = past_key_values.get_max_cache_shape()
+    else:
+        target_length = (
+            attention_mask.shape[-1]
+            if isinstance(attention_mask, torch.Tensor)
+            else cache_position[0] + sequence_length + 1
+        )
+
+    # Handle precomputed attention masks
+    if attention_mask is not None and attention_mask.dim() == 4:
+        return attention_mask
+
+    # Causal mask initialization
+    causal_mask = torch.full(
+        (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
+    )
+
+    # Standard causal masking (triu ensures tokens can only attend to past)
+    if sequence_length != 1:
+        causal_mask = torch.triu(causal_mask, diagonal=1)
+
+        # Apply block causal mask
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.to(causal_mask.device).bool()
+            cumsum = torch.cumsum(token_type_ids, dim=1)
+            block_causal_mask = cumsum[:, None, :] <= cumsum[:, :, None]
+
+            # Combine causal_mask with block-wise attention mask
+            causal_mask = torch.where(block_causal_mask, 0.0, causal_mask)
+            causal_mask = causal_mask[:, None, :, :]
+        else:
+            # Apply past cache position constraint
+            causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
+                -1, 1
+            )
+            causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
+    else:
+        # Apply past cache position constraint
+        causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(
+            -1, 1
+        )
+        causal_mask = causal_mask[None, None, :, :].expand(inputs_lead_dim, 1, -1, -1)
+
+    if attention_mask is not None:
+        causal_mask = causal_mask.clone()  # Copy to contiguous memory for in-place edits
+        mask_length = attention_mask.shape[-1]
+
+        # Apply padding mask
+        padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+            causal_mask.device
+        )
+        padding_mask = padding_mask == 0
+        causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+            padding_mask, min_dtype
+        )
+
+    return causal_mask
+
+
+def prepare_inputs_for_generation(
+    # self,
+    input_ids,
+    past_key_values=None,
+    inputs_embeds=None,
+    cache_position=None,
+    position_ids=None,
+    pixel_values=None,
+    attention_mask=None,
+    token_type_ids=None,
+    use_cache=True,
+    num_logits_to_keep=None,
+    labels=None,
+    self=None,
+    **kwargs,
+):
+    # create block causal attention
+    if cache_position[0] > 0 and input_ids.shape[1] > 0:
+        input_tensor = input_ids[:, -1:]
+        new_positions = (
+            torch.ones(
+                (position_ids.shape[0], input_ids.shape[1]),
+                dtype=position_ids.dtype,
+                device=position_ids.device,
+            ).cumsum(-1)
+            + position_ids[:, -1:]
+        )
+        position_ids = torch.cat([position_ids, new_positions], dim=-1)
+    else:
+        input_tensor = inputs_embeds
+    attention_mask = block_causal_update_causal_mask(
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        cache_position=cache_position,
+        input_tensor=input_tensor,
+        token_type_ids=token_type_ids,
+        dtype=self.dtype,
+        attn_implementation=self.config.text_config._attn_implementation,
+    )
+    # Overwritten -- custom `position_ids` and `pixel_values` handling
+    model_inputs = self.language_model.prepare_inputs_for_generation(
+        input_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        cache_position=cache_position,
+        use_cache=use_cache,
+        num_logits_to_keep=num_logits_to_keep,
+        token_type_ids=token_type_ids,
+        **kwargs,
+    )
+
+    # Position_ids in Paligemma are 1-indexed
+    if model_inputs.get("position_ids") is not None:
+        model_inputs["position_ids"] += 1
+    # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+    # Otherwise we need pixel values to be passed to model. NOTE: use_cache=False needs pixel_values always
+    if cache_position[0] == 0:
+        model_inputs["pixel_values"] = pixel_values
+    is_training = token_type_ids is not None and labels is not None
+    if cache_position[0] == 0 and isinstance(past_key_values, HybridCache):
+        input_tensor = inputs_embeds if inputs_embeds is not None else input_ids
+        causal_mask = self._update_causal_mask(
+            attention_mask, token_type_ids, past_key_values, cache_position, input_tensor, is_training
+        )
+        model_inputs["attention_mask"] = causal_mask
+
+    return model_inputs
+
+
+class PI0FAST(nn.Module):
+    def __init__(self, config: PI0FASTConfig):
+        super().__init__()
+        self.config = config
+
+        # TODO: move tokenizers in Policy
+        fast_tokenizer_path = "physical-intelligence/fast"
+        pi0_paligemma_path = "google/paligemma-3b-pt-224"
+        self.paligemma_tokenizer = AutoTokenizer.from_pretrained(pi0_paligemma_path)
+        self.processor = AutoProcessor.from_pretrained(pi0_paligemma_path)
+        self.fast_tokenizer = AutoProcessor.from_pretrained(fast_tokenizer_path, trust_remote_code=True)
+        self.fast_skip_tokens = self.config.fast_skip_tokens
+        self.max_input_seq_len = self.config.max_input_seq_len
+        self.action_horizon = self.config.chunk_size
+        self.action_dim = self.config.action_feature.shape[
+            0
+        ]  # self.config.max_action_dim  # self.config.action_feature.shape[0]
+        precision = config.precision
+        torch_precision = PRECISION.get(precision, torch.float32)
+        self.pad_token_id = (
+            self.paligemma_tokenizer.pad_token_id
+            if hasattr(self.paligemma_tokenizer, "pad_token_id")
+            else self.paligemma_tokenizer.eos_token_id
+        )
+
+        paligemma_config = CONFIG_MAPPING["paligemma"](
+            transformers_version="4.48.1",
+            _vocab_size=257152,
+            bos_token_id=2,
+            eos_token_id=1,
+            hidden_size=2048,
+            image_token_index=257152,
+            model_type="paligemma",
+            pad_token_id=0,
+            projection_dim=2048,
+            text_config={
+                "hidden_activation": "gelu_pytorch_tanh",
+                "hidden_size": 2048,
+                "intermediate_size": 16384,
+                "model_type": "gemma",
+                "num_attention_heads": 8,
+                "num_hidden_layers": 18,
+                "num_image_tokens": 256,
+                "num_key_value_heads": 1,
+                "torch_dtype": precision,
+                "vocab_size": 257152,
+                "_attn_implementation": "eager",
+            },
+            vision_config={
+                "hidden_size": 1152,
+                "intermediate_size": 4304,
+                "model_type": "siglip_vision_model",
+                "num_attention_heads": 16,
+                "num_hidden_layers": 27,
+                "num_image_tokens": 256,
+                "patch_size": 14,
+                "projection_dim": 2048,
+                "projector_hidden_act": "gelu_pytorch_tanh",
+                "torch_dtype": precision,
+                "vision_use_head": False,
+            },
+        )
+        self.pi0_paligemma = PaliGemmaForConditionalGeneration(config=paligemma_config)
+
+        self.pi0_paligemma.prepare_inputs_for_generation = partial(
+            prepare_inputs_for_generation, self=self.pi0_paligemma
+        )
+        # change important stuff in bf16
+        params_to_change_dtype = [
+            "language_model",
+            "vision_tower",
+            "multi_modal",
+        ]
+        for name, param in self.pi0_paligemma.named_parameters():
+            if any(selector in name for selector in params_to_change_dtype):
+                param.data = param.data.to(dtype=torch_precision)
+        self.set_requires_grad()
+        self.image_keys = self.config.image_features.keys()
+        self.ignore_index = self.pi0_paligemma.config.ignore_index
+        self.padding_side = self.config.padding_side
+
+    def set_requires_grad(self):
+        if self.config.freeze_vision_encoder:
+            self.pi0_paligemma.vision_tower.eval()
+            for params in self.pi0_paligemma.vision_tower.parameters():
+                params.requires_grad = False
+        # To avoid unused params issue with distributed training
+        if self.config.freeze_lm_head:
+            for name, params in self.pi0_paligemma.named_parameters():
+                if "embed_tokens" in name:  # lm heads and embedding layer are tied
+                    params.requires_grad = False
+
+    def embed_tokens(self, tokens: torch.Tensor):
+        return self.pi0_paligemma.language_model.model.embed_tokens(tokens)
+
+    def prepare_inputs_for_generation(self, *args, **kwargs):
+        return self.pi0_paligemma.prepare_inputs_for_generation(*args, **kwargs)
+
+    def prepare_images(self, batch):
+        """Preprocess LeRobot batch into Pi0 inputs"""
+        images = []
+        img_masks = []
+        present_img_keys = [key for key in self.image_keys if key in batch]
+        if len(present_img_keys) == 0:
+            raise ValueError(
+                f"All image features are missing from the batch. At least one expected. (batch: {batch.keys()}) (image_features:{self.config.image_features})"
+            )
+
+        # Preprocess image features present in the batch
+        num_empty_cameras = 0
+        for key in self.image_keys:
+            if key in present_img_keys:
+                img = batch[key]
+
+                if self.config.resize_imgs_with_padding is not None:
+                    img = resize_with_pad(
+                        img,
+                        *self.config.resize_imgs_with_padding,
+                        pad_value=0,
+                        interpolate_like_pi=self.config.interpolate_like_pi,
+                    )
+
+                # Normalize from range [0,1] to [-1,1] as expacted by siglip
+                img = img * 2.0 - 1.0
+
+                bsize = img.shape[0]
+                device = img.device
+                mask = torch.ones(bsize, dtype=torch.bool, device=device)
+            else:
+                if num_empty_cameras >= self.config.empty_cameras:
+                    continue
+                img = torch.ones_like(img) * -1
+                bsize = img.shape[0]
+                device = img.device
+                mask = torch.ones(bsize, dtype=torch.bool, device=device)
+                num_empty_cameras += 1
+
+            images.append(img)
+            img_masks.append(mask)
+        return images, img_masks
+
+    def normalize_actions(self, actions: torch.Tensor) -> torch.Tensor:
+        mins = actions.amin(dim=(1, 2), keepdim=True)  # [0]
+        maxs = actions.amax(dim=(1, 2), keepdim=True)  # [0]
+        return 2 * (actions - mins) / (maxs - mins + 1e-8) - 1
+
+    def _act_tokens_to_paligemma_tokens(self, tokens: torch.Tensor) -> torch.Tensor:
+        out = self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens - tokens
+        return out
+
+    def fast_tokenizer_wrapper(self, actions_norm):
+        """
+        A wrapper for self.fast_tokenizer that ensures batch processing,
+        conversion to PyTorch tensors, and returns a dictionary without padding.
+        """
+        batch_tokens = self.fast_tokenizer(actions_norm)
+        fast_out = self.processor.tokenizer.pad({"input_ids": batch_tokens}, return_tensors="pt")
+
+        return fast_out
+
+    def create_token_type_ids(self, padded_mask: torch.Tensor, prefix_len: int) -> torch.Tensor:
+        token_type_ids = torch.zeros_like(padded_mask, dtype=torch.bool)
+        # Compute cumulative sum mask
+        cumsum_mask = (padded_mask != 0).cumsum(dim=1)
+        # Suffix block (everything after prefix_len)
+        suffix_mask = cumsum_mask > prefix_len
+        token_type_ids = suffix_mask
+        return token_type_ids
+
+    def create_input_tokens(self, state, lang_text, actions=None):
+        bsize = state.shape[0]
+        device = state.device
+        bins = torch.linspace(-1, 1, 256 + 1, device=device)[:-1]
+        discretized = torch.bucketize(state, bins) - 1
+        discretized = discretized[:, :32]
+
+        prefix_texts = []
+        state_text = []
+        for txt, disc in zip(lang_text, discretized, strict=False):
+            cleaned = txt.lower().strip().replace("_", " ")
+            state_str = " ".join(str(val.item()) for val in disc)
+            prefix_texts.append(f"Task: {cleaned}, State: {state_str};\n")
+            state_text.append(f"State: {state_str};\n")
+
+        prefix_out = self.paligemma_tokenizer(
+            prefix_texts, add_special_tokens=True, return_tensors="pt", padding="longest", truncation=False
+        )
+        prefix_ids = prefix_out["input_ids"].to(device)
+        prefix_mask = prefix_out["attention_mask"].to(device)
+        prefix_lens = prefix_mask.sum(dim=1)[:, None].cpu()
+
+        if actions is not None:
+            actions_norm = self.normalize_actions(actions)
+            actions_pad = F.pad(
+                actions_norm, (0, max(0, self.config.max_action_dim - actions_norm.shape[2])), value=0
+            )[:, :, : self.config.max_action_dim]
+            fast_out = self.fast_tokenizer_wrapper(
+                actions_pad.cpu(),
+            )
+            act_ids = fast_out["input_ids"]
+            act_mask = fast_out["attention_mask"].to(device)
+
+            act_ids = self._act_tokens_to_paligemma_tokens(act_ids).to(device)
+            # Replace action with 0 to pad tokens
+            act_ids = torch.where(
+                act_ids == self.paligemma_tokenizer.vocab_size - 1 - self.fast_skip_tokens,
+                self.pad_token_id,
+                act_ids,
+            )
+
+            eos_token = torch.tensor(
+                [self.paligemma_tokenizer.eos_token_id], dtype=torch.long, device=device
+            ).expand(bsize, -1)
+            eos_mask = torch.tensor([1], dtype=torch.long, device=device).expand(bsize, -1)
+            bos = self.paligemma_tokenizer("Action: ", add_special_tokens=False, return_tensors="pt")
+            bos_token = bos["input_ids"].expand(act_ids.shape[0], -1).to(device)
+            bos_mask = bos["attention_mask"].expand(act_ids.shape[0], -1).to(device)
+            act_ids = torch.cat([bos_token, act_ids, eos_token], dim=1)
+            act_mask = torch.cat([bos_mask, act_mask, eos_mask], dim=1)
+            act_mask = act_mask.to(device)
+        else:
+            act_ids = torch.empty(bsize, self.pad_token_id, dtype=torch.long, device=device)
+            act_mask = torch.empty(bsize, 0, dtype=torch.long, device=device)
+        final_ids = torch.cat([prefix_ids, act_ids], dim=1)
+
+        final_mask = torch.cat([prefix_mask, act_mask], dim=1)
+        batch_inputs = {"input_ids": final_ids.tolist(), "attention_mask": final_mask.tolist()}
+
+        # Use tokenizer pad function
+        padded_output = self.paligemma_tokenizer.pad(
+            batch_inputs, padding="longest", max_length=180, return_tensors="pt"
+        )
+        padded_mask = padded_output["attention_mask"]
+
+        # define tensor of padding lengths
+        att_mask = (padded_mask != 0).cumsum(dim=1) > prefix_lens
+
+        token_type_ids = self.create_token_type_ids(padded_mask=padded_mask, prefix_len=prefix_lens)
+
+        padded_output["padded_mask"] = padded_output.pop("attention_mask")
+        padded_output["attention_mask"] = att_mask
+        # loss is computed not on prefix, and not on padding
+        padded_output["loss_mask"] = att_mask & padded_output["padded_mask"]
+        padded_output["token_type_ids"] = token_type_ids
+        return padded_output
+
+    def shift_padding_side(
+        self,
+        tokens: torch.Tensor,
+        ar_mask: torch.Tensor,
+        padding_mask: torch.Tensor,
+        loss_mask: torch.Tensor,
+        targets: torch.Tensor,
+        token_type_ids: torch.Tensor,
+        padding_side: str = "right",
+    ) -> tuple[torch.Tensor]:
+        if padding_side not in ["right", "left"]:
+            return tokens, ar_mask, padding_mask, loss_mask, targets, token_type_ids
+
+        new_tokens = torch.empty_like(tokens)
+        new_ar_masks = torch.empty_like(ar_mask)
+        new_padding_mask = torch.empty_like(padding_mask)
+        new_loss_mask = torch.empty_like(loss_mask)
+        new_targets = torch.empty_like(targets)
+        new_token_type_ids = torch.empty_like(token_type_ids)
+        batch_size = tokens.shape[0]
+        for i in range(batch_size):
+            padding_indices = torch.where(padding_mask[i] == 0)[0]
+            non_padding_indices = torch.where(padding_mask[i] == 1)[0]
+            if padding_side == "left":
+                new_indices = torch.cat((padding_indices, non_padding_indices), dim=0)
+            else:
+                new_indices = torch.cat((non_padding_indices, padding_indices), dim=0)
+            new_tokens[i] = tokens[i].index_select(0, new_indices)
+            new_ar_masks[i] = ar_mask[i].index_select(0, new_indices)
+            new_padding_mask[i] = padding_mask[i].index_select(0, new_indices)
+            new_loss_mask[i] = loss_mask[i].index_select(0, new_indices)
+            new_targets[i] = targets[i].index_select(0, new_indices)
+            new_token_type_ids[i] = token_type_ids[i].index_select(0, new_indices)
+
+        return new_tokens, new_ar_masks, new_padding_mask, new_loss_mask, new_targets, new_token_type_ids
+
+    def forward(self, batch: dict[str, Tensor]):
+        device = batch[OBS_ROBOT].device
+        # TODO: keep like this or move to the policy .forward
+        images, img_masks = self.prepare_images(batch)
+
+        padded_outs = self.create_input_tokens(
+            state=batch[OBS_ROBOT],
+            lang_text=batch["task"],
+            actions=batch[ACTION],
+        )
+
+        embs, pad_masks, _, targets, loss_mask, token_type_ids = self.embed_inputs(
+            images,
+            img_masks,
+            padded_outs["input_ids"],
+            padded_outs["padded_mask"],
+            padded_outs["attention_mask"],
+            padded_outs["loss_mask"],
+            padded_outs["token_type_ids"],
+            padding_side=self.padding_side,
+        )
+        position_ids = torch.cumsum(pad_masks, dim=1) - 1
+        token_type_ids = token_type_ids.to(dtype=torch.int64)
+        past_seen_tokens = 0
+        cache_position = torch.arange(past_seen_tokens, past_seen_tokens + embs.shape[1], device=embs.device)
+        pad_masks = block_causal_update_causal_mask(
+            attention_mask=pad_masks,
+            past_key_values=None,
+            cache_position=cache_position,
+            input_tensor=embs,
+            token_type_ids=token_type_ids,
+            dtype=self.pi0_paligemma.dtype,
+            attn_implementation=self.pi0_paligemma.config.text_config._attn_implementation,
+        )
+        outputs = self.pi0_paligemma.forward(
+            input_ids=None,
+            token_type_ids=None,
+            attention_mask=pad_masks,
+            position_ids=position_ids,
+            past_key_values=None,
+            inputs_embeds=embs,
+            use_cache=False,
+            labels=None,
+        )
+
+        logits = outputs.logits
+
+        loss_fct = nn.CrossEntropyLoss(reduction="none")
+
+        # Shift left for next-step prediction
+        logits = logits[:, :-1, :]
+        targets = targets[:, 1:].to(device)  # Shift targets
+        loss_mask = loss_mask[:, 1:].to(device)  # Ensure correct shape
+
+        # Compute per-token loss
+        token_loss = loss_fct(logits.reshape(-1, logits.shape[-1]), targets.reshape(-1))
+
+        # Apply loss mask
+        token_loss = token_loss * loss_mask.reshape(-1)
+
+        # Compute final loss
+        loss = token_loss.sum() / torch.clamp(loss_mask.sum(), min=1)
+
+        # Return loss dictionary
+        loss_dict = {"ce_loss": loss.item(), "loss": loss}
+        return loss_dict
+
+    def decode_actions_with_fast(
+        self,
+        tokens: list[list[int]],
+        *,
+        time_horizon: int | None = None,
+        action_dim: int | None = None,
+        relaxed_decoding: bool = True,
+    ) -> np.array:
+        """
+        Adapt original decoding in FAST to always return actions instead of zeros.
+        """
+        self.time_horizon = (
+            time_horizon or self.fast_tokenizer.time_horizon or self.fast_tokenizer.called_time_horizon
+        )
+        self.action_dim = (
+            action_dim or self.fast_tokenizer.action_dim or self.fast_tokenizer.called_action_dim
+        )
+
+        # Cache the time horizon and action dimension for the next call
+        self.called_time_horizon = self.time_horizon
+        self.called_action_dim = self.action_dim
+
+        assert self.time_horizon is not None and self.action_dim is not None, (
+            "Tokenizer not initialized, call encode() once or pass in time_horizon and action_dim."
+        )
+
+        decoded_actions = []
+        for token in tokens:
+            try:
+                decoded_tokens = self.fast_tokenizer.bpe_tokenizer.decode(token)
+                decoded_dct_coeff = np.array(list(map(ord, decoded_tokens))) + self.fast_tokenizer.min_token
+                if relaxed_decoding:
+                    # Expected sequence length
+                    expected_seq_len = self.time_horizon * self.action_dim
+                    diff = expected_seq_len - decoded_dct_coeff.shape[0]
+                    # Apply truncation if too long
+                    if diff < 0:
+                        decoded_dct_coeff = decoded_dct_coeff[:expected_seq_len]  # Truncate on the right
+                    # Apply padding if too short
+                    elif diff > 0:
+                        decoded_dct_coeff = np.pad(
+                            decoded_dct_coeff, (0, diff), mode="constant", constant_values=0
+                        )
+
+                decoded_dct_coeff = decoded_dct_coeff.reshape(-1, self.action_dim)
+                assert decoded_dct_coeff.shape == (
+                    self.time_horizon,
+                    self.action_dim,
+                ), (
+                    f"Decoded DCT coefficients have shape {decoded_dct_coeff.shape}, expected ({self.time_horizon}, {self.action_dim})"
+                )
+            except Exception as e:
+                print(f"Error decoding tokens: {e}")
+                print(f"Tokens: {token}")
+                decoded_dct_coeff = np.zeros((self.time_horizon, self.action_dim))
+            decoded_actions.append(idct(decoded_dct_coeff / self.fast_tokenizer.scale, axis=0, norm="ortho"))
+        return np.stack(decoded_actions)
+
+    def extract_actions(self, tokens: torch.Tensor, action_horizon: int, action_dim: int) -> torch.Tensor:
+        """
+        Extracts actions from predicted output tokens using the FAST model.
+
+        Args:
+            tokens (torch.Tensor): The input tensor of tokenized outputs.
+            action_horizon (int): The number of timesteps for actions.
+            action_dim (int): The dimensionality of each action.
+
+        Returns:
+            torch.Tensor: The extracted actions as a tensor of shape (action_horizon, action_dim).
+        """
+        # Decode predicted output tokens
+        decoded_tokens = self.paligemma_tokenizer.batch_decode(tokens, skip_special_tokens=True)
+        cleaned_tokens = [
+            tokens_sequence.replace("Action:", "").replace(":", "").strip().split("|")[0].strip()
+            for tokens_sequence in decoded_tokens
+        ]
+        raw_action_tokens = [
+            self.processor.tokenizer.encode(sample_tokens, return_tensors="pt", padding=False)
+            for sample_tokens in cleaned_tokens
+        ]  # something like this should be robust #looks good
+        action_tokens = [
+            self._act_tokens_to_paligemma_tokens(raw_action_token) for raw_action_token in raw_action_tokens
+        ]
+        # returns the tensor of decoded actions per sample in a list
+        decoded_actions = [
+            torch.tensor(
+                self.decode_actions_with_fast(
+                    tok.tolist(),
+                    time_horizon=action_horizon,
+                    action_dim=action_dim,
+                    relaxed_decoding=self.config.relaxed_action_decoding,
+                ),
+                device=tokens.device,
+            ).squeeze(0)
+            for tok in action_tokens
+        ]
+
+        return torch.stack(
+            decoded_actions,
+            dim=0,
+        )
+
+    def generate_actions(self, batch: dict[str, Tensor]):
+        # TODO: keep like this or move to the policy .forward
+        images, img_masks = self.prepare_images(batch)
+
+        padded_outs = self.create_input_tokens(state=batch[OBS_ROBOT], lang_text=batch["task"], actions=None)
+        embs, pad_masks, att_masks2, targets, loss_mask, token_type_ids = self.embed_inputs(
+            images,
+            img_masks,
+            padded_outs["input_ids"],
+            padded_outs["padded_mask"],
+            padded_outs["attention_mask"],
+            padded_outs["loss_mask"],
+            padded_outs["token_type_ids"],
+            padding_side="left",
+        )
+        token_type_ids = token_type_ids.to(dtype=torch.int64)
+        prefix_position_ids = torch.cumsum(pad_masks, dim=1) - 1
+        output_tokens = self.pi0_paligemma.generate(
+            input_ids=None,
+            attention_mask=pad_masks,
+            position_ids=prefix_position_ids,
+            past_key_values=None,
+            inputs_embeds=embs,
+            use_cache=self.config.use_cache,
+            max_new_tokens=self.config.max_decoding_steps,
+            do_sample=False,
+            num_beams=1,
+            token_type_ids=token_type_ids,
+        )
+        actions = self.extract_actions(output_tokens, self.action_horizon, self.action_dim)
+        return actions
+
+    def embed_image(self, image: torch.Tensor):
+        return self.pi0_paligemma.get_image_features(image)
+
+    def embed_inputs(
+        self,
+        images,
+        img_masks,
+        tokens,
+        pad_mask,
+        ar_mask,
+        loss_mask,
+        token_type_ids,
+        padding_side: str = "right",
+    ):
+        # TODO: avoid list in python and torch.cat ; prefer pre-allocation with torch.empty
+        # images are a list of same size
+        # vectorizing everything!
+        device = images[0].device
+        image_embedding_dim = images[0].shape[-1]  # TODO should be from self.config
+        all_images = torch.stack(images, dim=1).to(device)
+        b, n, c, h, w = all_images.shape
+        all_images = all_images.view(b * n, c, h, w)
+        embedded = self.embed_image(all_images).to(device)
+        b_n, p, image_embedding_dim = embedded.shape  # Extract current dimensions
+        m = b_n // b  # Compute the number of images per sample dynamically
+
+        # Reshape dynamically
+        embedded = embedded.view(b, m, p, image_embedding_dim)
+        tokens_embs = self.embed_tokens(tokens.to(device))
+
+        img_masks = torch.stack(img_masks, dim=1).unsqueeze(-1).to(device)
+        num_img_emb = embedded.shape[2]
+        img_pad_masks = img_masks.repeat(1, 1, num_img_emb).view(b, -1)
+        img_att_masks = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
+
+        image_target_tokens = (
+            torch.ones((b, n, num_img_emb), dtype=torch.long, device=device) * self.pad_token_id
+        ).reshape(b, -1)
+        image_loss_mask = torch.zeros((b, n, num_img_emb), dtype=torch.long, device=device).reshape(b, -1)
+
+        embedded = embedded.reshape(b, n * num_img_emb, image_embedding_dim)  # Shape: (B, N*P, D)
+
+        embs = torch.cat([embedded, tokens_embs], dim=1).to(device)
+        pad_masks = torch.cat([img_pad_masks, pad_mask.to(device)], dim=1)
+        att_masks = torch.cat([img_att_masks, ar_mask.to(device)], dim=1)
+        loss_masks = torch.cat([image_loss_mask, loss_mask.to(device)], dim=1)
+        targets = torch.cat([image_target_tokens, tokens.to(device)], dim=1)
+        token_type_ids = torch.cat([img_att_masks, token_type_ids.to(device)], dim=1)
+
+        # Shift pad tokens to the left (.generate()) or right (.train())
+        embs, att_masks, pad_masks, loss_masks, targets, token_type_ids = self.shift_padding_side(
+            embs, att_masks, pad_masks, loss_masks, targets, token_type_ids, padding_side=padding_side
+        )
+
+        targets = torch.where(targets == self.pad_token_id, self.ignore_index, targets)
+        return embs, pad_masks, att_masks, targets, loss_masks, token_type_ids
+
+
+def resize_with_pad(img, width, height, pad_value=0, interpolate_like_pi=True):
+    # assume no-op when width height fits already
+    if img.ndim != 4:
+        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
+
+    cur_height, cur_width = img.shape[2:]
+
+    ratio = max(cur_width / width, cur_height / height)
+    resized_height = int(cur_height / ratio)
+    resized_width = int(cur_width / ratio)
+
+    if interpolate_like_pi:
+        img = (img * 255.0).to(dtype=torch.uint8)
+        img = img.permute(0, 2, 3, 1)
+        original_device = img.device
+        img = img.to(device="cpu").numpy()
+        imgs = []
+        for sub_img in img:
+            sub_img = Image.fromarray(sub_img)
+            resized_img = sub_img.resize((resized_width, resized_height), resample=2)
+            resized_img = torch.from_numpy(np.array(resized_img))
+            imgs.append(resized_img)
+        img = torch.stack(imgs, dim=0)
+        img = img.permute(0, 3, 1, 2)
+        resized_img = img.to(device=original_device, dtype=torch.float32) / 255.0
+    else:
+        resized_img = F.interpolate(
+            img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+        )
+
+    pad_height = max(0, int(height - resized_height))
+    pad_width = max(0, int(width - resized_width))
+
+    # pad on left and top of image
+    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
+    return padded_img
--- a/lerobot/common/policies/sac/configuration_sac.py
+++ b/lerobot/common/policies/sac/configuration_sac.py
@@ -1,227 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team.
-# All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from dataclasses import dataclass, field
-
-from lerobot.common.optim.optimizers import MultiAdamConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
-
-
-@dataclass
-class ConcurrencyConfig:
-    actor: str = "threads"
-    learner: str = "threads"
-
-
-@dataclass
-class ActorLearnerConfig:
-    learner_host: str = "127.0.0.1"
-    learner_port: int = 50051
-    policy_parameters_push_frequency: int = 4
-
-
-@dataclass
-class CriticNetworkConfig:
-    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
-    activate_final: bool = True
-    final_activation: str | None = None
-
-
-@dataclass
-class ActorNetworkConfig:
-    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
-    activate_final: bool = True
-
-
-@dataclass
-class PolicyConfig:
-    use_tanh_squash: bool = True
-    log_std_min: int = -5
-    log_std_max: int = 2
-    init_final: float = 0.05
-
-
-@PreTrainedConfig.register_subclass("sac")
-@dataclass
-class SACConfig(PreTrainedConfig):
-    """Soft Actor-Critic (SAC) configuration.
-
-    SAC is an off-policy actor-critic deep RL algorithm based on the maximum entropy
-    reinforcement learning framework. It learns a policy and a Q-function simultaneously
-    using experience collected from the environment.
-
-    This configuration class contains all the parameters needed to define a SAC agent,
-    including network architectures, optimization settings, and algorithm-specific
-    hyperparameters.
-
-    Args:
-        actor_network: Configuration for the actor network architecture.
-        critic_network: Configuration for the critic network architecture.
-        policy: Configuration for the policy parameters.
-        n_obs_steps: Number of observation steps to consider.
-        normalization_mapping: Mapping of feature types to normalization modes.
-        dataset_stats: Statistics for normalizing different types of inputs.
-        input_features: Dictionary of input features with their types and shapes.
-        output_features: Dictionary of output features with their types and shapes.
-        camera_number: Number of cameras used for visual observations.
-        device: Device to run the model on (e.g., "cuda", "cpu").
-        storage_device: Device to store the model on.
-        vision_encoder_name: Name of the vision encoder model.
-        freeze_vision_encoder: Whether to freeze the vision encoder during training.
-        image_encoder_hidden_dim: Hidden dimension size for the image encoder.
-        shared_encoder: Whether to use a shared encoder for actor and critic.
-        num_discrete_actions: Number of discrete actions, eg for gripper actions.
-        concurrency: Configuration for concurrency settings.
-        actor_learner: Configuration for actor-learner architecture.
-        online_steps: Number of steps for online training.
-        online_env_seed: Seed for the online environment.
-        online_buffer_capacity: Capacity of the online replay buffer.
-        offline_buffer_capacity: Capacity of the offline replay buffer.
-        async_prefetch: Whether to use asynchronous prefetching for the buffers.
-        online_step_before_learning: Number of steps before learning starts.
-        policy_update_freq: Frequency of policy updates.
-        discount: Discount factor for the SAC algorithm.
-        temperature_init: Initial temperature value.
-        num_critics: Number of critics in the ensemble.
-        num_subsample_critics: Number of subsampled critics for training.
-        critic_lr: Learning rate for the critic network.
-        actor_lr: Learning rate for the actor network.
-        temperature_lr: Learning rate for the temperature parameter.
-        critic_target_update_weight: Weight for the critic target update.
-        utd_ratio: Update-to-data ratio for the UTD algorithm.
-        state_encoder_hidden_dim: Hidden dimension size for the state encoder.
-        latent_dim: Dimension of the latent space.
-        target_entropy: Target entropy for the SAC algorithm.
-        use_backup_entropy: Whether to use backup entropy for the SAC algorithm.
-        grad_clip_norm: Gradient clipping norm for the SAC algorithm.
-    """
-
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.MEAN_STD,
-            "STATE": NormalizationMode.MIN_MAX,
-            "ENV": NormalizationMode.MIN_MAX,
-            "ACTION": NormalizationMode.MIN_MAX,
-        }
-    )
-
-    dataset_stats: dict[str, dict[str, list[float]]] = field(
-        default_factory=lambda: {
-            "observation.image": {
-                "mean": [0.485, 0.456, 0.406],
-                "std": [0.229, 0.224, 0.225],
-            },
-            "observation.state": {
-                "min": [0.0, 0.0],
-                "max": [1.0, 1.0],
-            },
-            "action": {
-                "min": [0.0, 0.0, 0.0],
-                "max": [1.0, 1.0, 1.0],
-            },
-        }
-    )
-
-    # Architecture specifics
-    camera_number: int = 1
-    device: str = "cuda"
-    storage_device: str = "cpu"
-    # Set to "helper2424/resnet10" for hil serl
-    vision_encoder_name: str | None = None
-    freeze_vision_encoder: bool = True
-    image_encoder_hidden_dim: int = 32
-    shared_encoder: bool = True
-    num_discrete_actions: int | None = None
-
-    # Training parameter
-    online_steps: int = 1000000
-    online_env_seed: int = 10000
-    online_buffer_capacity: int = 100000
-    offline_buffer_capacity: int = 100000
-    async_prefetch: bool = False
-    online_step_before_learning: int = 100
-    policy_update_freq: int = 1
-
-    # SAC algorithm parameters
-    discount: float = 0.99
-    temperature_init: float = 1.0
-    num_critics: int = 2
-    num_subsample_critics: int | None = None
-    critic_lr: float = 3e-4
-    actor_lr: float = 3e-4
-    temperature_lr: float = 3e-4
-    critic_target_update_weight: float = 0.005
-    utd_ratio: int = 1  # If you want enable utd_ratio, you need to set it to >1
-    state_encoder_hidden_dim: int = 256
-    latent_dim: int = 256
-    target_entropy: float | None = None
-    use_backup_entropy: bool = True
-    grad_clip_norm: float = 40.0
-
-    # Network configuration
-    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    actor_network_kwargs: ActorNetworkConfig = field(default_factory=ActorNetworkConfig)
-    policy_kwargs: PolicyConfig = field(default_factory=PolicyConfig)
-    grasp_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
-    concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
-
-    def __post_init__(self):
-        super().__post_init__()
-        # Any validation specific to SAC configuration
-
-    def get_optimizer_preset(self) -> MultiAdamConfig:
-        return MultiAdamConfig(
-            weight_decay=0.0,
-            optimizer_groups={
-                "actor": {"lr": self.actor_lr},
-                "critic": {"lr": self.critic_lr},
-                "temperature": {"lr": self.temperature_lr},
-            },
-        )
-
-    def get_scheduler_preset(self) -> None:
-        return None
-
-    def validate_features(self) -> None:
-        has_image = any(key.startswith("observation.image") for key in self.input_features)
-        has_state = "observation.state" in self.input_features
-
-        if not (has_state or has_image):
-            raise ValueError(
-                "You must provide either 'observation.state' or an image observation (key starting with 'observation.image') in the input features"
-            )
-
-        if "action" not in self.output_features:
-            raise ValueError("You must provide 'action' in the output features")
-
-    @property
-    def image_features(self) -> list[str]:
-        return [key for key in self.input_features if "image" in key]
-
-    @property
-    def observation_delta_indices(self) -> list:
-        return None
-
-    @property
-    def action_delta_indices(self) -> list:
-        return None  # SAC typically predicts one action at a time
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None
--- a/lerobot/common/policies/sac/modeling_sac.py
+++ b/lerobot/common/policies/sac/modeling_sac.py
--- a/lerobot/common/policies/tdmpc/modeling_tdmpc.py
+++ b/lerobot/common/policies/tdmpc/modeling_tdmpc.py
@@ -35,15 +35,11 @@ import torch.nn as nn
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor

-from lerobot.common.constants import OBS_ENV, OBS_ROBOT
+from lerobot.common.constants import OBS_ENV_STATE, OBS_STATE
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
-from lerobot.common.policies.utils import (
-    get_device_from_parameters,
-    get_output_shape,
-    populate_queues,
-)
+from lerobot.common.policies.utils import get_device_from_parameters, get_output_shape, populate_queues


 class TDMPCPolicy(PreTrainedPolicy):
@@ -67,11 +63,7 @@ class TDMPCPolicy(PreTrainedPolicy):
    config_class = TDMPCConfig
    name = "tdmpc"

-    def __init__(
-        self,
-        config: TDMPCConfig,
-        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
+    def __init__(self, config: TDMPCConfig, dataset_stats: dict[str, dict[str, Tensor]] | None = None):
        """
        Args:
            config: Policy configuration class instance or None, in which case the default instantiation of
@@ -130,7 +122,7 @@ class TDMPCPolicy(PreTrainedPolicy):

        # When the action queue is depleted, populate it again by querying the policy.
        if len(self._queues["action"]) == 0:
-            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
+            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch if key in self._queues}

            # Remove the time dimensions as it is not handled yet.
            for key in batch:
@@ -197,20 +189,13 @@ class TDMPCPolicy(PreTrainedPolicy):

        # In the CEM loop we will need this for a call to estimate_value with the gaussian sampled
        # trajectories.
-        z = einops.repeat(
-            z,
-            "b d -> n b d",
-            n=self.config.n_gaussian_samples + self.config.n_pi_samples,
-        )
+        z = einops.repeat(z, "b d -> n b d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)

        # Model Predictive Path Integral (MPPI) with the cross-entropy method (CEM) as the optimization
        # algorithm.
        # The initial mean and standard deviation for the cross-entropy method (CEM).
        mean = torch.zeros(
-            self.config.horizon,
-            batch_size,
-            self.config.action_feature.shape[0],
-            device=device,
+            self.config.horizon, batch_size, self.config.action_feature.shape[0], device=device
        )
        # Maybe warm start CEM with the mean from the previous step.
        if self._prev_mean is not None:
@@ -306,10 +291,9 @@ class TDMPCPolicy(PreTrainedPolicy):
        if self.config.q_ensemble_size > 2:
            G += (
                running_discount
-                * torch.min(
-                    terminal_values[torch.randint(0, self.config.q_ensemble_size, size=(2,))],
-                    dim=0,
-                )[0]
+                * torch.min(terminal_values[torch.randint(0, self.config.q_ensemble_size, size=(2,))], dim=0)[
+                    0
+                ]
            )
        else:
            G += running_discount * torch.min(terminal_values, dim=0)[0]
@@ -345,10 +329,7 @@ class TDMPCPolicy(PreTrainedPolicy):
        # Apply random image augmentations.
        if self.config.image_features and self.config.max_random_shift_ratio > 0:
            observations["observation.image"] = flatten_forward_unflatten(
-                partial(
-                    random_shifts_aug,
-                    max_random_shift_ratio=self.config.max_random_shift_ratio,
-                ),
+                partial(random_shifts_aug, max_random_shift_ratio=self.config.max_random_shift_ratio),
                observations["observation.image"],
            )

@@ -572,10 +553,7 @@ class TDMPCTOLD(nn.Module):
        self._Qs = nn.ModuleList(
            [
                nn.Sequential(
-                    nn.Linear(
-                        config.latent_dim + config.action_feature.shape[0],
-                        config.mlp_dim,
-                    ),
+                    nn.Linear(config.latent_dim + config.action_feature.shape[0], config.mlp_dim),
                    nn.LayerNorm(config.mlp_dim),
                    nn.Tanh(),
                    nn.Linear(config.mlp_dim, config.mlp_dim),
@@ -724,26 +702,11 @@ class TDMPCObservationEncoder(nn.Module):
                    stride=2,
                ),
                nn.ReLU(),
-                nn.Conv2d(
-                    config.image_encoder_hidden_dim,
-                    config.image_encoder_hidden_dim,
-                    5,
-                    stride=2,
-                ),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 5, stride=2),
                nn.ReLU(),
-                nn.Conv2d(
-                    config.image_encoder_hidden_dim,
-                    config.image_encoder_hidden_dim,
-                    3,
-                    stride=2,
-                ),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
                nn.ReLU(),
-                nn.Conv2d(
-                    config.image_encoder_hidden_dim,
-                    config.image_encoder_hidden_dim,
-                    3,
-                    stride=2,
-                ),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
                nn.ReLU(),
            )
            dummy_shape = (1, *next(iter(config.image_features.values())).shape)
@@ -786,14 +749,13 @@ class TDMPCObservationEncoder(nn.Module):
        if self.config.image_features:
            feat.append(
                flatten_forward_unflatten(
-                    self.image_enc_layers,
-                    obs_dict[next(iter(self.config.image_features))],
+                    self.image_enc_layers, obs_dict[next(iter(self.config.image_features))]
                )
            )
        if self.config.env_state_feature:
-            feat.append(self.env_state_enc_layers(obs_dict[OBS_ENV]))
+            feat.append(self.env_state_enc_layers(obs_dict[OBS_ENV_STATE]))
        if self.config.robot_state_feature:
-            feat.append(self.state_enc_layers(obs_dict[OBS_ROBOT]))
+            feat.append(self.state_enc_layers(obs_dict[OBS_STATE]))
        return torch.stack(feat, dim=0).mean(0)


@@ -834,9 +796,7 @@ def update_ema_parameters(ema_net: nn.Module, net: nn.Module, alpha: float):
    """Update EMA parameters in place with ema_param <- alpha * ema_param + (1 - alpha) * param."""
    for ema_module, module in zip(ema_net.modules(), net.modules(), strict=True):
        for (n_p_ema, p_ema), (n_p, p) in zip(
-            ema_module.named_parameters(recurse=False),
-            module.named_parameters(recurse=False),
-            strict=True,
+            ema_module.named_parameters(recurse=False), module.named_parameters(recurse=False), strict=True
        ):
            assert n_p_ema == n_p, "Parameter names don't match for EMA model update"
            if isinstance(p, dict):
--- a/lerobot/common/policies/vqbet/configuration_vqbet.py
+++ b/lerobot/common/policies/vqbet/configuration_vqbet.py
@@ -193,12 +193,7 @@ class VQBeTConfig(PreTrainedConfig):

    @property
    def action_delta_indices(self) -> list:
-        return list(
-            range(
-                1 - self.n_obs_steps,
-                self.n_action_pred_token + self.action_chunk_size - 1,
-            )
-        )
+        return list(range(1 - self.n_obs_steps, self.n_action_pred_token + self.action_chunk_size - 1))

    @property
    def reward_delta_indices(self) -> None:
--- a/lerobot/common/policies/vqbet/modeling_vqbet.py
+++ b/lerobot/common/policies/vqbet/modeling_vqbet.py
@@ -29,11 +29,7 @@ from torch import Tensor, nn

 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pretrained import PreTrainedPolicy
-from lerobot.common.policies.utils import (
-    get_device_from_parameters,
-    get_output_shape,
-    populate_queues,
-)
+from lerobot.common.policies.utils import get_device_from_parameters, get_output_shape, populate_queues
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
 from lerobot.common.policies.vqbet.vqbet_utils import GPT, ResidualVQ

@@ -328,8 +324,7 @@ class VQBeTModel(nn.Module):

        # To input state and observation features into GPT layers, we first project the features to fit the shape of input size of GPT.
        self.state_projector = MLP(
-            config.robot_state_feature.shape[0],
-            hidden_channels=[self.config.gpt_input_dim],
+            config.robot_state_feature.shape[0], hidden_channels=[self.config.gpt_input_dim]
        )
        self.rgb_feature_projector = MLP(
            self.rgb_encoder.feature_dim, hidden_channels=[self.config.gpt_input_dim]
@@ -359,11 +354,7 @@ class VQBeTModel(nn.Module):
        )
        # Separate batch and sequence dims.
        img_features = einops.rearrange(
-            img_features,
-            "(b s n) ... -> b s n ...",
-            b=batch_size,
-            s=n_obs_steps,
-            n=self.num_images,
+            img_features, "(b s n) ... -> b s n ...", b=batch_size, s=n_obs_steps, n=self.num_images
        )

        # Arrange prior and current observation step tokens as shown in the class docstring.
@@ -400,11 +391,7 @@ class VQBeTModel(nn.Module):
        # Thus, it predicts a historical action sequence, in addition to current and future actions (predicting future actions : optional).
        if len_additional_action_token > 0:
            features = torch.cat(
-                [
-                    features[:, historical_act_pred_index],
-                    features[:, -len_additional_action_token:],
-                ],
-                dim=1,
+                [features[:, historical_act_pred_index], features[:, -len_additional_action_token:]], dim=1
            )
        else:
            features = features[:, historical_act_pred_index]
@@ -527,13 +514,7 @@ class VQBeTHead(nn.Module):

            cbet_secondary_logits = self.map_to_cbet_preds_secondary_bin(
                torch.cat(
-                    (
-                        x,
-                        F.one_hot(
-                            sampled_primary_centers,
-                            num_classes=self.config.vqvae_n_embed,
-                        ),
-                    ),
+                    (x, F.one_hot(sampled_primary_centers, num_classes=self.config.vqvae_n_embed)),
                    axis=1,
                )
            )
@@ -551,9 +532,7 @@ class VQBeTHead(nn.Module):
        else:
            cbet_logits = self.map_to_cbet_preds_bin(x)
            cbet_logits = einops.rearrange(
-                cbet_logits,
-                "(NT) (G C) -> (NT) G C",
-                G=self.vqvae_model.vqvae_num_layers,
+                cbet_logits, "(NT) (G C) -> (NT) G C", G=self.vqvae_model.vqvae_num_layers
            )
            cbet_probs = torch.softmax(cbet_logits / self.config.bet_softmax_temperature, dim=-1)
            NT, G, choices = cbet_probs.shape
@@ -751,9 +730,7 @@ class VQBeTRgbEncoder(nn.Module):


 def _replace_submodules(
-    root_module: nn.Module,
-    predicate: Callable[[nn.Module], bool],
-    func: Callable[[nn.Module], nn.Module],
+    root_module: nn.Module, predicate: Callable[[nn.Module], bool], func: Callable[[nn.Module], nn.Module]
 ) -> nn.Module:
    """
    Args:
--- a/lerobot/common/policies/vqbet/vqbet_utils.py
+++ b/lerobot/common/policies/vqbet/vqbet_utils.py
@@ -377,10 +377,7 @@ class ResidualVQ(nn.Module):
        self.layers = nn.ModuleList(
            [
                VectorQuantize(
-                    dim=codebook_dim,
-                    codebook_dim=codebook_dim,
-                    accept_image_fmap=accept_image_fmap,
-                    **kwargs,
+                    dim=codebook_dim, codebook_dim=codebook_dim, accept_image_fmap=accept_image_fmap, **kwargs
                )
                for _ in range(num_quantizers)
            ]
--- a/lerobot/common/robot_devices/cameras/configs.py
+++ b/lerobot/common/robot_devices/cameras/configs.py
@@ -1,114 +0,0 @@
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import abc
-from dataclasses import dataclass
-
-import draccus
-
-
-@dataclass
-class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
-    @property
-    def type(self) -> str:
-        return self.get_choice_name(self.__class__)
-
-
-@CameraConfig.register_subclass("opencv")
-@dataclass
-class OpenCVCameraConfig(CameraConfig):
-    """
-    Example of tested options for Intel Real Sense D405:
-
-    ```python
-    OpenCVCameraConfig(0, 30, 640, 480)
-    OpenCVCameraConfig(0, 60, 640, 480)
-    OpenCVCameraConfig(0, 90, 640, 480)
-    OpenCVCameraConfig(0, 30, 1280, 720)
-    ```
-    """
-
-    camera_index: int
-    fps: int | None = None
-    width: int | None = None
-    height: int | None = None
-    color_mode: str = "rgb"
-    channels: int | None = None
-    rotation: int | None = None
-    mock: bool = False
-
-    def __post_init__(self):
-        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."
-            )
-
-        self.channels = 3
-
-        if self.rotation not in [-90, None, 90, 180]:
-            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")
-
-
-@CameraConfig.register_subclass("intelrealsense")
-@dataclass
-class IntelRealSenseCameraConfig(CameraConfig):
-    """
-    Example of tested options for Intel Real Sense D405:
-
-    ```python
-    IntelRealSenseCameraConfig(128422271347, 30, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 60, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 90, 640, 480)
-    IntelRealSenseCameraConfig(128422271347, 30, 1280, 720)
-    IntelRealSenseCameraConfig(128422271347, 30, 640, 480, use_depth=True)
-    IntelRealSenseCameraConfig(128422271347, 30, 640, 480, rotation=90)
-    ```
-    """
-
-    name: str | None = None
-    serial_number: int | None = None
-    fps: int | None = None
-    width: int | None = None
-    height: int | None = None
-    color_mode: str = "rgb"
-    channels: int | None = None
-    use_depth: bool = False
-    force_hardware_reset: bool = True
-    rotation: int | None = None
-    mock: bool = False
-
-    def __post_init__(self):
-        # bool is stronger than is None, since it works with empty strings
-        if bool(self.name) and bool(self.serial_number):
-            raise ValueError(
-                f"One of them must be set: name or serial_number, but {self.name=} and {self.serial_number=} provided."
-            )
-
-        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."
-            )
-
-        self.channels = 3
-
-        at_least_one_is_not_none = self.fps is not None or self.width is not None or self.height is not None
-        at_least_one_is_none = self.fps is None or self.width is None or self.height is None
-        if at_least_one_is_not_none and at_least_one_is_none:
-            raise ValueError(
-                "For `fps`, `width` and `height`, either all of them need to be set, or none of them, "
-                f"but {self.fps=}, {self.width=}, {self.height=} were provided."
-            )
-
-        if self.rotation not in [-90, None, 90, 180]:
-            raise ValueError(f"`rotation` must be in [-90, None, 90, 180] (got {self.rotation})")
--- a/lerobot/common/robot_devices/cameras/intelrealsense.py
+++ b/lerobot/common/robot_devices/cameras/intelrealsense.py
@@ -1,546 +0,0 @@
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-This file contains utilities for recording frames from Intel Realsense cameras.
-"""
-
-import argparse
-import concurrent.futures
-import logging
-import math
-import shutil
-import threading
-import time
-import traceback
-from collections import Counter
-from pathlib import Path
-from threading import Thread
-
-import numpy as np
-from PIL import Image
-
-from lerobot.common.robot_devices.cameras.configs import IntelRealSenseCameraConfig
-from lerobot.common.robot_devices.utils import (
-    RobotDeviceAlreadyConnectedError,
-    RobotDeviceNotConnectedError,
-    busy_wait,
-)
-from lerobot.common.utils.utils import capture_timestamp_utc
-
-SERIAL_NUMBER_INDEX = 1
-
-
-def find_cameras(raise_when_empty=True, mock=False) -> list[dict]:
-    """
-    Find the names and the serial numbers of the Intel RealSense cameras
-    connected to the computer.
-    """
-    if mock:
-        import tests.cameras.mock_pyrealsense2 as rs
-    else:
-        import pyrealsense2 as rs
-
-    cameras = []
-    for device in rs.context().query_devices():
-        serial_number = int(device.get_info(rs.camera_info(SERIAL_NUMBER_INDEX)))
-        name = device.get_info(rs.camera_info.name)
-        cameras.append(
-            {
-                "serial_number": serial_number,
-                "name": name,
-            }
-        )
-
-    if raise_when_empty and len(cameras) == 0:
-        raise OSError(
-            "Not a single camera was detected. Try re-plugging, or re-installing `librealsense` and its python wrapper `pyrealsense2`, or updating the firmware."
-        )
-
-    return cameras
-
-
-def save_image(img_array, serial_number, frame_index, images_dir):
-    try:
-        img = Image.fromarray(img_array)
-        path = images_dir / f"camera_{serial_number}_frame_{frame_index:06d}.png"
-        path.parent.mkdir(parents=True, exist_ok=True)
-        img.save(str(path), quality=100)
-        logging.info(f"Saved image: {path}")
-    except Exception as e:
-        logging.error(f"Failed to save image for camera {serial_number} frame {frame_index}: {e}")
-
-
-def save_images_from_cameras(
-    images_dir: Path,
-    serial_numbers: list[int] | None = None,
-    fps=None,
-    width=None,
-    height=None,
-    record_time_s=2,
-    mock=False,
-):
-    """
-    Initializes all the cameras and saves images to the directory. Useful to visually identify the camera
-    associated to a given serial number.
-    """
-    if serial_numbers is None or len(serial_numbers) == 0:
-        camera_infos = find_cameras(mock=mock)
-        serial_numbers = [cam["serial_number"] for cam in camera_infos]
-
-    if mock:
-        import tests.cameras.mock_cv2 as cv2
-    else:
-        import cv2
-
-    print("Connecting cameras")
-    cameras = []
-    for cam_sn in serial_numbers:
-        print(f"{cam_sn=}")
-        config = IntelRealSenseCameraConfig(
-            serial_number=cam_sn, fps=fps, width=width, height=height, mock=mock
-        )
-        camera = IntelRealSenseCamera(config)
-        camera.connect()
-        print(
-            f"IntelRealSenseCamera({camera.serial_number}, fps={camera.fps}, width={camera.capture_width}, height={camera.capture_height}, color_mode={camera.color_mode})"
-        )
-        cameras.append(camera)
-
-    images_dir = Path(images_dir)
-    if images_dir.exists():
-        shutil.rmtree(
-            images_dir,
-        )
-    images_dir.mkdir(parents=True, exist_ok=True)
-
-    print(f"Saving images to {images_dir}")
-    frame_index = 0
-    start_time = time.perf_counter()
-    try:
-        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
-            while True:
-                now = time.perf_counter()
-
-                for camera in cameras:
-                    # If we use async_read when fps is None, the loop will go full speed, and we will end up
-                    # saving the same images from the cameras multiple times until the RAM/disk is full.
-                    image = camera.read() if fps is None else camera.async_read()
-                    if image is None:
-                        print("No Frame")
-
-                    bgr_converted_image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
-
-                    executor.submit(
-                        save_image,
-                        bgr_converted_image,
-                        camera.serial_number,
-                        frame_index,
-                        images_dir,
-                    )
-
-                if fps is not None:
-                    dt_s = time.perf_counter() - now
-                    busy_wait(1 / fps - dt_s)
-
-                if time.perf_counter() - start_time > record_time_s:
-                    break
-
-                print(f"Frame: {frame_index:04d}\tLatency (ms): {(time.perf_counter() - now) * 1000:.2f}")
-
-                frame_index += 1
-    finally:
-        print(f"Images have been saved to {images_dir}")
-        for camera in cameras:
-            camera.disconnect()
-
-
-class IntelRealSenseCamera:
-    """
-    The IntelRealSenseCamera class is similar to OpenCVCamera class but adds additional features for Intel Real Sense cameras:
-    - is instantiated with the serial number of the camera - won't randomly change as it can be the case of OpenCVCamera for Linux,
-    - can also be instantiated with the camera's name — if it's unique — using IntelRealSenseCamera.init_from_name(),
-    - depth map can be returned.
-
-    To find the camera indices of your cameras, you can run our utility script that will save a few frames for each camera:
-    ```bash
-    python lerobot/common/robot_devices/cameras/intelrealsense.py --images-dir outputs/images_from_intelrealsense_cameras
-    ```
-
-    When an IntelRealSenseCamera is instantiated, if no specific config is provided, the default fps, width, height and color_mode
-    of the given camera will be used.
-
-    Example of instantiating with a serial number:
-    ```python
-    from lerobot.common.robot_devices.cameras.configs import IntelRealSenseCameraConfig
-
-    config = IntelRealSenseCameraConfig(serial_number=128422271347)
-    camera = IntelRealSenseCamera(config)
-    camera.connect()
-    color_image = camera.read()
-    # when done using the camera, consider disconnecting
-    camera.disconnect()
-    ```
-
-    Example of instantiating with a name if it's unique:
-    ```
-    config = IntelRealSenseCameraConfig(name="Intel RealSense D405")
-    ```
-
-    Example of changing default fps, width, height and color_mode:
-    ```python
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, fps=30, width=1280, height=720)
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480)
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, fps=90, width=640, height=480, color_mode="bgr")
-    # Note: might error out upon `camera.connect()` if these settings are not compatible with the camera
-    ```
-
-    Example of returning depth:
-    ```python
-    config = IntelRealSenseCameraConfig(serial_number=128422271347, use_depth=True)
-    camera = IntelRealSenseCamera(config)
-    camera.connect()
-    color_image, depth_map = camera.read()
-    ```
-    """
-
-    def __init__(
-        self,
-        config: IntelRealSenseCameraConfig,
-    ):
-        self.config = config
-        if config.name is not None:
-            self.serial_number = self.find_serial_number_from_name(config.name)
-        else:
-            self.serial_number = config.serial_number
-
-        # Store the raw (capture) resolution from the config.
-        self.capture_width = config.width
-        self.capture_height = config.height
-
-        # If rotated by ±90, swap width and height.
-        if config.rotation in [-90, 90]:
-            self.width = config.height
-            self.height = config.width
-        else:
-            self.width = config.width
-            self.height = config.height
-
-        self.fps = config.fps
-        self.channels = config.channels
-        self.color_mode = config.color_mode
-        self.use_depth = config.use_depth
-        self.force_hardware_reset = config.force_hardware_reset
-        self.mock = config.mock
-
-        self.camera = None
-        self.is_connected = False
-        self.thread = None
-        self.stop_event = None
-        self.color_image = None
-        self.depth_map = None
-        self.logs = {}
-
-        if self.mock:
-            import tests.cameras.mock_cv2 as cv2
-        else:
-            import cv2
-
-        self.rotation = None
-        if config.rotation == -90:
-            self.rotation = cv2.ROTATE_90_COUNTERCLOCKWISE
-        elif config.rotation == 90:
-            self.rotation = cv2.ROTATE_90_CLOCKWISE
-        elif config.rotation == 180:
-            self.rotation = cv2.ROTATE_180
-
-    def find_serial_number_from_name(self, name):
-        camera_infos = find_cameras()
-        camera_names = [cam["name"] for cam in camera_infos]
-        this_name_count = Counter(camera_names)[name]
-        if this_name_count > 1:
-            # TODO(aliberts): Test this with multiple identical cameras (Aloha)
-            raise ValueError(
-                f"Multiple {name} cameras have been detected. Please use their serial number to instantiate them."
-            )
-
-        name_to_serial_dict = {cam["name"]: cam["serial_number"] for cam in camera_infos}
-        cam_sn = name_to_serial_dict[name]
-
-        return cam_sn
-
-    def connect(self):
-        if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is already connected."
-            )
-
-        if self.mock:
-            import tests.cameras.mock_pyrealsense2 as rs
-        else:
-            import pyrealsense2 as rs
-
-        config = rs.config()
-        config.enable_device(str(self.serial_number))
-
-        if self.fps and self.capture_width and self.capture_height:
-            # TODO(rcadene): can we set rgb8 directly?
-            config.enable_stream(
-                rs.stream.color,
-                self.capture_width,
-                self.capture_height,
-                rs.format.rgb8,
-                self.fps,
-            )
-        else:
-            config.enable_stream(rs.stream.color)
-
-        if self.use_depth:
-            if self.fps and self.capture_width and self.capture_height:
-                config.enable_stream(
-                    rs.stream.depth,
-                    self.capture_width,
-                    self.capture_height,
-                    rs.format.z16,
-                    self.fps,
-                )
-            else:
-                config.enable_stream(rs.stream.depth)
-
-        self.camera = rs.pipeline()
-        try:
-            profile = self.camera.start(config)
-            is_camera_open = True
-        except RuntimeError:
-            is_camera_open = False
-            traceback.print_exc()
-
-        # If the camera doesn't work, display the camera indices corresponding to
-        # valid cameras.
-        if not is_camera_open:
-            # Verify that the provided `serial_number` is valid before printing the traceback
-            camera_infos = find_cameras()
-            serial_numbers = [cam["serial_number"] for cam in camera_infos]
-            if self.serial_number not in serial_numbers:
-                raise ValueError(
-                    f"`serial_number` is expected to be one of these available cameras {serial_numbers}, but {self.serial_number} is provided instead. "
-                    "To find the serial number you should use, run `python lerobot/common/robot_devices/cameras/intelrealsense.py`."
-                )
-
-            raise OSError(f"Can't access IntelRealSenseCamera({self.serial_number}).")
-
-        color_stream = profile.get_stream(rs.stream.color)
-        color_profile = color_stream.as_video_stream_profile()
-        actual_fps = color_profile.fps()
-        actual_width = color_profile.width()
-        actual_height = color_profile.height()
-
-        # Using `math.isclose` since actual fps can be a float (e.g. 29.9 instead of 30)
-        if self.fps is not None and not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
-            # Using `OSError` since it's a broad that encompasses issues related to device communication
-            raise OSError(
-                f"Can't set {self.fps=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_fps}."
-            )
-        if self.capture_width is not None and self.capture_width != actual_width:
-            raise OSError(
-                f"Can't set {self.capture_width=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_width}."
-            )
-        if self.capture_height is not None and self.capture_height != actual_height:
-            raise OSError(
-                f"Can't set {self.capture_height=} for IntelRealSenseCamera({self.serial_number}). Actual value is {actual_height}."
-            )
-
-        self.fps = round(actual_fps)
-        self.capture_width = round(actual_width)
-        self.capture_height = round(actual_height)
-
-        self.is_connected = True
-
-    def read(self, temporary_color: str | None = None) -> np.ndarray | tuple[np.ndarray, np.ndarray]:
-        """Read a frame from the camera returned in the format height x width x channels (e.g. 480 x 640 x 3)
-        of type `np.uint8`, contrarily to the pytorch format which is float channel first.
-
-        When `use_depth=True`, returns a tuple `(color_image, depth_map)` with a depth map in the format
-        height x width (e.g. 480 x 640) of type np.uint16.
-
-        Note: Reading a frame is done every `camera.fps` times per second, and it is blocking.
-        If you are reading data from other sensors, we advise to use `camera.async_read()` which is non blocking version of `camera.read()`.
-        """
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
-            )
-
-        if self.mock:
-            import tests.cameras.mock_cv2 as cv2
-        else:
-            import cv2
-
-        start_time = time.perf_counter()
-
-        frame = self.camera.wait_for_frames(timeout_ms=5000)
-
-        color_frame = frame.get_color_frame()
-
-        if not color_frame:
-            raise OSError(f"Can't capture color image from IntelRealSenseCamera({self.serial_number}).")
-
-        color_image = np.asanyarray(color_frame.get_data())
-
-        requested_color_mode = self.color_mode if temporary_color is None else temporary_color
-        if requested_color_mode not in ["rgb", "bgr"]:
-            raise ValueError(
-                f"Expected color values are 'rgb' or 'bgr', but {requested_color_mode} is provided."
-            )
-
-        # IntelRealSense uses RGB format as default (red, green, blue).
-        if requested_color_mode == "bgr":
-            color_image = cv2.cvtColor(color_image, cv2.COLOR_RGB2BGR)
-
-        h, w, _ = color_image.shape
-        if h != self.capture_height or w != self.capture_width:
-            raise OSError(
-                f"Can't capture color image with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
-            )
-
-        if self.rotation is not None:
-            color_image = cv2.rotate(color_image, self.rotation)
-
-        # log the number of seconds it took to read the image
-        self.logs["delta_timestamp_s"] = time.perf_counter() - start_time
-
-        # log the utc time at which the image was received
-        self.logs["timestamp_utc"] = capture_timestamp_utc()
-
-        if self.use_depth:
-            depth_frame = frame.get_depth_frame()
-            if not depth_frame:
-                raise OSError(f"Can't capture depth image from IntelRealSenseCamera({self.serial_number}).")
-
-            depth_map = np.asanyarray(depth_frame.get_data())
-
-            h, w = depth_map.shape
-            if h != self.capture_height or w != self.capture_width:
-                raise OSError(
-                    f"Can't capture depth map with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
-                )
-
-            if self.rotation is not None:
-                depth_map = cv2.rotate(depth_map, self.rotation)
-
-            return color_image, depth_map
-        else:
-            return color_image
-
-    def read_loop(self):
-        while not self.stop_event.is_set():
-            if self.use_depth:
-                self.color_image, self.depth_map = self.read()
-            else:
-                self.color_image = self.read()
-
-    def async_read(self):
-        """Access the latest color image"""
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
-            )
-
-        if self.thread is None:
-            self.stop_event = threading.Event()
-            self.thread = Thread(target=self.read_loop, args=())
-            self.thread.daemon = True
-            self.thread.start()
-
-        num_tries = 0
-        while self.color_image is None:
-            # TODO(rcadene, aliberts): intelrealsense has diverged compared to opencv over here
-            num_tries += 1
-            time.sleep(1 / self.fps)
-            if num_tries > self.fps and (self.thread.ident is None or not self.thread.is_alive()):
-                raise Exception(
-                    "The thread responsible for `self.async_read()` took too much time to start. There might be an issue. Verify that `self.thread.start()` has been called."
-                )
-
-        if self.use_depth:
-            return self.color_image, self.depth_map
-        else:
-            return self.color_image
-
-    def disconnect(self):
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"IntelRealSenseCamera({self.serial_number}) is not connected. Try running `camera.connect()` first."
-            )
-
-        if self.thread is not None and self.thread.is_alive():
-            # wait for the thread to finish
-            self.stop_event.set()
-            self.thread.join()
-            self.thread = None
-            self.stop_event = None
-
-        self.camera.stop()
-        self.camera = None
-
-        self.is_connected = False
-
-    def __del__(self):
-        if getattr(self, "is_connected", False):
-            self.disconnect()
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        description="Save a few frames using `IntelRealSenseCamera` for all cameras connected to the computer, or a selected subset."
-    )
-    parser.add_argument(
-        "--serial-numbers",
-        type=int,
-        nargs="*",
-        default=None,
-        help="List of serial numbers used to instantiate the `IntelRealSenseCamera`. If not provided, find and use all available camera indices.",
-    )
-    parser.add_argument(
-        "--fps",
-        type=int,
-        default=30,
-        help="Set the number of frames recorded per seconds for all cameras. If not provided, use the default fps of each camera.",
-    )
-    parser.add_argument(
-        "--width",
-        type=str,
-        default=640,
-        help="Set the width for all cameras. If not provided, use the default width of each camera.",
-    )
-    parser.add_argument(
-        "--height",
-        type=str,
-        default=480,
-        help="Set the height for all cameras. If not provided, use the default height of each camera.",
-    )
-    parser.add_argument(
-        "--images-dir",
-        type=Path,
-        default="outputs/images_from_intelrealsense_cameras",
-        help="Set directory to save a few frames for each camera.",
-    )
-    parser.add_argument(
-        "--record-time-s",
-        type=float,
-        default=2.0,
-        help="Set the number of seconds used to record the frames. By default, 2 seconds.",
-    )
-    args = parser.parse_args()
-    save_images_from_cameras(**vars(args))
--- a/lerobot/common/robot_devices/cameras/opencv.py
+++ b/lerobot/common/robot_devices/cameras/opencv.py
@@ -1,518 +0,0 @@
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-This file contains utilities for recording frames from cameras. For more info look at `OpenCVCamera` docstring.
-"""
-
-import argparse
-import concurrent.futures
-import math
-import platform
-import shutil
-import threading
-import time
-from pathlib import Path
-from threading import Thread
-
-import numpy as np
-from PIL import Image
-
-from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
-from lerobot.common.robot_devices.utils import (
-    RobotDeviceAlreadyConnectedError,
-    RobotDeviceNotConnectedError,
-    busy_wait,
-)
-from lerobot.common.utils.utils import capture_timestamp_utc
-
-# The maximum opencv device index depends on your operating system. For instance,
-# if you have 3 cameras, they should be associated to index 0, 1, and 2. This is the case
-# on MacOS. However, on Ubuntu, the indices are different like 6, 16, 23.
-# When you change the USB port or reboot the computer, the operating system might
-# treat the same cameras as new devices. Thus we select a higher bound to search indices.
-MAX_OPENCV_INDEX = 60
-
-
-def find_cameras(raise_when_empty=False, max_index_search_range=MAX_OPENCV_INDEX, mock=False) -> list[dict]:
-    cameras = []
-    if platform.system() == "Linux":
-        print("Linux detected. Finding available camera indices through scanning '/dev/video*' ports")
-        possible_ports = [str(port) for port in Path("/dev").glob("video*")]
-        ports = _find_cameras(possible_ports, mock=mock)
-        for port in ports:
-            cameras.append(
-                {
-                    "port": port,
-                    "index": int(port.removeprefix("/dev/video")),
-                }
-            )
-    else:
-        print(
-            "Mac or Windows detected. Finding available camera indices through "
-            f"scanning all indices from 0 to {MAX_OPENCV_INDEX}"
-        )
-        possible_indices = range(max_index_search_range)
-        indices = _find_cameras(possible_indices, mock=mock)
-        for index in indices:
-            cameras.append(
-                {
-                    "port": None,
-                    "index": index,
-                }
-            )
-
-    return cameras
-
-
-def _find_cameras(
-    possible_camera_ids: list[int | str], raise_when_empty=False, mock=False
-) -> list[int | str]:
-    if mock:
-        import tests.cameras.mock_cv2 as cv2
-    else:
-        import cv2
-
-    camera_ids = []
-    for camera_idx in possible_camera_ids:
-        camera = cv2.VideoCapture(camera_idx)
-        is_open = camera.isOpened()
-        camera.release()
-
-        if is_open:
-            print(f"Camera found at index {camera_idx}")
-            camera_ids.append(camera_idx)
-
-    if raise_when_empty and len(camera_ids) == 0:
-        raise OSError(
-            "Not a single camera was detected. Try re-plugging, or re-installing `opencv2`, "
-            "or your camera driver, or make sure your camera is compatible with opencv2."
-        )
-
-    return camera_ids
-
-
-def is_valid_unix_path(path: str) -> bool:
-    """Note: if 'path' points to a symlink, this will return True only if the target exists"""
-    p = Path(path)
-    return p.is_absolute() and p.exists()
-
-
-def get_camera_index_from_unix_port(port: Path) -> int:
-    return int(str(port.resolve()).removeprefix("/dev/video"))
-
-
-def save_image(img_array, camera_index, frame_index, images_dir):
-    img = Image.fromarray(img_array)
-    path = images_dir / f"camera_{camera_index:02d}_frame_{frame_index:06d}.png"
-    path.parent.mkdir(parents=True, exist_ok=True)
-    img.save(str(path), quality=100)
-
-
-def save_images_from_cameras(
-    images_dir: Path,
-    camera_ids: list | None = None,
-    fps=None,
-    width=None,
-    height=None,
-    record_time_s=2,
-    mock=False,
-):
-    """
-    Initializes all the cameras and saves images to the directory. Useful to visually identify the camera
-    associated to a given camera index.
-    """
-    if camera_ids is None or len(camera_ids) == 0:
-        camera_infos = find_cameras(mock=mock)
-        camera_ids = [cam["index"] for cam in camera_infos]
-
-    print("Connecting cameras")
-    cameras = []
-    for cam_idx in camera_ids:
-        config = OpenCVCameraConfig(camera_index=cam_idx, fps=fps, width=width, height=height, mock=mock)
-        camera = OpenCVCamera(config)
-        camera.connect()
-        print(
-            f"OpenCVCamera({camera.camera_index}, fps={camera.fps}, width={camera.capture_width}, "
-            f"height={camera.capture_height}, color_mode={camera.color_mode})"
-        )
-        cameras.append(camera)
-
-    images_dir = Path(images_dir)
-    if images_dir.exists():
-        shutil.rmtree(
-            images_dir,
-        )
-    images_dir.mkdir(parents=True, exist_ok=True)
-
-    print(f"Saving images to {images_dir}")
-    frame_index = 0
-    start_time = time.perf_counter()
-    with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
-        while True:
-            now = time.perf_counter()
-
-            for camera in cameras:
-                # If we use async_read when fps is None, the loop will go full speed, and we will endup
-                # saving the same images from the cameras multiple times until the RAM/disk is full.
-                image = camera.read() if fps is None else camera.async_read()
-
-                executor.submit(
-                    save_image,
-                    image,
-                    camera.camera_index,
-                    frame_index,
-                    images_dir,
-                )
-
-            if fps is not None:
-                dt_s = time.perf_counter() - now
-                busy_wait(1 / fps - dt_s)
-
-            print(f"Frame: {frame_index:04d}\tLatency (ms): {(time.perf_counter() - now) * 1000:.2f}")
-
-            if time.perf_counter() - start_time > record_time_s:
-                break
-
-            frame_index += 1
-
-    print(f"Images have been saved to {images_dir}")
-
-
-class OpenCVCamera:
-    """
-    The OpenCVCamera class allows to efficiently record images from cameras. It relies on opencv2 to communicate
-    with the cameras. Most cameras are compatible. For more info, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
-
-    An OpenCVCamera instance requires a camera index (e.g. `OpenCVCamera(camera_index=0)`). When you only have one camera
-    like a webcam of a laptop, the camera index is expected to be 0, but it might also be very different, and the camera index
-    might change if you reboot your computer or re-plug your camera. This behavior depends on your operation system.
-
-    To find the camera indices of your cameras, you can run our utility script that will be save a few frames for each camera:
-    ```bash
-    python lerobot/common/robot_devices/cameras/opencv.py --images-dir outputs/images_from_opencv_cameras
-    ```
-
-    When an OpenCVCamera is instantiated, if no specific config is provided, the default fps, width, height and color_mode
-    of the given camera will be used.
-
-    Example of usage:
-    ```python
-    from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
-
-    config = OpenCVCameraConfig(camera_index=0)
-    camera = OpenCVCamera(config)
-    camera.connect()
-    color_image = camera.read()
-    # when done using the camera, consider disconnecting
-    camera.disconnect()
-    ```
-
-    Example of changing default fps, width, height and color_mode:
-    ```python
-    config = OpenCVCameraConfig(camera_index=0, fps=30, width=1280, height=720)
-    config = OpenCVCameraConfig(camera_index=0, fps=90, width=640, height=480)
-    config = OpenCVCameraConfig(camera_index=0, fps=90, width=640, height=480, color_mode="bgr")
-    # Note: might error out open `camera.connect()` if these settings are not compatible with the camera
-    ```
-    """
-
-    def __init__(self, config: OpenCVCameraConfig):
-        self.config = config
-        self.camera_index = config.camera_index
-        self.port = None
-
-        # Linux uses ports for connecting to cameras
-        if platform.system() == "Linux":
-            if isinstance(self.camera_index, int):
-                self.port = Path(f"/dev/video{self.camera_index}")
-            elif isinstance(self.camera_index, str) and is_valid_unix_path(self.camera_index):
-                self.port = Path(self.camera_index)
-                # Retrieve the camera index from a potentially symlinked path
-                self.camera_index = get_camera_index_from_unix_port(self.port)
-            else:
-                raise ValueError(f"Please check the provided camera_index: {self.camera_index}")
-
-        # Store the raw (capture) resolution from the config.
-        self.capture_width = config.width
-        self.capture_height = config.height
-
-        # If rotated by ±90, swap width and height.
-        if config.rotation in [-90, 90]:
-            self.width = config.height
-            self.height = config.width
-        else:
-            self.width = config.width
-            self.height = config.height
-
-        self.fps = config.fps
-        self.channels = config.channels
-        self.color_mode = config.color_mode
-        self.mock = config.mock
-
-        self.camera = None
-        self.is_connected = False
-        self.thread = None
-        self.stop_event = None
-        self.color_image = None
-        self.logs = {}
-
-        if self.mock:
-            import tests.cameras.mock_cv2 as cv2
-        else:
-            import cv2
-
-        self.rotation = None
-        if config.rotation == -90:
-            self.rotation = cv2.ROTATE_90_COUNTERCLOCKWISE
-        elif config.rotation == 90:
-            self.rotation = cv2.ROTATE_90_CLOCKWISE
-        elif config.rotation == 180:
-            self.rotation = cv2.ROTATE_180
-
-    def connect(self):
-        if self.is_connected:
-            raise RobotDeviceAlreadyConnectedError(f"OpenCVCamera({self.camera_index}) is already connected.")
-
-        if self.mock:
-            import tests.cameras.mock_cv2 as cv2
-        else:
-            import cv2
-
-            # Use 1 thread to avoid blocking the main thread. Especially useful during data collection
-            # when other threads are used to save the images.
-            cv2.setNumThreads(1)
-
-        backend = (
-            cv2.CAP_V4L2
-            if platform.system() == "Linux"
-            else cv2.CAP_DSHOW
-            if platform.system() == "Windows"
-            else cv2.CAP_AVFOUNDATION
-            if platform.system() == "Darwin"
-            else cv2.CAP_ANY
-        )
-
-        camera_idx = f"/dev/video{self.camera_index}" if platform.system() == "Linux" else self.camera_index
-        # First create a temporary camera trying to access `camera_index`,
-        # and verify it is a valid camera by calling `isOpened`.
-        tmp_camera = cv2.VideoCapture(camera_idx, backend)
-        is_camera_open = tmp_camera.isOpened()
-        # Release camera to make it accessible for `find_camera_indices`
-        tmp_camera.release()
-        del tmp_camera
-
-        # If the camera doesn't work, display the camera indices corresponding to
-        # valid cameras.
-        if not is_camera_open:
-            # Verify that the provided `camera_index` is valid before printing the traceback
-            cameras_info = find_cameras()
-            available_cam_ids = [cam["index"] for cam in cameras_info]
-            if self.camera_index not in available_cam_ids:
-                raise ValueError(
-                    f"`camera_index` is expected to be one of these available cameras {available_cam_ids}, but {self.camera_index} is provided instead. "
-                    "To find the camera index you should use, run `python lerobot/common/robot_devices/cameras/opencv.py`."
-                )
-
-            raise OSError(f"Can't access OpenCVCamera({camera_idx}).")
-
-        # Secondly, create the camera that will be used downstream.
-        # Note: For some unknown reason, calling `isOpened` blocks the camera which then
-        # needs to be re-created.
-        self.camera = cv2.VideoCapture(camera_idx, backend)
-
-        if self.fps is not None:
-            self.camera.set(cv2.CAP_PROP_FPS, self.fps)
-        if self.capture_width is not None:
-            self.camera.set(cv2.CAP_PROP_FRAME_WIDTH, self.capture_width)
-        if self.capture_height is not None:
-            self.camera.set(cv2.CAP_PROP_FRAME_HEIGHT, self.capture_height)
-
-        actual_fps = self.camera.get(cv2.CAP_PROP_FPS)
-        actual_width = self.camera.get(cv2.CAP_PROP_FRAME_WIDTH)
-        actual_height = self.camera.get(cv2.CAP_PROP_FRAME_HEIGHT)
-
-        # Using `math.isclose` since actual fps can be a float (e.g. 29.9 instead of 30)
-        if self.fps is not None and not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
-            # Using `OSError` since it's a broad that encompasses issues related to device communication
-            raise OSError(
-                f"Can't set {self.fps=} for OpenCVCamera({self.camera_index}). Actual value is {actual_fps}."
-            )
-        if self.capture_width is not None and not math.isclose(
-            self.capture_width, actual_width, rel_tol=1e-3
-        ):
-            raise OSError(
-                f"Can't set {self.capture_width=} for OpenCVCamera({self.camera_index}). Actual value is {actual_width}."
-            )
-        if self.capture_height is not None and not math.isclose(
-            self.capture_height, actual_height, rel_tol=1e-3
-        ):
-            raise OSError(
-                f"Can't set {self.capture_height=} for OpenCVCamera({self.camera_index}). Actual value is {actual_height}."
-            )
-
-        self.fps = round(actual_fps)
-        self.capture_width = round(actual_width)
-        self.capture_height = round(actual_height)
-        self.is_connected = True
-
-    def read(self, temporary_color_mode: str | None = None) -> np.ndarray:
-        """Read a frame from the camera returned in the format (height, width, channels)
-        (e.g. 480 x 640 x 3), contrarily to the pytorch format which is channel first.
-
-        Note: Reading a frame is done every `camera.fps` times per second, and it is blocking.
-        If you are reading data from other sensors, we advise to use `camera.async_read()` which is non blocking version of `camera.read()`.
-        """
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
-            )
-
-        start_time = time.perf_counter()
-
-        ret, color_image = self.camera.read()
-
-        if not ret:
-            raise OSError(f"Can't capture color image from camera {self.camera_index}.")
-
-        requested_color_mode = self.color_mode if temporary_color_mode is None else temporary_color_mode
-
-        if requested_color_mode not in ["rgb", "bgr"]:
-            raise ValueError(
-                f"Expected color values are 'rgb' or 'bgr', but {requested_color_mode} is provided."
-            )
-
-        # OpenCV uses BGR format as default (blue, green, red) for all operations, including displaying images.
-        # However, Deep Learning framework such as LeRobot uses RGB format as default to train neural networks,
-        # so we convert the image color from BGR to RGB.
-        if requested_color_mode == "rgb":
-            if self.mock:
-                import tests.cameras.mock_cv2 as cv2
-            else:
-                import cv2
-
-            color_image = cv2.cvtColor(color_image, cv2.COLOR_BGR2RGB)
-
-        h, w, _ = color_image.shape
-        if h != self.capture_height or w != self.capture_width:
-            raise OSError(
-                f"Can't capture color image with expected height and width ({self.height} x {self.width}). ({h} x {w}) returned instead."
-            )
-
-        if self.rotation is not None:
-            color_image = cv2.rotate(color_image, self.rotation)
-
-        # log the number of seconds it took to read the image
-        self.logs["delta_timestamp_s"] = time.perf_counter() - start_time
-
-        # log the utc time at which the image was received
-        self.logs["timestamp_utc"] = capture_timestamp_utc()
-
-        self.color_image = color_image
-
-        return color_image
-
-    def read_loop(self):
-        while not self.stop_event.is_set():
-            try:
-                self.color_image = self.read()
-            except Exception as e:
-                print(f"Error reading in thread: {e}")
-
-    def async_read(self):
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
-            )
-
-        if self.thread is None:
-            self.stop_event = threading.Event()
-            self.thread = Thread(target=self.read_loop, args=())
-            self.thread.daemon = True
-            self.thread.start()
-
-        num_tries = 0
-        while True:
-            if self.color_image is not None:
-                return self.color_image
-
-            time.sleep(1 / self.fps)
-            num_tries += 1
-            if num_tries > self.fps * 2:
-                raise TimeoutError("Timed out waiting for async_read() to start.")
-
-    def disconnect(self):
-        if not self.is_connected:
-            raise RobotDeviceNotConnectedError(
-                f"OpenCVCamera({self.camera_index}) is not connected. Try running `camera.connect()` first."
-            )
-
-        if self.thread is not None:
-            self.stop_event.set()
-            self.thread.join()  # wait for the thread to finish
-            self.thread = None
-            self.stop_event = None
-
-        self.camera.release()
-        self.camera = None
-        self.is_connected = False
-
-    def __del__(self):
-        if getattr(self, "is_connected", False):
-            self.disconnect()
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        description="Save a few frames using `OpenCVCamera` for all cameras connected to the computer, or a selected subset."
-    )
-    parser.add_argument(
-        "--camera-ids",
-        type=int,
-        nargs="*",
-        default=None,
-        help="List of camera indices used to instantiate the `OpenCVCamera`. If not provided, find and use all available camera indices.",
-    )
-    parser.add_argument(
-        "--fps",
-        type=int,
-        default=None,
-        help="Set the number of frames recorded per seconds for all cameras. If not provided, use the default fps of each camera.",
-    )
-    parser.add_argument(
-        "--width",
-        type=str,
-        default=None,
-        help="Set the width for all cameras. If not provided, use the default width of each camera.",
-    )
-    parser.add_argument(
-        "--height",
-        type=str,
-        default=None,
-        help="Set the height for all cameras. If not provided, use the default height of each camera.",
-    )
-    parser.add_argument(
-        "--images-dir",
-        type=Path,
-        default="outputs/images_from_opencv_cameras",
-        help="Set directory to save a few frames for each camera.",
-    )
-    parser.add_argument(
-        "--record-time-s",
-        type=float,
-        default=4.0,
-        help="Set the number of seconds used to record the frames. By default, 2 seconds.",
-    )
-    args = parser.parse_args()
-    save_images_from_cameras(**vars(args))
--- a/lerobot/common/robot_devices/cameras/utils.py
+++ b/lerobot/common/robot_devices/cameras/utils.py
@@ -1,71 +0,0 @@
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from typing import Protocol
-
-import numpy as np
-
-from lerobot.common.robot_devices.cameras.configs import (
-    CameraConfig,
-    IntelRealSenseCameraConfig,
-    OpenCVCameraConfig,
-)
-
-
-# Defines a camera type
-class Camera(Protocol):
-    def connect(self): ...
-    def read(self, temporary_color: str | None = None) -> np.ndarray: ...
-    def async_read(self) -> np.ndarray: ...
-    def disconnect(self): ...
-
-
-def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> list[Camera]:
-    cameras = {}
-
-    for key, cfg in camera_configs.items():
-        if cfg.type == "opencv":
-            from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
-
-            cameras[key] = OpenCVCamera(cfg)
-
-        elif cfg.type == "intelrealsense":
-            from lerobot.common.robot_devices.cameras.intelrealsense import (
-                IntelRealSenseCamera,
-            )
-
-            cameras[key] = IntelRealSenseCamera(cfg)
-        else:
-            raise ValueError(f"The camera type '{cfg.type}' is not valid.")
-
-    return cameras
-
-
-def make_camera(camera_type, **kwargs) -> Camera:
-    if camera_type == "opencv":
-        from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera
-
-        config = OpenCVCameraConfig(**kwargs)
-        return OpenCVCamera(config)
-
-    elif camera_type == "intelrealsense":
-        from lerobot.common.robot_devices.cameras.intelrealsense import (
-            IntelRealSenseCamera,
-        )
-
-        config = IntelRealSenseCameraConfig(**kwargs)
-        return IntelRealSenseCamera(config)
-
-    else:
-        raise ValueError(f"The camera type '{camera_type}' is not valid.")
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				`../../lerobot/common/robots/koch_follower/koch.md`
				`@@ -0,0 +1 @@`
				`../../lerobot/common/robots/lekiwi/lekiwi.md`
				`@@ -0,0 +1 @@`
				`../../lerobot/common/robots/so100_follower/so100.md`
				`@@ -0,0 +1 @@`
				`../../lerobot/common/robots/so101_follower/so101.md`
				`@@ -0,0 +1 @@`
				`from .motors_bus import Motor, MotorCalibration, MotorNormMode, MotorsBus`