push to specific repo

* fix(ci): move hub & lerobot artefacts to /mnt to avoid No space left on device in the future

* chore(ci): remove dh -h steps

.github/workflows/fast_tests.yml

@@ -60,12 +60,19 @@ jobs:
     runs-on: ubuntu-latest
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
       - uses: actions/checkout@v4
         with:
           persist-credentials: false
           lfs: true
 
+      # NOTE(Steven): Mount to `/mnt` to avoid the limited storage on `/home`. Consider cleaning default SDKs or using self-hosted runners for more space.
+      # (As of 2024-06-10, the runner's `/home` has only 6.2 GB free—8% of its 72 GB total.)
+      - name: Setup /mnt storage
+        run: sudo chown -R $USER:$USER /mnt
+
       # TODO(Steven): Evaluate the need of these dependencies
       - name: Install apt dependencies
         run: |

.github/workflows/full_tests.yml

@@ -58,12 +58,19 @@ jobs:
       github.event_name == 'workflow_dispatch'
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
       - uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
 
+      # NOTE(Steven): Mount to `/mnt` to avoid the limited storage on `/home`. Consider cleaning default SDKs or using self-hosted runners for more space.
+      # (As of 2024-06-10, the runner's `/home` has only 6.2 GB free—8% of its 72 GB total.)
+      - name: Setup /mnt storage
+        run: sudo chown -R $USER:$USER /mnt
+
       - name: Install apt dependencies
         run: |
           sudo apt-get update && sudo apt-get install -y build-essential \

.github/workflows/unbound_deps_tests.yml

@@ -45,12 +45,19 @@ jobs:
     runs-on: ubuntu-latest
     env:
       MUJOCO_GL: egl
+      HF_HOME: /mnt/cache/.cache/huggingface
+      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
       - uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
 
+      # NOTE(Steven): Mount to `/mnt` to avoid the limited storage on `/home`. Consider cleaning default SDKs or using self-hosted runners for more space.
+      # (As of 2024-06-10, the runner's `/home` has only 6.2 GB free—8% of its 72 GB total.)
+      - name: Setup /mnt storage
+        run: sudo chown -R $USER:$USER /mnt
+
       - name: Install apt dependencies
         run: |
           sudo apt-get update && sudo apt-get install -y build-essential \

.gitignore

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

benchmarks/video/benchmark.py

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

benchmarks/video/capture_camera_feed.py

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

benchmarks/video/run_video_benchmark.py

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

docs/source/_toctree.yml

@@ -47,8 +47,8 @@
 - sections:
   - local: envhub
     title: Environments from the Hub
-  - local: il_sim
-    title: Imitation Learning in Sim
+  - local: envhub_leisaac
+    title: Control & Train Robots in Sim (LeIsaac)
   - local: libero
     title: Using Libero
   - local: metaworld
@@ -79,6 +79,8 @@
     title: Hope Jr
   - local: reachy2
     title: Reachy 2
+  - local: unitree_g1
+    title: Unitree G1
   title: "Robots"
 - sections:
   - local: phone_teleop

docs/source/async.mdx

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

docs/source/envhub_leisaac.mdx

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

docs/source/il_robots.mdx

@@ -393,7 +393,7 @@ import time
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 episode_idx = 0
@@ -415,7 +415,7 @@ for idx in range(dataset.num_frames):
     }
     robot.send_action(action)
 
-    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+    precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
 
 robot.disconnect()
 ```

docs/source/il_sim.mdx

@@ -1,220 +0,0 @@
-# Imitation Learning in Sim
-
-This tutorial will explain how to train a neural network to control a robot in simulation with imitation learning.
-
-**You'll learn:**
-
-1. How to record a dataset in simulation with [gym-hil](https://github.com/huggingface/gym-hil) and visualize the dataset.
-2. How to train a policy using your data.
-3. How to evaluate your policy in simulation and visualize the results.
-
-For the simulation environment we use the same [repo](https://github.com/huggingface/gym-hil) that is also being used by the Human-In-the-Loop (HIL) reinforcement learning algorithm.
-This environment is based on [MuJoCo](https://mujoco.org) and allows you to record datasets in LeRobotDataset format.
-Teleoperation is easiest with a controller like the Logitech F710, but you can also use your keyboard if you are up for the challenge.
-
-## Installation
-
-First, install the `gym_hil` package within the LeRobot environment, go to your LeRobot folder and run this command:
-
-```bash
-pip install -e ".[hilserl]"
-```
-
-## Teleoperate and Record a Dataset
-
-To use `gym_hil` with LeRobot, you need to use a configuration file. An example config file can be found [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/sim_il/env_config.json).
-
-To teleoperate and collect a dataset, we need to modify this config file. Here's an example configuration for imitation learning data collection:
-
-```json
-{
-  "env": {
-    "type": "gym_manipulator",
-    "name": "gym_hil",
-    "task": "PandaPickCubeGamepad-v0",
-    "fps": 10
-  },
-  "dataset": {
-    "repo_id": "your_username/il_gym",
-    "root": null,
-    "task": "pick_cube",
-    "num_episodes_to_record": 30,
-    "replay_episode": null,
-    "push_to_hub": true
-  },
-  "mode": "record",
-  "device": "cuda"
-}
-```
-
-Key configuration points:
-
-- Set your `repo_id` in the `dataset` section: `"repo_id": "your_username/il_gym"`
-- Set `num_episodes_to_record: 30` to collect 30 demonstration episodes
-- Ensure `mode` is set to `"record"`
-- If you don't have an NVIDIA GPU, change `"device": "cuda"` to `"mps"` for macOS or `"cpu"`
-- To use keyboard instead of gamepad, change `"task"` to `"PandaPickCubeKeyboard-v0"`
-
-Then we can run this command to start:
-
-<hfoptions id="teleop_sim">
-<hfoption id="Linux">
-
-```bash
-python -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
-```
-
-</hfoption>
-<hfoption id="MacOS">
-
-```bash
-mjpython -m lerobot.rl.gym_manipulator --config_path path/to/env_config_gym_hil_il.json
-```
-
-</hfoption>
-</hfoptions>
-
-Once rendered you can teleoperate the robot with the gamepad or keyboard, below you can find the gamepad/keyboard controls.
-
-Note that to teleoperate the robot you have to hold the "Human Take Over Pause Policy" Button `RB` to enable control!
-
-**Gamepad Controls**
-
-<p align="center">
-  <img
-    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/gamepad_guide.jpg?raw=true"
-    alt="Figure shows the control mappings on a Logitech gamepad."
-    title="Gamepad Control Mapping"
-    width="100%"
-  ></img>
-</p>
-<p align="center">
-  <i>Gamepad button mapping for robot control and episode management</i>
-</p>
-
-**Keyboard controls**
-
-For keyboard controls use the `spacebar` to enable control and the following keys to move the robot:
-
-```bash
-  Arrow keys: Move in X-Y plane
-  Shift and Shift_R: Move in Z axis
-  Right Ctrl and Left Ctrl: Open and close gripper
-  ESC: Exit
-```
-
-## Visualize a dataset
-
-If you uploaded your dataset to the hub you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id.
-
-<p align="center">
-  <img
-    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/dataset_visualizer_sim.png"
-    alt="Figure shows the dataset visualizer"
-    title="Dataset visualization"
-    width="100%"
-  ></img>
-</p>
-<p align="center">
-  <i>Dataset visualizer</i>
-</p>
-
-## Train a policy
-
-To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-
-```bash
-lerobot-train \
-  --dataset.repo_id=${HF_USER}/il_gym \
-  --policy.type=act \
-  --output_dir=outputs/train/il_sim_test \
-  --job_name=il_sim_test \
-  --policy.device=cuda \
-  --wandb.enable=true
-```
-
-Let's explain the command:
-
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/il_gym`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-3. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
-4. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
-
-Training should take several hours, 100k steps (which is the default) will take about 1h on Nvidia A100. You will find checkpoints in `outputs/train/il_sim_test/checkpoints`.
-
-#### Train using Collab
-
-If your local computer doesn't have a powerful GPU you could utilize Google Collab to train your model by following the [ACT training notebook](./notebooks#training-act).
-
-#### Upload policy checkpoints
-
-Once training is done, upload the latest checkpoint with:
-
-```bash
-huggingface-cli upload ${HF_USER}/il_sim_test \
-  outputs/train/il_sim_test/checkpoints/last/pretrained_model
-```
-
-You can also upload intermediate checkpoints with:
-
-```bash
-CKPT=010000
-huggingface-cli upload ${HF_USER}/il_sim_test${CKPT} \
-  outputs/train/il_sim_test/checkpoints/${CKPT}/pretrained_model
-```
-
-## Evaluate your policy in Sim
-
-To evaluate your policy we have to use a configuration file. An example can be found [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/sim_il/eval_config.json).
-
-Here's an example evaluation configuration:
-
-```json
-{
-  "env": {
-    "type": "gym_manipulator",
-    "name": "gym_hil",
-    "task": "PandaPickCubeGamepad-v0",
-    "fps": 10
-  },
-  "dataset": {
-    "repo_id": "your_username/il_sim_dataset",
-    "dataset_root": null,
-    "task": "pick_cube"
-  },
-  "pretrained_policy_name_or_path": "your_username/il_sim_model",
-  "device": "cuda"
-}
-```
-
-Make sure to replace:
-
-- `repo_id` with the dataset you trained on (e.g., `your_username/il_sim_dataset`)
-- `pretrained_policy_name_or_path` with your model ID (e.g., `your_username/il_sim_model`)
-
-Then you can run this command to visualize your trained policy
-
-<hfoptions id="eval_policy">
-<hfoption id="Linux">
-
-```bash
-python -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
-```
-
-</hfoption>
-<hfoption id="MacOS">
-
-```bash
-mjpython -m lerobot.rl.eval_policy --config_path=path/to/eval_config_gym_hil.json
-```
-
-</hfoption>
-</hfoptions>
-
-> [!WARNING]
-> While the main workflow of training ACT in simulation is straightforward, there is significant room for exploring how to set up the task, define the initial state of the environment, and determine the type of data required during collection to learn the most effective policy. If your trained policy doesn't perform well, investigate the quality of the dataset it was trained on using our visualizers, as well as the action values and various hyperparameters related to ACT and the simulation.
-
-Congrats 🎉, you have finished this tutorial. If you want to continue with using LeRobot in simulation follow this [Tutorial on reinforcement learning in sim with HIL-SERL](https://huggingface.co/docs/lerobot/hilserl_sim)
-
-> [!TIP]
-> If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb).

docs/source/unitree_g1.mdx

@@ -0,0 +1,203 @@
+# Unitree G1 Robot Setup and Control
+
+This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
+
+## About the Unitree G1
+
+We offer support for both 29 and 23 DOF G1. In this first PR we introduce:
+
+- **`unitree g1` robot class, handling low level communication with the humanoid**
+- **ZMQ socket bridge** for remote communication over WiFi, allowing one to deploy policies remotely instead of over ethernet or directly on the Orin
+- **GR00T locomotion policy** for bipedal walking and balance
+
+---
+
+## Part 1: Connect to Robot over Ethernet
+
+### Step 1: Configure Your Computer's Ethernet Interface
+
+Set a static IP on the same subnet as the robot:
+
+```bash
+# Replace 'enp131s0' with your ethernet interface name (check with `ip a`)
+sudo ip addr flush dev enp131s0
+sudo ip addr add 192.168.123.200/24 dev enp131s0
+sudo ip link set enp131s0 up
+```
+
+**Note**: The robot's Ethernet IP is fixed at `192.168.123.164`. Your computer must use `192.168.123.x` where x ≠ 164.
+
+### Step 2: SSH into the Robot
+
+```bash
+ssh unitree@192.168.123.164
+# Password: 123
+```
+
+You should now be connected to the robot's onboard computer.
+
+---
+
+## Part 2: Enable WiFi on the Robot
+
+Once connected via Ethernet, follow these steps to enable WiFi:
+
+### Step 1: Enable WiFi Hardware
+
+```bash
+# Unblock WiFi radio
+sudo rfkill unblock wifi
+sudo rfkill unblock all
+
+# Bring up WiFi interface
+sudo ip link set wlan0 up
+
+# Enable NetworkManager control
+sudo nmcli radio wifi on
+sudo nmcli device set wlan0 managed yes
+sudo systemctl restart NetworkManager
+```
+
+### Step 2: Enable Internet Forwarding
+
+**On your laptop:**
+
+```bash
+# Enable IP forwarding
+sudo sysctl -w net.ipv4.ip_forward=1
+
+# Set up NAT (replace wlp132s0f0 with your WiFi interface)
+sudo iptables -t nat -A POSTROUTING -o wlp132s0f0 -s 192.168.123.0/24 -j MASQUERADE
+sudo iptables -A FORWARD -i wlp132s0f0 -o enp131s0 -m state --state RELATED,ESTABLISHED -j ACCEPT
+sudo iptables -A FORWARD -i enp131s0 -o wlp132s0f0 -j ACCEPT
+```
+
+**On the robot:**
+
+```bash
+# Add laptop as default gateway
+sudo ip route del default 2>/dev/null || true
+sudo ip route add default via 192.168.123.200 dev eth0
+echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf
+
+# Test connection
+ping -c 3 8.8.8.8
+```
+
+### Step 3: Connect to WiFi Network
+
+```bash
+# List available networks
+nmcli device wifi list
+
+# Connect to your WiFi (example)
+sudo nmcli connection add type wifi ifname wlan0 con-name "YourNetwork" ssid "YourNetwork"
+sudo nmcli connection modify "YourNetwork" wifi-sec.key-mgmt wpa-psk
+sudo nmcli connection modify "YourNetwork" wifi-sec.psk "YourPassword"
+sudo nmcli connection modify "YourNetwork" connection.autoconnect yes
+sudo nmcli connection up "YourNetwork"
+
+# Check WiFi IP address
+ip a show wlan0
+```
+
+### Step 4: SSH Over WiFi
+
+Once connected to WiFi, note the robot's IP address and disconnect the Ethernet cable. You can now SSH over WiFi:
+
+```bash
+ssh unitree@<YOUR_ROBOT_IP>
+# Password: 123
+```
+
+Replace `<YOUR_ROBOT_IP>` with your robot's actual WiFi IP address (e.g., `172.18.129.215`).
+
+---
+
+## Part 3: Robot Server Setup
+
+### Step 1: Install LeRobot on the Orin
+
+SSH into the robot and install LeRobot:
+
+```bash
+ssh unitree@<YOUR_ROBOT_IP>
+
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+**Note**: The Unitree SDK requires CycloneDDS v0.10.2 to be installed. See the [Unitree SDK documentation](https://github.com/unitreerobotics/unitree_sdk2_python) for details.
+
+### Step 2: Run the Robot Server
+
+On the robot:
+
+```bash
+python src/lerobot/robots/unitree_g1/run_g1_server.py
+```
+
+**Important**: Keep this terminal running. The server must be active for remote control.
+
+---
+
+## Part 4: Running GR00T Locomotion
+
+With the robot server running, you can now control the robot from your laptop.
+
+### Step 1: Install LeRobot on your machine
+
+```bash
+conda create -y -n lerobot python=3.10
+conda activate lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e '.[unitree_g1]'
+git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
+cd unitree_sdk2_python  && pip install -e .
+```
+
+### Step 2: Update Robot IP in Config
+
+Edit the config file to match your robot's WiFi IP:
+
+```python
+# In src/lerobot/robots/unitree_g1/config_unitree_g1.py
+robot_ip: str = "<YOUR_ROBOT_IP>"  # Replace with your robot's WiFi IP.
+```
+
+**Note**: When running directly on the G1 (not remotely), set `robot_ip: str = "127.0.0.1"` instead.
+
+### Step 3: Run the Locomotion Policy
+
+```bash
+# Run GR00T locomotion controller
+python examples/unitree_g1/gr00t_locomotion.py --repo-id "nepyope/GR00T-WholeBodyControl_g1"
+```
+
+### Step 4: Control with Remote
+
+- **Left stick**: Forward/backward and left/right movement
+- **Right stick**: Rotation
+- **R1 button**: Raise waist height
+- **R2 button**: Lower waist height
+
+Press `Ctrl+C` to stop the policy.
+
+---
+
+## Additional Resources
+
+- [Unitree SDK Documentation](https://github.com/unitreerobotics/unitree_sdk2_python)
+- [GR00T Policy Repository](https://huggingface.co/nepyope/GR00T-WholeBodyControl_g1)
+- [LeRobot Documentation](https://github.com/huggingface/lerobot)
+- [Unitree_IL_Lerobot](https://github.com/unitreerobotics/unitree_IL_lerobot)
+
+---
+
+_Last updated: December 2025_

examples/backward_compatibility/replay.py

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

examples/dataset/aggregate_egodex.py

@@ -0,0 +1,245 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Aggregate EgoDex shards into a single dataset.
+
+After distributed processing creates multiple shards, this script combines
+them into a single unified dataset.
+
+Reference: https://arxiv.org/abs/2505.11709, https://github.com/apple/ml-egodex
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+
+
+class AggregateEgoDexDatasets(PipelineStep):
+    """Datatrove pipeline step for aggregating EgoDex shards."""
+
+    def __init__(
+        self,
+        repo_ids: list[str],
+        aggregated_repo_id: str,
+        local_dir: Path | str | None = None,
+        push_to_hub: bool = False,
+        hf_repo_id: str | None = None,
+    ):
+        super().__init__()
+        self.repo_ids = repo_ids
+        self.aggr_repo_id = aggregated_repo_id
+        self.local_dir = Path(local_dir) if local_dir else None
+        self.push_to_hub = push_to_hub
+        self.hf_repo_id = hf_repo_id if hf_repo_id else aggregated_repo_id
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+
+        from lerobot.datasets.aggregate import aggregate_datasets
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+        from lerobot.utils.utils import init_logging
+
+        init_logging()
+
+        # Only worker 0 performs aggregation (aggregate_datasets handles parallelism internally)
+        if rank == 0:
+            logging.info(f"Starting aggregation of {len(self.repo_ids)} shards into {self.aggr_repo_id}")
+
+            # Build roots list if local_dir is specified
+            roots = None
+            if self.local_dir:
+                roots = [self.local_dir / repo_id for repo_id in self.repo_ids]
+                # Filter to only existing directories
+                existing_roots = [r for r in roots if r.exists()]
+                if len(existing_roots) != len(self.repo_ids):
+                    logging.warning(
+                        f"Only {len(existing_roots)} of {len(self.repo_ids)} shard directories found. "
+                        "Missing shards will be skipped."
+                    )
+                # Update repo_ids to match existing roots
+                existing_repo_ids = [
+                    repo_id for repo_id, r in zip(self.repo_ids, roots, strict=False) if r.exists()
+                ]
+                roots = existing_roots
+                self.repo_ids = existing_repo_ids
+
+            if len(self.repo_ids) == 0:
+                logging.error("No shard directories found. Nothing to aggregate.")
+                return
+
+            aggr_root = self.local_dir / self.aggr_repo_id if self.local_dir else None
+
+            aggregate_datasets(
+                repo_ids=self.repo_ids,
+                aggr_repo_id=self.aggr_repo_id,
+                roots=roots,
+                aggr_root=aggr_root,
+            )
+            logging.info("Aggregation complete!")
+
+            # Push to Hugging Face Hub if requested
+            if self.push_to_hub:
+                logging.info(f"Pushing to Hugging Face Hub as {self.hf_repo_id}...")
+                dataset = LeRobotDataset(
+                    repo_id=self.aggr_repo_id,
+                    root=aggr_root,
+                )
+                # Update repo_id for pushing to different HF account if specified
+                dataset.repo_id = self.hf_repo_id
+                dataset.push_to_hub(
+                    tags=["egodex", "hand", "dexterous", "lerobot"],
+                    license="cc-by-nc-nd-4.0",
+                )
+                logging.info("Push to hub complete!")
+        else:
+            logging.info(f"Worker {rank} skipping - only worker 0 performs aggregation")
+
+
+def make_aggregate_executor(
+    repo_id,
+    num_shards,
+    job_name,
+    logs_dir,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+    local_dir,
+    push_to_hub,
+    hf_repo_id,
+    slurm=True,
+):
+    """Create executor for aggregating EgoDex shards."""
+    # Generate repo IDs for all shards
+    repo_ids = [f"{repo_id}_world_{num_shards}_rank_{rank}" for rank in range(num_shards)]
+
+    kwargs = {
+        "pipeline": [
+            AggregateEgoDexDatasets(repo_ids, repo_id, local_dir, push_to_hub, hf_repo_id),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": 1,  # Only need 1 task for aggregation
+                "workers": 1,  # Only need 1 worker
+                "time": "24:00:00",  # 24 hours for aggregation
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        executor = SlurmPipelineExecutor(**kwargs)
+    else:
+        kwargs.update(
+            {
+                "tasks": 1,
+                "workers": 1,
+            }
+        )
+        executor = LocalPipelineExecutor(**kwargs)
+
+    return executor
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Aggregate EgoDex dataset shards into a single unified dataset."
+    )
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Repository identifier (base name without shard suffix, e.g., pepijn/egodex-test)",
+    )
+    parser.add_argument(
+        "--num-shards",
+        type=int,
+        required=True,
+        help="Number of shards to aggregate (must match --workers from slurm_port_egodex.py)",
+    )
+    parser.add_argument(
+        "--logs-dir",
+        type=Path,
+        default=Path("logs"),
+        help="Path to logs directory for datatrove",
+    )
+    parser.add_argument(
+        "--job-name",
+        type=str,
+        default="aggr_egodex",
+        help="Job name used in SLURM",
+    )
+    parser.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="Launch over SLURM. Use --slurm 0 to launch locally (for debugging)",
+    )
+    parser.add_argument(
+        "--partition",
+        type=str,
+        help="SLURM partition (ideally CPU partition)",
+    )
+    parser.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=16,
+        help="Number of CPUs for aggregation task",
+    )
+    parser.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="8G",
+        help="Memory per CPU for aggregation",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="Local directory where shards are stored. If not specified, uses default HF cache.",
+    )
+    parser.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Push aggregated dataset to Hugging Face Hub after aggregation.",
+    )
+    parser.add_argument(
+        "--hf-repo-id",
+        type=str,
+        default=None,
+        help="Hugging Face repo ID for upload (e.g., username/dataset-name). Defaults to --repo-id.",
+    )
+
+    args = parser.parse_args()
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+
+    aggregate_executor = make_aggregate_executor(**kwargs)
+    aggregate_executor.run()
+
+
+if __name__ == "__main__":
+    main()
+

examples/dataset/download_egodex.sh

@@ -0,0 +1,129 @@
+#!/bin/bash
+
+# Download EgoDex dataset
+# Reference: https://arxiv.org/abs/2505.11709, https://github.com/apple/ml-egodex
+#
+# Usage: ./download_egodex.sh [output_dir] [parts...]
+#
+# Examples:
+#   ./download_egodex.sh ./data test           # Download test set only (16 GB)
+#   ./download_egodex.sh ./data part1 part2    # Download training parts 1 and 2
+#   ./download_egodex.sh ./data all            # Download everything (~1.7 TB)
+#
+# Available parts:
+#   test   - Test set (16 GB)
+#   part1  - Training set part 1 (300 GB)
+#   part2  - Training set part 2 (300 GB)
+#   part3  - Training set part 3 (300 GB)
+#   part4  - Training set part 4 (300 GB)
+#   part5  - Training set part 5 (300 GB)
+#   extra  - Additional data (200 GB)
+#   all    - Download all parts (~1.7 TB total)
+
+set -e
+
+BASE_URL="https://ml-site.cdn-apple.com/datasets/egodex"
+
+# Map part names to filenames
+declare -A PART_FILES=(
+    ["test"]="test.zip"
+    ["part1"]="part1.zip"
+    ["part2"]="part2.zip"
+    ["part3"]="part3.zip"
+    ["part4"]="part4.zip"
+    ["part5"]="part5.zip"
+    ["extra"]="extra.zip"
+)
+
+ALL_PARTS=("test" "part1" "part2" "part3" "part4" "part5" "extra")
+
+usage() {
+    echo "Usage: $0 <output_dir> <parts...>"
+    echo ""
+    echo "Examples:"
+    echo "  $0 ./data test           # Download test set only (16 GB)"
+    echo "  $0 ./data part1 part2    # Download training parts 1 and 2"
+    echo "  $0 ./data all            # Download everything (~1.7 TB)"
+    echo ""
+    echo "Available parts: test, part1, part2, part3, part4, part5, extra, all"
+    exit 1
+}
+
+download_part() {
+    local output_dir="$1"
+    local part="$2"
+    local filename="${PART_FILES[$part]}"
+    local url="${BASE_URL}/${filename}"
+    local output_file="${output_dir}/${filename}"
+
+    echo "----------------------------------------"
+    echo "Downloading: ${part} (${filename})"
+    echo "URL: ${url}"
+    echo "Output: ${output_file}"
+    echo "----------------------------------------"
+
+    # Download with curl, showing progress
+    curl -L --progress-bar "${url}" -o "${output_file}"
+
+    # Unzip
+    echo "Extracting ${filename}..."
+    unzip -q "${output_file}" -d "${output_dir}"
+
+    # Optionally remove zip file to save space
+    # Uncomment the next line if you want to delete zips after extraction
+    # rm "${output_file}"
+
+    echo "Done: ${part}"
+    echo ""
+}
+
+# Check arguments
+if [ $# -lt 2 ]; then
+    usage
+fi
+
+OUTPUT_DIR="$1"
+shift
+
+# Create output directory
+mkdir -p "${OUTPUT_DIR}"
+
+# Determine which parts to download
+PARTS_TO_DOWNLOAD=()
+
+for arg in "$@"; do
+    if [ "$arg" == "all" ]; then
+        PARTS_TO_DOWNLOAD=("${ALL_PARTS[@]}")
+        break
+    elif [ -n "${PART_FILES[$arg]}" ]; then
+        PARTS_TO_DOWNLOAD+=("$arg")
+    else
+        echo "Error: Unknown part '${arg}'"
+        echo "Available parts: test, part1, part2, part3, part4, part5, extra, all"
+        exit 1
+    fi
+done
+
+if [ ${#PARTS_TO_DOWNLOAD[@]} -eq 0 ]; then
+    echo "Error: No valid parts specified"
+    usage
+fi
+
+echo "========================================"
+echo "EgoDex Dataset Download"
+echo "========================================"
+echo "Output directory: ${OUTPUT_DIR}"
+echo "Parts to download: ${PARTS_TO_DOWNLOAD[*]}"
+echo "========================================"
+echo ""
+
+# Download each part
+for part in "${PARTS_TO_DOWNLOAD[@]}"; do
+    download_part "${OUTPUT_DIR}" "${part}"
+done
+
+echo "========================================"
+echo "Download complete!"
+echo "Data saved to: ${OUTPUT_DIR}"
+echo "========================================"
+

examples/dataset/load_lerobot_dataset.py

@@ -34,105 +34,106 @@ from huggingface_hub import HfApi
 import lerobot
 from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 
-# We ported a number of existing datasets ourselves, use this to see the list:
-print("List of available datasets:")
-pprint(lerobot.available_datasets)
 
-# You can also browse through the datasets created/ported by the community on the hub using the hub api:
-hub_api = HfApi()
-repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
-pprint(repo_ids)
+def main():
+    # We ported a number of existing datasets ourselves, use this to see the list:
+    print("List of available datasets:")
+    pprint(lerobot.available_datasets)
 
-# Or simply explore them in your web browser directly at:
-# https://huggingface.co/datasets?other=LeRobot
+    # You can also browse through the datasets created/ported by the community on the hub using the hub api:
+    hub_api = HfApi()
+    repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
+    pprint(repo_ids)
 
-# Let's take this one for this example
-repo_id = "lerobot/aloha_mobile_cabinet"
-# We can have a look and fetch its metadata to know more about it:
-ds_meta = LeRobotDatasetMetadata(repo_id)
+    # Or simply explore them in your web browser directly at:
+    # https://huggingface.co/datasets?other=LeRobot
 
-# By instantiating just this class, you can quickly access useful information about the content and the
-# structure of the dataset without downloading the actual data yet (only metadata files — which are
-# lightweight).
-print(f"Total number of episodes: {ds_meta.total_episodes}")
-print(f"Average number of frames per episode: {ds_meta.total_frames / ds_meta.total_episodes:.3f}")
-print(f"Frames per second used during data collection: {ds_meta.fps}")
-print(f"Robot type: {ds_meta.robot_type}")
-print(f"keys to access images from cameras: {ds_meta.camera_keys=}\n")
+    # Let's take this one for this example
+    repo_id = "lerobot/aloha_mobile_cabinet"
+    # We can have a look and fetch its metadata to know more about it:
+    ds_meta = LeRobotDatasetMetadata(repo_id)
 
-print("Tasks:")
-print(ds_meta.tasks)
-print("Features:")
-pprint(ds_meta.features)
+    # By instantiating just this class, you can quickly access useful information about the content and the
+    # structure of the dataset without downloading the actual data yet (only metadata files — which are
+    # lightweight).
+    print(f"Total number of episodes: {ds_meta.total_episodes}")
+    print(f"Average number of frames per episode: {ds_meta.total_frames / ds_meta.total_episodes:.3f}")
+    print(f"Frames per second used during data collection: {ds_meta.fps}")
+    print(f"Robot type: {ds_meta.robot_type}")
+    print(f"keys to access images from cameras: {ds_meta.camera_keys=}\n")
 
-# You can also get a short summary by simply printing the object:
-print(ds_meta)
+    print("Tasks:")
+    print(ds_meta.tasks)
+    print("Features:")
+    pprint(ds_meta.features)
 
-# You can then load the actual dataset from the hub.
-# Either load any subset of episodes:
-dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
+    # You can also get a short summary by simply printing the object:
+    print(ds_meta)
 
-# And see how many frames you have:
-print(f"Selected episodes: {dataset.episodes}")
-print(f"Number of episodes selected: {dataset.num_episodes}")
-print(f"Number of frames selected: {dataset.num_frames}")
+    # You can then load the actual dataset from the hub.
+    # Either load any subset of episodes:
+    dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
 
-# Or simply load the entire dataset:
-dataset = LeRobotDataset(repo_id)
-print(f"Number of episodes selected: {dataset.num_episodes}")
-print(f"Number of frames selected: {dataset.num_frames}")
+    # And see how many frames you have:
+    print(f"Selected episodes: {dataset.episodes}")
+    print(f"Number of episodes selected: {dataset.num_episodes}")
+    print(f"Number of frames selected: {dataset.num_frames}")
 
-# The previous metadata class is contained in the 'meta' attribute of the dataset:
-print(dataset.meta)
+    # Or simply load the entire dataset:
+    dataset = LeRobotDataset(repo_id)
+    print(f"Number of episodes selected: {dataset.num_episodes}")
+    print(f"Number of frames selected: {dataset.num_frames}")
 
-# LeRobotDataset actually wraps an underlying Hugging Face dataset
-# (see https://huggingface.co/docs/datasets for more information).
-print(dataset.hf_dataset)
+    # The previous metadata class is contained in the 'meta' attribute of the dataset:
+    print(dataset.meta)
 
-# LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
-# with the latter, like iterating through the dataset.
-# The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
-# episodes, you can access the frame indices of any episode using dataset.meta.episodes. Here, we access
-# frame indices associated to the first episode:
-episode_index = 0
-from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
-to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
+    # LeRobotDataset actually wraps an underlying Hugging Face dataset
+    # (see https://huggingface.co/docs/datasets for more information).
+    print(dataset.hf_dataset)
 
-# Then we grab all the image frames from the first camera:
-camera_key = dataset.meta.camera_keys[0]
-frames = [dataset[idx][camera_key] for idx in range(from_idx, to_idx)]
+    # LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
+    # with the latter, like iterating through the dataset.
+    # The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
+    # episodes, you can access the frame indices of any episode using dataset.meta.episodes. Here, we access
+    # frame indices associated to the first episode:
+    episode_index = 0
+    from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
+    to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
 
-# The objects returned by the dataset are all torch.Tensors
-print(type(frames[0]))
-print(frames[0].shape)
+    # Then we grab all the image frames from the first camera:
+    camera_key = dataset.meta.camera_keys[0]
+    frames = [dataset[idx][camera_key] for idx in range(from_idx, to_idx)]
 
-# Since we're using pytorch, the shape is in pytorch, channel-first convention (c, h, w).
-# We can compare this shape with the information available for that feature
-pprint(dataset.features[camera_key])
-# In particular:
-print(dataset.features[camera_key]["shape"])
-# The shape is in (h, w, c) which is a more universal format.
+    # The objects returned by the dataset are all torch.Tensors
+    print(type(frames[0]))
+    print(frames[0].shape)
 
-# For many machine learning applications we need to load the history of past observations or trajectories of
-# future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
-# differences with the current loaded frame. For instance:
-delta_timestamps = {
-    # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
-    camera_key: [-1, -0.5, -0.20, 0],
-    # loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
-    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
-    # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
-    "action": [t / dataset.fps for t in range(64)],
-}
-# Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
-# timestamp, you still get a valid timestamp.
+    # Since we're using pytorch, the shape is in pytorch, channel-first convention (c, h, w).
+    # We can compare this shape with the information available for that feature
+    pprint(dataset.features[camera_key])
+    # In particular:
+    print(dataset.features[camera_key]["shape"])
+    # The shape is in (h, w, c) which is a more universal format.
 
-dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
-print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
-print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
-print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
+    # For many machine learning applications we need to load the history of past observations or trajectories of
+    # future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
+    # differences with the current loaded frame. For instance:
+    delta_timestamps = {
+        # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
+        camera_key: [-1, -0.5, -0.20, 0],
+        # loads 6 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
+        "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
+        # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
+        "action": [t / dataset.fps for t in range(64)],
+    }
+    # Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
+    # timestamp, you still get a valid timestamp.
+
+    dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
+    print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
+    print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
+    print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
 
-if __name__ == "__main__":
     dataloader = torch.utils.data.DataLoader(
         dataset,
         num_workers=4,
@@ -144,3 +145,7 @@ if __name__ == "__main__":
         print(f"{batch['observation.state'].shape=}")  # (32, 6, c)
         print(f"{batch['action'].shape=}")  # (32, 64, c)
         break
+
+
+if __name__ == "__main__":
+    main()

examples/dataset/slurm_port_egodex.py

@@ -0,0 +1,443 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Distributed EgoDex dataset porting using SLURM and datatrove.
+
+EgoDex is a large-scale dataset for egocentric dexterous manipulation collected
+with ARKit on Apple Vision Pro. This script converts EgoDex data to LeRobot format.
+
+Reference: https://arxiv.org/abs/2505.11709, https://github.com/apple/ml-egodex 
+"""
+
+import argparse
+from pathlib import Path
+
+import cv2
+import h5py
+import mediapy as mpy
+import numpy as np
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+# Image dimensions
+DEFAULT_IMAGE_HEIGHT = 1080
+DEFAULT_IMAGE_WIDTH = 1920
+
+class PortEgoDexShards(PipelineStep):
+    def __init__(
+        self,
+        raw_dir: Path | str,
+        repo_id: str,
+        local_dir: Path | str = None,
+        percentage: float = 100.0,
+    ):
+        super().__init__()
+        self.raw_dir = Path(raw_dir)
+        self.repo_id = repo_id
+        self.local_dir = Path(local_dir) if local_dir else Path("data/local_datasets")
+        self.percentage = percentage
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        from pathlib import Path
+
+        import cv2
+        import h5py
+        import mediapy as mpy
+        import numpy as np
+
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+        from lerobot.utils.utils import init_logging
+
+        def _get_state_for_single_frame(transforms_group, frame_idx):
+            """
+            Construct 48D hand state representation from EgoDex.
+            
+            State vector composition (per hand = 24D, total = 48D):
+            - Wrist 3D position (3)
+            - Wrist orientation in 6D representation (6)
+            - 5 fingertip 3D positions (15)
+            """
+            state_vector = []
+            fingertip_joints = {
+                "left": [
+                    "leftThumbTip",
+                    "leftIndexFingerTip",
+                    "leftMiddleFingerTip",
+                    "leftRingFingerTip",
+                    "leftLittleFingerTip",
+                ],
+                "right": [
+                    "rightThumbTip",
+                    "rightIndexFingerTip",
+                    "rightMiddleFingerTip",
+                    "rightRingFingerTip",
+                    "rightLittleFingerTip",
+                ],
+            }
+
+            for hand_side in ["left", "right"]:
+                hand_key = f"{hand_side}Hand"
+                hand_transform = transforms_group[hand_key][frame_idx]
+
+                # 1. Wrist 3D position
+                hand_position = hand_transform[:3, 3]
+                state_vector.extend(hand_position)
+
+                # 2. Wrist orientation in compact 6D representation
+                rotation_matrix = hand_transform[:3, :3]
+                rotation_6d = np.concatenate([rotation_matrix[:, 0], rotation_matrix[:, 1]])
+                state_vector.extend(rotation_6d)
+
+                # 3. 3D positions of 5 fingertips
+                for fingertip in fingertip_joints[hand_side]:
+                    fingertip_transform = transforms_group[fingertip][frame_idx]
+                    fingertip_pos = fingertip_transform[:3, 3]
+                    state_vector.extend(fingertip_pos)
+
+            # Also return camera extrinsics for optional coordinate frame transformations
+            return np.array(state_vector, dtype=np.float32), transforms_group["camera"][frame_idx]
+
+        def get_state_and_action_from_egodex_annotations(demo):
+            """
+            Convert EgoDex demo annotations into states and actions.
+            
+            The "action" is the state at time t+1 (next-pose prediction).
+            """
+            transforms_group = demo["transforms"]
+            total_frames = list(transforms_group.values())[0].shape[0]
+
+            states_list, extrinsics_list = [], []
+            for frame_idx in range(total_frames):
+                state_vector, extrinsics = _get_state_for_single_frame(transforms_group, frame_idx)
+                states_list.append(state_vector)
+                extrinsics_list.append(extrinsics.flatten())  # Flatten 4x4 to 16D
+
+            state = np.array(states_list, dtype=np.float32)
+            extrinsics = np.array(extrinsics_list, dtype=np.float32)
+
+            # Shift by 1 timestep to convert state to action
+            action = np.roll(state, -1, axis=0)
+
+            return state, action, extrinsics
+
+        def process_demo(hdf5_file_path, video_path):
+            """Process a single EgoDex demo and return frames for LeRobot."""
+            video = mpy.read_video(str(video_path))
+            video = np.asarray(video)
+            num_frames = video.shape[0]
+            frames = []
+
+            with h5py.File(hdf5_file_path, "r") as demo:
+                state, action, extrinsics = get_state_and_action_from_egodex_annotations(demo)
+
+                # Get natural language task description
+                if demo.attrs.get("llm_type") == "reversible":
+                    direction = demo.attrs.get("which_llm_description", "1")
+                    lang_instruction = demo.attrs.get(
+                        "llm_description" if direction == "1" else "llm_description2",
+                        "manipulation task",
+                    )
+                else:
+                    lang_instruction = demo.attrs.get("llm_description", "manipulation task")
+
+                for step_idx in range(num_frames):
+                    # Resize image to default dimensions
+                    image_resized = cv2.resize(
+                        video[step_idx],
+                        (DEFAULT_IMAGE_WIDTH, DEFAULT_IMAGE_HEIGHT),
+                        interpolation=cv2.INTER_AREA,
+                    )
+                    frame = {
+                        "task": lang_instruction,
+                        "observation.image": image_resized,
+                        "observation.state": state[step_idx],
+                        "observation.extrinsics": extrinsics[step_idx],
+                        "action": action[step_idx],
+                    }
+                    frames.append(frame)
+
+            return frames
+
+        init_logging()
+
+        # Define EgoDex features
+        EGODEX_FEATURES = {
+            "observation.image": {
+                "dtype": "video",
+                "shape": (DEFAULT_IMAGE_HEIGHT, DEFAULT_IMAGE_WIDTH, 3),
+                "names": ["height", "width", "rgb"],
+            },
+            "observation.state": {
+                "dtype": "float32",
+                "shape": (48,),
+                "names": [
+                    # Left hand wrist position (3)
+                    "left_wrist_x",
+                    "left_wrist_y",
+                    "left_wrist_z",
+                    # Left hand wrist rotation 6D (6)
+                    "left_rot_0",
+                    "left_rot_1",
+                    "left_rot_2",
+                    "left_rot_3",
+                    "left_rot_4",
+                    "left_rot_5",
+                    # Left fingertips (15)
+                    "left_thumb_x",
+                    "left_thumb_y",
+                    "left_thumb_z",
+                    "left_index_x",
+                    "left_index_y",
+                    "left_index_z",
+                    "left_middle_x",
+                    "left_middle_y",
+                    "left_middle_z",
+                    "left_ring_x",
+                    "left_ring_y",
+                    "left_ring_z",
+                    "left_little_x",
+                    "left_little_y",
+                    "left_little_z",
+                    # Right hand wrist position (3)
+                    "right_wrist_x",
+                    "right_wrist_y",
+                    "right_wrist_z",
+                    # Right hand wrist rotation 6D (6)
+                    "right_rot_0",
+                    "right_rot_1",
+                    "right_rot_2",
+                    "right_rot_3",
+                    "right_rot_4",
+                    "right_rot_5",
+                    # Right fingertips (15)
+                    "right_thumb_x",
+                    "right_thumb_y",
+                    "right_thumb_z",
+                    "right_index_x",
+                    "right_index_y",
+                    "right_index_z",
+                    "right_middle_x",
+                    "right_middle_y",
+                    "right_middle_z",
+                    "right_ring_x",
+                    "right_ring_y",
+                    "right_ring_z",
+                    "right_little_x",
+                    "right_little_y",
+                    "right_little_z",
+                ],
+            },
+            "observation.extrinsics": {
+                "dtype": "float32",
+                "shape": (16,),
+                "names": [f"extrinsic_{i}" for i in range(16)],
+            },
+            "action": {
+                "dtype": "float32",
+                "shape": (48,),
+                "names": [f"action_{i}" for i in range(48)],
+            },
+        }
+
+        # 1. Discover all HDF5 files
+        files = sorted(list(self.raw_dir.rglob("*.hdf5")))
+        if not files:
+            print(f"No HDF5 files found in {self.raw_dir}")
+            return
+
+        # 2. Apply percentage filter
+        if self.percentage < 100:
+            num_files = max(1, int(len(files) * self.percentage / 100))
+            files = files[:num_files]
+            print(f"Processing {self.percentage}% of dataset: {num_files} files")
+
+        # 3. Assign files to this worker
+        my_files = files[rank::world_size]
+        if not my_files:
+            print(f"Rank {rank} has no files to process.")
+            return
+
+        print(f"Rank {rank} processing {len(my_files)} files out of {len(files)} total.")
+
+        # 4. Create a LeRobot dataset for this shard
+        shard_repo_id = f"{self.repo_id}_world_{world_size}_rank_{rank}"
+        shard_root = self.local_dir / shard_repo_id if self.local_dir else None
+
+        dataset = LeRobotDataset.create(
+            repo_id=shard_repo_id,
+            fps=30,
+            robot_type="hand",
+            features=EGODEX_FEATURES,
+            root=shard_root,
+        )
+
+        # 5. Process each file
+        for input_h5 in my_files:
+            try:
+                # Derive corresponding video path
+                video_path = input_h5.with_suffix(".mp4")
+                if not video_path.exists():
+                    print(f"Warning: Video file not found for {input_h5}, skipping.")
+                    continue
+
+                # Process demo and add frames
+                frames = process_demo(input_h5, video_path)
+                for frame in frames:
+                    dataset.add_frame(frame)
+                dataset.save_episode()
+
+                # Clean up to avoid OOM
+                del frames
+
+            except Exception as e:
+                print(f"Error processing {input_h5}: {e}")
+                continue
+
+        # 6. Finalize the dataset
+        dataset.finalize()
+
+
+def make_port_executor(
+    raw_dir,
+    repo_id,
+    job_name,
+    logs_dir,
+    workers,
+    partition,
+    cpus_per_task,
+    mem_per_cpu,
+    local_dir,
+    percentage,
+    slurm=True,
+):
+    kwargs = {
+        "pipeline": [
+            PortEgoDexShards(raw_dir, repo_id, local_dir, percentage),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": workers,
+                "workers": workers,
+                "time": "10:00:00",  # EgoDex is large, allow more time
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        executor = SlurmPipelineExecutor(**kwargs)
+    else:
+        kwargs.update(
+            {
+                "tasks": workers,
+                "workers": 1,  # Run locally sequentially for debugging
+            }
+        )
+        executor = LocalPipelineExecutor(**kwargs)
+
+    return executor
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Convert EgoDex dataset to LeRobot format using SLURM."
+    )
+
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        required=True,
+        help="Directory containing input EgoDex data (HDF5 + MP4 files).",
+    )
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Repository identifier (e.g., user/egodex-lerobot).",
+    )
+    parser.add_argument(
+        "--logs-dir",
+        type=Path,
+        default=Path("logs"),
+        help="Path to logs directory.",
+    )
+    parser.add_argument(
+        "--job-name",
+        type=str,
+        default="port_egodex",
+        help="Job name used in SLURM.",
+    )
+    parser.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="Launch over SLURM. Use --slurm 0 to launch sequentially (useful for debugging).",
+    )
+    parser.add_argument(
+        "--workers",
+        type=int,
+        default=50,
+        help="Number of SLURM workers.",
+    )
+    parser.add_argument(
+        "--partition",
+        type=str,
+        help="SLURM partition.",
+    )
+    parser.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=4,
+        help="Number of CPUs per worker.",
+    )
+    parser.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="4G",
+        help="Memory per CPU.",
+    )
+    parser.add_argument(
+        "--percentage",
+        type=float,
+        default=100.0,
+        help="Percentage of dataset to process (e.g., 1.0 for 1%%). Useful for testing.",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="Local directory to save the LeRobot dataset. Defaults to data/local_datasets.",
+    )
+
+    args = parser.parse_args()
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+
+    port_executor = make_port_executor(**kwargs)
+    port_executor.run()
+
+
+if __name__ == "__main__":
+    main()
+

examples/lekiwi/evaluate.py

@@ -33,83 +33,68 @@ TASK_DESCRIPTION = "My task description"
 HF_MODEL_ID = "<hf_username>/<model_repo_id>"
 HF_DATASET_ID = "<hf_username>/<eval_dataset_repo_id>"
 
-# Create the robot configuration & robot
-robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 
-robot = LeKiwiClient(robot_config)
+def main():
+    # Create the robot configuration & robot
+    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
 
-# Create policy
-policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+    robot = LeKiwiClient(robot_config)
 
-# Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, ACTION)
-obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
-dataset_features = {**action_features, **obs_features}
+    # Create policy
+    policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_DATASET_ID,
-    fps=FPS,
-    features=dataset_features,
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
+    # Configure the dataset features
+    action_features = hw_to_dataset_features(robot.action_features, ACTION)
+    obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
+    dataset_features = {**action_features, **obs_features}
 
-# Build Policy Processors
-preprocessor, postprocessor = make_pre_post_processors(
-    policy_cfg=policy,
-    pretrained_path=HF_MODEL_ID,
-    dataset_stats=dataset.meta.stats,
-    # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
-    preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
-)
-
-# Connect the robot
-# To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
-robot.connect()
-
-# TODO(Steven): Update this example to use pipelines
-teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
-
-# Initialize the keyboard listener and rerun visualization
-listener, events = init_keyboard_listener()
-init_rerun(session_name="lekiwi_evaluate")
-
-if not robot.is_connected:
-    raise ValueError("Robot is not connected!")
-
-print("Starting evaluate loop...")
-recorded_episodes = 0
-while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-    log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
-
-    # Main record loop
-    record_loop(
-        robot=robot,
-        events=events,
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_DATASET_ID,
         fps=FPS,
-        policy=policy,
-        preprocessor=preprocessor,  # Pass the pre and post policy processors
-        postprocessor=postprocessor,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
-        teleop_action_processor=teleop_action_processor,
-        robot_action_processor=robot_action_processor,
-        robot_observation_processor=robot_observation_processor,
+        features=dataset_features,
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and (
-        (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-    ):
-        log_say("Reset the environment")
+    # Build Policy Processors
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=policy,
+        pretrained_path=HF_MODEL_ID,
+        dataset_stats=dataset.meta.stats,
+        # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
+        preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
+    )
+
+    # Connect the robot
+    # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
+    robot.connect()
+
+    # TODO(Steven): Update this example to use pipelines
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+
+    # Initialize the keyboard listener and rerun visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="lekiwi_evaluate")
+
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
+
+    print("Starting evaluate loop...")
+    recorded_episodes = 0
+    while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
+        log_say(f"Running inference, recording eval episode {recorded_episodes} of {NUM_EPISODES}")
+
+        # Main record loop
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
+            policy=policy,
+            preprocessor=preprocessor,  # Pass the pre and post policy processors
+            postprocessor=postprocessor,
+            dataset=dataset,
             control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
@@ -118,21 +103,42 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             robot_observation_processor=robot_observation_processor,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-record episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (
+            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+        ):
+            log_say("Reset the environment")
+            record_loop(
+                robot=robot,
+                events=events,
+                fps=FPS,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    recorded_episodes += 1
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-listener.stop()
+        # Save episode
+        dataset.save_episode()
+        recorded_episodes += 1
 
-dataset.finalize()
-dataset.push_to_hub()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/record.py

@@ -34,78 +34,62 @@ RESET_TIME_SEC = 10
 TASK_DESCRIPTION = "My task description"
 HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
 
-# Create the robot and teleoperator configurations
-robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
-leader_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
-keyboard_config = KeyboardTeleopConfig()
 
-# Initialize the robot and teleoperator
-robot = LeKiwiClient(robot_config)
-leader_arm = SO100Leader(leader_arm_config)
-keyboard = KeyboardTeleop(keyboard_config)
+def main():
+    # Create the robot and teleoperator configurations
+    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+    leader_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
+    keyboard_config = KeyboardTeleopConfig()
 
-# TODO(Steven): Update this example to use pipelines
-teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+    # Initialize the robot and teleoperator
+    robot = LeKiwiClient(robot_config)
+    leader_arm = SO100Leader(leader_arm_config)
+    keyboard = KeyboardTeleop(keyboard_config)
 
-# Configure the dataset features
-action_features = hw_to_dataset_features(robot.action_features, ACTION)
-obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
-dataset_features = {**action_features, **obs_features}
+    # TODO(Steven): Update this example to use pipelines
+    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
 
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_REPO_ID,
-    fps=FPS,
-    features=dataset_features,
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
+    # Configure the dataset features
+    action_features = hw_to_dataset_features(robot.action_features, ACTION)
+    obs_features = hw_to_dataset_features(robot.observation_features, OBS_STR)
+    dataset_features = {**action_features, **obs_features}
 
-# Connect the robot and teleoperator
-# To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
-robot.connect()
-leader_arm.connect()
-keyboard.connect()
-
-# Initialize the keyboard listener and rerun visualization
-listener, events = init_keyboard_listener()
-init_rerun(session_name="lekiwi_record")
-
-if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-    raise ValueError("Robot or teleop is not connected!")
-
-print("Starting record loop...")
-recorded_episodes = 0
-while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-    log_say(f"Recording episode {recorded_episodes}")
-
-    # Main record loop
-    record_loop(
-        robot=robot,
-        events=events,
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_REPO_ID,
         fps=FPS,
-        dataset=dataset,
-        teleop=[leader_arm, keyboard],
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
-        teleop_action_processor=teleop_action_processor,
-        robot_action_processor=robot_action_processor,
-        robot_observation_processor=robot_observation_processor,
+        features=dataset_features,
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and (
-        (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-    ):
-        log_say("Reset the environment")
+    # Connect the robot and teleoperator
+    # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
+    robot.connect()
+    leader_arm.connect()
+    keyboard.connect()
+
+    # Initialize the keyboard listener and rerun visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="lekiwi_record")
+
+    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
+        raise ValueError("Robot or teleop is not connected!")
+
+    print("Starting record loop...")
+    recorded_episodes = 0
+    while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
+        log_say(f"Recording episode {recorded_episodes}")
+
+        # Main record loop
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
+            dataset=dataset,
             teleop=[leader_arm, keyboard],
-            control_time_s=RESET_TIME_SEC,
+            control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=teleop_action_processor,
@@ -113,23 +97,45 @@ while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
             robot_observation_processor=robot_observation_processor,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-record episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (
+            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+        ):
+            log_say("Reset the environment")
+            record_loop(
+                robot=robot,
+                events=events,
+                fps=FPS,
+                teleop=[leader_arm, keyboard],
+                control_time_s=RESET_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+                teleop_action_processor=teleop_action_processor,
+                robot_action_processor=robot_action_processor,
+                robot_observation_processor=robot_observation_processor,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    recorded_episodes += 1
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-leader_arm.disconnect()
-keyboard.disconnect()
-listener.stop()
+        # Save episode
+        dataset.save_episode()
+        recorded_episodes += 1
 
-dataset.finalize()
-dataset.push_to_hub()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    leader_arm.disconnect()
+    keyboard.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/lekiwi/replay.py

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

examples/lekiwi/teleoperate.py

@@ -19,54 +19,60 @@ import time
 from lerobot.robots.lekiwi import LeKiwiClient, LeKiwiClientConfig
 from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop, KeyboardTeleopConfig
 from lerobot.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 FPS = 30
 
-# Create the robot and teleoperator configurations
-robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="my_lekiwi")
-teleop_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
-keyboard_config = KeyboardTeleopConfig(id="my_laptop_keyboard")
 
-# Initialize the robot and teleoperator
-robot = LeKiwiClient(robot_config)
-leader_arm = SO100Leader(teleop_arm_config)
-keyboard = KeyboardTeleop(keyboard_config)
+def main():
+    # Create the robot and teleoperator configurations
+    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="my_lekiwi")
+    teleop_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem585A0077581", id="my_awesome_leader_arm")
+    keyboard_config = KeyboardTeleopConfig(id="my_laptop_keyboard")
 
-# Connect to the robot and teleoperator
-# To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
-robot.connect()
-leader_arm.connect()
-keyboard.connect()
+    # Initialize the robot and teleoperator
+    robot = LeKiwiClient(robot_config)
+    leader_arm = SO100Leader(teleop_arm_config)
+    keyboard = KeyboardTeleop(keyboard_config)
 
-# Init rerun viewer
-init_rerun(session_name="lekiwi_teleop")
+    # Connect to the robot and teleoperator
+    # To connect you already should have this script running on LeKiwi: `python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi`
+    robot.connect()
+    leader_arm.connect()
+    keyboard.connect()
 
-if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-    raise ValueError("Robot or teleop is not connected!")
+    # Init rerun viewer
+    init_rerun(session_name="lekiwi_teleop")
 
-print("Starting teleop loop...")
-while True:
-    t0 = time.perf_counter()
+    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
+        raise ValueError("Robot or teleop is not connected!")
 
-    # Get robot observation
-    observation = robot.get_observation()
+    print("Starting teleop loop...")
+    while True:
+        t0 = time.perf_counter()
 
-    # Get teleop action
-    # Arm
-    arm_action = leader_arm.get_action()
-    arm_action = {f"arm_{k}": v for k, v in arm_action.items()}
-    # Keyboard
-    keyboard_keys = keyboard.get_action()
-    base_action = robot._from_keyboard_to_base_action(keyboard_keys)
+        # Get robot observation
+        observation = robot.get_observation()
 
-    action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
+        # Get teleop action
+        # Arm
+        arm_action = leader_arm.get_action()
+        arm_action = {f"arm_{k}": v for k, v in arm_action.items()}
+        # Keyboard
+        keyboard_keys = keyboard.get_action()
+        base_action = robot._from_keyboard_to_base_action(keyboard_keys)
 
-    # Send action to robot
-    _ = robot.send_action(action)
+        action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
 
-    # Visualize
-    log_rerun_data(observation=observation, action=action)
+        # Send action to robot
+        _ = robot.send_action(action)
 
-    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        # Visualize
+        log_rerun_data(observation=observation, action=action)
+
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/evaluate.py

@@ -52,125 +52,114 @@ TASK_DESCRIPTION = "My task description"
 HF_MODEL_ID = "<hf_username>/<model_repo_id>"
 HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
 
-# Create the robot configuration & robot
-camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
-robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem58760434471",
-    id="my_awesome_follower_arm",
-    cameras=camera_config,
-    use_degrees=True,
-)
 
-robot = SO100Follower(robot_config)
-
-# Create policy
-policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
-
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(robot.bus.motors.keys()),
-)
-
-# Build pipeline to convert EE action to joints action
-robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    steps=[
-        InverseKinematicsEEToJoints(
-            kinematics=kinematics_solver,
-            motor_names=list(robot.bus.motors.keys()),
-            initial_guess_current_joints=True,
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
-
-# Build pipeline to convert joints observation to EE observation
-robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
-    steps=[
-        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
-    ],
-    to_transition=observation_to_transition,
-    to_output=transition_to_observation,
-)
-
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_DATASET_ID,
-    fps=FPS,
-    features=combine_feature_dicts(
-        aggregate_pipeline_dataset_features(
-            pipeline=robot_joints_to_ee_pose_processor,
-            initial_features=create_initial_features(observation=robot.observation_features),
-            use_videos=True,
-        ),
-        # User for now should be explicit on the feature keys that were used for record
-        # Alternatively, the user can pass the processor step that has the right features
-        aggregate_pipeline_dataset_features(
-            pipeline=make_default_teleop_action_processor(),
-            initial_features=create_initial_features(
-                action={
-                    f"ee.{k}": PolicyFeature(type=FeatureType.ACTION, shape=(1,))
-                    for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]
-                }
-            ),
-            use_videos=True,
-        ),
-    ),
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
-
-# Build Policy Processors
-preprocessor, postprocessor = make_pre_post_processors(
-    policy_cfg=policy,
-    pretrained_path=HF_MODEL_ID,
-    dataset_stats=dataset.meta.stats,
-    # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
-    preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
-)
-
-# Connect the robot
-robot.connect()
-
-# Initialize the keyboard listener and rerun visualization
-listener, events = init_keyboard_listener()
-init_rerun(session_name="phone_so100_evaluate")
-
-if not robot.is_connected:
-    raise ValueError("Robot is not connected!")
-
-print("Starting evaluate loop...")
-episode_idx = 0
-for episode_idx in range(NUM_EPISODES):
-    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
-
-    # Main record loop
-    record_loop(
-        robot=robot,
-        events=events,
-        fps=FPS,
-        policy=policy,
-        preprocessor=preprocessor,  # Pass the pre and post policy processors
-        postprocessor=postprocessor,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
-        teleop_action_processor=make_default_teleop_action_processor(),
-        robot_action_processor=robot_ee_to_joints_processor,
-        robot_observation_processor=robot_joints_to_ee_pose_processor,
+def main():
+    # Create the robot configuration & robot
+    camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem58760434471",
+        id="my_awesome_follower_arm",
+        cameras=camera_config,
+        use_degrees=True,
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
-        log_say("Reset the environment")
+    robot = SO100Follower(robot_config)
+
+    # Create policy
+    policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(robot.bus.motors.keys()),
+    )
+
+    # Build pipeline to convert EE action to joints action
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=list(robot.bus.motors.keys()),
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # Build pipeline to convert joints observation to EE observation
+    robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[
+            ForwardKinematicsJointsToEE(
+                kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys())
+            )
+        ],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_DATASET_ID,
+        fps=FPS,
+        features=combine_feature_dicts(
+            aggregate_pipeline_dataset_features(
+                pipeline=robot_joints_to_ee_pose_processor,
+                initial_features=create_initial_features(observation=robot.observation_features),
+                use_videos=True,
+            ),
+            # User for now should be explicit on the feature keys that were used for record
+            # Alternatively, the user can pass the processor step that has the right features
+            aggregate_pipeline_dataset_features(
+                pipeline=make_default_teleop_action_processor(),
+                initial_features=create_initial_features(
+                    action={
+                        f"ee.{k}": PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+                        for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]
+                    }
+                ),
+                use_videos=True,
+            ),
+        ),
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
+    )
+
+    # Build Policy Processors
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=policy,
+        pretrained_path=HF_MODEL_ID,
+        dataset_stats=dataset.meta.stats,
+        # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
+        preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
+    )
+
+    # Connect the robot
+    robot.connect()
+
+    # Initialize the keyboard listener and rerun visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="phone_so100_evaluate")
+
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
+
+    print("Starting evaluate loop...")
+    episode_idx = 0
+    for episode_idx in range(NUM_EPISODES):
+        log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+
+        # Main record loop
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
+            policy=policy,
+            preprocessor=preprocessor,  # Pass the pre and post policy processors
+            postprocessor=postprocessor,
+            dataset=dataset,
             control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
@@ -179,21 +168,40 @@ for episode_idx in range(NUM_EPISODES):
             robot_observation_processor=robot_joints_to_ee_pose_processor,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-record episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+            log_say("Reset the environment")
+            record_loop(
+                robot=robot,
+                events=events,
+                fps=FPS,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+                teleop_action_processor=make_default_teleop_action_processor(),
+                robot_action_processor=robot_ee_to_joints_processor,
+                robot_observation_processor=robot_joints_to_ee_pose_processor,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-listener.stop()
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
 
-dataset.finalize()
-dataset.push_to_hub()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/record.py

@@ -50,133 +50,122 @@ RESET_TIME_SEC = 30
 TASK_DESCRIPTION = "My task description"
 HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
 
-# Create the robot and teleoperator configurations
-camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
-robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem5A460814411",
-    id="my_awesome_follower_arm",
-    cameras=camera_config,
-    use_degrees=True,
-)
-teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
 
-# Initialize the robot and teleoperator
-robot = SO100Follower(robot_config)
-phone = Phone(teleop_config)
+def main():
+    # Create the robot and teleoperator configurations
+    camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411",
+        id="my_awesome_follower_arm",
+        cameras=camera_config,
+        use_degrees=True,
+    )
+    teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
 
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(robot.bus.motors.keys()),
-)
+    # Initialize the robot and teleoperator
+    robot = SO100Follower(robot_config)
+    phone = Phone(teleop_config)
 
-# Build pipeline to convert phone action to EE action
-phone_to_robot_ee_pose_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    steps=[
-        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
-        EEReferenceAndDelta(
-            kinematics=kinematics_solver,
-            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
-            motor_names=list(robot.bus.motors.keys()),
-            use_latched_reference=True,
-        ),
-        EEBoundsAndSafety(
-            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
-            max_ee_step_m=0.20,
-        ),
-        GripperVelocityToJoint(speed_factor=20.0),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
-
-# Build pipeline to convert EE action to joints action
-robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    steps=[
-        InverseKinematicsEEToJoints(
-            kinematics=kinematics_solver,
-            motor_names=list(robot.bus.motors.keys()),
-            initial_guess_current_joints=True,
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
-
-# Build pipeline to convert joint observation to EE observation
-robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation](
-    steps=[
-        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
-    ],
-    to_transition=observation_to_transition,
-    to_output=transition_to_observation,
-)
-
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_REPO_ID,
-    fps=FPS,
-    features=combine_feature_dicts(
-        # Run the feature contract of the pipelines
-        # This tells you how the features would look like after the pipeline steps
-        aggregate_pipeline_dataset_features(
-            pipeline=phone_to_robot_ee_pose_processor,
-            initial_features=create_initial_features(action=phone.action_features),
-            use_videos=True,
-        ),
-        aggregate_pipeline_dataset_features(
-            pipeline=robot_joints_to_ee_pose,
-            initial_features=create_initial_features(observation=robot.observation_features),
-            use_videos=True,
-        ),
-    ),
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
-
-# Connect the robot and teleoperator
-robot.connect()
-phone.connect()
-
-# Initialize the keyboard listener and rerun visualization
-listener, events = init_keyboard_listener()
-init_rerun(session_name="phone_so100_record")
-
-if not robot.is_connected or not phone.is_connected:
-    raise ValueError("Robot or teleop is not connected!")
-
-
-print("Starting record loop. Move your phone to teleoperate the robot...")
-episode_idx = 0
-while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
-
-    # Main record loop
-    record_loop(
-        robot=robot,
-        events=events,
-        fps=FPS,
-        teleop=phone,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
-        teleop_action_processor=phone_to_robot_ee_pose_processor,
-        robot_action_processor=robot_ee_to_joints_processor,
-        robot_observation_processor=robot_joints_to_ee_pose,
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(robot.bus.motors.keys()),
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-        log_say("Reset the environment")
+    # Build pipeline to convert phone action to EE action
+    phone_to_robot_ee_pose_processor = RobotProcessorPipeline[
+        tuple[RobotAction, RobotObservation], RobotAction
+    ](
+        steps=[
+            MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+            EEReferenceAndDelta(
+                kinematics=kinematics_solver,
+                end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+                motor_names=list(robot.bus.motors.keys()),
+                use_latched_reference=True,
+            ),
+            EEBoundsAndSafety(
+                end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+                max_ee_step_m=0.20,
+            ),
+            GripperVelocityToJoint(speed_factor=20.0),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # Build pipeline to convert EE action to joints action
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=list(robot.bus.motors.keys()),
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # Build pipeline to convert joint observation to EE observation
+    robot_joints_to_ee_pose = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[
+            ForwardKinematicsJointsToEE(
+                kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys())
+            )
+        ],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_REPO_ID,
+        fps=FPS,
+        features=combine_feature_dicts(
+            # Run the feature contract of the pipelines
+            # This tells you how the features would look like after the pipeline steps
+            aggregate_pipeline_dataset_features(
+                pipeline=phone_to_robot_ee_pose_processor,
+                initial_features=create_initial_features(action=phone.action_features),
+                use_videos=True,
+            ),
+            aggregate_pipeline_dataset_features(
+                pipeline=robot_joints_to_ee_pose,
+                initial_features=create_initial_features(observation=robot.observation_features),
+                use_videos=True,
+            ),
+        ),
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
+    )
+
+    # Connect the robot and teleoperator
+    robot.connect()
+    phone.connect()
+
+    # Initialize the keyboard listener and rerun visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="phone_so100_record")
+
+    if not robot.is_connected or not phone.is_connected:
+        raise ValueError("Robot or teleop is not connected!")
+
+    print("Starting record loop. Move your phone to teleoperate the robot...")
+    episode_idx = 0
+    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+        # Main record loop
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
             teleop=phone,
-            control_time_s=RESET_TIME_SEC,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=phone_to_robot_ee_pose_processor,
@@ -184,22 +173,42 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
             robot_observation_processor=robot_joints_to_ee_pose,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-recording episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+            log_say("Reset the environment")
+            record_loop(
+                robot=robot,
+                events=events,
+                fps=FPS,
+                teleop=phone,
+                control_time_s=RESET_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+                teleop_action_processor=phone_to_robot_ee_pose_processor,
+                robot_action_processor=robot_ee_to_joints_processor,
+                robot_observation_processor=robot_joints_to_ee_pose,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-recording episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-phone.disconnect()
-listener.stop()
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
 
-dataset.finalize()
-dataset.push_to_hub()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    phone.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/replay.py

@@ -29,72 +29,78 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.utils.constants import ACTION
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 EPISODE_IDX = 0
 HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
 
-# Initialize the robot config
-robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
-)
 
-# Initialize the robot
-robot = SO100Follower(robot_config)
+def main():
+    # Initialize the robot config
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+    )
 
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(robot.bus.motors.keys()),
-)
+    # Initialize the robot
+    robot = SO100Follower(robot_config)
 
-# Build pipeline to convert EE action to joints action
-robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    steps=[
-        InverseKinematicsEEToJoints(
-            kinematics=kinematics_solver,
-            motor_names=list(robot.bus.motors.keys()),
-            initial_guess_current_joints=False,  # Because replay is open loop
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(robot.bus.motors.keys()),
+    )
 
-# Fetch the dataset to replay
-dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
-# Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
-episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
-actions = episode_frames.select_columns(ACTION)
+    # Build pipeline to convert EE action to joints action
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=list(robot.bus.motors.keys()),
+                initial_guess_current_joints=False,  # Because replay is open loop
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
 
-# Connect to the robot
-robot.connect()
+    # Fetch the dataset to replay
+    dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
+    # Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
+    episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
+    actions = episode_frames.select_columns(ACTION)
 
-if not robot.is_connected:
-    raise ValueError("Robot is not connected!")
+    # Connect to the robot
+    robot.connect()
 
-print("Starting replay loop...")
-log_say(f"Replaying episode {EPISODE_IDX}")
-for idx in range(len(episode_frames)):
-    t0 = time.perf_counter()
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
 
-    # Get recorded action from dataset
-    ee_action = {
-        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-    }
+    print("Starting replay loop...")
+    log_say(f"Replaying episode {EPISODE_IDX}")
+    for idx in range(len(episode_frames)):
+        t0 = time.perf_counter()
 
-    # Get robot observation
-    robot_obs = robot.get_observation()
+        # Get recorded action from dataset
+        ee_action = {
+            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        }
 
-    # Dataset EE -> robot joints
-    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+        # Get robot observation
+        robot_obs = robot.get_observation()
 
-    # Send action to robot
-    _ = robot.send_action(joint_action)
+        # Dataset EE -> robot joints
+        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+        # Send action to robot
+        _ = robot.send_action(joint_action)
 
-# Clean up
-robot.disconnect()
+        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+    # Clean up
+    robot.disconnect()
+
+
+if __name__ == "__main__":
+    main()

examples/phone_to_so100/teleoperate.py

@@ -32,82 +32,90 @@ from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.phone.config_phone import PhoneConfig, PhoneOS
 from lerobot.teleoperators.phone.phone_processor import MapPhoneActionToRobotAction
 from lerobot.teleoperators.phone.teleop_phone import Phone
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 FPS = 30
 
-# Initialize the robot and teleoperator
-robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
-)
-teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
 
-# Initialize the robot and teleoperator
-robot = SO100Follower(robot_config)
-teleop_device = Phone(teleop_config)
+def main():
+    # Initialize the robot and teleoperator
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+    )
+    teleop_config = PhoneConfig(phone_os=PhoneOS.IOS)  # or PhoneOS.ANDROID
 
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(robot.bus.motors.keys()),
-)
+    # Initialize the robot and teleoperator
+    robot = SO100Follower(robot_config)
+    teleop_device = Phone(teleop_config)
 
-# Build pipeline to convert phone action to ee pose action to joint action
-phone_to_robot_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    steps=[
-        MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
-        EEReferenceAndDelta(
-            kinematics=kinematics_solver,
-            end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
-            motor_names=list(robot.bus.motors.keys()),
-            use_latched_reference=True,
-        ),
-        EEBoundsAndSafety(
-            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
-            max_ee_step_m=0.10,
-        ),
-        GripperVelocityToJoint(
-            speed_factor=20.0,
-        ),
-        InverseKinematicsEEToJoints(
-            kinematics=kinematics_solver,
-            motor_names=list(robot.bus.motors.keys()),
-            initial_guess_current_joints=True,
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(robot.bus.motors.keys()),
+    )
 
-# Connect to the robot and teleoperator
-robot.connect()
-teleop_device.connect()
+    # Build pipeline to convert phone action to ee pose action to joint action
+    phone_to_robot_joints_processor = RobotProcessorPipeline[
+        tuple[RobotAction, RobotObservation], RobotAction
+    ](
+        steps=[
+            MapPhoneActionToRobotAction(platform=teleop_config.phone_os),
+            EEReferenceAndDelta(
+                kinematics=kinematics_solver,
+                end_effector_step_sizes={"x": 0.5, "y": 0.5, "z": 0.5},
+                motor_names=list(robot.bus.motors.keys()),
+                use_latched_reference=True,
+            ),
+            EEBoundsAndSafety(
+                end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+                max_ee_step_m=0.10,
+            ),
+            GripperVelocityToJoint(
+                speed_factor=20.0,
+            ),
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=list(robot.bus.motors.keys()),
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
 
-# Init rerun viewer
-init_rerun(session_name="phone_so100_teleop")
+    # Connect to the robot and teleoperator
+    robot.connect()
+    teleop_device.connect()
 
-if not robot.is_connected or not teleop_device.is_connected:
-    raise ValueError("Robot or teleop is not connected!")
+    # Init rerun viewer
+    init_rerun(session_name="phone_so100_teleop")
 
-print("Starting teleop loop. Move your phone to teleoperate the robot...")
-while True:
-    t0 = time.perf_counter()
+    if not robot.is_connected or not teleop_device.is_connected:
+        raise ValueError("Robot or teleop is not connected!")
 
-    # Get robot observation
-    robot_obs = robot.get_observation()
+    print("Starting teleop loop. Move your phone to teleoperate the robot...")
+    while True:
+        t0 = time.perf_counter()
 
-    # Get teleop action
-    phone_obs = teleop_device.get_action()
+        # Get robot observation
+        robot_obs = robot.get_observation()
 
-    # Phone -> EE pose -> Joints transition
-    joint_action = phone_to_robot_joints_processor((phone_obs, robot_obs))
+        # Get teleop action
+        phone_obs = teleop_device.get_action()
 
-    # Send action to robot
-    _ = robot.send_action(joint_action)
+        # Phone -> EE pose -> Joints transition
+        joint_action = phone_to_robot_joints_processor((phone_obs, robot_obs))
 
-    # Visualize
-    log_rerun_data(observation=phone_obs, action=joint_action)
+        # Send action to robot
+        _ = robot.send_action(joint_action)
 
-    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        # Visualize
+        log_rerun_data(observation=phone_obs, action=joint_action)
+
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+
+
+if __name__ == "__main__":
+    main()

examples/so100_to_so100_EE/evaluate.py

@@ -52,126 +52,114 @@ TASK_DESCRIPTION = "My task description"
 HF_MODEL_ID = "<hf_username>/<model_repo_id>"
 HF_DATASET_ID = "<hf_username>/<dataset_repo_id>"
 
-# Create the robot configuration & robot
-camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
-robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem5A460814411",
-    id="my_awesome_follower_arm",
-    cameras=camera_config,
-    use_degrees=True,
-)
 
-robot = SO100Follower(robot_config)
-
-# Create policy
-policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
-
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(robot.bus.motors.keys()),
-)
-
-# Build pipeline to convert EE action to joints action
-robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    steps=[
-        InverseKinematicsEEToJoints(
-            kinematics=kinematics_solver,
-            motor_names=list(robot.bus.motors.keys()),
-            initial_guess_current_joints=True,
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
-
-# Build pipeline to convert joints observation to EE observation
-robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
-    steps=[
-        ForwardKinematicsJointsToEE(kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys()))
-    ],
-    to_transition=observation_to_transition,
-    to_output=transition_to_observation,
-)
-
-
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_DATASET_ID,
-    fps=FPS,
-    features=combine_feature_dicts(
-        aggregate_pipeline_dataset_features(
-            pipeline=robot_joints_to_ee_pose_processor,
-            initial_features=create_initial_features(observation=robot.observation_features),
-            use_videos=True,
-        ),
-        # User for now should be explicit on the feature keys that were used for record
-        # Alternatively, the user can pass the processor step that has the right features
-        aggregate_pipeline_dataset_features(
-            pipeline=make_default_teleop_action_processor(),
-            initial_features=create_initial_features(
-                action={
-                    f"ee.{k}": PolicyFeature(type=FeatureType.ACTION, shape=(1,))
-                    for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]
-                }
-            ),
-            use_videos=True,
-        ),
-    ),
-    robot_type=robot.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
-
-# Build Policy Processors
-preprocessor, postprocessor = make_pre_post_processors(
-    policy_cfg=policy,
-    pretrained_path=HF_MODEL_ID,
-    dataset_stats=dataset.meta.stats,
-    # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
-    preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
-)
-
-# Connect the robot and teleoperator
-robot.connect()
-
-# Initialize the keyboard listener and rerun visualization
-listener, events = init_keyboard_listener()
-init_rerun(session_name="so100_so100_evaluate")
-
-if not robot.is_connected:
-    raise ValueError("Robot is not connected!")
-
-print("Starting evaluate loop...")
-episode_idx = 0
-for episode_idx in range(NUM_EPISODES):
-    log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
-
-    # Main record loop
-    record_loop(
-        robot=robot,
-        events=events,
-        fps=FPS,
-        policy=policy,
-        preprocessor=preprocessor,  # Pass the pre and post policy processors
-        postprocessor=postprocessor,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
-        teleop_action_processor=make_default_teleop_action_processor(),
-        robot_action_processor=robot_ee_to_joints_processor,
-        robot_observation_processor=robot_joints_to_ee_pose_processor,
+def main():
+    # Create the robot configuration & robot
+    camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411",
+        id="my_awesome_follower_arm",
+        cameras=camera_config,
+        use_degrees=True,
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
-        log_say("Reset the environment")
+    robot = SO100Follower(robot_config)
+
+    # Create policy
+    policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(robot.bus.motors.keys()),
+    )
+
+    # Build pipeline to convert EE action to joints action
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=list(robot.bus.motors.keys()),
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # Build pipeline to convert joints observation to EE observation
+    robot_joints_to_ee_pose_processor = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[
+            ForwardKinematicsJointsToEE(
+                kinematics=kinematics_solver, motor_names=list(robot.bus.motors.keys())
+            )
+        ],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_DATASET_ID,
+        fps=FPS,
+        features=combine_feature_dicts(
+            aggregate_pipeline_dataset_features(
+                pipeline=robot_joints_to_ee_pose_processor,
+                initial_features=create_initial_features(observation=robot.observation_features),
+                use_videos=True,
+            ),
+            # User for now should be explicit on the feature keys that were used for record
+            # Alternatively, the user can pass the processor step that has the right features
+            aggregate_pipeline_dataset_features(
+                pipeline=make_default_teleop_action_processor(),
+                initial_features=create_initial_features(
+                    action={
+                        f"ee.{k}": PolicyFeature(type=FeatureType.ACTION, shape=(1,))
+                        for k in ["x", "y", "z", "wx", "wy", "wz", "gripper_pos"]
+                    }
+                ),
+                use_videos=True,
+            ),
+        ),
+        robot_type=robot.name,
+        use_videos=True,
+        image_writer_threads=4,
+    )
+
+    # Build Policy Processors
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=policy,
+        pretrained_path=HF_MODEL_ID,
+        dataset_stats=dataset.meta.stats,
+        # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
+        preprocessor_overrides={"device_processor": {"device": str(policy.config.device)}},
+    )
+
+    # Connect the robot and teleoperator
+    robot.connect()
+
+    # Initialize the keyboard listener and rerun visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="so100_so100_evaluate")
+
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
+
+    print("Starting evaluate loop...")
+    episode_idx = 0
+    for episode_idx in range(NUM_EPISODES):
+        log_say(f"Running inference, recording eval episode {episode_idx + 1} of {NUM_EPISODES}")
+
+        # Main record loop
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
+            policy=policy,
+            preprocessor=preprocessor,  # Pass the pre and post policy processors
+            postprocessor=postprocessor,
+            dataset=dataset,
             control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
@@ -180,21 +168,40 @@ for episode_idx in range(NUM_EPISODES):
             robot_observation_processor=robot_joints_to_ee_pose_processor,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-record episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+            log_say("Reset the environment")
+            record_loop(
+                robot=robot,
+                events=events,
+                fps=FPS,
+                control_time_s=EPISODE_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+                teleop_action_processor=make_default_teleop_action_processor(),
+                robot_action_processor=robot_ee_to_joints_processor,
+                robot_observation_processor=robot_joints_to_ee_pose_processor,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-record episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-robot.disconnect()
-listener.stop()
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
 
-dataset.finalize()
-dataset.push_to_hub()
+    # Clean up
+    log_say("Stop recording")
+    robot.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/so100_to_so100_EE/record.py

@@ -48,134 +48,122 @@ RESET_TIME_SEC = 30
 TASK_DESCRIPTION = "My task description"
 HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
 
-# Create the robot and teleoperator configurations
-camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
-follower_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", cameras=camera_config, use_degrees=True
-)
-leader_config = SO100LeaderConfig(port="/dev/tty.usbmodem5A460819811", id="my_awesome_leader_arm")
 
-# Initialize the robot and teleoperator
-follower = SO100Follower(follower_config)
-leader = SO100Leader(leader_config)
+def main():
+    # Create the robot and teleoperator configurations
+    camera_config = {"front": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=FPS)}
+    follower_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411",
+        id="my_awesome_follower_arm",
+        cameras=camera_config,
+        use_degrees=True,
+    )
+    leader_config = SO100LeaderConfig(port="/dev/tty.usbmodem5A460819811", id="my_awesome_leader_arm")
 
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-follower_kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(follower.bus.motors.keys()),
-)
+    # Initialize the robot and teleoperator
+    follower = SO100Follower(follower_config)
+    leader = SO100Leader(leader_config)
 
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-leader_kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(leader.bus.motors.keys()),
-)
-
-# Build pipeline to convert follower joints to EE observation
-follower_joints_to_ee = RobotProcessorPipeline[RobotObservation, RobotObservation](
-    steps=[
-        ForwardKinematicsJointsToEE(
-            kinematics=follower_kinematics_solver, motor_names=list(follower.bus.motors.keys())
-        ),
-    ],
-    to_transition=observation_to_transition,
-    to_output=transition_to_observation,
-)
-
-# Build pipeline to convert leader joints to EE action
-leader_joints_to_ee = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    steps=[
-        ForwardKinematicsJointsToEE(
-            kinematics=leader_kinematics_solver, motor_names=list(leader.bus.motors.keys())
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
-
-# Build pipeline to convert EE action to follower joints
-ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    [
-        EEBoundsAndSafety(
-            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
-            max_ee_step_m=0.10,
-        ),
-        InverseKinematicsEEToJoints(
-            kinematics=follower_kinematics_solver,
-            motor_names=list(follower.bus.motors.keys()),
-            initial_guess_current_joints=True,
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
-
-# Create the dataset
-dataset = LeRobotDataset.create(
-    repo_id=HF_REPO_ID,
-    fps=FPS,
-    features=combine_feature_dicts(
-        # Run the feature contract of the pipelines
-        # This tells you how the features would look like after the pipeline steps
-        aggregate_pipeline_dataset_features(
-            pipeline=leader_joints_to_ee,
-            initial_features=create_initial_features(action=leader.action_features),
-            use_videos=True,
-        ),
-        aggregate_pipeline_dataset_features(
-            pipeline=follower_joints_to_ee,
-            initial_features=create_initial_features(observation=follower.observation_features),
-            use_videos=True,
-        ),
-    ),
-    robot_type=follower.name,
-    use_videos=True,
-    image_writer_threads=4,
-)
-
-
-# Connect the robot and teleoperator
-leader.connect()
-follower.connect()
-
-# Initialize the keyboard listener and rerun visualization
-listener, events = init_keyboard_listener()
-init_rerun(session_name="recording_phone")
-
-if not leader.is_connected or not follower.is_connected:
-    raise ValueError("Robot or teleop is not connected!")
-
-print("Starting record loop...")
-episode_idx = 0
-while episode_idx < NUM_EPISODES and not events["stop_recording"]:
-    log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
-
-    # Main record loop
-    record_loop(
-        robot=follower,
-        events=events,
-        fps=FPS,
-        teleop=leader,
-        dataset=dataset,
-        control_time_s=EPISODE_TIME_SEC,
-        single_task=TASK_DESCRIPTION,
-        display_data=True,
-        teleop_action_processor=leader_joints_to_ee,
-        robot_action_processor=ee_to_follower_joints,
-        robot_observation_processor=follower_joints_to_ee,
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    follower_kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(follower.bus.motors.keys()),
     )
 
-    # Reset the environment if not stopping or re-recording
-    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-        log_say("Reset the environment")
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    leader_kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(leader.bus.motors.keys()),
+    )
+
+    # Build pipeline to convert follower joints to EE observation
+    follower_joints_to_ee = RobotProcessorPipeline[RobotObservation, RobotObservation](
+        steps=[
+            ForwardKinematicsJointsToEE(
+                kinematics=follower_kinematics_solver, motor_names=list(follower.bus.motors.keys())
+            ),
+        ],
+        to_transition=observation_to_transition,
+        to_output=transition_to_observation,
+    )
+
+    # Build pipeline to convert leader joints to EE action
+    leader_joints_to_ee = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            ForwardKinematicsJointsToEE(
+                kinematics=leader_kinematics_solver, motor_names=list(leader.bus.motors.keys())
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # Build pipeline to convert EE action to follower joints
+    ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        [
+            EEBoundsAndSafety(
+                end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+                max_ee_step_m=0.10,
+            ),
+            InverseKinematicsEEToJoints(
+                kinematics=follower_kinematics_solver,
+                motor_names=list(follower.bus.motors.keys()),
+                initial_guess_current_joints=True,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
+
+    # Create the dataset
+    dataset = LeRobotDataset.create(
+        repo_id=HF_REPO_ID,
+        fps=FPS,
+        features=combine_feature_dicts(
+            # Run the feature contract of the pipelines
+            # This tells you how the features would look like after the pipeline steps
+            aggregate_pipeline_dataset_features(
+                pipeline=leader_joints_to_ee,
+                initial_features=create_initial_features(action=leader.action_features),
+                use_videos=True,
+            ),
+            aggregate_pipeline_dataset_features(
+                pipeline=follower_joints_to_ee,
+                initial_features=create_initial_features(observation=follower.observation_features),
+                use_videos=True,
+            ),
+        ),
+        robot_type=follower.name,
+        use_videos=True,
+        image_writer_threads=4,
+    )
+
+    # Connect the robot and teleoperator
+    leader.connect()
+    follower.connect()
+
+    # Initialize the keyboard listener and rerun visualization
+    listener, events = init_keyboard_listener()
+    init_rerun(session_name="recording_phone")
+
+    if not leader.is_connected or not follower.is_connected:
+        raise ValueError("Robot or teleop is not connected!")
+
+    print("Starting record loop...")
+    episode_idx = 0
+    while episode_idx < NUM_EPISODES and not events["stop_recording"]:
+        log_say(f"Recording episode {episode_idx + 1} of {NUM_EPISODES}")
+
+        # Main record loop
         record_loop(
             robot=follower,
             events=events,
             fps=FPS,
             teleop=leader,
-            control_time_s=RESET_TIME_SEC,
+            dataset=dataset,
+            control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=leader_joints_to_ee,
@@ -183,22 +171,42 @@ while episode_idx < NUM_EPISODES and not events["stop_recording"]:
             robot_observation_processor=follower_joints_to_ee,
         )
 
-    if events["rerecord_episode"]:
-        log_say("Re-recording episode")
-        events["rerecord_episode"] = False
-        events["exit_early"] = False
-        dataset.clear_episode_buffer()
-        continue
+        # Reset the environment if not stopping or re-recording
+        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+            log_say("Reset the environment")
+            record_loop(
+                robot=follower,
+                events=events,
+                fps=FPS,
+                teleop=leader,
+                control_time_s=RESET_TIME_SEC,
+                single_task=TASK_DESCRIPTION,
+                display_data=True,
+                teleop_action_processor=leader_joints_to_ee,
+                robot_action_processor=ee_to_follower_joints,
+                robot_observation_processor=follower_joints_to_ee,
+            )
 
-    # Save episode
-    dataset.save_episode()
-    episode_idx += 1
+        if events["rerecord_episode"]:
+            log_say("Re-recording episode")
+            events["rerecord_episode"] = False
+            events["exit_early"] = False
+            dataset.clear_episode_buffer()
+            continue
 
-# Clean up
-log_say("Stop recording")
-leader.disconnect()
-follower.disconnect()
-listener.stop()
+        # Save episode
+        dataset.save_episode()
+        episode_idx += 1
 
-dataset.finalize()
-dataset.push_to_hub()
+    # Clean up
+    log_say("Stop recording")
+    leader.disconnect()
+    follower.disconnect()
+    listener.stop()
+
+    dataset.finalize()
+    dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    main()

examples/so100_to_so100_EE/replay.py

@@ -30,72 +30,78 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 )
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.utils.constants import ACTION
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 EPISODE_IDX = 0
 HF_REPO_ID = "<hf_username>/<dataset_repo_id>"
 
-# Initialize the robot config
-robot_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
-)
 
-# Initialize the robot
-robot = SO100Follower(robot_config)
+def main():
+    # Initialize the robot config
+    robot_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+    )
 
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(robot.bus.motors.keys()),
-)
+    # Initialize the robot
+    robot = SO100Follower(robot_config)
 
-# Build pipeline to convert EE action to joints action
-robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    steps=[
-        InverseKinematicsEEToJoints(
-            kinematics=kinematics_solver,
-            motor_names=list(robot.bus.motors.keys()),
-            initial_guess_current_joints=False,  # Because replay is open loop
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(robot.bus.motors.keys()),
+    )
 
-# Fetch the dataset to replay
-dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
-# Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
-episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
-actions = episode_frames.select_columns(ACTION)
+    # Build pipeline to convert EE action to joints action
+    robot_ee_to_joints_processor = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        steps=[
+            InverseKinematicsEEToJoints(
+                kinematics=kinematics_solver,
+                motor_names=list(robot.bus.motors.keys()),
+                initial_guess_current_joints=False,  # Because replay is open loop
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
 
-# Connect to the robot
-robot.connect()
+    # Fetch the dataset to replay
+    dataset = LeRobotDataset(HF_REPO_ID, episodes=[EPISODE_IDX])
+    # Filter dataset to only include frames from the specified episode since episodes are chunked in dataset V3.0
+    episode_frames = dataset.hf_dataset.filter(lambda x: x["episode_index"] == EPISODE_IDX)
+    actions = episode_frames.select_columns(ACTION)
 
-if not robot.is_connected:
-    raise ValueError("Robot is not connected!")
+    # Connect to the robot
+    robot.connect()
 
-print("Starting replay loop...")
-log_say(f"Replaying episode {EPISODE_IDX}")
-for idx in range(len(episode_frames)):
-    t0 = time.perf_counter()
+    if not robot.is_connected:
+        raise ValueError("Robot is not connected!")
 
-    # Get recorded action from dataset
-    ee_action = {
-        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-    }
+    print("Starting replay loop...")
+    log_say(f"Replaying episode {EPISODE_IDX}")
+    for idx in range(len(episode_frames)):
+        t0 = time.perf_counter()
 
-    # Get robot observation
-    robot_obs = robot.get_observation()
+        # Get recorded action from dataset
+        ee_action = {
+            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+        }
 
-    # Dataset EE -> robot joints
-    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+        # Get robot observation
+        robot_obs = robot.get_observation()
 
-    # Send action to robot
-    _ = robot.send_action(joint_action)
+        # Dataset EE -> robot joints
+        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+        # Send action to robot
+        _ = robot.send_action(joint_action)
 
-# Clean up
-robot.disconnect()
+        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+
+    # Clean up
+    robot.disconnect()
+
+
+if __name__ == "__main__":
+    main()

examples/so100_to_so100_EE/teleoperate.py

@@ -32,90 +32,96 @@ from lerobot.robots.so100_follower.robot_kinematic_processor import (
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 from lerobot.teleoperators.so100_leader.config_so100_leader import SO100LeaderConfig
 from lerobot.teleoperators.so100_leader.so100_leader import SO100Leader
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
 FPS = 30
 
-# Initialize the robot and teleoperator config
-follower_config = SO100FollowerConfig(
-    port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
-)
-leader_config = SO100LeaderConfig(port="/dev/tty.usbmodem5A460819811", id="my_awesome_leader_arm")
 
-# Initialize the robot and teleoperator
-follower = SO100Follower(follower_config)
-leader = SO100Leader(leader_config)
+def main():
+    # Initialize the robot and teleoperator config
+    follower_config = SO100FollowerConfig(
+        port="/dev/tty.usbmodem5A460814411", id="my_awesome_follower_arm", use_degrees=True
+    )
+    leader_config = SO100LeaderConfig(port="/dev/tty.usbmodem5A460819811", id="my_awesome_leader_arm")
 
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-follower_kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(follower.bus.motors.keys()),
-)
+    # Initialize the robot and teleoperator
+    follower = SO100Follower(follower_config)
+    leader = SO100Leader(leader_config)
 
-# NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
-leader_kinematics_solver = RobotKinematics(
-    urdf_path="./SO101/so101_new_calib.urdf",
-    target_frame_name="gripper_frame_link",
-    joint_names=list(leader.bus.motors.keys()),
-)
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    follower_kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(follower.bus.motors.keys()),
+    )
 
-# Build pipeline to convert teleop joints to EE action
-leader_to_ee = RobotProcessorPipeline[RobotAction, RobotAction](
-    steps=[
-        ForwardKinematicsJointsToEE(
-            kinematics=leader_kinematics_solver, motor_names=list(leader.bus.motors.keys())
-        ),
-    ],
-    to_transition=robot_action_to_transition,
-    to_output=transition_to_robot_action,
-)
+    # NOTE: It is highly recommended to use the urdf in the SO-ARM100 repo: https://github.com/TheRobotStudio/SO-ARM100/blob/main/Simulation/SO101/so101_new_calib.urdf
+    leader_kinematics_solver = RobotKinematics(
+        urdf_path="./SO101/so101_new_calib.urdf",
+        target_frame_name="gripper_frame_link",
+        joint_names=list(leader.bus.motors.keys()),
+    )
 
-# build pipeline to convert EE action to robot joints
-ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
-    [
-        EEBoundsAndSafety(
-            end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
-            max_ee_step_m=0.10,
-        ),
-        InverseKinematicsEEToJoints(
-            kinematics=follower_kinematics_solver,
-            motor_names=list(follower.bus.motors.keys()),
-            initial_guess_current_joints=False,
-        ),
-    ],
-    to_transition=robot_action_observation_to_transition,
-    to_output=transition_to_robot_action,
-)
+    # Build pipeline to convert teleop joints to EE action
+    leader_to_ee = RobotProcessorPipeline[RobotAction, RobotAction](
+        steps=[
+            ForwardKinematicsJointsToEE(
+                kinematics=leader_kinematics_solver, motor_names=list(leader.bus.motors.keys())
+            ),
+        ],
+        to_transition=robot_action_to_transition,
+        to_output=transition_to_robot_action,
+    )
 
-# Connect to the robot and teleoperator
-follower.connect()
-leader.connect()
+    # build pipeline to convert EE action to robot joints
+    ee_to_follower_joints = RobotProcessorPipeline[tuple[RobotAction, RobotObservation], RobotAction](
+        [
+            EEBoundsAndSafety(
+                end_effector_bounds={"min": [-1.0, -1.0, -1.0], "max": [1.0, 1.0, 1.0]},
+                max_ee_step_m=0.10,
+            ),
+            InverseKinematicsEEToJoints(
+                kinematics=follower_kinematics_solver,
+                motor_names=list(follower.bus.motors.keys()),
+                initial_guess_current_joints=False,
+            ),
+        ],
+        to_transition=robot_action_observation_to_transition,
+        to_output=transition_to_robot_action,
+    )
 
-# Init rerun viewer
-init_rerun(session_name="so100_so100_EE_teleop")
+    # Connect to the robot and teleoperator
+    follower.connect()
+    leader.connect()
 
-print("Starting teleop loop...")
-while True:
-    t0 = time.perf_counter()
+    # Init rerun viewer
+    init_rerun(session_name="so100_so100_EE_teleop")
 
-    # Get robot observation
-    robot_obs = follower.get_observation()
+    print("Starting teleop loop...")
+    while True:
+        t0 = time.perf_counter()
 
-    # Get teleop observation
-    leader_joints_obs = leader.get_action()
+        # Get robot observation
+        robot_obs = follower.get_observation()
 
-    # teleop joints -> teleop EE action
-    leader_ee_act = leader_to_ee(leader_joints_obs)
+        # Get teleop observation
+        leader_joints_obs = leader.get_action()
 
-    # teleop EE -> robot joints
-    follower_joints_act = ee_to_follower_joints((leader_ee_act, robot_obs))
+        # teleop joints -> teleop EE action
+        leader_ee_act = leader_to_ee(leader_joints_obs)
 
-    # Send action to robot
-    _ = follower.send_action(follower_joints_act)
+        # teleop EE -> robot joints
+        follower_joints_act = ee_to_follower_joints((leader_ee_act, robot_obs))
 
-    # Visualize
-    log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
+        # Send action to robot
+        _ = follower.send_action(follower_joints_act)
 
-    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        # Visualize
+        log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
+
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/act/act_training_example.py

@@ -19,80 +19,86 @@ def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[flo
     return [i / fps for i in delta_indices]
 
 
-output_directory = Path("outputs/robot_learning_tutorial/act")
-output_directory.mkdir(parents=True, exist_ok=True)
+def main():
+    output_directory = Path("outputs/robot_learning_tutorial/act")
+    output_directory.mkdir(parents=True, exist_ok=True)
 
-# Select your device
-device = torch.device("mps")  # or "cuda" or "cpu"
+    # Select your device
+    device = torch.device("mps")  # or "cuda" or "cpu"
 
-dataset_id = "lerobot/svla_so101_pickplace"
+    dataset_id = "lerobot/svla_so101_pickplace"
 
-# This specifies the inputs the model will be expecting and the outputs it will produce
-dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-features = dataset_to_policy_features(dataset_metadata.features)
+    # This specifies the inputs the model will be expecting and the outputs it will produce
+    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+    features = dataset_to_policy_features(dataset_metadata.features)
 
-output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
-input_features = {key: ft for key, ft in features.items() if key not in output_features}
+    output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    input_features = {key: ft for key, ft in features.items() if key not in output_features}
 
-cfg = ACTConfig(input_features=input_features, output_features=output_features)
-policy = ACTPolicy(cfg)
-preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
+    cfg = ACTConfig(input_features=input_features, output_features=output_features)
+    policy = ACTPolicy(cfg)
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
 
-policy.train()
-policy.to(device)
+    policy.train()
+    policy.to(device)
 
-# To perform action chunking, ACT expects a given number of actions as targets
-delta_timestamps = {
-    "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
-}
+    # To perform action chunking, ACT expects a given number of actions as targets
+    delta_timestamps = {
+        "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
+    }
 
-# add image features if they are present
-delta_timestamps |= {
-    k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps) for k in cfg.image_features
-}
+    # add image features if they are present
+    delta_timestamps |= {
+        k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps)
+        for k in cfg.image_features
+    }
 
-# Instantiate the dataset
-dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
+    # Instantiate the dataset
+    dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
 
-# Create the optimizer and dataloader for offline training
-optimizer = cfg.get_optimizer_preset().build(policy.parameters())
-batch_size = 32
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    batch_size=batch_size,
-    shuffle=True,
-    pin_memory=device.type != "cpu",
-    drop_last=True,
-)
+    # Create the optimizer and dataloader for offline training
+    optimizer = cfg.get_optimizer_preset().build(policy.parameters())
+    batch_size = 32
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=True,
+        pin_memory=device.type != "cpu",
+        drop_last=True,
+    )
 
-# Number of training steps and logging frequency
-training_steps = 1
-log_freq = 1
+    # Number of training steps and logging frequency
+    training_steps = 1
+    log_freq = 1
 
-# Run training loop
-step = 0
-done = False
-while not done:
-    for batch in dataloader:
-        batch = preprocessor(batch)
-        loss, _ = policy.forward(batch)
-        loss.backward()
-        optimizer.step()
-        optimizer.zero_grad()
+    # Run training loop
+    step = 0
+    done = False
+    while not done:
+        for batch in dataloader:
+            batch = preprocessor(batch)
+            loss, _ = policy.forward(batch)
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
 
-        if step % log_freq == 0:
-            print(f"step: {step} loss: {loss.item():.3f}")
-        step += 1
-        if step >= training_steps:
-            done = True
-            break
+            if step % log_freq == 0:
+                print(f"step: {step} loss: {loss.item():.3f}")
+            step += 1
+            if step >= training_steps:
+                done = True
+                break
 
-# Save the policy checkpoint, alongside the pre/post processors
-policy.save_pretrained(output_directory)
-preprocessor.save_pretrained(output_directory)
-postprocessor.save_pretrained(output_directory)
+    # Save the policy checkpoint, alongside the pre/post processors
+    policy.save_pretrained(output_directory)
+    preprocessor.save_pretrained(output_directory)
+    postprocessor.save_pretrained(output_directory)
 
-# Save all assets to the Hub
-policy.push_to_hub("fracapuano/robot_learning_tutorial_act")
-preprocessor.push_to_hub("fracapuano/robot_learning_tutorial_act")
-postprocessor.push_to_hub("fracapuano/robot_learning_tutorial_act")
+    # Save all assets to the Hub
+    policy.push_to_hub("<user>/robot_learning_tutorial_act")
+    preprocessor.push_to_hub("<user>/robot_learning_tutorial_act")
+    postprocessor.push_to_hub("<user>/robot_learning_tutorial_act")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/act/act_using_example.py

@@ -8,50 +8,56 @@ from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 
-device = torch.device("mps")  # or "cuda" or "cpu"
-model_id = "fracapuano/robot_learning_tutorial_act"
-model = ACTPolicy.from_pretrained(model_id)
-
-dataset_id = "lerobot/svla_so101_pickplace"
-# This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
-dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-preprocess, postprocess = make_pre_post_processors(model.config, dataset_stats=dataset_metadata.stats)
-
-# # find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
-
-# # the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
-
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20
 
-# Robot and environment configuration
-# Camera keys must match the name and resolutions of the ones used for training!
-# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
-camera_config = {
-    "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-}
 
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-robot = SO100Follower(robot_cfg)
-robot.connect()
+def main():
+    device = torch.device("mps")  # or "cuda" or "cpu"
+    model_id = "<user>/robot_learning_tutorial_act"
+    model = ACTPolicy.from_pretrained(model_id)
 
-for _ in range(MAX_EPISODES):
-    for _ in range(MAX_STEPS_PER_EPISODE):
-        obs = robot.get_observation()
-        obs_frame = build_inference_frame(
-            observation=obs, ds_features=dataset_metadata.features, device=device
-        )
+    dataset_id = "lerobot/svla_so101_pickplace"
+    # This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
+    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+    preprocess, postprocess = make_pre_post_processors(model.config, dataset_stats=dataset_metadata.stats)
 
-        obs = preprocess(obs_frame)
+    # # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
 
-        action = model.select_action(obs)
-        action = postprocess(action)
+    # # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
 
-        action = make_robot_action(action, dataset_metadata.features)
+    # Robot and environment configuration
+    # Camera keys must match the name and resolutions of the ones used for training!
+    # You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+    camera_config = {
+        "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    }
 
-        robot.send_action(action)
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+    robot = SO100Follower(robot_cfg)
+    robot.connect()
 
-    print("Episode finished! Starting new episode...")
+    for _ in range(MAX_EPISODES):
+        for _ in range(MAX_STEPS_PER_EPISODE):
+            obs = robot.get_observation()
+            obs_frame = build_inference_frame(
+                observation=obs, ds_features=dataset_metadata.features, device=device
+            )
+
+            obs = preprocess(obs_frame)
+
+            action = model.select_action(obs)
+            action = postprocess(action)
+
+            action = make_robot_action(action, dataset_metadata.features)
+
+            robot.send_action(action)
+
+        print("Episode finished! Starting new episode...")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/async-inf/policy_server.py

@@ -1,11 +1,17 @@
 from lerobot.async_inference.configs import PolicyServerConfig
 from lerobot.async_inference.policy_server import serve
 
-host = ...  # something like "127.0.0.1" if you're exposing to localhost
-port = ...  # something like 8080
 
-config = PolicyServerConfig(
-    host=host,
-    port=port,
-)
-serve(config)
+def main():
+    host = ...  # something like "127.0.0.1" if you're exposing to localhost
+    port = ...  # something like 8080
+
+    config = PolicyServerConfig(
+        host=host,
+        port=port,
+    )
+    serve(config)
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/async-inf/robot_client.py

@@ -6,50 +6,56 @@ from lerobot.async_inference.robot_client import RobotClient
 from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
 from lerobot.robots.so100_follower import SO100FollowerConfig
 
-# these cameras must match the ones expected by the policy - find your cameras with lerobot-find-cameras
-# check the config.json on the Hub for the policy you are using to see the expected camera specs
-camera_cfg = {
-    "up": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "side": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-}
 
-# # find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+def main():
+    # these cameras must match the ones expected by the policy - find your cameras with lerobot-find-cameras
+    # check the config.json on the Hub for the policy you are using to see the expected camera specs
+    camera_cfg = {
+        "up": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "side": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    }
 
-# # the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
+    # # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
 
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_cfg)
+    # # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
 
-server_address = ...  # something like "127.0.0.1:8080" if using localhost
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_cfg)
 
-# 3. Create client configuration
-client_cfg = RobotClientConfig(
-    robot=robot_cfg,
-    server_address=server_address,
-    policy_device="mps",
-    policy_type="act",
-    pretrained_name_or_path="fracapuano/robot_learning_tutorial_act",
-    chunk_size_threshold=0.5,  # g
-    actions_per_chunk=50,  # make sure this is less than the max actions of the policy
-)
+    server_address = ...  # something like "127.0.0.1:8080" if using localhost
 
-# 4. Create and start client
-client = RobotClient(client_cfg)
+    # 3. Create client configuration
+    client_cfg = RobotClientConfig(
+        robot=robot_cfg,
+        server_address=server_address,
+        policy_device="mps",
+        policy_type="act",
+        pretrained_name_or_path="<user>/robot_learning_tutorial_act",
+        chunk_size_threshold=0.5,  # g
+        actions_per_chunk=50,  # make sure this is less than the max actions of the policy
+    )
 
-# 5. Provide a textual description of the task
-task = ...
+    # 4. Create and start client
+    client = RobotClient(client_cfg)
 
-if client.start():
-    # Start action receiver thread
-    action_receiver_thread = threading.Thread(target=client.receive_actions, daemon=True)
-    action_receiver_thread.start()
+    # 5. Provide a textual description of the task
+    task = ...
 
-    try:
-        # Run the control loop
-        client.control_loop(task)
-    except KeyboardInterrupt:
-        client.stop()
-        action_receiver_thread.join()
-        # (Optionally) plot the action queue size
-        visualize_action_queue_size(client.action_queue_size)
+    if client.start():
+        # Start action receiver thread
+        action_receiver_thread = threading.Thread(target=client.receive_actions, daemon=True)
+        action_receiver_thread.start()
+
+        try:
+            # Run the control loop
+            client.control_loop(task)
+        except KeyboardInterrupt:
+            client.stop()
+            action_receiver_thread.join()
+            # (Optionally) plot the action queue size
+            visualize_action_queue_size(client.action_queue_size)
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/diffusion/diffusion_training_example.py

@@ -19,81 +19,87 @@ def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[flo
     return [i / fps for i in delta_indices]
 
 
-output_directory = Path("outputs/robot_learning_tutorial/diffusion")
-output_directory.mkdir(parents=True, exist_ok=True)
+def main():
+    output_directory = Path("outputs/robot_learning_tutorial/diffusion")
+    output_directory.mkdir(parents=True, exist_ok=True)
 
-# Select your device
-device = torch.device("mps")  # or "cuda" or "cpu"
+    # Select your device
+    device = torch.device("mps")  # or "cuda" or "cpu"
 
-dataset_id = "lerobot/svla_so101_pickplace"
+    dataset_id = "lerobot/svla_so101_pickplace"
 
-# This specifies the inputs the model will be expecting and the outputs it will produce
-dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-features = dataset_to_policy_features(dataset_metadata.features)
+    # This specifies the inputs the model will be expecting and the outputs it will produce
+    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+    features = dataset_to_policy_features(dataset_metadata.features)
 
-output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
-input_features = {key: ft for key, ft in features.items() if key not in output_features}
+    output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    input_features = {key: ft for key, ft in features.items() if key not in output_features}
 
-cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
-policy = DiffusionPolicy(cfg)
-preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
+    cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
+    policy = DiffusionPolicy(cfg)
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
 
-policy.train()
-policy.to(device)
+    policy.train()
+    policy.to(device)
 
-# To perform action chunking, ACT expects a given number of actions as targets
-delta_timestamps = {
-    "observation.state": make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps),
-    "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
-}
+    # To perform action chunking, ACT expects a given number of actions as targets
+    delta_timestamps = {
+        "observation.state": make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps),
+        "action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
+    }
 
-# add image features if they are present
-delta_timestamps |= {
-    k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps) for k in cfg.image_features
-}
+    # add image features if they are present
+    delta_timestamps |= {
+        k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps)
+        for k in cfg.image_features
+    }
 
-# Instantiate the dataset
-dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
+    # Instantiate the dataset
+    dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
 
-# Create the optimizer and dataloader for offline training
-optimizer = cfg.get_optimizer_preset().build(policy.parameters())
-batch_size = 32
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    batch_size=batch_size,
-    shuffle=True,
-    pin_memory=device.type != "cpu",
-    drop_last=True,
-)
+    # Create the optimizer and dataloader for offline training
+    optimizer = cfg.get_optimizer_preset().build(policy.parameters())
+    batch_size = 32
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=True,
+        pin_memory=device.type != "cpu",
+        drop_last=True,
+    )
 
-# Number of training steps and logging frequency
-training_steps = 1
-log_freq = 1
+    # Number of training steps and logging frequency
+    training_steps = 1
+    log_freq = 1
 
-# Run training loop
-step = 0
-done = False
-while not done:
-    for batch in dataloader:
-        batch = preprocessor(batch)
-        loss, _ = policy.forward(batch)
-        loss.backward()
-        optimizer.step()
-        optimizer.zero_grad()
+    # Run training loop
+    step = 0
+    done = False
+    while not done:
+        for batch in dataloader:
+            batch = preprocessor(batch)
+            loss, _ = policy.forward(batch)
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
 
-        if step % log_freq == 0:
-            print(f"step: {step} loss: {loss.item():.3f}")
-        step += 1
-        if step >= training_steps:
-            done = True
-            break
+            if step % log_freq == 0:
+                print(f"step: {step} loss: {loss.item():.3f}")
+            step += 1
+            if step >= training_steps:
+                done = True
+                break
 
-# Save the policy checkpoint, alongside the pre/post processors
-policy.save_pretrained(output_directory)
-preprocessor.save_pretrained(output_directory)
-postprocessor.save_pretrained(output_directory)
+    # Save the policy checkpoint, alongside the pre/post processors
+    policy.save_pretrained(output_directory)
+    preprocessor.save_pretrained(output_directory)
+    postprocessor.save_pretrained(output_directory)
 
-# Save all assets to the Hub
-policy.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
-preprocessor.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
-postprocessor.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
+    # Save all assets to the Hub
+    policy.push_to_hub("<user>/robot_learning_tutorial_diffusion")
+    preprocessor.push_to_hub("<user>/robot_learning_tutorial_diffusion")
+    postprocessor.push_to_hub("<user>/robot_learning_tutorial_diffusion")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/diffusion/diffusion_using_example.py

@@ -8,53 +8,57 @@ from lerobot.policies.utils import build_inference_frame, make_robot_action
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
 
-device = torch.device("mps")  # or "cuda" or "cpu"
-model_id = "fracapuano/robot_learning_tutorial_diffusion"
-
-model = DiffusionPolicy.from_pretrained(model_id)
-
-dataset_id = "lerobot/svla_so101_pickplace"
-# This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
-dataset_metadata = LeRobotDatasetMetadata(dataset_id)
-preprocess, postprocess = make_pre_post_processors(
-    model.config, model_id, dataset_stats=dataset_metadata.stats
-)
-
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20
 
 
-# # find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+def main():
+    device = torch.device("mps")  # or "cuda" or "cpu"
+    model_id = "<user>/robot_learning_tutorial_diffusion"
 
-# # the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
+    model = DiffusionPolicy.from_pretrained(model_id)
 
-# Robot and environment configuration
-# Camera keys must match the name and resolutions of the ones used for training!
-# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
-camera_config = {
-    "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-}
+    dataset_id = "lerobot/svla_so101_pickplace"
+    # This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
+    dataset_metadata = LeRobotDatasetMetadata(dataset_id)
+    preprocess, postprocess = make_pre_post_processors(
+        model.config, model_id, dataset_stats=dataset_metadata.stats
+    )
 
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-robot = SO100Follower(robot_cfg)
-robot.connect()
+    # # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+
+    # # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
+
+    # Robot and environment configuration
+    # Camera keys must match the name and resolutions of the ones used for training!
+    # You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+    camera_config = {
+        "side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    }
+
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+    robot = SO100Follower(robot_cfg)
+    robot.connect()
+
+    for _ in range(MAX_EPISODES):
+        for _ in range(MAX_STEPS_PER_EPISODE):
+            obs = robot.get_observation()
+            obs_frame = build_inference_frame(
+                observation=obs, ds_features=dataset_metadata.features, device=device
+            )
+
+            obs = preprocess(obs_frame)
+
+            action = model.select_action(obs)
+            action = postprocess(action)
+            action = make_robot_action(action, dataset_metadata.features)
+            robot.send_action(action)
+
+        print("Episode finished! Starting new episode...")
 
 
-for _ in range(MAX_EPISODES):
-    for _ in range(MAX_STEPS_PER_EPISODE):
-        obs = robot.get_observation()
-        obs_frame = build_inference_frame(
-            observation=obs, ds_features=dataset_metadata.features, device=device
-        )
-
-        obs = preprocess(obs_frame)
-
-        action = model.select_action(obs)
-        action = postprocess(action)
-        action = make_robot_action(action, dataset_metadata.features)
-        robot.send_action(action)
-
-    print("Episode finished! Starting new episode...")
+if __name__ == "__main__":
+    main()

examples/tutorial/pi0/using_pi0_example.py

@@ -11,57 +11,63 @@ from lerobot.robots.so100_follower.so100_follower import SO100Follower
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20
 
-device = torch.device("mps")  # or "cuda" or "cpu"
-model_id = "lerobot/pi0_base"
 
-model = PI0Policy.from_pretrained(model_id)
+def main():
+    device = torch.device("mps")  # or "cuda" or "cpu"
+    model_id = "lerobot/pi0_base"
 
-preprocess, postprocess = make_pre_post_processors(
-    model.config,
-    model_id,
-    # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
-    preprocessor_overrides={"device_processor": {"device": str(device)}},
-)
+    model = PI0Policy.from_pretrained(model_id)
 
-# find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+    preprocess, postprocess = make_pre_post_processors(
+        model.config,
+        model_id,
+        # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
+        preprocessor_overrides={"device_processor": {"device": str(device)}},
+    )
 
-# the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
+    # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
 
-# Robot and environment configuration
-# Camera keys must match the name and resolutions of the ones used for training!
-# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
-camera_config = {
-    "base_0_rgb": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "left_wrist_0_rgb": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-    "right_wrist_0_rgb": OpenCVCameraConfig(index_or_path=2, width=640, height=480, fps=30),
-}
+    # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
 
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-robot = SO100Follower(robot_cfg)
-robot.connect()
+    # Robot and environment configuration
+    # Camera keys must match the name and resolutions of the ones used for training!
+    # You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+    camera_config = {
+        "base_0_rgb": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "left_wrist_0_rgb": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+        "right_wrist_0_rgb": OpenCVCameraConfig(index_or_path=2, width=640, height=480, fps=30),
+    }
 
-task = ""  # something like "pick the red block"
-robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+    robot = SO100Follower(robot_cfg)
+    robot.connect()
 
-# This is used to match the raw observation keys to the keys expected by the policy
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-dataset_features = {**action_features, **obs_features}
+    task = ""  # something like "pick the red block"
+    robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
 
-for _ in range(MAX_EPISODES):
-    for _ in range(MAX_STEPS_PER_EPISODE):
-        obs = robot.get_observation()
-        obs_frame = build_inference_frame(
-            observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
-        )
+    # This is used to match the raw observation keys to the keys expected by the policy
+    action_features = hw_to_dataset_features(robot.action_features, "action")
+    obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+    dataset_features = {**action_features, **obs_features}
 
-        obs = preprocess(obs_frame)
+    for _ in range(MAX_EPISODES):
+        for _ in range(MAX_STEPS_PER_EPISODE):
+            obs = robot.get_observation()
+            obs_frame = build_inference_frame(
+                observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
+            )
 
-        action = model.select_action(obs)
-        action = postprocess(action)
-        action = make_robot_action(action, dataset_features)
-        robot.send_action(action)
+            obs = preprocess(obs_frame)
 
-    print("Episode finished! Starting new episode...")
+            action = model.select_action(obs)
+            action = postprocess(action)
+            action = make_robot_action(action, dataset_features)
+            robot.send_action(action)
+
+        print("Episode finished! Starting new episode...")
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/rl/hilserl_example.py

@@ -20,6 +20,8 @@ from lerobot.teleoperators.utils import TeleopEvents
 
 LOG_EVERY = 10
 SEND_EVERY = 10
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
 
 
 def run_learner(
@@ -223,123 +225,123 @@ def make_policy_obs(obs, device: torch.device = "cpu"):
     }
 
 
-"""Main function - coordinates actor and learner processes."""
+def main():
+    """Main function - coordinates actor and learner processes."""
 
-device = "mps"  # or "cuda" or "cpu"
-output_directory = Path("outputs/robot_learning_tutorial/hil_serl")
-output_directory.mkdir(parents=True, exist_ok=True)
+    device = "mps"  # or "cuda" or "cpu"
+    output_directory = Path("outputs/robot_learning_tutorial/hil_serl")
+    output_directory.mkdir(parents=True, exist_ok=True)
 
-# find ports using lerobot-find-port
-follower_port = ...
-leader_port = ...
+    # find ports using lerobot-find-port
+    follower_port = ...
+    leader_port = ...
 
-# the robot ids are used the load the right calibration files
-follower_id = ...
-leader_id = ...
+    # the robot ids are used the load the right calibration files
+    follower_id = ...
+    leader_id = ...
 
-# A pretrained model (to be used in-distribution!)
-reward_classifier_id = "fracapuano/reward_classifier_hil_serl_example"
-reward_classifier = Classifier.from_pretrained(reward_classifier_id)
+    # A pretrained model (to be used in-distribution!)
+    reward_classifier_id = "<user>/reward_classifier_hil_serl_example"
+    reward_classifier = Classifier.from_pretrained(reward_classifier_id)
 
-reward_classifier.to(device)
-reward_classifier.eval()
+    reward_classifier.to(device)
+    reward_classifier.eval()
 
-MAX_EPISODES = 5
-MAX_STEPS_PER_EPISODE = 20
+    # Robot and environment configuration
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id)
+    teleop_cfg = SO100LeaderConfig(port=leader_port, id=leader_id)
+    processor_cfg = HILSerlProcessorConfig(control_mode="leader")
 
-# Robot and environment configuration
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id)
-teleop_cfg = SO100LeaderConfig(port=leader_port, id=leader_id)
-processor_cfg = HILSerlProcessorConfig(control_mode="leader")
+    env_cfg = HILSerlRobotEnvConfig(robot=robot_cfg, teleop=teleop_cfg, processor=processor_cfg)
 
-env_cfg = HILSerlRobotEnvConfig(robot=robot_cfg, teleop=teleop_cfg, processor=processor_cfg)
+    # Create robot environment
+    env, teleop_device = make_robot_env(env_cfg)
 
-# Create robot environment
-env, teleop_device = make_robot_env(env_cfg)
+    obs_features = hw_to_dataset_features(env.robot.observation_features, "observation")
+    action_features = hw_to_dataset_features(env.robot.action_features, "action")
 
-obs_features = hw_to_dataset_features(env.robot.observation_features, "observation")
-action_features = hw_to_dataset_features(env.robot.action_features, "action")
+    # Create SAC policy for action selection
+    policy_cfg = SACConfig(
+        device=device,
+        input_features=obs_features,
+        output_features=action_features,
+    )
 
-# Create SAC policy for action selection
-policy_cfg = SACConfig(
-    device=device,
-    input_features=obs_features,
-    output_features=action_features,
-)
+    policy_actor = SACPolicy(policy_cfg)
+    policy_learner = SACPolicy(policy_cfg)
 
-policy_actor = SACPolicy(policy_cfg)
-policy_learner = SACPolicy(policy_cfg)
+    demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
+    offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)
 
-demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
-offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)
+    # Online buffer: initialized from scratch
+    online_replay_buffer = ReplayBuffer(device=device, state_keys=list(obs_features.keys()))
+    # Offline buffer: Created from dataset (pre-populated it with demonstrations)
+    offline_replay_buffer = ReplayBuffer.from_lerobot_dataset(
+        lerobot_dataset=offline_dataset, device=device, state_keys=list(obs_features.keys())
+    )
 
-# Online buffer: initialized from scratch
-online_replay_buffer = ReplayBuffer(device=device, state_keys=list(obs_features.keys()))
-# Offline buffer: Created from dataset (pre-populated it with demonstrations)
-offline_replay_buffer = ReplayBuffer.from_lerobot_dataset(
-    lerobot_dataset=offline_dataset, device=device, state_keys=list(obs_features.keys())
-)
+    # Create communication channels between learner and actor processes
+    transitions_queue = mp.Queue(maxsize=10)
+    parameters_queue = mp.Queue(maxsize=2)
+    shutdown_event = mp.Event()
 
-# Create communication channels between learner and actor processes
-transitions_queue = mp.Queue(maxsize=10)
-parameters_queue = mp.Queue(maxsize=2)
-shutdown_event = mp.Event()
+    # Signal handler for graceful shutdown
+    def signal_handler(sig):
+        print(f"\nSignal {sig} received, shutting down...")
+        shutdown_event.set()
+
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTERM, signal_handler)
+
+    # Create processes
+    learner_process = mp.Process(
+        target=run_learner,
+        args=(
+            transitions_queue,
+            parameters_queue,
+            shutdown_event,
+            policy_learner,
+            online_replay_buffer,
+            offline_replay_buffer,
+        ),
+        kwargs={"device": device},  # can run on accelerated hardware for training
+    )
+
+    actor_process = mp.Process(
+        target=run_actor,
+        args=(
+            transitions_queue,
+            parameters_queue,
+            shutdown_event,
+            policy_actor,
+            reward_classifier,
+            env_cfg,
+            output_directory,
+        ),
+        kwargs={"device": "cpu"},  # actor is frozen, can run on CPU or accelerate for inference
+    )
+
+    learner_process.start()
+    actor_process.start()
+
+    try:
+        # Wait for actor to finish (it controls the episode loop)
+        actor_process.join()
+        shutdown_event.set()
+        learner_process.join(timeout=10)
+
+    except KeyboardInterrupt:
+        print("Main process interrupted")
+        shutdown_event.set()
+        actor_process.join(timeout=5)
+        learner_process.join(timeout=10)
+
+    finally:
+        if learner_process.is_alive():
+            learner_process.terminate()
+        if actor_process.is_alive():
+            actor_process.terminate()
 
 
-# Signal handler for graceful shutdown
-def signal_handler(sig):
-    print(f"\nSignal {sig} received, shutting down...")
-    shutdown_event.set()
-
-
-signal.signal(signal.SIGINT, signal_handler)
-signal.signal(signal.SIGTERM, signal_handler)
-
-# Create processes
-learner_process = mp.Process(
-    target=run_learner,
-    args=(
-        transitions_queue,
-        parameters_queue,
-        shutdown_event,
-        policy_learner,
-        online_replay_buffer,
-        offline_replay_buffer,
-    ),
-    kwargs={"device": device},  # can run on accelerated hardware for training
-)
-
-actor_process = mp.Process(
-    target=run_actor,
-    args=(
-        transitions_queue,
-        parameters_queue,
-        shutdown_event,
-        policy_actor,
-        reward_classifier,
-        env_cfg,
-        output_directory,
-    ),
-    kwargs={"device": "cpu"},  # actor is frozen, can run on CPU or accelerate for inference
-)
-
-learner_process.start()
-actor_process.start()
-
-try:
-    # Wait for actor to finish (it controls the episode loop)
-    actor_process.join()
-    shutdown_event.set()
-    learner_process.join(timeout=10)
-
-except KeyboardInterrupt:
-    print("Main process interrupted")
-    shutdown_event.set()
-    actor_process.join(timeout=5)
-    learner_process.join(timeout=10)
-
-finally:
-    if learner_process.is_alive():
-        learner_process.terminate()
-    if actor_process.is_alive():
-        actor_process.terminate()
+if __name__ == "__main__":
+    main()

examples/tutorial/rl/reward_classifier_example.py

@@ -4,59 +4,64 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.policies.factory import make_policy, make_pre_post_processors
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
 
-# Device to use for training
-device = "mps"  # or "cuda", or "cpu"
 
-# Load the dataset used for training
-repo_id = "lerobot/example_hil_serl_dataset"
-dataset = LeRobotDataset(repo_id)
+def main():
+    # Device to use for training
+    device = "mps"  # or "cuda", or "cpu"
 
-# Configure the policy to extract features from the image frames
-camera_keys = dataset.meta.camera_keys
+    # Load the dataset used for training
+    repo_id = "lerobot/example_hil_serl_dataset"
+    dataset = LeRobotDataset(repo_id)
 
-config = RewardClassifierConfig(
-    num_cameras=len(camera_keys),
-    device=device,
-    # backbone model to extract features from the image frames
-    model_name="microsoft/resnet-18",
-)
+    # Configure the policy to extract features from the image frames
+    camera_keys = dataset.meta.camera_keys
 
-# Make policy, preprocessor, and optimizer
-policy = make_policy(config, ds_meta=dataset.meta)
-optimizer = config.get_optimizer_preset().build(policy.parameters())
-preprocessor, _ = make_pre_post_processors(policy_cfg=config, dataset_stats=dataset.meta.stats)
+    config = RewardClassifierConfig(
+        num_cameras=len(camera_keys),
+        device=device,
+        # backbone model to extract features from the image frames
+        model_name="microsoft/resnet-18",
+    )
+
+    # Make policy, preprocessor, and optimizer
+    policy = make_policy(config, ds_meta=dataset.meta)
+    optimizer = config.get_optimizer_preset().build(policy.parameters())
+    preprocessor, _ = make_pre_post_processors(policy_cfg=config, dataset_stats=dataset.meta.stats)
+
+    classifier_id = "<user>/reward_classifier_hil_serl_example"
+
+    # Instantiate a dataloader
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=16, shuffle=True)
+
+    # Training loop
+    num_epochs = 5
+    for epoch in range(num_epochs):
+        total_loss = 0
+        total_accuracy = 0
+        for batch in dataloader:
+            # Preprocess the batch and move it to the correct device.
+            batch = preprocessor(batch)
+
+            # Forward pass
+            loss, output_dict = policy.forward(batch)
+
+            # Backward pass and optimization
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+            total_loss += loss.item()
+            total_accuracy += output_dict["accuracy"]
+
+        avg_loss = total_loss / len(dataloader)
+        avg_accuracy = total_accuracy / len(dataloader)
+        print(f"Epoch {epoch + 1}/{num_epochs}, Loss: {avg_loss:.4f}, Accuracy: {avg_accuracy:.2f}%")
+
+    print("Training finished!")
+
+    # You can now save the trained policy.
+    policy.push_to_hub(classifier_id)
 
 
-classifier_id = "fracapuano/reward_classifier_hil_serl_example"
-
-# Instantiate a dataloader
-dataloader = torch.utils.data.DataLoader(dataset, batch_size=16, shuffle=True)
-
-# Training loop
-num_epochs = 5
-for epoch in range(num_epochs):
-    total_loss = 0
-    total_accuracy = 0
-    for batch in dataloader:
-        # Preprocess the batch and move it to the correct device.
-        batch = preprocessor(batch)
-
-        # Forward pass
-        loss, output_dict = policy.forward(batch)
-
-        # Backward pass and optimization
-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
-
-        total_loss += loss.item()
-        total_accuracy += output_dict["accuracy"]
-
-    avg_loss = total_loss / len(dataloader)
-    avg_accuracy = total_accuracy / len(dataloader)
-    print(f"Epoch {epoch + 1}/{num_epochs}, Loss: {avg_loss:.4f}, Accuracy: {avg_accuracy:.2f}%")
-
-print("Training finished!")
-
-# You can now save the trained policy.
-policy.push_to_hub(classifier_id)
+if __name__ == "__main__":
+    main()

examples/tutorial/smolvla/using_smolvla_example.py

@@ -11,56 +11,62 @@ from lerobot.robots.so100_follower.so100_follower import SO100Follower
 MAX_EPISODES = 5
 MAX_STEPS_PER_EPISODE = 20
 
-device = torch.device("mps")  # or "cuda" or "cpu"
-model_id = "lerobot/smolvla_base"
 
-model = SmolVLAPolicy.from_pretrained(model_id)
+def main():
+    device = torch.device("mps")  # or "cuda" or "cpu"
+    model_id = "lerobot/smolvla_base"
 
-preprocess, postprocess = make_pre_post_processors(
-    model.config,
-    model_id,
-    # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
-    preprocessor_overrides={"device_processor": {"device": str(device)}},
-)
+    model = SmolVLAPolicy.from_pretrained(model_id)
 
-# find ports using lerobot-find-port
-follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+    preprocess, postprocess = make_pre_post_processors(
+        model.config,
+        model_id,
+        # This overrides allows to run on MPS, otherwise defaults to CUDA (if available)
+        preprocessor_overrides={"device_processor": {"device": str(device)}},
+    )
 
-# the robot ids are used the load the right calibration files
-follower_id = ...  # something like "follower_so100"
+    # find ports using lerobot-find-port
+    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
 
-# Robot and environment configuration
-# Camera keys must match the name and resolutions of the ones used for training!
-# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
-camera_config = {
-    "camera1": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
-    "camera2": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
-}
+    # the robot ids are used the load the right calibration files
+    follower_id = ...  # something like "follower_so100"
 
-robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-robot = SO100Follower(robot_cfg)
-robot.connect()
+    # Robot and environment configuration
+    # Camera keys must match the name and resolutions of the ones used for training!
+    # You can check the camera keys expected by a model in the info.json card on the model card on the Hub
+    camera_config = {
+        "camera1": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+        "camera2": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+    }
 
-task = ""  # something like "pick the red block"
-robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
+    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+    robot = SO100Follower(robot_cfg)
+    robot.connect()
 
-# This is used to match the raw observation keys to the keys expected by the policy
-action_features = hw_to_dataset_features(robot.action_features, "action")
-obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-dataset_features = {**action_features, **obs_features}
+    task = ""  # something like "pick the red block"
+    robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
 
-for _ in range(MAX_EPISODES):
-    for _ in range(MAX_STEPS_PER_EPISODE):
-        obs = robot.get_observation()
-        obs_frame = build_inference_frame(
-            observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
-        )
+    # This is used to match the raw observation keys to the keys expected by the policy
+    action_features = hw_to_dataset_features(robot.action_features, "action")
+    obs_features = hw_to_dataset_features(robot.observation_features, "observation")
+    dataset_features = {**action_features, **obs_features}
 
-        obs = preprocess(obs_frame)
+    for _ in range(MAX_EPISODES):
+        for _ in range(MAX_STEPS_PER_EPISODE):
+            obs = robot.get_observation()
+            obs_frame = build_inference_frame(
+                observation=obs, ds_features=dataset_features, device=device, task=task, robot_type=robot_type
+            )
 
-        action = model.select_action(obs)
-        action = postprocess(action)
-        action = make_robot_action(action, dataset_features)
-        robot.send_action(action)
+            obs = preprocess(obs_frame)
 
-    print("Episode finished! Starting new episode...")
+            action = model.select_action(obs)
+            action = postprocess(action)
+            action = make_robot_action(action, dataset_features)
+            robot.send_action(action)
+
+        print("Episode finished! Starting new episode...")
+
+
+if __name__ == "__main__":
+    main()

examples/unitree_g1/gr00t_locomotion.py

@@ -0,0 +1,347 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Example: GR00T Locomotion with Pre-loaded Policies
+
+This example demonstrates the NEW pattern for loading GR00T policies externally
+and passing them to the robot class.
+"""
+
+import argparse
+import logging
+import threading
+import time
+from collections import deque
+
+import numpy as np
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
+from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
+
+logger = logging.getLogger(__name__)
+
+GROOT_DEFAULT_ANGLES = np.zeros(29, dtype=np.float32)
+GROOT_DEFAULT_ANGLES[[0, 6]] = -0.1  # hip pitch
+GROOT_DEFAULT_ANGLES[[3, 9]] = 0.3  # knee
+GROOT_DEFAULT_ANGLES[[4, 10]] = -0.2  # ankle pitch
+
+MISSING_JOINTS = []
+G1_MODEL = "g1_23"  # or "g1_29"
+if G1_MODEL == "g1_23":
+    MISSING_JOINTS = [12, 14, 20, 21, 27, 28]  # waist yaw/pitch, wrist pitch/yaw
+
+LOCOMOTION_ACTION_SCALE = 0.25
+
+LOCOMOTION_CONTROL_DT = 0.02
+
+ANG_VEL_SCALE: float = 0.25
+DOF_POS_SCALE: float = 1.0
+DOF_VEL_SCALE: float = 0.05
+CMD_SCALE: list = [2.0, 2.0, 0.25]
+
+
+DEFAULT_GROOT_REPO_ID = "nepyope/GR00T-WholeBodyControl_g1"
+
+
+def load_groot_policies(
+    repo_id: str = DEFAULT_GROOT_REPO_ID,
+) -> tuple[ort.InferenceSession, ort.InferenceSession]:
+    """Load GR00T dual-policy system (Balance + Walk) from Hugging Face Hub.
+
+    Args:
+        repo_id: Hugging Face Hub repository ID containing the ONNX policies.
+    """
+    logger.info(f"Loading GR00T dual-policy system from Hugging Face Hub ({repo_id})...")
+
+    # Download ONNX policies from Hugging Face Hub
+    balance_path = hf_hub_download(
+        repo_id=repo_id,
+        filename="GR00T-WholeBodyControl-Balance.onnx",
+    )
+    walk_path = hf_hub_download(
+        repo_id=repo_id,
+        filename="GR00T-WholeBodyControl-Walk.onnx",
+    )
+
+    # Load ONNX policies
+    policy_balance = ort.InferenceSession(balance_path)
+    policy_walk = ort.InferenceSession(walk_path)
+
+    logger.info("GR00T policies loaded successfully")
+
+    return policy_balance, policy_walk
+
+
+class GrootLocomotionController:
+    """
+    Handles GR00T-style locomotion control for the Unitree G1 robot.
+
+    This controller manages:
+    - Dual-policy system (Balance + Walk)
+    - 29-joint observation processing
+    - 15D action output (legs + waist)
+    - Policy inference and motor command generation
+    """
+
+    def __init__(self, policy_balance, policy_walk, robot, config):
+        self.policy_balance = policy_balance
+        self.policy_walk = policy_walk
+        self.robot = robot
+        self.config = config
+
+        self.locomotion_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)  # vx, vy, theta_dot
+
+        # GR00T-specific state
+        self.groot_qj_all = np.zeros(29, dtype=np.float32)
+        self.groot_dqj_all = np.zeros(29, dtype=np.float32)
+        self.groot_action = np.zeros(15, dtype=np.float32)
+        self.groot_obs_single = np.zeros(86, dtype=np.float32)
+        self.groot_obs_history = deque(maxlen=6)
+        self.groot_obs_stacked = np.zeros(516, dtype=np.float32)
+        self.groot_height_cmd = 0.74  # Default base height
+        self.groot_orientation_cmd = np.array([0.0, 0.0, 0.0], dtype=np.float32)
+
+        # input to gr00t is 6 frames (6*86D=516)
+        for _ in range(6):
+            self.groot_obs_history.append(np.zeros(86, dtype=np.float32))
+
+        # Thread management
+        self.locomotion_running = False
+        self.locomotion_thread = None
+
+        logger.info("GrootLocomotionController initialized")
+
+    def groot_locomotion_run(self):
+        # get current observation
+        robot_state = self.robot.get_observation()
+
+        if robot_state is None:
+            return
+
+        # get command from remote controller
+        if robot_state.wireless_remote is not None:
+            self.robot.remote_controller.set(robot_state.wireless_remote)
+            if self.robot.remote_controller.button[0]:  # R1 - raise waist
+                self.groot_height_cmd += 0.001
+                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
+            if self.robot.remote_controller.button[4]:  # R2 - lower waist
+                self.groot_height_cmd -= 0.001
+                self.groot_height_cmd = np.clip(self.groot_height_cmd, 0.50, 1.00)
+        else:
+            self.robot.remote_controller.lx = 0.0
+            self.robot.remote_controller.ly = 0.0
+            self.robot.remote_controller.rx = 0.0
+            self.robot.remote_controller.ry = 0.0
+
+        self.locomotion_cmd[0] = self.robot.remote_controller.ly  # forward/backward
+        self.locomotion_cmd[1] = self.robot.remote_controller.lx * -1  # left/right
+        self.locomotion_cmd[2] = self.robot.remote_controller.rx * -1  # rotation rate
+
+        for i in range(29):
+            self.groot_qj_all[i] = robot_state.motor_state[i].q
+            self.groot_dqj_all[i] = robot_state.motor_state[i].dq
+
+        # adapt observation for g1_23dof
+        for idx in MISSING_JOINTS:
+            self.groot_qj_all[idx] = 0.0
+            self.groot_dqj_all[idx] = 0.0
+
+        # Scale joint positions and velocities
+        qj_obs = self.groot_qj_all.copy()
+        dqj_obs = self.groot_dqj_all.copy()
+
+        # express imu data in gravity frame of reference
+        quat = robot_state.imu_state.quaternion
+        ang_vel = np.array(robot_state.imu_state.gyroscope, dtype=np.float32)
+        gravity_orientation = self.robot.get_gravity_orientation(quat)
+
+        # scale joint positions and velocities before policy inference
+        qj_obs = (qj_obs - GROOT_DEFAULT_ANGLES) * DOF_POS_SCALE
+        dqj_obs = dqj_obs * DOF_VEL_SCALE
+        ang_vel_scaled = ang_vel * ANG_VEL_SCALE
+
+        # build single frame observation
+        self.groot_obs_single[:3] = self.locomotion_cmd * np.array(CMD_SCALE)
+        self.groot_obs_single[3] = self.groot_height_cmd
+        self.groot_obs_single[4:7] = self.groot_orientation_cmd
+        self.groot_obs_single[7:10] = ang_vel_scaled
+        self.groot_obs_single[10:13] = gravity_orientation
+        self.groot_obs_single[13:42] = qj_obs
+        self.groot_obs_single[42:71] = dqj_obs
+        self.groot_obs_single[71:86] = self.groot_action  # 15D previous actions
+
+        # Add to history and stack observations (6 frames × 86D = 516D)
+        self.groot_obs_history.append(self.groot_obs_single.copy())
+
+        # Stack all 6 frames into 516D vector
+        for i, obs_frame in enumerate(self.groot_obs_history):
+            start_idx = i * 86
+            end_idx = start_idx + 86
+            self.groot_obs_stacked[start_idx:end_idx] = obs_frame
+
+        # Run policy inference (ONNX) with 516D stacked observation
+
+        cmd_magnitude = np.linalg.norm(self.locomotion_cmd)
+
+        selected_policy = (
+            self.policy_balance if cmd_magnitude < 0.05 else self.policy_walk
+        )  # balance/standing policy for small commands, walking policy for movement commands
+
+        # run policy inference
+        ort_inputs = {selected_policy.get_inputs()[0].name: np.expand_dims(self.groot_obs_stacked, axis=0)}
+        ort_outs = selected_policy.run(None, ort_inputs)
+        self.groot_action = ort_outs[0].squeeze()
+
+        # transform action back to target joint positions
+        target_dof_pos_15 = GROOT_DEFAULT_ANGLES[:15] + self.groot_action * LOCOMOTION_ACTION_SCALE
+
+        # command motors
+        for i in range(15):
+            motor_idx = i
+            self.robot.msg.motor_cmd[motor_idx].q = target_dof_pos_15[i]
+            self.robot.msg.motor_cmd[motor_idx].qd = 0
+            self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
+            self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
+            self.robot.msg.motor_cmd[motor_idx].tau = 0
+
+        # adapt action for g1_23dof
+        for joint_idx in MISSING_JOINTS:
+            self.robot.msg.motor_cmd[joint_idx].q = 0.0
+            self.robot.msg.motor_cmd[joint_idx].qd = 0
+            self.robot.msg.motor_cmd[joint_idx].kp = self.robot.kp[joint_idx]
+            self.robot.msg.motor_cmd[joint_idx].kd = self.robot.kd[joint_idx]
+            self.robot.msg.motor_cmd[joint_idx].tau = 0
+
+        # send action to robot
+        self.robot.send_action(self.robot.msg)
+
+    def _locomotion_thread_loop(self):
+        """Background thread that runs the locomotion policy at specified rate."""
+        logger.info("Locomotion thread started")
+        while self.locomotion_running:
+            start_time = time.time()
+            try:
+                self.groot_locomotion_run()
+            except Exception as e:
+                logger.error(f"Error in locomotion loop: {e}")
+
+            # Sleep to maintain control rate
+            elapsed = time.time() - start_time
+            sleep_time = max(0, LOCOMOTION_CONTROL_DT - elapsed)
+            time.sleep(sleep_time)
+        logger.info("Locomotion thread stopped")
+
+    def start_locomotion_thread(self):
+        if self.locomotion_running:
+            logger.warning("Locomotion thread already running")
+            return
+
+        logger.info("Starting locomotion control thread...")
+        self.locomotion_running = True
+        self.locomotion_thread = threading.Thread(target=self._locomotion_thread_loop, daemon=True)
+        self.locomotion_thread.start()
+
+        logger.info("Locomotion control thread started!")
+
+    def stop_locomotion_thread(self):
+        if not self.locomotion_running:
+            return
+
+        logger.info("Stopping locomotion control thread...")
+        self.locomotion_running = False
+        if self.locomotion_thread:
+            self.locomotion_thread.join(timeout=2.0)
+        logger.info("Locomotion control thread stopped")
+
+    def reset_robot(self):
+        """Move robot legs to default standing position over 2 seconds (arms are not moved)."""
+        total_time = 3.0
+        num_step = int(total_time / self.robot.control_dt)
+
+        # Only control legs, not arms (first 12 joints)
+        default_pos = GROOT_DEFAULT_ANGLES  # First 12 values are leg angles
+        dof_size = len(default_pos)
+
+        # Get current lowstate
+        robot_state = self.robot.get_observation()
+
+        # Record the current leg positions
+        init_dof_pos = np.zeros(dof_size, dtype=np.float32)
+        for i in range(dof_size):
+            init_dof_pos[i] = robot_state.motor_state[i].q
+
+        # Move legs to default pos
+        for i in range(num_step):
+            alpha = i / num_step
+            for motor_idx in range(dof_size):
+                target_pos = default_pos[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].q = (
+                    init_dof_pos[motor_idx] * (1 - alpha) + target_pos * alpha
+                )
+                self.robot.msg.motor_cmd[motor_idx].qd = 0
+                self.robot.msg.motor_cmd[motor_idx].kp = self.robot.kp[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].kd = self.robot.kd[motor_idx]
+                self.robot.msg.motor_cmd[motor_idx].tau = 0
+            self.robot.msg.crc = self.robot.crc.Crc(self.robot.msg)
+            self.robot.lowcmd_publisher.Write(self.robot.msg)
+            time.sleep(self.robot.control_dt)
+        logger.info("Reached default position (legs only)")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="GR00T Locomotion Controller for Unitree G1")
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        default=DEFAULT_GROOT_REPO_ID,
+        help=f"Hugging Face Hub repo ID for GR00T policies (default: {DEFAULT_GROOT_REPO_ID})",
+    )
+    args = parser.parse_args()
+
+    # load policies
+    policy_balance, policy_walk = load_groot_policies(repo_id=args.repo_id)
+
+    # initialize robot
+    config = UnitreeG1Config()
+    robot = UnitreeG1(config)
+
+    # initialize gr00t locomotion controller
+    groot_controller = GrootLocomotionController(
+        policy_balance=policy_balance,
+        policy_walk=policy_walk,
+        robot=robot,
+        config=config,
+    )
+
+    # reset legs and start locomotion thread
+    try:
+        groot_controller.reset_robot()
+        groot_controller.start_locomotion_thread()
+
+        # log status
+        logger.info("Robot initialized with GR00T locomotion policies")
+        logger.info("Locomotion controller running in background thread")
+        logger.info("Press Ctrl+C to stop")
+
+        # keep robot alive
+        while True:
+            time.sleep(1.0)
+    except KeyboardInterrupt:
+        print("\nStopping locomotion...")
+        groot_controller.stop_locomotion_thread()
+        print("Done!")

get_calibration.py

@@ -1,107 +0,0 @@
-import json
-import time
-import math
-from pathlib import Path
-
-# ---- key → (section, name, id)
-MAP = {
-    # LEFT
-    "kLeftShoulderPitch.pos": ("left",  "shoulder_pitch", 0),
-    "kLeftShoulderYaw.pos":   ("left",  "shoulder_yaw",   1),
-    "kLeftShoulderRoll.pos":  ("left",  "shoulder_roll",  2),
-    "kLeftElbow.pos":         ("left",  "elbow_flex",     3),
-    "kLeftWristRoll.pos":     ("left",  "wrist_roll",     4),
-    "kLeftWristYaw.pos":      ("left",  "wrist_yaw",      5),
-    "kLeftWristyaw.pos":      ("left",  "wrist_yaw",      5),  # tolerate casing variant
-    "kLeftWristPitch.pos":    ("left",  "wrist_pitch",    6),
-
-    # RIGHT
-    "kRightShoulderPitch.pos": ("right", "shoulder_pitch", 0),
-    "kRightShoulderYaw.pos":   ("right", "shoulder_yaw",   1),
-    "kRightShoulderRoll.pos":  ("right", "shoulder_roll",  2),
-    "kRightElbow.pos":         ("right", "elbow_flex",     3),
-    "kRightWristRoll.pos":     ("right", "wrist_roll",     4),
-    "kRightWristYaw.pos":      ("right", "wrist_yaw",      5),
-    "kRightWristPitch.pos":    ("right", "wrist_pitch",    6),
-}
-
-# Output
-CALIB_PATH = Path("calibration.json")
-ROUND_TO_INT = False  # set True if you want int ranges
-
-# Init tracker: tracker["left"]["shoulder_pitch"] = {...}
-tracker = {"left": {}, "right": {}}
-for sec, name, idx in MAP.values():
-    if name not in tracker[sec]:
-        tracker[sec][name] = {
-            "id": idx,
-            "drive_mode": 0,
-            "homing_offset": 0,
-            "range_min": math.inf,
-            "range_max": -math.inf,
-        }
-
-def _to_float(x):
-    # unwrap numpy / torch scalars if present
-    if hasattr(x, "item"):
-        try:
-            x = x.item()
-        except Exception:
-            pass
-    return float(x)
-
-def update_tracker(obs: dict):
-    for k, v in obs.items():
-        if k not in MAP:
-            continue
-        sec, name, _ = MAP[k]
-        try:
-            x = _to_float(v)
-        except Exception:
-            continue
-        t = tracker[sec][name]
-        if x < t["range_min"]:
-            t["range_min"] = x
-        if x > t["range_max"]:
-            t["range_max"] = x
-
-def dump_calibration(path: Path):
-    out = {"left": {}, "right": {}}
-    for sec in ("left", "right"):
-        for name, d in tracker[sec].items():
-            mn, mx = d["range_min"], d["range_max"]
-            if ROUND_TO_INT:
-                mn = None if mn is math.inf else int(round(mn))
-                mx = None if mx is -math.inf else int(round(mx))
-            else:
-                mn = None if mn is math.inf else mn
-                mx = None if mx is -math.inf else mx
-            out[sec][name] = {
-                "id": d["id"],
-                "drive_mode": d["drive_mode"],
-                "homing_offset": d["homing_offset"],
-                "range_min": mn,
-                "range_max": mx,
-            }
-    path.write_text(json.dumps(out, indent=4))
-    print(f"Saved calibration to {path.resolve()}")
-
-from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1, G1_29_JointIndex
-from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
-
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-import time
-config = UnitreeG1Config(
-    motion_mode=False,
-    simulation_mode=False
-)
-
-robot = UnitreeG1(config)
-try:
-    while True:
-        observation = robot.get_observation()
-        update_tracker(observation)
-        robot.send_action(observation)  # mirror, if desired
-        time.sleep(0.01)
-except KeyboardInterrupt:
-    dump_calibration(CALIB_PATH)

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.2"
+version = "0.4.3"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -107,6 +107,10 @@ dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
+unitree_g1 = [
+    "pyzmq>=26.2.1,<28.0.0",
+    "onnxruntime>=1.16.0"
+]
 reachy2 = ["reachy2_sdk>=1.0.14,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [

src/lerobot/cameras/utils.py

@@ -43,11 +43,6 @@ def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[s
 
             cameras[key] = Reachy2Camera(cfg)
 
-        elif cfg.type == "zmq":
-            from .zmq import ZMQCamera
-
-            cameras[key] = ZMQCamera(cfg)
-
         else:
             try:
                 cameras[key] = cast(Camera, make_device_from_device_class(cfg))

src/lerobot/cameras/zmq/__init__.py

@@ -1,16 +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 .camera_zmq import ZMQCamera
-from .configuration_zmq import ZMQCameraConfig

src/lerobot/cameras/zmq/camera_zmq.py

@@ -1,623 +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.
-
-"""
-Provides the ZMQCamera class for capturing frames from remote cameras via ZeroMQ.
-"""
-
-import json
-import logging
-import os
-import threading
-import time
-from pathlib import Path
-from threading import Event, Lock, Thread
-from typing import Any
-import base64
-import cv2
-import numpy as np
-import zmq
-from numpy.typing import NDArray
-import base64
-import msgpack
-import msgpack_numpy as m
-from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
-
-from ..camera import Camera
-from ..configs import ColorMode
-from .configuration_zmq import ZMQCameraConfig
-
-logger = logging.getLogger(__name__)
-
-
-class ZMQCamera(Camera):
-    """
-    Manages camera interactions using ZeroMQ for remote frame streaming.
-
-    This class provides a high-level interface to connect to remote cameras
-    that stream JPEG-encoded images over ZeroMQ PUB/SUB sockets. It supports
-    both synchronous and asynchronous frame reading.
-
-    The camera server must be running and publishing JPEG images on the specified
-    address and port. Use the provided utility script to find available ZMQ cameras:
-    ```bash
-    lerobot-find-cameras zmq
-    ```
-
-    Example:
-        ```python
-        from lerobot.cameras.zmq import ZMQCamera
-        from lerobot.cameras.zmq.configuration_zmq import ZMQCameraConfig, ColorMode
-
-        # Basic usage
-        config = ZMQCameraConfig(
-            server_address="192.168.123.164",
-            port=5554,
-            camera_name="remote_cam"
-        )
-        camera = ZMQCamera(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
-        camera.disconnect()
-        ```
-    """
-
-    def __init__(self, config: ZMQCameraConfig):
-        """
-        Initializes the ZMQCamera instance.
-
-        Args:
-            config: The configuration settings for the ZMQ camera.
-        """
-        super().__init__(config)
-
-        self.config = config
-        self.server_address = config.server_address
-        self.port = config.port
-        self.camera_name = config.camera_name
-        self.color_mode = config.color_mode
-        self.timeout_ms = config.timeout_ms
-
-        self.context: zmq.Context | None = None
-        self.socket: zmq.Socket | None = None
-        self._connected = False
-
-        self.thread: Thread | None = None
-        self.stop_event: Event | None = None
-        self.frame_lock: Lock = Lock()
-        self.latest_frame: NDArray[Any] | None = None
-        self.new_frame_event: Event = Event()
-        
-        # Format type detected during connection (msgpack, json, or raw_jpeg)
-        self._format_type: str | None = None
-
-    def __str__(self) -> str:
-        return f"{self.__class__.__name__}({self.camera_name}@{self.server_address}:{self.port})"
-
-    @property
-    def is_connected(self) -> bool:
-        """Checks if the camera is currently connected."""
-        return self._connected and self.context is not None and self.socket is not None
-
-    def connect(self, warmup: bool = True) -> None:
-        """
-        Connects to the ZMQ camera server and configures settings.
-
-        Args:
-            warmup: If True (default), captures a warmup frame before returning.
-
-        Raises:
-            DeviceAlreadyConnectedError: If the camera is already connected.
-            RuntimeError: If connection to the ZMQ server fails.
-        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
-
-        logger.info(f"Connecting to {self}...")
-
-        try:
-            self.context = zmq.Context()
-            self.socket = self.context.socket(zmq.SUB)
-            self.socket.connect(f"tcp://{self.server_address}:{self.port}")
-            self.socket.setsockopt_string(zmq.SUBSCRIBE, "")
-            
-            # Set receive timeout
-            self.socket.setsockopt(zmq.RCVTIMEO, self.timeout_ms)
-
-            self._connected = True
-
-            # Try to receive one frame to validate connection and detect format
-            try:
-                # Try each format until one works
-                test_frame = None
-                for format_type in ["msgpack", "json", "raw_jpeg"]:
-                    try:
-                        test_frame = self.read(format=format_type)
-                        self._format_type = format_type
-                        logger.info(f"{self} detected format: {format_type}")
-                        break
-                    except Exception as e:
-                        logger.debug(f"{self} format '{format_type}' failed: {e}")
-                        continue
-                
-                if test_frame is None:
-                    raise RuntimeError("Failed to decode frame with any supported format (msgpack, json, raw_jpeg)")
-                
-                # Auto-detect resolution if not specified
-                if self.width is None or self.height is None:
-                    h, w = test_frame.shape[:2]
-                    self.height = h
-                    self.width = w
-                    logger.info(f"{self} auto-detected resolution: {w}x{h}")
-
-                logger.info(f"{self} connected successfully.")
-                
-                if warmup:
-                    logger.debug(f"Warming up {self}...")
-                    time.sleep(0.1)  # Brief warmup period
-                    
-            except Exception as e:
-                self._connected = False
-                if self.socket:
-                    self.socket.close()
-                if self.context:
-                    self.context.term()
-                self.socket = None
-                self.context = None
-                raise RuntimeError(f"Failed to receive initial frame from {self}: {e}")
-
-        except Exception as e:
-            self._connected = False
-            if self.socket:
-                self.socket.close()
-            if self.context:
-                self.context.term()
-            self.socket = None
-            self.context = None
-            raise RuntimeError(f"Failed to connect to {self}: {e}")
-
-    @staticmethod
-    def find_cameras(
-        subnet: str | None = None,
-        ports: list[int] | None = None,
-        timeout_ms: int = 200,
-    ) -> list[dict[str, Any]]:
-        """
-        Scans the local network for ZMQ cameras (fast parallel scan).
-
-        Uses threading to scan multiple hosts simultaneously. Without parallelization,
-        scanning 254 hosts would take 6+ minutes. With threads, takes ~10-15 seconds.
-
-        Args:
-            subnet: Network subnet to scan (e.g., "192.168.1.0/24"). If None, auto-detects.
-            ports: List of ports to scan. Defaults to [5554, 5555, 5556].
-            timeout_ms: Connection timeout per host in milliseconds. Default: 200ms.
-
-        Returns:
-            List of dicts containing camera info (address, port, format, resolution).
-            
-        Example:
-            >>> cameras = ZMQCamera.find_cameras()
-            >>> # Or specify: cameras = ZMQCamera.find_cameras(subnet="10.0.0.0/24", ports=[5554])
-        """
-        import socket
-        import ipaddress
-        from concurrent.futures import ThreadPoolExecutor, as_completed
-
-        if ports is None:
-            ports = [5554, 5555, 5556]
-
-        # Auto-detect local subnet
-        if subnet is None:
-            try:
-                s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
-                s.connect(("8.8.8.8", 80))
-                local_ip = s.getsockname()[0]
-                s.close()
-                subnet = ".".join(local_ip.split(".")[:-1]) + ".0/24"
-                logger.info(f"Auto-detected subnet: {subnet}")
-            except Exception as e:
-                logger.error(f"Failed to auto-detect subnet: {e}")
-                return []
-
-        # Parse subnet
-        try:
-            network = ipaddress.ip_network(subnet, strict=False)
-            hosts = list(network.hosts())
-            # Always include localhost (for MuJoCo sim, local servers)
-            hosts.insert(0, ipaddress.IPv4Address("127.0.0.1"))
-        except Exception as e:
-            logger.error(f"Invalid subnet '{subnet}': {e}")
-            return []
-
-        total = len(hosts) * len(ports)
-        logger.info(f"Scanning {len(hosts)} hosts × {len(ports)} ports = {total} targets (this takes ~10-15s)...")
-        
-        def test_target(host_ip: str, port: int) -> dict | None:
-            """Test one host:port for ZMQ camera."""
-            ctx = zmq.Context()
-            sock = ctx.socket(zmq.SUB)
-            sock.connect(f"tcp://{host_ip}:{port}")
-            sock.setsockopt_string(zmq.SUBSCRIBE, "")
-            sock.setsockopt(zmq.RCVTIMEO, timeout_ms)
-            
-            # Wait for subscription to establish (ZMQ "slow joiner" problem)
-            time.sleep(0.1)
-            
-            # Try receiving a few times
-            msg = None
-            for _ in range(3):
-                try:
-                    msg = sock.recv()
-                    break
-                except zmq.Again:
-                    time.sleep(0.05)
-            
-            if msg is None:
-                sock.close()
-                ctx.term()
-                return None
-            
-            # Try formats: msgpack → json → raw_jpeg
-            frame = fmt = None
-            
-            # Msgpack
-            try:
-                d = msgpack.unpackb(msg, object_hook=m.decode)
-                if isinstance(d, dict) and "images" in d and len(d["images"]) > 0:
-                    img = next(iter(d["images"].values()))
-                    if isinstance(img, str):
-                        frame = cv2.imdecode(np.frombuffer(base64.b64decode(img), np.uint8), cv2.IMREAD_COLOR)
-                    elif isinstance(img, np.ndarray):
-                        frame = img
-                    if frame is not None:
-                        fmt = "msgpack"
-            except:
-                pass
-            
-            # JSON
-            if frame is None:
-                try:
-                    d = json.loads(msg.decode('utf-8'))
-                    if isinstance(d, dict):
-                        for v in d.values():
-                            if isinstance(v, str) and len(v) > 100:
-                                try:
-                                    frame = cv2.imdecode(np.frombuffer(base64.b64decode(v), np.uint8), cv2.IMREAD_COLOR)
-                                    if frame is not None:
-                                        fmt = "json"
-                                        break
-                                except:
-                                    pass
-                except:
-                    pass
-            
-            # Raw JPEG
-            if frame is None:
-                try:
-                    frame = cv2.imdecode(np.frombuffer(msg, np.uint8), cv2.IMREAD_COLOR)
-                    if frame is not None:
-                        fmt = "raw_jpeg"
-                except:
-                    pass
-            
-            sock.close()
-            ctx.term()
-            
-            if frame is not None:
-                h, w = frame.shape[:2]
-                return {
-                    "name": f"ZMQ @ {host_ip}:{port}",
-                    "type": "ZMQ",
-                    "id": f"{host_ip}:{port}",
-                    "server_address": host_ip,
-                    "port": port,
-                    "camera_name": f"cam_{host_ip.replace('.', '_')}_{port}",
-                    "format": fmt,
-                    "default_stream_profile": {"width": w, "height": h, "format": fmt.upper()},
-                }
-            return None
-
-        # Parallel scan with thread pool
-        found = []
-        with ThreadPoolExecutor(max_workers=100) as ex:
-            futures = [ex.submit(test_target, str(h), p) for h in hosts for p in ports]
-            for i, fut in enumerate(as_completed(futures), 1):
-                if i % 100 == 0:
-                    logger.info(f"  Progress: {i}/{total} ({100*i//total}%)")
-                res = fut.result()
-                if res:
-                    found.append(res)
-                    logger.info(f"  ✓ {res['server_address']}:{res['port']} ({res['format']})")
-        
-        logger.info(f"Scan complete! Found {len(found)} camera(s).")
-        return found
-
-    def read(self, color_mode: ColorMode | None = None, format: str | None = None) -> NDArray[Any]:
-        """
-        Reads a single frame synchronously from the ZMQ camera.
-
-        Supports three message formats:
-        1. "msgpack": Msgpack with base64 JPEGs: {"timestamps": {...}, "images": {camera_name: "b64"}}
-           (used by MuJoCo sim)
-        2. "json": JSON with base64 JPEGs: {"state": 0.0, "camera_name": "b64jpeg"}
-           (used by LeKiwi-style servers)
-        3. "raw_jpeg": Raw JPEG bytes (used by Unitree G1 head camera)
-        
-        Args:
-            color_mode: Target color mode (RGB or BGR). If None, uses self.color_mode.
-            format: Message format to use. If None, uses auto-detected format from connect().
-                   One of: "msgpack", "json", "raw_jpeg"
-        
-        Returns:
-            np.ndarray: Decoded frame in shape (height, width, 3)
-        
-        Raises:
-            DeviceNotConnectedError: If camera is not connected
-            TimeoutError: If no frame received within timeout_ms
-            RuntimeError: If frame decoding fails
-        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
-        if self.socket is None:
-            raise DeviceNotConnectedError(f"{self} socket is not initialized")
-
-        # Use detected format if not specified
-        if format is None:
-            format = self._format_type
-        
-        if format is None:
-            raise RuntimeError(f"{self} format not specified and not auto-detected during connect()")
-
-        start_time = time.perf_counter()
-
-        try:
-            message = self.socket.recv()
-        except zmq.Again:
-            raise TimeoutError(f"{self} timeout waiting for frame after {self.timeout_ms}ms")
-        except Exception as e:
-            raise RuntimeError(f"{self} read failed: {e}")
-
-        frame = None
-
-        # Decode based on format
-        if format == "msgpack":
-            data = msgpack.unpackb(message, object_hook=m.decode)
-            if not isinstance(data, dict) or "images" not in data:
-                raise RuntimeError(f"{self} invalid msgpack format: expected dict with 'images' key")
-
-            images_dict = data["images"]
-            
-            # Prefer named camera if present
-            if self.camera_name in images_dict:
-                img_data = images_dict[self.camera_name]
-            elif len(images_dict) > 0:
-                # Fallback: first available camera
-                img_data = next(iter(images_dict.values()))
-            else:
-                raise RuntimeError(f"{self} no images found in msgpack message")
-
-            # Decode the image data
-            if isinstance(img_data, str):
-                color_bytes = base64.b64decode(img_data)
-                np_img = np.frombuffer(color_bytes, dtype=np.uint8)
-                frame = cv2.imdecode(np_img, cv2.IMREAD_COLOR)
-            elif isinstance(img_data, np.ndarray):
-                frame = img_data
-            else:
-                raise RuntimeError(f"{self} unknown image payload type: {type(img_data)}")
-        
-        elif format == "json":
-            data = json.loads(message.decode('utf-8'))
-            if not isinstance(data, dict) or self.camera_name not in data:
-                raise RuntimeError(f"{self} invalid JSON format: expected dict with '{self.camera_name}' key")
-            
-            img_b64 = data[self.camera_name]
-            if not isinstance(img_b64, str):
-                raise RuntimeError(f"{self} expected base64 string in JSON, got {type(img_b64)}")
-            
-            color_bytes = base64.b64decode(img_b64)
-            np_img = np.frombuffer(color_bytes, dtype=np.uint8)
-            frame = cv2.imdecode(np_img, cv2.IMREAD_COLOR)
-        
-        elif format == "raw_jpeg":
-            np_img = np.frombuffer(message, dtype=np.uint8)
-            frame = cv2.imdecode(np_img, cv2.IMREAD_COLOR)
-        
-        else:
-            raise ValueError(f"{self} unsupported format: {format}. Use 'msgpack', 'json', or 'raw_jpeg'")
-
-        if frame is None or not isinstance(frame, np.ndarray):
-            raise RuntimeError(f"{self} failed to decode image using format '{format}'")
-
-        processed_frame = self._postprocess_image(frame, color_mode)
-
-        read_duration_ms = (time.perf_counter() - start_time) * 1e3
-        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
-
-        return processed_frame
-
-    def _postprocess_image(self, image: NDArray[Any], color_mode: ColorMode | None = None) -> NDArray[Any]:
-        """
-        Applies color conversion to a raw frame.
-
-        Args:
-            image: The raw image frame (BGR format from cv2.imdecode).
-            color_mode: The target color mode (RGB or BGR). If None, uses self.color_mode.
-
-        Returns:
-            np.ndarray: The processed image frame.
-
-        Raises:
-            ValueError: If the requested color_mode is invalid.
-            RuntimeError: If the frame dimensions don't match expectations.
-        """
-        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 color mode '{requested_color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
-            )
-
-        h, w, c = image.shape
-
-        # Validate dimensions if they were specified
-        if self.height is not None and self.width is not None:
-            if h != self.height or w != self.width:
-                logger.warning(
-                    f"{self} frame dimensions ({w}x{h}) don't match configured ({self.width}x{self.height}). "
-                    "This might be expected if the server sends different resolutions."
-                )
-
-        if c != 3:
-            raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
-
-        processed_image = image
-        if requested_color_mode == ColorMode.RGB:
-            processed_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
-
-        return processed_image
-
-    def _read_loop(self) -> None:
-        """
-        Internal loop run by the background thread for asynchronous reading.
-
-        On each iteration:
-        1. Reads a frame from ZMQ
-        2. Stores result in latest_frame (thread-safe)
-        3. Sets new_frame_event to notify listeners
-
-        Stops on DeviceNotConnectedError, logs other errors and continues.
-        """
-        if self.stop_event is None:
-            raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
-
-        while not self.stop_event.is_set():
-            try:
-                frame = self.read()
-
-                with self.frame_lock:
-                    self.latest_frame = frame
-                self.new_frame_event.set()
-
-            except DeviceNotConnectedError:
-                break
-            except TimeoutError:
-                # Timeout is expected occasionally, just continue
-                logger.debug(f"{self} read timeout in background thread")
-            except Exception as e:
-                logger.warning(f"Error reading frame in background thread for {self}: {e}")
-
-    def _start_read_thread(self) -> None:
-        """Starts or restarts the background read thread if it's not running."""
-        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()
-
-    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)
-
-        self.thread = None
-        self.stop_event = None
-
-    def async_read(self, timeout_ms: float = 10000) -> NDArray[Any]:
-        """
-        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 ZMQ directly, but may wait
-        up to timeout_ms for the background thread to provide a frame.
-
-        Args:
-            timeout_ms: Maximum time in milliseconds to wait for a frame
-                to become available. Defaults to 2000ms.
-
-        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.
-        """
-        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()
-
-        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
-            thread_alive = self.thread is not None and self.thread.is_alive()
-            raise TimeoutError(
-                f"Timed out waiting for frame from {self} after {timeout_ms} ms. "
-                f"Read thread alive: {thread_alive}."
-            )
-
-        with self.frame_lock:
-            frame = self.latest_frame
-            self.new_frame_event.clear()
-
-        if frame is None:
-            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
-
-        return frame
-
-    def disconnect(self) -> None:
-        """
-        Disconnects from the ZMQ camera and cleans up resources.
-
-        Stops the background read thread (if running) and closes the ZMQ socket.
-
-        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.socket is not None:
-            self.socket.close()
-            self.socket = None
-
-        if self.context is not None:
-            self.context.term()
-            self.context = None
-
-        self._connected = False
-
-        logger.info(f"{self} disconnected.")
-

src/lerobot/cameras/zmq/configuration_zmq.py

@@ -1,78 +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 dataclasses import dataclass
-
-from ..configs import CameraConfig, ColorMode
-
-__all__ = ["ZMQCameraConfig", "ColorMode"]
-
-
-@CameraConfig.register_subclass("zmq")
-@dataclass
-class ZMQCameraConfig(CameraConfig):
-    """Configuration class for ZMQ-based remote camera streams.
-
-    This class provides configuration options for cameras accessed through ZeroMQ (ZMQ),
-    supporting remote camera streams over the network. The server must be running and
-    streaming JPEG-encoded images over a ZMQ PUB socket.
-
-    Example configurations:
-    ```python
-    # Basic configuration
-    ZMQCameraConfig(
-        server_address="192.168.123.164",
-        port=5554,
-        camera_name="remote_cam_1"
-    )
-
-    # With custom resolution
-    ZMQCameraConfig(
-        server_address="10.0.0.100",
-        port=5555,
-        camera_name="lab_cam",
-        width=1280,
-        height=480,
-        fps=30
-    )
-    ```
-
-    Attributes:
-        server_address: IP address or hostname of the ZMQ image server.
-        port: Port number where the ZMQ server is publishing images.
-        camera_name: Identifier name for this camera (for logging/debugging).
-        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
-        timeout_ms: Timeout in milliseconds for receiving frames. Defaults to 1000ms.
-    """
-
-    server_address: str
-    port: int = 5554
-    camera_name: str = "zmq_camera"
-    color_mode: ColorMode = ColorMode.RGB
-    timeout_ms: int = 5000
-
-    def __post_init__(self) -> None:
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
-            )
-
-        if self.timeout_ms <= 0:
-            raise ValueError(f"`timeout_ms` must be positive, but {self.timeout_ms} is provided.")
-
-        if not self.server_address:
-            raise ValueError("`server_address` cannot be empty.")
-
-        if self.port <= 0 or self.port > 65535:
-            raise ValueError(f"`port` must be between 1 and 65535, but {self.port} is provided.")

src/lerobot/datasets/image_writer.py

@@ -110,8 +110,8 @@ def worker_thread_loop(queue: queue.Queue):
         if item is None:
             queue.task_done()
             break
-        image_array, fpath = item
-        write_image(image_array, fpath)
+        image_array, fpath, compress_level = item
+        write_image(image_array, fpath, compress_level)
         queue.task_done()
 
 
@@ -169,11 +169,13 @@ class AsyncImageWriter:
                 p.start()
                 self.processes.append(p)
 
-    def save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path):
+    def save_image(
+        self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1
+    ):
         if isinstance(image, torch.Tensor):
             # Convert tensor to numpy array to minimize main process time
             image = image.cpu().numpy()
-        self.queue.put((image, fpath))
+        self.queue.put((image, fpath, compress_level))
 
     def wait_until_done(self):
         self.queue.join()

src/lerobot/datasets/lerobot_dataset.py

@@ -13,6 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import concurrent.futures
 import contextlib
 import logging
 import shutil
@@ -539,6 +540,15 @@ class LeRobotDatasetMetadata:
         return obj
 
 
+def _encode_video_worker(video_key: str, episode_index: int, root: Path, fps: int) -> Path:
+    temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
+    fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
+    img_dir = (root / fpath).parent
+    encode_video_frames(img_dir, temp_path, fps, overwrite=True)
+    shutil.rmtree(img_dir)
+    return temp_path
+
+
 class LeRobotDataset(torch.utils.data.Dataset):
     def __init__(
         self,
@@ -1071,6 +1081,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             ep_buffer[key] = current_ep_idx if key == "episode_index" else []
         return ep_buffer
 
+    # TODO(Steven): consider move this to utils
     def _get_image_file_path(self, episode_index: int, image_key: str, frame_index: int) -> Path:
         fpath = DEFAULT_IMAGE_PATH.format(
             image_key=image_key, episode_index=episode_index, frame_index=frame_index
@@ -1080,13 +1091,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
     def _get_image_file_dir(self, episode_index: int, image_key: str) -> Path:
         return self._get_image_file_path(episode_index, image_key, frame_index=0).parent
 
-    def _save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path) -> None:
+    def _save_image(
+        self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1
+    ) -> None:
         if self.image_writer is None:
             if isinstance(image, torch.Tensor):
                 image = image.cpu().numpy()
-            write_image(image, fpath)
+            write_image(image, fpath, compress_level=compress_level)
         else:
-            self.image_writer.save_image(image=image, fpath=fpath)
+            self.image_writer.save_image(image=image, fpath=fpath, compress_level=compress_level)
 
     def add_frame(self, frame: dict) -> None:
         """
@@ -1124,14 +1137,19 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 )
                 if frame_index == 0:
                     img_path.parent.mkdir(parents=True, exist_ok=True)
-                self._save_image(frame[key], img_path)
+                compress_level = 1 if self.features[key]["dtype"] == "video" else 6
+                self._save_image(frame[key], img_path, compress_level)
                 self.episode_buffer[key].append(str(img_path))
             else:
                 self.episode_buffer[key].append(frame[key])
 
         self.episode_buffer["size"] += 1
 
-    def save_episode(self, episode_data: dict | None = None) -> None:
+    def save_episode(
+        self,
+        episode_data: dict | None = None,
+        parallel_encoding: bool = True,
+    ) -> None:
         """
         This will save to disk the current episode in self.episode_buffer.
 
@@ -1143,6 +1161,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
             episode_data (dict | None, optional): Dict containing the episode data to save. If None, this will
                 save the current episode in self.episode_buffer, which is filled with 'add_frame'. Defaults to
                 None.
+            parallel_encoding (bool, optional): If True, encode videos in parallel using ProcessPoolExecutor.
+                Defaults to True on Linux, False on macOS as it tends to use all the CPU available already.
         """
         episode_buffer = episode_data if episode_data is not None else self.episode_buffer
 
@@ -1179,8 +1199,40 @@ class LeRobotDataset(torch.utils.data.Dataset):
         use_batched_encoding = self.batch_encoding_size > 1
 
         if has_video_keys and not use_batched_encoding:
-            for video_key in self.meta.video_keys:
-                ep_metadata.update(self._save_episode_video(video_key, episode_index))
+            num_cameras = len(self.meta.video_keys)
+            if parallel_encoding and num_cameras > 1:
+                # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
+                # num_cameras * num_threads = (total_cpu -1)
+                with concurrent.futures.ProcessPoolExecutor(max_workers=num_cameras) as executor:
+                    future_to_key = {
+                        executor.submit(
+                            _encode_video_worker,
+                            video_key,
+                            episode_index,
+                            self.root,
+                            self.fps,
+                        ): video_key
+                        for video_key in self.meta.video_keys
+                    }
+
+                    results = {}
+                    for future in concurrent.futures.as_completed(future_to_key):
+                        video_key = future_to_key[future]
+                        try:
+                            temp_path = future.result()
+                            results[video_key] = temp_path
+                        except Exception as exc:
+                            logging.error(f"Video encoding failed for {video_key}: {exc}")
+                            raise exc
+
+                for video_key in self.meta.video_keys:
+                    temp_path = results[video_key]
+                    ep_metadata.update(
+                        self._save_episode_video(video_key, episode_index, temp_path=temp_path)
+                    )
+            else:
+                for video_key in self.meta.video_keys:
+                    ep_metadata.update(self._save_episode_video(video_key, episode_index))
 
         # `meta.save_episode` need to be executed after encoding the videos
         self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats, ep_metadata)
@@ -1345,9 +1397,18 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         return metadata
 
-    def _save_episode_video(self, video_key: str, episode_index: int) -> dict:
+    def _save_episode_video(
+        self,
+        video_key: str,
+        episode_index: int,
+        temp_path: Path | None = None,
+    ) -> dict:
         # Encode episode frames into a temporary video
-        ep_path = self._encode_temporary_episode_video(video_key, episode_index)
+        if temp_path is None:
+            ep_path = self._encode_temporary_episode_video(video_key, episode_index)
+        else:
+            ep_path = temp_path
+
         ep_size_in_mb = get_file_size_in_mb(ep_path)
         ep_duration_in_s = get_video_duration_in_s(ep_path)
 
@@ -1465,11 +1526,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        temp_path = Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4"
-        img_dir = self._get_image_file_dir(episode_index, video_key)
-        encode_video_frames(img_dir, temp_path, self.fps, overwrite=True)
-        shutil.rmtree(img_dir)
-        return temp_path
+        return _encode_video_worker(video_key, episode_index, self.root, self.fps)
 
     @classmethod
     def create(

src/lerobot/datasets/utils.py

@@ -49,7 +49,7 @@ from lerobot.utils.utils import SuppressProgressBars, is_valid_numpy_dtype_strin
 
 DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
 DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
-DEFAULT_VIDEO_FILE_SIZE_IN_MB = 500  # Max size per file
+DEFAULT_VIDEO_FILE_SIZE_IN_MB = 200  # Max size per file
 
 INFO_PATH = "meta/info.json"
 STATS_PATH = "meta/stats.json"

src/lerobot/datasets/video_utils.py

@@ -311,6 +311,7 @@ def encode_video_frames(
     fast_decode: int = 0,
     log_level: int | None = av.logging.ERROR,
     overwrite: bool = False,
+    preset: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
     # Check encoder availability
@@ -359,6 +360,9 @@ def encode_video_frames(
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
         video_options[key] = value
 
+    if vcodec == "libsvtav1":
+        video_options["preset"] = str(preset) if preset is not None else "12"
+
     # Set logging level
     if log_level is not None:
         # "While less efficient, it is generally preferable to modify logging with Python's logging"

src/lerobot/envs/factory.py

@@ -111,6 +111,7 @@ def make_env(
 
         # import and surface clear import errors
         module = _import_hub_module(local_file, repo_id)
+
         # call the hub-provided make_env
         raw_result = _call_make_env(module, n_envs=n_envs, use_async_envs=use_async_envs)
 

src/lerobot/envs/utils.py

@@ -221,22 +221,7 @@ def _load_module_from_path(path: str, module_name: str | None = None):
     if spec is None:
         raise ImportError(f"Could not load module spec for {module_name} from {path}")
     module = importlib.util.module_from_spec(spec)
-    
-    # Add the module's directory to sys.path so it can import local modules
-    import sys
-    module_dir = os.path.dirname(os.path.abspath(path))
-    sys_path_modified = False
-    if module_dir not in sys.path:
-        sys.path.insert(0, module_dir)
-        sys_path_modified = True
-    
-    try:
-        spec.loader.exec_module(module)  # type: ignore
-    finally:
-        # Clean up sys.path after import
-        if sys_path_modified:
-            sys.path.remove(module_dir)
-    
+    spec.loader.exec_module(module)  # type: ignore
     return module
 
 

src/lerobot/policies/pi05/modeling_pi05.py

@@ -538,6 +538,8 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         if config.compile_model:
             torch.set_float32_matmul_precision("high")
             self.sample_actions = torch.compile(self.sample_actions, mode=config.compile_mode)
+            # Also compile the main forward pass used during training
+            self.forward = torch.compile(self.forward, mode=config.compile_mode)
 
         msg = """An incorrect transformer version is used, please create an issue on https://github.com/huggingface/lerobot/issues"""
 

src/lerobot/rl/actor.py

@@ -78,7 +78,7 @@ from lerobot.transport.utils import (
     transitions_to_bytes,
 )
 from lerobot.utils.random_utils import set_seed
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.transition import (
     Transition,
     move_state_dict_to_device,
@@ -398,7 +398,7 @@ def act_with_policy(
 
         if cfg.env.fps is not None:
             dt_time = time.perf_counter() - start_time
-            busy_wait(1 / cfg.env.fps - dt_time)
+            precise_sleep(1 / cfg.env.fps - dt_time)
 
 
 #  Communication Functions - Group all gRPC/messaging functions

src/lerobot/rl/gym_manipulator.py

@@ -74,7 +74,7 @@ from lerobot.teleoperators import (
 from lerobot.teleoperators.teleoperator import Teleoperator
 from lerobot.teleoperators.utils import TeleopEvents
 from lerobot.utils.constants import ACTION, DONE, OBS_IMAGES, OBS_STATE, REWARD
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import log_say
 
 logging.basicConfig(level=logging.INFO)
@@ -114,7 +114,7 @@ def reset_follower_position(robot_arm: Robot, target_position: np.ndarray) -> No
     for pose in trajectory:
         action_dict = dict(zip(current_position_dict, pose, strict=False))
         robot_arm.bus.sync_write("Goal_Position", action_dict)
-        busy_wait(0.015)
+        precise_sleep(0.015)
 
 
 class RobotEnv(gym.Env):
@@ -238,7 +238,7 @@ class RobotEnv(gym.Env):
             reset_follower_position(self.robot, np.array(self.reset_pose))
             log_say("Reset the environment done.", play_sounds=True)
 
-        busy_wait(self.reset_time_s - (time.perf_counter() - start_time))
+        precise_sleep(self.reset_time_s - (time.perf_counter() - start_time))
 
         super().reset(seed=seed, options=options)
 
@@ -713,7 +713,7 @@ def control_loop(
             transition = env_processor(transition)
 
         # Maintain fps timing
-        busy_wait(dt - (time.perf_counter() - step_start_time))
+        precise_sleep(dt - (time.perf_counter() - step_start_time))
 
     if dataset is not None and cfg.dataset.push_to_hub:
         logging.info("Pushing dataset to hub")
@@ -745,7 +745,7 @@ def replay_trajectory(
         )
         transition = action_processor(transition)
         env.step(transition[TransitionKey.ACTION])
-        busy_wait(1 / cfg.env.fps - (time.perf_counter() - start_time))
+        precise_sleep(1 / cfg.env.fps - (time.perf_counter() - start_time))
 
 
 @parser.wrap()

src/lerobot/robots/unitree_g1/arm_calibration.json

@@ -1,108 +0,0 @@
-{
-    "kLeftShoulderPitch.pos": {
-        "id": 0,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -3,
-        "range_max": 1
-    },
-    "kLeftShoulderYaw.pos": {
-        "id": 1,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -2.6,
-        "range_max": 2.6
-    },
-    "kLeftShoulderRoll.pos": {
-        "id": 2,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -0.1,
-        "range_max": 2.2
-    },
-    "kLeftElbow.pos": {
-        "id": 3,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -1,
-        "range_max": 1
-    },
-    "kLeftWristRoll.pos": {
-        "id": 4,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -1.9,
-        "range_max": 1.9
-    },
-    "kLeftWristYaw.pos": {
-        "id": 5,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": 0.0,
-        "range_max": 0.0
-    },
-    "kLeftWristyaw.pos": {
-        "id": 5,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": 0.0,
-        "range_max": 0.0
-    },
-    "kLeftWristPitch.pos": {
-        "id": 6,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": 0.0,
-        "range_max": 0.0
-    },
-
-    "kRightShoulderPitch.pos": {
-        "id": 0,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -3.0,
-        "range_max": 1
-    },
-    "kRightShoulderYaw.pos": {
-        "id": 1,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -2.6,
-        "range_max": 2.6
-    },
-    "kRightShoulderRoll.pos": {
-        "id": 2,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -2.2,
-        "range_max": 0.5
-    },
-    "kRightElbow.pos": {
-        "id": 3,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -1,
-        "range_max": 1
-    },
-    "kRightWristRoll.pos": {
-        "id": 4,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": -1.9,
-        "range_max": 1.9
-    },
-    "kRightWristYaw.pos": {
-        "id": 5,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": 0.0,
-        "range_max": 0.0
-    },
-    "kRightWristPitch.pos": {
-        "id": 6,
-        "drive_mode": 0,
-        "homing_offset": 0,
-        "range_min": 0.0,
-        "range_max": 0.0
-    }
-}

src/lerobot/robots/unitree_g1/assets/g1/.gitignore

@@ -1,2 +0,0 @@
-*.gv
-*.pdf

src/lerobot/robots/unitree_g1/assets/g1/README.md

@@ -1,33 +0,0 @@
-# Unitree G1 Description (URDF & MJCF)
-
-## Overview
-
-This package includes a universal humanoid robot description (URDF & MJCF) for the [Unitree G1](https://www.unitree.com/g1/), developed by [Unitree Robotics](https://www.unitree.com/).
-
-MJCF/URDF for the G1 robot:
-
-| MJCF/URDF file name           | `mode_machine` | Hip roll reduction ratio | Update status | dof#leg | dof#waist | dof#arm | dof#hand |
-| ----------------------------- | :------------: | :----------------------: | ------------- | :-----: | :-------: | :-----: | :------: |
-| `g1_23dof`                    |       1        |           14.5           | Beta          |   6*2   |     1     |   5*2   |    0     |
-| `g1_29dof`                    |       2        |           14.5           | Beta          |   6*2   |     3     |   7*2   |    0     |
-| `g1_29dof_with_hand`          |       2        |           14.5           | Beta          |   6*2   |     3     |   7*2   |   7*2    |
-| `g1_29dof_lock_waist`         |       3        |           14.5           | Beta          |   6*2   |     1     |   7*2   |    0     |
-| `g1_23dof_rev_1_0`            |       4        |           22.5           | Up-to-date    |   6*2   |     1     |   5*2   |    0     |
-| `g1_29dof_rev_1_0`            |       5        |           22.5           | Up-to-date    |   6*2   |     3     |   7*2   |    0     |
-| `g1_29dof_with_hand_rev_1_0`  |       5        |           22.5           | Up-to-date    |   6*2   |     3     |   7*2   |   7*2    |
-| `g1_29dof_lock_waist_rev_1_0` |       6        |           22.5           | Up-to-date    |   6*2   |     1     |   7*2   |    0     |
-| `g1_dual_arm`                 |       9        |           null           | Up-to-date    |    0    |     0     |   7*2   |    0     |
-
-## Visulization with [MuJoCo](https://github.com/google-deepmind/mujoco)
-
-1. Open MuJoCo Viewer
-
-   ```bash
-   pip install mujoco
-   python -m mujoco.viewer
-   ```
-
-2. Drag and drop the MJCF/URDF model file (`g1_XXX.xml`/`g1_XXX.urdf`) to the MuJoCo Viewer.
-
-## Note for teleoperate
-g1_body29_hand14 is modified from [g1_29dof_with_hand_rev_1_0](https://github.com/unitreerobotics/unitree_ros/blob/master/robots/g1_description/g1_29dof_with_hand_rev_1_0.urdf)

src/lerobot/robots/unitree_g1/assets/g1/g1_body23.urdf

@@ -1,903 +0,0 @@
-<robot name="g1_23dof">
-  <mujoco>
-    <compiler meshdir="meshes" discardvisual="false"/>
-  </mujoco>
-
-  <!-- [CAUTION] uncomment when convert to mujoco -->
-  <!-- <link name="world"></link>
-  <joint name="floating_base_joint" type="floating">
-    <parent link="world"/>
-    <child link="pelvis"/>
-  </joint> -->
-
-  <link name="pelvis">
-    <inertial>
-      <origin xyz="0 0 -0.07605" rpy="0 0 0"/>
-      <mass value="3.813"/>
-      <inertia ixx="0.010549" ixy="0" ixz="2.1E-06" iyy="0.0093089" iyz="0" izz="0.0079184"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/pelvis.STL"/>
-      </geometry>
-      <material name="dark">
-        <color rgba="0.2 0.2 0.2 1"/>
-      </material>
-    </visual>
-  </link>
-  <link name="pelvis_contour_link">
-    <inertial>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <mass value="0.001"/>
-      <inertia ixx="1e-7" ixy="0" ixz="0" iyy="1e-7" iyz="0" izz="1e-7"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/pelvis_contour_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/pelvis_contour_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="pelvis_contour_joint" type="fixed">
-    <parent link="pelvis"/>
-    <child link="pelvis_contour_link"/>
-  </joint>
-
-  <!-- Legs -->
-  <link name="left_hip_pitch_link">
-    <inertial>
-      <origin xyz="0.002741 0.047791 -0.02606" rpy="0 0 0"/>
-      <mass value="1.35"/>
-      <inertia ixx="0.001811" ixy="3.68E-05" ixz="-3.44E-05" iyy="0.0014193" iyz="0.000171" izz="0.0012812"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_hip_pitch_link.STL"/>
-      </geometry>
-      <material name="dark">
-        <color rgba="0.2 0.2 0.2 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_hip_pitch_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_hip_pitch_joint" type="revolute">
-    <origin xyz="0 0.064452 -0.1027" rpy="0 0 0"/>
-    <parent link="pelvis"/>
-    <child link="left_hip_pitch_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-2.5307" upper="2.8798" effort="88" velocity="32"/>
-  </joint>
-  <link name="left_hip_roll_link">
-    <inertial>
-      <origin xyz="0.029812 -0.001045 -0.087934" rpy="0 0 0"/>
-      <mass value="1.52"/>
-      <inertia ixx="0.0023773" ixy="-3.8E-06" ixz="-0.0003908" iyy="0.0024123" iyz="1.84E-05" izz="0.0016595"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_hip_roll_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_hip_roll_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_hip_roll_joint" type="revolute">
-    <origin xyz="0 0.052 -0.030465" rpy="0 -0.1749 0"/>
-    <parent link="left_hip_pitch_link"/>
-    <child link="left_hip_roll_link"/>
-    <axis xyz="1 0 0"/>
-    <limit lower="-0.5236" upper="2.9671" effort="88" velocity="32"/>
-  </joint>
-  <link name="left_hip_yaw_link">
-    <inertial>
-      <origin xyz="-0.057709 -0.010981 -0.15078" rpy="0 0 0"/>
-      <mass value="1.702"/>
-      <inertia ixx="0.0057774" ixy="-0.0005411" ixz="-0.0023948" iyy="0.0076124" iyz="-0.0007072" izz="0.003149"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_hip_yaw_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_hip_yaw_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_hip_yaw_joint" type="revolute">
-    <origin xyz="0.025001 0 -0.12412" rpy="0 0 0"/>
-    <parent link="left_hip_roll_link"/>
-    <child link="left_hip_yaw_link"/>
-    <axis xyz="0 0 1"/>
-    <limit lower="-2.7576" upper="2.7576" effort="88" velocity="32"/>
-  </joint>
-  <link name="left_knee_link">
-    <inertial>
-      <origin xyz="0.005457 0.003964 -0.12074" rpy="0 0 0"/>
-      <mass value="1.932"/>
-      <inertia ixx="0.011329" ixy="4.82E-05" ixz="-4.49E-05" iyy="0.011277" iyz="-0.0007146" izz="0.0015168"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_knee_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_knee_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_knee_joint" type="revolute">
-    <origin xyz="-0.078273 0.0021489 -0.17734" rpy="0 0.1749 0"/>
-    <parent link="left_hip_yaw_link"/>
-    <child link="left_knee_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-0.087267" upper="2.8798" effort="139" velocity="20"/>
-  </joint>
-  <link name="left_ankle_pitch_link">
-    <inertial>
-      <origin xyz="-0.007269 0 0.011137" rpy="0 0 0"/>
-      <mass value="0.074"/>
-      <inertia ixx="8.4E-06" ixy="0" ixz="-2.9E-06" iyy="1.89E-05" iyz="0" izz="1.26E-05"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_ankle_pitch_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_ankle_pitch_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_ankle_pitch_joint" type="revolute">
-    <origin xyz="0 -9.4445E-05 -0.30001" rpy="0 0 0"/>
-    <parent link="left_knee_link"/>
-    <child link="left_ankle_pitch_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-0.87267" upper="0.5236" effort="50" velocity="37"/>
-  </joint>
-  <link name="left_ankle_roll_link">
-    <inertial>
-      <origin xyz="0.026505 0 -0.016425" rpy="0 0 0"/>
-      <mass value="0.608"/>
-      <inertia ixx="0.0002231" ixy="2E-07" ixz="8.91E-05" iyy="0.0016161" iyz="-1E-07" izz="0.0016667"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_ankle_roll_link.STL"/>
-      </geometry>
-      <material name="dark">
-        <color rgba="0.2 0.2 0.2 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="-0.05 0.025 -0.03" rpy="0 0 0"/>
-      <geometry>
-        <sphere radius="0.005"/>
-      </geometry>
-    </collision>
-    <collision>
-      <origin xyz="-0.05 -0.025 -0.03" rpy="0 0 0"/>
-      <geometry>
-        <sphere radius="0.005"/>
-      </geometry>
-    </collision>
-    <collision>
-      <origin xyz="0.12 0.03 -0.03" rpy="0 0 0"/>
-      <geometry>
-        <sphere radius="0.005"/>
-      </geometry>
-    </collision>
-    <collision>
-      <origin xyz="0.12 -0.03 -0.03" rpy="0 0 0"/>
-      <geometry>
-        <sphere radius="0.005"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_ankle_roll_joint" type="revolute">
-    <origin xyz="0 0 -0.017558" rpy="0 0 0"/>
-    <parent link="left_ankle_pitch_link"/>
-    <child link="left_ankle_roll_link"/>
-    <axis xyz="1 0 0"/>
-    <limit lower="-0.2618" upper="0.2618" effort="50" velocity="37"/>
-  </joint>
-  <link name="right_hip_pitch_link">
-    <inertial>
-      <origin xyz="0.002741 -0.047791 -0.02606" rpy="0 0 0"/>
-      <mass value="1.35"/>
-      <inertia ixx="0.001811" ixy="-3.68E-05" ixz="-3.44E-05" iyy="0.0014193" iyz="-0.000171" izz="0.0012812"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_hip_pitch_link.STL"/>
-      </geometry>
-      <material name="dark">
-        <color rgba="0.2 0.2 0.2 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_hip_pitch_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_hip_pitch_joint" type="revolute">
-    <origin xyz="0 -0.064452 -0.1027" rpy="0 0 0"/>
-    <parent link="pelvis"/>
-    <child link="right_hip_pitch_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-2.5307" upper="2.8798" effort="88" velocity="32"/>
-  </joint>
-  <link name="right_hip_roll_link">
-    <inertial>
-      <origin xyz="0.029812 0.001045 -0.087934" rpy="0 0 0"/>
-      <mass value="1.52"/>
-      <inertia ixx="0.0023773" ixy="3.8E-06" ixz="-0.0003908" iyy="0.0024123" iyz="-1.84E-05" izz="0.0016595"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_hip_roll_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_hip_roll_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_hip_roll_joint" type="revolute">
-    <origin xyz="0 -0.052 -0.030465" rpy="0 -0.1749 0"/>
-    <parent link="right_hip_pitch_link"/>
-    <child link="right_hip_roll_link"/>
-    <axis xyz="1 0 0"/>
-    <limit lower="-2.9671" upper="0.5236" effort="88" velocity="32"/>
-  </joint>
-  <link name="right_hip_yaw_link">
-    <inertial>
-      <origin xyz="-0.057709 0.010981 -0.15078" rpy="0 0 0"/>
-      <mass value="1.702"/>
-      <inertia ixx="0.0057774" ixy="0.0005411" ixz="-0.0023948" iyy="0.0076124" iyz="0.0007072" izz="0.003149"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_hip_yaw_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_hip_yaw_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_hip_yaw_joint" type="revolute">
-    <origin xyz="0.025001 0 -0.12412" rpy="0 0 0"/>
-    <parent link="right_hip_roll_link"/>
-    <child link="right_hip_yaw_link"/>
-    <axis xyz="0 0 1"/>
-    <limit lower="-2.7576" upper="2.7576" effort="88" velocity="32"/>
-  </joint>
-  <link name="right_knee_link">
-    <inertial>
-      <origin xyz="0.005457 -0.003964 -0.12074" rpy="0 0 0"/>
-      <mass value="1.932"/>
-      <inertia ixx="0.011329" ixy="-4.82E-05" ixz="4.49E-05" iyy="0.011277" iyz="0.0007146" izz="0.0015168"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_knee_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_knee_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_knee_joint" type="revolute">
-    <origin xyz="-0.078273 -0.0021489 -0.17734" rpy="0 0.1749 0"/>
-    <parent link="right_hip_yaw_link"/>
-    <child link="right_knee_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-0.087267" upper="2.8798" effort="139" velocity="20"/>
-  </joint>
-  <link name="right_ankle_pitch_link">
-    <inertial>
-      <origin xyz="-0.007269 0 0.011137" rpy="0 0 0"/>
-      <mass value="0.074"/>
-      <inertia ixx="8.4E-06" ixy="0" ixz="-2.9E-06" iyy="1.89E-05" iyz="0" izz="1.26E-05"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_ankle_pitch_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_ankle_pitch_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_ankle_pitch_joint" type="revolute">
-    <origin xyz="0 9.4445E-05 -0.30001" rpy="0 0 0"/>
-    <parent link="right_knee_link"/>
-    <child link="right_ankle_pitch_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-0.87267" upper="0.5236" effort="50" velocity="37"/>
-  </joint>
-  <link name="right_ankle_roll_link">
-    <inertial>
-      <origin xyz="0.026505 0 -0.016425" rpy="0 0 0"/>
-      <mass value="0.608"/>
-      <inertia ixx="0.0002231" ixy="-2E-07" ixz="8.91E-05" iyy="0.0016161" iyz="1E-07" izz="0.0016667"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_ankle_roll_link.STL"/>
-      </geometry>
-      <material name="dark">
-        <color rgba="0.2 0.2 0.2 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="-0.05 0.025 -0.03" rpy="0 0 0"/>
-      <geometry>
-        <sphere radius="0.005"/>
-      </geometry>
-    </collision>
-    <collision>
-      <origin xyz="-0.05 -0.025 -0.03" rpy="0 0 0"/>
-      <geometry>
-        <sphere radius="0.005"/>
-      </geometry>
-    </collision>
-    <collision>
-      <origin xyz="0.12 0.03 -0.03" rpy="0 0 0"/>
-      <geometry>
-        <sphere radius="0.005"/>
-      </geometry>
-    </collision>
-    <collision>
-      <origin xyz="0.12 -0.03 -0.03" rpy="0 0 0"/>
-      <geometry>
-        <sphere radius="0.005"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_ankle_roll_joint" type="revolute">
-    <origin xyz="0 0 -0.017558" rpy="0 0 0"/>
-    <parent link="right_ankle_pitch_link"/>
-    <child link="right_ankle_roll_link"/>
-    <axis xyz="1 0 0"/>
-    <limit lower="-0.2618" upper="0.2618" effort="50" velocity="37"/>
-  </joint>
-
-  <!-- Torso -->
-  <link name="waist_yaw_fixed_link">
-    <inertial>
-      <origin xyz="0.003964 0 0.018769" rpy="0 0 0"/>
-      <mass value="0.244"/>
-      <inertia ixx="9.9587E-05" ixy="-1.833E-06" ixz="-1.2617E-05" iyy="0.00012411" iyz="-1.18E-07" izz="0.00015586"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/waist_yaw_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-  </link>
-  <joint name="waist_yaw_fixed_joint" type="fixed">
-    <origin xyz="0.0039635 0 -0.054" rpy="0 0 0"/>
-    <parent link="torso_link"/>
-    <child link="waist_yaw_fixed_link"/>
-  </joint>
-  <joint name="waist_yaw_joint" type="revolute">
-    <origin xyz="-0.0039635 0 0.054" rpy="0 0 0"/>
-    <parent link="pelvis"/>
-    <child link="torso_link"/>
-    <axis xyz="0 0 1"/>
-    <limit lower="-2.618" upper="2.618" effort="88" velocity="32"/>
-  </joint>
-  <link name="torso_link">
-    <inertial>
-      <origin xyz="0.002601 0.000257 0.153719" rpy="0 0 0"/>
-      <mass value="8.562"/>
-      <inertia ixx="0.065674966" ixy="-8.597E-05" ixz="-0.001737252" iyy="0.053535188" iyz="8.6899E-05" izz="0.030808125"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/torso_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/torso_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-
-  <!-- LOGO -->
-  <joint name="logo_joint" type="fixed">
-    <origin xyz="0.0039635 0 -0.054" rpy="0 0 0"/>
-    <parent link="torso_link"/>
-    <child link="logo_link"/>
-  </joint>
-  <link name="logo_link">
-    <inertial>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <mass value="0.001"/>
-      <inertia ixx="1e-7" ixy="0" ixz="0" iyy="1e-7" iyz="0" izz="1e-7"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/logo_link.STL"/>
-      </geometry>
-      <material name="dark">
-        <color rgba="0.2 0.2 0.2 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/logo_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-
-  <!-- Head -->
-  <link name="head_link">
-    <inertial>
-      <origin xyz="0.005267 0.000299 0.449869" rpy="0 0 0"/>
-      <mass value="1.036"/>
-      <inertia ixx="0.004085051" ixy="-2.543E-06" ixz="-6.9455E-05" iyy="0.004185212" iyz="-3.726E-06" izz="0.001807911"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/head_link.STL"/>
-      </geometry>
-      <material name="dark">
-        <color rgba="0.2 0.2 0.2 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/head_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="head_joint" type="fixed">
-    <origin xyz="0.0039635 0 -0.054" rpy="0 0 0"/>
-    <parent link="torso_link"/>
-    <child link="head_link"/>
-  </joint>
-
-  <!-- Waist Support -->
-  <link name="waist_support_link">
-    <inertial>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <mass value="0.001"/>
-      <inertia ixx="1e-7" ixy="0" ixz="0" iyy="1e-7" iyz="0" izz="1e-7"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/waist_support_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/waist_support_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="waist_support_joint" type="fixed">
-    <origin xyz="0.0039635 0 -0.054" rpy="0 0 0"/>
-    <parent link="torso_link"/>
-    <child link="waist_support_link"/>
-  </joint>
-
-  <!-- IMU -->
-  <link name="imu_in_torso"></link>
-  <joint name="imu_in_torso_joint" type="fixed">
-    <origin xyz="-0.03959 -0.00224 0.13792" rpy="0 0 0"/>
-    <parent link="torso_link"/>
-    <child link="imu_in_torso"/>
-  </joint>
-
-  <link name="imu_in_pelvis"></link>
-  <joint name="imu_in_pelvis_joint" type="fixed">
-    <origin xyz="0.04525 0 -0.08339" rpy="0 0 0"/>
-    <parent link="pelvis"/>
-    <child link="imu_in_pelvis"/>
-  </joint>
-
-  <!-- d435 -->
-  <link name="d435_link"></link>
-  <joint name="d435_joint" type="fixed">
-    <origin xyz="0.0576235 0.01753 0.41987" rpy="0 0.8307767239493009 0"/>
-    <parent link="torso_link"/>
-    <child link="d435_link"/>
-  </joint>
-
-  <!-- mid360 -->
-  <link name="mid360_link"></link>
-  <joint name="mid360_joint" type="fixed">
-    <origin xyz="0.0002835 0.00003 0.40618" rpy="0 0.04014257279586953 0"/>
-    <parent link="torso_link"/>
-    <child link="mid360_link"/>
-  </joint>
-
-  <!-- Arm -->
-  <link name="left_shoulder_pitch_link">
-    <inertial>
-      <origin xyz="0 0.035892 -0.011628" rpy="0 0 0"/>
-      <mass value="0.718"/>
-      <inertia ixx="0.0004291" ixy="-9.2E-06" ixz="6.4E-06" iyy="0.000453" iyz="2.26E-05" izz="0.000423"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_shoulder_pitch_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0.04 -0.01" rpy="0 1.5707963267948966 0"/>
-      <geometry>
-        <cylinder radius="0.03" length="0.05"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_shoulder_pitch_joint" type="revolute">
-    <origin xyz="0.0039563 0.10022 0.23778" rpy="0.27931 5.4949E-05 -0.00019159"/>
-    <parent link="torso_link"/>
-    <child link="left_shoulder_pitch_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-3.0892" upper="2.6704" effort="25" velocity="37"/>
-  </joint>
-  <link name="left_shoulder_roll_link">
-    <inertial>
-      <origin xyz="-0.000227 0.00727 -0.063243" rpy="0 0 0"/>
-      <mass value="0.643"/>
-      <inertia ixx="0.0006177" ixy="-1E-06" ixz="8.7E-06" iyy="0.0006912" iyz="-5.3E-06" izz="0.0003894"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_shoulder_roll_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="-0.004 0.006 -0.053" rpy="0 0 0"/>
-      <geometry>
-        <cylinder radius="0.03" length="0.03"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_shoulder_roll_joint" type="revolute">
-    <origin xyz="0 0.038 -0.013831" rpy="-0.27925 0 0"/>
-    <parent link="left_shoulder_pitch_link"/>
-    <child link="left_shoulder_roll_link"/>
-    <axis xyz="1 0 0"/>
-    <limit lower="-1.5882" upper="2.2515" effort="25" velocity="37"/>
-  </joint>
-  <link name="left_shoulder_yaw_link">
-    <inertial>
-      <origin xyz="0.010773 -0.002949 -0.072009" rpy="0 0 0"/>
-      <mass value="0.734"/>
-      <inertia ixx="0.0009988" ixy="7.9E-06" ixz="0.0001412" iyy="0.0010605" iyz="-2.86E-05" izz="0.0004354"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_shoulder_yaw_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_shoulder_yaw_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_shoulder_yaw_joint" type="revolute">
-    <origin xyz="0 0.00624 -0.1032" rpy="0 0 0"/>
-    <parent link="left_shoulder_roll_link"/>
-    <child link="left_shoulder_yaw_link"/>
-    <axis xyz="0 0 1"/>
-    <limit lower="-2.618" upper="2.618" effort="25" velocity="37"/>
-  </joint>
-  <link name="left_elbow_link">
-    <inertial>
-      <origin xyz="0.064956 0.004454 -0.010062" rpy="0 0 0"/>
-      <mass value="0.6"/>
-      <inertia ixx="0.0002891" ixy="6.53E-05" ixz="1.72E-05" iyy="0.0004152" iyz="-5.6E-06" izz="0.0004197"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_elbow_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_elbow_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="left_elbow_joint" type="revolute">
-    <origin xyz="0.015783 0 -0.080518" rpy="0 0 0"/>
-    <parent link="left_shoulder_yaw_link"/>
-    <child link="left_elbow_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-1.0472" upper="2.0944" effort="25" velocity="37"/>
-  </joint>
-  <joint name="left_wrist_roll_joint" type="revolute">
-    <origin xyz="0.100 0.00188791 -0.010" rpy="0 0 0"/>
-    <axis xyz="1 0 0"/>
-    <parent link="left_elbow_link"/>
-    <child link="left_wrist_roll_rubber_hand"/>
-    <limit effort="25" velocity="37" lower="-1.972222054" upper="1.972222054"/>
-  </joint>
-  <link name="left_wrist_roll_rubber_hand">
-    <inertial>
-      <origin xyz="0.10794656650 0.00163511945 0.00202244863" rpy="0 0 0"/>
-      <mass value="0.35692864"/>
-      <inertia ixx="0.00019613494735" ixy="-0.00000419816908" ixz="-0.00003950860580" iyy="0.00200280358206" iyz="0.00000249774203" izz="0.00194181412808"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_wrist_roll_rubber_hand.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/left_wrist_roll_rubber_hand.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <link name="right_shoulder_pitch_link">
-    <inertial>
-      <origin xyz="0 -0.035892 -0.011628" rpy="0 0 0"/>
-      <mass value="0.718"/>
-      <inertia ixx="0.0004291" ixy="9.2E-06" ixz="6.4E-06" iyy="0.000453" iyz="-2.26E-05" izz="0.000423"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_shoulder_pitch_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 -0.04 -0.01" rpy="0 1.5707963267948966 0"/>
-      <geometry>
-        <cylinder radius="0.03" length="0.05"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_shoulder_pitch_joint" type="revolute">
-    <origin xyz="0.0039563 -0.10021 0.23778" rpy="-0.27931 5.4949E-05 0.00019159"/>
-    <parent link="torso_link"/>
-    <child link="right_shoulder_pitch_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-3.0892" upper="2.6704" effort="25" velocity="37"/>
-  </joint>
-  <link name="right_shoulder_roll_link">
-    <inertial>
-      <origin xyz="-0.000227 -0.00727 -0.063243" rpy="0 0 0"/>
-      <mass value="0.643"/>
-      <inertia ixx="0.0006177" ixy="1E-06" ixz="8.7E-06" iyy="0.0006912" iyz="5.3E-06" izz="0.0003894"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_shoulder_roll_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="-0.004 -0.006 -0.053" rpy="0 0 0"/>
-      <geometry>
-        <cylinder radius="0.03" length="0.03"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_shoulder_roll_joint" type="revolute">
-    <origin xyz="0 -0.038 -0.013831" rpy="0.27925 0 0"/>
-    <parent link="right_shoulder_pitch_link"/>
-    <child link="right_shoulder_roll_link"/>
-    <axis xyz="1 0 0"/>
-    <limit lower="-2.2515" upper="1.5882" effort="25" velocity="37"/>
-  </joint>
-  <link name="right_shoulder_yaw_link">
-    <inertial>
-      <origin xyz="0.010773 0.002949 -0.072009" rpy="0 0 0"/>
-      <mass value="0.734"/>
-      <inertia ixx="0.0009988" ixy="-7.9E-06" ixz="0.0001412" iyy="0.0010605" iyz="2.86E-05" izz="0.0004354"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_shoulder_yaw_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_shoulder_yaw_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_shoulder_yaw_joint" type="revolute">
-    <origin xyz="0 -0.00624 -0.1032" rpy="0 0 0"/>
-    <parent link="right_shoulder_roll_link"/>
-    <child link="right_shoulder_yaw_link"/>
-    <axis xyz="0 0 1"/>
-    <limit lower="-2.618" upper="2.618" effort="25" velocity="37"/>
-  </joint>
-  <link name="right_elbow_link">
-    <inertial>
-      <origin xyz="0.064956 -0.004454 -0.010062" rpy="0 0 0"/>
-      <mass value="0.6"/>
-      <inertia ixx="0.0002891" ixy="-6.53E-05" ixz="1.72E-05" iyy="0.0004152" iyz="5.6E-06" izz="0.0004197"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_elbow_link.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_elbow_link.STL"/>
-      </geometry>
-    </collision>
-  </link>
-  <joint name="right_elbow_joint" type="revolute">
-    <origin xyz="0.015783 0 -0.080518" rpy="0 0 0"/>
-    <parent link="right_shoulder_yaw_link"/>
-    <child link="right_elbow_link"/>
-    <axis xyz="0 1 0"/>
-    <limit lower="-1.0472" upper="2.0944" effort="25" velocity="37"/>
-  </joint>
-  <joint name="right_wrist_roll_joint" type="revolute">
-    <origin xyz="0.100 -0.00188791 -0.010" rpy="0 0 0"/>
-    <axis xyz="1 0 0"/>
-    <parent link="right_elbow_link"/>
-    <child link="right_wrist_roll_rubber_hand"/>
-    <limit effort="25" velocity="37" lower="-1.972222054" upper="1.972222054"/>
-  </joint>
-  <link name="right_wrist_roll_rubber_hand">
-    <inertial>
-      <origin xyz="0.10794656650 -0.00163511945 0.00202244863" rpy="0 0 0"/>
-      <mass value="0.35692864"/>
-      <inertia ixx="0.00019613494735" ixy="0.00000419816908" ixz="-0.00003950860580" iyy="0.00200280358206" iyz="-0.00000249774203" izz="0.00194181412808"/>
-    </inertial>
-    <visual>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_wrist_roll_rubber_hand.STL"/>
-      </geometry>
-      <material name="white">
-        <color rgba="0.7 0.7 0.7 1"/>
-      </material>
-    </visual>
-    <collision>
-      <origin xyz="0 0 0" rpy="0 0 0"/>
-      <geometry>
-        <mesh filename="meshes/right_wrist_roll_rubber_hand.STL"/>
-      </geometry>
-    </collision>
-  </link>
-</robot>

src/lerobot/robots/unitree_g1/assets/g1/g1_body29_hand14.xml

@@ -1,408 +0,0 @@
-<mujoco model="g1">
-  <compiler angle="radian" meshdir="meshes"/>
-
-  <asset>
-    <mesh name="pelvis" file="pelvis.STL"/>
-    <mesh name="pelvis_contour_link" file="pelvis_contour_link.STL"/>
-    <mesh name="left_hip_pitch_link" file="left_hip_pitch_link.STL"/>
-    <mesh name="left_hip_roll_link" file="left_hip_roll_link.STL"/>
-    <mesh name="left_hip_yaw_link" file="left_hip_yaw_link.STL"/>
-    <mesh name="left_knee_link" file="left_knee_link.STL"/>
-    <mesh name="left_ankle_pitch_link" file="left_ankle_pitch_link.STL"/>
-    <mesh name="left_ankle_roll_link" file="left_ankle_roll_link.STL"/>
-    <mesh name="right_hip_pitch_link" file="right_hip_pitch_link.STL"/>
-    <mesh name="right_hip_roll_link" file="right_hip_roll_link.STL"/>
-    <mesh name="right_hip_yaw_link" file="right_hip_yaw_link.STL"/>
-    <mesh name="right_knee_link" file="right_knee_link.STL"/>
-    <mesh name="right_ankle_pitch_link" file="right_ankle_pitch_link.STL"/>
-    <mesh name="right_ankle_roll_link" file="right_ankle_roll_link.STL"/>
-    <mesh name="waist_yaw_link" file="waist_yaw_link_rev_1_0.STL"/>
-    <mesh name="waist_roll_link" file="waist_roll_link_rev_1_0.STL"/>
-    <mesh name="torso_link" file="torso_link_rev_1_0.STL"/>
-    <mesh name="logo_link" file="logo_link.STL"/>
-    <mesh name="head_link" file="head_link.STL"/>
-    <mesh name="left_shoulder_pitch_link" file="left_shoulder_pitch_link.STL"/>
-    <mesh name="left_shoulder_roll_link" file="left_shoulder_roll_link.STL"/>
-    <mesh name="left_shoulder_yaw_link" file="left_shoulder_yaw_link.STL"/>
-    <mesh name="left_elbow_link" file="left_elbow_link.STL"/>
-    <mesh name="left_wrist_roll_link" file="left_wrist_roll_link.STL"/>
-    <mesh name="left_wrist_pitch_link" file="left_wrist_pitch_link.STL"/>
-    <mesh name="left_wrist_yaw_link" file="left_wrist_yaw_link.STL"/>
-    <mesh name="left_hand_palm_link" file="left_hand_palm_link.STL"/>
-    <mesh name="left_hand_thumb_0_link" file="left_hand_thumb_0_link.STL"/>
-    <mesh name="left_hand_thumb_1_link" file="left_hand_thumb_1_link.STL"/>
-    <mesh name="left_hand_thumb_2_link" file="left_hand_thumb_2_link.STL"/>
-    <mesh name="left_hand_middle_0_link" file="left_hand_middle_0_link.STL"/>
-    <mesh name="left_hand_middle_1_link" file="left_hand_middle_1_link.STL"/>
-    <mesh name="left_hand_index_0_link" file="left_hand_index_0_link.STL"/>
-    <mesh name="left_hand_index_1_link" file="left_hand_index_1_link.STL"/>
-    <mesh name="right_shoulder_pitch_link" file="right_shoulder_pitch_link.STL"/>
-    <mesh name="right_shoulder_roll_link" file="right_shoulder_roll_link.STL"/>
-    <mesh name="right_shoulder_yaw_link" file="right_shoulder_yaw_link.STL"/>
-    <mesh name="right_elbow_link" file="right_elbow_link.STL"/>
-    <mesh name="right_wrist_roll_link" file="right_wrist_roll_link.STL"/>
-    <mesh name="right_wrist_pitch_link" file="right_wrist_pitch_link.STL"/>
-    <mesh name="right_wrist_yaw_link" file="right_wrist_yaw_link.STL"/>
-    <mesh name="right_hand_palm_link" file="right_hand_palm_link.STL"/>
-    <mesh name="right_hand_thumb_0_link" file="right_hand_thumb_0_link.STL"/>
-    <mesh name="right_hand_thumb_1_link" file="right_hand_thumb_1_link.STL"/>
-    <mesh name="right_hand_thumb_2_link" file="right_hand_thumb_2_link.STL"/>
-    <mesh name="right_hand_middle_0_link" file="right_hand_middle_0_link.STL"/>
-    <mesh name="right_hand_middle_1_link" file="right_hand_middle_1_link.STL"/>
-    <mesh name="right_hand_index_0_link" file="right_hand_index_0_link.STL"/>
-    <mesh name="right_hand_index_1_link" file="right_hand_index_1_link.STL"/>
-  </asset>
-
-  <worldbody>
-    <body name="pelvis" pos="0 0 0.793">
-      <inertial pos="0 0 -0.07605" quat="1 0 -0.000399148 0" mass="3.813" diaginertia="0.010549 0.0093089 0.0079184"/>
-      <joint name="floating_base_joint" type="free" limited="false" actuatorfrclimited="false"/>
-      <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.2 0.2 0.2 1" mesh="pelvis"/>
-      <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="pelvis_contour_link"/>
-      <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="pelvis_contour_link"/>
-      <site name="imu_in_pelvis" size="0.01" pos="0.04525 0 -0.08339"/>
-      <body name="left_hip_pitch_link" pos="0 0.064452 -0.1027">
-        <inertial pos="0.002741 0.047791 -0.02606" quat="0.954862 0.293964 0.0302556 0.030122" mass="1.35" diaginertia="0.00181517 0.00153422 0.00116212"/>
-        <joint name="left_hip_pitch_joint" pos="0 0 0" axis="0 1 0" range="-2.5307 2.8798" actuatorfrcrange="-88 88"/>
-        <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.2 0.2 0.2 1" mesh="left_hip_pitch_link"/>
-        <geom type="mesh" rgba="0.2 0.2 0.2 1" mesh="left_hip_pitch_link"/>
-        <body name="left_hip_roll_link" pos="0 0.052 -0.030465" quat="0.996179 0 -0.0873386 0">
-          <inertial pos="0.029812 -0.001045 -0.087934" quat="0.977808 -1.97119e-05 0.205576 -0.0403793" mass="1.52" diaginertia="0.00254986 0.00241169 0.00148755"/>
-          <joint name="left_hip_roll_joint" pos="0 0 0" axis="1 0 0" range="-0.5236 2.9671" actuatorfrcrange="-139 139"/>
-          <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hip_roll_link"/>
-          <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hip_roll_link"/>
-          <body name="left_hip_yaw_link" pos="0.025001 0 -0.12412">
-            <inertial pos="-0.057709 -0.010981 -0.15078" quat="0.600598 0.15832 0.223482 0.751181" mass="1.702" diaginertia="0.00776166 0.00717575 0.00160139"/>
-            <joint name="left_hip_yaw_joint" pos="0 0 0" axis="0 0 1" range="-2.7576 2.7576" actuatorfrcrange="-88 88"/>
-            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hip_yaw_link"/>
-            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hip_yaw_link"/>
-            <body name="left_knee_link" pos="-0.078273 0.0021489 -0.17734" quat="0.996179 0 0.0873386 0">
-              <inertial pos="0.005457 0.003964 -0.12074" quat="0.923418 -0.0327699 0.0158246 0.382067" mass="1.932" diaginertia="0.0113804 0.0112778 0.00146458"/>
-              <joint name="left_knee_joint" pos="0 0 0" axis="0 1 0" range="-0.087267 2.8798" actuatorfrcrange="-139 139"/>
-              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_knee_link"/>
-              <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_knee_link"/>
-              <body name="left_ankle_pitch_link" pos="0 -9.4445e-05 -0.30001">
-                <inertial pos="-0.007269 0 0.011137" quat="0.603053 0.369225 0.369225 0.603053" mass="0.074" diaginertia="1.89e-05 1.40805e-05 6.9195e-06"/>
-                <joint name="left_ankle_pitch_joint" pos="0 0 0" axis="0 1 0" range="-0.87267 0.5236" actuatorfrcrange="-50 50"/>
-                <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_ankle_pitch_link"/>
-                <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_ankle_pitch_link"/>
-                <body name="left_ankle_roll_link" pos="0 0 -0.017558">
-                  <inertial pos="0.026505 0 -0.016425" quat="-0.000481092 0.728482 -0.000618967 0.685065" mass="0.608" diaginertia="0.00167218 0.0016161 0.000217621"/>
-                  <joint name="left_ankle_roll_joint" pos="0 0 0" axis="1 0 0" range="-0.2618 0.2618" actuatorfrcrange="-50 50"/>
-                  <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.2 0.2 0.2 1" mesh="left_ankle_roll_link"/>
-                  <geom size="0.005" pos="-0.05 0.025 -0.03" rgba="0.2 0.2 0.2 1"/>
-                  <geom size="0.005" pos="-0.05 -0.025 -0.03" rgba="0.2 0.2 0.2 1"/>
-                  <geom size="0.005" pos="0.12 0.03 -0.03" rgba="0.2 0.2 0.2 1"/>
-                  <geom size="0.005" pos="0.12 -0.03 -0.03" rgba="0.2 0.2 0.2 1"/>
-                </body>
-              </body>
-            </body>
-          </body>
-        </body>
-      </body>
-      <body name="right_hip_pitch_link" pos="0 -0.064452 -0.1027">
-        <inertial pos="0.002741 -0.047791 -0.02606" quat="0.954862 -0.293964 0.0302556 -0.030122" mass="1.35" diaginertia="0.00181517 0.00153422 0.00116212"/>
-        <joint name="right_hip_pitch_joint" pos="0 0 0" axis="0 1 0" range="-2.5307 2.8798" actuatorfrcrange="-88 88"/>
-        <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.2 0.2 0.2 1" mesh="right_hip_pitch_link"/>
-        <geom type="mesh" rgba="0.2 0.2 0.2 1" mesh="right_hip_pitch_link"/>
-        <body name="right_hip_roll_link" pos="0 -0.052 -0.030465" quat="0.996179 0 -0.0873386 0">
-          <inertial pos="0.029812 0.001045 -0.087934" quat="0.977808 1.97119e-05 0.205576 0.0403793" mass="1.52" diaginertia="0.00254986 0.00241169 0.00148755"/>
-          <joint name="right_hip_roll_joint" pos="0 0 0" axis="1 0 0" range="-2.9671 0.5236" actuatorfrcrange="-139 139"/>
-          <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hip_roll_link"/>
-          <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hip_roll_link"/>
-          <body name="right_hip_yaw_link" pos="0.025001 0 -0.12412">
-            <inertial pos="-0.057709 0.010981 -0.15078" quat="0.751181 0.223482 0.15832 0.600598" mass="1.702" diaginertia="0.00776166 0.00717575 0.00160139"/>
-            <joint name="right_hip_yaw_joint" pos="0 0 0" axis="0 0 1" range="-2.7576 2.7576" actuatorfrcrange="-88 88"/>
-            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hip_yaw_link"/>
-            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hip_yaw_link"/>
-            <body name="right_knee_link" pos="-0.078273 -0.0021489 -0.17734" quat="0.996179 0 0.0873386 0">
-              <inertial pos="0.005457 -0.003964 -0.12074" quat="0.923439 0.0345276 0.0116333 -0.382012" mass="1.932" diaginertia="0.011374 0.0112843 0.00146452"/>
-              <joint name="right_knee_joint" pos="0 0 0" axis="0 1 0" range="-0.087267 2.8798" actuatorfrcrange="-139 139"/>
-              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_knee_link"/>
-              <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_knee_link"/>
-              <body name="right_ankle_pitch_link" pos="0 9.4445e-05 -0.30001">
-                <inertial pos="-0.007269 0 0.011137" quat="0.603053 0.369225 0.369225 0.603053" mass="0.074" diaginertia="1.89e-05 1.40805e-05 6.9195e-06"/>
-                <joint name="right_ankle_pitch_joint" pos="0 0 0" axis="0 1 0" range="-0.87267 0.5236" actuatorfrcrange="-50 50"/>
-                <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_ankle_pitch_link"/>
-                <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_ankle_pitch_link"/>
-                <body name="right_ankle_roll_link" pos="0 0 -0.017558">
-                  <inertial pos="0.026505 0 -0.016425" quat="0.000481092 0.728482 0.000618967 0.685065" mass="0.608" diaginertia="0.00167218 0.0016161 0.000217621"/>
-                  <joint name="right_ankle_roll_joint" pos="0 0 0" axis="1 0 0" range="-0.2618 0.2618" actuatorfrcrange="-50 50"/>
-                  <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.2 0.2 0.2 1" mesh="right_ankle_roll_link"/>
-                  <geom size="0.005" pos="-0.05 0.025 -0.03" rgba="0.2 0.2 0.2 1"/>
-                  <geom size="0.005" pos="-0.05 -0.025 -0.03" rgba="0.2 0.2 0.2 1"/>
-                  <geom size="0.005" pos="0.12 0.03 -0.03" rgba="0.2 0.2 0.2 1"/>
-                  <geom size="0.005" pos="0.12 -0.03 -0.03" rgba="0.2 0.2 0.2 1"/>
-                </body>
-              </body>
-            </body>
-          </body>
-        </body>
-      </body>
-      <body name="waist_yaw_link">
-        <inertial pos="0.003494 0.000233 0.018034" quat="0.289697 0.591001 -0.337795 0.672821" mass="0.214" diaginertia="0.000163531 0.000107714 0.000102205"/>
-        <joint name="waist_yaw_joint" pos="0 0 0" axis="0 0 1" range="-2.618 2.618" actuatorfrcrange="-88 88"/>
-        <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="waist_yaw_link"/>
-        <body name="waist_roll_link" pos="-0.0039635 0 0.044">
-          <inertial pos="0 2.3e-05 0" quat="0.5 0.5 -0.5 0.5" mass="0.086" diaginertia="8.245e-06 7.079e-06 6.339e-06"/>
-          <joint name="waist_roll_joint" pos="0 0 0" axis="1 0 0" range="-0.52 0.52" actuatorfrcrange="-50 50"/>
-          <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="waist_roll_link"/>
-          <body name="torso_link">
-            <inertial pos="0.00203158 0.000339683 0.184568" quat="0.999803 -6.03319e-05 0.0198256 0.00131986" mass="7.818" diaginertia="0.121847 0.109825 0.0273735"/>
-            <joint name="waist_pitch_joint" pos="0 0 0" axis="0 1 0" range="-0.52 0.52" actuatorfrcrange="-50 50"/>
-            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="torso_link"/>
-            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="torso_link"/>
-            <geom pos="0.0039635 0 -0.044" quat="1 0 0 0" type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.2 0.2 0.2 1" mesh="logo_link"/>
-            <geom pos="0.0039635 0 -0.044" quat="1 0 0 0" type="mesh" rgba="0.2 0.2 0.2 1" mesh="logo_link"/>
-            <geom pos="0.0039635 0 -0.044" type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.2 0.2 0.2 1" mesh="head_link"/>
-            <geom pos="0.0039635 0 -0.044" type="mesh" rgba="0.2 0.2 0.2 1" mesh="head_link"/>
-            <site name="imu_in_torso" size="0.01" pos="-0.03959 -0.00224 0.14792"/>
-            <body name="left_shoulder_pitch_link" pos="0.0039563 0.10022 0.24778" quat="0.990264 0.139201 1.38722e-05 -9.86868e-05">
-              <inertial pos="0 0.035892 -0.011628" quat="0.654152 0.0130458 -0.326267 0.68225" mass="0.718" diaginertia="0.000465864 0.000432842 0.000406394"/>
-              <joint name="left_shoulder_pitch_joint" pos="0 0 0" axis="0 1 0" range="-3.0892 2.6704" actuatorfrcrange="-25 25"/>
-              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_shoulder_pitch_link"/>
-              <geom size="0.03 0.025" pos="0 0.04 -0.01" quat="0.707107 0 0.707107 0" type="cylinder" rgba="0.7 0.7 0.7 1"/>
-              <body name="left_shoulder_roll_link" pos="0 0.038 -0.013831" quat="0.990268 -0.139172 0 0">
-                <inertial pos="-0.000227 0.00727 -0.063243" quat="0.701256 -0.0196223 -0.00710317 0.712604" mass="0.643" diaginertia="0.000691311 0.000618011 0.000388977"/>
-                <joint name="left_shoulder_roll_joint" pos="0 0 0" axis="1 0 0" range="-1.5882 2.2515" actuatorfrcrange="-25 25"/>
-                <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_shoulder_roll_link"/>
-                <geom size="0.03 0.015" pos="-0.004 0.006 -0.053" type="cylinder" rgba="0.7 0.7 0.7 1"/>
-                <body name="left_shoulder_yaw_link" pos="0 0.00624 -0.1032">
-                  <inertial pos="0.010773 -0.002949 -0.072009" quat="0.716879 -0.0964829 -0.0679942 0.687134" mass="0.734" diaginertia="0.00106187 0.00103217 0.000400661"/>
-                  <joint name="left_shoulder_yaw_joint" pos="0 0 0" axis="0 0 1" range="-2.618 2.618" actuatorfrcrange="-25 25"/>
-                  <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_shoulder_yaw_link"/>
-                  <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_shoulder_yaw_link"/>
-                  <body name="left_elbow_link" pos="0.015783 0 -0.080518">
-                    <inertial pos="0.064956 0.004454 -0.010062" quat="0.541765 0.636132 0.388821 0.388129" mass="0.6" diaginertia="0.000443035 0.000421612 0.000259353"/>
-                    <joint name="left_elbow_joint" pos="0 0 0" axis="0 1 0" range="-1.0472 2.0944" actuatorfrcrange="-25 25"/>
-                    <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_elbow_link"/>
-                    <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_elbow_link"/>
-                    <body name="left_wrist_roll_link" pos="0.1 0.00188791 -0.01">
-                      <inertial pos="0.0171394 0.000537591 4.8864e-07" quat="0.575338 0.411667 -0.574906 0.411094" mass="0.085445" diaginertia="5.48211e-05 4.96646e-05 3.57798e-05"/>
-                      <joint name="left_wrist_roll_joint" pos="0 0 0" axis="1 0 0" range="-1.97222 1.97222" actuatorfrcrange="-25 25"/>
-                      <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_wrist_roll_link"/>
-                      <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_wrist_roll_link"/>
-                      <body name="left_wrist_pitch_link" pos="0.038 0 0">
-                        <inertial pos="0.0229999 -0.00111685 -0.00111658" quat="0.249998 0.661363 0.293036 0.643608" mass="0.48405" diaginertia="0.000430353 0.000429873 0.000164648"/>
-                        <joint name="left_wrist_pitch_joint" pos="0 0 0" axis="0 1 0" range="-1.61443 1.61443" actuatorfrcrange="-5 5"/>
-                        <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_wrist_pitch_link"/>
-                        <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_wrist_pitch_link"/>
-                        <body name="left_wrist_yaw_link" pos="0.046 0 0">
-                          <inertial pos="0.0885506 0.00212216 -0.000374562" quat="0.487149 0.493844 0.513241 0.505358" mass="0.457415" diaginertia="0.00105989 0.000895419 0.000323842"/>
-                          <joint name="left_wrist_yaw_joint" pos="0 0 0" axis="0 0 1" range="-1.61443 1.61443" actuatorfrcrange="-5 5"/>
-                          <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_wrist_yaw_link"/>
-                          <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_wrist_yaw_link"/>
-                          <geom pos="0.0415 0.003 0" quat="1 0 0 0" type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hand_palm_link"/>
-                          <geom pos="0.0415 0.003 0" quat="1 0 0 0" type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hand_palm_link"/>
-                          <body name="left_hand_thumb_0_link" pos="0.067 0.003 0">
-                            <inertial pos="-0.000884246 -0.00863407 0.000944293" quat="0.462991 0.643965 -0.460173 0.398986" mass="0.0862366" diaginertia="1.6546e-05 1.60058e-05 1.43741e-05"/>
-                            <joint name="left_hand_thumb_0_joint" pos="0 0 0" axis="0 1 0" range="-1.0472 1.0472" actuatorfrcrange="-2.45 2.45"/>
-                            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hand_thumb_0_link"/>
-                            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hand_thumb_0_link"/>
-                            <body name="left_hand_thumb_1_link" pos="-0.0025 -0.0193 0">
-                              <inertial pos="-0.000827888 -0.0354744 -0.0003809" quat="0.685598 0.705471 -0.15207 0.0956069" mass="0.0588507" diaginertia="1.28514e-05 1.22902e-05 5.9666e-06"/>
-                              <joint name="left_hand_thumb_1_joint" pos="0 0 0" axis="0 0 1" range="-0.724312 1.0472" actuatorfrcrange="-1.4 1.4"/>
-                              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hand_thumb_1_link"/>
-                              <geom size="0.01 0.015 0.01" pos="-0.001 -0.032 0" type="box" rgba="0.7 0.7 0.7 1"/>
-                              <body name="left_hand_thumb_2_link" pos="0 -0.0458 0">
-                                <inertial pos="-0.00171735 -0.0262819 0.000107789" quat="0.703174 0.710977 -0.00017564 -0.00766553" mass="0.0203063" diaginertia="4.61314e-06 3.86645e-06 1.53495e-06"/>
-                                <joint name="left_hand_thumb_2_joint" pos="0 0 0" axis="0 0 1" range="0 1.74533" actuatorfrcrange="-1.4 1.4"/>
-                                <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hand_thumb_2_link"/>
-                                <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hand_thumb_2_link"/>
-                              </body>
-                            </body>
-                          </body>
-                          <body name="left_hand_middle_0_link" pos="0.1192 0.0046 -0.0285">
-                            <inertial pos="0.0354744 0.000827888 0.0003809" quat="0.391313 0.552395 0.417187 0.606373" mass="0.0588507" diaginertia="1.28514e-05 1.22902e-05 5.9666e-06"/>
-                            <joint name="left_hand_middle_0_joint" pos="0 0 0" axis="0 0 1" range="-1.5708 0" actuatorfrcrange="-1.4 1.4"/>
-                            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hand_middle_0_link"/>
-                            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hand_middle_0_link"/>
-                            <body name="left_hand_middle_1_link" pos="0.0458 0 0">
-                              <inertial pos="0.0262819 0.00171735 -0.000107789" quat="0.502612 0.491799 0.502639 0.502861" mass="0.0203063" diaginertia="4.61314e-06 3.86645e-06 1.53495e-06"/>
-                              <joint name="left_hand_middle_1_joint" pos="0 0 0" axis="0 0 1" range="-1.74533 0" actuatorfrcrange="-1.4 1.4"/>
-                              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hand_middle_1_link"/>
-                              <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hand_middle_1_link"/>
-                            </body>
-                          </body>
-                          <body name="left_hand_index_0_link" pos="0.1192 0.0046 0.0285">
-                            <inertial pos="0.0354744 0.000827888 0.0003809" quat="0.391313 0.552395 0.417187 0.606373" mass="0.0588507" diaginertia="1.28514e-05 1.22902e-05 5.9666e-06"/>
-                            <joint name="left_hand_index_0_joint" pos="0 0 0" axis="0 0 1" range="-1.5708 0" actuatorfrcrange="-1.4 1.4"/>
-                            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hand_index_0_link"/>
-                            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hand_index_0_link"/>
-                            <body name="left_hand_index_1_link" pos="0.0458 0 0">
-                              <inertial pos="0.0262819 0.00171735 -0.000107789" quat="0.502612 0.491799 0.502639 0.502861" mass="0.0203063" diaginertia="4.61314e-06 3.86645e-06 1.53495e-06"/>
-                              <joint name="left_hand_index_1_joint" pos="0 0 0" axis="0 0 1" range="-1.74533 0" actuatorfrcrange="-1.4 1.4"/>
-                              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="left_hand_index_1_link"/>
-                              <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="left_hand_index_1_link"/>
-                            </body>
-                          </body>
-                        </body>
-                      </body>
-                    </body>
-                  </body>
-                </body>
-              </body>
-            </body>
-            <body name="right_shoulder_pitch_link" pos="0.0039563 -0.10021 0.24778" quat="0.990264 -0.139201 1.38722e-05 9.86868e-05">
-              <inertial pos="0 -0.035892 -0.011628" quat="0.68225 -0.326267 0.0130458 0.654152" mass="0.718" diaginertia="0.000465864 0.000432842 0.000406394"/>
-              <joint name="right_shoulder_pitch_joint" pos="0 0 0" axis="0 1 0" range="-3.0892 2.6704" actuatorfrcrange="-25 25"/>
-              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_shoulder_pitch_link"/>
-              <geom size="0.03 0.025" pos="0 -0.04 -0.01" quat="0.707107 0 0.707107 0" type="cylinder" rgba="0.7 0.7 0.7 1"/>
-              <body name="right_shoulder_roll_link" pos="0 -0.038 -0.013831" quat="0.990268 0.139172 0 0">
-                <inertial pos="-0.000227 -0.00727 -0.063243" quat="0.712604 -0.00710317 -0.0196223 0.701256" mass="0.643" diaginertia="0.000691311 0.000618011 0.000388977"/>
-                <joint name="right_shoulder_roll_joint" pos="0 0 0" axis="1 0 0" range="-2.2515 1.5882" actuatorfrcrange="-25 25"/>
-                <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_shoulder_roll_link"/>
-                <geom size="0.03 0.015" pos="-0.004 -0.006 -0.053" type="cylinder" rgba="0.7 0.7 0.7 1"/>
-                <body name="right_shoulder_yaw_link" pos="0 -0.00624 -0.1032">
-                  <inertial pos="0.010773 0.002949 -0.072009" quat="0.687134 -0.0679942 -0.0964829 0.716879" mass="0.734" diaginertia="0.00106187 0.00103217 0.000400661"/>
-                  <joint name="right_shoulder_yaw_joint" pos="0 0 0" axis="0 0 1" range="-2.618 2.618" actuatorfrcrange="-25 25"/>
-                  <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_shoulder_yaw_link"/>
-                  <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_shoulder_yaw_link"/>
-                  <body name="right_elbow_link" pos="0.015783 0 -0.080518">
-                    <inertial pos="0.064956 -0.004454 -0.010062" quat="0.388129 0.388821 0.636132 0.541765" mass="0.6" diaginertia="0.000443035 0.000421612 0.000259353"/>
-                    <joint name="right_elbow_joint" pos="0 0 0" axis="0 1 0" range="-1.0472 2.0944" actuatorfrcrange="-25 25"/>
-                    <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_elbow_link"/>
-                    <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_elbow_link"/>
-                    <body name="right_wrist_roll_link" pos="0.1 -0.00188791 -0.01">
-                      <inertial pos="0.0171394 -0.000537591 4.8864e-07" quat="0.411667 0.575338 -0.411094 0.574906" mass="0.085445" diaginertia="5.48211e-05 4.96646e-05 3.57798e-05"/>
-                      <joint name="right_wrist_roll_joint" pos="0 0 0" axis="1 0 0" range="-1.97222 1.97222" actuatorfrcrange="-25 25"/>
-                      <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_wrist_roll_link"/>
-                      <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_wrist_roll_link"/>
-                      <body name="right_wrist_pitch_link" pos="0.038 0 0">
-                        <inertial pos="0.0229999 0.00111685 -0.00111658" quat="0.643608 0.293036 0.661363 0.249998" mass="0.48405" diaginertia="0.000430353 0.000429873 0.000164648"/>
-                        <joint name="right_wrist_pitch_joint" pos="0 0 0" axis="0 1 0" range="-1.61443 1.61443" actuatorfrcrange="-5 5"/>
-                        <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_wrist_pitch_link"/>
-                        <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_wrist_pitch_link"/>
-                        <body name="right_wrist_yaw_link" pos="0.046 0 0">
-                          <inertial pos="0.0885506 -0.00212216 -0.000374562" quat="0.505358 0.513241 0.493844 0.487149" mass="0.457415" diaginertia="0.00105989 0.000895419 0.000323842"/>
-                          <joint name="right_wrist_yaw_joint" pos="0 0 0" axis="0 0 1" range="-1.61443 1.61443" actuatorfrcrange="-5 5"/>
-                          <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_wrist_yaw_link"/>
-                          <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_wrist_yaw_link"/>
-                          <geom pos="0.0415 -0.003 0" quat="1 0 0 0" type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hand_palm_link"/>
-                          <geom pos="0.0415 -0.003 0" quat="1 0 0 0" type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hand_palm_link"/>
-                          <body name="right_hand_thumb_0_link" pos="0.067 -0.003 0">
-                            <inertial pos="-0.000884246 0.00863407 0.000944293" quat="0.643965 0.462991 -0.398986 0.460173" mass="0.0862366" diaginertia="1.6546e-05 1.60058e-05 1.43741e-05"/>
-                            <joint name="right_hand_thumb_0_joint" pos="0 0 0" axis="0 1 0" range="-1.0472 1.0472" actuatorfrcrange="-2.45 2.45"/>
-                            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hand_thumb_0_link"/>
-                            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hand_thumb_0_link"/>
-                            <body name="right_hand_thumb_1_link" pos="-0.0025 0.0193 0">
-                              <inertial pos="-0.000827888 0.0354744 -0.0003809" quat="0.705471 0.685598 -0.0956069 0.15207" mass="0.0588507" diaginertia="1.28514e-05 1.22902e-05 5.9666e-06"/>
-                              <joint name="right_hand_thumb_1_joint" pos="0 0 0" axis="0 0 1" range="-1.0472 0.724312" actuatorfrcrange="-1.4 1.4"/>
-                              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hand_thumb_1_link"/>
-                              <geom size="0.01 0.015 0.01" pos="-0.001 0.032 0" type="box" rgba="0.7 0.7 0.7 1"/>
-                              <body name="right_hand_thumb_2_link" pos="0 0.0458 0">
-                                <inertial pos="-0.00171735 0.0262819 0.000107789" quat="0.710977 0.703174 0.00766553 0.00017564" mass="0.0203063" diaginertia="4.61314e-06 3.86645e-06 1.53495e-06"/>
-                                <joint name="right_hand_thumb_2_joint" pos="0 0 0" axis="0 0 1" range="-1.74533 0" actuatorfrcrange="-1.4 1.4"/>
-                                <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hand_thumb_2_link"/>
-                                <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hand_thumb_2_link"/>
-                              </body>
-                            </body>
-                          </body>
-                          <body name="right_hand_middle_0_link" pos="0.1192 -0.0046 -0.0285">
-                            <inertial pos="0.0354744 -0.000827888 0.0003809" quat="0.606373 0.417187 0.552395 0.391313" mass="0.0588507" diaginertia="1.28514e-05 1.22902e-05 5.9666e-06"/>
-                            <joint name="right_hand_middle_0_joint" pos="0 0 0" axis="0 0 1" range="0 1.5708" actuatorfrcrange="-1.4 1.4"/>
-                            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hand_middle_0_link"/>
-                            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hand_middle_0_link"/>
-                            <body name="right_hand_middle_1_link" pos="0.0458 0 0">
-                              <inertial pos="0.0262819 -0.00171735 -0.000107789" quat="0.502861 0.502639 0.491799 0.502612" mass="0.0203063" diaginertia="4.61314e-06 3.86645e-06 1.53495e-06"/>
-                              <joint name="right_hand_middle_1_joint" pos="0 0 0" axis="0 0 1" range="0 1.74533" actuatorfrcrange="-1.4 1.4"/>
-                              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hand_middle_1_link"/>
-                              <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hand_middle_1_link"/>
-                            </body>
-                          </body>
-                          <body name="right_hand_index_0_link" pos="0.1192 -0.0046 0.0285">
-                            <inertial pos="0.0354744 -0.000827888 0.0003809" quat="0.606373 0.417187 0.552395 0.391313" mass="0.0588507" diaginertia="1.28514e-05 1.22902e-05 5.9666e-06"/>
-                            <joint name="right_hand_index_0_joint" pos="0 0 0" axis="0 0 1" range="0 1.5708" actuatorfrcrange="-1.4 1.4"/>
-                            <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hand_index_0_link"/>
-                            <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hand_index_0_link"/>
-                            <body name="right_hand_index_1_link" pos="0.0458 0 0">
-                              <inertial pos="0.0262819 -0.00171735 -0.000107789" quat="0.502861 0.502639 0.491799 0.502612" mass="0.0203063" diaginertia="4.61314e-06 3.86645e-06 1.53495e-06"/>
-                              <joint name="right_hand_index_1_joint" pos="0 0 0" axis="0 0 1" range="0 1.74533" actuatorfrcrange="-1.4 1.4"/>
-                              <geom type="mesh" contype="0" conaffinity="0" group="1" density="0" rgba="0.7 0.7 0.7 1" mesh="right_hand_index_1_link"/>
-                              <geom type="mesh" rgba="0.7 0.7 0.7 1" mesh="right_hand_index_1_link"/>
-                            </body>
-                          </body>
-                        </body>
-                      </body>
-                    </body>
-                  </body>
-                </body>
-              </body>
-            </body>
-          </body>
-        </body>
-      </body>
-    </body>
-  </worldbody>
-
-  <actuator>
-    <motor name="left_hip_pitch_joint" joint="left_hip_pitch_joint"/>
-    <motor name="left_hip_roll_joint" joint="left_hip_roll_joint"/>
-    <motor name="left_hip_yaw_joint" joint="left_hip_yaw_joint"/>
-    <motor name="left_knee_joint" joint="left_knee_joint"/>
-    <motor name="left_ankle_pitch_joint" joint="left_ankle_pitch_joint"/>
-    <motor name="left_ankle_roll_joint" joint="left_ankle_roll_joint"/>
-    <motor name="right_hip_pitch_joint" joint="right_hip_pitch_joint"/>
-    <motor name="right_hip_roll_joint" joint="right_hip_roll_joint"/>
-    <motor name="right_hip_yaw_joint" joint="right_hip_yaw_joint"/>
-    <motor name="right_knee_joint" joint="right_knee_joint"/>
-    <motor name="right_ankle_pitch_joint" joint="right_ankle_pitch_joint"/>
-    <motor name="right_ankle_roll_joint" joint="right_ankle_roll_joint"/>
-    <motor name="waist_yaw_joint" joint="waist_yaw_joint"/>
-    <motor name="waist_roll_joint" joint="waist_roll_joint"/>
-    <motor name="waist_pitch_joint" joint="waist_pitch_joint"/>
-    <motor name="left_shoulder_pitch_joint" joint="left_shoulder_pitch_joint"/>
-    <motor name="left_shoulder_roll_joint" joint="left_shoulder_roll_joint"/>
-    <motor name="left_shoulder_yaw_joint" joint="left_shoulder_yaw_joint"/>
-    <motor name="left_elbow_joint" joint="left_elbow_joint"/>
-    <motor name="left_wrist_roll_joint" joint="left_wrist_roll_joint"/>
-    <motor name="left_wrist_pitch_joint" joint="left_wrist_pitch_joint"/>
-    <motor name="left_wrist_yaw_joint" joint="left_wrist_yaw_joint"/>
-    <motor name="left_hand_thumb_0_joint" joint="left_hand_thumb_0_joint"/>
-    <motor name="left_hand_thumb_1_joint" joint="left_hand_thumb_1_joint"/>
-    <motor name="left_hand_thumb_2_joint" joint="left_hand_thumb_2_joint"/>
-    <motor name="left_hand_middle_0_joint" joint="left_hand_middle_0_joint"/>
-    <motor name="left_hand_middle_1_joint" joint="left_hand_middle_1_joint"/>
-    <motor name="left_hand_index_0_joint" joint="left_hand_index_0_joint"/>
-    <motor name="left_hand_index_1_joint" joint="left_hand_index_1_joint"/>
-    <motor name="right_shoulder_pitch_joint" joint="right_shoulder_pitch_joint"/>
-    <motor name="right_shoulder_roll_joint" joint="right_shoulder_roll_joint"/>
-    <motor name="right_shoulder_yaw_joint" joint="right_shoulder_yaw_joint"/>
-    <motor name="right_elbow_joint" joint="right_elbow_joint"/>
-    <motor name="right_wrist_roll_joint" joint="right_wrist_roll_joint"/>
-    <motor name="right_wrist_pitch_joint" joint="right_wrist_pitch_joint"/>
-    <motor name="right_wrist_yaw_joint" joint="right_wrist_yaw_joint"/>
-    <motor name="right_hand_thumb_0_joint" joint="right_hand_thumb_0_joint"/>
-    <motor name="right_hand_thumb_1_joint" joint="right_hand_thumb_1_joint"/>
-    <motor name="right_hand_thumb_2_joint" joint="right_hand_thumb_2_joint"/>
-    <motor name="right_hand_index_0_joint" joint="right_hand_index_0_joint"/>
-    <motor name="right_hand_index_1_joint" joint="right_hand_index_1_joint"/>
-    <motor name="right_hand_middle_0_joint" joint="right_hand_middle_0_joint"/>
-    <motor name="right_hand_middle_1_joint" joint="right_hand_middle_1_joint"/>
-  </actuator>
-
-  <sensor>
-    <gyro name="imu-torso-angular-velocity" site="imu_in_torso" noise="5e-4" cutoff="34.9"/>
-    <accelerometer name="imu-torso-linear-acceleration" site="imu_in_torso" noise="1e-2" cutoff="157"/>
-    <gyro name="imu-pelvis-angular-velocity" site="imu_in_pelvis" noise="5e-4" cutoff="34.9"/>
-    <accelerometer name="imu-pelvis-linear-acceleration" site="imu_in_pelvis" noise="1e-2" cutoff="157"/>
-  </sensor>
-
-
-  <!-- setup scene -->
-  <statistic center="1.0 0.7 1.0" extent="0.8"/>
-  <visual>
-    <headlight diffuse="0.6 0.6 0.6" ambient="0.1 0.1 0.1" specular="0.9 0.9 0.9"/>
-    <rgba haze="0.15 0.25 0.35 1"/>
-    <global azimuth="-140" elevation="-20"/>
-  </visual>
-  <asset>
-    <texture type="skybox" builtin="flat" rgb1="0 0 0" rgb2="0 0 0" width="512" height="3072"/>
-    <texture type="2d" name="groundplane" builtin="checker" mark="edge" rgb1="0.2 0.3 0.4" rgb2="0.1 0.2 0.3" markrgb="0.8 0.8 0.8" width="300" height="300"/>
-    <material name="groundplane" texture="groundplane" texuniform="true" texrepeat="5 5" reflectance="0.2"/>
-  </asset>
-  <worldbody>
-    <light pos="1 0 3.5" dir="0 0 -1" directional="true"/>
-    <geom name="floor" size="0 0 0.05" type="plane" material="groundplane"/>
-  </worldbody>
-</mujoco>