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 \

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/envhub_leisaac.mdx

@@ -139,7 +139,7 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     so101_leader,
 )
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging
 from lerobot.envs.factory import make_env
 
@@ -196,7 +196,7 @@ def teleop_loop(teleop: Teleoperator, env: gym.Env, fps: int):
             obs, info = env.reset()
 
         dt_s = time.perf_counter() - loop_start
-        busy_wait(1 / fps - dt_s)
+        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)")
 

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/unitree_g1.mdx

@@ -1,14 +1,14 @@
 # Unitree G1 Robot Setup and Control
 
-This guide covers the complete setup process for the Unitree G1 humanoid robot, from initial connection to running locomotion policies.
+This guide covers the complete setup process for the Unitree G1 humanoid, from initial connection to running gr00t_wbc locomotion.
 
-## 🤖 About the Unitree G1
+## About the Unitree G1
 
-The Unitree G1 humanoid comes in two flavors: 29-DOF and 23-DOF humanoid robot capable of whole-body control, manipulation, and locomotion. In this first PR we introduce:
+We offer support for both 29 and 23 DOF G1. In this first PR we introduce:
 
-- **Low-level motor control** via DDS (Data Distribution Service)
+- **`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 policiey** for bipedal walking and balance
+- **GR00T locomotion policy** for bipedal walking and balance
 
 ---
 
@@ -25,7 +25,7 @@ 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.
+**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
 
@@ -103,76 +103,63 @@ ip a show wlan0
 
 ### Step 4: SSH Over WiFi
 
-Once connected to WiFi, note the robot's IP address (e.g., `172.18.129.215`) and disconnect the Ethernet cable. You can now 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@172.18.129.215
+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
 
-The robot server introduced here acts as a DDS-to-ZMQ bridge, allowing your one to control the robot wirelessly.
+### Step 1: Install LeRobot on the Orin
 
-### Step 1: Copy Server Script to Robot
-
-From your laptop, copy the robot server script:
+SSH into the robot and install LeRobot:
 
 ```bash
-# Copy the server script and its dependencies
-scp src/lerobot/robots/unitree_g1/run_g1_server.py unitree@172.18.129.215:~/run_g1_server.py
-scp src/lerobot/robots/unitree_g1/g1_utils.py unitree@172.18.129.215:~/g1_utils.py
+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 .
 ```
 
-### Step 2: Install Dependencies on Robot
+**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.
 
-SSH into the robot and install required packages:
-
-```bash
-ssh unitree@172.18.129.215
-
-# Install build tools and Python dependencies
-sudo apt update
-sudo apt install -y build-essential python3-dev python3-pip
-
-# Install Python packages (pyzmq and Unitree SDK)
-pip3 install pyzmq
-pip3 install git+https://github.com/unitreerobotics/unitree_sdk2_python.git
-```
-
-> **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 3: Run the Robot Server
+### Step 2: Run the Robot Server
 
 On the robot:
 
 ```bash
-python3 ~/run_g1_server.py
+python src/lerobot/robots/unitree_g1/run_g1_server.py
 ```
 
-You should see output like:
-
-```
-Robot server listening on:
-  Commands:  tcp://*:6000 (PULL)
-  State:     tcp://*:6001 (PUB)
-DDS initialized, forwarding started...
-```
-
-> **Important**: Keep this terminal running. The server must be active for remote control.
+**Important**: Keep this terminal running. The server must be active for remote control.
 
 ---
 
-## 🚶 Part 4: Running GR00T Locomotion
+## Part 4: Running GR00T Locomotion
 
 With the robot server running, you can now control the robot from your laptop.
 
-### Step 1: Install LeRobot with Unitree G1 Support (on 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
@@ -181,40 +168,16 @@ 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 = "172.18.129.215"  # Your robot's WiFi IP
+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 (downloads policies from HuggingFace)
+# Run GR00T locomotion controller
 python examples/unitree_g1/gr00t_locomotion.py --repo-id "nepyope/GR00T-WholeBodyControl_g1"
-
-# Or use the default repo (same as above):
-python examples/unitree_g1/gr00t_locomotion.py
-```
-
-The script will:
-
-1. Download Balance and Walk policies from the Hub (cached locally after first run)
-2. Connect to the robot server over WiFi/ZMQ
-3. Initialize the robot and locomotion controller
-4. Gradually move legs to default standing position (3 seconds)
-5. Start locomotion control loop at 50Hz in background thread
-6. Accept commands from the wireless remote controller
-
-**Expected output:**
-
-```
-INFO - Loading GR00T Balance policy...
-INFO - Loading GR00T Walk policy...
-INFO - [UnitreeG1] Initialize UnitreeG1...
-INFO - [UnitreeG1] Connected to robot.
-INFO - Reached default position (legs only)
-INFO - Locomotion control thread started!
-INFO - Robot initialized with GR00T locomotion policies
-INFO - Locomotion controller running in background thread
-INFO - Press Ctrl+C to stop
 ```
 
 ### Step 4: Control with Remote
@@ -224,7 +187,7 @@ INFO - Press Ctrl+C to stop
 - **R1 button**: Raise waist height
 - **R2 button**: Lower waist height
 
-To stop, press `Ctrl+C` in the terminal.
+Press `Ctrl+C` to stop the policy.
 
 ---
 
@@ -237,4 +200,4 @@ To stop, press `Ctrl+C` in the terminal.
 
 ---
 
-_Last updated: November 2025_
+_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/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/replay.py

@@ -20,7 +20,7 @@ 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
@@ -58,7 +58,7 @@ def main():
         # Send action to robot
         _ = robot.send_action(action)
 
-        busy_wait(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
     robot.disconnect()
 

examples/lekiwi/teleoperate.py

@@ -19,7 +19,7 @@ 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
@@ -71,7 +71,7 @@ def main():
         # Visualize
         log_rerun_data(observation=observation, action=action)
 
-        busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 
 
 if __name__ == "__main__":

examples/phone_to_so100/replay.py

@@ -29,7 +29,7 @@ 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
@@ -96,7 +96,7 @@ def main():
         # Send action to robot
         _ = robot.send_action(joint_action)
 
-        busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
 
     # Clean up
     robot.disconnect()

examples/phone_to_so100/teleoperate.py

@@ -32,7 +32,7 @@ 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
@@ -114,7 +114,7 @@ def main():
         # Visualize
         log_rerun_data(observation=phone_obs, action=joint_action)
 
-        busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/replay.py

@@ -30,7 +30,7 @@ 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
@@ -97,7 +97,7 @@ def main():
         # Send action to robot
         _ = robot.send_action(joint_action)
 
-        busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
+        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
 
     # Clean up
     robot.disconnect()

examples/so100_to_so100_EE/teleoperate.py

@@ -32,7 +32,7 @@ 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
@@ -120,7 +120,7 @@ def main():
         # Visualize
         log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
 
-        busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
 
 
 if __name__ == "__main__":

examples/unitree_g1/gr00t_locomotion.py

@@ -1,5 +1,18 @@
 #!/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
 
@@ -15,7 +28,6 @@ from collections import deque
 
 import numpy as np
 import onnxruntime as ort
-import torch
 from huggingface_hub import hf_hub_download
 
 from lerobot.robots.unitree_g1.config_unitree_g1 import UnitreeG1Config
@@ -23,46 +35,15 @@ from lerobot.robots.unitree_g1.unitree_g1 import UnitreeG1
 
 logger = logging.getLogger(__name__)
 
-GROOT_DEFAULT_ANGLES = np.array(
-    [
-        -0.1,
-        0.0,
-        0.0,
-        0.3,
-        -0.2,
-        0.0,  # left leg
-        -0.1,
-        0.0,
-        0.0,
-        0.3,
-        -0.2,
-        0.0,  # right leg
-        0.0,
-        0.0,
-        0.0,  # waist
-        0.0,
-        0.0,
-        0.0,
-        0.0,
-        0.0,
-        0.0,
-        0.0,  # left arm
-        0.0,
-        0.0,
-        0.0,
-        0.0,
-        0.0,
-        0.0,
-        0.0,  # right arm
-    ],
-    dtype=np.float32,
-)
+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
 
-G1_MODEL = "g1_23"
+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
-elif G1_MODEL == "g1_29":
-    MISSING_JOINTS = []  # waist yaw/pitch, wrist pitch/yaw
 
 LOCOMOTION_ACTION_SCALE = 0.25
 
@@ -214,7 +195,6 @@ class GrootLocomotionController:
             self.groot_obs_stacked[start_idx:end_idx] = obs_frame
 
         # Run policy inference (ONNX) with 516D stacked observation
-        obs_tensor = torch.from_numpy(self.groot_obs_stacked).unsqueeze(0)
 
         cmd_magnitude = np.linalg.norm(self.locomotion_cmd)
 
@@ -223,7 +203,7 @@ class GrootLocomotionController:
         )  # balance/standing policy for small commands, walking policy for movement commands
 
         # run policy inference
-        ort_inputs = {selected_policy.get_inputs()[0].name: obs_tensor.cpu().numpy()}
+        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()
 

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" }
@@ -109,7 +109,7 @@ 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",
-    "unitree_sdk2py @ git+https://github.com/unitreerobotics/unitree_sdk2_python.git",
+    "onnxruntime>=1.16.0"
 ]
 reachy2 = ["reachy2_sdk>=1.0.14,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]

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/config_unitree_g1.py

@@ -18,93 +18,38 @@ from dataclasses import dataclass, field
 
 from ..config import RobotConfig
 
+_GAINS: dict[str, dict[str, list[float]]] = {
+    "left_leg": {
+        "kp": [150, 150, 150, 300, 40, 40],
+        "kd": [2, 2, 2, 4, 2, 2],
+    },  # pitch, roll, yaw, knee, ankle_pitch, ankle_roll
+    "right_leg": {"kp": [150, 150, 150, 300, 40, 40], "kd": [2, 2, 2, 4, 2, 2]},
+    "waist": {"kp": [250, 250, 250], "kd": [5, 5, 5]},  # yaw, roll, pitch
+    "left_arm": {"kp": [80, 80, 80, 80], "kd": [3, 3, 3, 3]},  # shoulder_pitch/roll/yaw, elbow
+    "left_wrist": {"kp": [40, 40, 40], "kd": [1.5, 1.5, 1.5]},  # roll, pitch, yaw
+    "right_arm": {"kp": [80, 80, 80, 80], "kd": [3, 3, 3, 3]},
+    "right_wrist": {"kp": [40, 40, 40], "kd": [1.5, 1.5, 1.5]},
+    "other": {"kp": [80, 80, 80, 80, 80, 80], "kd": [3, 3, 3, 3, 3, 3]},
+}
+
+
+def _build_gains() -> tuple[list[float], list[float]]:
+    """Build kp and kd lists from body-part groupings."""
+    kp = [v for g in _GAINS.values() for v in g["kp"]]
+    kd = [v for g in _GAINS.values() for v in g["kd"]]
+    return kp, kd
+
+
+_DEFAULT_KP, _DEFAULT_KD = _build_gains()
+
 
 @RobotConfig.register_subclass("unitree_g1")
 @dataclass
 class UnitreeG1Config(RobotConfig):
-    # id: str = "unitree_g1"
+    kp: list[float] = field(default_factory=lambda: _DEFAULT_KP.copy())
+    kd: list[float] = field(default_factory=lambda: _DEFAULT_KD.copy())
 
-    kp: list = field(
-        default_factory=lambda: [
-            150,
-            150,
-            150,
-            300,
-            40,
-            40,  # Left leg pitch, roll, yaw, knee, ankle pitch, ankle roll
-            150,
-            150,
-            150,
-            300,
-            40,
-            40,  # Right leg pitch, roll, yaw, knee, ankle pitch, ankle roll
-            250,
-            250,
-            250,  # Waist yaw, roll, pitch
-            80,
-            80,
-            80,
-            80,  # Left shoulder pitch, roll, yaw, elbow (kp_low)
-            40,
-            40,
-            40,  # Left wrist roll, pitch, yaw (kp_wrist)
-            80,
-            80,
-            80,
-            80,  # Right shoulder pitch, roll, yaw, elbow (kp_low)
-            40,
-            40,
-            40,  # Right wrist roll, pitch, yaw (kp_wrist)
-            80,
-            80,
-            80,
-            80,
-            80,
-            80,  # Other
-        ]
-    )
-
-    kd: list = field(
-        default_factory=lambda: [
-            2,
-            2,
-            2,
-            4,
-            2,
-            2,  # Left leg pitch, roll, yaw, knee, ankle pitch, ankle roll
-            2,
-            2,
-            2,
-            4,
-            2,
-            2,  # Right leg pitch, roll, yaw, knee, ankle pitch, ankle roll
-            5,
-            5,
-            5,  # Waist yaw, roll, pitch
-            3,
-            3,
-            3,
-            3,  # Left shoulder pitch, roll, yaw, elbow (kd_low)
-            1.5,
-            1.5,
-            1.5,  # Left wrist roll, pitch, yaw (kd_wrist)
-            3,
-            3,
-            3,
-            3,  # Right shoulder pitch, roll, yaw, elbow (kd_low)
-            1.5,
-            1.5,
-            1.5,  # Right wrist roll, pitch, yaw (kd_wrist)
-            3,
-            3,
-            3,
-            3,
-            3,
-            3,  # Other
-        ]
-    )
-
-    control_dt = 1.0 / 250.0  # 250Hz
+    control_dt: float = 1.0 / 250.0  # 250Hz
 
     # socket config for ZMQ bridge
-    robot_ip: str = "172.18.129.215"
+    robot_ip: str = "192.168.123.164"

src/lerobot/robots/unitree_g1/g1_utils.py

@@ -1,3 +1,19 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from enum import IntEnum
 
 # ruff: noqa: N801, N815

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -73,7 +73,9 @@ class IMUState:
 # g1 observation class
 @dataclass
 class G1_29_LowState:  # noqa: N801
-    motor_state: list[MotorState] = field(default_factory=lambda: [MotorState() for _ in range(G1_29_Num_Motors)])
+    motor_state: list[MotorState] = field(
+        default_factory=lambda: [MotorState() for _ in range(G1_29_Num_Motors)]
+    )
     imu_state: IMUState = field(default_factory=IMUState)
     wireless_remote: Any = None  # Raw wireless remote data
     mode_machine: int = 0  # Robot mode
@@ -147,7 +149,7 @@ class UnitreeG1(Robot):
         self.crc = CRC()
         self.msg = unitree_hg_msg_dds__LowCmd_()
         self.msg.mode_pr = 0
-        
+
         # Wait for first state message to arrive
         lowstate = None
         while lowstate is None:
@@ -194,7 +196,7 @@ class UnitreeG1(Robot):
 
                 # Capture wireless remote data
                 lowstate.wireless_remote = msg.wireless_remote
-                
+
                 # Capture mode_machine
                 lowstate.mode_machine = msg.mode_machine
 
@@ -202,7 +204,7 @@ class UnitreeG1(Robot):
 
             current_time = time.time()
             all_t_elapsed = current_time - start_time
-            sleep_time = max(0, (self.control_dt - all_t_elapsed))  # maintina constant control dt
+            sleep_time = max(0, (self.control_dt - all_t_elapsed))  # maintain constant control dt
             time.sleep(sleep_time)
 
     @cached_property
@@ -249,6 +251,7 @@ class UnitreeG1(Robot):
     def send_action(self, action: dict[str, Any]) -> dict[str, Any]:
         self.msg.crc = self.crc.Crc(action)
         self.lowcmd_publisher.Write(action)
+        return action
 
     def get_gravity_orientation(self, quaternion):  # get gravity orientation from quaternion
         """Get gravity orientation from quaternion."""

src/lerobot/robots/unitree_g1/unitree_sdk2_socket.py

@@ -1,12 +1,18 @@
-"""
-ZMQ socket wrapper that mimics the Unitree SDK Channel interface.
+#!/usr/bin/env python
 
-This module provides a drop-in replacement for the Unitree SDK's DDS-based
-ChannelPublisher and ChannelSubscriber, using ZMQ sockets instead. This allows
-remote communication with the robot over WiFi via the robot_server bridge.
-
-Uses JSON for secure serialization instead of pickle.
-"""
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 import base64
 import json

src/lerobot/scripts/lerobot_dataset_viz.py

@@ -65,7 +65,6 @@ import argparse
 import gc
 import logging
 import time
-from collections.abc import Iterator
 from pathlib import Path
 
 import numpy as np
@@ -78,19 +77,6 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD
 
 
-class EpisodeSampler(torch.utils.data.Sampler):
-    def __init__(self, dataset: LeRobotDataset, episode_index: int):
-        from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
-        to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
-        self.frame_ids = range(from_idx, to_idx)
-
-    def __iter__(self) -> Iterator:
-        return iter(self.frame_ids)
-
-    def __len__(self) -> int:
-        return len(self.frame_ids)
-
-
 def to_hwc_uint8_numpy(chw_float32_torch: torch.Tensor) -> np.ndarray:
     assert chw_float32_torch.dtype == torch.float32
     assert chw_float32_torch.ndim == 3
@@ -119,12 +105,10 @@ def visualize_dataset(
     repo_id = dataset.repo_id
 
     logging.info("Loading dataloader")
-    episode_sampler = EpisodeSampler(dataset, episode_index)
     dataloader = torch.utils.data.DataLoader(
         dataset,
         num_workers=num_workers,
         batch_size=batch_size,
-        sampler=episode_sampler,
     )
 
     logging.info("Starting Rerun")

src/lerobot/scripts/lerobot_find_joint_limits.py

@@ -50,7 +50,7 @@ from lerobot.teleoperators import (  # noqa: F401
     make_teleoperator_from_config,
     so100_leader,
 )
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 
 
 @dataclass
@@ -114,7 +114,7 @@ def find_joint_and_ee_bounds(cfg: FindJointLimitsConfig):
             print(f"Min joint pos position {np.round(min_pos, 4).tolist()}")
             break
 
-        busy_wait(0.01)
+        precise_sleep(0.01)
 
 
 def main():

src/lerobot/scripts/lerobot_record.py

@@ -119,7 +119,7 @@ from lerobot.utils.control_utils import (
     sanity_check_dataset_robot_compatibility,
 )
 from lerobot.utils.import_utils import register_third_party_devices
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import (
     get_safe_torch_device,
     init_logging,
@@ -364,7 +364,7 @@ def record_loop(
             log_rerun_data(observation=obs_processed, action=action_values)
 
         dt_s = time.perf_counter() - start_loop_t
-        busy_wait(1 / fps - dt_s)
+        precise_sleep(1 / fps - dt_s)
 
         timestamp = time.perf_counter() - start_episode_t
 

src/lerobot/scripts/lerobot_replay.py

@@ -62,7 +62,7 @@ from lerobot.robots import (  # noqa: F401
 )
 from lerobot.utils.constants import ACTION
 from lerobot.utils.import_utils import register_third_party_devices
-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,
@@ -121,7 +121,7 @@ def replay(cfg: ReplayConfig):
         _ = robot.send_action(processed_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()
 

src/lerobot/scripts/lerobot_teleoperate.py

@@ -89,7 +89,7 @@ from lerobot.teleoperators import (  # noqa: F401
     so101_leader,
 )
 from lerobot.utils.import_utils import register_third_party_devices
-from lerobot.utils.robot_utils import busy_wait
+from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.utils import init_logging, move_cursor_up
 from lerobot.utils.visualization_utils import init_rerun, log_rerun_data
 
@@ -170,12 +170,13 @@ def teleop_loop(
             # Display the final robot action that was sent
             for motor, value in robot_action_to_send.items():
                 print(f"{motor:<{display_len}} | {value:>7.2f}")
-            move_cursor_up(len(robot_action_to_send) + 5)
+            move_cursor_up(len(robot_action_to_send) + 3)
 
         dt_s = time.perf_counter() - loop_start
-        busy_wait(1 / fps - dt_s)
+        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)")
+        print(f"Teleop loop time: {loop_s * 1e3:.2f}ms ({1 / loop_s:.0f} Hz)")
+        move_cursor_up(1)
 
         if duration is not None and time.perf_counter() - start >= duration:
             return

src/lerobot/utils/robot_utils.py

@@ -16,14 +16,40 @@ import platform
 import time
 
 
-def busy_wait(seconds):
-    if platform.system() == "Darwin" or platform.system() == "Windows":
-        # On Mac and Windows, `time.sleep` is not accurate and we need to use this while loop trick,
-        # but it consumes CPU cycles.
+def precise_sleep(seconds: float, spin_threshold: float = 0.010, sleep_margin: float = 0.003):
+    """
+    Wait for `seconds` with better precision than time.sleep alone at the expense of more CPU usage.
+
+    Parameters:
+      - seconds: duration to wait
+      - spin_threshold: if remaining <= spin_threshold -> spin; otherwise sleep (seconds). Default 10ms
+      - sleep_margin: when sleeping leave this much time before deadline to avoid oversleep. Default 3ms
+
+    Note:
+        The default parameters are chosen to prioritize timing accuracy over CPU usage for the common 30 FPS use case.
+    """
+    if seconds <= 0:
+        return
+
+    system = platform.system()
+    # On macOS and Windows the scheduler / sleep granularity can make
+    # short sleeps inaccurate. Instead of burning CPU for the whole
+    # duration, sleep for most of the time and spin for the final few
+    # milliseconds to achieve good accuracy with much lower CPU usage.
+    if system in ("Darwin", "Windows"):
         end_time = time.perf_counter() + seconds
-        while time.perf_counter() < end_time:
-            pass
+        while True:
+            remaining = end_time - time.perf_counter()
+            if remaining <= 0:
+                break
+            # If there's more than a couple milliseconds left, sleep most
+            # of the remaining time and leave a small margin for the final spin.
+            if remaining > spin_threshold:
+                # Sleep but avoid sleeping past the end by leaving a small margin.
+                time.sleep(max(remaining - sleep_margin, 0))
+            else:
+                # Final short spin to hit precise timing without long sleeps.
+                pass
     else:
-        # On Linux time.sleep is accurate
-        if seconds > 0:
-            time.sleep(seconds)
+        # On Linux time.sleep is accurate enough for most uses
+        time.sleep(seconds)