Revert accidental changes to dataset files

* enable variable image sizes to pi0/pi0.5

* add square image assertion

.github/workflows/documentation-upload-pr.yml

@@ -31,7 +31,8 @@ jobs:
     name: Upload Preview and Comment
     if: >
       github.event.workflow_run.event == 'pull_request' &&
-      github.event.workflow_run.conclusion == 'success'
+      github.event.workflow_run.conclusion == 'success' &&
+      github.repository == 'huggingface/lerobot'
     uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@main
     with:
       package_name: lerobot

.github/workflows/documentation.yml

@@ -42,7 +42,9 @@ jobs:
   # This job builds and deploys the official documentation.
   build_main_docs:
     name: Build Main Docs
-    if: github.event_name == 'push' || github.event_name == 'workflow_dispatch'
+    if: >
+      (github.event_name == 'push' || github.event_name == 'workflow_dispatch') &&
+      github.repository == 'huggingface/lerobot'
     permissions:
       contents: read
     uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
@@ -58,7 +60,7 @@ jobs:
   # The result of this job triggers the 'Upload PR Documentation' workflow.
   build_pr_docs:
     name: Build PR Docs
-    if: github.event_name == 'pull_request'
+    if: github.event_name == 'pull_request' && github.repository == 'huggingface/lerobot'
     permissions:
       contents: read
       pull-requests: write

.github/workflows/fast_tests.yml

@@ -45,7 +45,6 @@ permissions:
 env:
   UV_VERSION: "0.8.0"
   PYTHON_VERSION: "3.10"
-  DOCKER_IMAGE_NAME: huggingface/lerobot-gpu
 
 # Ensures that only the latest commit for a PR or branch is built, canceling older runs.
 concurrency:
@@ -60,12 +59,19 @@ 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/nightly.yml

@@ -43,6 +43,7 @@ jobs:
     name: Build CPU Docker for Nightly
     runs-on:
       group: aws-general-8-plus
+    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME_CPU }}
     steps:
@@ -77,6 +78,7 @@ jobs:
     name: Build GPU Docker for Nightly
     runs-on:
       group: aws-general-8-plus
+    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME_GPU }}
     steps:

.github/workflows/release.yml

@@ -29,6 +29,7 @@ jobs:
   build-and-publish:
     name: Build and publish Python distributions
     runs-on: ubuntu-latest
+    if: github.repository == 'huggingface/lerobot'
     outputs:
       version: ${{ steps.extract_info.outputs.tag_version }}
     permissions:
@@ -83,11 +84,11 @@ jobs:
           fi
 
       - name: Remove Tags with Git dependencies
-        # TODO(Steven): Temporary patch to remove libero and pi from PyPi 0.4.0 release due to its reliance on git dependencies.
+        # TODO(Steven): Temporary patch to remove pi from PyPi 0.4.0 release due to its reliance on git dependencies.
         run: |
           echo "::info:: Checking for Git dependencies to remove from pyproject.toml..."
-          grep -E '@ git\+https|lerobot\[pi\]|lerobot\[libero\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
-          sed -E -i '/@ git\+https|lerobot\[pi\]|lerobot\[libero\]/d' pyproject.toml
+          grep -E '@ git\+https|lerobot\[pi\]' pyproject.toml | sed 's/^/::warning:: Removing line: /' || true
+          sed -E -i '/@ git\+https|lerobot\[pi\]/d' pyproject.toml
           echo "::info:: Git dependencies removed. Proceeding with build."
 
       - name: Install build dependencies

.github/workflows/stale.yml

@@ -45,6 +45,7 @@ jobs:
   stale:
     name: Close Stale Issues and PRs
     runs-on: ubuntu-latest
+    if: github.repository == 'huggingface/lerobot'
     permissions:
       actions: write
       contents: write # only for delete-branch option

.github/workflows/unbound_deps_tests.yml

@@ -43,14 +43,22 @@ jobs:
   full-tests:
     name: Full Unbound Tests
     runs-on: ubuntu-latest
+    if: github.repository == 'huggingface/lerobot'
     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 \
@@ -70,7 +78,7 @@ jobs:
           echo "Dependencies unbound:" && cat pyproject.toml
 
       - name: Install lerobot with all extras
-        run: uv sync --all-extras
+        run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
 
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv

README.md

@@ -185,6 +185,11 @@ _Replace `[...]` with your desired features._
 For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
+> [!NOTE]
+> For lerobot 0.4.0, if you want to install pi tags, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+>
+> This will be solved in the next patch release
+
 ### Weights & Biases
 
 To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
@@ -337,7 +342,3 @@ If you want, you can cite this work with:
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)
-
-```
-
-```

benchmarks/video/benchmark.py

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

benchmarks/video/capture_camera_feed.py

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

benchmarks/video/run_video_benchmark.py

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

docs/source/_toctree.yml

@@ -9,14 +9,14 @@
     title: Imitation Learning for Robots
   - local: cameras
     title: Cameras
+  - local: bring_your_own_policies
+    title: Bring Your Own Policies
   - local: integrate_hardware
     title: Bring Your Own Hardware
   - local: hilserl
     title: Train a Robot with RL
   - local: hilserl_sim
     title: Train RL in Simulation
-  - local: async
-    title: Use Async Inference
   - local: multi_gpu_training
     title: Multi GPU training
   title: "Tutorials"
@@ -39,10 +39,20 @@
     title: π₀.₅ (Pi05)
   - local: groot
     title: NVIDIA GR00T N1.5
+  - local: xvla
+    title: X-VLA
   title: "Policies"
 - sections:
-  - local: il_sim
-    title: Imitation Learning in Sim
+  - local: async
+    title: Use Async Inference
+  - local: rtc
+    title: Real-Time Chunking (RTC)
+  title: "Inference"
+- sections:
+  - local: envhub
+    title: Environments from the Hub
+  - local: envhub_leisaac
+    title: Control & Train Robots in Sim (LeIsaac)
   - local: libero
     title: Using Libero
   - local: metaworld
@@ -57,6 +67,8 @@
     title: Implement your own processor
   - local: processors_robots_teleop
     title: Processors for Robots and Teleoperators
+  - local: env_processor
+    title: Environment Processors
   title: "Robot Processors"
 - sections:
   - local: so101
@@ -71,11 +83,19 @@
     title: Hope Jr
   - local: reachy2
     title: Reachy 2
+  - local: unitree_g1
+    title: Unitree G1
+  - local: earthrover_mini_plus
+    title: Earth Rover Mini
   title: "Robots"
 - sections:
   - local: phone_teleop
     title: Phone
   title: "Teleoperators"
+- sections:
+  - local: torch_accelerators
+    title: PyTorch accelerators
+  title: "Supported Hardware"
 - sections:
   - local: notebooks
     title: Notebooks

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
 )
@@ -278,7 +278,7 @@ We found the default values of `actions_per_chunk` and `chunk_size_threshold` to
 2. **Adjust your `fps` based on inference latency.** While the server generates a new action chunk, the client is not idle and is stepping through its current action queue. If the two processes happen at fundamentally different speeds, the client might end up with an empty queue. As such, you should reduce your fps if you consistently run out of actions in queue.
 3. **Adjust `chunk_size_threshold`**.
    - Values closer to `0.0` result in almost sequential behavior. Values closer to `1.0` → send observation every step (more bandwidth, relies on good world-model).
-   - We found values around 0.5-0.6 to work well. If you want to tweak this, spin up a `RobotClient` setting the `--debug-visualize-queue-size` to `True`. This will plot the action queue size evolution at runtime, and you can use it to find the value of `chunk_size_threshold` that works best for your setup.
+   - We found values around 0.5-0.6 to work well. If you want to tweak this, spin up a `RobotClient` setting the `--debug_visualize_queue_size` to `True`. This will plot the action queue size evolution at runtime, and you can use it to find the value of `chunk_size_threshold` that works best for your setup.
 
 <p align="center">
   <img
@@ -289,7 +289,7 @@ We found the default values of `actions_per_chunk` and `chunk_size_threshold` to
 <p align="center">
   <i>
     The action queue size is plotted at runtime when the
-    `--debug-visualize-queue-size` flag is passed, for various levels of
+    `--debug_visualize_queue_size` flag is passed, for various levels of
     `chunk_size_threshold` (`g` in the SmolVLA paper).
   </i>
 </p>

docs/source/bring_your_own_policies.mdx

@@ -0,0 +1,175 @@
+# Bring Your Own Policies
+
+This tutorial explains how to integrate your own custom policy implementations into the LeRobot ecosystem, allowing you to leverage all LeRobot tools for training, evaluation, and deployment while using your own algorithms.
+
+## Step 1: Create a Policy Package
+
+Your custom policy should be organized as an installable Python package following LeRobot's plugin conventions.
+
+### Package Structure
+
+Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
+
+```bash
+lerobot_policy_my_custom_policy/
+├── pyproject.toml
+└── src/
+    └── lerobot_policy_my_custom_policy/
+        ├── __init__.py
+        ├── configuration_my_custom_policy.py
+        ├── modeling_my_custom_policy.py
+        └── processor_my_custom_policy.py
+```
+
+### Package Configuration
+
+Set up your `pyproject.toml`:
+
+```toml
+[project]
+name = "lerobot_policy_my_custom_policy"
+version = "0.1.0"
+dependencies = [
+    # your policy-specific dependencies
+]
+requires-python = ">= 3.11"
+
+[build-system]
+build-backend = # your-build-backend
+requires = # your-build-system
+```
+
+## Step 2: Define the Policy Configuration
+
+Create a configuration class that inherits from `PreTrainedConfig` and registers your policy type:
+
+```python
+# configuration_my_custom_policy.py
+from dataclasses import dataclass, field
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+
+@PreTrainedConfig.register_subclass("my_custom_policy")
+@dataclass
+class MyCustomPolicyConfig(PreTrainedConfig):
+    """Configuration class for MyCustomPolicy.
+
+    Args:
+        n_obs_steps: Number of observation steps to use as input
+        horizon: Action prediction horizon
+        n_action_steps: Number of action steps to execute
+        hidden_dim: Hidden dimension for the policy network
+        # Add your policy-specific parameters here
+    """
+    # ...PreTrainedConfig fields...
+    pass
+
+    def __post_init__(self):
+        super().__post_init__()
+        # Add any validation logic here
+
+    def validate_features(self) -> None:
+        """Validate input/output feature compatibility."""
+        # Implement validation logic for your policy's requirements
+        pass
+```
+
+## Step 3: Implement the Policy Class
+
+Create your policy implementation by inheriting from LeRobot's base `PreTrainedPolicy` class:
+
+```python
+# modeling_my_custom_policy.py
+import torch
+import torch.nn as nn
+from typing import Dict, Any
+
+from lerobot.policies.pretrained import PreTrainedPolicy
+from .configuration_my_custom_policy import MyCustomPolicyConfig
+
+class MyCustomPolicy(PreTrainedPolicy):
+    config_class = MyCustomPolicyConfig
+    name = "my_custom_policy"
+
+    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: Dict[str, Any] = None):
+        super().__init__(config, dataset_stats)
+        ...
+```
+
+## Step 4: Add Data Processors
+
+Create processor functions:
+
+```python
+# processor_my_custom_policy.py
+from typing import Dict, Any
+import torch
+
+
+def make_my_custom_policy_pre_post_processors(
+    config,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Create preprocessing and postprocessing functions for your policy."""
+    pass  # Define your preprocessing and postprocessing logic here
+
+```
+
+## Step 5: Package Initialization
+
+Expose your classes in the package's `__init__.py`:
+
+```python
+# __init__.py
+"""Custom policy package for LeRobot."""
+
+try:
+    import lerobot  # noqa: F401
+except ImportError:
+    raise ImportError(
+        "lerobot is not installed. Please install lerobot to use this policy package."
+    )
+
+from .configuration_my_custom_policy import MyCustomPolicyConfig
+from .modeling_my_custom_policy import MyCustomPolicy
+from .processor_my_custom_policy import make_my_custom_policy_pre_post_processors
+
+__all__ = [
+    "MyCustomPolicyConfig",
+    "MyCustomPolicy",
+    "make_my_custom_policy_pre_post_processors",
+]
+```
+
+## Step 6: Installation and Usage
+
+### Install Your Policy Package
+
+```bash
+cd lerobot_policy_my_custom_policy
+pip install -e .
+
+# Or install from PyPI if published
+pip install lerobot_policy_my_custom_policy
+```
+
+### Use Your Policy
+
+Once installed, your policy automatically integrates with LeRobot's training and evaluation tools:
+
+```bash
+lerobot-train \
+    --policy.type my_custom_policy \
+    --env.type pusht \
+    --steps 200000
+```
+
+## Examples and Community Contributions
+
+Check out these example policy implementations:
+
+- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) - Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)
+
+Share your policy implementations with the community! 🤗

docs/source/earthrover_mini_plus.mdx

@@ -0,0 +1,206 @@
+# EarthRover Mini Plus
+
+The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.
+
+## What You Need
+
+### Hardware
+
+- EarthRover Mini robot
+- Computer with Python 3.10 or newer
+- Internet connection
+
+### Setting Up the Frodobots SDK
+
+The robot needs the [Frodobots SDK](https://github.com/Frodobots/earth-rovers-sdk) running on your computer. Here's how:
+
+1. Download and install the SDK:
+
+```bash
+git clone https://github.com/Frodobots/earth-rovers-sdk.git
+cd earth-rovers-sdk
+pip install -r requirements.txt
+```
+
+2. Start the SDK:
+
+```bash
+hypercorn main:app --reload
+```
+
+3. Open your web browser and go to `http://localhost:8000`, then click "Join"
+
+The SDK gives you:
+
+- Live video from front and rear cameras
+
+> [!IMPORTANT]
+> The SDK must be running before you can use the robot.
+
+## Install LeRobot
+
+Follow our [Installation Guide](./installation) to install LeRobot.
+
+In addition to the base installation, install the EarthRover Mini dependencies:
+
+```bash
+pip install -e .
+```
+
+## How It Works
+
+The robot uses the internet to communicate:
+
+- **Movement commands**: Sent through the SDK
+- **Camera video**: Received from the SDK
+- **Robot info**: Battery, location, speed from the SDK
+
+You don't need to plug anything in - it all works through the SDK.
+
+## Calibration
+
+No calibration needed! The robot is ready to use as soon as the SDK is running.
+
+## Controlling the Robot
+
+You control the robot using your keyboard - just like playing a video game with WASD keys.
+
+### Keyboard Controls
+
+| Key | Action                           |
+| --- | -------------------------------- |
+| W   | Move forward                     |
+| S   | Move backward                    |
+| A   | Turn left (with forward motion)  |
+| D   | Turn right (with forward motion) |
+| Q   | Rotate left in place             |
+| E   | Rotate right in place            |
+| X   | Stop all movement                |
+| +/= | Increase speed                   |
+| -   | Decrease speed                   |
+| ESC | Disconnect                       |
+
+### Speed Settings
+
+You can adjust how fast the robot moves:
+
+- **Forward/backward speed**: Default is full speed (1.0)
+- **Turning speed**: Default is full speed (1.0)
+- **Speed changes**: Use +/- keys to adjust by 0.1 each time
+
+### Try It Out
+
+Test driving the robot before recording data:
+
+```python
+from lerobot.robots.earthrover_mini_plus import EarthRoverMiniPlus, EarthRoverMiniPlusConfig
+from lerobot.teleoperators.keyboard import KeyboardRoverTeleop, KeyboardRoverTeleopConfig
+
+# Initialize robot
+robot_config = EarthRoverMiniPlusConfig()
+robot = EarthRoverMiniPlus(robot_config)
+
+# Initialize teleoperator
+teleop_config = KeyboardRoverTeleopConfig(
+    linear_speed=1.0,
+    angular_speed=1.0,
+    speed_increment=0.1
+)
+teleop = KeyboardRoverTeleop(teleop_config)
+
+# Connect
+robot.connect()
+teleop.connect()
+
+# Teleoperate (use keyboard controls)
+try:
+    while True:
+        action = teleop.get_action()
+        robot.send_action(action)
+except KeyboardInterrupt:
+    pass
+finally:
+    robot.disconnect()
+    teleop.disconnect()
+```
+
+> [!TIP]
+> If you're using a Mac, you might need to give Terminal permission to access your keyboard for teleoperation. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
+
+## Recording Data
+
+Once you can drive the robot well, you can start recording data to train AI models. The system records:
+
+- **What you do**: How you move the robot (forward, backward, turning)
+- **What the robot sees**:
+  - Videos from both cameras
+  - Robot speed and direction
+  - Battery level and location
+  - GPS position and signal
+  - Other sensor data
+- **When it happened**: Timestamps for everything
+
+### Setting Up Hugging Face
+
+We use Hugging Face to store your data online. First, log in with your token from [Hugging Face settings](https://huggingface.co/settings/tokens):
+
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face username:
+
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+### Start Recording
+
+Use the standard recording command:
+
+```bash
+python src/lerobot/scripts/lerobot_record.py \
+    --robot.type=earthrover_mini_plus \
+    --teleop.type=keyboard_rover \
+    --dataset.repo_id=your_username/dataset_name \
+    --dataset.num_episodes=2 \
+    --dataset.fps=10 \
+    --dataset.single_task="Navigate around obstacles" \
+    --display_data=true
+```
+
+Replace `your_username/dataset_name` with your Hugging Face username and a name for your dataset.
+
+### What Gets Saved
+
+Your dataset includes:
+
+**Your Actions (2 things)**:
+
+- How much you moved forward/backward
+- How much you turned left/right
+
+**Robot Observations (12 things)**:
+
+- Front camera video
+- Rear camera video
+- Current speed
+- Battery level
+- Which way the robot is facing
+- GPS location (latitude, longitude, signal strength)
+- Network signal strength
+- Vibration level
+- Lamp status (on/off)
+
+### Where Your Data Goes
+
+On your computer: `~/.cache/huggingface/lerobot/{repo-id}`
+
+After recording, your data automatically uploads to your Hugging Face page:
+
+```bash
+echo https://huggingface.co/datasets/${HF_USER}/earthrover-navigation
+```
+
+Your dataset will be tagged with `LeRobot` for community discovery.

docs/source/env_processor.mdx

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

docs/source/envhub.mdx

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

docs/source/envhub_leisaac.mdx

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

docs/source/groot.mdx

@@ -40,7 +40,7 @@ python -c "import flash_attn; print(f'Flash Attention {flash_attn.__version__} i
 3. Install LeRobot by running:
 
 ```bash
-pip install lerobot[groot] # consider also installing libero,dev and test tags
+pip install lerobot[groot]
 ```
 
 ## Usage
@@ -83,6 +83,9 @@ accelerate launch \
 
 ### Libero Benchmark Results
 
+> [!NOTE]
+> Follow our instructions for Libero usage: [Libero](./libero)
+
 GR00T has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the GR00T N1.5 model for 30k steps on the Libero dataset and compared the results to the GR00T reference results.
 
 | Benchmark          | LeRobot Implementation | GR00T Reference |

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()
 ```
@@ -428,7 +428,7 @@ Your robot should replicate movements similar to those you recorded. For example
 
 ## 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:
+To train a policy to control your robot, use the [`lerobot-train`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_train.py) script. A few arguments are required. Here is an example command:
 
 ```bash
 lerobot-train \
@@ -485,7 +485,7 @@ huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
 
 ## Run inference and evaluate your policy
 
-You can use the `record` script from [`lerobot/record.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/record.py) with a policy checkpoint as input, to run inference and evaluate your policy. For instance, run this command or API example to run inference and record 10 evaluation episodes:
+You can use the `record` script from [`lerobot-record`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/scripts/lerobot_record.py) with a policy checkpoint as input, to run inference and evaluate your policy. For instance, run this command or API example to run inference and record 10 evaluation episodes:
 
 <hfoptions id="eval">
 <hfoption id="Command">

docs/source/il_sim.mdx

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

docs/source/installation.mdx

@@ -82,7 +82,7 @@ For a full list of optional dependencies, see:
 https://pypi.org/project/lerobot/
 
 > [!NOTE]
-> For lerobot 0.4.0, if you want to install libero or pi, you will have to do: `pip install "lerobot[pi,libero]@git+https://github.com/huggingface/lerobot.git"`
+> For lerobot 0.4.0, if you want to install pi, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`
 
 ### Troubleshooting
 
@@ -90,7 +90,7 @@ If you encounter build errors, you may need to install additional dependencies:
 To install these for linux run:
 
 ```bash
-sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config
+sudo apt-get install cmake build-essential python3-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev
 ```
 
 For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)

docs/source/libero.mdx

@@ -62,6 +62,11 @@ lerobot-eval \
 
 - Pass a comma-separated list to `--env.task` for multi-suite evaluation.
 
+### Control Mode
+
+LIBERO now supports two control modes: relative and absolute. This matters because different VLA checkpoints are trained with different mode of action to output hence control parameterizations.
+You can switch them with: `env.control_mode = "relative"` and `env.control_mode = "absolute"`
+
 ### Policy inputs and outputs
 
 When using LIBERO through LeRobot, policies interact with the environment via **observations** and **actions**:

docs/source/pi0.mdx

@@ -28,6 +28,11 @@ As described by Physical Intelligence, while AI has achieved remarkable success
    pip install -e ".[pi]"
    ```
 
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
 ## Training Data and Capabilities
 
 π₀ is trained on the largest robot interaction dataset to date, combining three key data sources:

docs/source/pi05.mdx

@@ -36,6 +36,11 @@ This diverse training mixture creates a "curriculum" that enables generalization
    pip install -e ".[pi]"
    ```
 
+   > [!NOTE]
+   > For lerobot 0.4.0, if you want to install pi tag, you will have to do: `pip install "lerobot[pi]@git+https://github.com/huggingface/lerobot.git"`.
+   >
+   > This will be solved in the next patch release
+
 ## Usage
 
 To use π₀.₅ in your LeRobot configuration, specify the policy type as:

docs/source/rtc.mdx

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

docs/source/so101.mdx

@@ -30,131 +30,6 @@ The follower arm uses 6x STS3215 motors with 1/345 gearing. The leader, however,
 | Wrist Roll          |   5   |  1 / 147   |
 | Gripper             |   6   |  1 / 147   |
 
-### Clean Parts
-
-Remove all support material from the 3D-printed parts. The easiest way to do this is using a small screwdriver to get underneath the support material.
-
-It is advisable to install one 3-pin cable in the motor after placing them before continuing assembly.
-
-### Joint 1
-
-- Place the first motor into the base.
-- Fasten the motor with 4 M2x6mm screws (smallest screws). Two from the top and two from the bottom.
-- Slide over the first motor holder and fasten it using two M2x6mm screws (one on each side).
-- Install both motor horns, securing the top horn with a M3x6mm screw.
-- Attach the shoulder part.
-- Tighten the shoulder part with 4 M3x6mm screws on top and 4 M3x6mm screws on the bottom
-- Add the shoulder motor holder.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint1_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Joint 2
-
-- Slide the second motor in from the top.
-- Fasten the second motor with 4 M2x6mm screws.
-- Attach both motor horns to motor 2, again use the M3x6mm horn screw.
-- Attach the upper arm with 4 M3x6mm screws on each side.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint2_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Joint 3
-
-- Insert motor 3 and fasten using 4 M2x6mm screws
-- Attach both motor horns to motor 3 and secure one again with a M3x6mm horn screw.
-- Connect the forearm to motor 3 using 4 M3x6mm screws on each side.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint3_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Joint 4
-
-- Slide over motor holder 4.
-- Slide in motor 4.
-- Fasten motor 4 with 4 M2x6mm screws and attach its motor horns, use a M3x6mm horn screw.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint4_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Joint 5
-
-- Insert motor 5 into the wrist holder and secure it with 2 M2x6mm front screws.
-- Install only one motor horn on the wrist motor and secure it with a M3x6mm horn screw.
-- Secure the wrist to motor 4 using 4 M3x6mm screws on both sides.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint5_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-### Gripper / Handle
-
-<hfoptions id="assembly">
-<hfoption id="Follower">
-
-- Attach the gripper to motor 5, attach it to the motor horn on the wrist using 4 M3x6mm screws.
-- Insert the gripper motor and secure it with 2 M2x6mm screws on each side.
-- Attach the motor horns and again use a M3x6mm horn screw.
-- Install the gripper claw and secure it with 4 M3x6mm screws on both sides.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Gripper_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-</hfoption>
-<hfoption id="Leader">
-
-- Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
-- Attach the handle to motor 5 using 1 M2x6mm screw.
-- Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
-- Attach the follower trigger with 4 M3x6mm screws.
-
-<div class="video-container">
-  <video controls width="600">
-    <source
-      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Leader_v2.mp4"
-      type="video/mp4"
-    />
-  </video>
-</div>
-
-</hfoption>
-</hfoptions>
-
 ## Configure the motors
 
 ### 1. Find the USB ports associated with each arm
@@ -340,6 +215,131 @@ leader.setup_motors()
 </hfoption>
 </hfoptions>
 
+### Clean Parts
+
+Remove all support material from the 3D-printed parts. The easiest way to do this is using a small screwdriver to get underneath the support material.
+
+It is advisable to install one 3-pin cable in the motor after placing them before continuing assembly.
+
+### Joint 1
+
+- Place the first motor into the base.
+- Fasten the motor with 4 M2x6mm screws (smallest screws). Two from the top and two from the bottom.
+- Slide over the first motor holder and fasten it using two M2x6mm screws (one on each side).
+- Install both motor horns, securing the top horn with a M3x6mm screw.
+- Attach the shoulder part.
+- Tighten the shoulder part with 4 M3x6mm screws on top and 4 M3x6mm screws on the bottom
+- Add the shoulder motor holder.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint1_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Joint 2
+
+- Slide the second motor in from the top.
+- Fasten the second motor with 4 M2x6mm screws.
+- Attach both motor horns to motor 2, again use the M3x6mm horn screw.
+- Attach the upper arm with 4 M3x6mm screws on each side.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint2_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Joint 3
+
+- Insert motor 3 and fasten using 4 M2x6mm screws
+- Attach both motor horns to motor 3 and secure one again with a M3x6mm horn screw.
+- Connect the forearm to motor 3 using 4 M3x6mm screws on each side.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint3_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Joint 4
+
+- Slide over motor holder 4.
+- Slide in motor 4.
+- Fasten motor 4 with 4 M2x6mm screws and attach its motor horns, use a M3x6mm horn screw.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint4_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Joint 5
+
+- Insert motor 5 into the wrist holder and secure it with 2 M2x6mm front screws.
+- Install only one motor horn on the wrist motor and secure it with a M3x6mm horn screw.
+- Secure the wrist to motor 4 using 4 M3x6mm screws on both sides.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Joint5_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+### Gripper / Handle
+
+<hfoptions id="assembly">
+<hfoption id="Follower">
+
+- Attach the gripper to motor 5, attach it to the motor horn on the wrist using 4 M3x6mm screws.
+- Insert the gripper motor and secure it with 2 M2x6mm screws on each side.
+- Attach the motor horns and again use a M3x6mm horn screw.
+- Install the gripper claw and secure it with 4 M3x6mm screws on both sides.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Gripper_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+</hfoption>
+<hfoption id="Leader">
+
+- Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
+- Attach the handle to motor 5 using 1 M2x6mm screw.
+- Insert the gripper motor, secure it with 2 M2x6mm screws on each side, attach a motor horn using a M3x6mm horn screw.
+- Attach the follower trigger with 4 M3x6mm screws.
+
+<div class="video-container">
+  <video controls width="600">
+    <source
+      src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Leader_v2.mp4"
+      type="video/mp4"
+    />
+  </video>
+</div>
+
+</hfoption>
+</hfoptions>
+
 ## Calibrate
 
 Next, you'll need to calibrate your robot to ensure that the leader and follower arms have the same position values when they are in the same physical position.

docs/source/torch_accelerators.mdx

@@ -0,0 +1,42 @@
+# PyTorch accelerators
+
+LeRobot supports multiple hardware acceleration options for both training and inference.
+
+These options include:
+
+- **CPU**: CPU executes all computations, no dedicated accelerator is used
+- **CUDA**: acceleration with NVIDIA & AMD GPUs
+- **MPS**: acceleration with Apple Silicon GPUs
+- **XPU**: acceleration with Intel integrated and discrete GPUs
+
+## Getting Started
+
+To use particular accelerator, a suitable version of PyTorch should be installed.
+
+For CPU, CUDA, and MPS backends follow instructions provided on [PyTorch installation page](https://pytorch.org/get-started/locally).
+For XPU backend, follow instructions from [PyTorch documentation](https://docs.pytorch.org/docs/stable/notes/get_start_xpu.html).
+
+### Verifying the installation
+
+After installation, accelerator availability can be verified by running
+
+```python
+import torch
+print(torch.<backend_name>.is_available())  # <backend_name> is cuda, mps, or xpu
+```
+
+## How to run training or evaluation
+
+To select the desired accelerator, use the `--policy.device` flag when running `lerobot-train` or `lerobot-eval`. For example, to use MPS on Apple Silicon, run:
+
+```bash
+lerobot-train
+    --policy.device=mps ...
+```
+
+```bash
+lerobot-eval \
+    --policy.device=mps ...
+```
+
+However, in most cases, presence of an accelerator is detected automatically and `policy.device` parameter can be omitted from CLI commands.

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_

docs/source/using_dataset_tools.mdx

@@ -11,13 +11,14 @@ LeRobot provides several utilities for manipulating datasets:
 3. **Merge Datasets** - Combine multiple datasets into one. The datasets must have identical features, and episodes are concatenated in the order specified in `repo_ids`
 4. **Add Features** - Add new features to a dataset
 5. **Remove Features** - Remove features from a dataset
+6. **Convert to Video** - Convert image-based datasets to video format for efficient storage
 
 The core implementation is in `lerobot.datasets.dataset_tools`.
 An example script detailing how to use the tools API is available in `examples/dataset/use_dataset_tools.py`.
 
 ## Command-Line Tool: lerobot-edit-dataset
 
-`lerobot-edit-dataset` is a command-line script for editing datasets. It can be used to delete episodes, split datasets, merge datasets, add features, and remove features.
+`lerobot-edit-dataset` is a command-line script for editing datasets. It can be used to delete episodes, split datasets, merge datasets, add features, remove features, and convert image datasets to video format.
 
 Run `lerobot-edit-dataset --help` for more information on the configuration of each operation.
 
@@ -86,9 +87,71 @@ lerobot-edit-dataset \
     --operation.feature_names "['observation.images.top']"
 ```
 
+#### Convert to Video
+
+Convert an image-based dataset to video format, creating a new LeRobotDataset where images are stored as videos. This is useful for reducing storage requirements and improving data loading performance. The new dataset will have the exact same structure as the original, but with images encoded as MP4 videos in the proper LeRobot format.
+
+```bash
+# Local-only: Save to a custom output directory (no hub push)
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.output_dir /path/to/output/pusht_video
+
+# Save with new repo_id (local storage)
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --new_repo_id lerobot/pusht_video \
+    --operation.type convert_to_video
+
+# Convert and push to Hugging Face Hub
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --new_repo_id lerobot/pusht_video \
+    --operation.type convert_to_video \
+    --push_to_hub true
+
+# Convert with custom video codec and quality settings
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.output_dir outputs/pusht_video \
+    --operation.vcodec libsvtav1 \
+    --operation.pix_fmt yuv420p \
+    --operation.g 2 \
+    --operation.crf 30
+
+# Convert only specific episodes
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.output_dir outputs/pusht_video \
+    --operation.episode_indices "[0, 1, 2, 5, 10]"
+
+# Convert with multiple workers for parallel processing
+lerobot-edit-dataset \
+    --repo_id lerobot/pusht_image \
+    --operation.type convert_to_video \
+    --operation.output_dir outputs/pusht_video \
+    --operation.num_workers 8
+```
+
+**Parameters:**
+
+- `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
+- `vcodec`: Video codec to use - options: `h264`, `hevc`, `libsvtav1` (default: `libsvtav1`)
+- `pix_fmt`: Pixel format - options: `yuv420p`, `yuv444p` (default: `yuv420p`)
+- `g`: Group of pictures (GOP) size - lower values give better quality but larger files (default: 2)
+- `crf`: Constant rate factor - lower values give better quality but larger files, 0 is lossless (default: 30)
+- `fast_decode`: Fast decode tuning option (default: 0)
+- `episode_indices`: List of specific episodes to convert (default: all episodes)
+- `num_workers`: Number of parallel workers for processing (default: 4)
+
+**Note:** The resulting dataset will be a proper LeRobotDataset with all cameras encoded as videos in the `videos/` directory, with parquet files containing only metadata (no raw image data). All episodes, stats, and tasks are preserved.
+
 ### Push to Hub
 
-Add the `--push_to_hub` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:
+Add the `--push_to_hub true` flag to any command to automatically upload the resulting dataset to the Hugging Face Hub:
 
 ```bash
 lerobot-edit-dataset \
@@ -96,7 +159,7 @@ lerobot-edit-dataset \
     --new_repo_id lerobot/pusht_after_deletion \
     --operation.type delete_episodes \
     --operation.episode_indices "[0, 2, 5]" \
-    --push_to_hub
+    --push_to_hub true
 ```
 
 There is also a tool for adding features to a dataset that is not yet covered in `lerobot-edit-dataset`.

docs/source/xvla.mdx

@@ -0,0 +1,528 @@
+# X-VLA: The First Soft-Prompted Robot Foundation Model for Any Robot, Any Task
+
+## Overview
+
+For years, robotics has aspired to build agents that can follow natural human instructions and operate dexterously across many environments and robot bodies. Recent breakthroughs in LLMs and VLMs suggest a path forward: extend these foundation-model architectures to embodied control by grounding them in actions. This has led to the rise of Vision-Language-Action (VLA) models, with the hope that a single generalist model could combine broad semantic understanding with robust manipulation skills.
+
+But training such models is difficult. Robot data is fragmented across platforms, sensors, embodiments, and collection protocols. Heterogeneity appears everywhere: different arm configurations, different action spaces, different camera setups, different visual domains, and different task distributions. These inconsistencies create major distribution shifts that make pretraining unstable and adaptation unreliable.
+
+Inspired by meta-learning and prompt learning, we ask: **"What if a VLA model could learn the structure of each robot and dataset the same way LLMs learn tasks, through prompts?"**
+
+**X-VLA** is a soft-prompted, flow-matching VLA framework that treats each hardware setup as a "task" and encodes it using a small set of learnable embeddings. These **Soft Prompts** capture embodiment and domain-specific variations, guiding the Transformer from the earliest stages of multimodal fusion. With this mechanism, X-VLA can reconcile diverse robot morphologies, data types, and sensor setups within a single unified architecture.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-architecture.png"
+    alt="XVLA Architecture"
+    style="max-width: 100%; height: auto; width: 800px;"
+  />
+</p>
+
+Built from pure Transformer encoders, X-VLA scales naturally with model size and dataset diversity. Across 6 simulation benchmarks and 3 real robots, Soft Prompts consistently outperform existing methods in handling hardware and domain differences. X-VLA-0.9B, trained on 290K episodes spanning seven robotic platforms, learns an embodiment-agnostic generalist policy in Phase I, and adapts efficiently to new robots in Phase II simply by learning a new set of prompts, while keeping the backbone frozen.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-architecture2.png"
+    alt="XVLA Architecture 2"
+    style="width: 60%; height: auto;"
+  />
+</p>
+
+With only 1% of parameters tuned (9M), X-VLA-0.9B achieves near-π₀ performance on LIBERO and Simpler-WidowX, despite using **300× fewer trainable parameters**. It also demonstrates strong real-world dexterity with minimal demonstrations, including folding cloths in under two minutes.
+
+<p align="center">
+  <img
+    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-fold.png"
+    alt="XVLA fold visualization"
+    style="width: 95%; max-width: 1100px; height: auto;"
+  />
+</p>
+
+X-VLA shows that generalist robot intelligence does not require increasingly complex architectures, only the right way to absorb heterogeneity. Soft Prompts offer a simple, scalable mechanism for unifying diverse robotic data, paving the way toward adaptable, cross-embodiment robot foundation models.
+
+## Installation
+
+After installing LeRobot, install the X-VLA dependencies:
+
+```bash
+pip install -e .[xvla]
+```
+
+After the new release, you'll be able to do:
+
+```bash
+pip install lerobot[xvla]
+```
+
+## Quick Start
+
+### Basic Usage
+
+To use X-VLA in your LeRobot configuration, specify the policy type as:
+
+```bash
+policy.type=xvla
+```
+
+### Evaluating Pre-trained Checkpoints
+
+Example evaluation with LIBERO:
+
+```bash
+lerobot-eval \
+  --policy.path="lerobot/xvla-libero" \
+  --env.type=libero \
+  --env.task=libero_spatial,libero_goal,libero_10 \
+  --env.control_mode=absolute \
+  --eval.batch_size=1 \
+  --eval.n_episodes=1 \
+  --env.episode_length=800 \
+  --seed=142
+```
+
+## Available Checkpoints
+
+### 🎯 Base Model
+
+**[lerobot/xvla-base](https://huggingface.co/lerobot/xvla-base)**
+
+A 0.9B parameter instantiation of X-VLA, trained with a carefully designed data processing and learning recipe. The training pipeline consists of two phases:
+
+- **Phase I: Pretraining** - Pretrained on 290K episodes from Droid, Robomind, and Agibot, spanning seven platforms across five types of robotic arms (single-arm to bi-manual setups). By leveraging soft prompts to absorb embodiment-specific variations, the model learns an embodiment-agnostic generalist policy.
+
+- **Phase II: Domain Adaptation** - Adapted to deployable policies for target domains. A new set of soft prompts is introduced and optimized to encode the hardware configuration of the novel domain, while the pretrained backbone remains frozen.
+
+### Simulation Checkpoints
+
+**[lerobot/xvla-libero](https://huggingface.co/lerobot/xvla-libero)**
+
+Achieves 93% success rate on LIBERO benchmarks. Fine-tuned from the base model for simulation tasks.
+
+**[lerobot/xvla-widowx](https://huggingface.co/lerobot/xvla-widowx)**
+
+Fine-tuned on BridgeData for pick-and-place experiments on compact WidowX platforms. Demonstrates robust manipulation capabilities.
+
+### 🤖 Real-World Checkpoints
+
+**[lerobot/xvla-folding](https://huggingface.co/lerobot/xvla-folding)**
+
+A fine-tuned dexterous manipulation model trained on the high-quality Soft-FOLD cloth folding dataset. Achieves 100% success rate over 2 hours of continuous cloth folding.
+
+**[lerobot/xvla-agibot-world](https://huggingface.co/lerobot/xvla-agibot-world)**
+
+Optimized for AgileX robot dexterous manipulation tasks.
+
+**[lerobot/xvla-google-robot](https://huggingface.co/lerobot/xvla-google-robot)**
+
+Adapted for Google Robot platforms.
+
+## Training X-VLA
+
+### Recommended Training Configuration
+
+When fine-tuning X-VLA for a new embodiment or task, we recommend not freezing the VLM, and also setting the `policy.dtype=bfloat16` to not hit OOM errors.
+
+```bash
+lerobot-train \
+  --dataset.repo_id=YOUR_DATASET \
+  --output_dir=./outputs/xvla_training \
+  --job_name=xvla_training \
+  --policy.path="lerobot/xvla-base" \
+  --policy.repo_id="HF_USER/xvla-your-robot" \
+  --policy.dtype=bfloat16 \
+  --steps=3000 \
+  --policy.device=cuda \
+  --policy.freeze_vision_encoder=false \
+  --policy.freeze_language_encoder=false \
+  --policy.train_policy_transformer=true \
+  --policy.train_soft_prompts=true \
+  --policy.action_mode=YOUR_ACTION_MODE
+```
+
+### Training Parameters Explained
+
+| Parameter                  | Default | Description                                    |
+| -------------------------- | ------- | ---------------------------------------------- |
+| `freeze_vision_encoder`    | `false` | Do not freeze the VLM vision encoder weights   |
+| `freeze_language_encoder`  | `false` | Do not freeze the VLM language encoder weights |
+| `train_policy_transformer` | `true`  | Allow policy transformer layers to train       |
+| `train_soft_prompts`       | `true`  | Allow soft prompts to train                    |
+
+**💡 Best Practice**: For Phase II adaptation to new embodiments, do not freeze the VLM encoders and also train the policy transformer and soft prompts.
+
+### Example: Training on Bimanual Robot
+
+```bash
+lerobot-train \
+  --dataset.repo_id=pepijn223/bimanual-so100-handover-cube \
+  --output_dir=./outputs/xvla_bimanual \
+  --job_name=xvla_so101_training \
+  --policy.path="lerobot/xvla-base" \
+  --policy.dtype=bfloat16 \
+  --policy.repo_id="YOUR_USERNAME/xvla-biso101" \
+  --steps=3000 \
+  --policy.device=cuda \
+  --policy.action_mode=so101_bimanual \
+  --policy.freeze_vision_encoder=false \
+  --policy.freeze_language_encoder=false \
+  --policy.train_policy_transformer=true \
+  --policy.train_soft_prompts=true
+```
+
+💡 **Best Performance:** If you have sufficient computational resources and want to achieve best X-VLA finetuning performance, you should follow the official finetuning strategy:
+
+**🔥 Full-finetune all components with a custom learning-rate scheme**
+
+To ensure stable optimization, the Vision-Language Model (VLM) must be trained with only 1/10 of the base learning rate, while all other components use the full LR.
+This LR ratio is crucial for achieving strong and stable finetuning performance. This is already done for you by default.
+❕Note
+
+Completely matching the official reported performance may require an additional warm-up LR schedule for soft-prompts, which can bring minor improvements.
+We encourage implementing this in your customized training pipeline for optimal results.
+
+## Core Concepts
+
+### 1. Action Modes
+
+X-VLA uses an **Action Registry** system to handle different action spaces and embodiments. The `action_mode` parameter defines how actions are processed, what loss functions are used, and how predictions are post-processed.
+
+#### Available Action Modes
+
+| Action Mode      | Action Dim              | Description                                 | Use Case                             |
+| ---------------- | ----------------------- | ------------------------------------------- | ------------------------------------ |
+| `ee6d`           | 20                      | End-effector with xyz, 6D rotation, gripper | Dual-arm setups with spatial control |
+| `joint`          | 14                      | Joint-space with gripper                    | Direct joint control robots          |
+| `agibot_ee6d`    | 20                      | AGI-bot variant with MSE loss               | AGI-bot platforms                    |
+| `so101_bimanual` | 20 (model), 12 (real)   | SO101 bimanual robot                        | Bimanual manipulation tasks          |
+| `auto`           | 20 (model), auto (real) | Auto-detects action dim from dataset        | **Recommended** for new robots       |
+
+#### Why Action Modes Matter
+
+When you have a pretrained checkpoint like `lerobot/xvla-base` trained with `action_dim=20`, and you want to train on a dataset with a different action dimension (e.g., 14 for bimanual arms), you can't simply trim the action dimension. The action mode orchestrates:
+
+1. **Loss Computation**: Different loss functions for different action components (MSE for joints, BCE for grippers, etc.)
+2. **Preprocessing**: Zeroing out gripper channels, padding dimensions
+3. **Postprocessing**: Applying sigmoid to gripper logits, trimming padding
+
+#### Example: BimanualSO101 Action Space
+
+The `so101_bimanual` action mode handles the mismatch between model output (20D) and real robot control (12D):
+
+```python
+# Model outputs 20 dimensions for compatibility
+dim_action = 20
+
+# Real robot only needs 12 dimensions
+# [left_arm (6), right_arm (6)] = [joints (5) + gripper (1)] × 2
+REAL_DIM = 12
+
+# Preprocessing: Pad 12D actions to 20D for training
+# Postprocessing: Trim 20D predictions to 12D for deployment
+```
+
+See the [action_hub.py](/home/jade_choghari/robot/lerobot/src/lerobot/policies/xvla/action_hub.py) implementation for details.
+
+#### Auto Action Mode (Recommended)
+
+The `auto` action mode is the easiest way to use X-VLA with any robot. It automatically detects your dataset's action dimension and handles padding/trimming:
+
+```bash
+lerobot-train \
+  --policy.path="lerobot/xvla-base" \
+  --policy.action_mode=auto \
+  --policy.max_action_dim=20 \
+  ...
+```
+
+**How it works:**
+
+- Reads `action_feature.shape[-1]` from your dataset (e.g., 7 for Franka)
+- Model outputs `max_action_dim` (default 20) for pretrained compatibility
+- Loss is computed **only on the real dimensions**: `MSE(pred[:,:,:real_dim], target[:,:,:real_dim])`
+- Postprocess trims output back to `real_dim` for robot control
+
+This eliminates the need to create custom action modes for most robots.
+
+### 2. Domain IDs
+
+Domain IDs are learnable identifiers for different robot configurations and camera setups. They allow X-VLA to distinguish between:
+
+- Different robots (Robot 1 vs Robot 2)
+- Different camera configurations (cam1 vs cam2)
+- Different combinations (Robot1-cam1-cam2 vs Robot1-cam1 vs Robot2-cam1)
+
+#### Setting Domain IDs
+
+**During Training**: By default, domain_id is set to 0 for general training.
+
+**During Evaluation**: Specify the domain_id that matches your checkpoint's training configuration.
+
+```python
+# Example: LIBERO checkpoint uses domain_id=3
+domain_id = 3
+```
+
+The domain_id is automatically added to observations by the `XVLAAddDomainIdProcessorStep` in the preprocessing pipeline.
+
+The `lerobot/xvla-base` model has been trained on the following domain IDs. It is recommended to choose one that most resembles your robot/configuration:
+
+#### Fine-tuning Datasets
+
+| Dataset Name     | Domain ID |
+| ---------------- | --------- |
+| Bridge           | 0         |
+| RT1              | 1         |
+| Calvin           | 2         |
+| libero           | 3         |
+| widowx-air       | 4         |
+| AIR-AGILEX-HQ    | 5         |
+| robotwin2_abs_ee | 6         |
+| robotwin2_clean  | 6         |
+| robocasa-human   | 7         |
+| VLABench         | 8         |
+| AGIBOT-challenge | 9         |
+| AIR-AGILEX       | 10        |
+| AIRBOT           | 18        |
+
+### 3. Processor Steps
+
+X-VLA requires specific preprocessing and postprocessing steps for proper operation.
+
+#### Required Preprocessing Steps
+
+1. **XVLAImageToFloatProcessorStep**: Converts images from [0, 255] to [0, 1] range
+2. **XVLAImageNetNormalizeProcessorStep**: Applies ImageNet normalization (required for VLM backbone)
+3. **XVLAAddDomainIdProcessorStep**: Adds domain_id to observations
+
+#### Example Custom Processor
+
+For LIBERO environments, a custom processor handles the specific observation format:
+
+```python
+from lerobot.policies.xvla.processor_xvla import LiberoProcessorStep
+
+processor = LiberoProcessorStep()
+# Handles robot_state dictionary, converts rotation matrices to 6D representation
+# Applies 180° image rotation for camera convention
+```
+
+### 4. Configuration Parameters
+
+Key configuration parameters for X-VLA:
+
+```python
+# Observation and action
+n_obs_steps: int = 1          # Number of observation timesteps
+chunk_size: int = 32           # Action sequence length
+n_action_steps: int = 32       # Number of action steps to execute
+
+# Model architecture
+hidden_size: int = 1024        # Transformer hidden dimension
+depth: int = 24                # Number of transformer layers
+num_heads: int = 16            # Number of attention heads
+num_domains: int = 30          # Maximum number of domain IDs
+len_soft_prompts: int = 32     # Length of soft prompt embeddings
+
+# Action space
+action_mode: str = "ee6d"      # Action space type (use "auto" for auto-detection)
+use_proprio: bool = True       # Use proprioceptive state
+max_state_dim: int = 32        # Maximum state dimension
+max_action_dim: int = 20       # Max action dim for padding (used by "auto" mode)
+
+# Vision
+num_image_views: int | None    # Number of camera views
+resize_imgs_with_padding: tuple[int, int] | None  # Target image size with padding
+
+# Training
+num_denoising_steps: int = 10  # Flow matching denoising steps
+```
+
+## Creating Custom Action Modes
+
+If your robot has a unique action space, you can create a custom action mode:
+
+### Step 1: Define Your Action Space
+
+```python
+from lerobot.policies.xvla.action_hub import BaseActionSpace, register_action
+import torch.nn as nn
+
+@register_action("my_custom_robot")
+class MyCustomActionSpace(BaseActionSpace):
+    """Custom action space for my robot."""
+
+    dim_action = 15  # Your robot's action dimension
+    gripper_idx = (7, 14)  # Gripper channel indices
+
+    def __init__(self):
+        super().__init__()
+        self.mse = nn.MSELoss()
+        self.bce = nn.BCEWithLogitsLoss()
+
+    def compute_loss(self, pred, target):
+        """Define your loss computation."""
+        # Example: MSE for joints, BCE for grippers
+        joints_loss = self.mse(pred[:, :, :7], target[:, :, :7])
+        gripper_loss = self.bce(pred[:, :, self.gripper_idx],
+                                target[:, :, self.gripper_idx])
+
+        return {
+            "joints_loss": joints_loss,
+            "gripper_loss": gripper_loss,
+        }
+
+    def preprocess(self, proprio, action, mode="train"):
+        """Preprocess actions before training."""
+        # Example: Zero out grippers in proprioception
+        proprio_m = proprio.clone()
+        action_m = action.clone() if action is not None else None
+        proprio_m[..., self.gripper_idx] = 0.0
+        if action_m is not None:
+            action_m[..., self.gripper_idx] = 0.0
+        return proprio_m, action_m
+
+    def postprocess(self, action):
+        """Post-process predictions for deployment."""
+        # Example: Apply sigmoid to gripper logits
+        action[..., self.gripper_idx] = torch.sigmoid(action[..., self.gripper_idx])
+        return action
+```
+
+### Step 2: Use Your Custom Action Mode
+
+```bash
+lerobot-train \
+  --policy.action_mode=my_custom_robot \
+  --dataset.repo_id=YOUR_DATASET \
+  --policy.path="lerobot/xvla-base" \
+  ...
+```
+
+## Advanced Topics
+
+### Multi-Camera Support
+
+X-VLA supports multiple camera views through the `num_image_views` parameter:
+
+```python
+# Configure for 3 camera views
+policy.num_image_views=3
+
+# Add empty cameras if you have fewer physical cameras
+policy.empty_cameras=1  # Adds 1 zero-padded camera view
+```
+
+### Custom Preprocessing Pipeline
+
+Create a custom preprocessing pipeline for your environment:
+
+```python
+from lerobot.processor import PolicyProcessorPipeline
+from lerobot.policies.xvla.processor_xvla import (
+    XVLAImageToFloatProcessorStep,
+    XVLAImageNetNormalizeProcessorStep,
+    XVLAAddDomainIdProcessorStep,
+)
+
+# Build custom pipeline
+preprocessor = PolicyProcessorPipeline(
+    steps=[
+        YourCustomProcessorStep(),  # Your custom processing
+        XVLAImageToFloatProcessorStep(),  # Required: convert to float
+        XVLAImageNetNormalizeProcessorStep(),  # Required: ImageNet norm
+        XVLAAddDomainIdProcessorStep(domain_id=5),  # Your domain ID
+    ]
+)
+```
+
+### Handling Different Action Dimensions
+
+When your dataset has fewer action dimensions than the pretrained model:
+
+**Option 1 (Recommended)**: Use `auto` action mode
+
+```bash
+# Automatically detects your dataset's action dimension
+# Works with any robot without custom code
+policy.action_mode=auto
+policy.max_action_dim=20  # Match pretrained model
+```
+
+**Option 2**: Use a predefined action mode with built-in padding
+
+```python
+# Model expects 20D, dataset has 12D
+# Action mode handles padding internally
+action_mode = "so101_bimanual"  # Pads 12 → 20
+```
+
+**Option 2**: Create a custom action mode that maps dimensions explicitly
+
+```python
+@register_action("my_mapped_action")
+class MappedActionSpace(BaseActionSpace):
+    dim_action = 20
+    REAL_DIM = 12
+
+    def _pad_to_model_dim(self, x):
+        # Custom padding logic
+        ...
+```
+
+## Troubleshooting
+
+### Common Issues
+
+**Issue**: "Action dimension mismatch"
+
+- **Solution**: Check that your `action_mode` matches your robot's action space. Create a custom action mode if needed.
+
+**Issue**: "Image values outside [0, 1] range"
+
+- **Solution**: Ensure images are preprocessed with `XVLAImageToFloatProcessorStep` before normalization.
+
+**Issue**: "Domain ID not found"
+
+- **Solution**: Make sure `XVLAAddDomainIdProcessorStep` is in your preprocessing pipeline with the correct domain_id.
+
+**Issue**: "Low success rate on new embodiment"
+
+- **Solution**:
+  1. Verify your action_mode is correct
+  2. Check that soft prompts are being trained (`train_soft_prompts=True`)
+  3. Ensure proper preprocessing (ImageNet normalization, domain_id)
+  4. Consider increasing training steps
+
+**Issue**: "Out of memory during training"
+
+- **Solution**:
+  1. Reduce `chunk_size` (e.g., from 32 to 16)
+  2. Enable gradient checkpointing
+  3. Reduce batch size
+  4. Freeze more components
+
+## Citation
+
+If you use X-VLA in your research, please cite:
+
+```bibtex
+@article{zheng2025x,
+  title   = {X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model},
+  author  = {Zheng, Jinliang and Li, Jianxiong and Wang, Zhihao and Liu, Dongxiu and Kang, Xirui
+             and Feng, Yuchun and Zheng, Yinan and Zou, Jiayin and Chen, Yilun and Zeng, Jia and others},
+  journal = {arXiv preprint arXiv:2510.10274},
+  year    = {2025}
+}
+```
+
+## Additional Resources
+
+- [X-VLA Paper](https://arxiv.org/pdf/2510.10274)
+- [LeRobot Documentation](https://github.com/huggingface/lerobot)
+- [Action Registry Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/action_hub.py)
+- [Processor Implementation](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/processor_xvla.py)
+- [Model Configuration](https://github.com/huggingface/lerobot/src/lerobot/policies/xvla/configuration_xvla.py)
+
+## Contributing
+
+We welcome contributions! If you've implemented a new action mode or processor for your robot, please consider submitting a PR to help the community.

examples/backward_compatibility/replay.py

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

examples/dataset/load_lerobot_dataset.py

@@ -34,105 +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/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/port_datasets/slurm_aggregate_shards.py

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

examples/port_datasets/slurm_port_shards.py

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

examples/port_datasets/slurm_upload.py

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

examples/rtc/eval_dataset.py

@@ -0,0 +1,951 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Evaluate Real-Time Chunking (RTC) performance on dataset samples.
+
+This script takes two random samples from a dataset:
+- Uses actions from the first sample as previous chunk
+- Generates new actions for the second sample with and without RTC
+
+It compares action predictions with and without RTC on dataset samples,
+measuring consistency and ground truth alignment.
+
+Usage:
+    # Basic usage with smolvla policy
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=mps \
+        --rtc.max_guidance_weight=10.0 \
+        --rtc.prefix_attention_schedule=EXP \
+        --seed=10
+
+    # Basic usage with pi0.5 policy
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi05_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=10 \
+        --device=mps
+        --seed=10
+
+    # Basic usage with pi0.5 policy with cuda device
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi05_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    # Basic usage with pi0 policy with cuda device
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lerobot/pi0_libero_finetuned \
+        --dataset.repo_id=HuggingFaceVLA/libero \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=lipsop/reuben_pi0 \
+        --dataset.repo_id=ReubenLim/so101_cube_in_cup \
+        --rtc.execution_horizon=8 \
+        --device=cuda
+
+    # With torch.compile for faster inference (PyTorch 2.0+)
+    # Note: CUDA graphs disabled by default due to in-place ops in denoising loop
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=mps \
+        --use_torch_compile=true \
+        --torch_compile_mode=max-autotune
+
+    # With torch.compile on CUDA (CUDA graphs disabled by default)
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --rtc.execution_horizon=8 \
+        --device=cuda \
+        --use_torch_compile=true \
+        --torch_compile_mode=reduce-overhead
+
+    # Enable CUDA graphs (advanced - may cause tensor aliasing errors)
+    uv run python examples/rtc/eval_dataset.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --dataset.repo_id=helper2424/check_rtc \
+        --use_torch_compile=true \
+        --torch_compile_backend=inductor \
+        --torch_compile_mode=max-autotune \
+        --torch_compile_disable_cudagraphs=false
+"""
+
+import gc
+import logging
+import os
+import random
+from dataclasses import dataclass, field
+
+import numpy as np
+import torch
+
+try:
+    import matplotlib.pyplot as plt
+
+    MATPLOTLIB_AVAILABLE = True
+except ImportError:
+    MATPLOTLIB_AVAILABLE = False
+    plt = None
+
+from lerobot.configs import parser
+from lerobot.configs.default import DatasetConfig
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.factory import resolve_delta_timestamps
+from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.debug_visualizer import RTCDebugVisualizer
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging
+
+
+def set_seed(seed: int):
+    """Set random seed for reproducibility."""
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+        torch.cuda.manual_seed_all(seed)
+    if torch.backends.mps.is_available():
+        torch.mps.manual_seed(seed)
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+
+
+def _check_matplotlib_available():
+    """Check if matplotlib is available, raise helpful error if not."""
+    if not MATPLOTLIB_AVAILABLE:
+        raise ImportError(
+            "matplotlib is required for RTC debug visualizations. "
+            "Please install it by running:\n"
+            "  uv pip install matplotlib"
+        )
+
+
+@dataclass
+class RTCEvalConfig(HubMixin):
+    """Configuration for RTC evaluation."""
+
+    # Policy configuration
+    policy: PreTrainedConfig | None = None
+
+    # Dataset configuration
+    dataset: DatasetConfig = field(default_factory=DatasetConfig)
+
+    # RTC configuration
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            enabled=True,
+            execution_horizon=20,
+            max_guidance_weight=10.0,
+            prefix_attention_schedule=RTCAttentionSchedule.EXP,
+            debug=True,
+            debug_maxlen=1000,
+        )
+    )
+
+    # Device configuration
+    device: str | None = field(
+        default=None,
+        metadata={"help": "Device to run on (cuda, cpu, mps, auto)"},
+    )
+
+    # Output configuration
+    output_dir: str = field(
+        default="rtc_debug_output",
+        metadata={"help": "Directory to save debug visualizations"},
+    )
+
+    # Seed configuration
+    seed: int = field(
+        default=42,
+        metadata={"help": "Random seed for reproducibility"},
+    )
+
+    inference_delay: int = field(
+        default=4,
+        metadata={"help": "Inference delay for RTC"},
+    )
+
+    # Torch compile configuration
+    use_torch_compile: bool = field(
+        default=False,
+        metadata={"help": "Use torch.compile for faster inference (PyTorch 2.0+)"},
+    )
+
+    torch_compile_backend: str = field(
+        default="inductor",
+        metadata={"help": "Backend for torch.compile (inductor, aot_eager, cudagraphs)"},
+    )
+
+    torch_compile_mode: str = field(
+        default="default",
+        metadata={"help": "Compilation mode (default, reduce-overhead, max-autotune)"},
+    )
+
+    torch_compile_disable_cudagraphs: bool = field(
+        default=True,
+        metadata={
+            "help": "Disable CUDA graphs in torch.compile. Required due to in-place tensor "
+            "operations in denoising loop (x_t += dt * v_t) which cause tensor aliasing issues."
+        },
+    )
+
+    def __post_init__(self):
+        # Parse policy path
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        else:
+            raise ValueError("Policy path is required (--policy.path)")
+
+        # Auto-detect device if not specified
+        if self.device is None or self.device == "auto":
+            if torch.cuda.is_available():
+                self.device = "cuda"
+            elif torch.backends.mps.is_available():
+                self.device = "mps"
+            else:
+                self.device = "cpu"
+            logging.info(f"Auto-detected device: {self.device}")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
+        return ["policy"]
+
+
+class RTCEvaluator:
+    """Evaluator for RTC on dataset samples."""
+
+    def __init__(self, cfg: RTCEvalConfig):
+        self.cfg = cfg
+        self.device = cfg.device
+
+        # Load dataset with proper delta_timestamps based on policy configuration
+        # Calculate delta_timestamps using the same logic as make_dataset factory
+        logging.info(f"Loading dataset: {cfg.dataset.repo_id}")
+
+        # Get dataset metadata to extract FPS
+        ds_meta = LeRobotDatasetMetadata(cfg.dataset.repo_id)
+
+        # Calculate delta_timestamps from policy's delta_indices
+        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
+
+        # Create dataset with calculated delta_timestamps
+        self.dataset = LeRobotDataset(
+            cfg.dataset.repo_id,
+            delta_timestamps=delta_timestamps,
+        )
+        logging.info(f"Dataset loaded: {len(self.dataset)} samples, {self.dataset.num_episodes} episodes")
+
+        # Create preprocessor/postprocessor
+        self.preprocessor, self.postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            preprocessor_overrides={
+                "device_processor": {"device": self.device},
+            },
+        )
+
+        logging.info("=" * 80)
+        logging.info("Ready to run evaluation with sequential policy loading:")
+        logging.info("  1. policy_prev_chunk - Generate reference chunk, then destroy")
+        logging.info("  2. policy_no_rtc - Generate without RTC, then destroy")
+        logging.info("  3. policy_rtc - Generate with RTC, then destroy")
+        logging.info("  Note: Only one policy in memory at a time for efficient memory usage")
+        logging.info("=" * 80)
+
+    def _init_policy(self, name: str, rtc_enabled: bool, rtc_debug: bool):
+        """Initialize a single policy instance with specified RTC configuration.
+
+        Args:
+            name: Name identifier for logging purposes
+            rtc_enabled: Whether to enable RTC for this policy
+            rtc_debug: Whether to enable debug tracking for this policy
+
+        Returns:
+            Configured policy instance with optional torch.compile applied
+        """
+        logging.info(f"Initializing {name}...")
+
+        # Load policy from pretrained
+        policy_class = get_policy_class(self.cfg.policy.type)
+
+        config = PreTrainedConfig.from_pretrained(self.cfg.policy.pretrained_path)
+
+        if self.cfg.policy.type == "pi05" or self.cfg.policy.type == "pi0":
+            config.compile_model = self.cfg.use_torch_compile
+
+        policy = policy_class.from_pretrained(self.cfg.policy.pretrained_path, config=config)
+        policy = policy.to(self.device)
+        policy.eval()
+
+        # Configure RTC
+        rtc_config = RTCConfig(
+            enabled=rtc_enabled,
+            execution_horizon=self.cfg.rtc.execution_horizon,
+            max_guidance_weight=self.cfg.rtc.max_guidance_weight,
+            prefix_attention_schedule=self.cfg.rtc.prefix_attention_schedule,
+            debug=rtc_debug,
+            debug_maxlen=self.cfg.rtc.debug_maxlen,
+        )
+        policy.config.rtc_config = rtc_config
+        policy.init_rtc_processor()
+
+        logging.info(f"  RTC enabled: {rtc_enabled}")
+        logging.info(f"  RTC debug: {rtc_debug}")
+        logging.info(f"  Policy config: {config}")
+
+        # Apply torch.compile to predict_action_chunk method if enabled
+        if self.cfg.use_torch_compile:
+            policy = self._apply_torch_compile(policy, name)
+
+        logging.info(f"✓ {name} initialized successfully")
+        return policy
+
+    def _apply_torch_compile(self, policy, policy_name: str):
+        """Apply torch.compile to the policy's predict_action_chunk method.
+
+        Args:
+            policy: Policy instance to compile
+            policy_name: Name for logging purposes
+
+        Returns:
+            Policy with compiled predict_action_chunk method
+        """
+
+        # PI models handle their own compilation
+        if policy.type == "pi05" or policy.type == "pi0":
+            return policy
+
+        try:
+            # Check if torch.compile is available (PyTorch 2.0+)
+            if not hasattr(torch, "compile"):
+                logging.warning(
+                    f"  [{policy_name}] torch.compile is not available. Requires PyTorch 2.0+. "
+                    f"Current version: {torch.__version__}. Skipping compilation."
+                )
+                return policy
+
+            logging.info(f"  [{policy_name}] Applying torch.compile to predict_action_chunk...")
+            logging.info(f"    Backend: {self.cfg.torch_compile_backend}")
+            logging.info(f"    Mode: {self.cfg.torch_compile_mode}")
+            logging.info(f"    Disable CUDA graphs: {self.cfg.torch_compile_disable_cudagraphs}")
+            logging.info("    Note: Debug tracker excluded from compilation via @torch._dynamo.disable")
+
+            # Compile the predict_action_chunk method
+            # - Debug tracker is excluded from compilation via @torch._dynamo.disable
+            # - CUDA graphs disabled to prevent tensor aliasing from in-place ops (x_t += dt * v_t)
+            compile_kwargs = {
+                "backend": self.cfg.torch_compile_backend,
+                "mode": self.cfg.torch_compile_mode,
+            }
+
+            # Disable CUDA graphs if requested (prevents tensor aliasing issues)
+            if self.cfg.torch_compile_disable_cudagraphs:
+                compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+            original_method = policy.predict_action_chunk
+            compiled_method = torch.compile(original_method, **compile_kwargs)
+            policy.predict_action_chunk = compiled_method
+            logging.info(f"  ✓ [{policy_name}] Successfully compiled predict_action_chunk")
+
+        except Exception as e:
+            logging.error(f"  [{policy_name}] Failed to apply torch.compile: {e}")
+            logging.warning(f"  [{policy_name}] Continuing without torch.compile")
+
+        return policy
+
+    def _destroy_policy(self, policy, policy_name: str):
+        """Explicitly destroy a policy and free all associated memory.
+
+        This method performs aggressive cleanup to ensure maximum memory is freed,
+        which is critical for large models (e.g., VLAs with billions of parameters).
+
+        Args:
+            policy: Policy instance to destroy
+            policy_name: Name for logging purposes
+        """
+        logging.info(f"  Destroying {policy_name} and freeing memory...")
+
+        try:
+            # Step 1: Move policy to CPU to free GPU/MPS memory
+            policy.cpu()
+
+            # Step 2: Delete the policy object
+            del policy
+
+            # Step 3: Force garbage collection to reclaim memory immediately
+            gc.collect()
+
+            # Step 4: Clear device-specific caches
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+                torch.cuda.synchronize()  # Ensure all operations complete
+
+            if torch.backends.mps.is_available():
+                torch.mps.empty_cache()
+
+            logging.info(f"  ✓ {policy_name} destroyed and memory freed")
+
+        except Exception as e:
+            logging.warning(f"  Warning: Error during {policy_name} cleanup: {e}")
+
+    def run_evaluation(self):
+        """Run evaluation on two random dataset samples using three separate policies.
+
+        Note: Policies are deinitalized after each step to free memory. Large models
+        (e.g., VLA models with billions of parameters) cannot fit three instances in
+        memory simultaneously. By deleting and garbage collecting after each step,
+        we ensure only one policy is loaded at a time.
+        """
+        # Create output directory
+        os.makedirs(self.cfg.output_dir, exist_ok=True)
+        logging.info(f"Output directory: {self.cfg.output_dir}")
+
+        logging.info("=" * 80)
+        logging.info("Starting RTC evaluation")
+        logging.info(f"Inference delay: {self.cfg.inference_delay}")
+        logging.info("=" * 80)
+
+        # Load two random samples from dataset
+        data_loader = torch.utils.data.DataLoader(self.dataset, batch_size=1, shuffle=True)
+        loader_iter = iter(data_loader)
+        first_sample = next(loader_iter)
+        second_sample = next(loader_iter)
+
+        preprocessed_first_sample = self.preprocessor(first_sample)
+        preprocessed_second_sample = self.preprocessor(second_sample)
+
+        # ============================================================================
+        # Step 1: Generate previous chunk using policy_prev_chunk
+        # ============================================================================
+        # This policy is only used to generate the reference chunk and then freed
+        logging.info("=" * 80)
+        logging.info("Step 1: Generating previous chunk with policy_prev_chunk")
+        logging.info("=" * 80)
+
+        # Initialize policy 1
+        policy_prev_chunk_policy = self._init_policy(
+            name="policy_prev_chunk",
+            rtc_enabled=False,
+            rtc_debug=False,
+        )
+        with torch.no_grad():
+            prev_chunk_left_over = policy_prev_chunk_policy.predict_action_chunk(
+                preprocessed_first_sample,
+            )[:, :25, :].squeeze(0)
+        logging.info(f"  Generated prev_chunk shape: {prev_chunk_left_over.shape}")
+
+        # Destroy policy_prev_chunk to free memory for large models
+        self._destroy_policy(policy_prev_chunk_policy, "policy_prev_chunk")
+
+        # ============================================================================
+        # Step 2: Generate actions WITHOUT RTC using policy_no_rtc
+        # ============================================================================
+        logging.info("=" * 80)
+        logging.info("Step 2: Generating actions WITHOUT RTC with policy_no_rtc")
+        logging.info("=" * 80)
+
+        set_seed(self.cfg.seed)
+
+        # Initialize policy 2
+        policy_no_rtc_policy = self._init_policy(
+            name="policy_no_rtc",
+            rtc_enabled=False,
+            rtc_debug=True,
+        )
+
+        # Sample noise (use same noise for both RTC and non-RTC for fair comparison)
+        noise_size = (1, policy_no_rtc_policy.config.chunk_size, policy_no_rtc_policy.config.max_action_dim)
+        noise = policy_no_rtc_policy.model.sample_noise(noise_size, self.device)
+        noise_clone = noise.clone()
+        policy_no_rtc_policy.rtc_processor.reset_tracker()
+        with torch.no_grad():
+            no_rtc_actions = policy_no_rtc_policy.predict_action_chunk(
+                preprocessed_second_sample,
+                noise=noise,
+            )
+        no_rtc_tracked_steps = policy_no_rtc_policy.rtc_processor.tracker.get_all_steps()
+        logging.info(f"  Tracked {len(no_rtc_tracked_steps)} steps without RTC")
+        logging.info(f"  Generated no_rtc_actions shape: {no_rtc_actions.shape}")
+
+        # Destroy policy_no_rtc to free memory before loading policy_rtc
+        self._destroy_policy(policy_no_rtc_policy, "policy_no_rtc")
+
+        # ============================================================================
+        # Step 3: Generate actions WITH RTC using policy_rtc
+        # ============================================================================
+        logging.info("=" * 80)
+        logging.info("Step 3: Generating actions WITH RTC with policy_rtc")
+        logging.info("=" * 80)
+
+        set_seed(self.cfg.seed)
+
+        # Initialize policy 3
+        policy_rtc_policy = self._init_policy(
+            name="policy_rtc",
+            rtc_enabled=True,
+            rtc_debug=True,
+        )
+        policy_rtc_policy.rtc_processor.reset_tracker()
+        with torch.no_grad():
+            rtc_actions = policy_rtc_policy.predict_action_chunk(
+                preprocessed_second_sample,
+                noise=noise_clone,
+                inference_delay=self.cfg.inference_delay,
+                prev_chunk_left_over=prev_chunk_left_over,
+                execution_horizon=self.cfg.rtc.execution_horizon,
+            )
+        rtc_tracked_steps = policy_rtc_policy.rtc_processor.get_all_debug_steps()
+        logging.info(f"  Tracked {len(rtc_tracked_steps)} steps with RTC")
+        logging.info(f"  Generated rtc_actions shape: {rtc_actions.shape}")
+
+        # Save num_steps before destroying policy (needed for plotting)
+        try:
+            num_steps = policy_rtc_policy.config.num_steps
+        except Exception as e:
+            logging.error(f"  Error getting num_steps: {e}")
+            num_steps = policy_rtc_policy.config.num_inference_steps
+            logging.warning(f"  Using num_inference_steps: {num_steps} instead of num_steps")
+
+        # Destroy policy_rtc after final use
+        self._destroy_policy(policy_rtc_policy, "policy_rtc")
+
+        # Plot and save results
+        logging.info("=" * 80)
+        logging.info("Plotting results...")
+        self.plot_tracked_data(rtc_tracked_steps, no_rtc_tracked_steps, prev_chunk_left_over, num_steps)
+
+        # Plot final actions comparison
+        logging.info("=" * 80)
+        logging.info("Plotting final actions comparison...")
+        self.plot_final_actions_comparison(rtc_actions, no_rtc_actions, prev_chunk_left_over)
+
+        logging.info("=" * 80)
+        logging.info("Evaluation completed successfully")
+
+    def plot_final_actions_comparison(self, rtc_actions, no_rtc_actions, prev_chunk_left_over):
+        """Plot final action predictions comparison on a single chart.
+
+        Args:
+            rtc_actions: Final actions from RTC policy
+            no_rtc_actions: Final actions from non-RTC policy
+            prev_chunk_left_over: Previous chunk used as ground truth
+        """
+        _check_matplotlib_available()
+
+        # Remove batch dimension if present
+        rtc_actions_plot = rtc_actions.squeeze(0).cpu() if len(rtc_actions.shape) == 3 else rtc_actions.cpu()
+        no_rtc_actions_plot = (
+            no_rtc_actions.squeeze(0).cpu() if len(no_rtc_actions.shape) == 3 else no_rtc_actions.cpu()
+        )
+        prev_chunk_plot = prev_chunk_left_over.cpu()
+
+        # Create figure with 6 subplots (one per action dimension)
+        fig, axes = plt.subplots(6, 1, figsize=(16, 12))
+        fig.suptitle("Final Action Predictions Comparison (Raw)", fontsize=16)
+
+        # Plot each action dimension
+        for dim_idx, ax in enumerate(axes):
+            # Plot previous chunk (ground truth) in red
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                prev_chunk_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="red",
+                label="Previous Chunk (Ground Truth)",
+                linewidth=2.5,
+                alpha=0.8,
+            )
+
+            # Plot no-RTC actions in blue
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                no_rtc_actions_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="blue",
+                label="No RTC",
+                linewidth=2,
+                alpha=0.7,
+            )
+
+            # Plot RTC actions in green
+            RTCDebugVisualizer.plot_waypoints(
+                [ax],
+                rtc_actions_plot[:, dim_idx : dim_idx + 1],
+                start_from=0,
+                color="green",
+                label="RTC",
+                linewidth=2,
+                alpha=0.7,
+            )
+
+            # Add vertical lines for inference delay and execution horizon
+            inference_delay = self.cfg.inference_delay
+            execution_horizon = self.cfg.rtc.execution_horizon
+
+            if inference_delay > 0:
+                ax.axvline(
+                    x=inference_delay - 1,
+                    color="orange",
+                    linestyle="--",
+                    alpha=0.5,
+                    label=f"Inference Delay ({inference_delay})",
+                )
+
+            if execution_horizon > 0:
+                ax.axvline(
+                    x=execution_horizon,
+                    color="purple",
+                    linestyle="--",
+                    alpha=0.5,
+                    label=f"Execution Horizon ({execution_horizon})",
+                )
+
+            ax.set_ylabel(f"Dim {dim_idx}", fontsize=10)
+            ax.grid(True, alpha=0.3)
+
+            # Set x-axis ticks to show all integer values
+            max_len = max(rtc_actions_plot.shape[0], no_rtc_actions_plot.shape[0], prev_chunk_plot.shape[0])
+            ax.set_xticks(range(0, max_len, max(1, max_len // 20)))  # Show ~20 ticks
+            ax.set_xlim(-0.5, max_len - 0.5)
+
+        axes[-1].set_xlabel("Step", fontsize=10)
+
+        # Collect legend handles and labels from first subplot
+        handles, labels = axes[0].get_legend_handles_labels()
+        # Remove duplicates while preserving order
+        seen = set()
+        unique_handles = []
+        unique_labels = []
+        for handle, label in zip(handles, labels, strict=True):
+            if label not in seen:
+                seen.add(label)
+                unique_handles.append(handle)
+                unique_labels.append(label)
+
+        # Add legend outside the plot area (to the right)
+        fig.legend(
+            unique_handles,
+            unique_labels,
+            loc="center right",
+            fontsize=9,
+            bbox_to_anchor=(1.0, 0.5),
+            framealpha=0.9,
+        )
+
+        # Save figure
+        output_path = os.path.join(self.cfg.output_dir, "final_actions_comparison.png")
+        fig.tight_layout(rect=[0, 0, 0.85, 1])  # Leave space for legend on right
+        fig.savefig(output_path, dpi=150, bbox_inches="tight")
+        logging.info(f"Saved final actions comparison to {output_path}")
+        plt.close(fig)
+
+    def plot_tracked_data(self, rtc_tracked_steps, no_rtc_tracked_steps, prev_chunk_left_over, num_steps):
+        _check_matplotlib_available()
+
+        # Create side-by-side figures for denoising visualization
+        fig_xt, axs_xt = self._create_figure("x_t Denoising: No RTC (left) vs RTC (right)")
+        fig_vt, axs_vt = self._create_figure("v_t Denoising: No RTC (left) vs RTC (right)")
+        fig_corr, axs_corr = self._create_figure("Correction: No RTC (left) vs RTC (right)")
+        fig_x1t, axs_x1t = self._create_figure(
+            "x1_t Predicted State & Error: No RTC (left - empty) vs RTC (right)"
+        )
+        self._plot_denoising_steps_from_tracker(
+            rtc_tracked_steps,
+            axs_xt[:, 1],  # Right column for x_t
+            axs_vt[:, 1],  # Right column for v_t
+            axs_corr[:, 1],  # Right column for correction
+            axs_x1t[:, 1],  # Right column for x1_t
+            num_steps,
+            add_labels=True,  # Add labels for RTC (right column)
+        )
+
+        self._plot_denoising_steps_from_tracker(
+            no_rtc_tracked_steps,
+            axs_xt[:, 0],  # Left column for x_t
+            axs_vt[:, 0],  # Left column for v_t
+            axs_corr[:, 0],  # Left column for correction
+            axs_x1t[:, 0],  # Left column for x1_t
+            num_steps,
+            add_labels=False,  # No labels for No RTC (left column)
+        )
+
+        # Plot no-RTC x_t data on right chart as orange dashed line for comparison
+        self._plot_no_rtc_xt_reference(no_rtc_tracked_steps, axs_xt[:, 1], num_steps)
+
+        # Plot ground truth on x_t axes
+        RTCDebugVisualizer.plot_waypoints(
+            axs_xt[:, 1], prev_chunk_left_over, start_from=0, color="red", label="Ground truth"
+        )
+
+        # Plot ground truth on x1_t axes
+        RTCDebugVisualizer.plot_waypoints(
+            axs_x1t[:, 1], prev_chunk_left_over, start_from=0, color="red", label="Ground truth"
+        )
+
+        # Plot ground truth on x_t axes (no labels for left column)
+        RTCDebugVisualizer.plot_waypoints(
+            axs_xt[:, 0], prev_chunk_left_over, start_from=0, color="red", label=None
+        )
+
+        RTCDebugVisualizer.plot_waypoints(
+            axs_x1t[:, 0], prev_chunk_left_over, start_from=0, color="red", label=None
+        )
+
+        # Add legends outside the plot area for each figure
+        self._add_figure_legend(fig_xt, axs_xt)
+        self._add_figure_legend(fig_vt, axs_vt)
+        self._add_figure_legend(fig_corr, axs_corr)
+        self._add_figure_legend(fig_x1t, axs_x1t)
+
+        # Save denoising plots
+        self._save_figure(fig_xt, os.path.join(self.cfg.output_dir, "denoising_xt_comparison.png"))
+        self._save_figure(fig_vt, os.path.join(self.cfg.output_dir, "denoising_vt_comparison.png"))
+        self._save_figure(fig_corr, os.path.join(self.cfg.output_dir, "denoising_correction_comparison.png"))
+        self._save_figure(fig_x1t, os.path.join(self.cfg.output_dir, "denoising_x1t_comparison.png"))
+
+    def _create_figure(self, title):
+        fig, axs = plt.subplots(6, 2, figsize=(24, 12))
+        fig.suptitle(title, fontsize=16)
+
+        for ax in axs[:, 0]:
+            ax.set_title("No RTC (N/A)" if ax == axs[0, 0] else "", fontsize=12)
+        for ax in axs[:, 1]:
+            ax.set_title("RTC" if ax == axs[0, 1] else "", fontsize=12)
+
+        return fig, axs
+
+    def _add_figure_legend(self, fig, axs):
+        """Add a legend outside the plot area on the right side.
+
+        Args:
+            fig: Matplotlib figure to add legend to
+            axs: Array of axes to collect legend handles from
+        """
+        # Collect all handles and labels from the first row of axes (right column)
+        handles, labels = axs[0, 1].get_legend_handles_labels()
+
+        # Remove duplicates while preserving order
+        seen = set()
+        unique_handles = []
+        unique_labels = []
+        for handle, label in zip(handles, labels, strict=True):
+            if label not in seen:
+                seen.add(label)
+                unique_handles.append(handle)
+                unique_labels.append(label)
+
+        # Add legend outside the plot area (to the right, close to charts)
+        if unique_handles:
+            fig.legend(
+                unique_handles,
+                unique_labels,
+                loc="center left",
+                fontsize=8,
+                bbox_to_anchor=(0.87, 0.5),
+                framealpha=0.9,
+                ncol=1,
+            )
+
+    def _save_figure(self, fig, path):
+        fig.tight_layout(rect=[0, 0, 0.85, 1])  # Leave space for legend/colorbar on right
+        fig.savefig(path, dpi=150, bbox_inches="tight")
+        logging.info(f"Saved figure to {path}")
+        plt.close(fig)
+
+    def _plot_denoising_steps_from_tracker(
+        self, tracked_steps, xt_axs, vt_axs, corr_axs, x1t_axs, num_steps, add_labels=True
+    ):
+        """Plot denoising steps from tracker data.
+
+        Args:
+            tracked_steps: List of DebugStep objects containing debug steps
+            xt_axs: Matplotlib axes for x_t plots (array of 6 axes)
+            vt_axs: Matplotlib axes for v_t plots (array of 6 axes)
+            corr_axs: Matplotlib axes for correction plots (array of 6 axes)
+            x1t_axs: Matplotlib axes for x1_t plots (array of 6 axes)
+            num_steps: Total number of denoising steps for colormap
+            add_labels: Whether to add legend labels for the plots
+        """
+
+        logging.info("=" * 80)
+        logging.info(f"Plotting {len(tracked_steps)} steps")
+
+        debug_steps = tracked_steps
+        if not debug_steps:
+            return
+
+        # Define colors for different denoise steps (using a colormap)
+        colors = plt.cm.viridis(np.linspace(0, 1, num_steps))
+
+        for step_idx, debug_step in enumerate(debug_steps):
+            color = colors[step_idx % len(colors)]
+            label = f"Step {step_idx}" if add_labels else None
+
+            # Plot x_t
+            if debug_step.x_t is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    xt_axs, debug_step.x_t, start_from=0, color=color, label=label
+                )
+
+            # Plot v_t
+            if debug_step.v_t is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    vt_axs, debug_step.v_t, start_from=0, color=color, label=label
+                )
+
+            # Plot correction on separate axes
+            if debug_step.correction is not None:
+                RTCDebugVisualizer.plot_waypoints(
+                    corr_axs,
+                    debug_step.correction,
+                    start_from=0,
+                    color=color,
+                    label=label,
+                )
+
+            # Plot x1_t (predicted state)
+            if x1t_axs is not None and debug_step.x1_t is not None:
+                x1t_label = f"x1_t Step {step_idx}" if add_labels else None
+                RTCDebugVisualizer.plot_waypoints(
+                    x1t_axs,
+                    debug_step.x1_t,
+                    start_from=0,
+                    color=color,
+                    label=x1t_label,
+                )
+
+            # Plot error in orange dashed
+            if x1t_axs is not None and debug_step.err is not None:
+                error_chunk = (
+                    debug_step.err[0].cpu().numpy()
+                    if len(debug_step.err.shape) == 3
+                    else debug_step.err.cpu().numpy()
+                )
+
+                num_dims = min(error_chunk.shape[-1], 6)
+                error_label = f"error Step {step_idx}" if add_labels else None
+                for j in range(num_dims):
+                    x1t_axs[j].plot(
+                        np.arange(0, error_chunk.shape[0]),
+                        error_chunk[:, j],
+                        color="orange",
+                        linestyle="--",
+                        alpha=0.7,
+                        label=error_label,
+                    )
+
+        # Recalculate axis limits after plotting to ensure proper scaling
+        self._rescale_axes(xt_axs)
+        self._rescale_axes(vt_axs)
+        self._rescale_axes(corr_axs)
+        self._rescale_axes(x1t_axs)
+
+    def _plot_no_rtc_xt_reference(self, no_rtc_tracked_steps, xt_axs, num_steps):
+        """Plot final no-RTC x_t data as orange dashed line on the RTC chart for comparison.
+
+        Args:
+            no_rtc_tracked_steps: List of DebugStep objects containing no-RTC debug steps
+            xt_axs: Matplotlib axes for x_t plots (array of 6 axes, right column)
+            num_steps: Total number of denoising steps for colormap
+        """
+        debug_steps = no_rtc_tracked_steps
+        if not debug_steps:
+            return
+
+        # Plot only the final x_t step as orange dashed line
+        final_step = debug_steps[-1]
+        logging.info("Plotting final no-RTC x_t step as orange dashed reference")
+
+        if final_step.x_t is not None:
+            x_t_chunk = (
+                final_step.x_t[0].cpu().numpy()
+                if len(final_step.x_t.shape) == 3
+                else final_step.x_t.cpu().numpy()
+            )
+
+            num_dims = min(x_t_chunk.shape[-1], 6)
+            for j in range(num_dims):
+                xt_axs[j].plot(
+                    np.arange(0, x_t_chunk.shape[0]),
+                    x_t_chunk[:, j],
+                    color="orange",
+                    linestyle="--",
+                    alpha=0.7,
+                    linewidth=2,
+                    label="No RTC (final)" if j == 0 else "",
+                )
+
+    def _rescale_axes(self, axes):
+        """Rescale axes to show all data with proper margins.
+
+        Args:
+            axes: Array of matplotlib axes to rescale
+        """
+        for ax in axes:
+            ax.relim()
+            ax.autoscale_view()
+
+            # Add 10% margin to y-axis for better visualization
+            ylim = ax.get_ylim()
+            y_range = ylim[1] - ylim[0]
+            if y_range > 0:  # Avoid division by zero
+                margin = y_range * 0.1
+                ax.set_ylim(ylim[0] - margin, ylim[1] + margin)
+
+            # Set x-axis ticks to show all integer values
+            xlim = ax.get_xlim()
+            max_len = int(xlim[1]) + 1
+            if max_len > 0:
+                ax.set_xticks(range(0, max_len, max(1, max_len // 20)))  # Show ~20 ticks
+                ax.set_xlim(-0.5, max_len - 0.5)
+
+
+@parser.wrap()
+def main(cfg: RTCEvalConfig):
+    """Main entry point for RTC evaluation."""
+    # Set random seed for reproducibility
+    set_seed(cfg.seed)
+
+    init_logging()
+
+    logging.info("=" * 80)
+    logging.info("RTC Dataset Evaluation")
+    logging.info(f"Config: {cfg}")
+    logging.info("=" * 80)
+
+    evaluator = RTCEvaluator(cfg)
+    evaluator.run_evaluation()
+
+
+if __name__ == "__main__":
+    main()

examples/rtc/eval_with_real_robot.py

@@ -0,0 +1,549 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Demo script showing how to use Real-Time Chunking (RTC) with action chunking policies on real robots.
+
+This script demonstrates:
+1. Creating a robot and policy (SmolVLA, Pi0, etc.) with RTC
+2. Consuming actions from the policy while the robot executes
+3. Periodically requesting new action chunks in the background using threads
+4. Managing action buffers and timing for real-time operation
+
+For simulation environments, see eval_with_simulation.py
+
+Usage:
+    # Run RTC with Real robot with RTC
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.device=mps \
+        --rtc.enabled=true \
+        --rtc.execution_horizon=20 \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+
+    # Run RTC with Real robot without RTC
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/smolvla_check_rtc_last3 \
+        --policy.device=mps \
+        --rtc.enabled=false \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+
+    # Run RTC with Real robot with pi0.5 policy
+    uv run examples/rtc/eval_with_real_robot.py \
+        --policy.path=helper2424/pi05_check_rtc \
+        --policy.device=mps \
+        --rtc.enabled=true \
+        --rtc.execution_horizon=20 \
+        --robot.type=so100_follower \
+        --robot.port=/dev/tty.usbmodem58FA0834591 \
+        --robot.id=so100_follower \
+        --robot.cameras="{ gripper: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, front: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}}" \
+        --task="Move green small object into the purple platform" \
+        --duration=120
+"""
+
+import logging
+import math
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+from threading import Event, Lock, Thread
+
+import torch
+from torch import Tensor
+
+from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.datasets.utils import build_dataset_frame, hw_to_dataset_features
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.action_queue import ActionQueue
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.latency_tracker import LatencyTracker
+from lerobot.processor.factory import (
+    make_default_robot_action_processor,
+    make_default_robot_observation_processor,
+)
+from lerobot.rl.process import ProcessSignalHandler
+from lerobot.robots import (  # noqa: F401
+    Robot,
+    RobotConfig,
+    koch_follower,
+    so100_follower,
+    so101_follower,
+)
+from lerobot.robots.utils import make_robot_from_config
+from lerobot.utils.constants import OBS_IMAGES
+from lerobot.utils.hub import HubMixin
+from lerobot.utils.utils import init_logging
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+class RobotWrapper:
+    def __init__(self, robot: Robot):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with self.lock:
+            return self.robot.get_observation()
+
+    def send_action(self, action: Tensor):
+        with self.lock:
+            self.robot.send_action(action)
+
+    def observation_features(self) -> list[str]:
+        with self.lock:
+            return self.robot.observation_features
+
+    def action_features(self) -> list[str]:
+        with self.lock:
+            return self.robot.action_features
+
+
+@dataclass
+class RTCDemoConfig(HubMixin):
+    """Configuration for RTC demo with action chunking policies and real robots."""
+
+    # Policy configuration
+    policy: PreTrainedConfig | None = None
+
+    # Robot configuration
+    robot: RobotConfig | None = None
+
+    # RTC configuration
+    rtc: RTCConfig = field(
+        default_factory=lambda: RTCConfig(
+            execution_horizon=10,
+            max_guidance_weight=1.0,
+            prefix_attention_schedule=RTCAttentionSchedule.EXP,
+        )
+    )
+
+    # Demo parameters
+    duration: float = 30.0  # Duration to run the demo (seconds)
+    fps: float = 10.0  # Action execution frequency (Hz)
+
+    # Compute device
+    device: str | None = None  # Device to run on (cuda, cpu, auto)
+
+    # Get new actions horizon. The amount of executed steps after which will be requested new actions.
+    # It should be higher than inference delay + execution horizon.
+    action_queue_size_to_get_new_actions: int = 30
+
+    # Task to execute
+    task: str = field(default="", metadata={"help": "Task to execute"})
+
+    # Torch compile configuration
+    use_torch_compile: bool = field(
+        default=False,
+        metadata={"help": "Use torch.compile for faster inference (PyTorch 2.0+)"},
+    )
+
+    torch_compile_backend: str = field(
+        default="inductor",
+        metadata={"help": "Backend for torch.compile (inductor, aot_eager, cudagraphs)"},
+    )
+
+    torch_compile_mode: str = field(
+        default="default",
+        metadata={"help": "Compilation mode (default, reduce-overhead, max-autotune)"},
+    )
+
+    torch_compile_disable_cudagraphs: bool = field(
+        default=True,
+        metadata={
+            "help": "Disable CUDA graphs in torch.compile. Required due to in-place tensor "
+            "operations in denoising loop (x_t += dt * v_t) which cause tensor aliasing issues."
+        },
+    )
+
+    def __post_init__(self):
+        # HACK: We parse again the cli args here to get the pretrained path if there was one.
+        policy_path = parser.get_path_arg("policy")
+        if policy_path:
+            cli_overrides = parser.get_cli_overrides("policy")
+            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            self.policy.pretrained_path = policy_path
+        else:
+            raise ValueError("Policy path is required")
+
+        # Validate that robot configuration is provided
+        if self.robot is None:
+            raise ValueError("Robot configuration must be provided")
+
+    @classmethod
+    def __get_path_fields__(cls) -> list[str]:
+        """This enables the parser to load config from the policy using `--policy.path=local/dir`"""
+        return ["policy"]
+
+
+def is_image_key(k: str) -> bool:
+    return k.startswith(OBS_IMAGES)
+
+
+def get_actions(
+    policy,
+    robot: RobotWrapper,
+    robot_observation_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: RTCDemoConfig,
+):
+    """Thread function to request action chunks from the policy.
+
+    Args:
+        policy: The policy instance (SmolVLA, Pi0, etc.)
+        robot: The robot instance for getting observations
+        robot_observation_processor: Processor for raw robot observations
+        action_queue: Queue to put new action chunks
+        shutdown_event: Event to signal shutdown
+        cfg: Demo configuration
+    """
+    try:
+        logger.info("[GET_ACTIONS] Starting get actions thread")
+
+        latency_tracker = LatencyTracker()  # Track latency of action chunks
+        fps = cfg.fps
+        time_per_chunk = 1.0 / fps
+
+        dataset_features = hw_to_dataset_features(robot.observation_features(), "observation")
+        policy_device = policy.config.device
+
+        # Load preprocessor and postprocessor from pretrained files
+        # The stats are embedded in the processor .safetensors files
+        logger.info(f"[GET_ACTIONS] Loading preprocessor/postprocessor from {cfg.policy.pretrained_path}")
+
+        preprocessor, postprocessor = make_pre_post_processors(
+            policy_cfg=cfg.policy,
+            pretrained_path=cfg.policy.pretrained_path,
+            dataset_stats=None,  # Will load from pretrained processor files
+            preprocessor_overrides={
+                "device_processor": {"device": cfg.policy.device},
+            },
+        )
+
+        logger.info("[GET_ACTIONS] Preprocessor/postprocessor loaded successfully with embedded stats")
+
+        get_actions_threshold = cfg.action_queue_size_to_get_new_actions
+
+        if not cfg.rtc.enabled:
+            get_actions_threshold = 0
+
+        while not shutdown_event.is_set():
+            if action_queue.qsize() <= get_actions_threshold:
+                current_time = time.perf_counter()
+                action_index_before_inference = action_queue.get_action_index()
+                prev_actions = action_queue.get_left_over()
+
+                inference_latency = latency_tracker.max()
+                inference_delay = math.ceil(inference_latency / time_per_chunk)
+
+                obs = robot.get_observation()
+
+                # Apply robot observation processor
+                obs_processed = robot_observation_processor(obs)
+
+                obs_with_policy_features = build_dataset_frame(
+                    dataset_features, obs_processed, prefix="observation"
+                )
+
+                for name in obs_with_policy_features:
+                    obs_with_policy_features[name] = torch.from_numpy(obs_with_policy_features[name])
+                    if "image" in name:
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].type(torch.float32) / 255
+                        )
+                        obs_with_policy_features[name] = (
+                            obs_with_policy_features[name].permute(2, 0, 1).contiguous()
+                        )
+                    obs_with_policy_features[name] = obs_with_policy_features[name].unsqueeze(0)
+                    obs_with_policy_features[name] = obs_with_policy_features[name].to(policy_device)
+
+                obs_with_policy_features["task"] = [cfg.task]  # Task should be a list, not a string!
+                obs_with_policy_features["robot_type"] = (
+                    robot.robot.name if hasattr(robot.robot, "name") else ""
+                )
+
+                preproceseded_obs = preprocessor(obs_with_policy_features)
+
+                # Generate actions WITH RTC
+                actions = policy.predict_action_chunk(
+                    preproceseded_obs,
+                    inference_delay=inference_delay,
+                    prev_chunk_left_over=prev_actions,
+                )
+
+                # Store original actions (before postprocessing) for RTC
+                original_actions = actions.squeeze(0).clone()
+
+                postprocessed_actions = postprocessor(actions)
+
+                postprocessed_actions = postprocessed_actions.squeeze(0)
+
+                new_latency = time.perf_counter() - current_time
+                new_delay = math.ceil(new_latency / time_per_chunk)
+                latency_tracker.add(new_latency)
+
+                if cfg.action_queue_size_to_get_new_actions < cfg.rtc.execution_horizon + new_delay:
+                    logger.warning(
+                        "[GET_ACTIONS] cfg.action_queue_size_to_get_new_actions Too small, It should be higher than inference delay + execution horizon."
+                    )
+
+                action_queue.merge(
+                    original_actions, postprocessed_actions, new_delay, action_index_before_inference
+                )
+            else:
+                # Small sleep to prevent busy waiting
+                time.sleep(0.1)
+
+        logger.info("[GET_ACTIONS] get actions thread shutting down")
+    except Exception as e:
+        logger.error(f"[GET_ACTIONS] Fatal exception in get_actions thread: {e}")
+        logger.error(traceback.format_exc())
+        sys.exit(1)
+
+
+def actor_control(
+    robot: RobotWrapper,
+    robot_action_processor,
+    action_queue: ActionQueue,
+    shutdown_event: Event,
+    cfg: RTCDemoConfig,
+):
+    """Thread function to execute actions on the robot.
+
+    Args:
+        robot: The robot instance
+        action_queue: Queue to get actions from
+        shutdown_event: Event to signal shutdown
+        cfg: Demo configuration
+    """
+    try:
+        logger.info("[ACTOR] Starting actor thread")
+
+        action_count = 0
+        action_interval = 1.0 / cfg.fps
+
+        while not shutdown_event.is_set():
+            start_time = time.perf_counter()
+
+            # Try to get an action from the queue with timeout
+            action = action_queue.get()
+
+            if action is not None:
+                action = action.cpu()
+                action_dict = {key: action[i].item() for i, key in enumerate(robot.action_features())}
+                action_processed = robot_action_processor((action_dict, None))
+                robot.send_action(action_processed)
+
+                action_count += 1
+
+            dt_s = time.perf_counter() - start_time
+            time.sleep(max(0, (action_interval - dt_s) - 0.001))
+
+        logger.info(f"[ACTOR] Actor thread shutting down. Total actions executed: {action_count}")
+    except Exception as e:
+        logger.error(f"[ACTOR] Fatal exception in actor_control thread: {e}")
+        logger.error(traceback.format_exc())
+        sys.exit(1)
+
+
+def _apply_torch_compile(policy, cfg: RTCDemoConfig):
+    """Apply torch.compile to the policy's predict_action_chunk method.
+
+    Args:
+        policy: Policy instance to compile
+        cfg: Configuration containing torch compile settings
+
+    Returns:
+        Policy with compiled predict_action_chunk method
+    """
+
+    # PI models handle their own compilation
+    if policy.type == "pi05" or policy.type == "pi0":
+        return policy
+
+    try:
+        # Check if torch.compile is available (PyTorch 2.0+)
+        if not hasattr(torch, "compile"):
+            logger.warning(
+                f"torch.compile is not available. Requires PyTorch 2.0+. "
+                f"Current version: {torch.__version__}. Skipping compilation."
+            )
+            return policy
+
+        logger.info("Applying torch.compile to predict_action_chunk...")
+        logger.info(f"  Backend: {cfg.torch_compile_backend}")
+        logger.info(f"  Mode: {cfg.torch_compile_mode}")
+        logger.info(f"  Disable CUDA graphs: {cfg.torch_compile_disable_cudagraphs}")
+
+        # Compile the predict_action_chunk method
+        # - CUDA graphs disabled to prevent tensor aliasing from in-place ops (x_t += dt * v_t)
+        compile_kwargs = {
+            "backend": cfg.torch_compile_backend,
+            "mode": cfg.torch_compile_mode,
+        }
+
+        # Disable CUDA graphs if requested (prevents tensor aliasing issues)
+        if cfg.torch_compile_disable_cudagraphs:
+            compile_kwargs["options"] = {"triton.cudagraphs": False}
+
+        original_method = policy.predict_action_chunk
+        compiled_method = torch.compile(original_method, **compile_kwargs)
+        policy.predict_action_chunk = compiled_method
+        logger.info("✓ Successfully compiled predict_action_chunk")
+
+    except Exception as e:
+        logger.error(f"Failed to apply torch.compile: {e}")
+        logger.warning("Continuing without torch.compile")
+
+    return policy
+
+
+@parser.wrap()
+def demo_cli(cfg: RTCDemoConfig):
+    """Main entry point for RTC demo with draccus configuration."""
+
+    # Initialize logging
+    init_logging()
+
+    logger.info(f"Using device: {cfg.device}")
+
+    # Setup signal handler for graceful shutdown
+    signal_handler = ProcessSignalHandler(use_threads=True, display_pid=False)
+    shutdown_event = signal_handler.shutdown_event
+
+    policy = None
+    robot = None
+    get_actions_thread = None
+    actor_thread = None
+
+    policy_class = get_policy_class(cfg.policy.type)
+
+    # Load config and set compile_model for pi0/pi05 models
+    config = PreTrainedConfig.from_pretrained(cfg.policy.pretrained_path)
+
+    if cfg.policy.type == "pi05" or cfg.policy.type == "pi0":
+        config.compile_model = cfg.use_torch_compile
+
+    policy = policy_class.from_pretrained(cfg.policy.pretrained_path, config=config)
+
+    # Turn on RTC
+    policy.config.rtc_config = cfg.rtc
+
+    # Init RTC processort, as by default if RTC disabled in the config
+    # The processor won't be created
+    policy.init_rtc_processor()
+
+    assert policy.name in ["smolvla", "pi05", "pi0"], "Only smolvla, pi05, and pi0 are supported for RTC"
+
+    policy = policy.to(cfg.device)
+    policy.eval()
+
+    # Apply torch.compile to predict_action_chunk method if enabled
+    if cfg.use_torch_compile:
+        policy = _apply_torch_compile(policy, cfg)
+
+    # Create robot
+    logger.info(f"Initializing robot: {cfg.robot.type}")
+    robot = make_robot_from_config(cfg.robot)
+    robot.connect()
+    robot_wrapper = RobotWrapper(robot)
+
+    # Create robot observation processor
+    robot_observation_processor = make_default_robot_observation_processor()
+    robot_action_processor = make_default_robot_action_processor()
+
+    # Create action queue for communication between threads
+    action_queue = ActionQueue(cfg.rtc)
+
+    # Start chunk requester thread
+    get_actions_thread = Thread(
+        target=get_actions,
+        args=(policy, robot_wrapper, robot_observation_processor, action_queue, shutdown_event, cfg),
+        daemon=True,
+        name="GetActions",
+    )
+    get_actions_thread.start()
+    logger.info("Started get actions thread")
+
+    # Start action executor thread
+    actor_thread = Thread(
+        target=actor_control,
+        args=(robot_wrapper, robot_action_processor, action_queue, shutdown_event, cfg),
+        daemon=True,
+        name="Actor",
+    )
+    actor_thread.start()
+    logger.info("Started actor thread")
+
+    logger.info("Started stop by duration thread")
+
+    # Main thread monitors for duration or shutdown
+    logger.info(f"Running demo for {cfg.duration} seconds...")
+    start_time = time.time()
+
+    while not shutdown_event.is_set() and (time.time() - start_time) < cfg.duration:
+        time.sleep(10)
+
+        # Log queue status periodically
+        if int(time.time() - start_time) % 5 == 0:
+            logger.info(f"[MAIN] Action queue size: {action_queue.qsize()}")
+
+        if time.time() - start_time > cfg.duration:
+            break
+
+    logger.info("Demo duration reached or shutdown requested")
+
+    # Signal shutdown
+    shutdown_event.set()
+
+    # Wait for threads to finish
+    if get_actions_thread and get_actions_thread.is_alive():
+        logger.info("Waiting for chunk requester thread to finish...")
+        get_actions_thread.join()
+
+    if actor_thread and actor_thread.is_alive():
+        logger.info("Waiting for action executor thread to finish...")
+        actor_thread.join()
+
+    # Cleanup robot
+    if robot:
+        robot.disconnect()
+        logger.info("Robot disconnected")
+
+    logger.info("Cleanup completed")
+
+
+if __name__ == "__main__":
+    demo_cli()
+    logging.info("RTC demo finished")

examples/so100_to_so100_EE/evaluate.py

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

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.0"
+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 = [
@@ -129,6 +133,7 @@ groot = [
     "ninja>=1.11.1,<2.0.0",
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
+xvla = ["lerobot[transformers-dep]"]
 hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
@@ -142,7 +147,7 @@ video_benchmark = ["scikit-image>=0.23.2,<0.26.0", "pandas>=2.2.2,<2.4.0"]
 # Simulation
 aloha = ["gym-aloha>=0.1.2,<0.2.0"]
 pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
-libero = ["lerobot[transformers-dep]", "libero @ git+https://github.com/huggingface/lerobot-libero.git@main#egg=libero"]
+libero = ["lerobot[transformers-dep]", "hf-libero>=0.1.3,<0.2.0"]
 metaworld = ["metaworld==3.0.0"]
 
 # All
@@ -157,6 +162,7 @@ all = [
     "lerobot[pi]",
     "lerobot[smolvla]",
     # "lerobot[groot]", TODO(Steven): Gr00t requires specific installation instructions for flash-attn
+    "lerobot[xvla]",
     "lerobot[hilserl]",
     "lerobot[async]",
     "lerobot[dev]",
@@ -356,9 +362,9 @@ ignore_errors = false
 # module = "lerobot.async_inference.*"
 # ignore_errors = false
 
-# [[tool.mypy.overrides]]
-# module = "lerobot.transport.*"
-# ignore_errors = false
+[[tool.mypy.overrides]]
+module = "lerobot.transport.*"
+ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.scripts.*"

src/lerobot/configs/types.py

@@ -43,3 +43,10 @@ class NormalizationMode(str, Enum):
 class PolicyFeature:
     type: FeatureType
     shape: tuple[int, ...]
+
+
+class RTCAttentionSchedule(str, Enum):
+    ZEROS = "ZEROS"
+    ONES = "ONES"
+    LINEAR = "LINEAR"
+    EXP = "EXP"

src/lerobot/datasets/aggregate.py

@@ -136,21 +136,40 @@ def update_meta_data(
         df["_orig_chunk"] = df[orig_chunk_col].copy()
         df["_orig_file"] = df[orig_file_col].copy()
 
-        # Update chunk and file indices to point to destination
-        df[orig_chunk_col] = video_idx["chunk"]
-        df[orig_file_col] = video_idx["file"]
-
-        # Apply per-source-file timestamp offsets
+        # Get mappings for this video key
         src_to_offset = video_idx.get("src_to_offset", {})
-        if src_to_offset:
-            # Apply offset based on original source file
+        src_to_dst = video_idx.get("src_to_dst", {})
+
+        # Apply per-source-file mappings
+        if src_to_dst:
+            # Map each episode to its correct destination file and apply offset
             for idx in df.index:
-                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
+                # Convert to Python int to avoid numpy type mismatch in dict lookup
+                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
+
+                # Get destination chunk/file for this source file
+                dst_chunk, dst_file = src_to_dst.get(src_key, (video_idx["chunk"], video_idx["file"]))
+                df.at[idx, orig_chunk_col] = dst_chunk
+                df.at[idx, orig_file_col] = dst_file
+
+                # Apply timestamp offset
+                offset = src_to_offset.get(src_key, 0)
+                df.at[idx, f"videos/{key}/from_timestamp"] += offset
+                df.at[idx, f"videos/{key}/to_timestamp"] += offset
+        elif src_to_offset:
+            # Fallback: use same destination for all, but apply per-file offsets
+            df[orig_chunk_col] = video_idx["chunk"]
+            df[orig_file_col] = video_idx["file"]
+            for idx in df.index:
+                # Convert to Python int to avoid numpy type mismatch in dict lookup
+                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
                 offset = src_to_offset.get(src_key, 0)
                 df.at[idx, f"videos/{key}/from_timestamp"] += offset
                 df.at[idx, f"videos/{key}/to_timestamp"] += offset
         else:
             # Fallback to simple offset (for backward compatibility)
+            df[orig_chunk_col] = video_idx["chunk"]
+            df[orig_file_col] = video_idx["file"]
             df[f"videos/{key}/from_timestamp"] = (
                 df[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
             )
@@ -268,6 +287,12 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
         videos_idx[key]["episode_duration"] = 0
         # Track offset for each source (chunk, file) pair
         videos_idx[key]["src_to_offset"] = {}
+        # Track destination (chunk, file) for each source (chunk, file) pair
+        videos_idx[key]["src_to_dst"] = {}
+        # Initialize dst_file_durations if not present
+        # dst_file_durations tracks duration of each destination file
+        if "dst_file_durations" not in videos_idx[key]:
+            videos_idx[key]["dst_file_durations"] = {}
 
     for key, video_idx in videos_idx.items():
         unique_chunk_file_pairs = {
@@ -282,9 +307,13 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
 
         chunk_idx = video_idx["chunk"]
         file_idx = video_idx["file"]
-        current_offset = video_idx["latest_duration"]
+        dst_file_durations = video_idx["dst_file_durations"]
 
         for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
+            # Convert to Python int to ensure consistent dict keys
+            src_chunk_idx = int(src_chunk_idx)
+            src_file_idx = int(src_file_idx)
+
             src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
                 video_key=key,
                 chunk_index=src_chunk_idx,
@@ -298,14 +327,17 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
             )
 
             src_duration = get_video_duration_in_s(src_path)
+            dst_key = (chunk_idx, file_idx)
 
             if not dst_path.exists():
-                # Store offset before incrementing
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
+                # New destination file: offset is 0
+                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = 0
+                videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
                 dst_path.parent.mkdir(parents=True, exist_ok=True)
                 shutil.copy(str(src_path), str(dst_path))
+                # Track duration of this destination file
+                dst_file_durations[dst_key] = src_duration
                 videos_idx[key]["episode_duration"] += src_duration
-                current_offset += src_duration
                 continue
 
             # Check file sizes before appending
@@ -313,10 +345,11 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
             dst_size = get_file_size_in_mb(dst_path)
 
             if dst_size + src_size >= video_files_size_in_mb:
-                # Rotate to a new file, this source becomes start of new destination
-                # So its offset should be 0
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = 0
+                # Rotate to a new file - offset is 0
                 chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, chunk_size)
+                dst_key = (chunk_idx, file_idx)
+                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = 0
+                videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
                 dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
                     video_key=key,
                     chunk_index=chunk_idx,
@@ -324,16 +357,20 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                 )
                 dst_path.parent.mkdir(parents=True, exist_ok=True)
                 shutil.copy(str(src_path), str(dst_path))
-                # Reset offset for next file
-                current_offset = src_duration
+                # Track duration of this new destination file
+                dst_file_durations[dst_key] = src_duration
             else:
-                # Append to existing video file - use current accumulated offset
-                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_offset
+                # Append to existing destination file
+                # Offset is the current duration of this destination file
+                current_dst_duration = dst_file_durations.get(dst_key, 0)
+                videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_dst_duration
+                videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
                 concatenate_video_files(
                     [dst_path, src_path],
                     dst_path,
                 )
-                current_offset += src_duration
+                # Update duration of this destination file
+                dst_file_durations[dst_key] = current_dst_duration + src_duration
 
             videos_idx[key]["episode_duration"] += src_duration
 

src/lerobot/datasets/dataset_tools.py

@@ -39,6 +39,7 @@ from lerobot.datasets.aggregate import aggregate_datasets
 from lerobot.datasets.compute_stats import aggregate_stats
 from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 from lerobot.datasets.utils import (
+    DATA_DIR,
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
@@ -962,28 +963,23 @@ def _copy_data_with_feature_changes(
     remove_features: list[str] | None = None,
 ) -> None:
     """Copy data while adding or removing features."""
-    if dataset.meta.episodes is None:
-        dataset.meta.episodes = load_episodes(dataset.meta.root)
+    data_dir = dataset.root / DATA_DIR
+    parquet_files = sorted(data_dir.glob("*/*.parquet"))
 
-    # Map file paths to episode indices to extract chunk/file indices
-    file_to_episodes: dict[Path, set[int]] = {}
-    for ep_idx in range(dataset.meta.total_episodes):
-        file_path = dataset.meta.get_data_file_path(ep_idx)
-        if file_path not in file_to_episodes:
-            file_to_episodes[file_path] = set()
-        file_to_episodes[file_path].add(ep_idx)
+    if not parquet_files:
+        raise ValueError(f"No parquet files found in {data_dir}")
 
     frame_idx = 0
 
-    for src_path in tqdm(sorted(file_to_episodes.keys()), desc="Processing data files"):
-        df = pd.read_parquet(dataset.root / src_path).reset_index(drop=True)
+    for src_path in tqdm(parquet_files, desc="Processing data files"):
+        df = pd.read_parquet(src_path).reset_index(drop=True)
 
-        # Get chunk_idx and file_idx from the source file's first episode
-        episodes_in_file = file_to_episodes[src_path]
-        first_ep_idx = min(episodes_in_file)
-        src_ep = dataset.meta.episodes[first_ep_idx]
-        chunk_idx = src_ep["data/chunk_index"]
-        file_idx = src_ep["data/file_index"]
+        relative_path = src_path.relative_to(dataset.root)
+        chunk_dir = relative_path.parts[1]
+        file_name = relative_path.parts[2]
+
+        chunk_idx = int(chunk_dir.split("-")[1])
+        file_idx = int(file_name.split("-")[1].split(".")[0])
 
         if remove_features:
             df = df.drop(columns=remove_features, errors="ignore")
@@ -1009,7 +1005,7 @@ def _copy_data_with_feature_changes(
                         df[feature_name] = feature_slice
             frame_idx = end_idx
 
-        # Write using the preserved chunk_idx and file_idx from source
+        # Write using the same chunk/file structure as source
         dst_path = new_meta.root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
         dst_path.parent.mkdir(parents=True, exist_ok=True)
 

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
@@ -430,9 +431,7 @@ class LeRobotDatasetMetadata:
         video_keys = [video_key] if video_key is not None else self.video_keys
         for key in video_keys:
             if not self.features[key].get("info", None):
-                video_path = self.root / self.video_path.format(
-                    video_key=video_key, chunk_index=0, file_index=0
-                )
+                video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
                 self.info["features"][key]["info"] = get_video_info(video_path)
 
     def update_chunk_settings(
@@ -541,6 +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,
@@ -714,6 +722,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
             self.download(download_videos)
             self.hf_dataset = self.load_hf_dataset()
 
+        # Create mapping from absolute indices to relative indices when only a subset of the episodes are loaded
+        # Build a mapping: absolute_index -> relative_index_in_filtered_dataset
+        self._absolute_to_relative_idx = None
+        if self.episodes is not None:
+            self._absolute_to_relative_idx = {
+                abs_idx.item() if isinstance(abs_idx, torch.Tensor) else abs_idx: rel_idx
+                for rel_idx, abs_idx in enumerate(self.hf_dataset["index"])
+            }
+
         # Setup delta_indices
         if self.delta_timestamps is not None:
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
@@ -832,7 +849,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
     def load_hf_dataset(self) -> datasets.Dataset:
         """hf_dataset contains all the observations, states, actions, rewards, etc."""
         features = get_hf_features_from_features(self.features)
-        hf_dataset = load_nested_dataset(self.root / "data", features=features)
+        hf_dataset = load_nested_dataset(self.root / "data", features=features, episodes=self.episodes)
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
@@ -849,10 +866,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Determine requested episodes
         if self.episodes is None:
-            # Requesting all episodes - check if we have all episodes from metadata
             requested_episodes = set(range(self.meta.total_episodes))
         else:
-            # Requesting specific episodes
             requested_episodes = set(self.episodes)
 
         # Check if all requested episodes are available in cached data
@@ -934,7 +949,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         query_timestamps = {}
         for key in self.meta.video_keys:
             if query_indices is not None and key in query_indices:
-                timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
+                if self._absolute_to_relative_idx is not None:
+                    relative_indices = [self._absolute_to_relative_idx[idx] for idx in query_indices[key]]
+                    timestamps = self.hf_dataset[relative_indices]["timestamp"]
+                else:
+                    timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
                 query_timestamps[key] = torch.stack(timestamps).tolist()
             else:
                 query_timestamps[key] = [current_ts]
@@ -942,11 +961,32 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return query_timestamps
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
-        return {
-            key: torch.stack(self.hf_dataset[q_idx][key])
-            for key, q_idx in query_indices.items()
-            if key not in self.meta.video_keys
-        }
+        """
+        Query dataset for indices across keys, skipping video keys.
+
+        Tries column-first [key][indices] for speed, falls back to row-first.
+
+        Args:
+            query_indices: Dict mapping keys to index lists to retrieve
+
+        Returns:
+            Dict with stacked tensors of queried data (video keys excluded)
+        """
+        result: dict = {}
+        for key, q_idx in query_indices.items():
+            if key in self.meta.video_keys:
+                continue
+            # Map absolute indices to relative indices if needed
+            relative_indices = (
+                q_idx
+                if self._absolute_to_relative_idx is None
+                else [self._absolute_to_relative_idx[idx] for idx in q_idx]
+            )
+            try:
+                result[key] = torch.stack(self.hf_dataset[key][relative_indices])
+            except (KeyError, TypeError, IndexError):
+                result[key] = torch.stack(self.hf_dataset[relative_indices][key])
+        return result
 
     def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict[str, torch.Tensor]:
         """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
@@ -1041,6 +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
@@ -1050,13 +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:
         """
@@ -1094,14 +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.
 
@@ -1113,6 +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
 
@@ -1149,8 +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)
@@ -1315,9 +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)
 
@@ -1435,11 +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(
@@ -1485,6 +1572,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_transforms = None
         obj.delta_timestamps = None
         obj.delta_indices = None
+        obj._absolute_to_relative_idx = None
         obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
         obj.writer = None
         obj.latest_episode = None

src/lerobot/datasets/utils.py

@@ -28,6 +28,7 @@ import numpy as np
 import packaging.version
 import pandas
 import pandas as pd
+import pyarrow.dataset as pa_ds
 import pyarrow.parquet as pq
 import torch
 from datasets import Dataset
@@ -48,7 +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"
@@ -103,7 +104,9 @@ def update_chunk_file_indices(chunk_idx: int, file_idx: int, chunks_size: int) -
     return chunk_idx, file_idx
 
 
-def load_nested_dataset(pq_dir: Path, features: datasets.Features | None = None) -> Dataset:
+def load_nested_dataset(
+    pq_dir: Path, features: datasets.Features | None = None, episodes: list[int] | None = None
+) -> Dataset:
     """Find parquet files in provided directory {pq_dir}/chunk-xxx/file-xxx.parquet
     Convert parquet files to pyarrow memory mapped in a cache folder for efficient RAM usage
     Concatenate all pyarrow references to return HF Dataset format
@@ -111,15 +114,26 @@ def load_nested_dataset(pq_dir: Path, features: datasets.Features | None = None)
     Args:
         pq_dir: Directory containing parquet files
         features: Optional features schema to ensure consistent loading of complex types like images
+        episodes: Optional list of episode indices to filter. Uses PyArrow predicate pushdown for efficiency.
     """
     paths = sorted(pq_dir.glob("*/*.parquet"))
     if len(paths) == 0:
         raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
 
-    # TODO(rcadene): set num_proc to accelerate conversion to pyarrow
     with SuppressProgressBars():
-        datasets = Dataset.from_parquet([str(path) for path in paths], features=features)
-    return datasets
+        # When no filtering needed, Dataset uses memory-mapped loading for efficiency
+        # PyArrow loads the entire dataset into memory
+        if episodes is None:
+            return Dataset.from_parquet([str(path) for path in paths], features=features)
+
+        arrow_dataset = pa_ds.dataset(paths, format="parquet")
+        filter_expr = pa_ds.field("episode_index").isin(episodes)
+        table = arrow_dataset.to_table(filter=filter_expr)
+
+        if features is not None:
+            table = table.cast(features.arrow_schema)
+
+        return Dataset(table)
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:

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/configs.py

@@ -21,7 +21,22 @@ import draccus
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.robots import RobotConfig
 from lerobot.teleoperators.config import TeleoperatorConfig
-from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
+from lerobot.utils.constants import (
+    ACTION,
+    LIBERO_KEY_EEF_MAT,
+    LIBERO_KEY_EEF_POS,
+    LIBERO_KEY_EEF_QUAT,
+    LIBERO_KEY_GRIPPER_QPOS,
+    LIBERO_KEY_GRIPPER_QVEL,
+    LIBERO_KEY_JOINTS_POS,
+    LIBERO_KEY_JOINTS_VEL,
+    LIBERO_KEY_PIXELS_AGENTVIEW,
+    LIBERO_KEY_PIXELS_EYE_IN_HAND,
+    OBS_ENV_STATE,
+    OBS_IMAGE,
+    OBS_IMAGES,
+    OBS_STATE,
+)
 
 
 @dataclass
@@ -230,7 +245,7 @@ class HILSerlRobotEnvConfig(EnvConfig):
 class LiberoEnv(EnvConfig):
     task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
     fps: int = 30
-    episode_length: int = 520
+    episode_length: int | None = None
     obs_type: str = "pixels_agent_pos"
     render_mode: str = "rgb_array"
     camera_name: str = "agentview_image,robot0_eye_in_hand_image"
@@ -246,28 +261,62 @@ class LiberoEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             ACTION: ACTION,
-            "agent_pos": OBS_STATE,
-            "pixels/agentview_image": f"{OBS_IMAGES}.image",
-            "pixels/robot0_eye_in_hand_image": f"{OBS_IMAGES}.image2",
+            LIBERO_KEY_EEF_POS: f"{OBS_STATE}.eef_pos",
+            LIBERO_KEY_EEF_QUAT: f"{OBS_STATE}.eef_quat",
+            LIBERO_KEY_EEF_MAT: f"{OBS_STATE}.eef_mat",
+            LIBERO_KEY_GRIPPER_QPOS: f"{OBS_STATE}.gripper_qpos",
+            LIBERO_KEY_GRIPPER_QVEL: f"{OBS_STATE}.gripper_qvel",
+            LIBERO_KEY_JOINTS_POS: f"{OBS_STATE}.joint_pos",
+            LIBERO_KEY_JOINTS_VEL: f"{OBS_STATE}.joint_vel",
+            LIBERO_KEY_PIXELS_AGENTVIEW: f"{OBS_IMAGES}.image",
+            LIBERO_KEY_PIXELS_EYE_IN_HAND: f"{OBS_IMAGES}.image2",
         }
     )
+    control_mode: str = "relative"  # or "absolute"
 
     def __post_init__(self):
         if self.obs_type == "pixels":
-            self.features["pixels/agentview_image"] = PolicyFeature(
+            self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
-            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
+            self.features[LIBERO_KEY_PIXELS_EYE_IN_HAND] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
         elif self.obs_type == "pixels_agent_pos":
-            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(8,))
-            self.features["pixels/agentview_image"] = PolicyFeature(
+            self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
-            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
+            self.features[LIBERO_KEY_PIXELS_EYE_IN_HAND] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
+            self.features[LIBERO_KEY_EEF_POS] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(3,),
+            )
+            self.features[LIBERO_KEY_EEF_QUAT] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(4,),
+            )
+            self.features[LIBERO_KEY_EEF_MAT] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(3, 3),
+            )
+            self.features[LIBERO_KEY_GRIPPER_QPOS] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(2,),
+            )
+            self.features[LIBERO_KEY_GRIPPER_QVEL] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(2,),
+            )
+            self.features[LIBERO_KEY_JOINTS_POS] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(7,),
+            )
+            self.features[LIBERO_KEY_JOINTS_VEL] = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(7,),
+            )
         else:
             raise ValueError(f"Unsupported obs_type: {self.obs_type}")
 

src/lerobot/envs/factory.py

@@ -14,11 +14,18 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import importlib
+from typing import Any
 
 import gymnasium as gym
 from gymnasium.envs.registration import registry as gym_registry
 
+from lerobot.configs.policies import PreTrainedConfig
 from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv
+from lerobot.envs.utils import _call_make_env, _download_hub_file, _import_hub_module, _normalize_hub_result
+from lerobot.policies.xvla.configuration_xvla import XVLAConfig
+from lerobot.processor import ProcessorStep
+from lerobot.processor.env_processor import LiberoProcessorStep
+from lerobot.processor.pipeline import PolicyProcessorPipeline
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -32,16 +39,65 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
         raise ValueError(f"Policy type '{env_type}' is not available.")
 
 
-def make_env(
-    cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False
-) -> dict[str, dict[int, gym.vector.VectorEnv]]:
-    """Makes a gym vector environment according to the config.
+def make_env_pre_post_processors(
+    env_cfg: EnvConfig,
+    policy_cfg: PreTrainedConfig,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+]:
+    """
+    Create preprocessor and postprocessor pipelines for environment observations.
+
+    This function creates processor pipelines that transform raw environment
+    observations and actions. By default, it returns identity processors that do nothing.
+    For specific environments like LIBERO, it adds environment-specific processing steps.
 
     Args:
-        cfg (EnvConfig): the config of the environment to instantiate.
+        env_cfg: The configuration of the environment.
+
+    Returns:
+        A tuple containing:
+            - preprocessor: Pipeline that processes environment observations
+            - postprocessor: Pipeline that processes environment outputs (currently identity)
+    """
+    # Preprocessor and Postprocessor steps are Identity for most environments
+    preprocessor_steps: list[ProcessorStep] = []
+    postprocessor_steps: list[ProcessorStep] = []
+    if isinstance(policy_cfg, XVLAConfig):
+        from lerobot.policies.xvla.processor_xvla import make_xvla_libero_pre_post_processors
+
+        return make_xvla_libero_pre_post_processors()
+
+    # For LIBERO environments, add the LiberoProcessorStep to preprocessor
+    if isinstance(env_cfg, LiberoEnv) or "libero" in env_cfg.type:
+        preprocessor_steps.append(LiberoProcessorStep())
+
+    preprocessor = PolicyProcessorPipeline(steps=preprocessor_steps)
+    postprocessor = PolicyProcessorPipeline(steps=postprocessor_steps)
+
+    return preprocessor, postprocessor
+
+
+def make_env(
+    cfg: EnvConfig | str,
+    n_envs: int = 1,
+    use_async_envs: bool = False,
+    hub_cache_dir: str | None = None,
+    trust_remote_code: bool = False,
+) -> dict[str, dict[int, gym.vector.VectorEnv]]:
+    """Makes a gym vector environment according to the config or Hub reference.
+
+    Args:
+        cfg (EnvConfig | str): Either an `EnvConfig` object describing the environment to build locally,
+            or a Hugging Face Hub repository identifier (e.g. `"username/repo"`). In the latter case,
+            the repo must include a Python file (usually `env.py`).
         n_envs (int, optional): The number of parallelized env to return. Defaults to 1.
         use_async_envs (bool, optional): Whether to return an AsyncVectorEnv or a SyncVectorEnv. Defaults to
             False.
+        hub_cache_dir (str | None): Optional cache path for downloaded hub files.
+        trust_remote_code (bool): **Explicit consent** to execute remote code from the Hub.
+            Default False — must be set to True to import/exec hub `env.py`.
 
     Raises:
         ValueError: if n_envs < 1
@@ -54,6 +110,21 @@ def make_env(
             - For single-task environments: a single suite entry (cfg.type) with task_id=0.
 
     """
+    # if user passed a hub id string (e.g., "username/repo", "username/repo@main:env.py")
+    # simplified: only support hub-provided `make_env`
+    if isinstance(cfg, str):
+        # _download_hub_file will raise the same RuntimeError if trust_remote_code is False
+        repo_id, file_path, local_file, revision = _download_hub_file(cfg, trust_remote_code, hub_cache_dir)
+
+        # 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)
+
+        # normalize the return into {suite: {task_id: vec_env}}
+        return _normalize_hub_result(raw_result)
+
     if n_envs < 1:
         raise ValueError("`n_envs` must be at least 1")
 
@@ -72,6 +143,8 @@ def make_env(
             init_states=cfg.init_states,
             gym_kwargs=cfg.gym_kwargs,
             env_cls=env_cls,
+            control_mode=cfg.control_mode,
+            episode_length=cfg.episode_length,
         )
     elif "metaworld" in cfg.type:
         from lerobot.envs.metaworld import create_metaworld_envs

src/lerobot/envs/libero.py

@@ -28,7 +28,6 @@ import torch
 from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv
-from robosuite.utils.transform_utils import quat2axisangle
 
 
 def _parse_camera_names(camera_name: str | Sequence[str]) -> list[str]:
@@ -81,10 +80,7 @@ def get_libero_dummy_action():
     return [0, 0, 0, 0, 0, 0, -1]
 
 
-OBS_STATE_DIM = 8
 ACTION_DIM = 7
-AGENT_POS_LOW = -1000.0
-AGENT_POS_HIGH = 1000.0
 ACTION_LOW = -1.0
 ACTION_HIGH = 1.0
 TASK_SUITE_MAX_STEPS: dict[str, int] = {
@@ -104,6 +100,7 @@ class LiberoEnv(gym.Env):
         task_suite: Any,
         task_id: int,
         task_suite_name: str,
+        episode_length: int | None = None,
         camera_name: str | Sequence[str] = "agentview_image,robot0_eye_in_hand_image",
         obs_type: str = "pixels",
         render_mode: str = "rgb_array",
@@ -115,6 +112,7 @@ class LiberoEnv(gym.Env):
         episode_index: int = 0,
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
+        control_mode: str = "relative",
     ):
         super().__init__()
         self.task_id = task_id
@@ -142,14 +140,19 @@ class LiberoEnv(gym.Env):
         self.camera_name_mapping = camera_name_mapping
         self.num_steps_wait = num_steps_wait
         self.episode_index = episode_index
+        self.episode_length = episode_length
         # Load once and keep
         self._init_states = get_task_init_states(task_suite, self.task_id) if self.init_states else None
         self._init_state_id = self.episode_index  # tie each sub-env to a fixed init state
 
         self._env = self._make_envs_task(task_suite, self.task_id)
         default_steps = 500
-        self._max_episode_steps = TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
-
+        self._max_episode_steps = (
+            TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
+            if self.episode_length is None
+            else self.episode_length
+        )
+        self.control_mode = control_mode
         images = {}
         for cam in self.camera_name:
             images[self.camera_name_mapping[cam]] = spaces.Box(
@@ -175,11 +178,36 @@ class LiberoEnv(gym.Env):
             self.observation_space = spaces.Dict(
                 {
                     "pixels": spaces.Dict(images),
-                    "agent_pos": spaces.Box(
-                        low=AGENT_POS_LOW,
-                        high=AGENT_POS_HIGH,
-                        shape=(OBS_STATE_DIM,),
-                        dtype=np.float64,
+                    "robot_state": spaces.Dict(
+                        {
+                            "eef": spaces.Dict(
+                                {
+                                    "pos": spaces.Box(low=-np.inf, high=np.inf, shape=(3,), dtype=np.float64),
+                                    "quat": spaces.Box(
+                                        low=-np.inf, high=np.inf, shape=(4,), dtype=np.float64
+                                    ),
+                                    "mat": spaces.Box(
+                                        low=-np.inf, high=np.inf, shape=(3, 3), dtype=np.float64
+                                    ),
+                                }
+                            ),
+                            "gripper": spaces.Dict(
+                                {
+                                    "qpos": spaces.Box(
+                                        low=-np.inf, high=np.inf, shape=(2,), dtype=np.float64
+                                    ),
+                                    "qvel": spaces.Box(
+                                        low=-np.inf, high=np.inf, shape=(2,), dtype=np.float64
+                                    ),
+                                }
+                            ),
+                            "joints": spaces.Dict(
+                                {
+                                    "pos": spaces.Box(low=-np.inf, high=np.inf, shape=(7,), dtype=np.float64),
+                                    "vel": spaces.Box(low=-np.inf, high=np.inf, shape=(7,), dtype=np.float64),
+                                }
+                            ),
+                        }
                     ),
                 }
             )
@@ -191,6 +219,7 @@ class LiberoEnv(gym.Env):
     def render(self):
         raw_obs = self._env.env._get_observations()
         image = self._format_raw_obs(raw_obs)["pixels"]["image"]
+        image = image[::-1, ::-1]  # flip both H and W for visualization
         return image
 
     def _make_envs_task(self, task_suite: Any, task_id: int = 0):
@@ -212,23 +241,48 @@ class LiberoEnv(gym.Env):
         images = {}
         for camera_name in self.camera_name:
             image = raw_obs[camera_name]
-            image = image[::-1, ::-1]  # rotate 180 degrees
             images[self.camera_name_mapping[camera_name]] = image
-        state = np.concatenate(
-            (
-                raw_obs["robot0_eef_pos"],
-                quat2axisangle(raw_obs["robot0_eef_quat"]),
-                raw_obs["robot0_gripper_qpos"],
-            )
-        )
-        agent_pos = state
+
+        eef_pos = raw_obs.get("robot0_eef_pos")
+        eef_quat = raw_obs.get("robot0_eef_quat")
+
+        # rotation matrix from controller
+        eef_mat = self._env.robots[0].controller.ee_ori_mat if eef_pos is not None else None
+        gripper_qpos = raw_obs.get("robot0_gripper_qpos")
+        gripper_qvel = raw_obs.get("robot0_gripper_qvel")
+        joint_pos = raw_obs.get("robot0_joint_pos")
+        joint_vel = raw_obs.get("robot0_joint_vel")
+        obs = {
+            "pixels": images,
+            "robot_state": {
+                "eef": {
+                    "pos": eef_pos,  # (3,)
+                    "quat": eef_quat,  # (4,)
+                    "mat": eef_mat,  # (3, 3)
+                },
+                "gripper": {
+                    "qpos": gripper_qpos,  # (2,)
+                    "qvel": gripper_qvel,  # (2,)
+                },
+                "joints": {
+                    "pos": joint_pos,  # (7,)
+                    "vel": joint_vel,  # (7,)
+                },
+            },
+        }
         if self.obs_type == "pixels":
             return {"pixels": images.copy()}
+
         if self.obs_type == "pixels_agent_pos":
-            return {
-                "pixels": images.copy(),
-                "agent_pos": agent_pos,
-            }
+            # Validate required fields are present
+            if eef_pos is None or eef_quat is None or gripper_qpos is None:
+                raise ValueError(
+                    f"Missing required robot state fields in raw observation. "
+                    f"Got eef_pos={eef_pos is not None}, eef_quat={eef_quat is not None}, "
+                    f"gripper_qpos={gripper_qpos is not None}"
+                )
+            return obs
+
         raise NotImplementedError(
             f"The observation type '{self.obs_type}' is not supported in LiberoEnv. "
             "Please switch to an image-based obs_type (e.g. 'pixels', 'pixels_agent_pos')."
@@ -246,6 +300,15 @@ class LiberoEnv(gym.Env):
         # Increasing this value can improve determinism and reproducibility across resets.
         for _ in range(self.num_steps_wait):
             raw_obs, _, _, _ = self._env.step(get_libero_dummy_action())
+
+        if self.control_mode == "absolute":
+            for robot in self._env.robots:
+                robot.controller.use_delta = False
+        elif self.control_mode == "relative":
+            for robot in self._env.robots:
+                robot.controller.use_delta = True
+        else:
+            raise ValueError(f"Invalid control mode: {self.control_mode}")
         observation = self._format_raw_obs(raw_obs)
         info = {"is_success": False}
         return observation, info
@@ -291,8 +354,10 @@ def _make_env_fns(
     task_id: int,
     n_envs: int,
     camera_names: list[str],
+    episode_length: int | None,
     init_states: bool,
     gym_kwargs: Mapping[str, Any],
+    control_mode: str,
 ) -> list[Callable[[], LiberoEnv]]:
     """Build n_envs factory callables for a single (suite, task_id)."""
 
@@ -304,7 +369,9 @@ def _make_env_fns(
             task_suite_name=suite_name,
             camera_name=camera_names,
             init_states=init_states,
+            episode_length=episode_length,
             episode_index=episode_index,
+            control_mode=control_mode,
             **local_kwargs,
         )
 
@@ -324,6 +391,8 @@ def create_libero_envs(
     camera_name: str | Sequence[str] = "agentview_image,robot0_eye_in_hand_image",
     init_states: bool = True,
     env_cls: Callable[[Sequence[Callable[[], Any]]], Any] | None = None,
+    control_mode: str = "relative",
+    episode_length: int | None = None,
 ) -> dict[str, dict[int, Any]]:
     """
     Create vectorized LIBERO environments with a consistent return shape.
@@ -355,24 +424,24 @@ def create_libero_envs(
         print(f"Restricting to task_ids={task_ids_filter}")
 
     out: dict[str, dict[int, Any]] = defaultdict(dict)
-
     for suite_name in suite_names:
         suite = _get_suite(suite_name)
         total = len(suite.tasks)
         selected = _select_task_ids(total, task_ids_filter)
-
         if not selected:
             raise ValueError(f"No tasks selected for suite '{suite_name}' (available: {total}).")
 
         for tid in selected:
             fns = _make_env_fns(
                 suite=suite,
+                episode_length=episode_length,
                 suite_name=suite_name,
                 task_id=tid,
                 n_envs=n_envs,
                 camera_names=camera_names,
                 init_states=init_states,
                 gym_kwargs=gym_kwargs,
+                control_mode=control_mode,
             )
             out[suite_name][tid] = env_cls(fns)
             print(f"Built vec env | suite={suite_name} | task_id={tid} | n_envs={n_envs}")

src/lerobot/envs/utils.py

@@ -13,6 +13,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import importlib.util
+import os
 import warnings
 from collections.abc import Mapping, Sequence
 from functools import singledispatch
@@ -22,14 +24,27 @@ import einops
 import gymnasium as gym
 import numpy as np
 import torch
+from huggingface_hub import hf_hub_download, snapshot_download
 from torch import Tensor
 
 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.envs.configs import EnvConfig
-from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
+from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.utils import get_channel_first_image_shape
 
 
+def _convert_nested_dict(d):
+    result = {}
+    for k, v in d.items():
+        if isinstance(v, dict):
+            result[k] = _convert_nested_dict(v)
+        elif isinstance(v, np.ndarray):
+            result[k] = torch.from_numpy(v)
+        else:
+            result[k] = v
+    return result
+
+
 def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
     # TODO(aliberts, rcadene): refactor this to use features from the environment (no hardcoding)
     """Convert environment observation to LeRobot format observation.
@@ -75,12 +90,14 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
 
         return_observations[OBS_ENV_STATE] = env_state
 
-    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
-    agent_pos = torch.from_numpy(observations["agent_pos"]).float()
-    if agent_pos.dim() == 1:
-        agent_pos = agent_pos.unsqueeze(0)
-    return_observations[OBS_STATE] = agent_pos
+    if "agent_pos" in observations:
+        agent_pos = torch.from_numpy(observations["agent_pos"]).float()
+        if agent_pos.dim() == 1:
+            agent_pos = agent_pos.unsqueeze(0)
+        return_observations[OBS_STATE] = agent_pos
 
+    if "robot_state" in observations:
+        return_observations[f"{OBS_STR}.robot_state"] = _convert_nested_dict(observations["robot_state"])
     return return_observations
 
 
@@ -195,3 +212,132 @@ def _(envs: Sequence) -> None:
 @close_envs.register
 def _(env: gym.Env) -> None:
     _close_single_env(env)
+
+
+# helper to safely load a python file as a module
+def _load_module_from_path(path: str, module_name: str | None = None):
+    module_name = module_name or f"hub_env_{os.path.basename(path).replace('.', '_')}"
+    spec = importlib.util.spec_from_file_location(module_name, path)
+    if spec is None:
+        raise ImportError(f"Could not load module spec for {module_name} from {path}")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+# helper to parse hub string (supports "user/repo", "user/repo@rev", optional path)
+# examples:
+#   "user/repo" -> will look for env.py at repo root
+#   "user/repo@main:envs/my_env.py" -> explicit revision and path
+def _parse_hub_url(hub_uri: str):
+    # very small parser: [repo_id][@revision][:path]
+    # repo_id is required (user/repo or org/repo)
+    revision = None
+    file_path = "env.py"
+    if "@" in hub_uri:
+        repo_and_rev, *rest = hub_uri.split(":", 1)
+        repo_id, rev = repo_and_rev.split("@", 1)
+        revision = rev
+        if rest:
+            file_path = rest[0]
+    else:
+        repo_id, *rest = hub_uri.split(":", 1)
+        if rest:
+            file_path = rest[0]
+    return repo_id, revision, file_path
+
+
+def _download_hub_file(
+    cfg_str: str,
+    trust_remote_code: bool,
+    hub_cache_dir: str | None,
+) -> tuple[str, str, str, str]:
+    """
+    Parse `cfg_str` (hub URL), enforce `trust_remote_code`, and return
+    (repo_id, file_path, local_file, revision).
+    """
+    if not trust_remote_code:
+        raise RuntimeError(
+            f"Refusing to execute remote code from the Hub for '{cfg_str}'. "
+            "Executing hub env modules runs arbitrary Python code from third-party repositories. "
+            "If you trust this repo and understand the risks, call `make_env(..., trust_remote_code=True)` "
+            "and prefer pinning to a specific revision: 'user/repo@<commit-hash>:env.py'."
+        )
+
+    repo_id, revision, file_path = _parse_hub_url(cfg_str)
+
+    try:
+        local_file = hf_hub_download(
+            repo_id=repo_id, filename=file_path, revision=revision, cache_dir=hub_cache_dir
+        )
+    except Exception as e:
+        # fallback to snapshot download
+        snapshot_dir = snapshot_download(repo_id=repo_id, revision=revision, cache_dir=hub_cache_dir)
+        local_file = os.path.join(snapshot_dir, file_path)
+        if not os.path.exists(local_file):
+            raise FileNotFoundError(
+                f"Could not find {file_path} in repository {repo_id}@{revision or 'main'}"
+            ) from e
+
+    return repo_id, file_path, local_file, revision
+
+
+def _import_hub_module(local_file: str, repo_id: str) -> Any:
+    """
+    Import the downloaded file as a module and surface helpful import error messages.
+    """
+    module_name = f"hub_env_{repo_id.replace('/', '_')}"
+    try:
+        module = _load_module_from_path(local_file, module_name=module_name)
+    except ModuleNotFoundError as e:
+        missing = getattr(e, "name", None) or str(e)
+        raise ModuleNotFoundError(
+            f"Hub env '{repo_id}:{os.path.basename(local_file)}' failed to import because the dependency "
+            f"'{missing}' is not installed locally.\n\n"
+        ) from e
+    except ImportError as e:
+        raise ImportError(
+            f"Failed to load hub env module '{repo_id}:{os.path.basename(local_file)}'. Import error: {e}\n\n"
+        ) from e
+    return module
+
+
+def _call_make_env(module: Any, n_envs: int, use_async_envs: bool) -> Any:
+    """
+    Ensure module exposes make_env and call it.
+    """
+    if not hasattr(module, "make_env"):
+        raise AttributeError(
+            f"The hub module {getattr(module, '__name__', 'hub_module')} must expose `make_env(n_envs=int, use_async_envs=bool)`."
+        )
+    entry_fn = module.make_env
+    return entry_fn(n_envs=n_envs, use_async_envs=use_async_envs)
+
+
+def _normalize_hub_result(result: Any) -> dict[str, dict[int, gym.vector.VectorEnv]]:
+    """
+    Normalize possible return types from hub `make_env` into the mapping:
+      { suite_name: { task_id: vector_env } }
+    Accepts:
+      - dict (assumed already correct)
+      - gym.vector.VectorEnv
+      - gym.Env (will be wrapped into SyncVectorEnv)
+    """
+    if isinstance(result, dict):
+        return result
+
+    # VectorEnv: use its spec.id if available
+    if isinstance(result, gym.vector.VectorEnv):
+        suite_name = getattr(result, "spec", None) and getattr(result.spec, "id", None) or "hub_env"
+        return {suite_name: {0: result}}
+
+    # Single Env: wrap into SyncVectorEnv
+    if isinstance(result, gym.Env):
+        vec = gym.vector.SyncVectorEnv([lambda: result])
+        suite_name = getattr(result, "spec", None) and getattr(result.spec, "id", None) or "hub_env"
+        return {suite_name: {0: vec}}
+
+    raise ValueError(
+        "Hub `make_env` must return either a mapping {suite: {task_id: vec_env}}, "
+        "a gym.vector.VectorEnv, or a single gym.Env."
+    )

src/lerobot/optim/optimizers.py

@@ -81,10 +81,14 @@ class AdamWConfig(OptimizerConfig):
     eps: float = 1e-8
     weight_decay: float = 1e-2
     grad_clip_norm: float = 10.0
+    fused: bool = False
 
     def build(self, params: dict) -> torch.optim.Optimizer:
         kwargs = asdict(self)
         kwargs.pop("grad_clip_norm")
+        # Fused optimizer only works on CUDA
+        if kwargs.get("fused") and not torch.cuda.is_available():
+            kwargs["fused"] = False
         return torch.optim.AdamW(params, **kwargs)
 
 
@@ -104,6 +108,107 @@ class SGDConfig(OptimizerConfig):
         return torch.optim.SGD(params, **kwargs)
 
 
+@OptimizerConfig.register_subclass("xvla-adamw")
+@dataclass
+class XVLAAdamWConfig(OptimizerConfig):
+    """Custom AdamW optimizer for XVLA with differential learning rates.
+
+    The Vision-Language Model (VLM) is trained with 1/10 of the base learning rate
+    for stable optimization, while all other components use the full LR.
+
+    This LR ratio is crucial for achieving strong and stable finetuning performance.
+
+    Soft-prompts can optionally use a separate learning rate with warm-up support.
+    Set `soft_prompt_lr_scale` to a value < 1.0 (e.g., 0.1) to start soft-prompts
+    at a lower LR. Combine with a warmup scheduler for optimal results.
+
+    Note:
+        Completely matching official reported performance may require an additional
+        warm-up LR schedule for soft-prompts, which can bring minor improvements.
+        When `soft_prompt_warmup_lr_scale` is set, soft-prompts start at
+        `lr * soft_prompt_warmup_lr_scale` and should be warmed up via the scheduler.
+
+    Parameter Groups:
+        - Group 0 (vlm): VLM parameters at lr * 0.1, weight_decay * 0.1
+        - Group 1 (soft_prompts): Soft-prompt parameters at lr * soft_prompt_lr_scale
+        - Group 2 (other): All other parameters at full lr
+    """
+
+    lr: float = 1e-4
+    betas: tuple[float, float] = (0.9, 0.99)
+    eps: float = 1e-8
+    weight_decay: float = 0.0
+    grad_clip_norm: float = 10.0
+    # Soft-prompt specific settings
+    soft_prompt_lr_scale: float = 1.0  # Scale factor for soft-prompt LR (1.0 = same as base LR)
+    soft_prompt_warmup_lr_scale: float | None = None  # If set, start soft-prompts at this scale (e.g., 0.01)
+
+    def build(self, params: dict) -> torch.optim.Optimizer:
+        """
+        Build AdamW optimizer with differential learning rates.
+
+        Expects `named_parameters()` as input (dict of name -> param).
+        Applies:
+        - lr * 0.1 for all VLM-related parameters
+        - lr * soft_prompt_lr_scale for soft-prompt parameters (with optional warmup)
+        - full lr for all other parameters
+
+        Args:
+            params: Dictionary of parameter names to parameters (from named_parameters())
+
+        Returns:
+            AdamW optimizer with parameter groups for VLM, soft-prompts, and other components
+        """
+        assert isinstance(params, dict), "Custom LR optimizer requires `named_parameters()` as inputs."
+
+        vlm_group, soft_prompt_group, other_group = [], [], []
+        for name, p in params.items():
+            if not p.requires_grad:
+                continue
+            if "vlm" in name.lower():
+                vlm_group.append(p)
+            elif "soft_prompt" in name.lower():
+                soft_prompt_group.append(p)
+            else:
+                other_group.append(p)
+
+        # Determine soft-prompt LR
+        soft_prompt_lr = self.lr * self.soft_prompt_lr_scale
+        if self.soft_prompt_warmup_lr_scale is not None:
+            # Start at warmup scale, scheduler will warm up to soft_prompt_lr
+            soft_prompt_lr = self.lr * self.soft_prompt_warmup_lr_scale
+
+        param_groups = [
+            {
+                "params": vlm_group,
+                "lr": self.lr * 0.1,
+                "weight_decay": self.weight_decay * 0.1,
+                "name": "vlm",
+            },
+            {
+                "params": soft_prompt_group,
+                "lr": soft_prompt_lr,
+                "weight_decay": self.weight_decay,
+                "name": "soft_prompts",
+            },
+            {
+                "params": other_group,
+                "lr": self.lr,
+                "weight_decay": self.weight_decay,
+                "name": "other",
+            },
+        ]
+
+        # Filter out empty groups
+        param_groups = [g for g in param_groups if len(g["params"]) > 0]
+
+        return torch.optim.AdamW(
+            param_groups,
+            betas=self.betas,
+            eps=self.eps,
+        )
+
+
 @OptimizerConfig.register_subclass("multi_adam")
 @dataclass
 class MultiAdamConfig(OptimizerConfig):

src/lerobot/policies/__init__.py

@@ -21,6 +21,7 @@ from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .smolvla.processor_smolvla import SmolVLANewLineProcessor
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
+from .xvla.configuration_xvla import XVLAConfig as XVLAConfig
 
 __all__ = [
     "ACTConfig",
@@ -31,4 +32,5 @@ __all__ = [
     "TDMPCConfig",
     "VQBeTConfig",
     "GrootConfig",
+    "XVLAConfig",
 ]

src/lerobot/policies/act/configuration_act.py

@@ -136,6 +136,7 @@ class ACTConfig(PreTrainedConfig):
     optimizer_lr: float = 1e-5
     optimizer_weight_decay: float = 1e-4
     optimizer_lr_backbone: float = 1e-5
+    optimizer_fused: bool = False  # Use CUDA fused AdamW kernel
 
     def __post_init__(self):
         super().__post_init__()
@@ -164,6 +165,7 @@ class ACTConfig(PreTrainedConfig):
         return AdamWConfig(
             lr=self.optimizer_lr,
             weight_decay=self.optimizer_weight_decay,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self) -> None:

src/lerobot/policies/factory.py

@@ -16,6 +16,7 @@
 
 from __future__ import annotations
 
+import importlib
 import logging
 from typing import Any, TypedDict
 
@@ -38,7 +39,9 @@ from lerobot.policies.sac.configuration_sac import SACConfig
 from lerobot.policies.sac.reward_model.configuration_classifier import RewardClassifierConfig
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.policies.tdmpc.configuration_tdmpc import TDMPCConfig
+from lerobot.policies.utils import validate_visual_features_consistency
 from lerobot.policies.vqbet.configuration_vqbet import VQBeTConfig
+from lerobot.policies.xvla.configuration_xvla import XVLAConfig
 from lerobot.processor import PolicyAction, PolicyProcessorPipeline
 from lerobot.processor.converters import (
     batch_to_transition,
@@ -106,8 +109,15 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.groot.modeling_groot import GrootPolicy
 
         return GrootPolicy
+    elif name == "xvla":
+        from lerobot.policies.xvla.modeling_xvla import XVLAPolicy
+
+        return XVLAPolicy
     else:
-        raise NotImplementedError(f"Policy with name {name} is not implemented.")
+        try:
+            return _get_policy_cls_from_policy_name(name=name)
+        except Exception as e:
+            raise ValueError(f"Policy type '{name}' is not available.") from e
 
 
 def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
@@ -149,8 +159,14 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return RewardClassifierConfig(**kwargs)
     elif policy_type == "groot":
         return GrootConfig(**kwargs)
+    elif policy_type == "xvla":
+        return XVLAConfig(**kwargs)
     else:
-        raise ValueError(f"Policy type '{policy_type}' is not available.")
+        try:
+            config_cls = PreTrainedConfig.get_choice_class(policy_type)
+            return config_cls(**kwargs)
+        except Exception as e:
+            raise ValueError(f"Policy type '{policy_type}' is not available.") from e
 
 
 class ProcessorConfigKwargs(TypedDict, total=False):
@@ -328,9 +344,24 @@ def make_pre_post_processors(
             config=policy_cfg,
             dataset_stats=kwargs.get("dataset_stats"),
         )
+    elif isinstance(policy_cfg, XVLAConfig):
+        from lerobot.policies.xvla.processor_xvla import (
+            make_xvla_pre_post_processors,
+        )
+
+        processors = make_xvla_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
 
     else:
-        raise NotImplementedError(f"Processor for policy type '{policy_cfg.type}' is not implemented.")
+        try:
+            processors = _make_processors_from_policy_config(
+                config=policy_cfg,
+                dataset_stats=kwargs.get("dataset_stats"),
+            )
+        except Exception as e:
+            raise ValueError(f"Processor for policy type '{policy_cfg.type}' is not implemented.") from e
 
     return processors
 
@@ -399,8 +430,7 @@ def make_policy(
             raise ValueError("env_cfg cannot be None when ds_meta is not provided")
         features = env_to_policy_features(env_cfg)
 
-    if not cfg.output_features:
-        cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
+    cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
     if not cfg.input_features:
         cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
     kwargs["config"] = cfg
@@ -420,20 +450,69 @@ def make_policy(
     # policy = torch.compile(policy, mode="reduce-overhead")
 
     if not rename_map:
-        expected_features = set(cfg.input_features.keys()) | set(cfg.output_features.keys())
-        provided_features = set(features.keys())
-        if expected_features and provided_features != expected_features:
-            missing = expected_features - provided_features
-            extra = provided_features - expected_features
-            # TODO (jadechoghari): provide a dynamic rename map suggestion to the user.
-            raise ValueError(
-                f"Feature mismatch between dataset/environment and policy config.\n"
-                f"- Missing features: {sorted(missing) if missing else 'None'}\n"
-                f"- Extra features: {sorted(extra) if extra else 'None'}\n\n"
-                f"Please ensure your dataset and policy use consistent feature names.\n"
-                f"If your dataset uses different observation keys (e.g., cameras named differently), "
-                f"use the `--rename_map` argument, for example:\n"
-                f'  --rename_map=\'{{"observation.images.left": "observation.images.camera1", '
-                f'"observation.images.top": "observation.images.camera2"}}\''
-            )
+        validate_visual_features_consistency(cfg, features)
+        # TODO: (jadechoghari) - add a check_state(cfg, features) and check_action(cfg, features)
+
     return policy
+
+
+def _get_policy_cls_from_policy_name(name: str) -> type[PreTrainedConfig]:
+    """Get policy class from its registered name using dynamic imports.
+
+    This is used as a helper function to import policies from 3rd party lerobot plugins.
+
+    Args:
+        name: The name of the policy.
+    Returns:
+        The policy class corresponding to the given name.
+    """
+    if name not in PreTrainedConfig.get_known_choices():
+        raise ValueError(
+            f"Unknown policy name '{name}'. Available policies: {PreTrainedConfig.get_known_choices()}"
+        )
+
+    config_cls = PreTrainedConfig.get_choice_class(name)
+    config_cls_name = config_cls.__name__
+
+    model_name = config_cls_name.removesuffix("Config")  # e.g., DiffusionConfig -> Diffusion
+    if model_name == config_cls_name:
+        raise ValueError(
+            f"The config class name '{config_cls_name}' does not follow the expected naming convention."
+            f"Make sure it ends with 'Config'!"
+        )
+    cls_name = model_name + "Policy"  # e.g., DiffusionConfig -> DiffusionPolicy
+    module_path = config_cls.__module__.replace(
+        "configuration_", "modeling_"
+    )  # e.g., configuration_diffusion -> modeling_diffusion
+
+    module = importlib.import_module(module_path)
+    policy_cls = getattr(module, cls_name)
+    return policy_cls
+
+
+def _make_processors_from_policy_config(
+    config: PreTrainedConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[Any, Any]:
+    """Create pre- and post-processors from a policy configuration using dynamic imports.
+
+    This is used as a helper function to import processor factories from 3rd party lerobot plugins.
+
+    Args:
+        config: The policy configuration object.
+        dataset_stats: Dataset statistics for normalization.
+    Returns:
+        A tuple containing the input (pre-processor) and output (post-processor) pipelines.
+    """
+
+    policy_type = config.type
+    function_name = f"make_{policy_type}_pre_post_processors"
+    module_path = config.__class__.__module__.replace(
+        "configuration_", "processor_"
+    )  # e.g., configuration_diffusion -> processor_diffusion
+    logging.debug(
+        f"Instantiating pre/post processors using function '{function_name}' from module '{module_path}'"
+    )
+    module = importlib.import_module(module_path)
+    function = getattr(module, function_name)
+    return function(config, dataset_stats=dataset_stats)

src/lerobot/policies/groot/configuration_groot.py

@@ -94,6 +94,7 @@ class GrootConfig(PreTrainedConfig):
     optimizer_betas: tuple[float, float] = (0.95, 0.999)
     optimizer_eps: float = 1e-8
     optimizer_weight_decay: float = 1e-5
+    optimizer_fused: bool = False  # Use CUDA fused AdamW kernel
     warmup_ratio: float = 0.05
     use_bf16: bool = True
 
@@ -174,6 +175,7 @@ class GrootConfig(PreTrainedConfig):
             betas=self.optimizer_betas,
             eps=self.optimizer_eps,
             weight_decay=self.optimizer_weight_decay,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self) -> CosineDecayWithWarmupSchedulerConfig:

src/lerobot/policies/pi0/configuration_pi0.py

@@ -20,8 +20,11 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.utils.constants import OBS_IMAGES
 
+DEFAULT_IMAGE_SIZE = 224
+
 
 @PreTrainedConfig.register_subclass("pi0")
 @dataclass
@@ -47,7 +50,13 @@ class PI0Config(PreTrainedConfig):
     min_period: float = 4e-3
     max_period: float = 4.0
 
-    image_resolution: tuple[int, int] = (224, 224)  # see openpi `preprocessing_pytorch.py`
+    # Real-Time Chunking (RTC) configuration
+    rtc_config: RTCConfig | None = None
+
+    image_resolution: tuple[int, int] = (
+        DEFAULT_IMAGE_SIZE,
+        DEFAULT_IMAGE_SIZE,
+    )  # see openpi `preprocessing_pytorch.py`
 
     # Add empty images. Used to add empty cameras when no image features are present.
     empty_cameras: int = 0
@@ -65,6 +74,7 @@ class PI0Config(PreTrainedConfig):
     gradient_checkpointing: bool = False  # Enable gradient checkpointing for memory optimization
     compile_model: bool = False  # Whether to use torch.compile for model optimization
     compile_mode: str = "max-autotune"  # Torch compile mode
+    optimizer_fused: bool = False  # Use CUDA fused AdamW kernel
     device: str | None = None  # Device to use for the model (None = auto-detect)
 
     # Optimizer settings: see openpi `AdamW``
@@ -132,6 +142,7 @@ class PI0Config(PreTrainedConfig):
             eps=self.optimizer_eps,
             weight_decay=self.optimizer_weight_decay,
             grad_clip_norm=self.optimizer_grad_clip_norm,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self):

src/lerobot/policies/pi0/modeling_pi0.py

@@ -19,11 +19,12 @@ import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal
+from typing import TYPE_CHECKING, Literal, TypedDict
 
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
+from typing_extensions import Unpack
 
 from lerobot.utils.import_utils import _transformers_available
 
@@ -40,8 +41,9 @@ else:
     PaliGemmaForConditionalGeneration = None
 
 from lerobot.configs.policies import PreTrainedConfig
-from lerobot.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.policies.pi0.configuration_pi0 import DEFAULT_IMAGE_SIZE, PI0Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
+from lerobot.policies.rtc.modeling_rtc import RTCProcessor
 from lerobot.utils.constants import (
     ACTION,
     OBS_LANGUAGE_ATTENTION_MASK,
@@ -51,6 +53,12 @@ from lerobot.utils.constants import (
 )
 
 
+class ActionSelectKwargs(TypedDict, total=False):
+    inference_delay: int | None
+    prev_chunk_left_over: Tensor | None
+    execution_horizon: int | None
+
+
 def get_safe_dtype(target_dtype, device_type):
     """Get a safe dtype for the given device type."""
     if device_type == "mps" and target_dtype == torch.float64:
@@ -329,6 +337,7 @@ class PaliGemmaWithExpertModel(
         action_expert_config,
         use_adarms=None,
         precision: Literal["bfloat16", "float32"] = "bfloat16",
+        image_size: int = DEFAULT_IMAGE_SIZE,
     ):
         if use_adarms is None:
             use_adarms = [False, False]
@@ -348,6 +357,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
+        vlm_config_hf.vision_config.image_size = image_size
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
@@ -503,18 +513,25 @@ class PaliGemmaWithExpertModel(
 class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
     """Core PI0 PyTorch model."""
 
-    def __init__(self, config: PI0Config):
+    def __init__(self, config: PI0Config, rtc_processor: RTCProcessor | None = None):
         super().__init__()
         self.config = config
+        self.rtc_processor = rtc_processor
 
         paligemma_config = get_gemma_config(config.paligemma_variant)
         action_expert_config = get_gemma_config(config.action_expert_variant)
 
+        if config.image_resolution[0] != config.image_resolution[1]:
+            raise ValueError(
+                f"PaliGemma expects square image resolution, invalid resolution: {config.image_resolution}"
+            )
+
         self.paligemma_with_expert = PaliGemmaWithExpertModel(
             paligemma_config,
             action_expert_config,
             use_adarms=[False, False],
             precision=config.dtype,
+            image_size=config.image_resolution[0],
         )
 
         self.action_in_proj = nn.Linear(config.max_action_dim, action_expert_config.width)
@@ -560,6 +577,9 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI0Pytorch model")
 
+    def _rtc_enabled(self):
+        return self.config.rtc_config is not None and self.config.rtc_config.enabled
+
     def _apply_checkpoint(self, func, *args, **kwargs):
         """Helper method to apply gradient checkpointing if enabled."""
         if self.gradient_checkpointing_enabled and self.training:
@@ -756,7 +776,15 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
 
     @torch.no_grad()  # see openpi `sample_actions` (slightly adapted)
     def sample_actions(
-        self, images, img_masks, lang_tokens, lang_masks, state, noise=None, num_steps=None
+        self,
+        images,
+        img_masks,
+        lang_tokens,
+        lang_masks,
+        state,
+        noise=None,
+        num_steps=None,
+        **kwargs: Unpack[ActionSelectKwargs],
     ) -> Tensor:
         """Do a full inference forward and compute the action."""
         if num_steps is None:
@@ -792,21 +820,41 @@ class PI0Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         )
 
         dt = -1.0 / num_steps
-        dt = torch.tensor(dt, dtype=torch.float32, device=device)
 
         x_t = noise
-        time = torch.tensor(1.0, dtype=torch.float32, device=device)
-        while time >= -dt / 2:
-            expanded_time = time.expand(bsize)
-            v_t = self.denoise_step(
-                state,
-                prefix_pad_masks,
-                past_key_values,
-                x_t,
-                expanded_time,
-            )
+        for step in range(num_steps):
+            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
+
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                return self.denoise_step(
+                    state=state,
+                    prefix_pad_masks=prefix_pad_masks,
+                    past_key_values=past_key_values,
+                    x_t=input_x_t,
+                    timestep=current_timestep,
+                )
+
+            if self._rtc_enabled():
+                inference_delay = kwargs.get("inference_delay")
+                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+                execution_horizon = kwargs.get("execution_horizon")
+
+                v_t = self.rtc_processor.denoise_step(
+                    x_t=x_t,
+                    prev_chunk_left_over=prev_chunk_left_over,
+                    inference_delay=inference_delay,
+                    time=time,
+                    original_denoise_step_partial=denoise_step_partial_call,
+                    execution_horizon=execution_horizon,
+                )
+            else:
+                v_t = denoise_step_partial_call(x_t)
+
             x_t = x_t + dt * v_t
-            time += dt
+
+            if self.rtc_processor is not None and self.rtc_processor.is_debug_enabled():
+                self.rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
 
         return x_t
 
@@ -869,7 +917,8 @@ class PI0Policy(PreTrainedPolicy):
         self.config = config
 
         # Initialize the core PI0 model
-        self.model = PI0Pytorch(config)
+        self.init_rtc_processor()
+        self.model = PI0Pytorch(config, rtc_processor=self.rtc_processor)
 
         # Enable gradient checkpointing if requested
         if config.gradient_checkpointing:
@@ -1059,6 +1108,22 @@ class PI0Policy(PreTrainedPolicy):
             ACTION: deque(maxlen=self.config.n_action_steps),
         }
 
+    def init_rtc_processor(self):
+        """Initialize RTC processor if RTC is enabled in config."""
+        self.rtc_processor = None
+
+        # Create processor if config provided
+        # If RTC is not enabled - we can still track the denoising data
+        if self.config.rtc_config is not None:
+            self.rtc_processor = RTCProcessor(self.config.rtc_config)
+
+            model_value = getattr(self, "model", None)
+            if model_value is not None:
+                model_value.rtc_processor = self.rtc_processor
+
+    def _rtc_enabled(self) -> bool:
+        return self.config.rtc_config is not None and self.config.rtc_config.enabled
+
     def _preprocess_images(self, batch: dict[str, Tensor]) -> tuple[list[Tensor], list[Tensor]]:
         """Preprocess images for the model.
 
@@ -1137,6 +1202,10 @@ class PI0Policy(PreTrainedPolicy):
     @torch.no_grad()
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select a single action given environment observations."""
+        assert not self._rtc_enabled(), (
+            "RTC is not supported for select_action, use it with predict_action_chunk"
+        )
+
         self.eval()
 
         # Action queue logic for n_action_steps > 1
@@ -1148,7 +1217,7 @@ class PI0Policy(PreTrainedPolicy):
         return self._action_queue.popleft()
 
     @torch.no_grad()
-    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
+    def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor:
         """Predict a chunk of actions given environment observations."""
         self.eval()
 
@@ -1157,8 +1226,8 @@ class PI0Policy(PreTrainedPolicy):
         lang_tokens, lang_masks = batch[f"{OBS_LANGUAGE_TOKENS}"], batch[f"{OBS_LANGUAGE_ATTENTION_MASK}"]
         state = self.prepare_state(batch)
 
-        # Sample actions using the model
-        actions = self.model.sample_actions(images, img_masks, lang_tokens, lang_masks, state)
+        # Sample actions using the model (pass through RTC kwargs)
+        actions = self.model.sample_actions(images, img_masks, lang_tokens, lang_masks, state, **kwargs)
 
         # Unpad actions to actual action dimension
         original_action_dim = self.config.output_features[ACTION].shape[0]

src/lerobot/policies/pi05/configuration_pi05.py

@@ -20,6 +20,9 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+
+DEFAULT_IMAGE_SIZE = 224
 
 
 @PreTrainedConfig.register_subclass("pi05")
@@ -46,7 +49,13 @@ class PI05Config(PreTrainedConfig):
     min_period: float = 4e-3
     max_period: float = 4.0
 
-    image_resolution: tuple[int, int] = (224, 224)  # see openpi `preprocessing_pytorch.py`
+    # Real-Time Chunking (RTC) configuration
+    rtc_config: RTCConfig | None = None
+
+    image_resolution: tuple[int, int] = (
+        DEFAULT_IMAGE_SIZE,
+        DEFAULT_IMAGE_SIZE,
+    )  # see openpi `preprocessing_pytorch.py`
 
     # Add empty images. Used to add empty cameras when no image features are present.
     empty_cameras: int = 0
@@ -65,6 +74,7 @@ class PI05Config(PreTrainedConfig):
     gradient_checkpointing: bool = False  # Enable gradient checkpointing for memory optimization
     compile_model: bool = False  # Whether to use torch.compile for model optimization
     compile_mode: str = "max-autotune"  # Torch compile mode
+    optimizer_fused: bool = False  # Use CUDA fused AdamW kernel
     device: str | None = None  # Device to use for the model (None = auto-detect)
 
     # Optimizer settings: see openpi `AdamW`
@@ -132,6 +142,7 @@ class PI05Config(PreTrainedConfig):
             eps=self.optimizer_eps,
             weight_decay=self.optimizer_weight_decay,
             grad_clip_norm=self.optimizer_grad_clip_norm,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self):

src/lerobot/policies/pi05/modeling_pi05.py

@@ -19,11 +19,12 @@ import logging
 import math
 from collections import deque
 from pathlib import Path
-from typing import TYPE_CHECKING, Literal
+from typing import TYPE_CHECKING, Literal, TypedDict
 
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
+from typing_extensions import Unpack
 
 from lerobot.utils.import_utils import _transformers_available
 
@@ -40,8 +41,9 @@ else:
     PaliGemmaForConditionalGeneration = None
 
 from lerobot.configs.policies import PreTrainedConfig
-from lerobot.policies.pi05.configuration_pi05 import PI05Config
+from lerobot.policies.pi05.configuration_pi05 import DEFAULT_IMAGE_SIZE, PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy, T
+from lerobot.policies.rtc.modeling_rtc import RTCProcessor
 from lerobot.utils.constants import (
     ACTION,
     OBS_LANGUAGE_ATTENTION_MASK,
@@ -50,6 +52,12 @@ from lerobot.utils.constants import (
 )
 
 
+class ActionSelectKwargs(TypedDict, total=False):
+    inference_delay: int | None
+    prev_chunk_left_over: Tensor | None
+    execution_horizon: int | None
+
+
 def get_safe_dtype(target_dtype, device_type):
     """Get a safe dtype for the given device type."""
     if device_type == "mps" and target_dtype == torch.float64:
@@ -328,6 +336,7 @@ class PaliGemmaWithExpertModel(
         action_expert_config,
         use_adarms=None,
         precision: Literal["bfloat16", "float32"] = "bfloat16",
+        image_size: int = DEFAULT_IMAGE_SIZE,
     ):
         if use_adarms is None:
             use_adarms = [False, False]
@@ -347,6 +356,7 @@ class PaliGemmaWithExpertModel(
         vlm_config_hf.text_config.vocab_size = 257152
         vlm_config_hf.text_config.use_adarms = use_adarms[0]
         vlm_config_hf.text_config.adarms_cond_dim = vlm_config.width if use_adarms[0] else None
+        vlm_config_hf.vision_config.image_size = image_size
         vlm_config_hf.vision_config.intermediate_size = 4304
         vlm_config_hf.vision_config.projection_dim = 2048
         vlm_config_hf.vision_config.projector_hidden_act = "gelu_fast"
@@ -502,18 +512,25 @@ class PaliGemmaWithExpertModel(
 class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
     """Core PI05 PyTorch model."""
 
-    def __init__(self, config: PI05Config):
+    def __init__(self, config: PI05Config, rtc_processor: RTCProcessor | None = None):
         super().__init__()
         self.config = config
+        self.rtc_processor = rtc_processor
 
         paligemma_config = get_gemma_config(config.paligemma_variant)
         action_expert_config = get_gemma_config(config.action_expert_variant)
 
+        if config.image_resolution[0] != config.image_resolution[1]:
+            raise ValueError(
+                f"PaliGemma expects square image resolution, invalid resolution: {config.image_resolution}"
+            )
+
         self.paligemma_with_expert = PaliGemmaWithExpertModel(
             paligemma_config,
             action_expert_config,
             use_adarms=[False, True],
             precision=config.dtype,
+            image_size=config.image_resolution[0],
         )
 
         self.action_in_proj = nn.Linear(config.max_action_dim, action_expert_config.width)
@@ -529,6 +546,8 @@ 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"""
 
@@ -556,6 +575,9 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         self.paligemma_with_expert.gemma_expert.model.gradient_checkpointing = False
         logging.info("Disabled gradient checkpointing for PI05Pytorch model")
 
+    def _rtc_enabled(self):
+        return self.config.rtc_config is not None and self.config.rtc_config.enabled
+
     def _apply_checkpoint(self, func, *args, **kwargs):
         """Helper method to apply gradient checkpointing if enabled."""
         if self.gradient_checkpointing_enabled and self.training:
@@ -731,7 +753,16 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         return F.mse_loss(u_t, v_t, reduction="none")
 
     @torch.no_grad()  # see openpi `sample_actions` (slightly adapted)
-    def sample_actions(self, images, img_masks, tokens, masks, noise=None, num_steps=None) -> Tensor:
+    def sample_actions(
+        self,
+        images,
+        img_masks,
+        tokens,
+        masks,
+        noise=None,
+        num_steps=None,
+        **kwargs: Unpack[ActionSelectKwargs],
+    ) -> Tensor:
         """Do a full inference forward and compute the action."""
         if num_steps is None:
             num_steps = self.config.num_inference_steps
@@ -764,20 +795,40 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         )
 
         dt = -1.0 / num_steps
-        dt = torch.tensor(dt, dtype=torch.float32, device=device)
 
         x_t = noise
-        time = torch.tensor(1.0, dtype=torch.float32, device=device)
-        while time >= -dt / 2:
-            expanded_time = time.expand(bsize)
-            v_t = self.denoise_step(
-                prefix_pad_masks,
-                past_key_values,
-                x_t,
-                expanded_time,
-            )
+        for step in range(num_steps):
+            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
+
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                return self.denoise_step(
+                    prefix_pad_masks=prefix_pad_masks,
+                    past_key_values=past_key_values,
+                    x_t=input_x_t,
+                    timestep=current_timestep,
+                )
+
+            if self._rtc_enabled():
+                inference_delay = kwargs.get("inference_delay")
+                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+                execution_horizon = kwargs.get("execution_horizon")
+
+                v_t = self.rtc_processor.denoise_step(
+                    x_t=x_t,
+                    prev_chunk_left_over=prev_chunk_left_over,
+                    inference_delay=inference_delay,
+                    time=time,
+                    original_denoise_step_partial=denoise_step_partial_call,
+                    execution_horizon=execution_horizon,
+                )
+            else:
+                v_t = denoise_step_partial_call(x_t)
+
             x_t = x_t + dt * v_t
-            time += dt
+
+            if self.rtc_processor is not None and self.rtc_processor.is_debug_enabled():
+                self.rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
 
         return x_t
 
@@ -839,7 +890,8 @@ class PI05Policy(PreTrainedPolicy):
         self.config = config
 
         # Initialize the core PI05 model
-        self.model = PI05Pytorch(config)
+        self.init_rtc_processor()
+        self.model = PI05Pytorch(config, rtc_processor=self.rtc_processor)
 
         # Enable gradient checkpointing if requested
         if config.gradient_checkpointing:
@@ -1035,6 +1087,22 @@ class PI05Policy(PreTrainedPolicy):
             ACTION: deque(maxlen=self.config.n_action_steps),
         }
 
+    def init_rtc_processor(self):
+        """Initialize RTC processor if RTC is enabled in config."""
+        self.rtc_processor = None
+
+        # Create processor if config provided
+        # If RTC is not enabled - we can still track the denoising data
+        if self.config.rtc_config is not None:
+            self.rtc_processor = RTCProcessor(self.config.rtc_config)
+
+            model_value = getattr(self, "model", None)
+            if model_value is not None:
+                model_value.rtc_processor = self.rtc_processor
+
+    def _rtc_enabled(self) -> bool:
+        return self.config.rtc_config is not None and self.config.rtc_config.enabled
+
     def _preprocess_images(self, batch: dict[str, Tensor]) -> tuple[list[Tensor], list[Tensor]]:
         """Preprocess images for the model.
 
@@ -1109,6 +1177,10 @@ class PI05Policy(PreTrainedPolicy):
     @torch.no_grad()
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select a single action given environment observations."""
+        assert not self._rtc_enabled(), (
+            "RTC is not supported for select_action, use it with predict_action_chunk"
+        )
+
         self.eval()
 
         # Action queue logic for n_action_steps > 1
@@ -1120,7 +1192,7 @@ class PI05Policy(PreTrainedPolicy):
         return self._action_queue.popleft()
 
     @torch.no_grad()
-    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
+    def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor:
         """Predict a chunk of actions given environment observations."""
         self.eval()
 
@@ -1128,8 +1200,8 @@ class PI05Policy(PreTrainedPolicy):
         images, img_masks = self._preprocess_images(batch)
         tokens, masks = batch[f"{OBS_LANGUAGE_TOKENS}"], batch[f"{OBS_LANGUAGE_ATTENTION_MASK}"]
 
-        # Sample actions using the model (no separate state needed for PI05)
-        actions = self.model.sample_actions(images, img_masks, tokens, masks)
+        # Sample actions using the model (pass through RTC kwargs, no separate state needed for PI05)
+        actions = self.model.sample_actions(images, img_masks, tokens, masks, **kwargs)
 
         # Unpad actions to actual action dimension
         original_action_dim = self.config.output_features[ACTION].shape[0]

src/lerobot/policies/rtc/README.md

@@ -0,0 +1,38 @@
+# Real-Time Chunking (RTC)
+
+This module contains the LeRobot implementation of **Real-Time Chunking (RTC)**, an inference-time technique for flow-matching based policies.
+
+**Note**: RTC is not a policy itself, but rather an inference enhancement that works with flow-matching based policies including [π₀](../pi0/), [π₀.₅](../pi05/), and [SmolVLA](../smolvla/).
+
+---
+
+## Citation
+
+If you use Real-Time Chunking in your work, please cite:
+
+```bibtex
+@misc{openpi2024,
+  author       = {Physical Intelligence Lab},
+  title        = {OpenPI: PyTorch Implementation of π0 and π0.5 Policies},
+  year         = {2024},
+  publisher    = {GitHub},
+  howpublished = {\url{https://github.com/Physical-Intelligence/openpi}},
+  license      = {Apache-2.0}
+}
+
+@misc{black2025realtimeexecutionactionchunking,
+      title={Real-Time Execution of Action Chunking Flow Policies},
+      author={Kevin Black and Manuel Y. Galliker and Sergey Levine},
+      year={2025},
+      eprint={2506.07339},
+      archivePrefix={arXiv},
+      primaryClass={cs.RO},
+      url={https://arxiv.org/abs/2506.07339},
+}
+```
+
+---
+
+## License
+
+This implementation follows the **Apache 2.0 License**, consistent with the LeRobot project.

src/lerobot/policies/rtc/action_queue.py

@@ -0,0 +1,219 @@
+#!/usr/bin/env python
+
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Action queue management for Real-Time Chunking (RTC).
+
+This module provides ActionQueue, a thread-safe queue for managing action chunks
+in real-time control scenarios. It supports both RTC-enabled and non-RTC modes,
+handling action merging and leftover tracking.
+"""
+
+import logging
+from threading import Lock
+
+import torch
+from torch import Tensor
+
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+
+logger = logging.getLogger(__name__)
+
+
+class ActionQueue:
+    """Thread-safe queue for managing action chunks in real-time control.
+
+    This queue handles two types of action sequences:
+    - Original actions: Used for RTC to compute leftovers from previous chunks
+    - Processed actions: Post-processed actions ready for robot execution
+
+    The queue operates in two modes:
+    1. RTC-enabled: Replaces the entire queue with new actions, accounting for inference delay
+    2. RTC-disabled: Appends new actions to the queue, maintaining continuity
+
+    Args:
+        cfg (RTCConfig): Configuration for Real-Time Chunking behavior.
+
+    Attributes:
+        queue (Tensor | None): Processed actions for robot rollout (time_steps, action_dim).
+        original_queue (Tensor | None): Original actions for RTC computation (time_steps, action_dim).
+        last_index (int): Current consumption index in the queue.
+    """
+
+    def __init__(self, cfg: RTCConfig):
+        """Initialize the action queue.
+
+        Args:
+            cfg: RTC configuration controlling queue behavior.
+        """
+        self.queue = None  # Processed actions for robot rollout
+        self.original_queue = None  # Original actions for RTC
+        self.lock = Lock()
+        self.last_index = 0
+        self.cfg = cfg
+
+    def get(self) -> Tensor | None:
+        """Get the next action from the queue.
+
+        Returns:
+            Tensor | None: The next action (action_dim,) or None if queue is empty.
+                          Returns a clone to prevent external modifications.
+        """
+        with self.lock:
+            if self.queue is None or self.last_index >= len(self.queue):
+                return None
+
+            action = self.queue[self.last_index]
+            self.last_index += 1
+            return action.clone()
+
+    def qsize(self) -> int:
+        """Get the number of remaining actions in the queue.
+
+        Returns:
+            int: Number of unconsumed actions.
+        """
+        if self.queue is None:
+            return 0
+        length = len(self.queue)
+        return length - self.last_index
+
+    def empty(self) -> bool:
+        """Check if the queue is empty.
+
+        Returns:
+            bool: True if no actions remain, False otherwise.
+        """
+        if self.queue is None:
+            return True
+
+        length = len(self.queue)
+        return length - self.last_index <= 0
+
+    def get_action_index(self) -> int:
+        """Get the current action consumption index.
+
+        Returns:
+            int: Index of the next action to be consumed.
+        """
+        return self.last_index
+
+    def get_left_over(self) -> Tensor | None:
+        """Get leftover original actions for RTC prev_chunk_left_over.
+
+        These are the unconsumed actions from the current chunk, which will be
+        used by RTC to compute corrections for the next chunk.
+
+        Returns:
+            Tensor | None: Remaining original actions (remaining_steps, action_dim),
+                          or None if no original queue exists.
+        """
+        with self.lock:
+            if self.original_queue is None:
+                return None
+            return self.original_queue[self.last_index :]
+
+    def merge(
+        self,
+        original_actions: Tensor,
+        processed_actions: Tensor,
+        real_delay: int,
+        action_index_before_inference: int | None = 0,
+    ):
+        """Merge new actions into the queue.
+
+        This method operates differently based on RTC mode:
+        - RTC enabled: Replaces the queue, accounting for inference delay
+        - RTC disabled: Appends to the queue, maintaining continuity
+
+        Args:
+            original_actions: Unprocessed actions from policy (time_steps, action_dim).
+            processed_actions: Post-processed actions for robot (time_steps, action_dim).
+            real_delay: Number of time steps of inference delay.
+            action_index_before_inference: Index before inference started, for validation.
+        """
+        with self.lock:
+            self._check_delays(real_delay, action_index_before_inference)
+
+            if self.cfg.enabled:
+                self._replace_actions_queue(original_actions, processed_actions, real_delay)
+                return
+
+            self._append_actions_queue(original_actions, processed_actions)
+
+    def _replace_actions_queue(self, original_actions: Tensor, processed_actions: Tensor, real_delay: int):
+        """Replace the queue with new actions (RTC mode).
+
+        Discards the first `real_delay` actions since they correspond to the time
+        spent during inference, when the robot was executing previous actions.
+
+        Args:
+            original_actions: Unprocessed actions from policy.
+            processed_actions: Post-processed actions for robot.
+            real_delay: Number of time steps to skip due to inference delay.
+        """
+        self.original_queue = original_actions[real_delay:].clone()
+        self.queue = processed_actions[real_delay:].clone()
+
+        logger.debug(f"original_actions shape: {self.original_queue.shape}")
+        logger.debug(f"processed_actions shape: {self.queue.shape}")
+        logger.debug(f"real_delay: {real_delay}")
+
+        self.last_index = 0
+
+    def _append_actions_queue(self, original_actions: Tensor, processed_actions: Tensor):
+        """Append new actions to the queue (non-RTC mode).
+
+        Removes already-consumed actions and appends new ones, maintaining
+        queue continuity without replacement.
+
+        Args:
+            original_actions: Unprocessed actions from policy.
+            processed_actions: Post-processed actions for robot.
+        """
+        if self.queue is None:
+            self.original_queue = original_actions.clone()
+            self.queue = processed_actions.clone()
+            return
+
+        self.original_queue = torch.cat([self.original_queue, original_actions.clone()])
+        self.original_queue = self.original_queue[self.last_index :]
+
+        self.queue = torch.cat([self.queue, processed_actions.clone()])
+        self.queue = self.queue[self.last_index :]
+
+        self.last_index = 0
+
+    def _check_delays(self, real_delay: int, action_index_before_inference: int | None = None):
+        """Validate that computed delays match expectations.
+
+        Compares the delay computed from inference latency with the actual
+        number of actions consumed during inference.
+
+        Args:
+            real_delay: Delay computed from inference latency.
+            action_index_before_inference: Action index when inference started.
+        """
+        if action_index_before_inference is None:
+            return
+
+        indexes_diff = self.last_index - action_index_before_inference
+        if indexes_diff != real_delay:
+            # Let's check that action index difference (real delay calculated based on action queue)
+            # is the same as delay calculated based on inference latency
+            logger.warning(
+                f"[ACTION_QUEUE] Indexes diff is not equal to real delay. "
+                f"Indexes diff: {indexes_diff}, real delay: {real_delay}"
+            )

src/lerobot/policies/rtc/configuration_rtc.py

@@ -0,0 +1,55 @@
+#!/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.
+
+"""
+Real Time Chunking (RTC) and Bidirectional Decoding (BID) configuration classes.
+
+Based on:
+- Real Time Chunking: https://www.physicalintelligence.company/research/real_time_chunking
+"""
+
+from dataclasses import dataclass
+
+from lerobot.configs.types import RTCAttentionSchedule
+
+
+@dataclass
+class RTCConfig:
+    """Configuration for Real Time Chunking (RTC) inference.
+
+    RTC improves real-time inference by treating chunk generation as an inpainting problem,
+    strategically handling overlapping timesteps between action chunks using prefix attention.
+    """
+
+    # Infrastructure
+    enabled: bool = False
+
+    # Core RTC settings
+    # Todo change to exp
+    prefix_attention_schedule: RTCAttentionSchedule = RTCAttentionSchedule.LINEAR
+    max_guidance_weight: float = 10.0
+    execution_horizon: int = 10
+
+    # Debug settings
+    debug: bool = False
+    debug_maxlen: int = 100
+
+    def __post_init__(self):
+        """Validate RTC configuration parameters."""
+        if self.max_guidance_weight <= 0:
+            raise ValueError(f"max_guidance_weight must be positive, got {self.max_guidance_weight}")
+        if self.debug_maxlen <= 0:
+            raise ValueError(f"debug_maxlen must be positive, got {self.debug_maxlen}")

src/lerobot/policies/rtc/debug_tracker.py

@@ -0,0 +1,233 @@
+#!/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.
+
+"""Debug information handler for Real-Time Chunking (RTC)."""
+
+from dataclasses import dataclass, field
+from typing import Any
+
+import torch
+from torch import Tensor
+
+
+@dataclass
+class DebugStep:
+    """Container for debug information from a single denoising step.
+
+    Attributes:
+        step_idx (int): Step index/counter.
+        x_t (Tensor | None): Current latent/state tensor.
+        v_t (Tensor | None): Velocity from denoiser.
+        x1_t (Tensor | None): Denoised prediction (x_t - time * v_t).
+        correction (Tensor | None): Correction gradient tensor.
+        err (Tensor | None): Weighted error term.
+        weights (Tensor | None): Prefix attention weights.
+        guidance_weight (float | Tensor | None): Applied guidance weight.
+        time (float | Tensor | None): Time parameter.
+        inference_delay (int | None): Inference delay parameter.
+        execution_horizon (int | None): Execution horizon parameter.
+        metadata (dict[str, Any]): Additional metadata.
+    """
+
+    step_idx: int = 0
+    x_t: Tensor | None = None
+    v_t: Tensor | None = None
+    x1_t: Tensor | None = None
+    correction: Tensor | None = None
+    err: Tensor | None = None
+    weights: Tensor | None = None
+    guidance_weight: float | Tensor | None = None
+    time: float | Tensor | None = None
+    inference_delay: int | None = None
+    execution_horizon: int | None = None
+    metadata: dict[str, Any] = field(default_factory=dict)
+
+    def to_dict(self, include_tensors: bool = False) -> dict[str, Any]:
+        """Convert debug step to dictionary.
+
+        Args:
+            include_tensors (bool): If True, include tensor values. If False, only include
+                tensor statistics (shape, mean, std, min, max).
+
+        Returns:
+            Dictionary representation of the debug step.
+        """
+        result = {
+            "step_idx": self.step_idx,
+            "guidance_weight": (
+                self.guidance_weight.item()
+                if isinstance(self.guidance_weight, Tensor)
+                else self.guidance_weight
+            ),
+            "time": self.time.item() if isinstance(self.time, Tensor) else self.time,
+            "inference_delay": self.inference_delay,
+            "execution_horizon": self.execution_horizon,
+            "metadata": self.metadata.copy(),
+        }
+
+        # Add tensor information
+        tensor_fields = ["x_t", "v_t", "x1_t", "correction", "err", "weights"]
+        for field_name in tensor_fields:
+            tensor = getattr(self, field_name)
+            if tensor is not None:
+                if include_tensors:
+                    result[field_name] = tensor.detach().cpu()
+                else:
+                    result[f"{field_name}_stats"] = {
+                        "shape": tuple(tensor.shape),
+                        "mean": tensor.mean().item(),
+                        "std": tensor.std().item(),
+                        "min": tensor.min().item(),
+                        "max": tensor.max().item(),
+                    }
+
+        return result
+
+
+class Tracker:
+    """Collects and manages debug information for RTC processing.
+
+    This tracker stores debug information from recent denoising steps in a dictionary,
+    using time as the key for efficient lookups and updates.
+
+    Args:
+        enabled (bool): Whether debug collection is enabled.
+        maxlen (int | None): Optional sliding window size. If provided, only the
+            most recent ``maxlen`` debug steps are kept. If ``None``, keeps all.
+    """
+
+    def __init__(self, enabled: bool = False, maxlen: int = 100):
+        self.enabled = enabled
+        self._steps = {} if enabled else None  # Dictionary with time as key
+        self._maxlen = maxlen
+        self._step_counter = 0
+
+    def reset(self) -> None:
+        """Clear all recorded debug information."""
+        if self.enabled and self._steps is not None:
+            self._steps.clear()
+        self._step_counter = 0
+
+    @torch._dynamo.disable
+    def track(
+        self,
+        time: float | Tensor,
+        x_t: Tensor | None = None,
+        v_t: Tensor | None = None,
+        x1_t: Tensor | None = None,
+        correction: Tensor | None = None,
+        err: Tensor | None = None,
+        weights: Tensor | None = None,
+        guidance_weight: float | Tensor | None = None,
+        inference_delay: int | None = None,
+        execution_horizon: int | None = None,
+        **metadata,
+    ) -> None:
+        """Track debug information for a denoising step at a given time.
+
+        If a step with the given time already exists, it will be updated with the new data.
+        Otherwise, a new step will be created. Only non-None fields are updated/set.
+
+        Note: This method is excluded from torch.compile to avoid graph breaks from
+        operations like .item() which are incompatible with compiled graphs.
+
+        Args:
+            time (float | Tensor): Time parameter - used as the key to identify the step.
+            x_t (Tensor | None): Current latent/state tensor.
+            v_t (Tensor | None): Velocity from denoiser.
+            x1_t (Tensor | None): Denoised prediction.
+            correction (Tensor | None): Correction gradient tensor.
+            err (Tensor | None): Weighted error term.
+            weights (Tensor | None): Prefix attention weights.
+            guidance_weight (float | Tensor | None): Applied guidance weight.
+            inference_delay (int | None): Inference delay parameter.
+            execution_horizon (int | None): Execution horizon parameter.
+            **metadata: Additional metadata to store.
+        """
+        if not self.enabled:
+            return
+
+        # Convert time to float and round to avoid float precision issues
+        time_value = time.item() if isinstance(time, Tensor) else time
+        time_key = round(time_value, 6)  # Use rounded time as dictionary key
+
+        # Check if step with this time already exists
+        if time_key in self._steps:
+            # Update existing step with non-None fields
+            existing_step = self._steps[time_key]
+            if x_t is not None:
+                existing_step.x_t = x_t.detach().clone()
+            if v_t is not None:
+                existing_step.v_t = v_t.detach().clone()
+            if x1_t is not None:
+                existing_step.x1_t = x1_t.detach().clone()
+            if correction is not None:
+                existing_step.correction = correction.detach().clone()
+            if err is not None:
+                existing_step.err = err.detach().clone()
+            if weights is not None:
+                existing_step.weights = weights.detach().clone()
+            if guidance_weight is not None:
+                existing_step.guidance_weight = guidance_weight
+            if inference_delay is not None:
+                existing_step.inference_delay = inference_delay
+            if execution_horizon is not None:
+                existing_step.execution_horizon = execution_horizon
+            if metadata:
+                existing_step.metadata.update(metadata)
+        else:
+            # Create new step
+            step = DebugStep(
+                step_idx=self._step_counter,
+                x_t=x_t.detach().clone() if x_t is not None else None,
+                v_t=v_t.detach().clone() if v_t is not None else None,
+                x1_t=x1_t.detach().clone() if x1_t is not None else None,
+                correction=correction.detach().clone() if correction is not None else None,
+                err=err.detach().clone() if err is not None else None,
+                weights=weights.detach().clone() if weights is not None else None,
+                guidance_weight=guidance_weight,
+                time=time_value,
+                inference_delay=inference_delay,
+                execution_horizon=execution_horizon,
+                metadata=metadata,
+            )
+
+            # Add to dictionary
+            self._steps[time_key] = step
+            self._step_counter += 1
+
+            # Enforce maxlen if set
+            if self._maxlen is not None and len(self._steps) > self._maxlen:
+                # Remove oldest entry (first key in dict - Python 3.7+ preserves insertion order)
+                oldest_key = next(iter(self._steps))
+                del self._steps[oldest_key]
+
+    def get_all_steps(self) -> list[DebugStep]:
+        """Get all recorded debug steps.
+
+        Returns:
+            List of all DebugStep objects (may be empty if disabled).
+        """
+        if not self.enabled or self._steps is None:
+            return []
+
+        return list(self._steps.values())
+
+    def __len__(self) -> int:
+        """Return the number of recorded debug steps."""
+        if not self.enabled or self._steps is None:
+            return 0
+        return len(self._steps)

src/lerobot/policies/rtc/debug_visualizer.py

@@ -0,0 +1,113 @@
+#!/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.
+
+"""Visualization utilities for RTC debug information."""
+
+import torch
+
+
+class RTCDebugVisualizer:
+    """Visualizer for RTC debug information.
+
+    This class provides methods to visualize debug information collected by the Tracker,
+    including corrections, errors, weights, and guidance weights over denoising steps.
+    """
+
+    @staticmethod
+    def plot_waypoints(
+        axes,
+        tensor,
+        start_from: int = 0,
+        color: str = "blue",
+        label: str = "",
+        alpha: float = 0.7,
+        linewidth: float = 2,
+        marker: str | None = None,
+        markersize: int = 4,
+    ):
+        """Plot trajectories across multiple dimensions.
+
+        This function plots a tensor's values across time for multiple dimensions,
+        with each dimension plotted on a separate axis.
+
+        Args:
+            axes: Array of matplotlib axes (one for each dimension).
+            tensor: The tensor to plot (can be torch.Tensor or numpy array).
+                   Shape should be (time_steps, num_dims) or (batch, time_steps, num_dims).
+            start_from: Starting index for the x-axis.
+            color: Color for the plot lines.
+            label: Label for the plot legend.
+            alpha: Transparency level for the plot.
+            linewidth: Width of the plot lines.
+            marker: Marker style for data points (e.g., 'o', 's', '^').
+            markersize: Size of the markers.
+        """
+        import numpy as np
+
+        # Handle None tensor
+        if tensor is None:
+            return
+
+        # Convert tensor to numpy if needed
+        tensor_np = tensor.detach().cpu().numpy() if isinstance(tensor, torch.Tensor) else tensor
+
+        # Handle different tensor shapes
+        if tensor_np.ndim == 3:
+            # If batch dimension present, take first batch
+            tensor_np = tensor_np[0]
+        elif tensor_np.ndim == 1:
+            # If 1D, reshape to (time_steps, 1)
+            tensor_np = tensor_np.reshape(-1, 1)
+
+        # Get dimensions
+        time_steps, num_dims = tensor_np.shape
+
+        # Create x-axis indices
+        x_indices = np.arange(start_from, start_from + time_steps)
+
+        # Plot each dimension on its corresponding axis
+        num_axes = len(axes) if hasattr(axes, "__len__") else 1
+        for dim_idx in range(min(num_dims, num_axes)):
+            ax = axes[dim_idx] if hasattr(axes, "__len__") else axes
+
+            # Plot the trajectory
+            if marker:
+                ax.plot(
+                    x_indices,
+                    tensor_np[:, dim_idx],
+                    color=color,
+                    label=label if dim_idx == 0 else "",  # Only show label once
+                    alpha=alpha,
+                    linewidth=linewidth,
+                    marker=marker,
+                    markersize=markersize,
+                )
+            else:
+                ax.plot(
+                    x_indices,
+                    tensor_np[:, dim_idx],
+                    color=color,
+                    label=label if dim_idx == 0 else "",  # Only show label once
+                    alpha=alpha,
+                    linewidth=linewidth,
+                )
+
+            # Add grid and labels if not already present
+            if not ax.xaxis.get_label().get_text():
+                ax.set_xlabel("Step", fontsize=10)
+            if not ax.yaxis.get_label().get_text():
+                ax.set_ylabel(f"Dim {dim_idx}", fontsize=10)
+            ax.grid(True, alpha=0.3)

src/lerobot/policies/rtc/latency_tracker.py

@@ -0,0 +1,72 @@
+#!/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.
+
+"""Latency tracking utilities for Real-Time Chunking (RTC)."""
+
+from collections import deque
+
+import numpy as np
+
+
+class LatencyTracker:
+    """Tracks recent latencies and provides max/percentile queries.
+
+    Args:
+        maxlen (int | None): Optional sliding window size. If provided, only the
+            most recent ``maxlen`` latencies are kept. If ``None``, keeps all.
+    """
+
+    def __init__(self, maxlen: int = 100):
+        self._values = deque(maxlen=maxlen)
+        self.reset()
+
+    def reset(self) -> None:
+        """Clear all recorded latencies."""
+        self._values.clear()
+        self.max_latency = 0.0
+
+    def add(self, latency: float) -> None:
+        """Add a latency sample (seconds)."""
+        # Ensure numeric and non-negative
+        val = float(latency)
+
+        if val < 0:
+            return
+        self._values.append(val)
+        self.max_latency = max(self.max_latency, val)
+
+    def __len__(self) -> int:
+        return len(self._values)
+
+    def max(self) -> float | None:
+        """Return the maximum latency or None if empty."""
+        return self.max_latency
+
+    def percentile(self, q: float) -> float | None:
+        """Return the q-quantile (q in [0,1]) of recorded latencies or None if empty."""
+        if not self._values:
+            return 0.0
+        q = float(q)
+        if q <= 0.0:
+            return min(self._values)
+        if q >= 1.0:
+            return self.max_latency
+        vals = np.array(list(self._values), dtype=np.float32)
+        return float(np.quantile(vals, q))
+
+    def p95(self) -> float | None:
+        """Return the 95th percentile latency or None if empty."""
+        return self.percentile(0.95)

src/lerobot/policies/rtc/modeling_rtc.py

@@ -0,0 +1,297 @@
+#!/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.
+
+"""
+Real-Time Chunking (RTC) implementation for LeRobot.
+
+Based on Physical Intelligence's Kinetix implementation:
+https://github.com/Physical-Intelligence/real-time-chunking-kinetix/blob/main/src/model.py#L214
+"""
+
+import logging
+import math
+
+import torch
+from torch import Tensor
+
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.policies.rtc.debug_tracker import Tracker
+
+logger = logging.getLogger(__name__)
+
+
+class RTCProcessor:
+    """Real-Time Chunking processor for action chunking policies.
+
+    This class implements RTC techniques including velocity calculation,
+    prefix attention, and adaptive chunk processing.
+    """
+
+    def __init__(self, rtc_config: RTCConfig):
+        self.rtc_config = rtc_config
+
+        self.tracker = None
+
+        if rtc_config.debug:
+            self.tracker = Tracker(
+                enabled=rtc_config.debug,
+                maxlen=rtc_config.debug_maxlen,
+            )
+
+    # ====================== Tracker Proxy Methods ======================
+    def track(
+        self,
+        time: float | Tensor,
+        x_t: Tensor | None = None,
+        v_t: Tensor | None = None,
+        x1_t: Tensor | None = None,
+        correction: Tensor | None = None,
+        err: Tensor | None = None,
+        weights: Tensor | None = None,
+        guidance_weight: float | Tensor | None = None,
+        inference_delay: int | None = None,
+        execution_horizon: int | None = None,
+        **metadata,
+    ) -> None:
+        """Proxy method to track debug information.
+
+        If tracker is None or disabled, this method does nothing.
+        Otherwise, it forwards the call to tracker.track().
+        """
+        if self.tracker is not None:
+            self.tracker.track(
+                time=time,
+                x_t=x_t,
+                v_t=v_t,
+                x1_t=x1_t,
+                correction=correction,
+                err=err,
+                weights=weights,
+                guidance_weight=guidance_weight,
+                inference_delay=inference_delay,
+                execution_horizon=execution_horizon,
+                **metadata,
+            )
+
+    def get_all_debug_steps(self) -> list:
+        """Get all debug steps from tracker.
+
+        Returns empty list if tracker is disabled or None.
+        """
+        if self.tracker is not None:
+            return self.tracker.get_all_steps()
+        return []
+
+    def is_debug_enabled(self) -> bool:
+        """Check if debug tracking is enabled.
+
+        Returns True if tracker exists and is enabled.
+        """
+        return self.tracker is not None and self.tracker.enabled
+
+    def reset_tracker(self) -> None:
+        """Reset the tracker, clearing all recorded steps.
+
+        Does nothing if tracker is None.
+        """
+        if self.tracker is not None:
+            self.tracker.reset()
+
+    # ====================== End Tracker Proxy Methods ======================
+
+    def denoise_step(
+        self,
+        x_t,
+        prev_chunk_left_over,
+        inference_delay,
+        time,
+        original_denoise_step_partial,
+        execution_horizon=None,
+    ) -> Tensor:
+        """RTC guidance wrapper around an existing denoiser.
+
+        This method wraps an original denoising callable that only takes ``x_t`` and
+        returns a base denoised velocity ``v_t``. It then applies Real-Time Chunking
+        (RTC) prefix guidance using the leftover prefix from the previous chunk.
+
+        Args:
+            x_t (Tensor): Current latent/state to denoise. Shape ``(B, T, A)`` or ``(T, A)``.
+            prev_chunk_left_over (Tensor | None): Unexecuted prefix from the previous
+                chunk. Shape ``(B, T_prev, A)`` or ``(T_prev, A)``. If ``None``, no guidance
+                is applied and the method returns ``v_t`` from the original denoiser.
+            inference_delay (int): Number of timesteps from the prefix to use for guidance.
+            time (float | Tensor): Scalar in [0, 1] indicating normalized time. Must be
+                broadcastable with ``x_t``.
+            original_denoise_step_partial (Callable[[Tensor], Tensor]): Callable that
+                computes the base denoised velocity given only ``x_t``.
+            execution_horizon (int | None): Horizon used to build prefix weights. If
+                ``None``, defaults to ``self.rtc_config.execution_horizon``.
+
+        Returns:
+            Tensor: Guided velocity with the same shape as ``v_t``.
+
+        Notes:
+            - If inputs are 2D, a batch dimension is temporarily added and removed at the end.
+            - If ``prev_chunk_left_over`` is shorter than the current chunk length ``T``, it is
+              right-padded with zeros to match ``T``.
+            - Prefix weights are constructed via ``get_prefix_weights(inference_delay, execution_horizon, T)``
+              and broadcast to ``(B, T, A)``.
+            - Guidance correction is computed via autograd using ``x1_t = x_t + time * v_t`` and
+              ``error = (prev_chunk_left_over - x1_t) * weights``.
+            - The final guidance weight is clamped by ``max_guidance_weight`` from the config.
+
+        Reference:
+            https://www.physicalintelligence.company/download/real_time_chunking.pdf
+        """
+
+        # In the original implementation, the time goes from 0 to 1 and
+        # In our implementation, the time goes from 1 to 0
+        # So we need to invert the time
+        tau = 1 - time
+
+        if prev_chunk_left_over is None:
+            # First step, no guidance - return v_t
+            v_t = original_denoise_step_partial(x_t)
+            return v_t
+
+        x_t = x_t.clone().detach()
+
+        squeezed = False
+        if len(x_t.shape) < 3:
+            # Add batch dimension
+            x_t = x_t.unsqueeze(0)
+            squeezed = True
+
+        if len(prev_chunk_left_over.shape) < 3:
+            # Add batch dimension
+            prev_chunk_left_over = prev_chunk_left_over.unsqueeze(0)
+
+        if execution_horizon is None:
+            execution_horizon = self.rtc_config.execution_horizon
+
+        # If the previous action chunk is to short then it doesn't make sense to use long execution horizon
+        # because there is nothing to merge
+        if execution_horizon > prev_chunk_left_over.shape[1]:
+            execution_horizon = prev_chunk_left_over.shape[1]
+
+        batch_size = x_t.shape[0]
+        action_chunk_size = x_t.shape[1]
+        action_dim = x_t.shape[2]
+
+        if prev_chunk_left_over.shape[1] < action_chunk_size or prev_chunk_left_over.shape[2] < action_dim:
+            padded = torch.zeros(batch_size, action_chunk_size, action_dim).to(x_t.device)
+            padded[:, : prev_chunk_left_over.shape[1], : prev_chunk_left_over.shape[2]] = prev_chunk_left_over
+            prev_chunk_left_over = padded
+
+        assert prev_chunk_left_over.shape == x_t.shape, (
+            "The padded previous chunk must be the same size as the input tensor"
+        )
+
+        weights = (
+            self.get_prefix_weights(inference_delay, execution_horizon, action_chunk_size)
+            .to(x_t.device)
+            .unsqueeze(0)
+            .unsqueeze(-1)
+        )
+
+        with torch.enable_grad():
+            v_t = original_denoise_step_partial(x_t)
+            x_t.requires_grad_(True)
+
+            x1_t = x_t - time * v_t  # noqa: N806
+            err = (prev_chunk_left_over - x1_t) * weights
+            grad_outputs = err.clone().detach()
+            correction = torch.autograd.grad(x1_t, x_t, grad_outputs, retain_graph=False)[0]
+
+        max_guidance_weight = torch.as_tensor(self.rtc_config.max_guidance_weight)
+        tau_tensor = torch.as_tensor(tau)
+        squared_one_minus_tau = (1 - tau_tensor) ** 2
+        inv_r2 = (squared_one_minus_tau + tau_tensor**2) / (squared_one_minus_tau)
+        c = torch.nan_to_num((1 - tau_tensor) / tau_tensor, posinf=max_guidance_weight)
+        guidance_weight = torch.nan_to_num(c * inv_r2, posinf=max_guidance_weight)
+        guidance_weight = torch.minimum(guidance_weight, max_guidance_weight)
+
+        result = v_t - guidance_weight * correction
+
+        # Remove the batch dimension if it was added
+        if squeezed:
+            result = result.squeeze(0)
+            correction = correction.squeeze(0)
+            x1_t = x1_t.squeeze(0)
+            err = err.squeeze(0)
+
+        self.track(
+            time=time,
+            x1_t=x1_t,
+            correction=correction,
+            err=err,
+            weights=weights,
+            guidance_weight=guidance_weight,
+            inference_delay=inference_delay,
+            execution_horizon=execution_horizon,
+        )
+
+        return result
+
+    def get_prefix_weights(self, start, end, total):
+        start = min(start, end)
+
+        if self.rtc_config.prefix_attention_schedule == RTCAttentionSchedule.ZEROS:
+            weights = torch.zeros(total)
+            weights[:start] = 1.0
+        elif self.rtc_config.prefix_attention_schedule == RTCAttentionSchedule.ONES:
+            weights = torch.ones(total)
+            weights[end:] = 0.0
+        elif self.rtc_config.prefix_attention_schedule == RTCAttentionSchedule.LINEAR:
+            lin_weights = self._linweights(start, end, total)
+            weights = self._add_trailing_zeros(lin_weights, total, end)
+            weights = self._add_leading_ones(weights, start, total)
+        elif self.rtc_config.prefix_attention_schedule == RTCAttentionSchedule.EXP:
+            lin_weights = self._linweights(start, end, total)
+            lin_weights = lin_weights * torch.expm1(lin_weights).div(math.e - 1)
+            weights = self._add_trailing_zeros(lin_weights, total, end)
+            weights = self._add_leading_ones(weights, start, total)
+
+        return weights
+
+    def _linweights(self, start, end, total):
+        skip_steps_at_end = max(total - end, 0)
+
+        linspace_steps = total - skip_steps_at_end - start
+
+        if end <= start or linspace_steps <= 0:
+            return torch.tensor([])
+
+        return torch.linspace(1, 0, linspace_steps + 2)[1:-1]
+
+    def _add_trailing_zeros(self, weights, total, end):
+        zeros_len = total - end
+
+        if zeros_len <= 0:
+            return weights
+
+        zeros = torch.zeros(zeros_len)
+        return torch.cat([weights, zeros])
+
+    def _add_leading_ones(self, weights, start, total):
+        ones_len = min(start, total)
+
+        if ones_len <= 0:
+            return weights
+
+        ones = torch.ones(ones_len)
+        return torch.cat([ones, weights])

src/lerobot/policies/smolvla/configuration_smolvla.py

@@ -20,6 +20,7 @@ from lerobot.optim.optimizers import AdamWConfig
 from lerobot.optim.schedulers import (
     CosineDecayWithWarmupSchedulerConfig,
 )
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
 from lerobot.utils.constants import OBS_IMAGES
 
 
@@ -78,6 +79,7 @@ class SmolVLAConfig(PreTrainedConfig):
     optimizer_eps: float = 1e-8
     optimizer_weight_decay: float = 1e-10
     optimizer_grad_clip_norm: float = 10
+    optimizer_fused: bool = False
 
     scheduler_warmup_steps: int = 1_000
     scheduler_decay_steps: int = 30_000
@@ -102,6 +104,9 @@ class SmolVLAConfig(PreTrainedConfig):
     min_period: float = 4e-3  # sensitivity range for the timestep used in sine-cosine positional encoding
     max_period: float = 4.0
 
+    # Real-Time Chunking (RTC) configuration
+    rtc_config: RTCConfig | None = None
+
     def __post_init__(self):
         super().__post_init__()
 
@@ -132,6 +137,7 @@ class SmolVLAConfig(PreTrainedConfig):
             eps=self.optimizer_eps,
             weight_decay=self.optimizer_weight_decay,
             grad_clip_norm=self.optimizer_grad_clip_norm,
+            fused=self.optimizer_fused,
         )
 
     def get_scheduler_preset(self):

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -54,12 +54,15 @@ policy = SmolVLAPolicy.from_pretrained("lerobot/smolvla_base")
 
 import math
 from collections import deque
+from typing import TypedDict
 
 import torch
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
+from typing_extensions import Unpack
 
 from lerobot.policies.pretrained import PreTrainedPolicy
+from lerobot.policies.rtc.modeling_rtc import RTCProcessor
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.policies.smolvla.smolvlm_with_expert import SmolVLMWithExpertModel
 from lerobot.policies.utils import (
@@ -69,6 +72,12 @@ from lerobot.utils.constants import ACTION, OBS_LANGUAGE_ATTENTION_MASK, OBS_LAN
 from lerobot.utils.utils import get_safe_dtype
 
 
+class ActionSelectKwargs(TypedDict, total=False):
+    inference_delay: int | None
+    prev_chunk_left_over: Tensor | None
+    execution_horizon: int | None
+
+
 def create_sinusoidal_pos_embedding(
     time: torch.tensor, dimension: int, min_period: float, max_period: float, device="cpu"
 ) -> Tensor:
@@ -232,8 +241,8 @@ class SmolVLAPolicy(PreTrainedPolicy):
         super().__init__(config)
         config.validate_features()
         self.config = config
-
-        self.model = VLAFlowMatching(config)
+        self.init_rtc_processor()
+        self.model = VLAFlowMatching(config, rtc_processor=self.rtc_processor)
         self.reset()
 
     def reset(self):
@@ -242,10 +251,28 @@ class SmolVLAPolicy(PreTrainedPolicy):
             ACTION: deque(maxlen=self.config.n_action_steps),
         }
 
+    def init_rtc_processor(self):
+        """Initialize RTC processor if RTC is enabled in config."""
+        self.rtc_processor = None
+
+        # Lets create processor if the config provided
+        # If RTC is not enabled - we still can track the denoising data
+        if self.config.rtc_config is not None:
+            self.rtc_processor = RTCProcessor(self.config.rtc_config)
+
+            # In case of calling init_rtc_processor after the model is created
+            # We need to set the rtc_processor to the model
+            # During the normal initialization process the model is not created yet
+            model_value = getattr(self, "model", None)
+            if model_value is not None:
+                model_value.rtc_processor = self.rtc_processor
+
     def get_optim_params(self) -> dict:
         return self.parameters()
 
-    def _get_action_chunk(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+    def _get_action_chunk(
+        self, batch: dict[str, Tensor], noise: Tensor | None = None, **kwargs: Unpack[ActionSelectKwargs]
+    ) -> Tensor:
         # TODO: Check if this for loop is needed.
         # Context: In fact, self.queues contains only ACTION field, and in inference, we don't have action in the batch
         # In the case of offline inference, we have the action in the batch
@@ -260,7 +287,9 @@ class SmolVLAPolicy(PreTrainedPolicy):
         lang_tokens = batch[f"{OBS_LANGUAGE_TOKENS}"]
         lang_masks = batch[f"{OBS_LANGUAGE_ATTENTION_MASK}"]
 
-        actions = self.model.sample_actions(images, img_masks, lang_tokens, lang_masks, state, noise=noise)
+        actions = self.model.sample_actions(
+            images, img_masks, lang_tokens, lang_masks, state, noise=noise, **kwargs
+        )
 
         # Unpad actions
         original_action_dim = self.config.action_feature.shape[0]
@@ -278,30 +307,37 @@ class SmolVLAPolicy(PreTrainedPolicy):
         return batch
 
     @torch.no_grad()
-    def predict_action_chunk(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+    def predict_action_chunk(
+        self, batch: dict[str, Tensor], noise: Tensor | None = None, **kwargs: Unpack[ActionSelectKwargs]
+    ) -> Tensor:
         self.eval()
 
         batch = self._prepare_batch(batch)
         self._queues = populate_queues(self._queues, batch, exclude_keys=[ACTION])
 
-        actions = self._get_action_chunk(batch, noise)
+        actions = self._get_action_chunk(batch, noise, **kwargs)
         return actions
 
     @torch.no_grad()
-    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+    def select_action(
+        self, batch: dict[str, Tensor], noise: Tensor | None = None, **kwargs: Unpack[ActionSelectKwargs]
+    ) -> Tensor:
         """Select a single action given environment observations.
 
         This method wraps `select_actions` in order to return one action at a time for execution in the
         environment. It works by managing the actions in a queue and only calling `select_actions` when the
         queue is empty.
         """
+
+        assert not self._rtc_enabled(), (
+            "RTC is not supported for select_action, use it with predict_action_chunk"
+        )
+
         self.eval()
         batch = self._prepare_batch(batch)
         self._queues = populate_queues(self._queues, batch, exclude_keys=[ACTION])
 
-        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
-        # querying the policy.
-        if len(self._queues[ACTION]) == 0:
+        if self._check_get_actions_condition():
             actions = self._get_action_chunk(batch, noise)
 
             # `self.predict_action_chunk` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
@@ -310,6 +346,12 @@ class SmolVLAPolicy(PreTrainedPolicy):
 
         return self._queues[ACTION].popleft()
 
+    def _check_get_actions_condition(self) -> bool:
+        return len(self._queues[ACTION]) == 0
+
+    def _rtc_enabled(self) -> bool:
+        return self.config.rtc_config is not None and self.config.rtc_config.enabled
+
     def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> dict[str, Tensor]:
         """Do a full training forward pass to compute the loss"""
         if self.config.adapt_to_pi_aloha:
@@ -471,7 +513,7 @@ class VLAFlowMatching(nn.Module):
     └──────────────────────────────┘
     """
 
-    def __init__(self, config: SmolVLAConfig):
+    def __init__(self, config: SmolVLAConfig, rtc_processor: RTCProcessor | None = None):
         super().__init__()
         self.config = config
 
@@ -485,7 +527,6 @@ class VLAFlowMatching(nn.Module):
             num_vlm_layers=self.config.num_vlm_layers,
             self_attn_every_n_layers=self.config.self_attn_every_n_layers,
             expert_width_multiplier=self.config.expert_width_multiplier,
-            device=self.config.device,
         )
         self.state_proj = nn.Linear(
             self.config.max_state_dim, self.vlm_with_expert.config.text_config.hidden_size
@@ -510,6 +551,10 @@ class VLAFlowMatching(nn.Module):
         self.add_image_special_tokens = self.config.add_image_special_tokens
         self.image_end_token = torch.tensor([self.fake_image_token], dtype=torch.long)
         self.prefix_length = self.config.prefix_length
+        self.rtc_processor = rtc_processor
+
+    def _rtc_enabled(self):
+        return self.config.rtc_config is not None and self.config.rtc_config.enabled
 
     def set_requires_grad(self):
         for params in self.state_proj.parameters():
@@ -706,7 +751,16 @@ class VLAFlowMatching(nn.Module):
         losses = F.mse_loss(u_t, v_t, reduction="none")
         return losses
 
-    def sample_actions(self, images, img_masks, lang_tokens, lang_masks, state, noise=None) -> Tensor:
+    def sample_actions(
+        self,
+        images,
+        img_masks,
+        lang_tokens,
+        lang_masks,
+        state,
+        noise=None,
+        **kwargs: Unpack[ActionSelectKwargs],
+    ) -> Tensor:
         """Do a full inference forward and compute the action (batch_size x num_steps x num_motors)"""
         bsize = state.shape[0]
         device = state.device
@@ -729,22 +783,43 @@ class VLAFlowMatching(nn.Module):
             use_cache=self.config.use_cache,
             fill_kv_cache=True,
         )
-        dt = -1.0 / self.config.num_steps
-        dt = torch.tensor(dt, dtype=torch.float32, device=device)
+        num_steps = self.config.num_steps
+        dt = -1.0 / num_steps
 
         x_t = noise
-        time = torch.tensor(1.0, dtype=torch.float32, device=device)
-        while time >= -dt / 2:
-            expanded_time = time.expand(bsize)
-            v_t = self.denoise_step(
-                prefix_pad_masks,
-                past_key_values,
-                x_t,
-                expanded_time,
-            )
-            # Euler step
-            x_t += dt * v_t
-            time += dt
+        for step in range(num_steps):
+            time = 1.0 + step * dt
+            time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
+
+            def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+                return self.denoise_step(
+                    x_t=input_x_t,
+                    prefix_pad_masks=prefix_pad_masks,
+                    past_key_values=past_key_values,
+                    timestep=current_timestep,
+                )
+
+            if self._rtc_enabled():
+                inference_delay = kwargs.get("inference_delay")
+                prev_chunk_left_over = kwargs.get("prev_chunk_left_over")
+                execution_horizon = kwargs.get("execution_horizon")
+
+                v_t = self.rtc_processor.denoise_step(
+                    x_t=x_t,
+                    prev_chunk_left_over=prev_chunk_left_over,
+                    inference_delay=inference_delay,
+                    time=time,
+                    original_denoise_step_partial=denoise_step_partial_call,
+                    execution_horizon=execution_horizon,
+                )
+            else:
+                v_t = denoise_step_partial_call(x_t)
+
+            x_t = x_t + dt * v_t
+
+            if self.rtc_processor is not None and self.rtc_processor.is_debug_enabled():
+                self.rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
+
         return x_t
 
     def denoise_step(

src/lerobot/policies/utils.py

@@ -22,6 +22,8 @@ import numpy as np
 import torch
 from torch import nn
 
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.datasets.utils import build_dataset_frame
 from lerobot.processor import PolicyAction, RobotAction, RobotObservation
 from lerobot.utils.constants import ACTION, OBS_STR
@@ -198,3 +200,42 @@ def make_robot_action(action_tensor: PolicyAction, ds_features: dict[str, dict])
         f"{name}": float(action_tensor[i]) for i, name in enumerate(action_names)
     }
     return act_processed_policy
+
+
+def raise_feature_mismatch_error(
+    provided_features: set[str],
+    expected_features: set[str],
+) -> None:
+    """
+    Raises a standardized ValueError for feature mismatches between dataset/environment and policy config.
+    """
+    missing = expected_features - provided_features
+    extra = provided_features - expected_features
+    # TODO (jadechoghari): provide a dynamic rename map suggestion to the user.
+    raise ValueError(
+        f"Feature mismatch between dataset/environment and policy config.\n"
+        f"- Missing features: {sorted(missing) if missing else 'None'}\n"
+        f"- Extra features: {sorted(extra) if extra else 'None'}\n\n"
+        f"Please ensure your dataset and policy use consistent feature names.\n"
+        f"If your dataset uses different observation keys (e.g., cameras named differently), "
+        f"use the `--rename_map` argument, for example:\n"
+        f'  --rename_map=\'{{"observation.images.left": "observation.images.camera1", '
+        f'"observation.images.top": "observation.images.camera2"}}\''
+    )
+
+
+def validate_visual_features_consistency(
+    cfg: PreTrainedConfig,
+    features: dict[str, PolicyFeature],
+) -> None:
+    """
+    Validates visual feature consistency between a policy config and provided dataset/environment features.
+
+    Args:
+        cfg (PreTrainedConfig): The model or policy configuration containing input_features and type.
+        features (Dict[str, PolicyFeature]): A mapping of feature names to PolicyFeature objects.
+    """
+    expected_visuals = {k for k, v in cfg.input_features.items() if v.type == FeatureType.VISUAL}
+    provided_visuals = {k for k, v in features.items() if v.type == FeatureType.VISUAL}
+    if not provided_visuals.issubset(expected_visuals):
+        raise_feature_mismatch_error(provided_visuals, expected_visuals)

src/lerobot/policies/xvla/__init__.py

@@ -0,0 +1,6 @@
+# register the processor steps
+from lerobot.policies.xvla.processor_xvla import (
+    XVLAAddDomainIdProcessorStep,
+    XVLAImageNetNormalizeProcessorStep,
+    XVLAImageToFloatProcessorStep,
+)

src/lerobot/policies/xvla/action_hub.py

@@ -0,0 +1,588 @@
+# ------------------------------------------------------------------------------
+# Copyright 2025 2toINF and HuggingFace Inc. (https://github.com/2toINF)
+#
+# 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 __future__ import annotations
+
+from collections.abc import Iterable
+
+import torch
+import torch.nn as nn
+
+# =============================================================================
+# Registry
+# =============================================================================
+ACTION_REGISTRY: dict[str, type[BaseActionSpace]] = {}
+
+
+def register_action(name: str):
+    """Decorator for registering a new action space."""
+
+    def _wrap(cls):
+        key = name.lower()
+        if key in ACTION_REGISTRY:
+            raise KeyError(f"ActionSpace '{key}' already registered -> {ACTION_REGISTRY[key]}")
+        ACTION_REGISTRY[key] = cls
+        cls.name = key
+        return cls
+
+    return _wrap
+
+
+def build_action_space(name: str, **kwargs) -> BaseActionSpace:
+    """Instantiate a registered action space by name."""
+    key = name.lower()
+    if key not in ACTION_REGISTRY:
+        raise KeyError(f"Unknown action space '{name}'. Available: {list(ACTION_REGISTRY.keys())}")
+    return ACTION_REGISTRY[key](**kwargs)
+
+
+# =============================================================================
+# Base class
+# =============================================================================
+class BaseActionSpace(nn.Module):
+    """
+    Abstract base class for all action-space definitions.
+
+    Each subclass defines:
+      - `dim_action`: dimension of the action vector.
+      - `gripper_idx`: indices of gripper channels.
+      - `compute_loss(pred, target)`: supervised loss for this space.
+      - `preprocess(proprio, action, mode)`: pre-step modifications.
+      - `postprocess(action)`: post-step corrections (e.g. apply sigmoid).
+    """
+
+    name: str = "base"
+    dim_action: int = 0
+    gripper_idx: tuple[int, ...] = ()
+
+    def __init__(self):
+        super().__init__()
+
+    # ---------------------------------------------------------------------
+    # Core supervised loss
+    # ---------------------------------------------------------------------
+    def compute_loss(self, pred: torch.Tensor, target: torch.Tensor) -> dict[str, torch.Tensor]:
+        raise NotImplementedError
+
+    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> dict[str, torch.Tensor]:
+        """Alias for compute_loss."""
+        return self.compute_loss(pred, target)
+
+    # ---------------------------------------------------------------------
+    # Space-level hooks
+    # ---------------------------------------------------------------------
+    def preprocess(
+        self,
+        proprio: torch.Tensor,
+        action: torch.Tensor,
+        mode: str = "train",
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """Default: return unchanged."""
+        return proprio, action
+
+    def postprocess(self, action: torch.Tensor) -> torch.Tensor:
+        """Default: return unchanged."""
+        return action
+
+
+# =============================================================================
+# Utilities
+# =============================================================================
+def _ensure_indices_valid(dim_action: int, idx: Iterable[int], name: str) -> None:
+    bad = [i for i in idx if i < 0 or i >= dim_action]
+    if bad:
+        raise IndexError(f"{name} contains out-of-range indices {bad} for action dim dim_action={dim_action}")
+
+
+# =============================================================================
+# Implementations
+# =============================================================================
+@register_action("ee6d")
+class EE6DActionSpace(BaseActionSpace):
+    """End-effector layout with xyz, 6D rotation, and gripper channels."""
+
+    dim_action = 20
+    gripper_idx = (9, 19)
+    GRIPPER_SCALE = 1.0
+    XYZ_SCALE = 500.0
+    ROT_SCALE = 10.0
+
+    POS_IDX_1 = (0, 1, 2)
+    POS_IDX_2 = (10, 11, 12)
+    ROT_IDX_1 = (3, 4, 5, 6, 7, 8)
+    ROT_IDX_2 = (13, 14, 15, 16, 17, 18)
+
+    def __init__(self):
+        super().__init__()
+        self.mse = nn.MSELoss()
+        self.bce = nn.BCEWithLogitsLoss()
+
+    def compute_loss(self, pred, target):
+        assert pred.shape == target.shape, "pred/target shapes must match"
+        batch_size, seq_len, action_dim = pred.shape
+        _ensure_indices_valid(action_dim, self.gripper_idx, "gripper_idx")
+
+        # Gripper BCE
+        g_losses = [self.bce(pred[:, :, gi], target[:, :, gi]) for gi in self.gripper_idx]
+        gripper_loss = sum(g_losses) / len(self.gripper_idx) * self.GRIPPER_SCALE
+
+        # XYZ position
+        pos_loss = (
+            self.mse(pred[:, :, self.POS_IDX_1], target[:, :, self.POS_IDX_1])
+            + self.mse(pred[:, :, self.POS_IDX_2], target[:, :, self.POS_IDX_2])
+        ) * self.XYZ_SCALE
+
+        # Rotation 6D
+        rot_loss = (
+            self.mse(pred[:, :, self.ROT_IDX_1], target[:, :, self.ROT_IDX_1])
+            + self.mse(pred[:, :, self.ROT_IDX_2], target[:, :, self.ROT_IDX_2])
+        ) * self.ROT_SCALE
+
+        return {
+            "position_loss": pos_loss,
+            "rotate6D_loss": rot_loss,
+            "gripper_loss": gripper_loss,
+        }
+
+    def preprocess(self, proprio, action, mode="train"):
+        """Zero-out gripper channels in proprio/action."""
+        proprio_m = proprio.clone()
+        action_m = action.clone()
+        proprio_m[..., self.gripper_idx] = 0.0
+        action_m[..., self.gripper_idx] = 0.0
+        return proprio_m, action_m
+
+    def postprocess(self, action: torch.Tensor) -> torch.Tensor:
+        """Apply sigmoid to gripper logits."""
+        if action.size(-1) > max(self.gripper_idx):
+            action[..., self.gripper_idx] = torch.sigmoid(action[..., self.gripper_idx])
+        return action
+
+
+@register_action("joint")
+class JointActionSpace(BaseActionSpace):
+    """Joint-space layout with joints + gripper only."""
+
+    dim_action = 14
+    gripper_idx = (6, 13)
+    GRIPPER_SCALE = 0.1
+    JOINTS_SCALE = 1.0
+
+    def __init__(self):
+        super().__init__()
+        self.mse = nn.MSELoss()
+        self.bce = nn.BCEWithLogitsLoss()
+
+    def compute_loss(self, pred, target):
+        assert pred.shape == target.shape
+        batch_size, seq_len, action_dim = pred.shape
+        _ensure_indices_valid(action_dim, self.gripper_idx, "gripper_idx")
+
+        g_losses = [self.bce(pred[:, :, gi], target[:, :, gi]) for gi in self.gripper_idx]
+        gripper_loss = sum(g_losses) / len(self.gripper_idx) * self.GRIPPER_SCALE
+
+        joints_idx = tuple(i for i in range(action_dim) if i not in set(self.gripper_idx))
+        joints_loss = self.mse(pred[:, :, joints_idx], target[:, :, joints_idx]) * self.JOINTS_SCALE
+
+        return {
+            "joints_loss": joints_loss,
+            "gripper_loss": gripper_loss,
+        }
+
+    def preprocess(self, proprio, action, mode="train"):
+        """Zero-out gripper channels in proprio/action."""
+        proprio_m = proprio.clone()
+        action_m = action.clone()
+        proprio_m[..., self.gripper_idx] = 0.0
+        action_m[..., self.gripper_idx] = 0.0
+        return proprio_m, action_m
+
+    def postprocess(self, action: torch.Tensor) -> torch.Tensor:
+        """Apply sigmoid to gripper logits."""
+        if action.size(-1) > max(self.gripper_idx):
+            action[..., self.gripper_idx] = torch.sigmoid(action[..., self.gripper_idx])
+        return action
+
+
+@register_action("agibot_ee6d")
+class AGIBOTEE6DActionSpace(BaseActionSpace):
+    """AGI-bot variant of EE6DActionSpace using MSE for all components."""
+
+    dim_action = 20
+    gripper_idx = (9, 19)
+    GRIPPER_SCALE = 10.0
+    XYZ_SCALE = 500.0
+    ROT_SCALE = 10.0
+    POS_IDX_1 = (0, 1, 2)
+    POS_IDX_2 = (10, 11, 12)
+    ROT_IDX_1 = (3, 4, 5, 6, 7, 8)
+    ROT_IDX_2 = (13, 14, 15, 16, 17, 18)
+
+    def __init__(self):
+        super().__init__()
+        self.mse = nn.MSELoss()
+
+    def compute_loss(self, pred, target):
+        assert pred.shape == target.shape
+        batch_size, seq_len, action_dim = pred.shape
+        _ensure_indices_valid(action_dim, self.gripper_idx, "gripper_idx")
+
+        gripper_loss = (
+            self.mse(pred[:, :, self.gripper_idx], target[:, :, self.gripper_idx]) * self.GRIPPER_SCALE
+        )
+        pos_loss = (
+            self.mse(pred[:, :, self.POS_IDX_1], target[:, :, self.POS_IDX_1])
+            + self.mse(pred[:, :, self.POS_IDX_2], target[:, :, self.POS_IDX_2])
+        ) * self.XYZ_SCALE
+        rot_loss = (
+            self.mse(pred[:, :, self.ROT_IDX_1], target[:, :, self.ROT_IDX_1])
+            + self.mse(pred[:, :, self.ROT_IDX_2], target[:, :, self.ROT_IDX_2])
+        ) * self.ROT_SCALE
+
+        return {
+            "position_loss": pos_loss,
+            "rotate6D_loss": rot_loss,
+            "gripper_loss": gripper_loss,
+        }
+
+    def preprocess(self, proprio, action, mode="train"):
+        """No preprocessing applied in AGIBOT variant."""
+        return proprio, action
+
+    def postprocess(self, action: torch.Tensor) -> torch.Tensor:
+        """AGIBOT does not postprocess."""
+        return action
+
+
+@register_action("franka_joint7")
+class FrankaJoint7ActionSpace(BaseActionSpace):
+    """
+    Franka Panda joint-space: 7 joints, with gripper.
+
+    - Real robot action dim: 7
+    - Model-facing dim: 20 (padded with zeros)
+      compatible with pretrained VLA models expecting 20D.
+    """
+
+    dim_action = 20  # model dimension
+    REAL_DIM = 7  # actual Franka joints
+
+    JOINTS_SCALE = 1.0
+
+    def __init__(self):
+        super().__init__()
+        self.mse = nn.MSELoss()
+
+    def _pad_to_model_dim(self, x: torch.Tensor) -> torch.Tensor:
+        """Pad 7 → 20 dims (zeros for the dummy channels)."""
+        if x is None:
+            return None
+        if x.size(-1) == self.dim_action:
+            return x
+        if x.size(-1) != self.REAL_DIM:
+            raise ValueError(
+                f"Expected last dim to be {self.REAL_DIM} or {self.dim_action}, got {x.size(-1)}"
+            )
+
+        pad_shape = list(x.shape[:-1]) + [self.dim_action - self.REAL_DIM]  # 13 zeros
+        pad = x.new_zeros(pad_shape)
+        return torch.cat([x, pad], dim=-1)
+
+    def _trim_to_real_dim(self, x: torch.Tensor) -> torch.Tensor:
+        """Trim model output 20 → 7 dims."""
+        return x[..., : self.REAL_DIM]
+
+    def compute_loss(self, pred, target):
+        """
+        pred :  [B, T, 20]
+        target : [B, T, 7] or [B, T, 20]
+
+        Only compute MSE on the first 7 dims.
+        """
+        pred = self._pad_to_model_dim(pred)
+        target = self._pad_to_model_dim(target)
+
+        assert pred.shape == target.shape
+
+        joints_loss = (
+            self.mse(
+                pred[:, :, : self.REAL_DIM],  # use only the first 7 joints
+                target[:, :, : self.REAL_DIM],
+            )
+            * self.JOINTS_SCALE
+        )
+
+        return {"joints_loss": joints_loss}
+
+    def preprocess(self, proprio, action, mode="train"):
+        """
+        During training:
+        - Pad [7] → [20]
+        """
+        return proprio, self._pad_to_model_dim(action)
+
+    def postprocess(self, action: torch.Tensor) -> torch.Tensor:
+        """
+        After model prediction:
+        - Trim [20] → [7] for real robot control.
+        """
+        return self._trim_to_real_dim(action)
+
+
+@register_action("auto")
+class AutoActionSpace(BaseActionSpace):
+    """
+    Auto-detecting action space that adapts to any action dimension.
+
+    - Auto-detects the real action dimension from the policy feature
+    - Model outputs max_dim for compatibility with pretrained models
+    - Loss is computed only on the first real_dim dimensions
+    - Postprocess trims output back to real_dim
+
+    Args:
+        real_dim: The actual action dimension from the dataset/policy feature
+        max_dim: The model's output dimension for pretrained VLA compatibility
+    """
+
+    JOINTS_SCALE = 1.0
+
+    def __init__(self, real_dim: int, max_dim: int):
+        super().__init__()
+        self.real_dim = real_dim
+        self.dim_action = max_dim  # Model-facing dimension
+        self.mse = nn.MSELoss()
+
+    def _pad_to_model_dim(self, x: torch.Tensor) -> torch.Tensor:
+        """Pad real_dim → max_dim (zeros for the dummy channels)."""
+        if x is None:
+            return None
+        if x.size(-1) == self.dim_action:
+            return x
+        if x.size(-1) != self.real_dim:
+            # If dimension doesn't match either, pad/trim to real_dim first
+            if x.size(-1) < self.real_dim:
+                pad_shape = list(x.shape[:-1]) + [self.real_dim - x.size(-1)]
+                pad = x.new_zeros(pad_shape)
+                x = torch.cat([x, pad], dim=-1)
+            else:
+                x = x[..., : self.real_dim]
+
+        pad_shape = list(x.shape[:-1]) + [self.dim_action - self.real_dim]
+        pad = x.new_zeros(pad_shape)
+        return torch.cat([x, pad], dim=-1)
+
+    def _trim_to_real_dim(self, x: torch.Tensor) -> torch.Tensor:
+        """Trim model output max_dim → real_dim."""
+        return x[..., : self.real_dim]
+
+    def compute_loss(self, pred: torch.Tensor, target: torch.Tensor) -> dict[str, torch.Tensor]:
+        """
+        Compute loss only on the first real_dim dimensions.
+
+        pred:   [B, T, max_dim] from the model
+        target: [B, T, real_dim] or [B, T, max_dim]
+
+        Loss = MSE(pred[:,:,:real_dim], target[:,:,:real_dim])
+        """
+        pred = self._pad_to_model_dim(pred)
+        target = self._pad_to_model_dim(target)
+        assert pred.shape == target.shape, f"Shape mismatch: pred {pred.shape} vs target {target.shape}"
+
+        # only compute loss on the real dimensions
+        joints_loss = (
+            self.mse(
+                pred[:, :, : self.real_dim],
+                target[:, :, : self.real_dim],
+            )
+            * self.JOINTS_SCALE
+        )
+
+        return {"joints_loss": joints_loss}
+
+    def preprocess(self, proprio: torch.Tensor, action: torch.Tensor, mode: str = "train"):
+        """
+        Pad action from real_dim to max_dim for the model.
+        """
+        return proprio, self._pad_to_model_dim(action)
+
+    def postprocess(self, action: torch.Tensor) -> torch.Tensor:
+        """
+        Trim model output from max_dim to real_dim for real robot control.
+        """
+        return self._trim_to_real_dim(action)
+
+
+@register_action("so101_bimanual")
+class BimanualSO101ActionSpace(BaseActionSpace):
+    """
+    Bimanual SO101 robot: 2 arms with 5 joints each + gripper.
+
+    Layout (real robot):
+    [left_arm (5 joints + gripper), right_arm (5 joints + gripper)]
+    - Left arm:  shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper
+    - Right arm: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper
+
+    Real action dim: 12
+    Model-facing dim: 20 (extra 8 dummy dims at the end)
+    """
+
+    # Model output / training dimension (to match pretrained policy)
+    dim_action = 20
+
+    # Real robot action dimension
+    REAL_DIM = 12
+
+    # Indices of real vs dummy channels
+    REAL_IDXS = tuple(range(REAL_DIM))  # 0..11
+    DUMMY_IDXS = tuple(range(REAL_DIM, dim_action))  # 12..19
+
+    # Grippers live in the real part
+    gripper_idx = (5, 11)  # left_gripper at idx 5, right_gripper at idx 11
+    GRIPPER_SCALE = 1.0
+    JOINTS_SCALE = 1.0
+
+    # Indices for left and right arm joints (excluding grippers)
+    LEFT_ARM_JOINTS = (0, 1, 2, 3, 4)
+    RIGHT_ARM_JOINTS = (6, 7, 8, 9, 10)
+
+    def __init__(self):
+        super().__init__()
+        self.mse = nn.MSELoss()
+        self.bce = nn.BCEWithLogitsLoss()
+
+    # ---------- helpers ----------
+
+    def _pad_to_model_dim(self, x: torch.Tensor) -> torch.Tensor:
+        """If last dim is REAL_DIM (12), pad zeros to reach dim_action (20)."""
+        if x is None:
+            return None
+        if x.size(-1) == self.dim_action:
+            return x
+        if x.size(-1) != self.REAL_DIM:
+            raise ValueError(
+                f"Expected last dim to be {self.REAL_DIM} or {self.dim_action}, got {x.size(-1)}"
+            )
+        pad_shape = list(x.shape[:-1]) + [self.dim_action - self.REAL_DIM]
+        pad = x.new_zeros(pad_shape)
+        return torch.cat([x, pad], dim=-1)
+
+    def _trim_to_real_dim(self, x: torch.Tensor) -> torch.Tensor:
+        """Keep only the first REAL_DIM (12) dims for the real robot."""
+        return x[..., : self.REAL_DIM]
+
+    # ---------- loss ----------
+
+    def compute_loss(self, pred, target):
+        """
+        pred:  [B, T, 20] from the model
+        target: [B, T, 12] or [B, T, 20]
+        We pad target → 20 and compute loss only on the real dims.
+        """
+        # Ensure both are [B, T, 20]
+        pred = self._pad_to_model_dim(pred)
+        target = self._pad_to_model_dim(target)
+        assert pred.shape == target.shape
+
+        # ---- MSE for all real dims (0–11) ----
+        real_dims = 12
+
+        joints_loss = (
+            self.mse(
+                pred[:, :, :real_dims],
+                target[:, :, :real_dims],
+            )
+            * self.JOINTS_SCALE
+        )
+
+        left_arm_loss = self.mse(pred[:, :, :6], target[:, :, :6])
+        right_arm_loss = self.mse(pred[:, :, 6:12], target[:, :, 6:12])
+
+        gripper_loss = (
+            self.mse(
+                pred[:, :, [5, 11]],
+                target[:, :, [5, 11]],
+            )
+            * self.GRIPPER_SCALE
+        )
+
+        return {
+            "joints_loss": joints_loss,
+            "gripper_loss": gripper_loss,
+            "left_arm_loss": left_arm_loss,
+            "right_arm_loss": right_arm_loss,
+        }
+
+    # ---------- preprocess / postprocess ----------
+
+    def preprocess(self, proprio, action, mode="train"):
+        """
+        - If proprio/action are 12-dim, pad them to 20 for the model.
+        - Zero-out gripper channels in proprio/action to focus learning on joints.
+        """
+        proprio_m = self._pad_to_model_dim(proprio.clone())
+        action_m = self._pad_to_model_dim(action.clone()) if action is not None else None
+
+        proprio_m[..., self.gripper_idx] = 0.0
+        if action_m is not None:
+            action_m[..., self.gripper_idx] = 0.0
+
+        return proprio_m, action_m
+
+    def postprocess(self, action: torch.Tensor) -> torch.Tensor:
+        """
+        - Model outputs [*, 20]
+        - Apply sigmoid to gripper logits
+        - Return only the first 12 dims for the real robot:
+          ["left_shoulder_pan.pos",
+           "left_shoulder_lift.pos",
+           "left_elbow_flex.pos",
+           "left_wrist_flex.pos",
+           "left_wrist_roll.pos",
+           "left_gripper.pos",
+           "right_shoulder_pan.pos",
+           "right_shoulder_lift.pos",
+           "right_elbow_flex.pos",
+           "right_wrist_flex.pos",
+           "right_wrist_roll.pos",
+           "right_gripper.pos"]
+        """
+        # Ensure we at least have the real dims + grippers
+        if action.size(-1) < self.REAL_DIM:
+            raise ValueError(f"Expected at least {self.REAL_DIM} dims in action, got {action.size(-1)}")
+
+        # Apply sigmoid on gripper channels in model space (indices 5 and 11)
+        if action.size(-1) > max(self.gripper_idx):
+            action[..., self.gripper_idx] = torch.sigmoid(action[..., self.gripper_idx])
+
+        # Return only the real 12-dim control vector for the env
+        return self._trim_to_real_dim(action)
+
+
+# =============================================================================
+# Exports
+# =============================================================================
+__all__ = [
+    "BaseActionSpace",
+    "build_action_space",
+    "register_action",
+    "EE6DActionSpace",
+    "JointActionSpace",
+    "AGIBOTEE6DActionSpace",
+    "FrankaJoint7ActionSpace",
+    "AutoActionSpace",
+    "BimanualSO101ActionSpace",
+    "ACTION_REGISTRY",
+]

src/lerobot/policies/xvla/configuration_florence2.py

@@ -0,0 +1,353 @@
+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import warnings
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+""" Florence-2 configuration"""
+
+logger = logging.get_logger(__name__)
+
+
+class Florence2VisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Florence2VisionModel`]. It is used to instantiate a Florence2VisionModel
+    according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the Florence2VisionModel architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        drop_path_rate (`float`, *optional*, defaults to 0.1):
+            The dropout rate of the drop path layer.
+        patch_size (`List[int]`, *optional*, defaults to [7, 3, 3, 3]):
+            The patch size of the image.
+        patch_stride (`List[int]`, *optional*, defaults to [4, 2, 2, 2]):
+            The patch stride of the image.
+        patch_padding (`List[int]`, *optional*, defaults to [3, 1, 1, 1]):
+            The patch padding of the image.
+        patch_prenorm (`List[bool]`, *optional*, defaults to [false, true, true, true]):
+            Whether to apply layer normalization before the patch embedding layer.
+        enable_checkpoint (`bool`, *optional*, defaults to False):
+            Whether to enable checkpointing.
+        dim_embed (`List[int]`, *optional*, defaults to [256, 512, 1024, 2048]):
+            The dimension of the embedding layer.
+        num_heads (`List[int]`, *optional*, defaults to [8, 16, 32, 64]):
+            The number of attention heads.
+        num_groups (`List[int]`, *optional*, defaults to [8, 16, 32, 64]):
+            The number of groups.
+        depths (`List[int]`, *optional*, defaults to [1, 1, 9, 1]):
+            The depth of the model.
+        window_size (`int`, *optional*, defaults to 12):
+            The window size of the model.
+        projection_dim (`int`, *optional*, defaults to 1024):
+            The dimension of the projection layer.
+        visual_temporal_embedding (`dict`, *optional*):
+            The configuration of the visual temporal embedding.
+        image_pos_embed (`dict`, *optional*):
+            The configuration of the image position embedding.
+        image_feature_source (`List[str]`, *optional*, defaults to ["spatial_avg_pool", "temporal_avg_pool"]):
+            The source of the image feature.
+    Example:
+
+    ```python
+    >>> from transformers import Florence2VisionConfig, Florence2VisionModel
+
+    >>> # Initializing a Florence2 Vision style configuration
+    >>> configuration = Florence2VisionConfig()
+
+    >>> # Initializing a model (with random weights)
+    >>> model = Florence2VisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "davit"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        drop_path_rate=0.1,
+        patch_size=None,
+        patch_stride=None,
+        patch_padding=None,
+        patch_prenorm=None,
+        enable_checkpoint=False,
+        dim_embed=None,
+        num_heads=None,
+        num_groups=None,
+        depths=None,
+        window_size=12,
+        projection_dim=1024,
+        visual_temporal_embedding=None,
+        image_pos_embed=None,
+        image_feature_source=None,
+        **kwargs,
+    ):
+        self.drop_path_rate = drop_path_rate
+        self.patch_size = patch_size if patch_size is not None else [7, 3, 3, 3]
+        self.patch_stride = patch_stride if patch_stride is not None else [4, 2, 2, 2]
+        self.patch_padding = patch_padding if patch_padding is not None else [3, 1, 1, 1]
+        self.patch_prenorm = patch_prenorm if patch_prenorm is not None else [False, True, True, True]
+        self.enable_checkpoint = enable_checkpoint
+        self.dim_embed = dim_embed if dim_embed is not None else [256, 512, 1024, 2048]
+        self.num_heads = num_heads if num_heads is not None else [8, 16, 32, 64]
+        self.num_groups = num_groups if num_groups is not None else [8, 16, 32, 64]
+        self.depths = depths if depths is not None else [1, 1, 9, 1]
+        self.window_size = window_size
+        self.projection_dim = projection_dim
+
+        if visual_temporal_embedding is None:
+            visual_temporal_embedding = {
+                "type": "COSINE",
+                "max_temporal_embeddings": 100,
+            }
+        self.visual_temporal_embedding = visual_temporal_embedding
+
+        if image_pos_embed is None:
+            image_pos_embed = {
+                "type": "learned_abs_2d",
+                "max_pos_embeddings": 1000,
+            }
+        self.image_pos_embed = image_pos_embed
+
+        self.image_feature_source = (
+            image_feature_source
+            if image_feature_source is not None
+            else ["spatial_avg_pool", "temporal_avg_pool"]
+        )
+
+        super().__init__(**kwargs)
+
+
+class Florence2LanguageConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Florence2LanguagePreTrainedModel`]. It is used to instantiate a BART
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the BART
+    [facebook/bart-large](https://huggingface.co/facebook/bart-large) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 51289):
+            Vocabulary size of the Florence2Language model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Florence2LanguageModel`].
+        d_model (`int`, *optional*, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (`int`, *optional*, defaults to 12):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 12):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (`int`, *optional*, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        scale_embedding (`bool`, *optional*, defaults to `False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        num_labels (`int`, *optional*, defaults to 3):
+            The number of labels to use in [`Florence2LanguageForSequenceClassification`].
+        forced_eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the token to force as the last generated token when `max_length` is reached. Usually set to
+            `eos_token_id`.
+
+    Example:
+
+    ```python
+    >>> from transformers import Florence2LanguageConfig, Florence2LanguageModel
+
+    >>> # Initializing a Florence2 Language style configuration
+    >>> configuration = Florence2LanguageConfig()
+
+    >>> # Initializing a model (with random weights)
+    >>> model = Florence2LanguageModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "florence2_language"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
+
+    def __init__(
+        self,
+        vocab_size=51289,
+        max_position_embeddings=1024,
+        encoder_layers=12,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        decoder_layers=12,
+        decoder_ffn_dim=4096,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        activation_function="gelu",
+        d_model=1024,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        classifier_dropout=0.0,
+        scale_embedding=False,
+        use_cache=True,
+        num_labels=3,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        is_encoder_decoder=True,
+        decoder_start_token_id=2,
+        forced_eos_token_id=2,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+        super().__init__(
+            num_labels=num_labels,
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            forced_eos_token_id=forced_eos_token_id,
+            **kwargs,
+        )
+
+        # ensure backward compatibility for BART CNN models
+        if self.forced_bos_token_id is None and kwargs.get("force_bos_token_to_be_generated", False):
+            self.forced_bos_token_id = self.bos_token_id
+            warnings.warn(
+                f"Please make sure the config includes `forced_bos_token_id={self.bos_token_id}` in future versions. "
+                "The config can simply be saved and uploaded again to be fixed.",
+                stacklevel=2,
+            )
+
+
+class Florence2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Florence2ForConditionalGeneration`]. It is used to instantiate an
+    Florence-2 model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vision_config (`Florence2VisionConfig`,  *optional*):
+            Custom vision config or dict
+        text_config (`Union[AutoConfig, dict]`, *optional*):
+            The config object of the text backbone.
+        ignore_index (`int`, *optional*, defaults to -100):
+            The ignore index for the loss function.
+        vocab_size (`int`, *optional*, defaults to 51289):
+            Vocabulary size of the Florence2model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`~Florence2ForConditionalGeneration`]
+        projection_dim (`int`, *optional*, defaults to 1024):
+            Dimension of the multimodal projection space.
+
+    Example:
+
+    ```python
+    >>> from transformers import Florence2ForConditionalGeneration, Florence2Config, CLIPVisionConfig, BartConfig
+
+    >>> # Initializing a clip-like vision config
+    >>> vision_config = CLIPVisionConfig()
+
+    >>> # Initializing a Bart config
+    >>> text_config = BartConfig()
+
+    >>> # Initializing a Florence-2 configuration
+    >>> configuration = Florence2Config(vision_config, text_config)
+
+    >>> # Initializing a model from the florence-2 configuration
+    >>> model = Florence2ForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "florence2"
+    is_composition = False
+
+    def __init__(
+        self,
+        vision_config=None,
+        text_config=None,
+        ignore_index=-100,
+        vocab_size=51289,
+        projection_dim=1024,
+        **kwargs,
+    ):
+        self.ignore_index = ignore_index
+        self.vocab_size = vocab_size
+        self.projection_dim = projection_dim
+        if vision_config is not None:
+            vision_config = Florence2VisionConfig(**vision_config)
+        self.vision_config = vision_config
+
+        self.text_config = text_config
+        if text_config is not None:
+            self.text_config = Florence2LanguageConfig(**text_config)
+
+        super().__init__(**kwargs)

src/lerobot/policies/xvla/configuration_xvla.py

@@ -0,0 +1,203 @@
+#!/usr/bin/env python
+
+# ------------------------------------------------------------------------------
+# Copyright 2025 The HuggingFace Inc. team and 2toINF (https://github.com/2toINF)
+#
+# 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 __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.optim.optimizers import XVLAAdamWConfig
+from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+from lerobot.utils.constants import OBS_IMAGES
+
+# Conditional import for type checking and lazy loading
+from lerobot.utils.import_utils import _transformers_available
+
+if TYPE_CHECKING or _transformers_available:
+    from .configuration_florence2 import Florence2Config
+else:
+    Florence2Config = None
+
+
+@PreTrainedConfig.register_subclass("xvla")
+@dataclass
+class XVLAConfig(PreTrainedConfig):
+    """
+    Configuration class for the XVLA (Extended Vision-Language-Action) policy so it can
+    plug into the LeRobot training stack.
+
+    The config mirrors the knobs exposed in the original XVLA repository but also
+    declares the input/output feature contract required by LeRobot.
+    """
+
+    # Input / output structure
+    n_obs_steps: int = 1
+    chunk_size: int = 32
+    n_action_steps: int = 32
+    dtype: str = "float32"  # Options: "bfloat16", "float32"
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.IDENTITY,
+            "ACTION": NormalizationMode.IDENTITY,
+        }
+    )
+
+    # Florence2 backbone and tokenizer configuration
+    florence_config: dict[str, Any] = field(default_factory=dict)
+    tokenizer_name: str = "facebook/bart-large"
+    tokenizer_max_length: int = 64
+    tokenizer_padding_side: str = "right"
+    pad_language_to: str = "max_length"
+
+    # Transformer head
+    hidden_size: int = 1024
+    depth: int = 24
+    num_heads: int = 16
+    mlp_ratio: float = 4.0
+    num_domains: int = 30
+    len_soft_prompts: int = 32
+    dim_time: int = 32
+    max_len_seq: int = 512
+    use_hetero_proj: bool = False
+
+    # Action & proprioception
+    action_mode: str = "ee6d"
+    num_denoising_steps: int = 10
+    use_proprio: bool = True
+    max_state_dim: int = 32
+    max_action_dim: int = 20  # Maximum action dimension for padding (used by "auto" action mode)
+    domain_feature_key: str | None = None
+
+    # Vision preprocessing
+    resize_imgs_with_padding: tuple[int, int] | None = None
+    num_image_views: int | None = None
+    empty_cameras: int = 0
+
+    # Freezing options for VLM components
+    # By default, VLM encoders are frozen and only policy transformer + soft prompts train
+    freeze_vision_encoder: bool = False  # Freeze VLM vision encoder weights
+    freeze_language_encoder: bool = False  # Freeze VLM language encoder weights
+    train_policy_transformer: bool = True  # Allow policy transformer to train
+    train_soft_prompts: bool = True  # Allow soft prompts to train
+
+    # Training presets
+    optimizer_lr: float = 1e-4
+    optimizer_betas: tuple[float, float] = (0.9, 0.99)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 0.0
+    optimizer_grad_clip_norm: float = 10.0
+    # Soft-prompt LR settings (for optional warm-up)
+    optimizer_soft_prompt_lr_scale: float = 1.0  # Scale factor for soft-prompt LR
+    optimizer_soft_prompt_warmup_lr_scale: float | None = None  # Start scale for warmup (e.g., 0.01)
+
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+
+        if self.chunk_size <= 0:
+            raise ValueError("`chunk_size` must be strictly positive.")
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"`n_action_steps` ({self.n_action_steps}) must be <= `chunk_size` ({self.chunk_size})."
+            )
+        if self.num_image_views is not None and self.num_image_views <= 0:
+            raise ValueError("`num_image_views` must be > 0 when specified.")
+        if self.dtype not in ["bfloat16", "float32"]:
+            raise ValueError(f"Invalid dtype: {self.dtype}")
+        self._florence_config_obj: Florence2Config | None = None
+
+    def get_florence_config(self) -> Florence2Config:
+        """
+        Build (and cache) the Florence2 transformer config that should back the VLM.
+        """
+        if self._florence_config_obj is None:
+            config_dict = dict(self.florence_config)
+            if "vision_config" not in config_dict or config_dict["vision_config"] is None:
+                raise ValueError("vision_config is required")
+
+            if "text_config" not in config_dict or config_dict["text_config"] is None:
+                raise ValueError("text_config is required")
+            self._florence_config_obj = Florence2Config(**config_dict)
+        return self._florence_config_obj
+
+    def validate_features(self) -> None:
+        if not self.image_features:
+            raise ValueError("XVLA requires at least one visual feature in the inputs.")
+        if self.use_proprio and self.robot_state_feature is None:
+            raise ValueError("`use_proprio=True` requires a proprioceptive state feature.")
+        if self.num_image_views is None:
+            self.num_image_views = len(self.image_features) + self.empty_cameras
+        else:
+            self.num_image_views = max(self.num_image_views, len(self.image_features) + self.empty_cameras)
+
+        if self.empty_cameras > 0:
+            height, width = (480, 640)
+            if self.resize_imgs_with_padding is not None:
+                height, width = self.resize_imgs_with_padding
+            for idx in range(self.empty_cameras):
+                key = f"{OBS_IMAGES}.empty_camera_{idx}"
+                if key not in self.input_features:
+                    self.input_features[key] = PolicyFeature(
+                        type=FeatureType.VISUAL,
+                        shape=(3, height, width),
+                    )
+
+    def get_optimizer_preset(self) -> XVLAAdamWConfig:
+        """Return the XVLA-specific optimizer with differential learning rates.
+
+        This optimizer applies:
+        - 1/10 LR for VLM parameters (stable optimization)
+        - Full LR for transformer/action head
+        - Configurable LR for soft-prompts (with optional warm-up)
+        """
+        return XVLAAdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+            soft_prompt_lr_scale=self.optimizer_soft_prompt_lr_scale,
+            soft_prompt_warmup_lr_scale=self.optimizer_soft_prompt_warmup_lr_scale,
+        )
+
+    def get_scheduler_preset(self) -> CosineDecayWithWarmupSchedulerConfig:
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> list[int] | None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> list[int]:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> list[int] | None:
+        return None

src/lerobot/policies/xvla/modeling_xvla.py

@@ -0,0 +1,548 @@
+#!/usr/bin/env python
+
+# ------------------------------------------------------------------------------
+# Copyright 2025 The HuggingFace Inc. team and 2toINF (https://github.com/2toINF)
+#
+# 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 __future__ import annotations
+
+import builtins
+import logging
+import os
+from collections import deque
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor, nn
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.policies.pretrained import PreTrainedPolicy, T
+from lerobot.policies.utils import populate_queues
+from lerobot.utils.constants import ACTION, OBS_LANGUAGE_TOKENS, OBS_STATE
+
+from .action_hub import build_action_space
+from .configuration_florence2 import Florence2Config
+from .configuration_xvla import XVLAConfig
+from .modeling_florence2 import Florence2ForConditionalGeneration
+from .soft_transformer import SoftPromptedTransformer
+
+
+class XVLAModel(nn.Module):
+    """
+    XVLA backbone that stitches Florence-2 embeddings with the temporal/action transformer head.
+    """
+
+    def __init__(
+        self,
+        config: XVLAConfig,
+        florence_config: Florence2Config,
+        proprio_dim: int,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.chunk_size: int = config.chunk_size
+        self.use_proprio: bool = config.use_proprio
+
+        # Build action space with auto-detection for "auto" mode
+        if config.action_mode.lower() == "auto":
+            # Auto-detect real action dim from config.action_feature
+            real_dim = (
+                config.action_feature.shape[-1]
+                if config.action_feature is not None
+                else config.max_action_dim
+            )
+            self.action_space = build_action_space(
+                config.action_mode.lower(),
+                real_dim=real_dim,
+                max_dim=config.max_action_dim,
+            )
+        else:
+            self.action_space = build_action_space(config.action_mode.lower())
+
+        self.dim_action = self.action_space.dim_action
+        self.dim_proprio = proprio_dim
+
+        self.vlm = Florence2ForConditionalGeneration(florence_config)
+        if hasattr(self.vlm, "language_model"):
+            lm = self.vlm.language_model
+            if hasattr(lm, "model") and hasattr(lm.model, "decoder"):
+                del lm.model.decoder
+            if hasattr(lm, "lm_head"):
+                del lm.lm_head
+
+        projection_dim = getattr(self.vlm.config, "projection_dim", None)
+        if projection_dim is None:
+            raise ValueError("Florence2 config must provide `projection_dim` for multimodal fusion.")
+
+        self.transformer = SoftPromptedTransformer(
+            hidden_size=config.hidden_size,
+            multi_modal_input_size=projection_dim,
+            depth=config.depth,
+            num_heads=config.num_heads,
+            mlp_ratio=config.mlp_ratio,
+            num_domains=config.num_domains,
+            dim_action=self.dim_action,
+            dim_propio=self.dim_proprio,
+            len_soft_prompts=config.len_soft_prompts,
+            dim_time=config.dim_time,
+            max_len_seq=config.max_len_seq,
+            use_hetero_proj=config.use_hetero_proj,
+        )
+
+        # Apply freezing based on config
+        self._apply_freezing()
+
+        # Apply dtype casting based on config
+        self._apply_dtype()
+
+    def _get_target_dtype(self) -> torch.dtype:
+        """Get the target dtype based on config."""
+        if self.config.dtype == "bfloat16":
+            return torch.bfloat16
+        return torch.float32
+
+    def _apply_dtype(self) -> None:
+        """
+        Apply dtype casting to model components based on config.
+        """
+        target_dtype = self._get_target_dtype()
+        self.to(dtype=target_dtype)
+
+    def _apply_freezing(self) -> None:
+        """
+        Freeze VLM vision and language encoders based on config options.
+        Keep only policy transformer and soft prompts trainable.
+        """
+        # Freeze vision encoder
+        if self.config.freeze_vision_encoder and hasattr(self.vlm, "vision_tower"):
+            for param in self.vlm.vision_tower.parameters():
+                param.requires_grad = False
+
+        # Freeze language encoder
+        if self.config.freeze_language_encoder and hasattr(self.vlm, "language_model"):
+            lm = self.vlm.language_model
+            # Freeze encoder
+            if hasattr(lm, "model") and hasattr(lm.model, "encoder"):
+                for param in lm.model.encoder.parameters():
+                    param.requires_grad = False
+            # Freeze shared embeddings
+            if hasattr(lm, "model") and hasattr(lm.model, "shared"):
+                for param in lm.model.shared.parameters():
+                    param.requires_grad = False
+
+        # Freeze or unfreeze policy transformer
+        if not self.config.train_policy_transformer:
+            for name, param in self.transformer.named_parameters():
+                if "soft_prompts" not in name:
+                    param.requires_grad = False
+
+        # Freeze or unfreeze soft prompts
+        if not self.config.train_soft_prompts and hasattr(self.transformer, "soft_prompt_hub"):
+            for param in self.transformer.soft_prompt_hub.parameters():
+                param.requires_grad = False
+
+    def forward_vlm(
+        self,
+        input_ids: torch.LongTensor,
+        pixel_values: torch.FloatTensor,
+        image_mask: torch.Tensor,
+    ) -> dict[str, torch.Tensor]:
+        """
+        Encode text and multi-view images via Florence2 encoder.
+        """
+        batch_size, num_views = pixel_values.shape[:2]
+        flat_mask = image_mask.view(-1).to(dtype=torch.bool)
+        flat_images = pixel_values.flatten(0, 1)
+        num_valid = int(flat_mask.sum().item())
+        if num_valid == 0:
+            raise ValueError("At least one image view must be valid per batch.")
+
+        valid_images = flat_images[flat_mask]
+        valid_feats = self.vlm._encode_image(valid_images)
+        tokens_per_view, hidden_dim = valid_feats.shape[1:]
+
+        image_features = valid_feats.new_zeros((batch_size * num_views, tokens_per_view, hidden_dim))
+        image_features[flat_mask] = valid_feats
+        image_features = image_features.view(batch_size, num_views, tokens_per_view, hidden_dim)
+        inputs_embeds = self.vlm.get_input_embeddings()(input_ids)
+        merged_embeds, attention_mask = self.vlm._merge_input_ids_with_image_features(
+            image_features[:, 0],
+            inputs_embeds,
+        )
+
+        enc_out = self.vlm.language_model.model.encoder(
+            attention_mask=attention_mask,
+            inputs_embeds=merged_embeds,
+        )[0]
+
+        aux_visual_inputs = image_features[:, 1:].reshape(batch_size, -1, hidden_dim)
+        return {"vlm_features": enc_out, "aux_visual_inputs": aux_visual_inputs}
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        image_input: torch.FloatTensor,
+        image_mask: torch.Tensor,
+        domain_id: torch.LongTensor,
+        proprio: torch.Tensor,
+        action: torch.Tensor,
+    ) -> dict[str, torch.Tensor]:
+        """
+        Forward pass for the XVLA model.
+        """
+        target_dtype = self._get_target_dtype()
+        image_input = image_input.to(dtype=target_dtype)
+        proprio = proprio.to(dtype=target_dtype)
+        action = action.to(dtype=target_dtype)
+
+        enc = self.forward_vlm(input_ids, image_input, image_mask)
+
+        batch_size = input_ids.shape[0]
+        t = (
+            torch.rand(1, device=input_ids.device, dtype=target_dtype)
+            + torch.arange(batch_size, device=input_ids.device, dtype=target_dtype) / batch_size
+        ) % (1 - 1e-5)
+
+        action_noisy = torch.randn_like(action) * t.view(-1, 1, 1) + action * (1 - t).view(-1, 1, 1)
+        proprio_m, action_noisy_m = self.action_space.preprocess(proprio, action_noisy)
+
+        pred_action = self.transformer(
+            domain_id=domain_id,
+            action_with_noise=action_noisy_m,
+            t=t,
+            proprio=proprio_m,
+            **enc,
+        )
+        return self.action_space.compute_loss(pred_action, action)
+
+    @torch.no_grad()
+    def generate_actions(
+        self,
+        input_ids: torch.LongTensor,
+        image_input: torch.FloatTensor,
+        image_mask: torch.Tensor,
+        domain_id: torch.LongTensor,
+        proprio: torch.Tensor,
+        steps: int,
+    ) -> torch.Tensor:
+        self.eval()
+
+        target_dtype = self._get_target_dtype()
+        image_input = image_input.to(dtype=target_dtype)
+        proprio = proprio.to(dtype=target_dtype)
+
+        enc = self.forward_vlm(input_ids, image_input, image_mask)
+
+        batch_size = input_ids.shape[0]
+        action_dim = self.dim_action
+
+        x1 = torch.randn(batch_size, self.chunk_size, action_dim, device=proprio.device, dtype=target_dtype)
+        action = torch.zeros_like(x1)
+
+        steps = max(1, int(steps))
+        for i in range(steps, 0, -1):
+            t = torch.full((batch_size,), i / steps, device=proprio.device, dtype=target_dtype)
+            x_t = x1 * t.view(-1, 1, 1) + action * (1 - t).view(-1, 1, 1)
+            proprio_m, x_t_m = self.action_space.preprocess(proprio, x_t)
+            action = self.transformer(
+                domain_id=domain_id,
+                action_with_noise=x_t_m,
+                proprio=proprio_m,
+                t=t,
+                **enc,
+            )
+        return self.action_space.postprocess(action)
+
+
+class XVLAPolicy(PreTrainedPolicy):
+    """LeRobot-compliant wrapper built around the XVLA model."""
+
+    config_class = XVLAConfig
+    name = "xvla"
+
+    def __init__(self, config: XVLAConfig):
+        super().__init__(config)
+        config.validate_features()
+        florence_config = config.get_florence_config()
+        proprio_dim = config.max_state_dim if config.use_proprio else 0
+        self.model = XVLAModel(config=config, florence_config=florence_config, proprio_dim=proprio_dim)
+        self.reset()
+
+    def reset(self) -> None:
+        self._queues = {
+            ACTION: deque(maxlen=self.config.n_action_steps),
+        }
+
+    def get_optim_params(self) -> dict:
+        """Return trainable named parameters for optimization.
+
+        Returns a dict of name -> param for all trainable parameters.
+        This enables the xvla-adamw optimizer to apply differential learning rates
+        based on parameter names (e.g., 1/10 LR for VLM components).
+        """
+        return dict(filter(lambda kv: kv[1].requires_grad, self.named_parameters()))
+
+    def _prepare_state(self, batch: dict[str, Tensor], batch_size: int, device: torch.device) -> Tensor:
+        if not self.config.use_proprio or OBS_STATE not in batch:
+            return torch.zeros(batch_size, 0, device=device)
+        state = batch[OBS_STATE]
+        if state.ndim > 2:
+            state = state[:, -1, :]
+        return pad_vector(state, self.model.dim_proprio)
+
+    def _prepare_images(self, batch: dict[str, Tensor]) -> tuple[Tensor, Tensor]:
+        present_img_keys = [key for key in self.config.image_features if key in batch]
+        if len(present_img_keys) == 0:
+            raise ValueError(
+                "All image features are missing from the batch. "
+                f"Batch keys: {list(batch.keys())}, expected at least one of {list(self.config.image_features)}."
+            )
+
+        images = []
+        masks = []
+        for key in present_img_keys:
+            img = batch[key][:, -1] if batch[key].ndim == 5 else batch[key]
+            if self.config.resize_imgs_with_padding is not None:
+                img = resize_with_pad(img, *self.config.resize_imgs_with_padding)
+            images.append(img)
+            masks.append(torch.ones(img.size(0), dtype=torch.bool, device=img.device))
+
+        stacked_imgs = torch.stack(images, dim=1)
+        stacked_masks = torch.stack(masks, dim=1)
+
+        total_views = self.config.num_image_views or stacked_imgs.size(1)
+        total_views = max(total_views, stacked_imgs.size(1))
+        num_pad = total_views - stacked_imgs.size(1)
+        if num_pad > 0:
+            pad_shape = (stacked_imgs.size(0), num_pad, *stacked_imgs.shape[2:])
+            pad_imgs = stacked_imgs.new_zeros(pad_shape)
+            pad_masks = stacked_masks.new_zeros((stacked_masks.size(0), num_pad))
+            stacked_imgs = torch.cat([stacked_imgs, pad_imgs], dim=1)
+            stacked_masks = torch.cat([stacked_masks, pad_masks], dim=1)
+
+        return stacked_imgs, stacked_masks
+
+    def _get_domain_id(self, batch: dict[str, Tensor], batch_size: int, device: torch.device) -> Tensor:
+        candidate = None
+        if self.config.domain_feature_key and self.config.domain_feature_key in batch:
+            candidate = batch[self.config.domain_feature_key]
+        elif "domain_id" in batch:
+            candidate = batch["domain_id"]
+
+        if candidate is None:
+            return torch.zeros(batch_size, dtype=torch.long, device=device)
+
+        if not isinstance(candidate, torch.Tensor):
+            candidate = torch.as_tensor(candidate, device=device)
+        else:
+            candidate = candidate.to(device=device)
+
+        if candidate.ndim == 0:
+            candidate = candidate.expand(batch_size)
+        if candidate.ndim > 1:
+            candidate = candidate.view(candidate.shape[0], -1)[:, 0]
+        if candidate.shape[0] != batch_size:
+            candidate = candidate.expand(batch_size)
+        return candidate.to(dtype=torch.long)
+
+    def _prepare_action_targets(self, batch: dict[str, Tensor]) -> Tensor:
+        if ACTION not in batch:
+            raise ValueError("Batch is missing action targets required for training.")
+        actions = batch[ACTION]
+        if actions.ndim == 2:
+            actions = actions.unsqueeze(1)
+        actions = pad_tensor_along_dim(actions, self.config.chunk_size, dim=1)
+        if actions.shape[-1] != self.model.dim_action:
+            actions = pad_vector(actions, self.model.dim_action)
+        return actions
+
+    def _build_model_inputs(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        input_ids = batch[OBS_LANGUAGE_TOKENS]
+        batch_size = input_ids.shape[0]
+        images, image_mask = self._prepare_images(batch)
+        domain_id = self._get_domain_id(batch, batch_size, images.device)
+        proprio = self._prepare_state(batch, batch_size, images.device)
+        return {
+            "input_ids": input_ids,
+            "image_input": images,
+            "image_mask": image_mask,
+            "domain_id": domain_id,
+            "proprio": proprio,
+        }
+
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
+        inputs = self._build_model_inputs(batch)
+        targets = self._prepare_action_targets(batch)
+        losses = self.model(action=targets, **inputs)
+        total_loss = sum(losses.values())
+
+        log_dict = {k: v.detach().item() for k, v in losses.items()}
+        log_dict["loss"] = total_loss.detach().item()
+        return total_loss, log_dict
+
+    def _get_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
+        inputs = self._build_model_inputs(batch)
+        actions = self.model.generate_actions(**inputs, steps=self.config.num_denoising_steps)
+        return actions
+
+    @torch.no_grad()
+    def predict_action_chunk(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:  # noqa: ARG002
+        self.eval()
+        self._queues = populate_queues(self._queues, batch, exclude_keys=[ACTION])
+        return self._get_action_chunk(batch)
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:  # noqa: ARG002
+        self.eval()
+        self._queues = populate_queues(self._queues, batch, exclude_keys=[ACTION])
+
+        if len(self._queues[ACTION]) == 0:
+            actions = self._get_action_chunk(batch)
+            self._queues[ACTION].extend(actions.transpose(0, 1)[: self.config.n_action_steps])
+
+        return self._queues[ACTION].popleft()
+
+    @classmethod
+    def from_pretrained(
+        cls: builtins.type[T],
+        pretrained_name_or_path: str | Path,
+        *,
+        config: PreTrainedConfig | None = None,
+        force_download: bool = False,
+        resume_download: bool | None = None,
+        proxies: dict | None = None,
+        token: str | bool | None = None,
+        cache_dir: str | Path | None = None,
+        local_files_only: bool = False,
+        revision: str | None = None,
+        strict: bool = False,
+        **kwargs,
+    ):
+        """
+        Loads XVLA model weights with:
+        - automatic prefix 'model.' added to all keys
+        - skip list for layers that should remain randomly initialized
+        """
+        import safetensors.torch
+
+        # step 1: load config
+        # TODO: jadechoghari, fix this
+        if config is None:
+            config = PreTrainedConfig.from_pretrained(
+                pretrained_name_or_path=pretrained_name_or_path,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                token=token,
+                cache_dir=cache_dir,
+                local_files_only=local_files_only,
+                revision=revision,
+                **kwargs,
+            )
+
+        model_id = str(pretrained_name_or_path)
+        instance = cls(config, **kwargs)
+        # step 2: locate model.safetensors
+        if os.path.isdir(model_id):
+            logging.info("Loading weights from local directory")
+            model_file = os.path.join(model_id, "model.safetensors")
+        else:
+            try:
+                from huggingface_hub import hf_hub_download
+                from huggingface_hub.utils import HfHubHTTPError
+
+                model_file = hf_hub_download(
+                    repo_id=model_id,
+                    filename="model.safetensors",
+                    revision=revision,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    token=token,
+                    local_files_only=local_files_only,
+                )
+            except HfHubHTTPError as e:
+                raise FileNotFoundError(f"model.safetensors not found on the Hub at {model_id}") from e
+
+        logging.info(f"Loading checkpoint from {model_file}")
+        # step 3: load state dict
+        state_dict = safetensors.torch.load_file(model_file)
+        encoder_key = "model.vlm.language_model.model.encoder.embed_tokens.weight"
+        shared_key = "model.vlm.language_model.model.shared.weight"
+        if encoder_key in state_dict:
+            state_dict[shared_key] = state_dict[encoder_key]
+            # or deepcopy
+        # step 4: load into instance
+        instance.load_state_dict(state_dict, strict=True)
+        logging.info("Loaded XVLA checkpoint")
+        # step 5: finalize
+        # Reapply dtype after loading state dict
+        instance.model._apply_dtype()
+        instance.to(config.device)
+        instance.eval()
+        return instance
+
+
+def resize_with_pad(img: torch.Tensor, height: int, width: int, pad_value: float = 0.0) -> torch.Tensor:
+    if img.ndim != 4:
+        raise ValueError(f"(b,c,h,w) expected, but got {img.shape}")
+
+    current_height, current_width = img.shape[2:]
+    if current_height == height and current_width == width:
+        return img
+
+    ratio = max(current_width / width, current_height / height)
+    resized_height = int(current_height / ratio)
+    resized_width = int(current_width / ratio)
+    resized_img = F.interpolate(
+        img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+    )
+
+    pad_height = max(0, height - resized_height)
+    pad_width = max(0, width - resized_width)
+    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
+    return padded_img
+
+
+def pad_vector(vector: Tensor, new_dim: int) -> Tensor:
+    if vector.shape[-1] == new_dim:
+        return vector
+    if new_dim == 0:
+        shape = list(vector.shape)
+        shape[-1] = 0
+        return vector.new_zeros(*shape)
+    shape = list(vector.shape)
+    current_dim = shape[-1]
+    shape[-1] = new_dim
+    new_vector = vector.new_zeros(*shape)
+    length = min(current_dim, new_dim)
+    new_vector[..., :length] = vector[..., :length]
+    return new_vector
+
+
+def pad_tensor_along_dim(tensor: Tensor, target_len: int, dim: int = 1) -> Tensor:
+    current_len = tensor.size(dim)
+    if current_len == target_len:
+        return tensor
+    if current_len > target_len:
+        slices = [slice(None)] * tensor.dim()
+        slices[dim] = slice(0, target_len)
+        return tensor[tuple(slices)]
+    pad_shape = list(tensor.shape)
+    pad_shape[dim] = target_len - current_len
+    pad_tensor = tensor.new_zeros(pad_shape)
+    return torch.cat([tensor, pad_tensor], dim=dim)

src/lerobot/policies/xvla/processor_xvla.py

@@ -0,0 +1,554 @@
+# ------------------------------------------------------------------------------
+# Copyright 2025 The HuggingFace Inc. team and 2toINF (https://github.com/2toINF)
+#
+# 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 typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.datasets.factory import IMAGENET_STATS
+from lerobot.policies.xvla.configuration_xvla import XVLAConfig
+from lerobot.policies.xvla.utils import rotate6d_to_axis_angle
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    NormalizerProcessorStep,
+    ObservationProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+    TokenizerProcessorStep,
+    UnnormalizerProcessorStep,
+)
+from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
+from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.utils.constants import (
+    OBS_IMAGES,
+    OBS_STATE,
+    POLICY_POSTPROCESSOR_DEFAULT_NAME,
+    POLICY_PREPROCESSOR_DEFAULT_NAME,
+)
+
+
+def make_xvla_pre_post_processors(
+    config: XVLAConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Build the LeRobot processor pipelines for XVLA.
+    """
+
+    features = {**config.input_features, **config.output_features}
+    input_steps = [
+        RenameObservationsProcessorStep(rename_map={}),
+        AddBatchDimensionProcessorStep(),
+        TokenizerProcessorStep(
+            tokenizer_name=config.tokenizer_name,
+            max_length=config.tokenizer_max_length,
+            padding=config.pad_language_to,
+            padding_side=config.tokenizer_padding_side,
+        ),
+        XVLAImageToFloatProcessorStep(),
+        XVLAImageNetNormalizeProcessorStep(),
+        XVLAAddDomainIdProcessorStep(),
+        DeviceProcessorStep(device=config.device),
+        NormalizerProcessorStep(
+            features=features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+    ]
+    output_steps = [
+        UnnormalizerProcessorStep(
+            features=config.output_features,
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
+        DeviceProcessorStep(device="cpu"),
+    ]
+
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )
+
+
+# Custom XVLA processor steps
+@dataclass
+class LiberoProcessorStep(ObservationProcessorStep):
+    """
+    Processes LIBERO observations into the LeRobot format.
+
+    This step handles the specific observation structure from LIBERO environments,
+    which includes nested robot_state dictionaries and image observations.
+
+    **State Processing:**
+    -   Processes the `robot_state` dictionary which contains nested end-effector,
+        gripper, and joint information.
+    -   Extracts and concatenates:
+        - End-effector position (3D)
+        - End-effector quaternion converted to axis-angle (3D)
+        - Gripper joint positions (2D)
+    -   Maps the concatenated state to `"observation.state"`.
+
+    **Image Processing:**
+    -   Rotates images by 180 degrees by flipping both height and width dimensions.
+    -   This accounts for the HuggingFaceVLA/libero camera orientation convention.
+    """
+
+    def _process_observation(self, observation):
+        """
+        Processes both image and robot_state observations from LIBERO.
+        """
+        processed_obs = observation.copy()
+        for key in list(processed_obs.keys()):
+            if key.startswith(f"{OBS_IMAGES}."):
+                img = processed_obs[key]
+
+                if key == f"{OBS_IMAGES}.image":
+                    # Flip both H and W
+                    img = torch.flip(img, dims=[2, 3])
+
+                processed_obs[key] = img
+        # Process robot_state into a flat state vector
+        if "observation.robot_state" in processed_obs:
+            robot_state = processed_obs.pop("observation.robot_state")
+
+            # Extract components
+            eef_pos = robot_state["eef"]["pos"]  # (B, 3,)
+            eef_mat = robot_state["eef"]["mat"]  # (B, 3, 3)
+            eef_rot6d = self._mat_to_rotate6d(eef_mat)  # (B, 6)
+
+            extra = torch.zeros((eef_pos.shape[0], 1), dtype=torch.float32, device=eef_pos.device)
+
+            proprio_state = torch.cat((eef_pos, eef_rot6d, extra), dim=-1)  # (B, 10)
+            state = torch.cat((proprio_state, torch.zeros_like(proprio_state)), dim=-1)  # (B, 20)
+            # ensure float32
+            state = state.float()
+            if state.dim() == 1:
+                state = state.unsqueeze(0)
+
+            processed_obs[OBS_STATE] = state
+        return processed_obs
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """
+        Transforms feature keys from the LIBERO format to the LeRobot standard.
+        """
+        new_features: dict[PipelineFeatureType, dict[str, PolicyFeature]] = {}
+
+        # copy over non-STATE features
+        for ft, feats in features.items():
+            if ft != PipelineFeatureType.STATE:
+                new_features[ft] = feats.copy()
+
+        # rebuild STATE features
+        state_feats = {}
+
+        # add our new flattened state
+        state_feats["observation.state"] = PolicyFeature(
+            key="observation.state",
+            shape=(20,),
+            dtype="float32",
+        )
+
+        new_features[PipelineFeatureType.STATE] = state_feats
+
+        return new_features
+
+    def _mat_to_rotate6d(self, rot_mats: torch.Tensor) -> torch.Tensor:
+        """
+        Convert batched rotation matrices (B, 3, 3) into 6D rotation representation (B, 6).
+
+        Args:
+            rot_mats (Tensor): Rotation matrices of shape (B, 3, 3)
+
+        Returns:
+            Tensor: 6D rotation representation, shape (B, 6)
+
+        Raises:
+            TypeError: if input is not a torch tensor
+            ValueError: if shape is not (B, 3, 3)
+        """
+
+        if not isinstance(rot_mats, torch.Tensor):
+            raise TypeError(f"mat_to_rot6d expects a torch.Tensor, got {type(rot_mats)}")
+
+        if rot_mats.ndim != 3 or rot_mats.shape[1:] != (3, 3):
+            raise ValueError(f"mat_to_rot6d expects shape (B, 3, 3), got {tuple(rot_mats.shape)}")
+
+        rot_mats = rot_mats.to(torch.float32)
+
+        col1 = rot_mats[:, :3, 0]  # (B, 3)
+        col2 = rot_mats[:, :3, 1]  # (B, 3)
+
+        rot6d = torch.cat([col1, col2], dim=-1)  # (B, 6)
+
+        return rot6d
+
+    def observation(self, observation):
+        return self._process_observation(observation)
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="xvla_image_scale")
+class XVLAImageScaleProcessorStep(ProcessorStep):
+    """Scale image observations by 255 to convert from [0, 1] to [0, 255] range.
+
+    This processor step multiplies all image observations by 255, which is required
+    for XVLA models that expect images in uint8-like range.
+
+    Args:
+        image_keys: List of observation keys that contain images to scale.
+                   If None, will automatically detect keys starting with "observation.images."
+    """
+
+    image_keys: list[str] | None = None
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Scale image observations by 255."""
+        new_transition = transition.copy()
+        obs = new_transition.get(TransitionKey.OBSERVATION, {})
+        if obs is None:
+            return new_transition
+
+        # Make a copy of observations to avoid modifying the original
+        obs = obs.copy()
+
+        # Determine which keys to scale
+        keys_to_scale = self.image_keys
+        if keys_to_scale is None:
+            # Auto-detect image keys
+            keys_to_scale = [k for k in obs if k.startswith("observation.images.")]
+
+        # Scale each image
+        for key in keys_to_scale:
+            if key in obs and isinstance(obs[key], torch.Tensor):
+                obs[key] = obs[key] * 255
+
+        new_transition[TransitionKey.OBSERVATION] = obs
+        return new_transition
+
+    def transform_features(self, features):
+        """Image scaling doesn't change feature structure."""
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        """Return serializable configuration."""
+        return {
+            "image_keys": self.image_keys,
+        }
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="xvla_image_to_float")
+class XVLAImageToFloatProcessorStep(ProcessorStep):
+    """Convert image observations from [0, 255] to [0, 1] range.
+
+    This processor step divides image observations by 255 to convert from uint8-like
+    range [0, 255] to float range [0, 1]. This is typically used when loading images
+    that are stored as uint8 values.
+
+    Args:
+        image_keys: List of observation keys that contain images to convert.
+                   If None, will automatically detect keys starting with "observation.images."
+        validate_range: If True, validates that input values are in [0, 255] range (default: True)
+
+    Raises:
+        ValueError: If validate_range is True and image values are not in [0, 255] range.
+    """
+
+    image_keys: list[str] | None = None
+    validate_range: bool = True
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Convert image observations from [0, 255] to [0, 1]."""
+        new_transition = transition.copy()
+        obs = new_transition.get(TransitionKey.OBSERVATION, {})
+        if obs is None:
+            return new_transition
+
+        # Make a copy of observations to avoid modifying the original
+        obs = obs.copy()
+
+        # Determine which keys to convert
+        keys_to_convert = self.image_keys
+        if keys_to_convert is None:
+            # Auto-detect image keys
+            keys_to_convert = [k for k in obs if k.startswith("observation.images.")]
+
+        # Convert each image
+        for key in keys_to_convert:
+            if key in obs and isinstance(obs[key], torch.Tensor):
+                tensor = obs[key]
+
+                min_val = tensor.min().item()
+                max_val = tensor.max().item()
+
+                if max_val <= 1.0:
+                    obs[key] = tensor.float()  # ensure float dtype, but no division
+                    continue
+                # Validate that values are in [0, 255] range if requested
+                if self.validate_range and (min_val < 0.0 or max_val > 255.0):
+                    raise ValueError(
+                        f"Image '{key}' has values outside [0, 255] range: "
+                        f"min={min_val:.4f}, max={max_val:.4f}. "
+                        f"Cannot convert to [0, 1] range."
+                    )
+
+                # Convert to float and divide by 255
+                obs[key] = tensor.float() / 255.0
+
+        new_transition[TransitionKey.OBSERVATION] = obs
+        return new_transition
+
+    def transform_features(self, features):
+        """Image conversion doesn't change feature structure."""
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        """Return serializable configuration."""
+        return {
+            "image_keys": self.image_keys,
+            "validate_range": self.validate_range,
+        }
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="xvla_imagenet_normalize")
+class XVLAImageNetNormalizeProcessorStep(ProcessorStep):
+    """Normalize image observations using ImageNet statistics.
+
+    This processor step applies ImageNet normalization (mean and std) to image observations.
+    It validates that input values are in the [0, 1] range before normalizing.
+
+    The normalization formula is: (image - mean) / std
+
+    Args:
+        image_keys: List of observation keys that contain images to normalize.
+                   If None, will automatically detect keys starting with "observation.images."
+
+    Raises:
+        ValueError: If image values are not in the [0, 1] range.
+    """
+
+    image_keys: list[str] | None = None
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Normalize image observations using ImageNet statistics."""
+        new_transition = transition.copy()
+        obs = new_transition.get(TransitionKey.OBSERVATION, {})
+        if obs is None:
+            return new_transition
+
+        # Make a copy of observations to avoid modifying the original
+        obs = obs.copy()
+
+        # Determine which keys to normalize
+        keys_to_normalize = self.image_keys
+        if keys_to_normalize is None:
+            # Auto-detect image keys
+            keys_to_normalize = [k for k in obs if k.startswith("observation.images.")]
+
+        # Normalize each image
+        for key in keys_to_normalize:
+            if key in obs and isinstance(obs[key], torch.Tensor):
+                tensor = obs[key]
+
+                # Validate that values are in [0, 1] range
+                min_val = tensor.min().item()
+                max_val = tensor.max().item()
+                if min_val < 0.0 or max_val > 1.0:
+                    raise ValueError(
+                        f"Image '{key}' has values outside [0, 1] range: "
+                        f"min={min_val:.4f}, max={max_val:.4f}. "
+                        f"ImageNet normalization requires input values in [0, 1]."
+                    )
+
+                # Apply ImageNet normalization
+                mean = torch.tensor(IMAGENET_STATS["mean"], device=tensor.device, dtype=tensor.dtype)
+                std = torch.tensor(IMAGENET_STATS["std"], device=tensor.device, dtype=tensor.dtype)
+
+                # Expand mean/std to match tensor dims (e.g., BCHW or BNCHW)
+                while mean.dim() < tensor.dim():
+                    mean = mean.unsqueeze(0)
+                    std = std.unsqueeze(0)
+
+                # Normalize: (image - mean) / std
+                obs[key] = (tensor - mean) / std
+
+        new_transition[TransitionKey.OBSERVATION] = obs
+        return new_transition
+
+    def transform_features(self, features):
+        """ImageNet normalization doesn't change feature structure."""
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        """Return serializable configuration."""
+        return {
+            "image_keys": self.image_keys,
+        }
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="xvla_add_domain_id")
+class XVLAAddDomainIdProcessorStep(ProcessorStep):
+    """Add domain_id to complementary data.
+
+    This processor step adds a domain_id tensor to the complementary data,
+    which is used by XVLA to identify different robot embodiments or task domains.
+
+    Args:
+        domain_id: The domain ID to add (default: 3)
+    """
+
+    domain_id: int = 0
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Add domain_id to complementary data."""
+        new_transition = transition.copy()
+        comp = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        comp = {} if comp is None else comp.copy()
+
+        # Infer batch size from observation tensors
+        obs = new_transition.get(TransitionKey.OBSERVATION, {})
+        batch_size = 1
+        if obs:
+            for v in obs.values():
+                if isinstance(v, torch.Tensor):
+                    batch_size = v.shape[0]
+                    break
+
+        # Add domain_id tensor
+        comp["domain_id"] = torch.tensor([int(self.domain_id)] * batch_size, dtype=torch.long)
+
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = comp
+        return new_transition
+
+    def transform_features(self, features):
+        """Domain ID addition doesn't change feature structure."""
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        """Return serializable configuration."""
+        return {
+            "domain_id": self.domain_id,
+        }
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="xvla_rotation_6d_to_axis_angle")
+class XVLARotation6DToAxisAngleProcessorStep(ProcessorStep):
+    """Convert 6D rotation representation to axis-angle and reorganize action dimensions.
+
+    This processor step takes actions with 6D rotation representation and converts them to
+    axis-angle representation, reorganizing the action dimensions as:
+    - action[:, :3] -> target_eef (end-effector position)
+    - action[:, 3:9] -> 6D rotation (converted to axis-angle, 3D)
+    - action[:, 9:10] -> gripper action
+
+    Final output: [target_eef (3), axis_angle (3), gripper (1)] = 7D action
+
+    Args:
+        expected_action_dim: Expected input action dimension (default: 10, supports 6D rotation + extras)
+    """
+
+    expected_action_dim: int = 10
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        """Convert 6D rotation to axis-angle in action."""
+        new_transition = transition.copy()
+        action = new_transition.get(TransitionKey.ACTION)
+
+        if action is None or not isinstance(action, torch.Tensor):
+            return new_transition
+
+        # Convert to numpy for processing
+        device = action.device
+        dtype = action.dtype
+        action_np = action.cpu().numpy()
+
+        # Extract components
+        # action shape: (B, D) where D >= 10
+        target_eef = action_np[:, :3]  # (B, 3)
+        rotation_6d = action_np[:, 3:9]  # (B, 6)
+        target_act = action_np[:, 9:10]  # (B, 1)
+
+        # Convert 6D rotation to axis-angle
+        target_axis = rotate6d_to_axis_angle(rotation_6d)  # (B, 3)
+
+        # Concatenate: [eef (3), axis_angle (3), gripper (1)] = 7D
+        action_np = np.concatenate([target_eef, target_axis, target_act], axis=-1)
+
+        # Convert gripper action to -1 or 1
+        action_np[:, -1] = np.where(action_np[:, -1] > 0.5, 1.0, -1.0)
+
+        # Convert back to tensor
+        action = torch.from_numpy(action_np).to(device=device, dtype=dtype)
+
+        new_transition[TransitionKey.ACTION] = action
+        return new_transition
+
+    def transform_features(self, features):
+        """Rotation conversion changes action dimension from 10 to 7."""
+        # Note: This is a simplified version. In practice, you might want to
+        # update the action feature shape in the features dict.
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        """Return serializable configuration."""
+        return {
+            "expected_action_dim": self.expected_action_dim,
+        }
+
+
+def make_xvla_libero_pre_post_processors() -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """
+    Build the LeRobot processor pipelines for XVLA with LIBERO environment.
+    """
+    pre_processor_steps: list[ProcessorStep] = []
+    post_processor_steps: list[ProcessorStep] = []
+    pre_processor_steps.extend(
+        [LiberoProcessorStep(), XVLAImageNetNormalizeProcessorStep(), XVLAAddDomainIdProcessorStep()]
+    )
+    post_processor_steps.extend([XVLARotation6DToAxisAngleProcessorStep()])
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=pre_processor_steps,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=post_processor_steps,
+        ),
+    )