profile

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

.github/workflows/fast_tests.yml

@@ -60,17 +60,12 @@ 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
 
-      - name: Setup /mnt storage
-        run: sudo chown -R $USER:$USER /mnt
-
       # TODO(Steven): Evaluate the need of these dependencies
       - name: Install apt dependencies
         run: |
@@ -85,14 +80,8 @@ jobs:
           version: ${{ env.UV_VERSION }}
           python-version: ${{ env.PYTHON_VERSION }}
 
-      - name: Check disk usage
-        run: df -h
-
       - name: Install lerobot with test extras
         run: uv sync --extra "test"
 
-      - name: Check disk usage
-        run: df -h
-
       - name: Run pytest
         run: uv run pytest tests -vv --maxfail=10

.github/workflows/full_tests.yml

@@ -58,17 +58,12 @@ 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
 
-      - 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 \
@@ -85,21 +80,12 @@ jobs:
       - name: Install lerobot with all extras
         run: uv sync --all-extras --no-extra groot # TODO(Steven): Make flash-attn optional
 
-      - name: Check disk usage
-        run: df -h
-
       - name: Run pytest (all extras)
         run: uv run pytest tests -vv --maxfail=10
 
-      - name: Check disk usage
-        run: df -h
-
       - name: Run end-to-end tests
         run: uv run make test-end-to-end
 
-      - name: Check disk usage
-        run: df -h
-
   # This job builds a GPU enabled image for testing
   # It runs everytime a PR is approved or a push to main
   # TODO(Steven): For now we skip this job for community PRs

.github/workflows/unbound_deps_tests.yml

@@ -45,15 +45,11 @@ jobs:
     runs-on: ubuntu-latest
     env:
       MUJOCO_GL: egl
-      HF_HOME: /mnt/cache/.cache/huggingface
-      HF_LEROBOT_HOME: /mnt/cache/.cache/huggingface/lerobot
     steps:
       - uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
-      - name: Setup /mnt storage
-        run: sudo chown -R $USER:$USER /mnt
 
       - name: Install apt dependencies
         run: |

benchmarks/video/benchmark.py

@@ -0,0 +1,94 @@
+#!/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

@@ -0,0 +1,102 @@
+#!/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

@@ -21,13 +21,11 @@ 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
@@ -37,13 +35,13 @@ 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,
+    decode_video_frames_torchvision,
     encode_video_frames,
 )
 from lerobot.utils.constants import OBS_IMAGE
-from lerobot.utils.utils import TimerManager
 
 BASE_ENCODING = OrderedDict(
     [
@@ -88,7 +86,7 @@ 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")
@@ -99,21 +97,21 @@ 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)
@@ -127,7 +125,7 @@ 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
@@ -151,6 +149,18 @@ 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,
@@ -162,8 +172,8 @@ def benchmark_decoding(
     num_workers: int = 4,
     save_frames: bool = False,
 ) -> dict:
-    def process_sample(sample: int, lock: Lock):
-        time_benchmark = TimerManager(log=False)
+    def process_sample(sample: int):
+        time_benchmark = TimeBenchmark()
         timestamps = sample_timestamps(timestamps_mode, ep_num_images, fps)
         num_frames = len(timestamps)
         result = {
@@ -172,13 +182,13 @@ def benchmark_decoding(
             "mse_values": [],
         }
 
-        with time_benchmark, lock:
+        with time_benchmark:
             frames = decode_video_frames(video_path, timestamps=timestamps, tolerance_s=5e-1, backend=backend)
-        result["load_time_video_ms"] = (time_benchmark.last * 1000) / num_frames
+        result["load_time_video_ms"] = time_benchmark.result_ms / num_frames
 
         with time_benchmark:
             original_frames = load_original_frames(imgs_dir, timestamps, fps)
-        result["load_time_images_ms"] = (time_benchmark.last * 1000) / num_frames
+        result["load_time_images_ms"] = time_benchmark.result_ms / num_frames
 
         frames_np, original_frames_np = frames.numpy(), original_frames.numpy()
         for i in range(num_frames):
@@ -205,10 +215,8 @@ 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, shared_lock) for i in range(num_samples)]
+        futures = [executor.submit(process_sample, i) 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"])
@@ -350,27 +358,24 @@ def main(
                 imgs_dir = output_dir / "images" / dataset.repo_id.replace("/", "_")
                 # We only use the first episode
                 save_first_episode(imgs_dir, dataset)
-                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():
+                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
                         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)
@@ -404,9 +409,9 @@ if __name__ == "__main__":
         nargs="*",
         default=[
             "lerobot/pusht_image",
-            "lerobot/aloha_mobile_shrimp_image",
-            "lerobot/paris_street",
-            "lerobot/kitchen",
+            "aliberts/aloha_mobile_shrimp_image",
+            "aliberts/paris_street",
+            "aliberts/kitchen",
         ],
         help="Datasets repo-ids to test against. First episodes only are used. Must be images.",
     )
@@ -414,7 +419,7 @@ if __name__ == "__main__":
         "--vcodec",
         type=str,
         nargs="*",
-        default=["h264", "hevc", "libsvtav1"],
+        default=["libx264", "hevc", "libsvtav1"],
         help="Video codecs to be tested",
     )
     parser.add_argument(
@@ -463,7 +468,7 @@ if __name__ == "__main__":
         "--backends",
         type=str,
         nargs="*",
-        default=["torchcodec", "pyav"],
+        default=["pyav", "video_reader"],
         help="Torchvision decoding backend to be tested.",
     )
     parser.add_argument(

docs/source/_toctree.yml

@@ -15,6 +15,8 @@
     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"
@@ -37,20 +39,12 @@
     title: π₀.₅ (Pi05)
   - local: groot
     title: NVIDIA GR00T N1.5
-  - local: xvla
-    title: X-VLA
   title: "Policies"
-- sections:
-  - 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: il_sim
+    title: Imitation Learning in Sim
   - local: libero
     title: Using Libero
   - local: metaworld
@@ -65,8 +59,6 @@
     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

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="<user>/smolvla_async",
+    pretrained_name_or_path="fracapuano/smolvla_async",
     chunk_size_threshold=0.5,
     actions_per_chunk=50,  # make sure this is less than the max actions of the policy
 )

docs/source/env_processor.mdx

@@ -1,418 +0,0 @@
-# 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_leisaac.mdx

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

docs/source/il_robots.mdx

@@ -393,7 +393,7 @@ import time
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
 from lerobot.robots.so100_follower.so100_follower import SO100Follower
-from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 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)
 
-    precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
 
 robot.disconnect()
 ```

docs/source/il_sim.mdx

@@ -0,0 +1,220 @@
+# 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/libero.mdx

@@ -62,11 +62,6 @@ 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/rtc.mdx

@@ -1,188 +0,0 @@
-# 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/xvla.mdx

@@ -1,543 +0,0 @@
-# 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.
-
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-architecture.png" width="400">
-
-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.
-
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-architecture2.png" width="400">
-
-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.
-
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/xvla-fold.png" width="400">
-
-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 the following freezing strategy:
-
-```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" \
-  --steps=3000 \
-  --policy.device=cuda \
-  --policy.freeze_vision_encoder=True \
-  --policy.freeze_language_encoder=True \
-  --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`    | `True`  | Freeze the VLM vision encoder weights    |
-| `freeze_language_encoder`  | `True`  | 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, freeze the VLM encoders and only train the policy transformer and soft prompts. This provides excellent sample efficiency with minimal compute.
-
-### 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.repo_id="YOUR_USERNAME/xvla-biso101" \
-  --steps=3000 \
-  --policy.device=cuda \
-  --policy.action_mode=so101_bimanual \
-  --policy.freeze_vision_encoder=True \
-  --policy.freeze_language_encoder=True \
-  --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.
-To enable this behavior, you must:
-
-1. Implement a custom optimizer and register it in your training config
-
-```
-from dataclasses import dataclass, asdict
-from lerobot.optim.optimizers import OptimizerConfig
-import torch
-
-@OptimizerConfig.register_subclass("xvla-adamw")
-@dataclass
-class XVLAAdamW(OptimizerConfig):
-    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
-
-    def build(self, params: dict) -> torch.optim.Optimizer:
-        """
-        Expect `named_parameters()` as input.
-        Apply lr = lr / 10 for all VLM-related parameters.
-        """
-        assert isinstance(params, dict), \
-            "Custom LR optimizer requires `named_parameters()` as inputs."
-        kwargs = asdict(self)
-        kwargs.pop("grad_clip_norm")
-        vlm_group, other_group = [], []
-        for name, p in params.items():
-            if not p.requires_grad:
-                continue
-            if "vlm" in name.lower():
-                vlm_group.append(p)
-            else:
-                other_group.append(p)
-
-        param_groups = [
-            {"params": vlm_group, "lr": self.lr * 0.1, "weight_decay": self.weight_decay * 0.1},
-            {"params": other_group, "lr": self.lr, "weight_decay": self.weight_decay},
-        ]
-
-        return torch.optim.AdamW(param_groups, **kwargs)
-```
-
-2. Modify X-VLA’s get_optim_params to return named parameters
-
-Replace:
-
-```
-def get_optim_params(self) -> dict:
-    """Return only trainable parameters for optimization."""
-    return filter(lambda p: p.requires_grad, self.parameters())
-```
-
-with:
-
-```
-def get_optim_params(self):
-    """Return trainable named parameters."""
-    return filter(lambda kv: kv[1].requires_grad, self.named_parameters())
-```
-
-This ensures the optimizer receives a dict of named parameters, allowing it to correctly detect VLM modules and apply the 1/10 LR rule.
-
-❕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                    |
-| `franka_joint7`  | 7                     | Franka Panda 7-joint control                | Franka robots without gripper        |
-| `so101_bimanual` | 20 (model), 12 (real) | SO101 bimanual robot                        | Bimanual manipulation tasks          |
-
-#### 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.
-
-### 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.
-
-### 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_proprio: bool = True       # Use proprioceptive state
-max_state_dim: int = 32        # Maximum state dimension
-
-# 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**: Use padding (automatic in most action modes)
-
-```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) (coming soon)
-- [LeRobot Documentation](https://github.com/huggingface/lerobot)
-- [Action Registry Implementation](/home/jade_choghari/robot/lerobot/src/lerobot/policies/xvla/action_hub.py)
-- [Processor Implementation](/home/jade_choghari/robot/lerobot/src/lerobot/policies/xvla/processor_xvla.py)
-- [Model Configuration](/home/jade_choghari/robot/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 precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 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
-        precise_sleep(1 / dataset.fps - dt_s)
+        busy_wait(1 / dataset.fps - dt_s)
 
     robot.disconnect()
 

examples/dataset/load_lerobot_dataset.py

@@ -34,106 +34,105 @@ 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)
 
-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)
+# 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)
 
-    # 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)
+# Or simply explore them in your web browser directly at:
+# https://huggingface.co/datasets?other=LeRobot
 
-    # Or simply explore them in your web browser directly at:
-    # https://huggingface.co/datasets?other=LeRobot
+# 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)
 
-    # 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)
+# 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")
 
-    # 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")
+print("Tasks:")
+print(ds_meta.tasks)
+print("Features:")
+pprint(ds_meta.features)
 
-    print("Tasks:")
-    print(ds_meta.tasks)
-    print("Features:")
-    pprint(ds_meta.features)
+# You can also get a short summary by simply printing the object:
+print(ds_meta)
 
-    # You can also get a short summary by simply printing the object:
-    print(ds_meta)
+# 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 then load the actual dataset from the hub.
-    # Either load any subset of episodes:
-    dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
+# 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}")
 
-    # 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}")
+# 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}")
 
-    # 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}")
+# The previous metadata class is contained in the 'meta' attribute of the dataset:
+print(dataset.meta)
 
-    # The previous metadata class is contained in the 'meta' attribute of the dataset:
-    print(dataset.meta)
+# LeRobotDataset actually wraps an underlying Hugging Face dataset
+# (see https://huggingface.co/docs/datasets for more information).
+print(dataset.hf_dataset)
 
-    # LeRobotDataset actually wraps an underlying Hugging Face dataset
-    # (see https://huggingface.co/docs/datasets for more information).
-    print(dataset.hf_dataset)
+# 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]
 
-    # 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]
+# 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)]
 
-    # 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)]
+# The objects returned by the dataset are all torch.Tensors
+print(type(frames[0]))
+print(frames[0].shape)
 
-    # The objects returned by the dataset are all torch.Tensors
-    print(type(frames[0]))
-    print(frames[0].shape)
+# 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.
 
-    # 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.
+# 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.
 
-    # 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)
+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,
@@ -145,7 +144,3 @@ def 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

@@ -33,68 +33,83 @@ 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")
 
-def main():
-    # Create the robot configuration & robot
-    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+robot = LeKiwiClient(robot_config)
 
-    robot = LeKiwiClient(robot_config)
+# Create policy
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
-    # Create policy
-    policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
+# 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}
 
-    # 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 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,
+)
 
-    # Create the dataset
-    dataset = LeRobotDataset.create(
-        repo_id=HF_DATASET_ID,
+# 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,
-        features=dataset_features,
-        robot_type=robot.name,
-        use_videos=True,
-        image_writer_threads=4,
+        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,
     )
 
-    # 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
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (
+        (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+    ):
+        log_say("Reset the environment")
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
             control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
@@ -103,42 +118,21 @@ def main():
             robot_observation_processor=robot_observation_processor,
         )
 
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (
-            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-        ):
-            log_say("Reset the environment")
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=teleop_action_processor,
-                robot_action_processor=robot_action_processor,
-                robot_observation_processor=robot_observation_processor,
-            )
+    if events["rerecord_episode"]:
+        log_say("Re-record episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+    # Save episode
+    dataset.save_episode()
+    recorded_episodes += 1
 
-        # Save episode
-        dataset.save_episode()
-        recorded_episodes += 1
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+listener.stop()
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
-
-    dataset.finalize()
-    dataset.push_to_hub()
-
-
-if __name__ == "__main__":
-    main()
+dataset.finalize()
+dataset.push_to_hub()

examples/lekiwi/record.py

@@ -34,62 +34,78 @@ 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()
 
-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()
+# Initialize the robot and teleoperator
+robot = LeKiwiClient(robot_config)
+leader_arm = SO100Leader(leader_arm_config)
+keyboard = KeyboardTeleop(keyboard_config)
 
-    # Initialize the robot and teleoperator
-    robot = LeKiwiClient(robot_config)
-    leader_arm = SO100Leader(leader_arm_config)
-    keyboard = KeyboardTeleop(keyboard_config)
+# TODO(Steven): Update this example to use pipelines
+teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
 
-    # TODO(Steven): Update this example to use pipelines
-    teleop_action_processor, robot_action_processor, robot_observation_processor = make_default_processors()
+# 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}
 
-    # 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 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,
+)
 
-    # Create the dataset
-    dataset = LeRobotDataset.create(
-        repo_id=HF_REPO_ID,
+# 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,
-        features=dataset_features,
-        robot_type=robot.name,
-        use_videos=True,
-        image_writer_threads=4,
+        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,
     )
 
-    # 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
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (
+        (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
+    ):
+        log_say("Reset the environment")
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
-            dataset=dataset,
             teleop=[leader_arm, keyboard],
-            control_time_s=EPISODE_TIME_SEC,
+            control_time_s=RESET_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=teleop_action_processor,
@@ -97,45 +113,23 @@ def main():
             robot_observation_processor=robot_observation_processor,
         )
 
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (
-            (recorded_episodes < NUM_EPISODES - 1) or events["rerecord_episode"]
-        ):
-            log_say("Reset the environment")
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                teleop=[leader_arm, keyboard],
-                control_time_s=RESET_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=teleop_action_processor,
-                robot_action_processor=robot_action_processor,
-                robot_observation_processor=robot_observation_processor,
-            )
+    if events["rerecord_episode"]:
+        log_say("Re-record episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+    # Save episode
+    dataset.save_episode()
+    recorded_episodes += 1
 
-        # Save episode
-        dataset.save_episode()
-        recorded_episodes += 1
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+leader_arm.disconnect()
+keyboard.disconnect()
+listener.stop()
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    leader_arm.disconnect()
-    keyboard.disconnect()
-    listener.stop()
-
-    dataset.finalize()
-    dataset.push_to_hub()
-
-
-if __name__ == "__main__":
-    main()
+dataset.finalize()
+dataset.push_to_hub()

examples/lekiwi/replay.py

@@ -20,48 +20,42 @@ 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 precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 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")
 
-def main():
-    # Initialize the robot config
-    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
+# Initialize the robot
+robot = LeKiwiClient(robot_config)
 
-    # Initialize the robot
-    robot = LeKiwiClient(robot_config)
+# 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)
 
-    # 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)
+# Connect to the robot
+robot.connect()
 
-    # Connect to the robot
-    robot.connect()
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+print("Starting replay loop...")
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(len(episode_frames)):
+    t0 = time.perf_counter()
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+    # Get recorded action from dataset
+    action = {
+        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+    }
 
-        # Get recorded action from dataset
-        action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+    # Send action to robot
+    _ = robot.send_action(action)
 
-        # Send action to robot
-        _ = robot.send_action(action)
+    busy_wait(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
-        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-
-    robot.disconnect()
-
-
-if __name__ == "__main__":
-    main()
+robot.disconnect()

examples/lekiwi/teleoperate.py

@@ -19,60 +19,54 @@ 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 precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 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")
 
-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")
+# Initialize the robot and teleoperator
+robot = LeKiwiClient(robot_config)
+leader_arm = SO100Leader(teleop_arm_config)
+keyboard = KeyboardTeleop(keyboard_config)
 
-    # Initialize the robot and teleoperator
-    robot = LeKiwiClient(robot_config)
-    leader_arm = SO100Leader(teleop_arm_config)
-    keyboard = KeyboardTeleop(keyboard_config)
+# 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()
 
-    # 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()
+# Init rerun viewer
+init_rerun(session_name="lekiwi_teleop")
 
-    # Init rerun viewer
-    init_rerun(session_name="lekiwi_teleop")
+if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
+    raise ValueError("Robot or teleop is not connected!")
 
-    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+print("Starting teleop loop...")
+while True:
+    t0 = time.perf_counter()
 
-    print("Starting teleop loop...")
-    while True:
-        t0 = time.perf_counter()
+    # Get robot observation
+    observation = robot.get_observation()
 
-        # Get robot observation
-        observation = robot.get_observation()
+    # 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 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)
+    action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
 
-        action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
+    # Send action to robot
+    _ = robot.send_action(action)
 
-        # Send action to robot
-        _ = robot.send_action(action)
+    # Visualize
+    log_rerun_data(observation=observation, action=action)
 
-        # Visualize
-        log_rerun_data(observation=observation, action=action)
-
-        precise_sleep(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))
-
-
-if __name__ == "__main__":
-    main()
+    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/phone_to_so100/evaluate.py

@@ -52,114 +52,125 @@ 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,
+)
 
-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,
-    )
+robot = SO100Follower(robot_config)
 
-    robot = SO100Follower(robot_config)
+# Create policy
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
-    # 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()),
+)
 
-    # 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,
-            ),
+# 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,
         ),
-        robot_type=robot.name,
-        use_videos=True,
-        image_writer_threads=4,
+    ],
+    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,
     )
 
-    # 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
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+        log_say("Reset the environment")
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
             control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
@@ -168,40 +179,21 @@ def main():
             robot_observation_processor=robot_joints_to_ee_pose_processor,
         )
 
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
-            log_say("Reset the environment")
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=make_default_teleop_action_processor(),
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose_processor,
-            )
+    if events["rerecord_episode"]:
+        log_say("Re-record episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+listener.stop()
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
-
-    dataset.finalize()
-    dataset.push_to_hub()
-
-
-if __name__ == "__main__":
-    main()
+dataset.finalize()
+dataset.push_to_hub()

examples/phone_to_so100/record.py

@@ -50,122 +50,133 @@ 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
 
-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
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+phone = Phone(teleop_config)
 
-    # Initialize the robot and teleoperator
-    robot = SO100Follower(robot_config)
-    phone = Phone(teleop_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
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.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
-    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 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,
-            ),
+# 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,
         ),
-        robot_type=robot.name,
-        use_videos=True,
-        image_writer_threads=4,
+        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,
     )
 
-    # 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
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
             teleop=phone,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
+            control_time_s=RESET_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=phone_to_robot_ee_pose_processor,
@@ -173,42 +184,22 @@ def main():
             robot_observation_processor=robot_joints_to_ee_pose,
         )
 
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-            log_say("Reset the environment")
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                teleop=phone,
-                control_time_s=RESET_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=phone_to_robot_ee_pose_processor,
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose,
-            )
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
 
-        if events["rerecord_episode"]:
-            log_say("Re-recording episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+phone.disconnect()
+listener.stop()
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    phone.disconnect()
-    listener.stop()
-
-    dataset.finalize()
-    dataset.push_to_hub()
-
-
-if __name__ == "__main__":
-    main()
+dataset.finalize()
+dataset.push_to_hub()

examples/phone_to_so100/replay.py

@@ -29,78 +29,72 @@ 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 precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 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
+)
 
-def main():
-    # 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)
 
-    # Initialize the robot
-    robot = SO100Follower(robot_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
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.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
-    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=False,  # Because replay is open loop
+        ),
+    ],
+    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=False,  # Because replay is open loop
-            ),
-        ],
-        to_transition=robot_action_observation_to_transition,
-        to_output=transition_to_robot_action,
-    )
+# 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)
 
-    # 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)
+# Connect to the robot
+robot.connect()
 
-    # Connect to the robot
-    robot.connect()
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+print("Starting replay loop...")
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(len(episode_frames)):
+    t0 = time.perf_counter()
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+    # Get recorded action from dataset
+    ee_action = {
+        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+    }
 
-        # Get recorded action from dataset
-        ee_action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+    # Get robot observation
+    robot_obs = robot.get_observation()
 
-        # Get robot observation
-        robot_obs = robot.get_observation()
+    # Dataset EE -> robot joints
+    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-        # Dataset EE -> robot joints
-        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+    # Send action to robot
+    _ = robot.send_action(joint_action)
 
-        # Send action to robot
-        _ = robot.send_action(joint_action)
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
 
-        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
-
-    # Clean up
-    robot.disconnect()
-
-
-if __name__ == "__main__":
-    main()
+# Clean up
+robot.disconnect()

examples/phone_to_so100/teleoperate.py

@@ -32,90 +32,82 @@ 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 precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 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
 
-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
+# Initialize the robot and teleoperator
+robot = SO100Follower(robot_config)
+teleop_device = Phone(teleop_config)
 
-    # Initialize the robot and teleoperator
-    robot = SO100Follower(robot_config)
-    teleop_device = Phone(teleop_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
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.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
-    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 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,
+)
 
-    # 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,
-    )
+# Connect to the robot and teleoperator
+robot.connect()
+teleop_device.connect()
 
-    # Connect to the robot and teleoperator
-    robot.connect()
-    teleop_device.connect()
+# Init rerun viewer
+init_rerun(session_name="phone_so100_teleop")
 
-    # Init rerun viewer
-    init_rerun(session_name="phone_so100_teleop")
+if not robot.is_connected or not teleop_device.is_connected:
+    raise ValueError("Robot or teleop is not connected!")
 
-    if not robot.is_connected or not teleop_device.is_connected:
-        raise ValueError("Robot or teleop is not connected!")
+print("Starting teleop loop. Move your phone to teleoperate the robot...")
+while True:
+    t0 = time.perf_counter()
 
-    print("Starting teleop loop. Move your phone to teleoperate the robot...")
-    while True:
-        t0 = time.perf_counter()
+    # Get robot observation
+    robot_obs = robot.get_observation()
 
-        # Get robot observation
-        robot_obs = robot.get_observation()
+    # Get teleop action
+    phone_obs = teleop_device.get_action()
 
-        # Get teleop action
-        phone_obs = teleop_device.get_action()
+    # Phone -> EE pose -> Joints transition
+    joint_action = phone_to_robot_joints_processor((phone_obs, robot_obs))
 
-        # Phone -> EE pose -> Joints transition
-        joint_action = phone_to_robot_joints_processor((phone_obs, robot_obs))
+    # Send action to robot
+    _ = robot.send_action(joint_action)
 
-        # Send action to robot
-        _ = robot.send_action(joint_action)
+    # Visualize
+    log_rerun_data(observation=phone_obs, action=joint_action)
 
-        # 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()
+    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/port_datasets/slurm_aggregate_shards.py

@@ -15,12 +15,16 @@
 # 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_droid import DROID_SHARDS
+from port_datasets.droid_rlds.port_droid import DROID_SHARDS
+
+from lerobot.datasets.aggregate import aggregate_datasets
+from lerobot.utils.utils import init_logging
 
 
 class AggregateDatasets(PipelineStep):
@@ -34,11 +38,6 @@ 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_droid import DROID_SHARDS
+from port_datasets.droid_rlds.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_droid import port_droid, validate_dataset
+        from port_datasets.droid_rlds.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_droid import DROID_SHARDS
+from port_datasets.droid_rlds.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, private=False, slurm=True
+    repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
 ):
     kwargs = {
         "pipeline": [
-            UploadDataset(repo_id, private=private),
+            UploadDataset(repo_id),
         ],
         "logging_dir": str(logs_dir / job_name),
     }
@@ -267,12 +267,6 @@ 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/PROFILING_GUIDE.md

@@ -0,0 +1,263 @@
+# RTC Profiling Guide
+
+This guide explains how to profile RTC (Real-Time Chunking) performance to identify bottlenecks and understand why RTC might be slower than expected.
+
+## Quick Start
+
+### 1. Profile with Real Robot (Profiled Version)
+
+Use `eval_with_real_robot_profiled.py` to profile actual robot execution:
+
+```bash
+# With RTC enabled
+uv run examples/rtc/eval_with_real_robot_profiled.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=30
+
+# Without RTC for comparison
+uv run examples/rtc/eval_with_real_robot_profiled.py \
+    --policy.path=helper2424/pi05_check_rtc \
+    --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: 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=30
+```
+
+**Output**: At the end of execution, you'll see a detailed breakdown of timing for each component:
+- `get_actions.policy_inference` - Time spent in policy inference
+- `get_actions.preprocessing` - Time spent preprocessing observations
+- `get_actions.postprocessing` - Time spent postprocessing actions
+- `get_actions.action_queue_merge` - Time spent merging actions with RTC
+- `robot.get_observation` - Time to get observations from robot
+- `robot.send_action` - Time to send actions to robot
+- And more...
+
+### 2. Profile Without Robot (Comparison Script)
+
+Use `profile_rtc_comparison.py` to profile just the policy inference without needing a robot:
+
+```bash
+uv run examples/rtc/profile_rtc_comparison.py \
+    --policy_path=helper2424/pi05_check_rtc \
+    --device=mps \
+    --num_iterations=50 \
+    --execution_horizon=20
+```
+
+**Output**: Side-by-side comparison of performance with and without RTC, including:
+- Mean/min/max inference times
+- Throughput (iterations per second)
+- Verdict on whether RTC is faster or slower
+
+### 3. Enable Detailed Method-Level Profiling
+
+For even more granular profiling, add the `--enable_detailed_profiling` flag:
+
+```bash
+uv run examples/rtc/profile_rtc_comparison.py \
+    --policy_path=helper2424/pi05_check_rtc \
+    --device=mps \
+    --num_iterations=50 \
+    --execution_horizon=20 \
+    --enable_detailed_profiling
+```
+
+This will show timing for individual methods within the policy.
+
+## Understanding the Output
+
+### Key Metrics to Look At
+
+1. **get_actions.policy_inference** - This should be the largest component
+   - If RTC is enabled, this includes the RTC guidance overhead
+   - Compare this with/without RTC to see the overhead
+
+2. **get_actions.preprocessing** - Image preprocessing and normalization
+   - Should be relatively fast
+   - If slow, consider optimizing image processing
+
+3. **get_actions.postprocessing** - Action denormalization
+   - Should be minimal
+   - If slow, check postprocessor implementation
+
+4. **get_actions.action_queue_merge** - RTC-specific merging logic
+   - Only present when RTC is enabled
+   - If this is taking significant time, the RTC algorithm may need optimization
+
+5. **robot.get_observation** - Robot communication overhead
+   - If slow, check camera/sensor latency
+   - Consider reducing image resolution
+
+6. **robot.send_action** - Action execution overhead
+   - Should be very fast
+   - If slow, check robot communication
+
+### Expected Performance
+
+For a typical Pi0 policy on Apple Silicon (MPS):
+- **Without RTC**: ~100-200ms per inference
+- **With RTC**: Should be similar or slightly faster due to action reuse
+- **Preprocessing**: ~5-20ms depending on number of cameras
+- **Postprocessing**: ~1-5ms
+
+If RTC is significantly slower, likely causes:
+1. **RTC overhead exceeds benefits** - The guidance computation is expensive
+2. **Execution horizon too small** - Not reusing enough actions to amortize overhead
+3. **No compilation** - Try with `--use_torch_compile`
+4. **Large prev_actions buffer** - Copying/processing previous actions is slow
+
+## Profiling Your Own Code
+
+### Using the Profiling Decorator
+
+Add profiling to your own methods:
+
+```python
+from lerobot.utils.profiling import profile_method, enable_profiling, print_profiling_summary
+
+# Enable profiling
+enable_profiling()
+
+# Decorate methods you want to profile
+@profile_method
+def my_slow_function(x):
+    # ... your code ...
+    return result
+
+# At end of execution
+print_profiling_summary()
+```
+
+### Using Profile Context Manager
+
+For profiling specific code blocks:
+
+```python
+from lerobot.utils.profiling import profile_section, enable_profiling
+
+enable_profiling()
+
+with profile_section("data_loading"):
+    data = load_data()
+
+with profile_section("model_inference"):
+    output = model(data)
+```
+
+### Adding Profiling to Policy Methods
+
+To profile specific parts of the Pi0 policy, you can add decorators:
+
+```python
+# In src/lerobot/policies/pi0/modeling_pi0.py
+from lerobot.utils.profiling import profile_method, profile_section
+
+class Pi0Policy:
+    @profile_method
+    def predict_action_chunk(self, obs, inference_delay=0, prev_chunk_left_over=None):
+        # ... existing code ...
+        pass
+
+    def _generate_actions_with_rtc(self, ...):
+        with profile_section("rtc.guidance_computation"):
+            # ... guidance code ...
+            pass
+        
+        with profile_section("rtc.action_merging"):
+            # ... merging code ...
+            pass
+```
+
+## Analyzing Results
+
+### Comparison Checklist
+
+When comparing RTC vs non-RTC performance, check:
+
+- [ ] Is `policy_inference` time higher with RTC?
+- [ ] Is `action_queue_merge` taking significant time?
+- [ ] Are you running enough iterations to amortize warmup?
+- [ ] Is torch.compile enabled for fair comparison?
+- [ ] Is the execution horizon large enough? (should be >= 10-20)
+- [ ] Are you testing on the same hardware/device?
+
+### Common Bottlenecks
+
+1. **Image preprocessing dominates** 
+   - Solution: Reduce image resolution, use fewer cameras, or optimize preprocessing
+
+2. **Action queue operations are slow**
+   - Solution: Review queue implementation, consider using ring buffer
+
+3. **RTC guidance is expensive**
+   - Solution: Reduce guidance weight, simplify guidance computation, use torch.compile
+
+4. **Robot communication is slow**
+   - Solution: Increase baud rate, reduce action frequency, optimize protocol
+
+5. **Memory allocation overhead**
+   - Solution: Pre-allocate buffers, reuse tensors, avoid unnecessary copies
+
+## Advanced: Adding Custom Metrics
+
+You can add custom timing metrics to the profiled script:
+
+```python
+from lerobot.utils.profiling import record_timing
+
+start = time.perf_counter()
+# ... your code ...
+duration = time.perf_counter() - start
+record_timing("my_custom_metric", duration)
+```
+
+## Troubleshooting
+
+### Profiling shows RTC is slower by >50%
+
+1. Check if torch.compile is enabled: `--use_torch_compile`
+2. Increase execution horizon: `--rtc.execution_horizon=30`
+3. Verify inference_delay is calculated correctly
+4. Profile with `--enable_detailed_profiling` to find exact bottleneck
+
+### Profiling output is empty
+
+1. Make sure profiling is enabled with `enable_profiling()`
+2. Verify you're running enough iterations (at least 10)
+3. Check that code is actually executing (not short-circuited)
+
+### Inconsistent results between runs
+
+1. Run more iterations: `--num_iterations=100`
+2. Increase warmup iterations
+3. Check for thermal throttling on device
+4. Ensure no other processes competing for resources
+
+## Next Steps
+
+1. Run both profiling scripts (with/without robot)
+2. Compare timing breakdowns
+3. Identify the largest bottleneck
+4. Focus optimization efforts on that component
+5. Re-run profiling to verify improvements
+
+## Questions?
+
+If profiling reveals unexpected bottlenecks or you need help interpreting results, please share:
+- The full profiling output
+- Your configuration (RTC enabled/disabled, execution horizon, etc.)
+- Hardware specs (device type, memory, etc.)
+- Policy type and size
+

examples/rtc/PROFILING_QUICK_START.md

@@ -0,0 +1,208 @@
+# RTC Profiling - Quick Start
+
+Quick reference for profiling Pi0 with RTC to identify performance bottlenecks.
+
+## 🚀 Quick Commands
+
+### 1. Profile with Real Robot
+
+```bash
+# With RTC enabled (profiled version)
+uv run examples/rtc/eval_with_real_robot_profiled.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.cameras="{ gripper: {type: opencv, index_or_path: 0}, front: {type: opencv, index_or_path: 1}}" \
+    --task="Pick up object" \
+    --duration=30
+```
+
+### 2. Compare RTC vs No-RTC (No Robot Needed)
+
+```bash
+uv run examples/rtc/profile_rtc_comparison.py \
+    --policy_path=helper2424/pi05_check_rtc \
+    --device=mps \
+    --num_iterations=50 \
+    --execution_horizon=20
+```
+
+### 3. Detailed RTC Method Profiling
+
+```bash
+uv run examples/rtc/profile_pi0_rtc_detailed.py \
+    --policy_path=helper2424/pi05_check_rtc \
+    --device=mps \
+    --num_iterations=20 \
+    --execution_horizon=20 \
+    --enable_rtc_profiling
+```
+
+## 📊 What Each Tool Does
+
+| Tool | Purpose | Needs Robot? |
+|------|---------|--------------|
+| `eval_with_real_robot_profiled.py` | Profile actual robot execution with RTC | ✅ Yes |
+| `profile_rtc_comparison.py` | Compare RTC vs no-RTC side-by-side | ❌ No |
+| `profile_pi0_rtc_detailed.py` | Deep dive into RTC internals | ❌ No |
+
+## 🔍 Key Metrics to Watch
+
+### Overall Performance
+- **iteration.policy_inference** - Total policy inference time
+- **iteration.preprocessing** - Image preprocessing time
+- **iteration.postprocessing** - Action denormalization time
+
+### RTC-Specific (with `--enable_rtc_profiling`)
+- **rtc.denoise_step.base_denoising** - Time without RTC overhead
+- **rtc.denoise_step.autograd_correction** - Gradient computation time
+- **rtc.denoise_step.guidance_computation** - Total RTC guidance overhead
+
+### Robot Communication
+- **robot.get_observation** - Time to get robot state
+- **robot.send_action** - Time to send action command
+
+## 🎯 Quick Diagnosis
+
+### RTC is slower than expected?
+
+1. **Check if torch.compile is enabled**
+   ```bash
+   # Add this flag
+   --use_torch_compile
+   ```
+
+2. **Try larger execution horizon**
+   ```bash
+   # Increase to amortize RTC overhead
+   --rtc.execution_horizon=30
+   ```
+
+3. **Profile to find bottleneck**
+   ```bash
+   uv run examples/rtc/profile_pi0_rtc_detailed.py \
+       --policy_path=helper2424/pi05_check_rtc \
+       --device=mps \
+       --enable_rtc_profiling
+   ```
+
+### Preprocessing is slow?
+
+- Reduce image resolution in robot config
+- Use fewer cameras
+- Check camera FPS settings
+
+### Policy inference is slow?
+
+- Enable torch.compile
+- Check device (MPS vs CUDA vs CPU)
+- Try smaller model if available
+
+## 📈 Expected Performance
+
+### Typical timings on Apple Silicon (MPS):
+
+| Component | Time (ms) | Notes |
+|-----------|-----------|-------|
+| Policy inference | 100-200 | Depends on model size |
+| Preprocessing | 5-20 | Depends on #cameras |
+| Postprocessing | 1-5 | Usually fast |
+| RTC overhead | 10-50 | Should be < 50% of base |
+
+### When RTC helps:
+- ✅ Execution horizon ≥ 10
+- ✅ Inference time > action execution rate
+- ✅ Using torch.compile
+- ✅ Proper inference_delay calculation
+
+### When RTC might not help:
+- ❌ Very fast inference already
+- ❌ Small execution horizon (< 5)
+- ❌ No compilation (interpreted mode)
+- ❌ Inference delay not accounted for
+
+## 🛠️ Adding Profiling to Your Code
+
+### Quick snippet:
+
+```python
+from lerobot.utils.profiling import enable_profiling, print_profiling_summary, profile_section
+
+# Enable at start
+enable_profiling()
+
+# Profile sections
+with profile_section("my_operation"):
+    # ... your code ...
+    pass
+
+# Print at end
+print_profiling_summary()
+```
+
+### Profile specific methods:
+
+```python
+from lerobot.utils.profiling import profile_method
+
+@profile_method
+def my_slow_function():
+    # ... your code ...
+    pass
+```
+
+## 📝 Example Output
+
+```
+PROFILING SUMMARY
+================================================================================
+Function                                                    Count    Mean (ms)
+--------------------------------------------------------------------------------
+iteration.policy_inference                                    20       150.23
+iteration.preprocessing                                       20        12.45
+rtc.denoise_step.guidance_computation                        200        15.67
+rtc.denoise_step.autograd_correction                         200         8.23
+rtc.denoise_step.base_denoising                             200       120.45
+================================================================================
+```
+
+## 🚨 Common Issues
+
+### "No profiling data available"
+- Did you call `enable_profiling()`?
+- Running enough iterations?
+
+### Inconsistent results
+- Increase `--num_iterations`
+- Check for thermal throttling
+- Close other applications
+
+### Can't find bottleneck
+- Enable `--enable_rtc_profiling` for detailed breakdown
+- Check both preprocessing and inference
+- Compare with and without RTC
+
+## 📖 More Details
+
+See `PROFILING_GUIDE.md` for comprehensive documentation.
+
+## 🤔 Still Slow?
+
+1. Run comparison: `profile_rtc_comparison.py`
+2. Run detailed profiling: `profile_pi0_rtc_detailed.py --enable_rtc_profiling`
+3. Share output for help (include device, model, settings)
+
+## ✅ Quick Checklist
+
+Before asking for help, verify:
+
+- [ ] Ran comparison script (with/without RTC)
+- [ ] Tried torch.compile
+- [ ] Tested different execution horizons (10, 20, 30)
+- [ ] Profiled with detailed RTC profiling
+- [ ] Checked preprocessing vs inference split
+- [ ] Verified hardware (device type, thermal state)
+

examples/rtc/PROFILING_README.md

@@ -0,0 +1,352 @@
+# RTC Profiling Toolkit
+
+Complete toolkit for profiling Pi0 with RTC to identify performance bottlenecks.
+
+## 📦 What's Included
+
+### Scripts
+
+1. **`eval_with_real_robot_profiled.py`**
+   - Profiled version of the real robot eval script
+   - Adds timing measurements throughout execution
+   - Works with actual robot hardware
+   - Same usage as original but with profiling output
+
+2. **`profile_rtc_comparison.py`**
+   - Side-by-side comparison of RTC vs no-RTC
+   - No robot needed (uses mock observations)
+   - Shows clear verdict on whether RTC is helping
+   - Great for quick performance checks
+
+3. **`profile_pi0_rtc_detailed.py`**
+   - Most detailed profiling available
+   - Can enable RTC method-level profiling
+   - Provides insights and recommendations
+   - Perfect for deep-dive investigations
+
+4. **`add_rtc_profiling.py`**
+   - Monkey-patching utility for RTC internals
+   - Profiles individual RTC operations
+   - Can be applied without modifying source
+   - Shows exactly where RTC spends time
+
+### Utilities
+
+5. **`src/lerobot/utils/profiling.py`**
+   - Core profiling utilities
+   - Decorators for method profiling
+   - Context managers for code blocks
+   - Statistics collection and reporting
+
+### Documentation
+
+6. **`PROFILING_GUIDE.md`** - Comprehensive guide
+7. **`PROFILING_QUICK_START.md`** - Quick reference
+
+## 🚀 Quick Start
+
+### Step 1: Compare Performance
+
+Run this first to see if RTC is actually slower:
+
+```bash
+uv run examples/rtc/profile_rtc_comparison.py \
+    --policy_path=helper2424/pi05_check_rtc \
+    --device=mps \
+    --num_iterations=50 \
+    --execution_horizon=20
+```
+
+**Expected output:**
+```
+COMPARISON SUMMARY
+================================================================================
+Metric                         Without RTC        With RTC      Difference
+--------------------------------------------------------------------------------
+Mean time (ms)                       150.23         165.45          +15.22
+Throughput (iter/s)                    6.66           6.05           -0.61
+================================================================================
+VERDICT
+✗ RTC is SLOWER by 10.1%
+  Mean time increased by 15.22 ms
+  
+  Possible reasons:
+  - RTC overhead exceeds benefits at current execution horizon
+  - No torch.compile enabled
+```
+
+### Step 2: Identify Bottleneck
+
+If RTC is slower, find out why:
+
+```bash
+uv run examples/rtc/profile_pi0_rtc_detailed.py \
+    --policy_path=helper2424/pi05_check_rtc \
+    --device=mps \
+    --num_iterations=20 \
+    --execution_horizon=20 \
+    --enable_rtc_profiling
+```
+
+**Expected output:**
+```
+PROFILING SUMMARY
+================================================================================
+Function                                             Count    Mean (ms)    Total (s)
+------------------------------------------------------------------------------------
+iteration.policy_inference                              20      150.23         3.00
+rtc.denoise_step.guidance_computation                  200       15.67         3.13
+rtc.denoise_step.autograd_correction                   200        8.23         1.65
+iteration.preprocessing                                 20       12.45         0.25
+================================================================================
+
+KEY INSIGHTS
+================================================================================
+Time breakdown:
+  Policy inference:  150.23 ms (87.2%)
+  Preprocessing:     12.45 ms (7.2%)
+  Postprocessing:    2.10 ms (1.2%)
+
+RTC breakdown:
+  Base denoising:    120.45 ms
+  Guidance compute:  15.67 ms
+  Autograd correct:  8.23 ms
+  RTC overhead:      23.90 ms (19.8% of base)
+
+Recommendations:
+  ⚠ RTC autograd overhead is significant
+    → This is expected, but consider increasing execution_horizon
+    → Try torch.compile if not already enabled
+  💡 torch.compile not enabled
+    → Try --use_torch_compile for potential speedup
+================================================================================
+```
+
+### Step 3: Try Optimizations
+
+Based on recommendations:
+
+```bash
+# Try with torch.compile
+uv run examples/rtc/profile_rtc_comparison.py \
+    --policy_path=helper2424/pi05_check_rtc \
+    --device=mps \
+    --num_iterations=50 \
+    --execution_horizon=20 \
+    --use_torch_compile
+
+# Try larger execution horizon
+uv run examples/rtc/profile_rtc_comparison.py \
+    --policy_path=helper2424/pi05_check_rtc \
+    --device=mps \
+    --num_iterations=50 \
+    --execution_horizon=30
+```
+
+### Step 4: Profile Real Robot (Optional)
+
+Test with actual hardware:
+
+```bash
+uv run examples/rtc/eval_with_real_robot_profiled.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.cameras="{...}" \
+    --task="Pick up object" \
+    --duration=30
+```
+
+## 🎯 Common Scenarios
+
+### "RTC is 2x slower!"
+
+This usually means:
+- RTC overhead is high but not getting benefits
+- Need to enable torch.compile
+- Execution horizon too small
+- Inference delay not calculated correctly
+
+**Try:**
+1. `--use_torch_compile`
+2. Increase `--execution_horizon` to 30+
+3. Check inference_delay calculation
+
+### "RTC is only slightly slower"
+
+This is expected! RTC overhead is about 10-30% typically.
+The benefit comes during **execution**, not single inference:
+- Actions are reused across chunks
+- Overall system latency is reduced
+- Robot gets smoother actions
+
+### "Want to optimize specific part"
+
+Use the profiling utilities:
+
+```python
+from lerobot.utils.profiling import enable_profiling, profile_section, print_profiling_summary
+
+enable_profiling()
+
+with profile_section("my_custom_operation"):
+    # Your code here
+    pass
+
+print_profiling_summary()
+```
+
+## 📊 Understanding Results
+
+### Key Metrics
+
+**Policy Inference Time**
+- Time for forward pass through model
+- Should be largest component (70-90%)
+- Includes RTC guidance if enabled
+
+**Preprocessing Time**
+- Image normalization, resizing
+- Should be < 20% of total
+- If high: reduce image resolution
+
+**RTC Guidance Overhead**
+- Extra time for RTC guidance computation
+- Typically 10-30% of base inference
+- If > 50%: RTC may not be beneficial at current settings
+
+**Autograd Correction**
+- Time computing gradients for RTC
+- Usually 5-15% of base inference
+- Can be reduced with torch.compile
+
+### Expected Ranges (Apple Silicon MPS)
+
+| Metric | Good | Acceptable | Poor |
+|--------|------|------------|------|
+| Policy inference | 100-150ms | 150-250ms | >250ms |
+| Preprocessing | <20ms | 20-50ms | >50ms |
+| RTC overhead | 10-30% | 30-50% | >50% |
+
+## 🔧 Optimization Guide
+
+### If RTC overhead is too high:
+
+1. **Enable compilation:**
+   ```bash
+   --use_torch_compile
+   ```
+   Expected improvement: 20-40% faster
+
+2. **Increase execution horizon:**
+   ```bash
+   --execution_horizon=30  # or higher
+   ```
+   Amortizes RTC cost over more actions
+
+3. **Check guidance weight:**
+   ```python
+   # In config
+   rtc.max_guidance_weight=1.0  # try 0.5 for less overhead
+   ```
+
+### If preprocessing is slow:
+
+1. **Reduce image resolution:**
+   ```python
+   # In robot config
+   cameras={
+       "gripper": {"width": 320, "height": 240}  # instead of 640x480
+   }
+   ```
+
+2. **Use fewer cameras:**
+   - Profile which cameras are essential
+   - Remove unnecessary views
+
+### If inference is generally slow:
+
+1. Use torch.compile (if not already)
+2. Check device is correct (MPS vs CUDA)
+3. Verify model is in eval mode
+4. Check for unnecessary gradient tracking
+
+## 🐛 Troubleshooting
+
+### Empty profiling output
+```python
+# Make sure to enable profiling!
+from lerobot.utils.profiling import enable_profiling
+enable_profiling()
+```
+
+### Inconsistent timings
+- Run more iterations (50-100)
+- Check thermal throttling
+- Close background apps
+- Use `--warmup_iterations=10`
+
+### Can't find bottleneck
+1. Start with `profile_rtc_comparison.py`
+2. Then run `profile_pi0_rtc_detailed.py --enable_rtc_profiling`
+3. Compare with/without RTC
+4. Check each component separately
+
+## 📖 Full Documentation
+
+- **`PROFILING_GUIDE.md`** - Complete reference with examples
+- **`PROFILING_QUICK_START.md`** - Quick commands and tips
+
+## 🤝 Getting Help
+
+If you're still experiencing issues:
+
+1. Run comparison script and save output
+2. Run detailed profiling and save output
+3. Include:
+   - Policy path
+   - Device type
+   - RTC settings (execution_horizon, etc.)
+   - Hardware specs
+   - Full profiling output
+
+## 🎓 Learning More
+
+### Profiling your own code:
+
+```python
+from lerobot.utils.profiling import profile_method, enable_profiling
+
+enable_profiling()
+
+@profile_method
+def my_function():
+    # Automatically profiled
+    pass
+```
+
+### RTC internals:
+
+```python
+from examples.rtc.add_rtc_profiling import monkey_patch_rtc_profiling
+
+enable_profiling()
+monkey_patch_rtc_profiling()
+
+# Now RTC methods are profiled
+policy.predict_action_chunk(...)
+```
+
+## ✨ Next Steps
+
+1. Run `profile_rtc_comparison.py` to establish baseline
+2. Use `profile_pi0_rtc_detailed.py` to find bottlenecks
+3. Apply optimizations (torch.compile, larger horizon)
+4. Re-run comparison to verify improvements
+5. Test with real robot using profiled version
+
+Happy profiling! 🚀
+

examples/rtc/README.md

@@ -0,0 +1,251 @@
+# Real-Time Chunking (RTC) Examples
+
+This directory contains examples and evaluation scripts for Real-Time Chunking (RTC), a technique for improving action chunking policies in real-time robot control.
+
+## Overview
+
+Real-Time Chunking addresses the challenge of maintaining consistency and reactivity when using action chunking policies with non-negligible inference latency. It uses a guidance technique during diffusion sampling to blend new action predictions with previously planned actions.
+
+**Key Benefits:**
+
+- Maintains consistency between consecutive action chunks
+- Reduces jitter and improves smoothness
+- Adapts to inference delays dynamically
+
+**Reference:** [Physical Intelligence - Real-Time Chunking](https://www.physicalintelligence.company/download/real_time_chunking.pdf)
+
+## Scripts
+
+### 1. `eval_dataset.py`
+
+Offline evaluation on dataset samples with detailed visualization and validation.
+
+**Features:**
+
+- Compare RTC vs non-RTC predictions on two random dataset samples
+- Validate RTC behavior (delay region, blend region, post-horizon region)
+- Generate debug visualizations:
+  - Denoising step comparisons (x_t, v_t, x1_t, corrections)
+  - Final action predictions comparison
+- Support for torch.compile() optimization
+- Memory-efficient sequential policy loading for large models
+
+**Usage:**
+
+```bash
+# 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 \
+    --seed=10
+
+# With Pi0.5 policy on CUDA
+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
+
+# With Pi0 policy
+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
+
+# With torch.compile for faster inference
+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=max-autotune
+
+# 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
+```
+
+**Key Parameters:**
+
+- `--policy.path`: Path to pretrained policy
+- `--dataset.repo_id`: Dataset to evaluate on
+- `--rtc.execution_horizon`: Number of steps to maintain consistency (default: 20)
+- `--rtc.max_guidance_weight`: Maximum guidance weight (default: 10.0)
+- `--rtc.prefix_attention_schedule`: Schedule type (ZEROS, ONES, LINEAR, EXP)
+- `--inference_delay`: Inference delay for RTC (default: 4)
+- `--seed`: Random seed for reproducibility (default: 42)
+- `--output_dir`: Directory to save visualizations (default: rtc_debug_output)
+- `--device`: Device to use (cuda, cpu, mps, auto)
+- `--use_torch_compile`: Enable torch.compile() for faster inference
+
+**Output:**
+
+The script generates several visualization files in `rtc_debug_output/`:
+
+- `denoising_xt_comparison.png` - Noisy state evolution during denoising
+- `denoising_vt_comparison.png` - Velocity predictions during denoising
+- `denoising_x1t_comparison.png` - Predicted final states during denoising
+- `denoising_correction_comparison.png` - RTC guidance corrections applied
+- `final_actions_comparison.png` - Final action predictions (prev_chunk, no_rtc, rtc)
+
+The script also validates RTC behavior and reports:
+
+- ✅ Delay region [0:inference_delay]: RTC = prev_chunk
+- ✅ Blend region [inference_delay:execution_horizon]: prev_chunk ≤ RTC ≤ no_rtc
+- ✅ Post-horizon [execution_horizon:]: RTC = no_rtc
+
+### 2. `eval_with_real_robot.py`
+
+Real-time evaluation on physical robots or simulation environments.
+
+**Features:**
+
+- Run policy with RTC on real robot or simulation
+- Multi-threaded action execution and inference
+- Action queue management with proper timing
+- Latency tracking and adaptive inference delay
+- Support for both robots and gym environments
+- Support for torch.compile() optimization
+
+**Usage:**
+
+```bash
+# With real robot
+uv run python examples/rtc/eval_with_real_robot.py \
+    --policy.path=lerobot/smolvla_base \
+    --robot.type=so100 \
+    --task="pick up the cup" \
+    --duration=30.0
+
+# With simulation environment
+uv run python examples/rtc/eval_with_real_robot.py \
+    --policy.path=lerobot/smolvla_base \
+    --env.type=pusht \
+    --duration=60.0
+
+# With policy compilation (CUDA only, not MPS)
+uv run python examples/rtc/eval_with_real_robot.py \
+    --policy.path=lerobot/smolvla_base \
+    --robot.type=so100 \
+    --use_torch_compile=true \
+    --torch_compile_mode=max-autotune
+```
+
+**Key Parameters:**
+
+- `--policy.path`: Path to pretrained policy
+- `--robot.type` or `--env.type`: Robot or environment to use
+- `--task`: Task description (for VLA models)
+- `--rtc.execution_horizon`: Number of steps to maintain consistency (default: 10)
+- `--rtc.max_guidance_weight`: Maximum guidance weight (default: 1.0)
+- `--rtc.prefix_attention_schedule`: Schedule type (ZEROS, ONES, LINEAR, EXP)
+- `--duration`: How long to run (seconds, default: 30.0)
+- `--fps`: Action execution frequency (Hz, default: 10.0)
+- `--action_queue_size_to_get_new_actions`: Queue size threshold to request new actions (default: 30)
+- `--device`: Device to use (cuda, cpu, mps, auto)
+- `--use_torch_compile`: Enable torch.compile() for faster inference
+
+## Understanding RTC Parameters
+
+### `execution_horizon`
+
+Number of timesteps from previous chunk to maintain consistency with. Higher values mean more consistency but potentially less reactivity.
+
+**Typical values:** 8-12 steps for dataset evaluation, 10 steps for real-time execution
+
+### `max_guidance_weight`
+
+Upper bound on guidance strength. Higher values give stronger consistency but may over-constrain new predictions.
+
+**Typical values:**
+
+- Dataset evaluation: 10.0-100.0 (can be higher for analysis)
+- Real-time execution: 1.0-10.0 (more conservative)
+
+### `prefix_attention_schedule`
+
+How to weight consistency across the overlap region:
+
+- `ZEROS`: Binary (full weight up to inference_delay, then zero)
+- `ONES`: Full weight across entire execution_horizon
+- `LINEAR`: Linear decay from inference_delay to execution_horizon
+- `EXP`: Exponential decay (recommended)
+
+**Recommended:** `EXP`
+
+### `inference_delay`
+
+Number of timesteps from the prefix to use for guidance. Typically calculated dynamically based on inference latency in real-time execution, but fixed for dataset evaluation.
+
+**Typical values:** 3-5 steps for dataset evaluation
+
+### `action_queue_size_to_get_new_actions` (real-time only)
+
+Threshold for requesting new action chunks. Should be higher than `inference_delay + execution_horizon` to ensure smooth operation.
+
+**Typical values:** 20-30 steps
+
+## Validation Rules (Dataset Evaluation)
+
+The dataset evaluation script validates that RTC behavior matches expectations:
+
+1. **Delay Region [0:inference_delay]**: RTC actions should equal previous chunk
+   - Ensures consistency during the inference delay period
+
+2. **Blend Region [inference_delay:execution_horizon]**: RTC should be between prev_chunk and no_rtc
+   - Smooth transition from previous plan to new predictions
+
+3. **Post-Horizon [execution_horizon:]**: RTC should equal no_rtc
+   - Full adoption of new predictions after execution horizon
+
+## Tips
+
+1. **Start with dataset evaluation** (`eval_dataset.py`) to understand RTC behavior and tune parameters before running on robot
+2. **Use visualizations** to debug unexpected behavior - check denoising steps and final actions
+3. **Tune execution_horizon** based on your inference latency and action frequency
+4. **Monitor validation output** - failures indicate potential implementation issues or misconfigured parameters
+5. **Compare different schedules** - EXP usually works best but LINEAR can be more interpretable
+
+## Troubleshooting
+
+### Validation fails in delay region
+
+- Check that `prev_chunk_left_over` is properly passed to the policy
+- Verify RTC guidance is being applied during denoising
+- Look at denoising visualizations to see where guidance diverges
+
+### Validation fails in post-horizon region
+
+- RTC and no_rtc use different noise - verify same noise is being used for comparison
+- Check that weights are correctly zeroed out after execution horizon
+- Review prefix_attention_schedule visualization
+
+### Poor performance on real robot
+
+- Increase `action_queue_size_to_get_new_actions` if you see warnings
+- Reduce `max_guidance_weight` if robot is too conservative
+- Try different `prefix_attention_schedule` values
+- Enable torch.compile() for faster inference (CUDA only)
+
+### Memory issues with large models
+
+- The dataset evaluation script loads policies sequentially to minimize memory
+- For real-time execution, only one policy is loaded
+- Use smaller batch sizes if needed
+
+## Related Documentation
+
+- [RTC Implementation](../../src/lerobot/policies/rtc/modeling_rtc.py)
+- [RTC Configuration](../../src/lerobot/policies/rtc/configuration_rtc.py)
+- [Action Queue](../../src/lerobot/policies/rtc/action_queue.py)
+- [Physical Intelligence Paper](https://www.physicalintelligence.company/download/real_time_chunking.pdf)

examples/rtc/add_rtc_profiling.py

@@ -0,0 +1,202 @@
+#!/usr/bin/env python
+
+"""
+Script to add profiling instrumentation to RTCProcessor.
+
+This script shows which methods to profile in the RTC code to identify bottlenecks.
+You can either:
+1. Apply these changes directly to modeling_rtc.py
+2. Use monkey patching to add profiling without modifying source
+3. Use as reference for manual instrumentation
+
+Usage:
+    # Option 1: Monkey patch (no source changes)
+    python examples/rtc/add_rtc_profiling.py
+
+    # Option 2: Apply changes to source
+    # Copy the profiled methods below into src/lerobot/policies/rtc/modeling_rtc.py
+"""
+
+import logging
+
+import torch
+from torch import Tensor
+
+from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+from lerobot.utils.profiling import ProfileContext, enable_profiling, is_profiling_enabled
+
+logger = logging.getLogger(__name__)
+
+
+def profile_denoise_step(self, x_t, prev_chunk_left_over, inference_delay, time, original_denoise_step_partial, execution_horizon=None) -> Tensor:
+    """Profiled version of denoise_step."""
+    
+    if not is_profiling_enabled():
+        # Call original implementation if profiling disabled
+        return self._original_denoise_step(x_t, prev_chunk_left_over, inference_delay, time, original_denoise_step_partial, execution_horizon)
+    
+    with ProfileContext("rtc.denoise_step.total"):
+        # 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
+            with ProfileContext("rtc.denoise_step.base_denoising"):
+                v_t = original_denoise_step_partial(x_t)
+            return v_t
+
+        with ProfileContext("rtc.denoise_step.setup"):
+            x_t = x_t.clone().detach()
+
+            squeezed = False
+            if len(x_t.shape) < 3:
+                x_t = x_t.unsqueeze(0)
+                squeezed = True
+
+            if len(prev_chunk_left_over.shape) < 3:
+                prev_chunk_left_over = prev_chunk_left_over.unsqueeze(0)
+
+            if execution_horizon is None:
+                execution_horizon = self.rtc_config.execution_horizon
+
+            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]
+
+        # Padding
+        with ProfileContext("rtc.denoise_step.padding"):
+            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
+
+        # Get prefix weights
+        with ProfileContext("rtc.denoise_step.get_prefix_weights"):
+            weights = (
+                self.get_prefix_weights(inference_delay, execution_horizon, action_chunk_size)
+                .to(x_t.device)
+                .unsqueeze(0)
+                .unsqueeze(-1)
+            )
+
+        # Main RTC guidance computation
+        with ProfileContext("rtc.denoise_step.guidance_computation"):
+            with torch.enable_grad():
+                # Base denoising
+                with ProfileContext("rtc.denoise_step.base_denoising"):
+                    v_t = original_denoise_step_partial(x_t)
+                
+                x_t.requires_grad_(True)
+
+                # Compute x1_t
+                with ProfileContext("rtc.denoise_step.compute_x1_t"):
+                    x1_t = x_t - time * v_t
+
+                # Compute error
+                with ProfileContext("rtc.denoise_step.compute_error"):
+                    err = (prev_chunk_left_over - x1_t) * weights
+                    grad_outputs = err.clone().detach()
+
+                # Compute correction via autograd
+                with ProfileContext("rtc.denoise_step.autograd_correction"):
+                    correction = torch.autograd.grad(x1_t, x_t, grad_outputs, retain_graph=False)[0]
+
+        # Compute guidance weight
+        with ProfileContext("rtc.denoise_step.compute_guidance_weight"):
+            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)
+
+        # Apply guidance
+        with ProfileContext("rtc.denoise_step.apply_guidance"):
+            result = v_t - guidance_weight * correction
+
+        # Cleanup
+        with ProfileContext("rtc.denoise_step.cleanup"):
+            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 monkey_patch_rtc_profiling():
+    """Apply profiling to RTCProcessor via monkey patching.
+    
+    This modifies the RTCProcessor class at runtime to add profiling
+    without changing source files.
+    """
+    logger.info("Applying RTC profiling monkey patch...")
+    
+    # Save original method
+    RTCProcessor._original_denoise_step = RTCProcessor.denoise_step
+    
+    # Replace with profiled version
+    RTCProcessor.denoise_step = profile_denoise_step
+    
+    logger.info("✓ RTC profiling enabled")
+
+
+def print_usage():
+    """Print usage instructions."""
+    print("\n" + "="*80)
+    print("RTC PROFILING INSTRUMENTATION")
+    print("="*80)
+    print("\nThis script provides profiling for RTCProcessor methods.")
+    print("\nOption 1: Monkey Patch (Recommended)")
+    print("-" * 40)
+    print("Add to your script:")
+    print("""
+    from lerobot.utils.profiling import enable_profiling, print_profiling_summary
+    from examples.rtc.add_rtc_profiling import monkey_patch_rtc_profiling
+
+    # Enable profiling
+    enable_profiling()
+    monkey_patch_rtc_profiling()
+
+    # ... run your code ...
+
+    # Print results
+    print_profiling_summary()
+    """)
+    
+    print("\nOption 2: Manual Source Modification")
+    print("-" * 40)
+    print("1. Copy profile_denoise_step() from this file")
+    print("2. Replace denoise_step() in src/lerobot/policies/rtc/modeling_rtc.py")
+    print("3. Add profiling imports at top of file")
+    
+    print("\nKey Metrics to Watch:")
+    print("-" * 40)
+    print("- rtc.denoise_step.base_denoising     - Time for base policy inference")
+    print("- rtc.denoise_step.autograd_correction - Time computing gradients")
+    print("- rtc.denoise_step.guidance_computation - Total guidance overhead")
+    print("- rtc.denoise_step.get_prefix_weights  - Time computing weights")
+    print("="*80 + "\n")
+
+
+if __name__ == "__main__":
+    print_usage()
+

examples/rtc/eval_dataset.py

@@ -39,9 +39,8 @@ Usage:
     uv run python examples/rtc/eval_dataset.py \
         --policy.path=lerobot/pi05_libero_finetuned \
         --dataset.repo_id=HuggingFaceVLA/libero \
-        --rtc.execution_horizon=10 \
+        --rtc.execution_horizon=8 \
         --device=mps
-        --seed=10
 
     # Basic usage with pi0.5 policy with cuda device
     uv run python examples/rtc/eval_dataset.py \
@@ -142,7 +141,7 @@ def _check_matplotlib_available():
         raise ImportError(
             "matplotlib is required for RTC debug visualizations. "
             "Please install it by running:\n"
-            "  uv pip install matplotlib"
+            "  uv pip install -e '.[matplotlib-dep]'"
         )
 
 
@@ -544,6 +543,11 @@ class RTCEvaluator:
         logging.info("Plotting results...")
         self.plot_tracked_data(rtc_tracked_steps, no_rtc_tracked_steps, prev_chunk_left_over, num_steps)
 
+        # Validate RTC behavior
+        # logging.info("=" * 80)
+        # logging.info("Validating RTC behavior...")
+        # self.validate_rtc_behavior(rtc_actions, no_rtc_actions, prev_chunk_left_over)
+
         # Plot final actions comparison
         logging.info("=" * 80)
         logging.info("Plotting final actions comparison...")
@@ -552,6 +556,159 @@ class RTCEvaluator:
         logging.info("=" * 80)
         logging.info("Evaluation completed successfully")
 
+    def validate_rtc_behavior(self, rtc_actions, no_rtc_actions, prev_chunk_left_over):
+        """Validate RTC behavior by comparing final action predictions with expected values.
+
+        Validation rules:
+        1. During delay [0:inference_delay]: RTC should equal prev_chunk
+        2. After delay, within execution horizon [inference_delay:execution_horizon]:
+           RTC should be between prev_chunk and no_rtc
+        3. After execution horizon [execution_horizon:]: RTC should equal no_rtc
+
+        Args:
+            rtc_actions: Final actions from RTC policy (batch, time, action_dim)
+            no_rtc_actions: Final actions from non-RTC policy (batch, time, action_dim)
+            prev_chunk_left_over: Previous chunk used as ground truth (time, action_dim)
+        """
+        # Remove batch dimension if present and move to CPU
+        rtc_actions_t = rtc_actions.squeeze(0).cpu() if len(rtc_actions.shape) == 3 else rtc_actions.cpu()
+        no_rtc_actions_t = (
+            no_rtc_actions.squeeze(0).cpu() if len(no_rtc_actions.shape) == 3 else no_rtc_actions.cpu()
+        )
+        prev_chunk = prev_chunk_left_over.cpu()
+
+        logging.info(f"  rtc_actions shape: {rtc_actions_t.shape}")
+        logging.info(f"  no_rtc_actions shape: {no_rtc_actions_t.shape}")
+        logging.info(f"  prev_chunk shape: {prev_chunk.shape}")
+
+        # Determine chunk length for comparison
+        chunk_len = min(rtc_actions_t.shape[0], no_rtc_actions_t.shape[0], prev_chunk.shape[0])
+        inference_delay = self.cfg.inference_delay
+        execution_horizon = self.cfg.rtc.execution_horizon
+
+        # Tolerance for floating point comparison
+        rtol = 1e-2  # Relative tolerance
+
+        validation_passed = True
+        warnings = []
+
+        logging.info("  Validating RTC behavior:")
+        logging.info(f"    Chunk length: {chunk_len}")
+        logging.info(f"    Inference delay: {inference_delay}")
+        logging.info(f"    Execution horizon: {execution_horizon}")
+        logging.info(f"    Tolerance: rtol={rtol}")
+
+        # ============================================================================
+        # Rule 1: During delay [0:inference_delay], RTC should equal prev_chunk
+        # ============================================================================
+        if inference_delay > 0:
+            delay_end = min(inference_delay, chunk_len)
+            rtc_delay = rtc_actions_t[:delay_end]
+            prev_delay = prev_chunk[:delay_end]
+
+            logging.info(f"  rtc_delay: {rtc_delay.shape}")
+            logging.info(f"  prev_delay: {prev_delay.shape}")
+
+            if not torch.allclose(rtc_delay, prev_delay, rtol=rtol):
+                max_diff = torch.max(torch.abs(rtc_delay - prev_delay)).item()
+                mean_diff = torch.mean(torch.abs(rtc_delay - prev_delay)).item()
+                logging.info(f"  rtc_delay: {rtc_delay}")
+                logging.info(f"  prev_delay: {prev_delay}")
+                logging.info(f"  max_diff: {max_diff}")
+                logging.info(f"  mean_diff: {mean_diff}")
+                warnings.append(
+                    f"    ⚠ VALIDATION FAILED: During delay [0:{delay_end}], "
+                    f"RTC does NOT equal prev_chunk!\n"
+                    f"      Max difference: {max_diff:.6f}\n"
+                    f"      Mean difference: {mean_diff:.6f}"
+                )
+                validation_passed = False
+            else:
+                logging.info(f"    ✓ During delay [0:{delay_end}]: RTC equals prev_chunk")
+
+        # ============================================================================
+        # Rule 2: After delay, within execution horizon [inference_delay:execution_horizon]
+        #         RTC should be between prev_chunk and no_rtc
+        # ============================================================================
+        blend_start = inference_delay
+        blend_end = min(execution_horizon, chunk_len)
+
+        if blend_end > blend_start:
+            rtc_blend = rtc_actions_t[blend_start:blend_end]
+            prev_blend = prev_chunk[blend_start:blend_end]
+            no_rtc_blend = no_rtc_actions_t[blend_start:blend_end]
+
+            # Check if RTC is between prev_chunk and no_rtc (element-wise)
+            # For each element, check if it's between the min and max of prev_chunk and no_rtc
+            min_bound = torch.minimum(prev_blend, no_rtc_blend)
+            max_bound = torch.maximum(prev_blend, no_rtc_blend)
+
+            within_bounds = torch.logical_and(rtc_blend >= min_bound, rtc_blend <= max_bound)
+
+            if not torch.all(within_bounds):
+                violations = torch.sum(~within_bounds).item()
+                total_elements = within_bounds.numel()
+                violation_pct = 100.0 * violations / total_elements
+
+                # Find max violation
+                lower_violations = torch.maximum(torch.tensor(0.0), min_bound - rtc_blend)
+                upper_violations = torch.maximum(torch.tensor(0.0), rtc_blend - max_bound)
+                max_violation = torch.max(torch.maximum(lower_violations, upper_violations)).item()
+
+                warnings.append(
+                    f"    ⚠ VALIDATION FAILED: In blend region [{blend_start}:{blend_end}], "
+                    f"RTC is NOT always between prev_chunk and no_rtc!\n"
+                    f"      Violations: {violations}/{total_elements} elements ({violation_pct:.1f}%)\n"
+                    f"      Max violation distance: {max_violation:.6f}"
+                )
+                validation_passed = False
+            else:
+                logging.info(
+                    f"    ✓ Blend region [{blend_start}:{blend_end}]: RTC is between prev_chunk and no_rtc"
+                )
+
+        # ============================================================================
+        # Rule 3: After execution horizon [execution_horizon:], RTC should equal no_rtc
+        # ============================================================================
+        if execution_horizon < chunk_len:
+            rtc_after = rtc_actions_t[execution_horizon:chunk_len]
+            no_rtc_after = no_rtc_actions_t[execution_horizon:chunk_len]
+
+            logging.info(f"  rtc_after: {rtc_after}")
+            logging.info(f"  no_rtc_after: {no_rtc_after}")
+
+            if not torch.allclose(rtc_after, no_rtc_after, rtol=rtol):
+                max_diff = torch.max(torch.abs(rtc_after - no_rtc_after)).item()
+                mean_diff = torch.mean(torch.abs(rtc_after - no_rtc_after)).item()
+                warnings.append(
+                    f"    ⚠ VALIDATION FAILED: After execution horizon [{execution_horizon}:{chunk_len}], "
+                    f"RTC does NOT equal no_rtc!\n"
+                    f"      Max difference: {max_diff:.6f}\n"
+                    f"      Mean difference: {mean_diff:.6f}"
+                )
+                validation_passed = False
+            else:
+                logging.info(
+                    f"    ✓ After execution horizon [{execution_horizon}:{chunk_len}]: RTC equals no_rtc"
+                )
+
+        # ============================================================================
+        # Report results
+        # ============================================================================
+        logging.info("=" * 80)
+        if validation_passed:
+            logging.info("  ✅ VALIDATION PASSED: All RTC behavior checks passed!")
+            logging.info("    • During delay: RTC = prev_chunk ✓")
+            logging.info("    • Blend region: prev_chunk ≤ RTC ≤ no_rtc ✓")
+            logging.info("    • After execution horizon: RTC = no_rtc ✓")
+        else:
+            logging.error("  ❌ VALIDATION FAILED: RTC behavior does not match expected!")
+            logging.error("")
+            for warning in warnings:
+                logging.error(warning)
+            logging.error("")
+            logging.error("  Please check the implementation of RTC guidance.")
+
     def plot_final_actions_comparison(self, rtc_actions, no_rtc_actions, prev_chunk_left_over):
         """Plot final action predictions comparison on a single chart.
 
@@ -638,34 +795,16 @@ class RTCEvaluator:
             ax.set_xticks(range(0, max_len, max(1, max_len // 20)))  # Show ~20 ticks
             ax.set_xlim(-0.5, max_len - 0.5)
 
+            # Add legend only to first subplot
+            if dim_idx == 0:
+                ax.legend(loc="best", fontsize=9)
+
         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")
+        fig.tight_layout()
+        fig.savefig(output_path, dpi=150)
         logging.info(f"Saved final actions comparison to {output_path}")
         plt.close(fig)
 
@@ -686,7 +825,6 @@ class RTCEvaluator:
             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(
@@ -696,7 +834,6 @@ class RTCEvaluator:
             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
@@ -712,21 +849,15 @@ class RTCEvaluator:
             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)
+        # Plot ground truth on x_t axes
         RTCDebugVisualizer.plot_waypoints(
-            axs_xt[:, 0], prev_chunk_left_over, start_from=0, color="red", label=None
+            axs_xt[:, 0], prev_chunk_left_over, start_from=0, color="red", label="Ground truth"
         )
 
         RTCDebugVisualizer.plot_waypoints(
-            axs_x1t[:, 0], prev_chunk_left_over, start_from=0, color="red", label=None
+            axs_x1t[:, 0], prev_chunk_left_over, start_from=0, color="red", label="Ground truth"
         )
 
-        # 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"))
@@ -744,47 +875,13 @@ class RTCEvaluator:
 
         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")
+        fig.tight_layout()
+        fig.savefig(path, dpi=150)
         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
-    ):
+    def _plot_denoising_steps_from_tracker(self, tracked_steps, xt_axs, vt_axs, corr_axs, x1t_axs, num_steps):
         """Plot denoising steps from tracker data.
 
         Args:
@@ -794,7 +891,6 @@ class RTCEvaluator:
             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)
@@ -809,18 +905,17 @@ class RTCEvaluator:
 
         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
+                    xt_axs, debug_step.x_t, start_from=0, color=color, label=f"Step {step_idx}"
                 )
 
             # 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
+                    vt_axs, debug_step.v_t, start_from=0, color=color, label=f"Step {step_idx}"
                 )
 
             # Plot correction on separate axes
@@ -830,18 +925,17 @@ class RTCEvaluator:
                     debug_step.correction,
                     start_from=0,
                     color=color,
-                    label=label,
+                    label=f"Step {step_idx}",
                 )
 
             # 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,
+                    label=f"x1_t Step {step_idx}",
                 )
 
             # Plot error in orange dashed
@@ -853,7 +947,6 @@ class RTCEvaluator:
                 )
 
                 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]),
@@ -861,7 +954,7 @@ class RTCEvaluator:
                         color="orange",
                         linestyle="--",
                         alpha=0.7,
-                        label=error_label,
+                        label=f"error Step {step_idx}",
                     )
 
         # Recalculate axis limits after plotting to ensure proper scaling

examples/rtc/eval_with_real_robot_profiled.py

@@ -0,0 +1,631 @@
+#!/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.
+
+"""
+Profiled version of eval_with_real_robot.py for performance analysis.
+
+This version adds detailed timing measurements for:
+- Policy inference
+- Preprocessing
+- Postprocessing
+- Action queue operations
+- Robot communication
+- Thread execution times
+
+Usage: Same as eval_with_real_robot.py but with profiling output.
+"""
+
+import logging
+import math
+import sys
+import time
+import traceback
+from collections import defaultdict
+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 ProfileTimer:
+    """Context manager and utility class for timing code sections."""
+
+    def __init__(self, name: str, stats_dict: dict):
+        self.name = name
+        self.stats_dict = stats_dict
+        self.start_time = None
+
+    def __enter__(self):
+        self.start_time = time.perf_counter()
+        return self
+
+    def __exit__(self, *args):
+        elapsed = time.perf_counter() - self.start_time
+        if self.name not in self.stats_dict:
+            self.stats_dict[self.name] = []
+        self.stats_dict[self.name].append(elapsed)
+
+
+class ProfilingStats:
+    """Global profiling statistics collector."""
+
+    def __init__(self):
+        self.stats = defaultdict(list)
+        self.lock = Lock()
+
+    def record(self, name: str, duration: float):
+        with self.lock:
+            self.stats[name].append(duration)
+
+    def timer(self, name: str):
+        """Return a context manager for timing."""
+        return ProfileTimer(name, self.stats)
+
+    def get_summary(self) -> dict[str, dict[str, float]]:
+        """Get summary statistics for all timings."""
+        with self.lock:
+            summary = {}
+            for name, times in self.stats.items():
+                if times:
+                    summary[name] = {
+                        "count": len(times),
+                        "mean": sum(times) / len(times),
+                        "min": min(times),
+                        "max": max(times),
+                        "total": sum(times),
+                    }
+            return summary
+
+    def print_summary(self):
+        """Print formatted summary of all timings."""
+        summary = self.get_summary()
+        
+        logger.info("\n" + "=" * 80)
+        logger.info("PROFILING SUMMARY")
+        logger.info("=" * 80)
+        
+        # Sort by total time (descending)
+        sorted_items = sorted(summary.items(), key=lambda x: x[1]["total"], reverse=True)
+        
+        for name, stats in sorted_items:
+            logger.info(f"\n{name}:")
+            logger.info(f"  Count:     {stats['count']}")
+            logger.info(f"  Mean:      {stats['mean']*1000:.2f} ms")
+            logger.info(f"  Min:       {stats['min']*1000:.2f} ms")
+            logger.info(f"  Max:       {stats['max']*1000:.2f} ms")
+            logger.info(f"  Total:     {stats['total']:.2f} s")
+            logger.info(f"  Hz:        {stats['count']/stats['total']:.2f}")
+        
+        logger.info("\n" + "=" * 80)
+
+
+# Global profiling stats
+profiling_stats = ProfilingStats()
+
+
+class RobotWrapper:
+    def __init__(self, robot: Robot):
+        self.robot = robot
+        self.lock = Lock()
+
+    def get_observation(self) -> dict[str, Tensor]:
+        with profiling_stats.timer("robot.get_observation"):
+            with self.lock:
+                return self.robot.get_observation()
+
+    def send_action(self, action: Tensor):
+        with profiling_stats.timer("robot.send_action"):
+            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 with profiling.
+
+    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
+        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
+
+        inference_count = 0
+
+        while not shutdown_event.is_set():
+            if action_queue.qsize() <= get_actions_threshold:
+                with profiling_stats.timer("get_actions.total_iteration"):
+                    inference_count += 1
+                    logger.info(f"[GET_ACTIONS] Starting inference #{inference_count}")
+
+                    current_time = time.perf_counter()
+                    action_index_before_inference = action_queue.get_action_index()
+                    
+                    with profiling_stats.timer("get_actions.get_prev_actions"):
+                        prev_actions = action_queue.get_left_over()
+
+                    inference_latency = latency_tracker.max()
+                    inference_delay = math.ceil(inference_latency / time_per_chunk)
+
+                    # Get observation
+                    obs = robot.get_observation()
+
+                    # Apply robot observation processor
+                    with profiling_stats.timer("get_actions.robot_obs_processing"):
+                        obs_processed = robot_observation_processor(obs)
+
+                    # Build dataset frame
+                    with profiling_stats.timer("get_actions.build_dataset_frame"):
+                        obs_with_policy_features = build_dataset_frame(
+                            dataset_features, obs_processed, prefix="observation"
+                        )
+
+                    # Convert to tensors and normalize
+                    with profiling_stats.timer("get_actions.tensor_conversion"):
+                        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]
+                        obs_with_policy_features["robot_type"] = (
+                            robot.robot.name if hasattr(robot.robot, "name") else ""
+                        )
+
+                    # Preprocessing
+                    with profiling_stats.timer("get_actions.preprocessing"):
+                        preproceseded_obs = preprocessor(obs_with_policy_features)
+
+                    # Policy inference
+                    with profiling_stats.timer("get_actions.policy_inference"):
+                        actions = policy.predict_action_chunk(
+                            preproceseded_obs,
+                            inference_delay=inference_delay,
+                            prev_chunk_left_over=prev_actions,
+                        )
+
+                    # Clone for RTC
+                    with profiling_stats.timer("get_actions.clone_actions"):
+                        original_actions = actions.squeeze(0).clone()
+
+                    # Postprocessing
+                    with profiling_stats.timer("get_actions.postprocessing"):
+                        postprocessed_actions = postprocessor(actions)
+                        postprocessed_actions = postprocessed_actions.squeeze(0)
+
+                    # Update latency tracker
+                    new_latency = time.perf_counter() - current_time
+                    new_delay = math.ceil(new_latency / time_per_chunk)
+                    latency_tracker.add(new_latency)
+
+                    logger.info(
+                        f"[GET_ACTIONS] Inference #{inference_count} completed in {new_latency*1000:.2f}ms "
+                        f"(delay={new_delay} chunks)"
+                    )
+
+                    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."
+                        )
+
+                    # Merge into action queue
+                    with profiling_stats.timer("get_actions.action_queue_merge"):
+                        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 with profiling.
+
+    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()
+
+            with profiling_stats.timer("actor.total_iteration"):
+                # Get action from queue
+                with profiling_stats.timer("actor.queue_get"):
+                    action = action_queue.get()
+
+                if action is not None:
+                    # Process action
+                    with profiling_stats.timer("actor.action_processing"):
+                        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))
+                    
+                    # Send to robot (includes robot.send_action timing)
+                    robot.send_action(action_processed)
+                    action_count += 1
+
+            # Sleep to maintain target FPS
+            dt_s = time.perf_counter() - start_time
+            sleep_time = max(0, (action_interval - dt_s) - 0.001)
+            if sleep_time > 0:
+                time.sleep(sleep_time)
+
+        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
+        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 profiling."""
+
+    # Initialize logging
+    init_logging()
+
+    logger.info(f"Using device: {cfg.device}")
+    logger.info("=" * 80)
+    logger.info("PROFILING MODE ENABLED")
+    logger.info("=" * 80)
+
+    # 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 processor
+    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")
+
+    # Print profiling summary
+    profiling_stats.print_summary()
+
+    logger.info("Cleanup completed")
+
+
+if __name__ == "__main__":
+    demo_cli()
+    logging.info("RTC demo finished")
+

examples/rtc/profile_pi0_rtc_detailed.py

@@ -0,0 +1,358 @@
+#!/usr/bin/env python
+
+"""
+Comprehensive profiling script for Pi0 with RTC.
+
+This script demonstrates how to use all the profiling tools to identify
+bottlenecks in Pi0 policy inference with RTC enabled.
+
+It profiles:
+1. Overall inference time
+2. RTC-specific operations (guidance, weights, etc.)
+3. Preprocessing/postprocessing
+4. Individual method timings
+
+Usage:
+    uv run examples/rtc/profile_pi0_rtc_detailed.py \
+        --policy_path=helper2424/pi05_check_rtc \
+        --device=mps \
+        --num_iterations=20 \
+        --execution_horizon=20 \
+        --enable_rtc_profiling
+"""
+
+import argparse
+import logging
+import sys
+import time
+
+import numpy as np
+import torch
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.utils.profiling import (
+    ProfileContext,
+    clear_profiling_stats,
+    enable_profiling,
+    get_profiling_stats,
+    print_profiling_summary,
+)
+
+# Import monkey patching for RTC profiling
+try:
+    from examples.rtc.add_rtc_profiling import monkey_patch_rtc_profiling
+except ImportError:
+    logging.warning("Could not import add_rtc_profiling, detailed RTC profiling disabled")
+    monkey_patch_rtc_profiling = None
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+def create_mock_observation(policy_config, device: str) -> dict:
+    """Create a mock observation matching policy requirements.
+    
+    Args:
+        policy_config: Policy configuration
+        device: Device to create tensors on
+        
+    Returns:
+        Mock observation dictionary
+    """
+    obs = {}
+    
+    # Create mock state observation
+    state_dim = 10  # Typical robot state dimension
+    obs["observation.state"] = torch.randn(1, state_dim, device=device)
+    
+    # Create mock images if needed
+    # For Pi0, we typically need at least one image
+    image_height = 224
+    image_width = 224
+    
+    # Common image keys for Pi0
+    image_keys = ["observation.images.gripper", "observation.images.front"]
+    
+    for key in image_keys:
+        # Images should be [B, C, H, W] and normalized to [0, 1]
+        obs[key] = torch.rand(1, 3, image_height, image_width, device=device)
+    
+    # Add task
+    obs["task"] = ["Pick up the object"]
+    
+    # Add language tokens and attention mask (required for Pi0)
+    # These are mock values - in real usage they come from tokenizer
+    max_seq_len = 32
+    obs["observation.language_tokens"] = torch.randint(0, 1000, (1, max_seq_len), device=device)
+    obs["observation.language_attention_mask"] = torch.ones(1, max_seq_len, device=device)
+    
+    return obs
+
+
+def profile_single_iteration(
+    policy,
+    preprocessor,
+    postprocessor,
+    observation: dict,
+    prev_actions: torch.Tensor | None,
+    use_rtc: bool,
+    inference_delay: int = 0,
+) -> tuple[torch.Tensor, torch.Tensor | None, dict]:
+    """Profile a single inference iteration.
+    
+    Args:
+        policy: Policy instance
+        preprocessor: Observation preprocessor
+        postprocessor: Action postprocessor
+        observation: Input observation
+        prev_actions: Previous action chunk (for RTC)
+        use_rtc: Whether RTC is enabled
+        inference_delay: Inference delay in timesteps
+        
+    Returns:
+        Tuple of (actions, new_prev_actions, timings)
+    """
+    timings = {}
+    
+    with ProfileContext("iteration.total"):
+        # Preprocessing
+        with ProfileContext("iteration.preprocessing"):
+            preprocessed_obs = preprocessor(observation)
+        
+        # Policy inference
+        with ProfileContext("iteration.policy_inference"):
+            if use_rtc:
+                actions = policy.predict_action_chunk(
+                    preprocessed_obs,
+                    inference_delay=inference_delay,
+                    prev_chunk_left_over=prev_actions,
+                )
+            else:
+                actions = policy.predict_action_chunk(preprocessed_obs)
+        
+        # Clone for next iteration (if RTC)
+        new_prev_actions = None
+        if use_rtc:
+            with ProfileContext("iteration.prepare_prev_actions"):
+                execution_horizon = policy.config.rtc_config.execution_horizon
+                if actions.shape[1] > execution_horizon:
+                    new_prev_actions = actions[:, execution_horizon:].clone()
+        
+        # Postprocessing
+        with ProfileContext("iteration.postprocessing"):
+            processed_actions = postprocessor(actions)
+    
+    return processed_actions, new_prev_actions, timings
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Detailed profiling for Pi0 with RTC")
+    parser.add_argument("--policy_path", type=str, required=True, help="Path to pretrained policy")
+    parser.add_argument("--device", type=str, default="cuda", help="Device (cuda/cpu/mps)")
+    parser.add_argument("--num_iterations", type=int, default=20, help="Number of iterations")
+    parser.add_argument("--execution_horizon", type=int, default=10, help="RTC execution horizon")
+    parser.add_argument("--warmup_iterations", type=int, default=5, help="Warmup iterations")
+    parser.add_argument("--enable_rtc_profiling", action="store_true", help="Enable detailed RTC profiling")
+    parser.add_argument("--use_torch_compile", action="store_true", help="Use torch.compile")
+    
+    args = parser.parse_args()
+    
+    logger.info("="*80)
+    logger.info("DETAILED PI0 RTC PROFILING")
+    logger.info("="*80)
+    logger.info(f"Policy: {args.policy_path}")
+    logger.info(f"Device: {args.device}")
+    logger.info(f"Iterations: {args.num_iterations}")
+    logger.info(f"Execution Horizon: {args.execution_horizon}")
+    logger.info(f"RTC Profiling: {args.enable_rtc_profiling}")
+    logger.info("="*80 + "\n")
+    
+    # Enable profiling
+    enable_profiling()
+    
+    # Apply RTC profiling if requested
+    if args.enable_rtc_profiling:
+        if monkey_patch_rtc_profiling is not None:
+            monkey_patch_rtc_profiling()
+            logger.info("✓ Detailed RTC profiling enabled\n")
+        else:
+            logger.warning("⚠ Could not enable detailed RTC profiling\n")
+    
+    # Load policy
+    logger.info("Loading policy...")
+    config = PreTrainedConfig.from_pretrained(args.policy_path)
+    
+    if hasattr(config, "compile_model"):
+        config.compile_model = args.use_torch_compile
+    
+    policy_class = get_policy_class(config.type)
+    policy = policy_class.from_pretrained(args.policy_path, config=config)
+    
+    # Configure RTC
+    policy.config.rtc_config = RTCConfig(
+        enabled=True,
+        execution_horizon=args.execution_horizon,
+        max_guidance_weight=1.0,
+        prefix_attention_schedule=RTCAttentionSchedule.EXP,
+    )
+    policy.init_rtc_processor()
+    
+    policy = policy.to(args.device)
+    policy.eval()
+    
+    logger.info(f"✓ Policy loaded: {config.type}\n")
+    
+    # Create preprocessor and postprocessor
+    logger.info("Loading preprocessor/postprocessor...")
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=config,
+        pretrained_path=args.policy_path,
+        dataset_stats=None,
+        preprocessor_overrides={
+            "device_processor": {"device": args.device},
+        },
+    )
+    logger.info("✓ Preprocessor/postprocessor loaded\n")
+    
+    # Create mock observation
+    logger.info("Creating mock observation...")
+    observation = create_mock_observation(config, args.device)
+    logger.info("✓ Mock observation created\n")
+    
+    # Warmup
+    logger.info(f"Warming up ({args.warmup_iterations} iterations)...")
+    prev_actions = None
+    for i in range(args.warmup_iterations):
+        with torch.no_grad():
+            _, prev_actions, _ = profile_single_iteration(
+                policy=policy,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                observation=observation,
+                prev_actions=prev_actions,
+                use_rtc=True,
+                inference_delay=0,
+            )
+    
+    # Clear warmup stats
+    clear_profiling_stats()
+    logger.info("✓ Warmup complete\n")
+    
+    # Profiled run WITH RTC
+    logger.info(f"Running profiled iterations WITH RTC ({args.num_iterations} iterations)...")
+    prev_actions = None
+    iteration_times = []
+    
+    for i in range(args.num_iterations):
+        start = time.perf_counter()
+        
+        with torch.no_grad():
+            _, prev_actions, _ = profile_single_iteration(
+                policy=policy,
+                preprocessor=preprocessor,
+                postprocessor=postprocessor,
+                observation=observation,
+                prev_actions=prev_actions,
+                use_rtc=True,
+                inference_delay=0,
+            )
+        
+        # Sync CUDA if needed
+        if args.device.startswith("cuda"):
+            torch.cuda.synchronize()
+        
+        elapsed = time.perf_counter() - start
+        iteration_times.append(elapsed)
+        
+        if (i + 1) % 5 == 0:
+            logger.info(f"  Completed {i+1}/{args.num_iterations}")
+    
+    logger.info("✓ Profiling complete\n")
+    
+    # Print summary statistics
+    logger.info("\n" + "="*80)
+    logger.info("ITERATION TIMING SUMMARY")
+    logger.info("="*80)
+    
+    times_arr = np.array(iteration_times)
+    logger.info(f"Mean time:       {np.mean(times_arr)*1000:.2f} ms")
+    logger.info(f"Median time:     {np.median(times_arr)*1000:.2f} ms")
+    logger.info(f"Std dev:         {np.std(times_arr)*1000:.2f} ms")
+    logger.info(f"Min time:        {np.min(times_arr)*1000:.2f} ms")
+    logger.info(f"Max time:        {np.max(times_arr)*1000:.2f} ms")
+    logger.info(f"Total time:      {np.sum(times_arr):.2f} s")
+    logger.info(f"Throughput:      {len(times_arr)/np.sum(times_arr):.2f} iter/s")
+    logger.info("="*80 + "\n")
+    
+    # Print detailed profiling breakdown
+    print_profiling_summary(sort_by="total")
+    
+    # Print key insights
+    stats = get_profiling_stats()
+    
+    logger.info("\n" + "="*80)
+    logger.info("KEY INSIGHTS")
+    logger.info("="*80)
+    
+    # Find bottlenecks
+    if stats:
+        policy_inference_time = stats.get("iteration.policy_inference", {}).get("mean", 0)
+        preprocessing_time = stats.get("iteration.preprocessing", {}).get("mean", 0)
+        postprocessing_time = stats.get("iteration.postprocessing", {}).get("mean", 0)
+        
+        total_time = policy_inference_time + preprocessing_time + postprocessing_time
+        
+        if total_time > 0:
+            logger.info(f"\nTime breakdown:")
+            logger.info(f"  Policy inference:  {policy_inference_time*1000:.2f} ms ({policy_inference_time/total_time*100:.1f}%)")
+            logger.info(f"  Preprocessing:     {preprocessing_time*1000:.2f} ms ({preprocessing_time/total_time*100:.1f}%)")
+            logger.info(f"  Postprocessing:    {postprocessing_time*1000:.2f} ms ({postprocessing_time/total_time*100:.1f}%)")
+        
+        # RTC-specific insights
+        if args.enable_rtc_profiling:
+            rtc_guidance = stats.get("rtc.denoise_step.guidance_computation", {}).get("mean", 0)
+            rtc_autograd = stats.get("rtc.denoise_step.autograd_correction", {}).get("mean", 0)
+            rtc_base = stats.get("rtc.denoise_step.base_denoising", {}).get("mean", 0)
+            
+            if rtc_guidance > 0:
+                logger.info(f"\nRTC breakdown:")
+                logger.info(f"  Base denoising:    {rtc_base*1000:.2f} ms")
+                logger.info(f"  Guidance compute:  {rtc_guidance*1000:.2f} ms")
+                logger.info(f"  Autograd correct:  {rtc_autograd*1000:.2f} ms")
+                logger.info(f"  RTC overhead:      {(rtc_guidance - rtc_base)*1000:.2f} ms")
+        
+        # Recommendations
+        logger.info("\nRecommendations:")
+        
+        if preprocessing_time > policy_inference_time * 0.3:
+            logger.info("  ⚠ Preprocessing is taking >30% of time")
+            logger.info("    → Consider reducing image resolution")
+            logger.info("    → Consider using fewer cameras")
+        
+        if args.enable_rtc_profiling and rtc_autograd > rtc_base * 0.5:
+            logger.info("  ⚠ RTC autograd overhead is significant")
+            logger.info("    → This is expected, but consider increasing execution_horizon")
+            logger.info("    → Try torch.compile if not already enabled")
+        
+        if not args.use_torch_compile:
+            logger.info("  💡 torch.compile not enabled")
+            logger.info("    → Try --use_torch_compile for potential speedup")
+    
+    logger.info("="*80 + "\n")
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except KeyboardInterrupt:
+        logger.info("\n\nProfiling interrupted by user")
+        sys.exit(0)
+    except Exception as e:
+        logger.error(f"\n\nError during profiling: {e}")
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)
+

examples/rtc/profile_rtc_comparison.py

@@ -0,0 +1,347 @@
+#!/usr/bin/env python
+
+"""
+Script to compare performance with and without RTC enabled.
+
+This script helps identify whether RTC is actually improving or degrading performance
+by running multiple inference passes and collecting detailed timing statistics.
+
+Usage:
+    # Profile with mock data (no robot needed)
+    uv run examples/rtc/profile_rtc_comparison.py \
+        --policy_path=helper2424/pi05_check_rtc \
+        --device=mps \
+        --num_iterations=50
+
+    # Profile with specific RTC config
+    uv run examples/rtc/profile_rtc_comparison.py \
+        --policy_path=helper2424/pi05_check_rtc \
+        --device=mps \
+        --num_iterations=50 \
+        --execution_horizon=20
+"""
+
+import argparse
+import logging
+import time
+from dataclasses import dataclass
+
+import numpy as np
+import torch
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import RTCAttentionSchedule
+from lerobot.policies.factory import get_policy_class, make_pre_post_processors
+from lerobot.policies.rtc.configuration_rtc import RTCConfig
+from lerobot.utils.profiling import (
+    clear_profiling_stats,
+    enable_profiling,
+    get_profiling_stats,
+    print_profiling_summary,
+)
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ProfileResults:
+    """Results from profiling run."""
+
+    mode: str  # "with_rtc" or "without_rtc"
+    mean_time: float
+    std_time: float
+    min_time: float
+    max_time: float
+    times: list[float]
+    throughput: float  # iterations per second
+
+
+def create_mock_observation(policy, device: str) -> dict:
+    """Create a mock observation for testing.
+
+    Args:
+        policy: Policy instance
+        device: Device to create tensors on
+
+    Returns:
+        Mock observation dictionary
+    """
+    # Get expected input shapes from policy config
+    # This is a simplified version - adjust based on actual policy requirements
+    obs = {}
+
+    # Mock image observations (if needed)
+    if hasattr(policy.config, "input_shapes"):
+        for key, shape in policy.config.input_shapes.items():
+            if "image" in key:
+                # Typical image shape: (batch, channels, height, width)
+                obs[key] = torch.randn(1, *shape, device=device)
+            else:
+                obs[key] = torch.randn(1, *shape, device=device)
+
+    # Add task if needed
+    if "task" in policy.config.__dict__ or hasattr(policy, "accepts_task"):
+        obs["task"] = ["Pick up the object"]
+
+    # Mock state observation
+    obs["observation.state"] = torch.randn(1, 10, device=device)  # Adjust size as needed
+
+    return obs
+
+
+def profile_inference(
+    policy, observation: dict, num_iterations: int, use_rtc: bool, execution_horizon: int = 10
+) -> ProfileResults:
+    """Profile policy inference with or without RTC.
+
+    Args:
+        policy: Policy instance
+        observation: Observation dictionary
+        num_iterations: Number of inference iterations to run
+        use_rtc: Whether to enable RTC
+        execution_horizon: Execution horizon for RTC
+
+    Returns:
+        ProfileResults with timing statistics
+    """
+    mode = "with_rtc" if use_rtc else "without_rtc"
+    logger.info(f"\n{'='*80}")
+    logger.info(f"Profiling: {mode.upper()}")
+    logger.info(f"{'='*80}")
+
+    # Configure RTC
+    if use_rtc:
+        policy.config.rtc_config.enabled = True
+        policy.config.rtc_config.execution_horizon = execution_horizon
+        policy.init_rtc_processor()
+    else:
+        policy.config.rtc_config.enabled = False
+
+    times = []
+    prev_actions = None
+
+    # Warmup
+    logger.info("Warming up (5 iterations)...")
+    for _ in range(5):
+        with torch.no_grad():
+            if use_rtc:
+                _ = policy.predict_action_chunk(
+                    observation, inference_delay=0, prev_chunk_left_over=prev_actions
+                )
+            else:
+                _ = policy.predict_action_chunk(observation)
+
+    # Actual profiling
+    logger.info(f"Running {num_iterations} profiled iterations...")
+    for i in range(num_iterations):
+        start = time.perf_counter()
+
+        with torch.no_grad():
+            if use_rtc:
+                actions = policy.predict_action_chunk(
+                    observation, inference_delay=0, prev_chunk_left_over=prev_actions
+                )
+                # Simulate consuming some actions for next iteration
+                if actions.shape[1] > execution_horizon:
+                    prev_actions = actions[:, execution_horizon:].clone()
+                else:
+                    prev_actions = None
+            else:
+                actions = policy.predict_action_chunk(observation)
+
+        # Synchronize if using CUDA
+        if observation["observation.state"].device.type == "cuda":
+            torch.cuda.synchronize()
+
+        elapsed = time.perf_counter() - start
+        times.append(elapsed)
+
+        if (i + 1) % 10 == 0:
+            logger.info(f"  Completed {i+1}/{num_iterations} iterations")
+
+    # Calculate statistics
+    times_arr = np.array(times)
+    results = ProfileResults(
+        mode=mode,
+        mean_time=float(np.mean(times_arr)),
+        std_time=float(np.std(times_arr)),
+        min_time=float(np.min(times_arr)),
+        max_time=float(np.max(times_arr)),
+        times=times,
+        throughput=num_iterations / sum(times),
+    )
+
+    logger.info(f"\nResults for {mode}:")
+    logger.info(f"  Mean time: {results.mean_time*1000:.2f} ms")
+    logger.info(f"  Std dev:   {results.std_time*1000:.2f} ms")
+    logger.info(f"  Min time:  {results.min_time*1000:.2f} ms")
+    logger.info(f"  Max time:  {results.max_time*1000:.2f} ms")
+    logger.info(f"  Throughput: {results.throughput:.2f} iter/s")
+
+    return results
+
+
+def compare_results(results_without_rtc: ProfileResults, results_with_rtc: ProfileResults):
+    """Compare and print results from both runs.
+
+    Args:
+        results_without_rtc: Results from run without RTC
+        results_with_rtc: Results from run with RTC
+    """
+    logger.info(f"\n{'='*80}")
+    logger.info("COMPARISON SUMMARY")
+    logger.info(f"{'='*80}")
+
+    mean_diff = results_with_rtc.mean_time - results_without_rtc.mean_time
+    mean_diff_pct = (mean_diff / results_without_rtc.mean_time) * 100
+
+    throughput_diff = results_with_rtc.throughput - results_without_rtc.throughput
+    throughput_diff_pct = (throughput_diff / results_without_rtc.throughput) * 100
+
+    logger.info(f"\n{'Metric':<30} {'Without RTC':>15} {'With RTC':>15} {'Difference':>15}")
+    logger.info("-" * 80)
+    logger.info(
+        f"{'Mean time (ms)':<30} "
+        f"{results_without_rtc.mean_time*1000:>15.2f} "
+        f"{results_with_rtc.mean_time*1000:>15.2f} "
+        f"{mean_diff*1000:>+15.2f}"
+    )
+    logger.info(
+        f"{'Std dev (ms)':<30} "
+        f"{results_without_rtc.std_time*1000:>15.2f} "
+        f"{results_with_rtc.std_time*1000:>15.2f} "
+        f"{(results_with_rtc.std_time - results_without_rtc.std_time)*1000:>+15.2f}"
+    )
+    logger.info(
+        f"{'Min time (ms)':<30} "
+        f"{results_without_rtc.min_time*1000:>15.2f} "
+        f"{results_with_rtc.min_time*1000:>15.2f} "
+        f"{(results_with_rtc.min_time - results_without_rtc.min_time)*1000:>+15.2f}"
+    )
+    logger.info(
+        f"{'Max time (ms)':<30} "
+        f"{results_without_rtc.max_time*1000:>15.2f} "
+        f"{results_with_rtc.max_time*1000:>15.2f} "
+        f"{(results_with_rtc.max_time - results_without_rtc.max_time)*1000:>+15.2f}"
+    )
+    logger.info(
+        f"{'Throughput (iter/s)':<30} "
+        f"{results_without_rtc.throughput:>15.2f} "
+        f"{results_with_rtc.throughput:>15.2f} "
+        f"{throughput_diff:>+15.2f}"
+    )
+
+    logger.info(f"\n{'='*80}")
+    logger.info("VERDICT")
+    logger.info(f"{'='*80}")
+
+    if mean_diff_pct < -5:
+        logger.info(f"✓ RTC is FASTER by {abs(mean_diff_pct):.1f}%")
+        logger.info(f"  Mean time reduced by {abs(mean_diff)*1000:.2f} ms")
+    elif mean_diff_pct > 5:
+        logger.info(f"✗ RTC is SLOWER by {mean_diff_pct:.1f}%")
+        logger.info(f"  Mean time increased by {mean_diff*1000:.2f} ms")
+        logger.info("\n  Possible reasons:")
+        logger.info("  - RTC overhead exceeds benefits at current execution horizon")
+        logger.info("  - Inference delay calculation not accounting for RTC processing")
+        logger.info("  - Additional tensor operations in RTC guidance")
+    else:
+        logger.info(f"≈ Performance is SIMILAR (difference: {mean_diff_pct:+.1f}%)")
+
+    logger.info(f"{'='*80}\n")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Profile RTC performance")
+    parser.add_argument(
+        "--policy_path", type=str, required=True, help="Path to pretrained policy"
+    )
+    parser.add_argument(
+        "--device", type=str, default="cuda", help="Device to run on (cuda/cpu/mps)"
+    )
+    parser.add_argument(
+        "--num_iterations", type=int, default=50, help="Number of inference iterations"
+    )
+    parser.add_argument(
+        "--execution_horizon", type=int, default=10, help="RTC execution horizon"
+    )
+    parser.add_argument(
+        "--enable_detailed_profiling",
+        action="store_true",
+        help="Enable detailed method-level profiling",
+    )
+    parser.add_argument(
+        "--use_torch_compile", action="store_true", help="Use torch.compile for faster inference"
+    )
+
+    args = parser.parse_args()
+
+    # Load policy
+    logger.info(f"Loading policy from {args.policy_path}")
+    config = PreTrainedConfig.from_pretrained(args.policy_path)
+    policy_class = get_policy_class(config.type)
+
+    # Set compile flag if needed
+    if hasattr(config, "compile_model"):
+        config.compile_model = args.use_torch_compile
+
+    policy = policy_class.from_pretrained(args.policy_path, config=config)
+
+    # Initialize RTC config
+    policy.config.rtc_config = RTCConfig(
+        execution_horizon=args.execution_horizon,
+        max_guidance_weight=1.0,
+        prefix_attention_schedule=RTCAttentionSchedule.EXP,
+    )
+
+    policy = policy.to(args.device)
+    policy.eval()
+
+    logger.info(f"Policy loaded: {config.type}")
+    logger.info(f"Device: {args.device}")
+    logger.info(f"Execution horizon: {args.execution_horizon}")
+
+    # Create mock observation
+    logger.info("Creating mock observation...")
+    observation = create_mock_observation(policy, args.device)
+
+    # Enable detailed profiling if requested
+    if args.enable_detailed_profiling:
+        enable_profiling()
+        logger.info("Detailed profiling enabled")
+
+    # Profile without RTC
+    results_without_rtc = profile_inference(
+        policy=policy,
+        observation=observation,
+        num_iterations=args.num_iterations,
+        use_rtc=False,
+        execution_horizon=args.execution_horizon,
+    )
+
+    if args.enable_detailed_profiling:
+        logger.info("\nDetailed profiling stats (WITHOUT RTC):")
+        print_profiling_summary()
+        clear_profiling_stats()
+
+    # Profile with RTC
+    results_with_rtc = profile_inference(
+        policy=policy,
+        observation=observation,
+        num_iterations=args.num_iterations,
+        use_rtc=True,
+        execution_horizon=args.execution_horizon,
+    )
+
+    if args.enable_detailed_profiling:
+        logger.info("\nDetailed profiling stats (WITH RTC):")
+        print_profiling_summary()
+
+    # Compare results
+    compare_results(results_without_rtc, results_with_rtc)
+
+
+if __name__ == "__main__":
+    main()
+

examples/so100_to_so100_EE/evaluate.py

@@ -52,114 +52,126 @@ 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,
+)
 
-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,
-    )
+robot = SO100Follower(robot_config)
 
-    robot = SO100Follower(robot_config)
+# Create policy
+policy = ACTPolicy.from_pretrained(HF_MODEL_ID)
 
-    # 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()),
+)
 
-    # 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,
-            ),
+# 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,
         ),
-        robot_type=robot.name,
-        use_videos=True,
-        image_writer_threads=4,
+    ],
+    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,
     )
 
-    # 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
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
+        log_say("Reset the environment")
         record_loop(
             robot=robot,
             events=events,
             fps=FPS,
-            policy=policy,
-            preprocessor=preprocessor,  # Pass the pre and post policy processors
-            postprocessor=postprocessor,
-            dataset=dataset,
             control_time_s=EPISODE_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
@@ -168,40 +180,21 @@ def main():
             robot_observation_processor=robot_joints_to_ee_pose_processor,
         )
 
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and ((episode_idx < NUM_EPISODES - 1) or events["rerecord_episode"]):
-            log_say("Reset the environment")
-            record_loop(
-                robot=robot,
-                events=events,
-                fps=FPS,
-                control_time_s=EPISODE_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=make_default_teleop_action_processor(),
-                robot_action_processor=robot_ee_to_joints_processor,
-                robot_observation_processor=robot_joints_to_ee_pose_processor,
-            )
+    if events["rerecord_episode"]:
+        log_say("Re-record episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
 
-        if events["rerecord_episode"]:
-            log_say("Re-record episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+# Clean up
+log_say("Stop recording")
+robot.disconnect()
+listener.stop()
 
-    # Clean up
-    log_say("Stop recording")
-    robot.disconnect()
-    listener.stop()
-
-    dataset.finalize()
-    dataset.push_to_hub()
-
-
-if __name__ == "__main__":
-    main()
+dataset.finalize()
+dataset.push_to_hub()

examples/so100_to_so100_EE/record.py

@@ -48,122 +48,134 @@ 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")
 
-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")
+# Initialize the robot and teleoperator
+follower = SO100Follower(follower_config)
+leader = SO100Leader(leader_config)
 
-    # 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
+follower_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(follower.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()),
-    )
+# 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
-    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,
-            ),
+# 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())
         ),
-        robot_type=follower.name,
-        use_videos=True,
-        image_writer_threads=4,
+    ],
+    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,
     )
 
-    # 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
+    # Reset the environment if not stopping or re-recording
+    if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
+        log_say("Reset the environment")
         record_loop(
             robot=follower,
             events=events,
             fps=FPS,
             teleop=leader,
-            dataset=dataset,
-            control_time_s=EPISODE_TIME_SEC,
+            control_time_s=RESET_TIME_SEC,
             single_task=TASK_DESCRIPTION,
             display_data=True,
             teleop_action_processor=leader_joints_to_ee,
@@ -171,42 +183,22 @@ def main():
             robot_observation_processor=follower_joints_to_ee,
         )
 
-        # Reset the environment if not stopping or re-recording
-        if not events["stop_recording"] and (episode_idx < NUM_EPISODES - 1 or events["rerecord_episode"]):
-            log_say("Reset the environment")
-            record_loop(
-                robot=follower,
-                events=events,
-                fps=FPS,
-                teleop=leader,
-                control_time_s=RESET_TIME_SEC,
-                single_task=TASK_DESCRIPTION,
-                display_data=True,
-                teleop_action_processor=leader_joints_to_ee,
-                robot_action_processor=ee_to_follower_joints,
-                robot_observation_processor=follower_joints_to_ee,
-            )
+    if events["rerecord_episode"]:
+        log_say("Re-recording episode")
+        events["rerecord_episode"] = False
+        events["exit_early"] = False
+        dataset.clear_episode_buffer()
+        continue
 
-        if events["rerecord_episode"]:
-            log_say("Re-recording episode")
-            events["rerecord_episode"] = False
-            events["exit_early"] = False
-            dataset.clear_episode_buffer()
-            continue
+    # Save episode
+    dataset.save_episode()
+    episode_idx += 1
 
-        # Save episode
-        dataset.save_episode()
-        episode_idx += 1
+# Clean up
+log_say("Stop recording")
+leader.disconnect()
+follower.disconnect()
+listener.stop()
 
-    # Clean up
-    log_say("Stop recording")
-    leader.disconnect()
-    follower.disconnect()
-    listener.stop()
-
-    dataset.finalize()
-    dataset.push_to_hub()
-
-
-if __name__ == "__main__":
-    main()
+dataset.finalize()
+dataset.push_to_hub()

examples/so100_to_so100_EE/replay.py

@@ -30,78 +30,72 @@ 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 precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 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
+)
 
-def main():
-    # 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)
 
-    # Initialize the robot
-    robot = SO100Follower(robot_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
+kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(robot.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
-    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=False,  # Because replay is open loop
+        ),
+    ],
+    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=False,  # Because replay is open loop
-            ),
-        ],
-        to_transition=robot_action_observation_to_transition,
-        to_output=transition_to_robot_action,
-    )
+# 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)
 
-    # 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)
+# Connect to the robot
+robot.connect()
 
-    # Connect to the robot
-    robot.connect()
+if not robot.is_connected:
+    raise ValueError("Robot is not connected!")
 
-    if not robot.is_connected:
-        raise ValueError("Robot is not connected!")
+print("Starting replay loop...")
+log_say(f"Replaying episode {EPISODE_IDX}")
+for idx in range(len(episode_frames)):
+    t0 = time.perf_counter()
 
-    print("Starting replay loop...")
-    log_say(f"Replaying episode {EPISODE_IDX}")
-    for idx in range(len(episode_frames)):
-        t0 = time.perf_counter()
+    # Get recorded action from dataset
+    ee_action = {
+        name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
+    }
 
-        # Get recorded action from dataset
-        ee_action = {
-            name: float(actions[idx][ACTION][i]) for i, name in enumerate(dataset.features[ACTION]["names"])
-        }
+    # Get robot observation
+    robot_obs = robot.get_observation()
 
-        # Get robot observation
-        robot_obs = robot.get_observation()
+    # Dataset EE -> robot joints
+    joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
 
-        # Dataset EE -> robot joints
-        joint_action = robot_ee_to_joints_processor((ee_action, robot_obs))
+    # Send action to robot
+    _ = robot.send_action(joint_action)
 
-        # Send action to robot
-        _ = robot.send_action(joint_action)
+    busy_wait(1.0 / dataset.fps - (time.perf_counter() - t0))
 
-        precise_sleep(1.0 / dataset.fps - (time.perf_counter() - t0))
-
-    # Clean up
-    robot.disconnect()
-
-
-if __name__ == "__main__":
-    main()
+# Clean up
+robot.disconnect()

examples/so100_to_so100_EE/teleoperate.py

@@ -32,96 +32,90 @@ 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 precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 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")
 
-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")
+# Initialize the robot and teleoperator
+follower = SO100Follower(follower_config)
+leader = SO100Leader(leader_config)
 
-    # 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
+follower_kinematics_solver = RobotKinematics(
+    urdf_path="./SO101/so101_new_calib.urdf",
+    target_frame_name="gripper_frame_link",
+    joint_names=list(follower.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()),
-    )
+# 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
-    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 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,
+)
 
-    # 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,
-    )
+# 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 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,
-    )
+# Connect to the robot and teleoperator
+follower.connect()
+leader.connect()
 
-    # Connect to the robot and teleoperator
-    follower.connect()
-    leader.connect()
+# Init rerun viewer
+init_rerun(session_name="so100_so100_EE_teleop")
 
-    # Init rerun viewer
-    init_rerun(session_name="so100_so100_EE_teleop")
+print("Starting teleop loop...")
+while True:
+    t0 = time.perf_counter()
 
-    print("Starting teleop loop...")
-    while True:
-        t0 = time.perf_counter()
+    # Get robot observation
+    robot_obs = follower.get_observation()
 
-        # Get robot observation
-        robot_obs = follower.get_observation()
+    # Get teleop observation
+    leader_joints_obs = leader.get_action()
 
-        # Get teleop observation
-        leader_joints_obs = leader.get_action()
+    # teleop joints -> teleop EE action
+    leader_ee_act = leader_to_ee(leader_joints_obs)
 
-        # teleop joints -> teleop EE action
-        leader_ee_act = leader_to_ee(leader_joints_obs)
+    # teleop EE -> robot joints
+    follower_joints_act = ee_to_follower_joints((leader_ee_act, robot_obs))
 
-        # teleop EE -> robot joints
-        follower_joints_act = ee_to_follower_joints((leader_ee_act, robot_obs))
+    # Send action to robot
+    _ = follower.send_action(follower_joints_act)
 
-        # Send action to robot
-        _ = follower.send_action(follower_joints_act)
+    # Visualize
+    log_rerun_data(observation=leader_ee_act, action=follower_joints_act)
 
-        # 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()
+    busy_wait(max(1.0 / FPS - (time.perf_counter() - t0), 0.0))

examples/tutorial/act/act_training_example.py

@@ -19,86 +19,80 @@ def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[flo
     return [i / fps for i in delta_indices]
 
 
-def main():
-    output_directory = Path("outputs/robot_learning_tutorial/act")
-    output_directory.mkdir(parents=True, exist_ok=True)
+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("<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()
+# 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")

examples/tutorial/act/act_using_example.py

@@ -8,56 +8,50 @@ 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),
+}
 
-def main():
-    device = torch.device("mps")  # or "cuda" or "cpu"
-    model_id = "<user>/robot_learning_tutorial_act"
-    model = ACTPolicy.from_pretrained(model_id)
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+robot = SO100Follower(robot_cfg)
+robot.connect()
 
-    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)
+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
+        )
 
-    # # find ports using lerobot-find-port
-    follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
+        obs = preprocess(obs_frame)
 
-    # # the robot ids are used the load the right calibration files
-    follower_id = ...  # something like "follower_so100"
+        action = model.select_action(obs)
+        action = postprocess(action)
 
-    # 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),
-    }
+        action = make_robot_action(action, dataset_metadata.features)
 
-    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-    robot = SO100Follower(robot_cfg)
-    robot.connect()
+        robot.send_action(action)
 
-    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()
+    print("Episode finished! Starting new episode...")

examples/tutorial/async-inf/policy_server.py

@@ -1,17 +1,11 @@
 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
 
-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()
+config = PolicyServerConfig(
+    host=host,
+    port=port,
+)
+serve(config)

examples/tutorial/async-inf/robot_client.py

@@ -6,56 +6,50 @@ 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),
+}
 
-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),
-    }
+# # find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
 
-    # # 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"
 
-    # # 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_cfg)
 
-    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_cfg)
+server_address = ...  # something like "127.0.0.1:8080" if using localhost
 
-    server_address = ...  # something like "127.0.0.1:8080" if using localhost
+# 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
+)
 
-    # 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
-    )
+# 4. Create and start client
+client = RobotClient(client_cfg)
 
-    # 4. Create and start client
-    client = RobotClient(client_cfg)
+# 5. Provide a textual description of the task
+task = ...
 
-    # 5. Provide a textual description of the task
-    task = ...
+if client.start():
+    # Start action receiver thread
+    action_receiver_thread = threading.Thread(target=client.receive_actions, daemon=True)
+    action_receiver_thread.start()
 
-    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()
+    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)

examples/tutorial/diffusion/diffusion_training_example.py

@@ -19,87 +19,81 @@ def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[flo
     return [i / fps for i in delta_indices]
 
 
-def main():
-    output_directory = Path("outputs/robot_learning_tutorial/diffusion")
-    output_directory.mkdir(parents=True, exist_ok=True)
+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("<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()
+# 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")

examples/tutorial/diffusion/diffusion_using_example.py

@@ -8,57 +8,53 @@ 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
 
 
-def main():
-    device = torch.device("mps")  # or "cuda" or "cpu"
-    model_id = "<user>/robot_learning_tutorial_diffusion"
+# # find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
 
-    model = DiffusionPolicy.from_pretrained(model_id)
+# # the robot ids are used the load the right calibration files
+follower_id = ...  # something like "follower_so100"
 
-    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 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),
+}
 
-    # # 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...")
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+robot = SO100Follower(robot_cfg)
+robot.connect()
 
 
-if __name__ == "__main__":
-    main()
+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...")

examples/tutorial/pi0/using_pi0_example.py

@@ -11,63 +11,57 @@ 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"
 
-def main():
-    device = torch.device("mps")  # or "cuda" or "cpu"
-    model_id = "lerobot/pi0_base"
+model = PI0Policy.from_pretrained(model_id)
 
-    model = PI0Policy.from_pretrained(model_id)
+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)}},
+)
 
-    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)}},
-    )
+# find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
 
-    # 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"
 
-    # 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 = {
+    "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),
+}
 
-    # 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),
-    }
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+robot = SO100Follower(robot_cfg)
+robot.connect()
 
-    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-    robot = SO100Follower(robot_cfg)
-    robot.connect()
+task = ""  # something like "pick the red block"
+robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
 
-    task = ""  # something like "pick the red block"
-    robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
+# 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}
 
-    # 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}
+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
+        )
 
-    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
-            )
+        obs = preprocess(obs_frame)
 
-            obs = preprocess(obs_frame)
+        action = model.select_action(obs)
+        action = postprocess(action)
+        action = make_robot_action(action, dataset_features)
+        robot.send_action(action)
 
-            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()
+    print("Episode finished! Starting new episode...")

examples/tutorial/rl/hilserl_example.py

@@ -20,8 +20,6 @@ from lerobot.teleoperators.utils import TeleopEvents
 
 LOG_EVERY = 10
 SEND_EVERY = 10
-MAX_EPISODES = 5
-MAX_STEPS_PER_EPISODE = 20
 
 
 def run_learner(
@@ -225,123 +223,123 @@ def make_policy_obs(obs, device: torch.device = "cpu"):
     }
 
 
-def main():
-    """Main function - coordinates actor and learner processes."""
+"""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 = "<user>/reward_classifier_hil_serl_example"
-    reward_classifier = Classifier.from_pretrained(reward_classifier_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)
 
-    reward_classifier.to(device)
-    reward_classifier.eval()
+reward_classifier.to(device)
+reward_classifier.eval()
 
-    # 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")
+MAX_EPISODES = 5
+MAX_STEPS_PER_EPISODE = 20
 
-    env_cfg = HILSerlRobotEnvConfig(robot=robot_cfg, teleop=teleop_cfg, processor=processor_cfg)
+# 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")
 
-    # Create robot environment
-    env, teleop_device = make_robot_env(env_cfg)
+env_cfg = HILSerlRobotEnvConfig(robot=robot_cfg, teleop=teleop_cfg, processor=processor_cfg)
 
-    obs_features = hw_to_dataset_features(env.robot.observation_features, "observation")
-    action_features = hw_to_dataset_features(env.robot.action_features, "action")
+# Create robot environment
+env, teleop_device = make_robot_env(env_cfg)
 
-    # Create SAC policy for action selection
-    policy_cfg = SACConfig(
-        device=device,
-        input_features=obs_features,
-        output_features=action_features,
-    )
+obs_features = hw_to_dataset_features(env.robot.observation_features, "observation")
+action_features = hw_to_dataset_features(env.robot.action_features, "action")
 
-    policy_actor = SACPolicy(policy_cfg)
-    policy_learner = SACPolicy(policy_cfg)
+# Create SAC policy for action selection
+policy_cfg = SACConfig(
+    device=device,
+    input_features=obs_features,
+    output_features=action_features,
+)
 
-    demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
-    offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)
+policy_actor = SACPolicy(policy_cfg)
+policy_learner = SACPolicy(policy_cfg)
 
-    # 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())
-    )
+demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
+offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)
 
-    # Create communication channels between learner and actor processes
-    transitions_queue = mp.Queue(maxsize=10)
-    parameters_queue = mp.Queue(maxsize=2)
-    shutdown_event = mp.Event()
+# 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())
+)
 
-    # 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()
+# Create communication channels between learner and actor processes
+transitions_queue = mp.Queue(maxsize=10)
+parameters_queue = mp.Queue(maxsize=2)
+shutdown_event = mp.Event()
 
 
-if __name__ == "__main__":
-    main()
+# 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()

examples/tutorial/rl/reward_classifier_example.py

@@ -4,64 +4,59 @@ 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"
 
-def main():
-    # 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)
 
-    # Load the dataset used for training
-    repo_id = "lerobot/example_hil_serl_dataset"
-    dataset = LeRobotDataset(repo_id)
+# Configure the policy to extract features from the image frames
+camera_keys = dataset.meta.camera_keys
 
-    # Configure the policy to extract features from the image frames
-    camera_keys = dataset.meta.camera_keys
+config = RewardClassifierConfig(
+    num_cameras=len(camera_keys),
+    device=device,
+    # backbone model to extract features from the image frames
+    model_name="microsoft/resnet-18",
+)
 
-    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)
+# 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)
 
 
-if __name__ == "__main__":
-    main()
+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)

examples/tutorial/smolvla/using_smolvla_example.py

@@ -11,62 +11,56 @@ 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"
 
-def main():
-    device = torch.device("mps")  # or "cuda" or "cpu"
-    model_id = "lerobot/smolvla_base"
+model = SmolVLAPolicy.from_pretrained(model_id)
 
-    model = SmolVLAPolicy.from_pretrained(model_id)
+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)}},
+)
 
-    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)}},
-    )
+# find ports using lerobot-find-port
+follower_port = ...  # something like "/dev/tty.usbmodem58760431631"
 
-    # 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"
 
-    # 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 = {
+    "camera1": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
+    "camera2": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
+}
 
-    # 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),
-    }
+robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
+robot = SO100Follower(robot_cfg)
+robot.connect()
 
-    robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
-    robot = SO100Follower(robot_cfg)
-    robot.connect()
+task = ""  # something like "pick the red block"
+robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
 
-    task = ""  # something like "pick the red block"
-    robot_type = ""  # something like "so100_follower" for multi-embodiment datasets
+# 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}
 
-    # 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}
+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
+        )
 
-    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
-            )
+        obs = preprocess(obs_frame)
 
-            obs = preprocess(obs_frame)
+        action = model.select_action(obs)
+        action = postprocess(action)
+        action = make_robot_action(action, dataset_features)
+        robot.send_action(action)
 
-            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()
+    print("Episode finished! Starting new episode...")

pyproject.toml

@@ -25,7 +25,7 @@ discord = "https://discord.gg/s3KuuzsPFb"
 
 [project]
 name = "lerobot"
-version = "0.4.3"
+version = "0.4.2"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 readme = "README.md"
 license = { text = "Apache-2.0" }
@@ -98,6 +98,7 @@ pygame-dep = ["pygame>=2.5.1,<2.7.0"]
 placo-dep = ["placo>=0.9.6,<0.10.0"]
 transformers-dep = ["transformers>=4.53.0,<5.0.0"]
 grpcio-dep = ["grpcio==1.73.1", "protobuf==6.31.0"] # TODO: Bumb dependency (compatible with wandb)
+matplotlib-dep = ["matplotlib>=3.10.3,<4.0.0"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
@@ -129,11 +130,10 @@ 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
-async = ["lerobot[grpcio-dep]", "matplotlib>=3.10.3,<4.0.0"]
+async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
 
 # Development
 dev = ["pre-commit>=3.7.0,<5.0.0", "debugpy>=1.8.1,<1.9.0", "lerobot[grpcio-dep]", "grpcio-tools==1.73.1"]
@@ -158,7 +158,6 @@ 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]",

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, compress_level = item
-        write_image(image_array, fpath, compress_level)
+        image_array, fpath = item
+        write_image(image_array, fpath)
         queue.task_done()
 
 
@@ -169,13 +169,11 @@ class AsyncImageWriter:
                 p.start()
                 self.processes.append(p)
 
-    def save_image(
-        self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path, compress_level: int = 1
-    ):
+    def save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path):
         if isinstance(image, torch.Tensor):
             # Convert tensor to numpy array to minimize main process time
             image = image.cpu().numpy()
-        self.queue.put((image, fpath, compress_level))
+        self.queue.put((image, fpath))
 
     def wait_until_done(self):
         self.queue.join()

src/lerobot/datasets/lerobot_dataset.py

@@ -13,7 +13,6 @@
 # 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
@@ -540,15 +539,6 @@ 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,
@@ -722,15 +712,6 @@ 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)
@@ -849,7 +830,7 @@ 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, episodes=self.episodes)
+        hf_dataset = load_nested_dataset(self.root / "data", features=features)
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
@@ -866,8 +847,10 @@ 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
@@ -949,11 +932,7 @@ 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:
-                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"]
+                timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
                 query_timestamps[key] = torch.stack(timestamps).tolist()
             else:
                 query_timestamps[key] = [current_ts]
@@ -976,16 +955,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         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])
+                result[key] = torch.stack(self.hf_dataset[key][q_idx])
             except (KeyError, TypeError, IndexError):
-                result[key] = torch.stack(self.hf_dataset[relative_indices][key])
+                result[key] = torch.stack(self.hf_dataset[q_idx][key])
         return result
 
     def _query_videos(self, query_timestamps: dict[str, list[float]], ep_idx: int) -> dict[str, torch.Tensor]:
@@ -1081,7 +1054,6 @@ 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
@@ -1091,15 +1063,13 @@ 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, compress_level: int = 1
-    ) -> None:
+    def _save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path) -> None:
         if self.image_writer is None:
             if isinstance(image, torch.Tensor):
                 image = image.cpu().numpy()
-            write_image(image, fpath, compress_level=compress_level)
+            write_image(image, fpath)
         else:
-            self.image_writer.save_image(image=image, fpath=fpath, compress_level=compress_level)
+            self.image_writer.save_image(image=image, fpath=fpath)
 
     def add_frame(self, frame: dict) -> None:
         """
@@ -1137,19 +1107,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 )
                 if frame_index == 0:
                     img_path.parent.mkdir(parents=True, exist_ok=True)
-                compress_level = 1 if self.features[key]["dtype"] == "video" else 6
-                self._save_image(frame[key], img_path, compress_level)
+                self._save_image(frame[key], img_path)
                 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,
-        parallel_encoding: bool = True,
-    ) -> None:
+    def save_episode(self, episode_data: dict | None = None) -> None:
         """
         This will save to disk the current episode in self.episode_buffer.
 
@@ -1161,8 +1126,6 @@ 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
 
@@ -1199,40 +1162,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         use_batched_encoding = self.batch_encoding_size > 1
 
         if has_video_keys and not use_batched_encoding:
-            num_cameras = len(self.meta.video_keys)
-            if parallel_encoding and num_cameras > 1:
-                # TODO(Steven): Ideally we would like to control the number of threads per encoding such that:
-                # 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))
+            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)
@@ -1397,18 +1328,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         return metadata
 
-    def _save_episode_video(
-        self,
-        video_key: str,
-        episode_index: int,
-        temp_path: Path | None = None,
-    ) -> dict:
+    def _save_episode_video(self, video_key: str, episode_index: int) -> dict:
         # Encode episode frames into a temporary video
-        if temp_path is None:
-            ep_path = self._encode_temporary_episode_video(video_key, episode_index)
-        else:
-            ep_path = temp_path
-
+        ep_path = self._encode_temporary_episode_video(video_key, episode_index)
         ep_size_in_mb = get_file_size_in_mb(ep_path)
         ep_duration_in_s = get_video_duration_in_s(ep_path)
 
@@ -1526,7 +1448,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        return _encode_video_worker(video_key, episode_index, self.root, self.fps)
+        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
 
     @classmethod
     def create(
@@ -1572,7 +1498,6 @@ 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

@@ -28,7 +28,6 @@ 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
@@ -49,7 +48,7 @@ from lerobot.utils.utils import SuppressProgressBars, is_valid_numpy_dtype_strin
 
 DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
 DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
-DEFAULT_VIDEO_FILE_SIZE_IN_MB = 200  # Max size per file
+DEFAULT_VIDEO_FILE_SIZE_IN_MB = 500  # Max size per file
 
 INFO_PATH = "meta/info.json"
 STATS_PATH = "meta/stats.json"
@@ -104,9 +103,7 @@ 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, episodes: list[int] | None = None
-) -> Dataset:
+def load_nested_dataset(pq_dir: Path, features: datasets.Features | 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
@@ -114,26 +111,15 @@ def load_nested_dataset(
     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():
-        # 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)
+        datasets = Dataset.from_parquet([str(path) for path in paths], features=features)
+    return datasets
 
 
 def get_parquet_num_frames(parquet_path: str | Path) -> int:

src/lerobot/datasets/video_utils.py

@@ -311,7 +311,6 @@ 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
@@ -360,9 +359,6 @@ 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

@@ -21,22 +21,7 @@ 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,
-    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,
-)
+from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 
 
 @dataclass
@@ -245,7 +230,7 @@ class HILSerlRobotEnvConfig(EnvConfig):
 class LiberoEnv(EnvConfig):
     task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
     fps: int = 30
-    episode_length: int | None = None
+    episode_length: int = 520
     obs_type: str = "pixels_agent_pos"
     render_mode: str = "rgb_array"
     camera_name: str = "agentview_image,robot0_eye_in_hand_image"
@@ -261,62 +246,28 @@ class LiberoEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             ACTION: ACTION,
-            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",
+            "agent_pos": OBS_STATE,
+            "pixels/agentview_image": f"{OBS_IMAGES}.image",
+            "pixels/robot0_eye_in_hand_image": f"{OBS_IMAGES}.image2",
         }
     )
-    control_mode: str = "relative"  # or "absolute"
 
     def __post_init__(self):
         if self.obs_type == "pixels":
-            self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
+            self.features["pixels/agentview_image"] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
-            self.features[LIBERO_KEY_PIXELS_EYE_IN_HAND] = PolicyFeature(
+            self.features["pixels/robot0_eye_in_hand_image"] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
         elif self.obs_type == "pixels_agent_pos":
-            self.features[LIBERO_KEY_PIXELS_AGENTVIEW] = PolicyFeature(
+            self.features["agent_pos"] = PolicyFeature(type=FeatureType.STATE, shape=(8,))
+            self.features["pixels/agentview_image"] = PolicyFeature(
                 type=FeatureType.VISUAL, shape=(self.observation_height, self.observation_width, 3)
             )
-            self.features[LIBERO_KEY_PIXELS_EYE_IN_HAND] = PolicyFeature(
+            self.features["pixels/robot0_eye_in_hand_image"] = 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

@@ -14,18 +14,12 @@
 # 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:
@@ -39,46 +33,6 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
         raise ValueError(f"Policy type '{env_type}' is not available.")
 
 
-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:
-        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,
@@ -143,8 +97,6 @@ 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

@@ -28,6 +28,7 @@ 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]:
@@ -80,7 +81,10 @@ 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] = {
@@ -100,7 +104,6 @@ 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",
@@ -112,7 +115,6 @@ 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
@@ -140,19 +142,14 @@ 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)
-            if self.episode_length is None
-            else self.episode_length
-        )
-        self.control_mode = control_mode
+        self._max_episode_steps = TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
+
         images = {}
         for cam in self.camera_name:
             images[self.camera_name_mapping[cam]] = spaces.Box(
@@ -178,36 +175,11 @@ class LiberoEnv(gym.Env):
             self.observation_space = spaces.Dict(
                 {
                     "pixels": spaces.Dict(images),
-                    "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),
-                                }
-                            ),
-                        }
+                    "agent_pos": spaces.Box(
+                        low=AGENT_POS_LOW,
+                        high=AGENT_POS_HIGH,
+                        shape=(OBS_STATE_DIM,),
+                        dtype=np.float64,
                     ),
                 }
             )
@@ -219,7 +191,6 @@ 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):
@@ -241,48 +212,23 @@ 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
-
-        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,)
-                },
-            },
-        }
+        state = np.concatenate(
+            (
+                raw_obs["robot0_eef_pos"],
+                quat2axisangle(raw_obs["robot0_eef_quat"]),
+                raw_obs["robot0_gripper_qpos"],
+            )
+        )
+        agent_pos = state
         if self.obs_type == "pixels":
             return {"pixels": images.copy()}
-
         if self.obs_type == "pixels_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
-
+            return {
+                "pixels": images.copy(),
+                "agent_pos": agent_pos,
+            }
         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')."
@@ -300,15 +246,6 @@ 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
@@ -354,10 +291,8 @@ 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)."""
 
@@ -369,9 +304,7 @@ 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,
         )
 
@@ -391,8 +324,6 @@ 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.
@@ -424,24 +355,24 @@ 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

@@ -29,22 +29,10 @@ 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, OBS_STR
+from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
 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.
@@ -90,14 +78,12 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
 
         return_observations[OBS_ENV_STATE] = env_state
 
-    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
+    # 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 "robot_state" in observations:
-        return_observations[f"{OBS_STR}.robot_state"] = _convert_nested_dict(observations["robot_state"])
     return return_observations
 
 

src/lerobot/policies/__init__.py

@@ -21,7 +21,6 @@ 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",
@@ -32,5 +31,4 @@ __all__ = [
     "TDMPCConfig",
     "VQBeTConfig",
     "GrootConfig",
-    "XVLAConfig",
 ]

src/lerobot/policies/factory.py

@@ -40,7 +40,6 @@ 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,
@@ -108,10 +107,6 @@ 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.")
 
@@ -155,8 +150,6 @@ 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.")
 
@@ -336,15 +329,6 @@ 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.")

src/lerobot/policies/rtc/README.md

@@ -1,38 +1,49 @@
-# Real-Time Chunking (RTC)
+# Real-Time Chunking (RTC) Module
 
-This module contains the LeRobot implementation of **Real-Time Chunking (RTC)**, an inference-time technique for flow-matching based policies.
+This module implements Real-Time Chunking and related adaptive inference techniques for robotics policies in LeRobot.
 
-**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/).
+## Overview
 
----
+Real-Time Chunking (RTC) addresses the challenge of real-time inference in action chunking policies by treating chunk generation as an inpainting problem. It strategically handles overlapping timesteps between action chunks using prefix attention mechanisms.
 
-## Citation
+It is particularly effective for handling long-horizon inference in robotics policies.
 
-If you use Real-Time Chunking in your work, please cite:
+## Integration with Policies
 
-```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}
-}
+RTC can be integrated with any policy that supports flow mathicng for chunking:
 
-@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},
-}
+- **SmolVLA**: Vision-language-action model with RTC support
+- **Pi0**: Action prediction model with adaptive chunking
+- **Pi05**: Action prediction model with adaptive chunking
+
+## Original Implementation
+
+This implementation is based on Physical Intelligence's Kinetix RTC:
+
+- [Original RTC implementation](https://github.com/Physical-Intelligence/real-time-chunking-kinetix/blob/main/src/model.py#L214)
+- [Kinetix GitHub Repository](https://github.com/Physical-Intelligence/real-time-chunking-kinetix)
+
+## References
+
+- [Real Time Chunking Paper](https://www.physicalintelligence.company/research/real_time_chunking)
+- [Physical Intelligence Kinetix](https://github.com/Physical-Intelligence/real-time-chunking-kinetix)
+
+## How to run
+
+### Check with data from the dataset
+
+```bash
+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 \
+--seed=42
 ```
 
----
+This script will evaluate RTC on a data from a dataset and save the results to a file, u can check the results in the `rtc_debug_output` directory.
 
-## License
+The example output should look like this:
+![Flow Matching with RTC](./flow_matching.png)
 
-This implementation follows the **Apache 2.0 License**, consistent with the LeRobot project.
+It shows how flow matching works with RTC and without it. The chart shows values of action predictions for each timestep. The colour shows the the generation progress. The blue ones - earlier timesteps, the yellow ones - later timesteps. The red line is the ground truth (previous action chunk).

src/lerobot/policies/rtc/debug_visualizer.py

@@ -111,3 +111,7 @@ class RTCDebugVisualizer:
             if not ax.yaxis.get_label().get_text():
                 ax.set_ylabel(f"Dim {dim_idx}", fontsize=10)
             ax.grid(True, alpha=0.3)
+
+            # Add legend if label provided and this is the first dimension
+            if label and dim_idx == 0:
+                ax.legend(loc="best", fontsize=8)

src/lerobot/policies/xvla/__init__.py

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

src/lerobot/policies/xvla/action_hub.py

@@ -1,454 +0,0 @@
-# ------------------------------------------------------------------------------
-# 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, no gripper."""
-
-    dim_action = 7
-    JOINTS_SCALE = 1.0
-
-    def __init__(self):
-        super().__init__()
-        self.mse = nn.MSELoss()
-
-    def compute_loss(self, pred, target):
-        assert pred.shape == target.shape, "pred/target shapes must match"
-        joints_loss = self.mse(pred, target) * self.JOINTS_SCALE
-        return {"joints_loss": joints_loss}
-
-    def preprocess(self, proprio, action, mode="train"):
-        """No preprocessing needed for 7 joint actions."""
-        return proprio, action
-
-    def postprocess(self, action: torch.Tensor) -> torch.Tensor:
-        """Return directly (no sigmoid since no gripper)."""
-        return 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",
-    "BimanualSO101ActionSpace",
-    "ACTION_REGISTRY",
-]

src/lerobot/policies/xvla/configuration_florence2.py

@@ -1,353 +0,0 @@
-# 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

@@ -1,190 +0,0 @@
-#!/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 AdamWConfig
-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.MEAN_STD,
-        }
-    )
-
-    # 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
-    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 = True  # Freeze VLM vision encoder weights
-    freeze_language_encoder: bool = True  # 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, 0.95)
-    optimizer_eps: float = 1e-8
-    optimizer_weight_decay: float = 1e-4
-    optimizer_grad_clip_norm: float = 10.0
-
-    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) -> AdamWConfig:
-        return AdamWConfig(
-            lr=self.optimizer_lr,
-            betas=self.optimizer_betas,
-            eps=self.optimizer_eps,
-            weight_decay=self.optimizer_weight_decay,
-            grad_clip_norm=self.optimizer_grad_clip_norm,
-        )
-
-    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

@@ -1,526 +0,0 @@
-#!/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
-        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]:
-        enc = self.forward_vlm(input_ids, image_input, image_mask)
-
-        batch_size = input_ids.shape[0]
-        t = (
-            torch.rand(1, device=input_ids.device)
-            + torch.arange(batch_size, device=input_ids.device) / 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()
-        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=proprio.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=proprio.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 only trainable parameters for optimization."""
-        return filter(lambda p: p.requires_grad, self.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 _trim_action_dim(self, actions: Tensor) -> Tensor:
-        feature = self.config.action_feature
-        if feature is None:
-            return actions
-        desired_dim = self.model.dim_action
-        if desired_dim == actions.shape[-1]:
-            return actions
-        if desired_dim < actions.shape[-1]:
-            return actions[..., :desired_dim]
-        return pad_vector(actions, desired_dim)
-
-    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)
-        actions = self._trim_action_dim(actions)
-        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

@@ -1,551 +0,0 @@
-# ------------------------------------------------------------------------------
-# 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]
-
-                # Validate that values are in [0, 255] range if requested
-                if self.validate_range:
-                    min_val = tensor.min().item()
-                    max_val = tensor.max().item()
-                    if 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,
-        ),
-    )

src/lerobot/policies/xvla/soft_transformer.py

@@ -1,415 +0,0 @@
-# ------------------------------------------------------------------------------
-# Copyright 2025 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 math
-from collections.abc import Iterable
-from functools import partial
-from typing import Final
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as functional
-
-# ------------------------------- Small utils ----------------------------------
-
-
-def _to_2tuple(x) -> tuple:
-    """Minimal replacement for timm.layers.to_2tuple."""
-    if isinstance(x, Iterable) and not isinstance(x, (str, bytes)):
-        t = tuple(x)
-        return (t[0], t[1]) if len(t) >= 2 else (t[0], t[0])
-    return (x, x)
-
-
-def _has_sdp_attention() -> bool:
-    """Check if we can use PyTorch fused scaled_dot_product_attention."""
-    return hasattr(functional, "scaled_dot_product_attention")
-
-
-# ---------------------------------- MLP --------------------------------------
-
-
-class Mlp(nn.Module):
-    """
-    MLP used in ViT-style blocks.
-
-    Supports Linear or 1x1 Conv 'linear_layer' for token/channel mixing.
-    """
-
-    def __init__(
-        self,
-        in_features: int,
-        hidden_features: int | None = None,
-        out_features: int | None = None,
-        norm_layer: type[nn.Module] | None = None,
-        bias: bool | tuple[bool, bool] = True,
-        drop: float | tuple[float, float] = 0.0,
-        use_conv: bool = False,
-    ) -> None:
-        super().__init__()
-        out_features = out_features or in_features
-        hidden_features = hidden_features or in_features
-        bias = _to_2tuple(bias)
-        drop_probs = _to_2tuple(drop)
-        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
-
-        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
-        self.act = nn.GELU(approximate="tanh")
-        self.drop1 = nn.Dropout(drop_probs[0])
-        self.norm = norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
-        self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
-        self.drop2 = nn.Dropout(drop_probs[1])
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        # Expect [B, T, C] for Linear variant; caller is responsible for shapes.
-        x = self.fc1(x)
-        x = self.act(x)
-        x = self.drop1(x)
-        x = self.norm(x)
-        x = self.fc2(x)
-        x = self.drop2(x)
-        return x
-
-
-# -------------------------------- Attention ----------------------------------
-
-
-class Attention(nn.Module):
-    """
-    Multi-Head Self-Attention with optional fused SDPA fallback.
-
-    If PyTorch provides `scaled_dot_product_attention`, it will be used
-    (usually faster and more stable); otherwise we use a manual implementation.
-    """
-
-    fused_attn: Final[bool]
-
-    def __init__(
-        self,
-        dim: int,
-        num_heads: int = 8,
-        qkv_bias: bool = False,
-        qk_norm: bool = False,
-        attn_drop: float = 0.0,
-        proj_drop: float = 0.0,
-        norm_layer: type[nn.Module] = nn.LayerNorm,
-    ) -> None:
-        super().__init__()
-        assert dim % num_heads == 0, "dim should be divisible by num_heads"
-        self.num_heads = num_heads
-        self.head_dim = dim // num_heads
-        self.scale = self.head_dim**-0.5
-        self.fused_attn = _has_sdp_attention()
-
-        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
-        self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
-        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
-        self.attn_drop = nn.Dropout(attn_drop)
-        self.proj = nn.Linear(dim, dim)
-        self.proj_drop = nn.Dropout(proj_drop)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        """
-        Parameters
-        ----------
-        x : Tensor, shape [batch_size, seq_len, channels]
-            Input sequence.
-
-        Returns
-        -------
-        Tensor, shape [batch_size, seq_len, channels]
-            Output sequence after MHSA + projection.
-        """
-        batch_size, seq_len, channels = x.shape
-        qkv = (
-            self.qkv(x)
-            .reshape(batch_size, seq_len, 3, self.num_heads, self.head_dim)
-            .permute(2, 0, 3, 1, 4)  # 3 x [batch_size, num_heads, seq_len, head_dim]
-        )
-        q, k, v = qkv.unbind(0)  # each: [batch_size, num_heads, seq_len, head_dim]
-        q, k = self.q_norm(q), self.k_norm(k)
-
-        if self.fused_attn:
-            x = functional.scaled_dot_product_attention(
-                q,
-                k,
-                v,
-                dropout_p=self.attn_drop.p if self.training else 0.0,
-            )  # [batch_size, num_heads, seq_len, head_dim]
-        else:
-            q = q * self.scale
-            attn = q @ k.transpose(-2, -1)  # [batch_size, num_heads, seq_len, seq_len]
-            attn = attn.softmax(dim=-1)
-            attn = self.attn_drop(attn)
-            x = attn @ v  # [batch_size, num_heads, seq_len, head_dim]
-
-        x = x.transpose(1, 2).reshape(batch_size, seq_len, channels)  # [batch_size, seq_len, channels]
-        x = self.proj(x)
-        x = self.proj_drop(x)
-        return x
-
-
-# ------------------------------- Utilities -----------------------------------
-
-
-def basic_init(module: nn.Module) -> None:
-    """
-    Apply a basic initialization scheme to Linear layers.
-
-    - Weight: Xavier uniform initialization.
-    - Bias: Set to zero.
-    """
-    if isinstance(module, nn.Linear):
-        nn.init.xavier_uniform_(module.weight)
-        if module.bias is not None:
-            nn.init.constant_(module.bias, 0.0)
-
-
-def timestep_embedding(t: torch.Tensor, dim: int, max_period: int = 100) -> torch.Tensor:
-    """
-    Create sinusoidal timestep embeddings.
-
-    Parameters
-    ----------
-    t : torch.Tensor
-        Shape [B]. Each element is a timestep index, may be fractional.
-    dim : int
-        Dimensionality of the output embedding.
-    max_period : int, default=100
-        Controls the minimum frequency of the sinusoids.
-
-    Returns
-    -------
-    torch.Tensor
-        Shape [B, dim]. Sinusoidal embeddings.
-    """
-    half = dim // 2
-    freqs = torch.exp(
-        -math.log(max_period) * torch.arange(start=0, end=half, dtype=t.dtype, device=t.device) / half
-    )
-    args = t[:, None] * freqs[None]
-    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
-    if dim % 2 == 1:
-        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
-    return embedding
-
-
-# ------------------------------- Core Layers ----------------------------------
-
-
-class DomainAwareLinear(nn.Module):
-    """
-    Linear layer with domain-conditioned parameters (per-sample).
-
-    Each domain has its own weight and bias vectors, stored in embeddings.
-    """
-
-    def __init__(self, input_size: int, output_size: int, num_domains: int = 20) -> None:
-        super().__init__()
-        self.input_size = input_size
-        self.output_size = output_size
-        self.fc = nn.Embedding(num_domains, output_size * input_size)
-        self.bias = nn.Embedding(num_domains, output_size)
-        nn.init.xavier_uniform_(self.fc.weight)
-        nn.init.zeros_(self.bias.weight)
-
-    def forward(self, x: torch.Tensor, domain_id: torch.LongTensor) -> torch.Tensor:
-        """
-        Parameters
-        ----------
-        x : Tensor
-            [B, I] or [B, T, I]
-        domain_id : LongTensor
-            [B], domain indices.
-
-        Returns
-        -------
-        Tensor
-            [batch_size, output_size] or [batch_size, seq_len, output_size]
-        """
-        batch_size = domain_id.shape[0]
-        squeeze_seq = False
-        if x.dim() == 2:
-            x = x.unsqueeze(1)
-            squeeze_seq = True
-        weight = self.fc(domain_id).view(batch_size, self.input_size, self.output_size)
-        bias = self.bias(domain_id).view(batch_size, self.output_size)
-        y = torch.matmul(x, weight) + bias.view(batch_size, 1, self.output_size)
-        if squeeze_seq:
-            y = y.squeeze(1)
-        return y
-
-
-class TransformerBlock(nn.Module):
-    """
-    Standard Transformer block (pre-LN): LN → MHSA → residual, LN → MLP → residual.
-    """
-
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float = 4.0) -> None:
-        super().__init__()
-        self.norm1 = nn.LayerNorm(hidden_size)
-        self.norm2 = nn.LayerNorm(hidden_size)
-        self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, attn_drop=0.1)
-        self.mlp = Mlp(
-            in_features=hidden_size,
-            hidden_features=int(hidden_size * mlp_ratio),
-            drop=0.1,
-        )
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        """
-        Parameters
-        ----------
-        x : Tensor, [B, T, H]
-
-        Returns
-        -------
-        Tensor, [B, T, H]
-        """
-        x = x + self.attn(self.norm1(x))
-        x = x + self.mlp(self.norm2(x))
-        return x
-
-
-# --------------------------- Main Model ---------------------------------------
-
-
-class SoftPromptedTransformer(nn.Module):
-    """
-    Multi-modal, domain-aware Transformer with optional soft prompts.
-
-    See parameter and forward I/O descriptions inside the docstrings.
-    """
-
-    def __init__(
-        self,
-        hidden_size: int = 768,
-        multi_modal_input_size: int = 768,
-        depth: int = 24,
-        num_heads: int = 16,
-        mlp_ratio: float = 4.0,
-        num_domains: int = 20,
-        dim_action: int = 20,
-        dim_propio: int = 20,
-        dim_time: int = 32,
-        len_soft_prompts: int = 32,
-        max_len_seq: int = 512,
-        use_hetero_proj: bool = False,
-    ) -> None:
-        super().__init__()
-        self.hidden_size = hidden_size
-        self.dim_action = dim_action
-        self.dim_time = dim_time
-        self.len_soft_prompts = len_soft_prompts
-        self.use_hetero_proj = use_hetero_proj
-
-        self.blocks = nn.ModuleList(
-            [TransformerBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio) for _ in range(depth)]
-        )
-
-        if use_hetero_proj:
-            self.vlm_proj = DomainAwareLinear(multi_modal_input_size, hidden_size, num_domains=num_domains)
-            self.aux_visual_proj = DomainAwareLinear(
-                multi_modal_input_size, hidden_size, num_domains=num_domains
-            )
-        else:
-            self.vlm_proj = nn.Linear(multi_modal_input_size, hidden_size)
-            self.aux_visual_proj = nn.Linear(multi_modal_input_size, hidden_size)
-
-        self.pos_emb = nn.Parameter(torch.zeros(1, max_len_seq, hidden_size), requires_grad=True)
-        nn.init.normal_(self.pos_emb, std=0.02)
-
-        self.norm = nn.LayerNorm(hidden_size)
-        self.action_encoder = DomainAwareLinear(
-            dim_action + dim_time + dim_propio, hidden_size, num_domains=num_domains
-        )
-        self.action_decoder = DomainAwareLinear(hidden_size, dim_action, num_domains=num_domains)
-
-        if len_soft_prompts > 0:
-            self.soft_prompt_hub = nn.Embedding(num_domains, len_soft_prompts * hidden_size)
-            nn.init.normal_(self.soft_prompt_hub.weight, std=0.02)
-
-        self.apply(basic_init)
-
-    def forward(
-        self,
-        domain_id: torch.LongTensor,
-        vlm_features: torch.Tensor,
-        aux_visual_inputs: torch.Tensor,
-        action_with_noise: torch.Tensor,
-        proprio: torch.Tensor,
-        t: torch.Tensor,
-    ) -> torch.Tensor:
-        """
-        Forward pass.
-
-        Inputs
-        ------
-        domain_id : [B]
-        vlm_features : [B, T_vlm, D]
-        aux_visual_inputs : [B, T_aux, D]
-        action_with_noise : [B, T_action, dim_action]
-        proprio : [B, dim_propio]
-        t : [B]
-
-        Returns
-        -------
-        Tensor
-            Predicted actions, [batch_size, num_actions, dim_action]
-        """
-        batch_size, num_actions = action_with_noise.shape[:2]
-
-        # Encode (action + proprio + time) → tokens
-        time_emb = timestep_embedding(t, self.dim_time)  # [batch_size, dim_time]
-        time_tokens = time_emb.unsqueeze(1).expand(batch_size, num_actions, self.dim_time)
-        proprio_tokens = proprio.unsqueeze(1).expand(batch_size, num_actions, proprio.shape[-1])
-        action_tokens = torch.cat([action_with_noise, proprio_tokens, time_tokens], dim=-1)
-        x = self.action_encoder(action_tokens, domain_id)  # [batch_size, num_actions, hidden_size]
-
-        # Project visual streams and concatenate
-        if self.use_hetero_proj:
-            x = torch.cat(
-                [
-                    x,
-                    self.vlm_proj(vlm_features, domain_id),
-                    self.aux_visual_proj(aux_visual_inputs, domain_id),
-                ],
-                dim=1,
-            )
-        else:
-            x = torch.cat([x, self.vlm_proj(vlm_features), self.aux_visual_proj(aux_visual_inputs)], dim=1)
-
-        # Add positional embeddings (truncate if needed)
-        seq_len = x.shape[1]
-        if seq_len > self.pos_emb.shape[1]:
-            raise ValueError(f"Sequence length {seq_len} exceeds max_len_seq={self.pos_emb.shape[1]}.")
-        x = x + self.pos_emb[:, :seq_len, :]
-
-        # Append soft prompts
-        if self.len_soft_prompts > 0:
-            soft_prompts = self.soft_prompt_hub(domain_id).view(
-                batch_size, self.len_soft_prompts, self.hidden_size
-            )
-            x = torch.cat([x, soft_prompts], dim=1)
-
-        # Transformer backbone
-        for block in self.blocks:
-            x = block(x)
-
-        # Decode only the action segment
-        return self.action_decoder(self.norm(x[:, :num_actions]), domain_id)

src/lerobot/policies/xvla/utils.py

@@ -1,138 +0,0 @@
-import math
-
-import numpy as np
-
-
-def mat2quat(rmat):
-    """
-    Converts given rotation matrix to quaternion.
-
-    Args:
-        rmat (np.array): 3x3 rotation matrix
-
-    Returns:
-        np.array: (x,y,z,w) float quaternion angles
-    """
-    mat = np.asarray(rmat).astype(np.float32)[:3, :3]
-
-    m00 = mat[0, 0]
-    m01 = mat[0, 1]
-    m02 = mat[0, 2]
-    m10 = mat[1, 0]
-    m11 = mat[1, 1]
-    m12 = mat[1, 2]
-    m20 = mat[2, 0]
-    m21 = mat[2, 1]
-    m22 = mat[2, 2]
-    # symmetric matrix k
-    k = np.array(
-        [
-            [m00 - m11 - m22, np.float32(0.0), np.float32(0.0), np.float32(0.0)],
-            [m01 + m10, m11 - m00 - m22, np.float32(0.0), np.float32(0.0)],
-            [m02 + m20, m12 + m21, m22 - m00 - m11, np.float32(0.0)],
-            [m21 - m12, m02 - m20, m10 - m01, m00 + m11 + m22],
-        ]
-    )
-    k /= 3.0
-    # quaternion is Eigen vector of k that corresponds to largest eigenvalue
-    w, v = np.linalg.eigh(k)
-    inds = np.array([3, 0, 1, 2])
-    q1 = v[inds, np.argmax(w)]
-    if q1[0] < 0.0:
-        np.negative(q1, q1)
-    inds = np.array([1, 2, 3, 0])
-    return q1[inds]
-
-
-def quat2axisangle(quat):
-    """
-    Converts quaternion to axis-angle format.
-    Returns a unit vector direction scaled by its angle in radians.
-
-    Args:
-        quat (np.array): (x,y,z,w) vec4 float angles
-
-    Returns:
-        np.array: (ax,ay,az) axis-angle exponential coordinates
-    """
-    # clip quaternion
-    if quat[3] > 1.0:
-        quat[3] = 1.0
-    elif quat[3] < -1.0:
-        quat[3] = -1.0
-
-    den = np.sqrt(1.0 - quat[3] * quat[3])
-    if math.isclose(den, 0.0):
-        # This is (close to) a zero degree rotation, immediately return
-        return np.zeros(3)
-
-    return (quat[:3] * 2.0 * math.acos(quat[3])) / den
-
-
-def rotate6d_to_axis_angle(r6d):
-    """
-    r6d: np.ndarray, shape (N, 6)
-    return: np.ndarray, shape (N, 3), axis-angle vectors
-    """
-    flag = 0
-    if len(r6d.shape) == 1:
-        r6d = r6d[None, ...]
-        flag = 1
-
-    a1 = r6d[:, 0:3]
-    a2 = r6d[:, 3:6]
-
-    # b1
-    b1 = a1 / (np.linalg.norm(a1, axis=-1, keepdims=True) + 1e-6)
-
-    # b2
-    dot_prod = np.sum(b1 * a2, axis=-1, keepdims=True)
-    b2_orth = a2 - dot_prod * b1
-    b2 = b2_orth / (np.linalg.norm(b2_orth, axis=-1, keepdims=True) + 1e-6)
-
-    # b3
-    b3 = np.cross(b1, b2, axis=-1)
-
-    rotation_matrix = np.stack([b1, b2, b3], axis=-1)  # shape: (N, 3, 3)
-
-    axis_angle_list = []
-    for i in range(rotation_matrix.shape[0]):
-        quat = mat2quat(rotation_matrix[i])
-        axis_angle = quat2axisangle(quat)
-        axis_angle_list.append(axis_angle)
-
-    axis_angle_array = np.stack(axis_angle_list, axis=0)  # shape: (N, 3)
-
-    if flag == 1:
-        axis_angle_array = axis_angle_array[0]
-
-    return axis_angle_array
-
-
-def mat_to_rotate6d(abs_action):
-    if len(abs_action.shape) == 2:
-        return np.concatenate([abs_action[:3, 0], abs_action[:3, 1]], axis=-1)
-    elif len(abs_action.shape) == 3:
-        return np.concatenate([abs_action[:, :3, 0], abs_action[:, :3, 1]], axis=-1)
-    else:
-        raise NotImplementedError
-
-
-def drop_path(x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True):
-    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
-
-    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
-    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
-    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
-    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
-    'survival rate' as the argument.
-
-    """
-    if drop_prob == 0.0 or not training:
-        return x
-    keep_prob = 1 - drop_prob
-    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
-    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
-    if keep_prob > 0.0 and scale_by_keep:
-        random_tensor.div_(keep_prob)
-    return x * random_tensor

src/lerobot/processor/env_processor.py

@@ -1,154 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-from dataclasses import dataclass
-
-import torch
-
-from lerobot.configs.types import PipelineFeatureType, PolicyFeature
-from lerobot.utils.constants import OBS_IMAGES, OBS_STATE
-
-from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
-
-
-@dataclass
-@ProcessorStepRegistry.register(name="libero_processor")
-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]
-
-                # 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_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 a single state vector
-            state = torch.cat((eef_pos, eef_axisangle, gripper_qpos), dim=-1)
-
-            # 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=(8,),  # [eef_pos(3), axis_angle(3), gripper(2)]
-            dtype="float32",
-            description=("Concatenated end-effector position (3), axis-angle (3), and gripper qpos (2)."),
-        )
-
-        new_features[PipelineFeatureType.STATE] = state_feats
-
-        return new_features
-
-    def observation(self, observation):
-        return self._process_observation(observation)
-
-    def _quat2axisangle(self, quat: torch.Tensor) -> torch.Tensor:
-        """
-        Convert batched quaternions to axis-angle format.
-        Only accepts torch tensors of shape (B, 4).
-
-        Args:
-            quat (Tensor): (B, 4) tensor of quaternions in (x, y, z, w) format
-
-        Returns:
-            Tensor: (B, 3) axis-angle vectors
-
-        Raises:
-            TypeError: if input is not a torch tensor
-            ValueError: if shape is not (B, 4)
-        """
-
-        if not isinstance(quat, torch.Tensor):
-            raise TypeError(f"_quat2axisangle expected a torch.Tensor, got {type(quat)}")
-
-        if quat.ndim != 2 or quat.shape[1] != 4:
-            raise ValueError(f"_quat2axisangle expected shape (B, 4), got {tuple(quat.shape)}")
-
-        quat = quat.to(dtype=torch.float32)
-        device = quat.device
-        batch_size = quat.shape[0]
-
-        w = quat[:, 3].clamp(-1.0, 1.0)
-
-        den = torch.sqrt(torch.clamp(1.0 - w * w, min=0.0))
-
-        result = torch.zeros((batch_size, 3), device=device)
-
-        mask = den > 1e-10
-
-        if mask.any():
-            angle = 2.0 * torch.acos(w[mask])  # (M,)
-            axis = quat[mask, :3] / den[mask].unsqueeze(1)
-            result[mask] = axis * angle.unsqueeze(1)
-
-        return result

src/lerobot/rl/actor.py

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

src/lerobot/rl/gym_manipulator.py

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

src/lerobot/scripts/lerobot_dataset_viz.py

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

src/lerobot/scripts/lerobot_eval.py

@@ -71,7 +71,7 @@ from tqdm import trange
 
 from lerobot.configs import parser
 from lerobot.configs.eval import EvalPipelineConfig
-from lerobot.envs.factory import make_env, make_env_pre_post_processors
+from lerobot.envs.factory import make_env
 from lerobot.envs.utils import (
     add_envs_task,
     check_env_attributes_and_types,
@@ -94,8 +94,6 @@ from lerobot.utils.utils import (
 def rollout(
     env: gym.vector.VectorEnv,
     policy: PreTrainedPolicy,
-    env_preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
-    env_postprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     postprocessor: PolicyProcessorPipeline[PolicyAction, PolicyAction],
     seeds: list[int] | None = None,
@@ -167,19 +165,11 @@ def rollout(
         # Infer "task" from attributes of environments.
         # TODO: works with SyncVectorEnv but not AsyncVectorEnv
         observation = add_envs_task(env, observation)
-
-        # Apply environment-specific preprocessing (e.g., LiberoProcessorStep for LIBERO)
-        observation = env_preprocessor(observation)
-
         observation = preprocessor(observation)
         with torch.inference_mode():
             action = policy.select_action(observation)
         action = postprocessor(action)
 
-        action_transition = {"action": action}
-        action_transition = env_postprocessor(action_transition)
-        action = action_transition["action"]
-
         # Convert to CPU / numpy.
         action_numpy: np.ndarray = action.to("cpu").numpy()
         assert action_numpy.ndim == 2, "Action dimensions should be (batch, action_dim)"
@@ -249,8 +239,6 @@ def rollout(
 def eval_policy(
     env: gym.vector.VectorEnv,
     policy: PreTrainedPolicy,
-    env_preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
-    env_postprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     postprocessor: PolicyProcessorPipeline[PolicyAction, PolicyAction],
     n_episodes: int,
@@ -331,8 +319,6 @@ def eval_policy(
         rollout_data = rollout(
             env=env,
             policy=policy,
-            env_preprocessor=env_preprocessor,
-            env_postprocessor=env_postprocessor,
             preprocessor=preprocessor,
             postprocessor=postprocessor,
             seeds=list(seeds) if seeds else None,
@@ -531,16 +517,10 @@ def eval_main(cfg: EvalPipelineConfig):
         pretrained_path=cfg.policy.pretrained_path,
         preprocessor_overrides=preprocessor_overrides,
     )
-
-    # Create environment-specific preprocessor and postprocessor (e.g., for LIBERO environments)
-    env_preprocessor, env_postprocessor = make_env_pre_post_processors(env_cfg=cfg.env, policy_cfg=cfg.policy)
-
     with torch.no_grad(), torch.autocast(device_type=device.type) if cfg.policy.use_amp else nullcontext():
         info = eval_policy_all(
             envs=envs,
             policy=policy,
-            env_preprocessor=env_preprocessor,
-            env_postprocessor=env_postprocessor,
             preprocessor=preprocessor,
             postprocessor=postprocessor,
             n_episodes=cfg.eval.n_episodes,
@@ -581,8 +561,6 @@ def eval_one(
     env: gym.vector.VectorEnv,
     *,
     policy: PreTrainedPolicy,
-    env_preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
-    env_postprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     postprocessor: PolicyProcessorPipeline[PolicyAction, PolicyAction],
     n_episodes: int,
@@ -598,8 +576,6 @@ def eval_one(
     task_result = eval_policy(
         env=env,
         policy=policy,
-        env_preprocessor=env_preprocessor,
-        env_postprocessor=env_postprocessor,
         preprocessor=preprocessor,
         postprocessor=postprocessor,
         n_episodes=n_episodes,
@@ -624,8 +600,6 @@ def run_one(
     env,
     *,
     policy,
-    env_preprocessor,
-    env_postprocessor,
     preprocessor,
     postprocessor,
     n_episodes: int,
@@ -648,8 +622,6 @@ def run_one(
     metrics = eval_one(
         env,
         policy=policy,
-        env_preprocessor=env_preprocessor,
-        env_postprocessor=env_postprocessor,
         preprocessor=preprocessor,
         postprocessor=postprocessor,
         n_episodes=n_episodes,
@@ -667,8 +639,6 @@ def run_one(
 def eval_policy_all(
     envs: dict[str, dict[int, gym.vector.VectorEnv]],
     policy,
-    env_preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
-    env_postprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     preprocessor: PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     postprocessor: PolicyProcessorPipeline[PolicyAction, PolicyAction],
     n_episodes: int,
@@ -724,8 +694,6 @@ def eval_policy_all(
     task_runner = partial(
         run_one,
         policy=policy,
-        env_preprocessor=env_preprocessor,
-        env_postprocessor=env_postprocessor,
         preprocessor=preprocessor,
         postprocessor=postprocessor,
         n_episodes=n_episodes,

src/lerobot/scripts/lerobot_find_joint_limits.py

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

src/lerobot/scripts/lerobot_record.py

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

src/lerobot/scripts/lerobot_replay.py

@@ -62,7 +62,7 @@ from lerobot.robots import (  # noqa: F401
 )
 from lerobot.utils.constants import ACTION
 from lerobot.utils.import_utils import register_third_party_devices
-from lerobot.utils.robot_utils import precise_sleep
+from lerobot.utils.robot_utils import busy_wait
 from lerobot.utils.utils import (
     init_logging,
     log_say,
@@ -121,7 +121,7 @@ def replay(cfg: ReplayConfig):
         _ = robot.send_action(processed_action)
 
         dt_s = time.perf_counter() - start_episode_t
-        precise_sleep(1 / dataset.fps - dt_s)
+        busy_wait(1 / dataset.fps - dt_s)
 
     robot.disconnect()
 

src/lerobot/scripts/lerobot_teleoperate.py

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

src/lerobot/scripts/lerobot_train.py

@@ -29,7 +29,7 @@ from lerobot.configs.train import TrainPipelineConfig
 from lerobot.datasets.factory import make_dataset
 from lerobot.datasets.sampler import EpisodeAwareSampler
 from lerobot.datasets.utils import cycle
-from lerobot.envs.factory import make_env, make_env_pre_post_processors
+from lerobot.envs.factory import make_env
 from lerobot.envs.utils import close_envs
 from lerobot.optim.factory import make_optimizer_and_scheduler
 from lerobot.policies.factory import make_policy, make_pre_post_processors
@@ -259,10 +259,6 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         logging.info(colored("Output dir:", "yellow", attrs=["bold"]) + f" {cfg.output_dir}")
         if cfg.env is not None:
             logging.info(f"{cfg.env.task=}")
-            logging.info("Creating environment processors")
-            env_preprocessor, env_postprocessor = make_env_pre_post_processors(
-                env_cfg=cfg.env, policy_cfg=cfg.policy
-            )
         logging.info(f"{cfg.steps=} ({format_big_number(cfg.steps)})")
         logging.info(f"{dataset.num_frames=} ({format_big_number(dataset.num_frames)})")
         logging.info(f"{dataset.num_episodes=}")
@@ -278,7 +274,6 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         sampler = EpisodeAwareSampler(
             dataset.meta.episodes["dataset_from_index"],
             dataset.meta.episodes["dataset_to_index"],
-            episode_indices_to_use=dataset.episodes,
             drop_n_last_frames=cfg.policy.drop_n_last_frames,
             shuffle=True,
         )
@@ -389,8 +384,6 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
                     eval_info = eval_policy_all(
                         envs=eval_env,  # dict[suite][task_id] -> vec_env
                         policy=accelerator.unwrap_model(policy),
-                        env_preprocessor=env_preprocessor,
-                        env_postprocessor=env_postprocessor,
                         preprocessor=preprocessor,
                         postprocessor=postprocessor,
                         n_episodes=cfg.eval.n_episodes,

src/lerobot/utils/constants.py

@@ -70,15 +70,3 @@ LOOKAHEAD_BACKTRACKTABLE = 100
 
 # openpi
 OPENPI_ATTENTION_MASK_VALUE = -2.3819763e38  # TODO(pepijn): Modify this when extending support to fp8 models
-
-# Constants for LIBERO observation keys
-LIBERO_KEY_EEF_POS = "robot_state/eef/pos"
-LIBERO_KEY_EEF_QUAT = "robot_state/eef/quat"
-LIBERO_KEY_EEF_MAT = "robot_state/eef/mat"
-LIBERO_KEY_EEF_AXISANGLE = "robot_state/eef/axisangle"
-LIBERO_KEY_GRIPPER_QPOS = "robot_state/gripper/qpos"
-LIBERO_KEY_GRIPPER_QVEL = "robot_state/gripper/qvel"
-LIBERO_KEY_JOINTS_POS = "robot_state/joints/pos"
-LIBERO_KEY_JOINTS_VEL = "robot_state/joints/vel"
-LIBERO_KEY_PIXELS_AGENTVIEW = "pixels/agentview_image"
-LIBERO_KEY_PIXELS_EYE_IN_HAND = "pixels/robot0_eye_in_hand_image"

src/lerobot/utils/profiling.py

@@ -0,0 +1,206 @@
+"""
+Profiling utilities for performance analysis.
+
+Usage:
+    from lerobot.utils.profiling import profile_method, get_profiling_stats, print_profiling_summary
+
+    @profile_method
+    def my_slow_function(x):
+        return x * 2
+
+    # At end of execution:
+    print_profiling_summary()
+"""
+
+import functools
+import logging
+import time
+from collections import defaultdict
+from threading import Lock
+from typing import Any, Callable
+
+logger = logging.getLogger(__name__)
+
+# Global profiling statistics storage
+_profiling_stats: dict[str, list[float]] = defaultdict(list)
+_profiling_lock = Lock()
+_profiling_enabled = False
+
+
+def enable_profiling():
+    """Enable profiling globally."""
+    global _profiling_enabled
+    _profiling_enabled = True
+    logger.info("Profiling enabled")
+
+
+def disable_profiling():
+    """Disable profiling globally."""
+    global _profiling_enabled
+    _profiling_enabled = False
+    logger.info("Profiling disabled")
+
+
+def is_profiling_enabled() -> bool:
+    """Check if profiling is enabled."""
+    return _profiling_enabled
+
+
+def record_timing(name: str, duration: float):
+    """Record a timing measurement.
+
+    Args:
+        name: Name/identifier for this timing
+        duration: Duration in seconds
+    """
+    if not _profiling_enabled:
+        return
+
+    with _profiling_lock:
+        _profiling_stats[name].append(duration)
+
+
+def profile_method(func: Callable) -> Callable:
+    """Decorator to profile a method or function.
+
+    Args:
+        func: Function to profile
+
+    Returns:
+        Wrapped function that records execution time
+    """
+
+    @functools.wraps(func)
+    def wrapper(*args, **kwargs) -> Any:
+        if not _profiling_enabled:
+            return func(*args, **kwargs)
+
+        start = time.perf_counter()
+        try:
+            result = func(*args, **kwargs)
+            return result
+        finally:
+            duration = time.perf_counter() - start
+            # Use fully qualified name
+            name = f"{func.__module__}.{func.__qualname__}"
+            record_timing(name, duration)
+
+    return wrapper
+
+
+class ProfileContext:
+    """Context manager for profiling code blocks.
+
+    Usage:
+        with ProfileContext("my_operation"):
+            # ... code to profile ...
+    """
+
+    def __init__(self, name: str):
+        self.name = name
+        self.start = None
+
+    def __enter__(self):
+        if _profiling_enabled:
+            self.start = time.perf_counter()
+        return self
+
+    def __exit__(self, *args):
+        if _profiling_enabled and self.start is not None:
+            duration = time.perf_counter() - self.start
+            record_timing(self.name, duration)
+
+
+def get_profiling_stats() -> dict[str, dict[str, float]]:
+    """Get summary statistics for all profiled functions.
+
+    Returns:
+        Dictionary mapping function names to their stats (count, mean, min, max, total)
+    """
+    with _profiling_lock:
+        summary = {}
+        for name, times in _profiling_stats.items():
+            if times:
+                summary[name] = {
+                    "count": len(times),
+                    "mean": sum(times) / len(times),
+                    "min": min(times),
+                    "max": max(times),
+                    "total": sum(times),
+                    "mean_ms": (sum(times) / len(times)) * 1000,
+                    "min_ms": min(times) * 1000,
+                    "max_ms": max(times) * 1000,
+                }
+        return summary
+
+
+def clear_profiling_stats():
+    """Clear all profiling statistics."""
+    with _profiling_lock:
+        _profiling_stats.clear()
+    logger.info("Profiling stats cleared")
+
+
+def print_profiling_summary(sort_by: str = "total"):
+    """Print formatted summary of profiling statistics.
+
+    Args:
+        sort_by: Sort key ('total', 'mean', 'count', 'max')
+    """
+    summary = get_profiling_stats()
+
+    if not summary:
+        logger.info("No profiling data available")
+        return
+
+    logger.info("\n" + "=" * 100)
+    logger.info("PROFILING SUMMARY")
+    logger.info("=" * 100)
+
+    # Sort by requested key
+    sorted_items = sorted(summary.items(), key=lambda x: x[1].get(sort_by, 0), reverse=True)
+
+    # Print header
+    logger.info(
+        f"{'Function':<60} {'Count':>8} {'Mean (ms)':>12} {'Min (ms)':>12} {'Max (ms)':>12} {'Total (s)':>12}"
+    )
+    logger.info("-" * 100)
+
+    # Print each function's stats
+    for name, stats in sorted_items:
+        # Shorten long names
+        display_name = name if len(name) <= 60 else "..." + name[-57:]
+
+        logger.info(
+            f"{display_name:<60} "
+            f"{stats['count']:>8} "
+            f"{stats['mean_ms']:>12.2f} "
+            f"{stats['min_ms']:>12.2f} "
+            f"{stats['max_ms']:>12.2f} "
+            f"{stats['total']:>12.2f}"
+        )
+
+    logger.info("=" * 100)
+
+    # Print summary
+    total_time = sum(s["total"] for s in summary.values())
+    total_calls = sum(s["count"] for s in summary.values())
+    logger.info(f"\nTotal profiled time: {total_time:.2f}s across {total_calls} calls")
+    logger.info("=" * 100 + "\n")
+
+
+def profile_section(name: str):
+    """Return a context manager for profiling a code section.
+
+    Args:
+        name: Name for this section
+
+    Returns:
+        ProfileContext instance
+
+    Usage:
+        with profile_section("data_loading"):
+            data = load_data()
+    """
+    return ProfileContext(name)
+

src/lerobot/utils/robot_utils.py

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

tests/policies/smolvla/test_smolvla_rtc.py

@@ -23,15 +23,13 @@ from lerobot.configs.types import FeatureType, PolicyFeature, RTCAttentionSchedu
 from lerobot.policies.factory import make_pre_post_processors  # noqa: E402
 from lerobot.policies.rtc.configuration_rtc import RTCConfig  # noqa: E402
 from lerobot.policies.smolvla.configuration_smolvla import SmolVLAConfig  # noqa: F401
+from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy  # noqa: F401
 from lerobot.utils.random_utils import set_seed  # noqa: E402
-from tests.utils import require_cuda, require_package  # noqa: E402
+from tests.utils import require_cuda  # noqa: E402
 
 
-@require_package("transformers")
 @require_cuda
 def test_smolvla_rtc_initialization():
-    from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy  # noqa: F401
-
     """Test SmolVLA policy can initialize RTC processor."""
     set_seed(42)
 
@@ -65,11 +63,8 @@ def test_smolvla_rtc_initialization():
     print("✓ SmolVLA RTC initialization: Test passed")
 
 
-@require_package("transformers")
 @require_cuda
 def test_smolvla_rtc_initialization_without_rtc_config():
-    from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy  # noqa: F401
-
     """Test SmolVLA policy can initialize without RTC config."""
     set_seed(42)
 
@@ -87,12 +82,9 @@ def test_smolvla_rtc_initialization_without_rtc_config():
     print("✓ SmolVLA RTC initialization without RTC config: Test passed")
 
 
-@require_package("transformers")
 @require_cuda
 @pytest.mark.skipif(True, reason="Requires pretrained SmolVLA model weights")
 def test_smolvla_rtc_inference_with_prev_chunk():
-    from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy  # noqa: F401
-
     """Test SmolVLA policy inference with RTC and previous chunk."""
     set_seed(42)
 
@@ -170,12 +162,9 @@ def test_smolvla_rtc_inference_with_prev_chunk():
     print("✓ SmolVLA RTC inference with prev_chunk: Test passed")
 
 
-@require_package("transformers")
 @require_cuda
 @pytest.mark.skipif(True, reason="Requires pretrained SmolVLA model weights")
 def test_smolvla_rtc_inference_without_prev_chunk():
-    from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy  # noqa: F401
-
     """Test SmolVLA policy inference with RTC but no previous chunk (RTC should have no effect)."""
     set_seed(42)
 
@@ -244,12 +233,9 @@ def test_smolvla_rtc_inference_without_prev_chunk():
     print("✓ SmolVLA RTC inference without prev_chunk: Test passed")
 
 
-@require_package("transformers")
 @require_cuda
 @pytest.mark.skipif(True, reason="Requires pretrained SmolVLA model weights")
 def test_smolvla_rtc_validation_rules():
-    from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy  # noqa: F401
-
     """Test SmolVLA policy with RTC follows all three validation rules."""
     set_seed(42)
 

tests/policies/xvla/test_xvla_original_vs_lerobot.py

@@ -1,361 +0,0 @@
-#!/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.
-
-"""Test script to verify XVLA policy integration with LeRobot vs the original implementation, only meant to be run locally!"""
-# ruff: noqa: E402
-
-import gc
-import random
-from copy import deepcopy
-from typing import Any
-
-import numpy as np
-import pytest
-import torch
-
-pytest.importorskip("transformers")
-
-from lerobot.policies.xvla.configuration_xvla import XVLAConfig
-from lerobot.policies.xvla.modeling_xvla import XVLAPolicy
-from lerobot.policies.xvla.processor_xvla import make_xvla_pre_post_processors
-from lerobot.processor import PolicyAction, PolicyProcessorPipeline  # noqa: E402
-from lerobot.utils.constants import OBS_IMAGES, OBS_STATE  # noqa: E402
-from tests.utils import require_cuda  # noqa: E402
-
-# Constants
-DUMMY_ACTION_DIM = 7  # Standard robot arm action dimension
-DUMMY_STATE_DIM = 20  # Proprioceptive state dimension
-IMAGE_HEIGHT = 224
-IMAGE_WIDTH = 224
-NUM_VIEWS = 2  # Number of camera views
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-MODEL_PATH_LEROBOT = "lerobot/xvla-widowx"
-LIBERO_DOMAIN_ID = 0  # Domain ID for examples purposes
-
-# Expected values from original XVLA implementation (reference values)
-EXPECTED_ACTIONS_SHAPE = (30, 20)
-EXPECTED_ACTIONS_MEAN = 0.117606
-EXPECTED_ACTIONS_STD = 0.245411
-EXPECTED_ACTIONS_FIRST_5 = torch.tensor([0.2742, 0.4977, 0.0500, 0.7040, -0.2653])
-
-
-def cleanup_memory():
-    """Clean up GPU/MPS memory to prevent OOM errors between tests."""
-    print("\nCleaning up memory...")
-    gc.collect()
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
-        torch.cuda.synchronize()
-    if torch.backends.mps.is_available():
-        torch.mps.empty_cache()
-    print("Memory cleanup complete.")
-
-
-def set_seed_all(seed: int):
-    """Set random seed for all RNG sources to ensure 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)
-
-    # Set deterministic behavior
-    torch.backends.cudnn.deterministic = True
-    torch.backends.cudnn.benchmark = False
-    torch.use_deterministic_algorithms(True, warn_only=True)
-
-
-def instantiate_lerobot_xvla(
-    from_pretrained: bool = False,
-    model_path: str = MODEL_PATH_LEROBOT,
-) -> tuple[
-    Any,  # Policy
-    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
-    PolicyProcessorPipeline[PolicyAction, PolicyAction],
-]:
-    """Instantiate LeRobot XVLA policy with preprocessor and postprocessor."""
-    if from_pretrained:
-        policy = XVLAPolicy.from_pretrained(
-            pretrained_name_or_path=model_path,
-            strict=False,
-        )
-    else:
-        config = XVLAConfig(
-            base_model_path=model_path,
-            n_action_steps=DUMMY_ACTION_DIM,
-            chunk_size=DUMMY_ACTION_DIM,
-            device=DEVICE,
-            num_image_views=NUM_VIEWS,
-        )  # add resize_imgs_with_padding=IMAGE_SIZE, IMAGE_SIZE?
-        policy = XVLAPolicy(config)
-
-    policy.to(DEVICE)
-    policy.config.device = DEVICE
-    preprocessor, postprocessor = make_xvla_pre_post_processors(
-        config=policy.config,
-        dataset_stats=None,  # Pass None for dataset_stats to disable normalization (original XVLA doesn't normalize)
-    )
-
-    return policy, preprocessor, postprocessor
-
-
-def create_dummy_data(device=DEVICE):
-    """Create dummy data for testing both implementations."""
-    batch_size = 1
-    prompt = "Pick up the red block and place it in the bin"
-
-    # Create random RGB images in [0, 255] uint8 range (as PIL images would be)
-    # Then convert to [0, 1] float32 range for LeRobot
-    def fake_rgb(h, w):
-        arr = np.random.randint(0, 255, (h, w, 3), dtype=np.uint8)
-        t = torch.from_numpy(arr).permute(2, 0, 1)  # CHW
-        return t
-
-    batch = {
-        f"{OBS_IMAGES}.image": torch.stack(
-            [fake_rgb(IMAGE_HEIGHT, IMAGE_WIDTH) for _ in range(batch_size)]
-        ).to(device),
-        f"{OBS_IMAGES}.image2": torch.stack(
-            [fake_rgb(IMAGE_HEIGHT, IMAGE_WIDTH) for _ in range(batch_size)]
-        ).to(device),
-        OBS_STATE: torch.randn(batch_size, DUMMY_STATE_DIM, dtype=torch.float32, device=device),
-        "task": [prompt for _ in range(batch_size)],
-    }
-
-    return batch
-
-
-# Pytest fixtures
-@pytest.fixture(scope="module")
-def xvla_components():
-    """Fixture to instantiate and provide all XVLA components for tests."""
-    print(f"\nTesting with DEVICE='{DEVICE}'")
-    print("\n[Setup] Instantiating LeRobot XVLA policy...")
-    policy_obj, preprocessor_obj, postprocessor_obj = instantiate_lerobot_xvla(from_pretrained=True)
-    print("✔️ Model loaded successfully")
-    yield policy_obj, preprocessor_obj, postprocessor_obj
-    cleanup_memory()
-
-
-@pytest.fixture(scope="module")
-def policy(xvla_components):
-    """Fixture to provide the XVLA policy for tests."""
-    return xvla_components[0]
-
-
-@pytest.fixture(scope="module")
-def preprocessor(xvla_components):
-    """Fixture to provide the XVLA preprocessor for tests."""
-    return xvla_components[1]
-
-
-@require_cuda
-def test_xvla_preprocessor_alignment(policy, preprocessor):
-    """Test that LeRobot XVLA preprocessor produces expected outputs."""
-    print("\n" + "=" * 80)
-    print("Test: XVLA Preprocessor Outputs")
-    print("=" * 80)
-
-    set_seed_all(42)
-
-    print("\nCreating dummy data...")
-    batch = create_dummy_data()
-
-    print("\n[LeRobot] Preprocessing...")
-    lerobot_observation = preprocessor(deepcopy(batch))
-    lerobot_inputs = policy._build_model_inputs(lerobot_observation)
-
-    print("\nVerifying preprocessor outputs:")
-    print("-" * 80)
-
-    # Expected shapes from tester.txt
-    expected_shapes = {
-        "domain_id": (1,),
-        "input_ids": (1, 50),
-        "proprio": (1, 20),
-        "image_mask": (1, 2),
-        "image_input": (1, 2, 3, 224, 224),
-    }
-
-    for key, expected_shape in expected_shapes.items():
-        if key in lerobot_inputs:
-            actual_shape = tuple(lerobot_inputs[key].shape)
-            print(f"\nKey: {key}")
-            print(f"Expected shape: {expected_shape}")
-            print(f"Actual shape: {actual_shape}")
-
-            if actual_shape == expected_shape:
-                print("Shape matches!")
-            else:
-                print("Shape mismatch!")
-
-            assert actual_shape == expected_shape, f"Shape mismatch for {key}"
-        else:
-            print(f"\nKey '{key}' not found in inputs!")
-
-    print("\nAll preprocessor outputs have correct shapes!")
-
-
-@require_cuda
-def test_xvla_action_generation(policy, preprocessor):
-    """Test XVLA LeRobot implementation generates expected actions."""
-    print("\n" + "=" * 80)
-    print("Test: XVLA Action Generation Against Expected Values")
-    print("=" * 80)
-
-    set_seed_all(42)
-
-    print("\nCreating dummy data...")
-    batch = create_dummy_data()
-
-    print("\n[LeRobot] Running inference...")
-    lerobot_observation = preprocessor(deepcopy(batch))
-    lerobot_inputs = policy._build_model_inputs(lerobot_observation)
-
-    # Reset seed for inference
-    torch.manual_seed(42)
-    with torch.no_grad():
-        lerobot_actions = policy.model.generate_actions(**lerobot_inputs, steps=10)
-        lerobot_actions = lerobot_actions.squeeze(0).float().cpu()
-
-    print(f"LeRobot actions shape: {lerobot_actions.shape}")
-    print(f"LeRobot actions mean: {lerobot_actions.mean().item():.6f}")
-    print(f"LeRobot actions std: {lerobot_actions.std().item():.6f}")
-    print(f"LeRobot actions first 5: {lerobot_actions[0, :5]}")
-
-    print("\nExpected values (from original XVLA):")
-    print(f"Expected actions shape: {EXPECTED_ACTIONS_SHAPE}")
-    print(f"Expected actions mean: {EXPECTED_ACTIONS_MEAN:.6f}")
-    print(f"Expected actions std: {EXPECTED_ACTIONS_STD:.6f}")
-    print(f"Expected actions first 5: {EXPECTED_ACTIONS_FIRST_5}")
-
-    print("\nAction Comparison:")
-    print("-" * 80)
-
-    # Compare shapes
-    actual_shape = tuple(lerobot_actions.shape)
-    assert actual_shape == EXPECTED_ACTIONS_SHAPE, (
-        f"Shape mismatch: {actual_shape} vs {EXPECTED_ACTIONS_SHAPE}"
-    )
-    print(f"✔️ Shape matches: {actual_shape}")
-
-    # Compare statistics
-    actual_mean = lerobot_actions.mean().item()
-    actual_std = lerobot_actions.std().item()
-
-    mean_diff = abs(actual_mean - EXPECTED_ACTIONS_MEAN)
-    std_diff = abs(actual_std - EXPECTED_ACTIONS_STD)
-
-    print(f"\nMean: {actual_mean:.6f} (expected: {EXPECTED_ACTIONS_MEAN:.6f}, diff: {mean_diff:.6e})")
-    print(f"Std: {actual_std:.6f} (expected: {EXPECTED_ACTIONS_STD:.6f}, diff: {std_diff:.6e})")
-
-    # Compare first 5 actions
-    actual_first_5 = lerobot_actions[0, :5]
-    first_5_diff = torch.abs(actual_first_5 - EXPECTED_ACTIONS_FIRST_5)
-
-    print("\nFirst 5 actions comparison:")
-    print(f"  Actual:   {actual_first_5}")
-    print(f"  Expected: {EXPECTED_ACTIONS_FIRST_5}")
-    print(f"  Max diff: {first_5_diff.max().item():.6e}")
-    print(f"  Mean diff: {first_5_diff.mean().item():.6e}")
-
-    # Check with different tolerances
-    tolerances = [1e-5, 1e-4, 1e-3, 1e-2]
-    for tol in tolerances:
-        is_close = torch.allclose(actual_first_5, EXPECTED_ACTIONS_FIRST_5, atol=tol)
-        status = "Success" if is_close else "Failure"
-        print(f"{status}: First 5 actions close (atol={tol}): {is_close}")
-
-    # Assert with reasonable tolerance
-    tolerance = 1e-3
-    assert torch.allclose(actual_first_5, EXPECTED_ACTIONS_FIRST_5, atol=tolerance), (
-        f"First 5 actions differ by more than tolerance ({tolerance})"
-    )
-    print(f"\nSuccess: Actions match expected values within tolerance ({tolerance})!")
-
-
-@require_cuda
-def test_xvla_inference_reproducibility(policy, preprocessor):
-    """Test that XVLA inference is reproducible with the same seed."""
-    print("\n" + "=" * 80)
-    print("Test: XVLA Inference Reproducibility")
-    print("=" * 80)
-
-    print("\nCreating dummy data...")
-    batch = create_dummy_data()
-
-    # First inference
-    print("\n[Run 1] Running inference...")
-    set_seed_all(42)
-    lerobot_observation = preprocessor(deepcopy(batch))
-    lerobot_inputs = policy._build_model_inputs(lerobot_observation)
-    with torch.no_grad():
-        actions_1 = policy.model.generate_actions(**lerobot_inputs, steps=10)
-        actions_1 = actions_1.squeeze(0).float().cpu()
-
-    # Second inference with same seed
-    print("\n[Run 2] Running inference with same seed...")
-    set_seed_all(42)
-    lerobot_observation = preprocessor(deepcopy(batch))
-    lerobot_inputs = policy._build_model_inputs(lerobot_observation)
-    with torch.no_grad():
-        actions_2 = policy.model.generate_actions(**lerobot_inputs, steps=10)
-        actions_2 = actions_2.squeeze(0).float().cpu()
-
-    print("\nComparing two runs:")
-    print("-" * 80)
-    if torch.allclose(actions_1, actions_2, atol=1e-8):
-        print("Inference is perfectly reproducible!")
-    else:
-        diff = torch.abs(actions_1 - actions_2)
-        print("Small differences detected:")
-        print(f"  Max diff: {diff.max().item():.6e}")
-        print(f"  Mean diff: {diff.mean().item():.6e}")
-
-    assert torch.allclose(actions_1, actions_2, atol=1e-6), "Inference should be reproducible!"
-
-    print("\nInference is reproducible!")
-
-
-if __name__ == "__main__":
-    print("\n" + "=" * 80)
-    print("XVLA LeRobot Validation Test Suite")
-    print("=" * 80)
-
-    try:
-        # Initialize model once for all tests
-        print("\n[Setup] Instantiating LeRobot XVLA policy...")
-        policy, preprocessor, postprocessor = instantiate_lerobot_xvla(from_pretrained=True)
-        print("✔️ Model loaded successfully")
-
-        # Run all tests with the same model instance
-        test_xvla_preprocessor_alignment(policy, preprocessor)
-        test_xvla_action_generation(policy, preprocessor)
-        test_xvla_inference_reproducibility(policy, preprocessor)
-
-        print("\n" + "=" * 80)
-        print("All tests passed!")
-        print("=" * 80)
-
-        cleanup_memory()
-    except Exception as e:
-        print("\n" + "=" * 80)
-        print(f"Test failed with error: {e}")
-        print("=" * 80)
-        cleanup_memory()
-        raise

tests/processor/test_libero_processor.py

@@ -1,72 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import numpy as np
-import torch
-
-from lerobot.envs.utils import preprocess_observation
-from lerobot.processor.env_processor import LiberoProcessorStep
-from lerobot.processor.pipeline import PolicyProcessorPipeline
-
-seed = 42
-np.random.seed(seed)
-
-B = 5
-obs1 = {
-    "pixels": {
-        "image": (np.random.rand(B, 256, 256, 3) * 255).astype(np.uint8),
-        "image2": (np.random.rand(B, 256, 256, 3) * 255).astype(np.uint8),
-    },
-    "robot_state": {
-        "eef": {
-            "pos": np.random.randn(B, 3),
-            "quat": np.random.randn(B, 4),
-            "mat": np.random.randn(B, 3, 3),
-        },
-        "gripper": {
-            "qpos": np.random.randn(B, 2),
-            "qvel": np.random.randn(B, 2),
-        },
-        "joints": {
-            "pos": np.random.randn(B, 7),
-            "vel": np.random.randn(B, 7),
-        },
-    },
-}
-
-observation = preprocess_observation(obs1)
-libero_preprocessor = PolicyProcessorPipeline(
-    steps=[
-        LiberoProcessorStep(),
-    ]
-)
-processed_obs = libero_preprocessor(observation)
-assert "observation.state" in processed_obs
-state = processed_obs["observation.state"]
-assert isinstance(state, torch.Tensor)
-assert state.dtype == torch.float32
-
-assert state.shape[0] == B
-assert state.shape[1] == 8
-
-assert "observation.images.image" in processed_obs
-assert "observation.images.image2" in processed_obs
-
-assert isinstance(processed_obs["observation.images.image"], torch.Tensor)
-assert isinstance(processed_obs["observation.images.image2"], torch.Tensor)
-
-assert processed_obs["observation.images.image"].shape == (B, 3, 256, 256)
-assert processed_obs["observation.images.image2"].shape == (B, 3, 256, 256)