Merge branch 'main' into refactor/lerobot_train_rabc

.github/workflows/full_tests.yml

@@ -101,11 +101,9 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     if: |
-      github.repository == 'huggingface/lerobot' && (
-        (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
-        github.event_name == 'push' ||
-        github.event_name == 'workflow_dispatch'
-      )
+      (github.event_name == 'pull_request_review' && github.event.review.state == 'approved' && github.event.pull_request.head.repo.fork == false) ||
+      github.event_name == 'push' ||
+      github.event_name == 'workflow_dispatch'
     outputs:
       image_tag: ${{ steps.set_tag.outputs.image_tag }}
     env:

.github/workflows/unbound_deps_tests.yml

@@ -91,7 +91,6 @@ jobs:
     name: Build and Push Docker
     runs-on:
       group: aws-general-8-plus
-    if: github.repository == 'huggingface/lerobot'
     outputs:
       image_tag: ${{ env.DOCKER_IMAGE_NAME }}
     env:

benchmarks/tokens/README.md

@@ -1,134 +0,0 @@
-# Action tokenizer benchmark
-
-## Questions
-
-What is the trade-off between:
-
-- **Compression**: how many tokens are needed to represent an action chunk (e.g. horizon × action_dim floats)?
-- **Reconstruction quality**: how well does encode-then-decode preserve the original actions?
-- **Speed**: how long does encoding and decoding take per chunk?
-
-How to choose an action tokenizer?
-
-- Which tokenizer architecture (e.g. dct + BPE, DCT + BPE)?
-- Which **action horizon** and **encoded dimensions** to use?
-- Which **normalization** (QUANTILES, MEAN_STD, MIN_MAX) and **delta transform** (relative vs absolute actions)?
-- How do reconstruction error and compression ratio vary across datasets and tokenizer settings?
-
-This benchmark loads action chunks from a LeRobot dataset using the same pipeline as `lerobot-train-tokenizer`, runs a trained action tokenizer in encode/decode mode, and reports reconstruction error, compression stats, and timing. Results are saved as JSON under `outputs/` for comparison and analysis.
-
-## Variables
-
-**Dataset & chunking**
-
-- **repo_id**: LeRobot dataset (e.g. `lerobot/pusht`). Action statistics and normalization are taken from the dataset metadata when available.
-- **action_horizon**: Number of future steps per action chunk (must match the tokenizer’s training).
-- **encoded_dims**: Dimension ranges to encode (e.g. `0:6` or `0:6,7:14`). Must match the tokenizer.
-- **max_episodes**: Cap on episodes to load (default: all).
-- **sample_fraction**: Fraction of chunks to sample per episode (default `0.2`) to keep runtime manageable.
-
-**Transform & normalization**
-
-- **normalization_mode**: `IDENTITY`, `MEAN_STD`, `MIN_MAX`, `QUANTILES`, `QUANTILE10`. Should match the tokenizer’s training.
-- **delta_dims**: Comma-separated dimension indices for delta (relative) transform.
-- **use_delta_transform**: Whether to convert actions to relative to current state for those dimensions.
-- **state_key**: Dataset key for state (e.g. `observation.state`) used when applying delta transform.
-
-**Tokenizer & evaluation**
-
-- **action_tokenizer_path**: Path or HuggingFace repo id of the trained tokenizer (e.g. `outputs/wavetoken`).
-- **max_chunks_for_reconstruction**: Max number of chunks to use for reconstruction and timing (default `500`) to limit runtime.
-
-### Main parameters
-
-| parameter                        | default                      | description                                      |
-| -------------------------------- | ---------------------------- | ------------------------------------------------ |
-| **action_tokenizer_path**        | (required)                   | Path or Hub id of the trained action tokenizer.  |
-| **repo_id**                      | (required)                   | LeRobot dataset repo id.                         |
-| **action_horizon**               | `10`                         | Future steps per chunk.                          |
-| **encoded_dims**                 | `0:6`                        | Dimension ranges to encode (e.g. `0:6,7:14`).   |
-| **normalization_mode**           | `QUANTILES`                  | Normalization mode for actions.                  |
-| **max_episodes**                 | all                          | Max episodes to load.                            |
-| **sample_fraction**              | `0.2`                        | Fraction of chunks sampled per episode.          |
-| **max_chunks_for_reconstruction**| `500`                        | Chunks used for reconstruction and timing.       |
-| **output_dir**                   | `outputs/action_tokenizer_benchmark` | Directory for results JSON.              |
-
-## Metrics
-
-**Reconstruction (lower is better)**
-
-- **reconstruction_mae**: Mean absolute error between original and decoded action chunks.
-- **reconstruction_mse**: Mean squared error.
-- **reconstruction_rmse**: Root mean squared error.
-- **reconstruction_max_abs_error**: Maximum absolute error over all dimensions and samples.
-- **per_dimension_mae**: MAE per action dimension (list of length `action_dim`).
-
-**Compression**
-
-- **compression_ratio**: Ratio (action_horizon × action_dim) / mean number of tokens. Higher means more compression.
-- **mean_token_length**, **std_token_length**: Mean and standard deviation of token count per chunk.
-- **min_token_length**, **max_token_length**: Min and max token count.
-- **p50_token_length**, **p99_token_length**: 50th and 99th percentile token counts.
-
-**Timing (seconds per chunk)**
-
-- **mean_encode_time_sec**: Mean time to encode one chunk.
-- **mean_decode_time_sec**: Mean time to decode one chunk.
-
-The JSON output also includes **num_chunks_evaluated** and **total_chunks_available** for context.
-
-## How the benchmark works
-
-1. **Load dataset**: LeRobot dataset is loaded for the given `repo_id` and `root`.
-2. **Build action chunks**: For each episode (up to `max_episodes`), action chunks are built with the same logic as `lerobot-train-tokenizer`: sliding window of length `action_horizon`, optional delta transform, and per-episode sampling with `sample_fraction`.
-3. **Extract and normalize**: Only `encoded_dims` are kept. Normalization is applied using the dataset’s action stats when available, according to `normalization_mode`.
-4. **Encode / decode**: A random sample of chunks (size `max_chunks_for_reconstruction`) is encoded and then decoded with the tokenizer. Encode and decode times are recorded per chunk.
-5. **Compute metrics**: Reconstruction metrics are computed between original and decoded chunks; compression and timing stats are aggregated.
-6. **Save results**: A JSON file is written to `output_dir` with name `{timestamp}_{repo_id}_action_tokenizer_results.json`, containing the full config and all metrics.
-
-The pipeline (chunking, dimensions, normalization, delta) must match how the tokenizer was trained; otherwise reconstruction error can be large or the tokenizer may raise.
-
-## Caveats
-
-- The tokenizer’s **action_horizon** and **action_dim** (and optionally DCT settings) are fixed at training time. The benchmark infers dimensions from the dataset and encoded dims; the tokenizer path must correspond to a model trained with the same horizon and encoded dimensions.
-- Reconstruction is evaluated in **normalized space** (the same space the tokenizer sees). For interpretation in raw action space, you would need to invert normalization outside this script.
-- Only one tokenizer and one dataset are evaluated per run. To compare tokenizers or datasets, run the script multiple times and compare the saved JSON files.
-
-## Example
-
-Quick run with a local tokenizer and a small number of episodes:
-
-```bash
-python benchmarks/tokens/run_action_tokenizer_benchmark.py \
-    --action-tokenizer-path=outputs/wavetoken \
-    --repo-id=lerobot/pusht \
-    --action-horizon=10 \
-    --max-episodes=50 \
-    --output-dir=outputs/action_tokenizer_benchmark
-```
-
-With delta transform and custom encoded dimensions:
-
-```bash
-python benchmarks/tokens/run_action_tokenizer_benchmark.py \
-    --action-tokenizer-path=outputs/wavetoken \
-    --repo-id=lerobot/pusht \
-    --action-horizon=10 \
-    --encoded-dims=0:6,7:14 \
-    --delta-dims=0,1,2,3,4,5 \
-    --use-delta-transform \
-    --normalization-mode=QUANTILES \
-    --max-chunks-for-reconstruction=500 \
-    --output-dir=outputs/action_tokenizer_benchmark
-```
-
-Results are written to e.g. `outputs/action_tokenizer_benchmark/2026-02-12_14-30-00_lerobot_pusht_action_tokenizer_results.json`.
-
-## Results
-
-Results are stored as JSON in the directory given by `--output-dir` (default: `outputs/action_tokenizer_benchmark`). Each file contains:
-
-- **config**: All script arguments (tokenizer path, repo_id, action_horizon, encoded_dims, normalization_mode, etc.) for reproducibility.
-- **metrics**: All reconstruction, compression, and timing metrics described above.
-
-To compare runs, load and diff or aggregate these JSON files with your own scripts or notebooks.

benchmarks/tokens/run_action_tokenizer_benchmark.py

@@ -1,442 +0,0 @@
-#!/usr/bin/env python
-# Copyright 2026 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.
-"""Benchmark action tokenization: reconstruction error, compression ratio, and timing.
-
-Loads action chunks from a LeRobot dataset, encodes/decodes them with a trained action
-tokenizer, and reports:
-- Reconstruction: MAE, MSE, RMSE, max absolute error, per-dimension MAE
-- Jerk: mean absolute jerk (original and reconstructed), jerk reconstruction MAE
-- Compression: ratio (input size / mean tokens), token length stats
-- Timing: mean encode/decode time per chunk
-
-Results are saved to outputs/action_tokenizer_benchmark/<timestamp>_results.json.
-
-Example:
-
-```bash
-python benchmarks/tokens/run_action_tokenizer_benchmark.py \
-    --action-tokenizer-path=outputs/wavetoken \
-    --repo-id=lerobot/pusht \
-    --action-horizon=10 \
-    --max-episodes=50 \
-    --output-dir=outputs/action_tokenizer_benchmark
-```
-"""
-
-import argparse
-import json
-import time
-from pathlib import Path
-
-import numpy as np
-
-from lerobot.configs.types import NormalizationMode
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.utils.constants import ACTION, OBS_STATE
-
-# Optional: use same helpers as train script if we want to avoid duplication
-from lerobot.scripts.lerobot_train_tokenizer import (
-    apply_normalization,
-    process_episode,
-)
-
-
-def load_action_chunks(
-    repo_id: str,
-    root: str | None,
-    action_horizon: int,
-    max_episodes: int | None,
-    sample_fraction: float,
-    encoded_dims: str,
-    delta_dims: str | None,
-    use_delta_transform: bool,
-    state_key: str,
-    normalization_mode: NormalizationMode,
-):
-    """Load and normalize action chunks from a LeRobot dataset (same pipeline as training)."""
-    dataset = LeRobotDataset(repo_id=repo_id, root=root)
-    num_episodes = dataset.num_episodes
-    if max_episodes is not None:
-        num_episodes = min(max_episodes, num_episodes)
-
-    # Parse encoded dims
-    encoded_dim_ranges = []
-    for range_str in encoded_dims.split(","):
-        start, end = map(int, range_str.strip().split(":"))
-        encoded_dim_ranges.append((start, end))
-    total_encoded_dims = sum(end - start for start, end in encoded_dim_ranges)
-
-    delta_dim_list = None
-    if delta_dims is not None and delta_dims.strip():
-        delta_dim_list = [int(d.strip()) for d in delta_dims.split(",")]
-
-    all_chunks = []
-    for ep_idx in range(num_episodes):
-        chunks = process_episode(
-            (
-                dataset,
-                ep_idx,
-                action_horizon,
-                delta_dim_list,
-                sample_fraction,
-                state_key,
-                use_delta_transform,
-            )
-        )
-        if chunks is not None:
-            all_chunks.append(chunks)
-
-    if not all_chunks:
-        raise ValueError("No action chunks collected. Check action_horizon and dataset.")
-
-    all_chunks = np.concatenate(all_chunks, axis=0)
-
-    # Extract encoded dimensions only
-    encoded_chunks = []
-    for start, end in encoded_dim_ranges:
-        encoded_chunks.append(all_chunks[:, :, start:end])
-    encoded_chunks = np.concatenate(encoded_chunks, axis=-1)
-
-    # Normalize
-    norm_stats = dataset.meta.stats
-    if norm_stats is not None and ACTION in norm_stats:
-        action_stats = norm_stats[ACTION]
-        encoded_dim_indices = []
-        for start, end in encoded_dim_ranges:
-            encoded_dim_indices.extend(range(start, end))
-        encoded_dim_indices = np.array(encoded_dim_indices)
-        encoded_stats = {}
-        for stat_name, stat_values in action_stats.items():
-            if isinstance(stat_values, (list, np.ndarray)):
-                stat_array = np.array(stat_values)
-                if len(stat_array) > max(encoded_dim_indices):
-                    encoded_stats[stat_name] = stat_array[encoded_dim_indices]
-        if encoded_stats:
-            try:
-                encoded_chunks = apply_normalization(
-                    encoded_chunks, encoded_stats, normalization_mode, eps=1e-8
-                )
-            except ValueError:
-                pass
-
-    return encoded_chunks, total_encoded_dims, action_horizon, dataset.repo_id
-
-
-def compute_reconstruction_metrics(original: np.ndarray, reconstructed: np.ndarray):
-    """Compute reconstruction error metrics (original and reconstructed same shape [N, T, D])."""
-    diff = reconstructed - original
-    mae = float(np.mean(np.abs(diff)))
-    mse = float(np.mean(diff**2))
-    rmse = float(np.sqrt(mse))
-    max_abs_err = float(np.max(np.abs(diff)))
-
-    # Per-dimension MAE (over N and T)
-    per_dim_mae = np.mean(np.abs(diff), axis=(0, 1))
-    per_dim_mae = per_dim_mae.tolist()
-
-    return {
-        "reconstruction_mae": mae,
-        "reconstruction_mse": mse,
-        "reconstruction_rmse": rmse,
-        "reconstruction_max_abs_error": max_abs_err,
-        "per_dimension_mae": per_dim_mae,
-    }
-
-
-def compute_jerk_metrics(original: np.ndarray, reconstructed: np.ndarray) -> dict:
-    """Compute jerk (3rd derivative of action w.r.t. time) metrics.
-
-    Args:
-        original: Action chunks [N, T, D].
-        reconstructed: Reconstructed action chunks [N, T, D].
-
-    Returns:
-        Dict with mean absolute jerk for original, reconstructed, and jerk reconstruction MAE.
-    """
-    # Jerk = 3rd discrete difference along time axis; need T >= 4
-    if original.shape[1] < 4:
-        return {}
-    jerk_orig = np.diff(original, n=3, axis=1)  # (N, T-3, D)
-    jerk_recon = np.diff(reconstructed, n=3, axis=1)
-    mae_jerk_orig = float(np.mean(np.abs(jerk_orig)))
-    mae_jerk_recon = float(np.mean(np.abs(jerk_recon)))
-    jerk_reconstruction_mae = float(np.mean(np.abs(jerk_recon - jerk_orig)))
-    return {
-        "jerk_mae_original": mae_jerk_orig,
-        "jerk_mae_reconstructed": mae_jerk_recon,
-        "jerk_reconstruction_mae": jerk_reconstruction_mae,
-    }
-
-
-def run_benchmark(
-    action_chunks: np.ndarray,
-    action_horizon: int,
-    action_dim: int,
-    tokenizer_path: str,
-    max_chunks_for_reconstruction: int | None = 500,
-):
-    """Encode/decode action chunks and compute metrics."""
-    from transformers import AutoProcessor
-
-    processor = AutoProcessor.from_pretrained(tokenizer_path, trust_remote_code=True)
-
-    n_chunks = len(action_chunks)
-    sample_size = n_chunks
-    if max_chunks_for_reconstruction is not None:
-        sample_size = min(max_chunks_for_reconstruction, n_chunks)
-    rng = np.random.RandomState(42)
-    indices = rng.choice(n_chunks, size=sample_size, replace=False)
-    sample_chunks = action_chunks[indices]
-
-    # Encode
-    token_lengths = []
-    encode_times = []
-    all_tokens = []
-    for i in range(len(sample_chunks)):
-        chunk = sample_chunks[i : i + 1]
-        t0 = time.perf_counter()
-        tokens = processor(chunk)[0]
-        encode_times.append(time.perf_counter() - t0)
-        if isinstance(tokens, list):
-            token_lengths.append(len(tokens))
-            all_tokens.append(tokens)
-        else:
-            n = tokens.shape[0] if hasattr(tokens, "shape") else len(tokens)
-            token_lengths.append(n)
-            all_tokens.append(tokens.tolist() if hasattr(tokens, "tolist") else list(tokens))
-
-    # Decode (processor keeps time_horizon/action_dim from encode)
-    decoded_list = []
-    decode_times = []
-    for i, tok_list in enumerate(all_tokens):
-        t0 = time.perf_counter()
-        recon = processor.decode(
-            [tok_list],
-            time_horizon=action_horizon,
-            action_dim=action_dim,
-        )
-        decode_times.append(time.perf_counter() - t0)
-        decoded_list.append(recon)
-    decoded = np.concatenate(decoded_list, axis=0)
-
-    # Reconstruction metrics
-    metrics = compute_reconstruction_metrics(sample_chunks, decoded)
-
-    # Jerk metrics (3rd derivative along time)
-    jerk_metrics = compute_jerk_metrics(sample_chunks, decoded)
-    metrics.update(jerk_metrics)
-
-    # Compression
-    token_lengths = np.array(token_lengths)
-    input_size = action_horizon * action_dim
-    compression_ratio = input_size / float(np.mean(token_lengths))
-    metrics["compression_ratio"] = compression_ratio
-    metrics["mean_token_length"] = float(np.mean(token_lengths))
-    metrics["std_token_length"] = float(np.std(token_lengths))
-    metrics["min_token_length"] = int(np.min(token_lengths))
-    metrics["max_token_length"] = int(np.max(token_lengths))
-    metrics["p50_token_length"] = float(np.percentile(token_lengths, 50))
-    metrics["p99_token_length"] = float(np.percentile(token_lengths, 99))
-
-    # Timing (seconds per chunk)
-    metrics["mean_encode_time_sec"] = float(np.mean(encode_times))
-    metrics["mean_decode_time_sec"] = float(np.mean(decode_times))
-    metrics["num_chunks_evaluated"] = sample_size
-    metrics["total_chunks_available"] = n_chunks
-
-    return metrics
-
-
-def main(
-    action_tokenizer_path: str,
-    repo_id: str,
-    root: str | None = None,
-    action_horizon: int = 10,
-    max_episodes: int | None = 100,
-    sample_fraction: float = 0.2,
-    encoded_dims: str = "0:6",
-    delta_dims: str | None = None,
-    use_delta_transform: bool = False,
-    state_key: str = OBS_STATE,
-    normalization_mode: str = "QUANTILES",
-    max_chunks_for_reconstruction: int | None = 500,
-    output_dir: str | None = None,
-):
-    if output_dir is None:
-        output_dir = "outputs/action_tokenizer_benchmark"
-    output_path = Path(output_dir)
-    output_path.mkdir(parents=True, exist_ok=True)
-
-    try:
-        norm_mode = NormalizationMode(normalization_mode)
-    except ValueError:
-        norm_mode = NormalizationMode.QUANTILES
-
-    print("Loading action chunks...")
-    encoded_chunks, action_dim, horizon, _ = load_action_chunks(
-        repo_id=repo_id,
-        root=root,
-        action_horizon=action_horizon,
-        max_episodes=max_episodes,
-        sample_fraction=sample_fraction,
-        encoded_dims=encoded_dims,
-        delta_dims=delta_dims,
-        use_delta_transform=use_delta_transform,
-        state_key=state_key,
-        normalization_mode=norm_mode,
-    )
-    print(f"Loaded {len(encoded_chunks)} chunks, shape {encoded_chunks.shape} (H={horizon}, D={action_dim})")
-
-    print("Running tokenizer benchmark...")
-    metrics = run_benchmark(
-        action_chunks=encoded_chunks,
-        action_horizon=horizon,
-        action_dim=action_dim,
-        tokenizer_path=action_tokenizer_path,
-        max_chunks_for_reconstruction=max_chunks_for_reconstruction,
-    )
-
-    # Attach config for reproducibility
-    results = {
-        "config": {
-            "action_tokenizer_path": action_tokenizer_path,
-            "repo_id": repo_id,
-            "action_horizon": action_horizon,
-            "max_episodes": max_episodes,
-            "sample_fraction": sample_fraction,
-            "encoded_dims": encoded_dims,
-            "delta_dims": delta_dims,
-            "use_delta_transform": use_delta_transform,
-            "state_key": state_key,
-            "normalization_mode": normalization_mode,
-        },
-        "metrics": metrics,
-    }
-
-    timestamp = time.strftime("%Y-%m-%d_%H-%M-%S")
-    safe_repo = repo_id.replace("/", "_")
-    out_file = output_path / f"{timestamp}_{safe_repo}_action_tokenizer_results.json"
-    with open(out_file, "w") as f:
-        json.dump(results, f, indent=2)
-
-    print(f"Results saved to {out_file}")
-    print("Metrics:")
-    for k, v in metrics.items():
-        if isinstance(v, list):
-            print(f"  {k}: (length {len(v)})")
-        else:
-            print(f"  {k}: {v}")
-
-    return results
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        description="Benchmark action tokenization (reconstruction error, compression, timing)."
-    )
-    parser.add_argument(
-        "--action-tokenizer-path",
-        type=str,
-        required=True,
-        help="Path or HuggingFace repo id of the trained action tokenizer (e.g. outputs/wavetoken).",
-    )
-    parser.add_argument(
-        "--repo-id",
-        type=str,
-        required=True,
-        help="LeRobot dataset repo id (e.g. lerobot/pusht).",
-    )
-    parser.add_argument(
-        "--root",
-        type=str,
-        default=None,
-        help="Root directory for LeRobot datasets.",
-    )
-    parser.add_argument(
-        "--action-horizon",
-        type=int,
-        default=10,
-        help="Number of future steps per action chunk.",
-    )
-    parser.add_argument(
-        "--max-episodes",
-        type=int,
-        default=None,
-        help="Max episodes to use (default: all).",
-    )
-    parser.add_argument(
-        "--sample-fraction",
-        type=float,
-        default=0.2,
-        help="Fraction of chunks to sample per episode.",
-    )
-    parser.add_argument(
-        "--encoded-dims",
-        type=str,
-        default="0:6",
-        help="Dimension ranges to encode (e.g. 0:6,7:14).",
-    )
-    parser.add_argument(
-        "--delta-dims",
-        type=str,
-        default=None,
-        help="Comma-separated dimensions for delta transform.",
-    )
-    parser.add_argument(
-        "--use-delta-transform",
-        action="store_true",
-        help="Apply delta (relative) transform to specified dimensions.",
-    )
-    parser.add_argument(
-        "--state-key",
-        type=str,
-        default=OBS_STATE,
-        help="Dataset key for state (for delta transform).",
-    )
-    parser.add_argument(
-        "--normalization-mode",
-        type=str,
-        default="QUANTILES",
-        choices=[m.value for m in NormalizationMode],
-        help="Normalization mode for actions.",
-    )
-    parser.add_argument(
-        "--max-chunks-for-reconstruction",
-        type=int,
-        default=500,
-        help="Max chunks to use for reconstruction metrics (default: 500).",
-    )
-    parser.add_argument(
-        "--output-dir",
-        type=str,
-        default="outputs/action_tokenizer_benchmark",
-        help="Directory to save results JSON (default: outputs/action_tokenizer_benchmark).",
-    )
-    args = parser.parse_args()
-    main(
-        action_tokenizer_path=args.action_tokenizer_path,
-        repo_id=args.repo_id,
-        root=args.root,
-        action_horizon=args.action_horizon,
-        max_episodes=args.max_episodes,
-        sample_fraction=args.sample_fraction,
-        encoded_dims=args.encoded_dims,
-        delta_dims=args.delta_dims,
-        use_delta_transform=args.use_delta_transform,
-        state_key=args.state_key,
-        normalization_mode=args.normalization_mode,
-        max_chunks_for_reconstruction=args.max_chunks_for_reconstruction,
-        output_dir=args.output_dir,
-    )

docs/source/_toctree.yml

@@ -7,6 +7,8 @@
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
+  - local: cameras
+    title: Cameras
   - local: bring_your_own_policies
     title: Bring Your Own Policies
   - local: integrate_hardware
@@ -27,8 +29,6 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     title: Using the Dataset Tools
-  - local: dataset_subtask
-    title: Using Subtasks in the Dataset
   title: "Datasets"
 - sections:
   - local: act
@@ -101,17 +101,11 @@
     title: Earth Rover Mini
   - local: omx
     title: OMX
-  - local: openarm
-    title: OpenArm
   title: "Robots"
 - sections:
   - local: phone_teleop
     title: Phone
   title: "Teleoperators"
-- sections:
-  - local: cameras
-    title: Cameras
-  title: "Sensors"
 - sections:
   - local: torch_accelerators
     title: PyTorch accelerators
@@ -121,8 +115,6 @@
     title: Notebooks
   - local: feetech
     title: Updating Feetech Firmware
-  - local: damiao
-    title: Damiao Motors and CAN Bus
   title: "Resources"
 - sections:
   - local: contributing

docs/source/cameras.mdx

@@ -1,22 +1,12 @@
 # Cameras
 
-LeRobot offers multiple options for video capture:
+LeRobot offers multiple options for video capture, including phone cameras, built-in laptop cameras, external webcams, and Intel RealSense cameras. To efficiently record frames from most cameras, you can use either the `OpenCVCamera` or `RealSenseCamera` class. For additional compatibility details on the `OpenCVCamera` class, refer to the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
 
-| Class             | Supported Cameras                   |
-| ----------------- | ----------------------------------- |
-| `OpenCVCamera`    | Phone, built-in laptop, USB webcams |
-| `ZMQCamera`       | Network-connected cameras           |
-| `RealSenseCamera` | Intel RealSense (with depth)        |
-| `Reachy2Camera`   | Reachy 2 robot cameras              |
+### Finding your camera
 
-> [!TIP]
-> For `OpenCVCamera` compatibility details, see the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
+To instantiate a camera, you need a camera identifier. This identifier might change if you reboot your computer or re-plug your camera, a behavior mostly dependant on your operating system.
 
-### Find your camera
-
-Every camera requires a unique identifier to be instantiated, allowing you to distinguish between multiple connected devices.
-
-`OpenCVCamera` and `RealSenseCamera` support auto-discovery. Run the command below to list available devices and their identifiers. Note that these identifiers may change after rebooting your computer or re-plugging the camera, depending on your operating system.
+To find the camera indices of the cameras plugged into your system, run the following script:
 
 ```bash
 lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
@@ -24,7 +14,7 @@ lerobot-find-cameras opencv # or realsense for Intel Realsense cameras
 
 The output will look something like this if you have two cameras connected:
 
-```bash
+```
 --- Detected Cameras ---
 Camera #0:
   Name: OpenCV Camera @ 0
@@ -43,37 +33,13 @@ Camera #0:
 > [!WARNING]
 > When using Intel RealSense cameras in `macOS`, you could get this [error](https://github.com/IntelRealSense/librealsense/issues/12307): `Error finding RealSense cameras: failed to set power state`, this can be solved by running the same command with `sudo` permissions. Note that using RealSense cameras in `macOS` is unstable.
 
-`ZMQCamera` and `Reachy2Camera` do not support auto-discovery. They must be configured manually by providing their network address and port or robot SDK settings.
+## Use Cameras
 
-## Use cameras
+Below are two examples, demonstrating how to work with the API.
 
-### Frame access modes
-
-All camera classes implement three access modes for capturing frames:
-
-| Method                    | Behavior                                                                                                                                                   | Blocks?        | Best For                                 |
-| ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------- | ---------------------------------------- |
-| `read()`                  | Waits for the camera hardware to return a frame. May block for a long time depending on the camera and SDK.                                                | Yes            | Simple scripts, sequential capture       |
-| `async_read(timeout_ms)`  | Returns the latest unconsumed frame from background thread. Blocks only if buffer is empty, up to `timeout_ms`. Raises `TimeoutError` if no frame arrives. | With a timeout | Control loops synchronized to camera FPS |
-| `read_latest(max_age_ms)` | Peeks at the most recent frame in buffer (may be stale). Raises `TimeoutError` if frame is older than `max_age_ms`.                                        | No             | UI visualization, logging, monitoring    |
-
-### Usage examples
-
-The following examples show how to use the camera API to configure and capture frames from different camera types.
-
-- **Blocking and non-blocking frame capture** using an OpenCV-based camera
+- **Asynchronous frame capture** using an OpenCV-based camera
 - **Color and depth capture** using an Intel RealSense camera
 
-> [!WARNING]
-> Failing to cleanly disconnect cameras can cause resource leaks. Use the context manager protocol to ensure automatic cleanup:
->
-> ```python
-> with OpenCVCamera(config) as camera:
->     ...
-> ```
->
-> You can also call `connect()` and `disconnect()` manually, but always use a `finally` block for the latter.
-
 <hfoptions id="shell_restart">
 <hfoption id="Open CV Camera">
 
@@ -94,30 +60,16 @@ config = OpenCVCameraConfig(
 )
 
 # Instantiate and connect an `OpenCVCamera`, performing a warm-up read (default).
-with OpenCVCamera(config) as camera:
-
-    # Read a frame synchronously — blocks until hardware delivers a new frame
-    frame = camera.read()
-    print(f"read() call returned frame with shape:", frame.shape)
-
-    # Read a frame asynchronously with a timeout — returns the latest unconsumed frame or waits up to timeout_ms for a new one
-    try:
-        for i in range(10):
-            frame = camera.async_read(timeout_ms=200)
-            print(f"async_read call returned frame {i} with shape:", frame.shape)
-    except TimeoutError as e:
-        print(f"No frame received within timeout: {e}")
-
-    # Instantly return a frame - returns the most recent frame captured by the camera
-    try:
-        initial_frame = camera.read_latest(max_age_ms=1000)
-        for i in range(10):
-            frame = camera.read_latest(max_age_ms=1000)
-            print(f"read_latest call returned frame {i} with shape:", frame.shape)
-            print(f"Was a new frame received by the camera? {not (initial_frame == frame).any()}")
-    except TimeoutError as e:
-        print(f"Frame too old: {e}")
+camera = OpenCVCamera(config)
+camera.connect()
 
+# Read frames asynchronously in a loop via `async_read(timeout_ms)`
+try:
+    for i in range(10):
+        frame = camera.async_read(timeout_ms=200)
+        print(f"Async frame {i} shape:", frame.shape)
+finally:
+    camera.disconnect()
 ```
 <!-- prettier-ignore-end -->
 
@@ -159,10 +111,10 @@ finally:
 </hfoption>
 </hfoptions>
 
-## Use your phone's camera
+## Use your phone
 
 <hfoptions id="use phone">
-<hfoption id="iPhone & macOS">
+<hfoption id="Mac">
 
 To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
 
@@ -172,49 +124,83 @@ To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
 
 For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).
 
+Your iPhone should be detected automatically when running the camera setup script in the next section.
+
 </hfoption>
-<hfoption id="OBS virtual camera">
+<hfoption id="Linux">
 
-If you want to use your phone as a camera using OBS, follow these steps to set up a virtual camera.
+If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
 
-1. _(Linux only) Install `v4l2loopback-dkms` and `v4l-utils`_. These packages create virtual camera devices and verify their settings. Install with:
+1. _Install `v4l2loopback-dkms` and `v4l-utils`_. Those packages are required to create virtual camera devices (`v4l2loopback`) and verify their settings with the `v4l2-ctl` utility from `v4l-utils`. Install them using:
 
-```bash
+<!-- prettier-ignore-start -->
+```python
 sudo apt install v4l2loopback-dkms v4l-utils
 ```
+<!-- prettier-ignore-end -->
 
-2. _Install the [DroidCam app](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
-3. _Download and install [OBS Studio](https://obsproject.com)_.
-4. _Download and install the [DroidCam OBS plugin](https://droidcam.app/obs)_.
-5. _Start OBS Studio_.
+2. _Install [DroidCam](https://droidcam.app) on your phone_. This app is available for both iOS and Android.
+3. _Install [OBS Studio](https://obsproject.com)_. This software will help you manage the camera feed. Install it using [Flatpak](https://flatpak.org):
 
-6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480` to avoid the watermarks.
-7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video` or `OBS > Preferences... > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it.
+<!-- prettier-ignore-start -->
+```python
+flatpak install flathub com.obsproject.Studio
+```
+<!-- prettier-ignore-end -->
+
+4. _Install the DroidCam OBS plugin_. This plugin integrates DroidCam with OBS Studio. Install it with:
+
+<!-- prettier-ignore-start -->
+```python
+flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
+```
+<!-- prettier-ignore-end -->
+
+5. _Start OBS Studio_. Launch with:
+
+<!-- prettier-ignore-start -->
+```python
+flatpak run com.obsproject.Studio
+```
+<!-- prettier-ignore-end -->
+
+6. _Add your phone as a source_. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480`.
+7. _Adjust resolution settings_. In OBS Studio, go to `File > Settings > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it in.
 8. _Start virtual camera_. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
-9. _Verify the virtual camera setup and resolution_.
-   - **Linux**: Use `v4l2-ctl` to list devices and check resolution:
-     ```bash
-     v4l2-ctl --list-devices  # find VirtualCam and note its /dev/videoX path
-     v4l2-ctl -d /dev/videoX --get-fmt-video  # replace with your VirtualCam path
-     ```
-     You should see `VirtualCam` listed and resolution `640x480`.
-   - **macOS**: Open Photo Booth or FaceTime and select "OBS Virtual Camera" as the input.
-   - **Windows**: The native Camera app doesn't support virtual cameras. Use a video conferencing app (Zoom, Teams) or run `lerobot-find-cameras opencv` directly to verify.
+9. _Verify the virtual camera setup_. Use `v4l2-ctl` to list the devices:
 
-<details>
-<summary><strong>Troubleshooting</strong></summary>
+<!-- prettier-ignore-start -->
+```python
+v4l2-ctl --list-devices
+```
+<!-- prettier-ignore-end -->
 
-> The virtual camera resolution is incorrect.
+You should see an entry like:
 
-Delete the virtual camera source and recreate it. The resolution cannot be changed after creation.
+```
+VirtualCam (platform:v4l2loopback-000):
+/dev/video1
+```
 
-> Error reading frame in background thread for OpenCVCamera(X): OpenCVCamera(X) frame width=640 or height=480 do not match configured width=1920 or height=1080.
+10. _Check the camera resolution_. Use `v4l2-ctl` to ensure that the virtual camera output resolution is `640x480`. Change `/dev/video1` to the port of your virtual camera from the output of `v4l2-ctl --list-devices`.
 
-This error is caused by OBS Virtual Camera advertising a `1920x1080` resolution despite rescaling. The only fix for now is to comment out the width and height check in `_postprocess_image()`.
+<!-- prettier-ignore-start -->
+```python
+v4l2-ctl -d /dev/video1 --get-fmt-video
+```
+<!-- prettier-ignore-end -->
 
-</details>
+You should see an entry like:
+
+```
+>>> Format Video Capture:
+>>>	Width/Height      : 640/480
+>>>	Pixel Format      : 'YUYV' (YUYV 4:2:2)
+```
+
+Troubleshooting: If the resolution is not correct you will have to delete the Virtual Camera port and try again as it cannot be changed.
+
+If everything is set up correctly, you can proceed with the rest of the tutorial.
 
 </hfoption>
 </hfoptions>
-
-If everything is set up correctly, your phone will appear as a standard OpenCV camera and can be used with `OpenCVCamera`.

docs/source/damiao.mdx

@@ -1,165 +0,0 @@
-# Damiao Motors and CAN Bus
-
-This guide covers setup and usage of Damiao motors with LeRobot via CAN bus communication.
-
-Currently, only Linux is supported, as the OpenArms CAN adapter only has drivers for Linux.
-
-## Linux CAN Setup
-
-Before using Damiao motors, you need to set up the CAN interface on your Linux system.
-
-### Install CAN Utilities
-
-```bash
-sudo apt-get install can-utils
-```
-
-### Configure CAN Interface (Manual)
-
-For standard CAN FD (recommended for OpenArms):
-
-```bash
-sudo ip link set can0 down
-sudo ip link set can0 type can bitrate 1000000 dbitrate 5000000 fd on
-sudo ip link set can0 up
-```
-
-For standard CAN (without FD):
-
-```bash
-sudo ip link set can0 down
-sudo ip link set can0 type can bitrate 1000000
-sudo ip link set can0 up
-```
-
-### Configure CAN Interface (Using LeRobot)
-
-LeRobot provides a utility script to setup and test CAN interfaces:
-
-```bash
-# Setup multiple interfaces (e.g., OpenArms Followers with 2 CAN buses)
-lerobot-setup-can --mode=setup --interfaces=can0,can1
-```
-
-## Debugging CAN Communication
-
-Use the built-in debug tools to test motor communication:
-
-```bash
-# Test motors on all interfaces
-lerobot-setup-can --mode=test --interfaces=can0,can1
-
-# Run speed/latency test
-lerobot-setup-can --mode=speed --interfaces=can0
-```
-
-The test mode will scan for motors (IDs 0x01-0x08) and report which ones respond. Example output:
-
-```
-can0: UP (CAN FD)
-  Motor 0x01 (joint_1): ✓ FOUND
-    → Response 0x11 [FD]: 00112233...
-  Motor 0x02 (joint_2): ✓ FOUND
-  Motor 0x03 (joint_3): ✗ No response
-  ...
-  Summary: 2/8 motors found
-```
-
-## Usage
-
-### Basic Setup
-
-```python
-from lerobot.motors import Motor
-from lerobot.motors.damiao import DamiaoMotorsBus
-
-# Define your motors with send/receive CAN IDs
-motors = {
-    "joint_1": Motor(id=0x01, motor_type_str="dm8009", recv_id=0x11),
-    "joint_2": Motor(id=0x02, motor_type_str="dm4340", recv_id=0x12),
-    "joint_3": Motor(id=0x03, motor_type_str="dm4310", recv_id=0x13),
-}
-
-# Create the bus
-bus = DamiaoMotorsBus(
-    port="can0",  # Linux socketcan interface
-    motors=motors,
-)
-
-# Connect
-bus.connect()
-```
-
-### Reading Motor States
-
-```python
-# Read single motor position (degrees)
-position = bus.read("Present_Position", "joint_1")
-
-# Read from multiple motors
-positions = bus.sync_read("Present_Position")  # All motors
-positions = bus.sync_read("Present_Position", ["joint_1", "joint_2"])
-
-# Read all states at once (position, velocity, torque)
-states = bus.sync_read_all_states()
-# Returns: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
-```
-
-### Writing Motor Commands
-
-```python
-# Enable torque
-bus.enable_torque()
-
-# Set goal position (degrees)
-bus.write("Goal_Position", "joint_1", 45.0)
-
-# Set positions for multiple motors
-bus.sync_write("Goal_Position", {
-    "joint_1": 45.0,
-    "joint_2": -30.0,
-    "joint_3": 90.0,
-})
-
-# Disable torque
-bus.disable_torque()
-```
-
-## Configuration Options
-
-| Parameter      | Default   | Description                                                 |
-| -------------- | --------- | ----------------------------------------------------------- |
-| `port`         | -         | CAN interface (`can0`) or serial port (`/dev/cu.usbmodem*`) |
-| `use_can_fd`   | `True`    | Enable CAN FD for higher data rates                         |
-| `bitrate`      | `1000000` | Nominal bitrate (1 Mbps)                                    |
-| `data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)                                |
-
-## Motor Configuration
-
-Each motor requires:
-
-- `id`: CAN ID for sending commands
-- `motor_type`: One of the supported motor types (e.g., `"dm8009"`, `"dm4340"`)
-- `recv_id`: CAN ID for receiving responses
-
-OpenArms default IDs follow the pattern: send ID `0x0N`, receive ID `0x1N` where N is the joint number.
-
-## Troubleshooting
-
-### No Response from Motors
-
-1. **Check power**
-2. **Verify CAN wiring**: Check CAN-H, CAN-L, and GND connections
-3. **Check motor IDs**: Use Damiao Debugging Tools to verify/configure IDs
-4. **Test CAN interface**: Run `candump can0` to see if messages are being received
-5. **Run diagnostics**: `lerobot-setup-can --mode=test --interfaces=can0`
-
-### Motor Timeout Parameter
-
-If motors were configured with timeout=0, they won't respond to commands. Use Damiao Debugging Tools to set a non-zero timeout value.
-
-### Verify CAN FD Status
-
-```bash
-ip -d link show can0 | grep fd
-```

docs/source/dataset_subtask.mdx

@@ -1,278 +0,0 @@
-# Using Subtasks in LeRobot Datasets
-
-Subtask support in robotics datasets has proven effective in improving robot reasoning and understanding. Subtasks are particularly useful for:
-
-- **Hierarchical policies**: Building policies that include subtask predictions to visualize robot reasoning in real time
-- **Reward modeling**: Helping reward models understand task progression (e.g., SARM-style stage-aware reward models)
-- **Task decomposition**: Breaking down complex manipulation tasks into atomic, interpretable steps
-
-LeRobotDataset now supports subtasks as part of its dataset structure, alongside tasks.
-
-## What are Subtasks?
-
-While a **task** describes the overall goal (e.g., "Pick up the apple and place it in the basket"), **subtasks** break down the execution into finer-grained steps:
-
-1. "Approach the apple"
-2. "Grasp the apple"
-3. "Lift the apple"
-4. "Move to basket"
-5. "Release the apple"
-
-Each frame in the dataset can be annotated with its corresponding subtask, enabling models to learn and predict these intermediate stages.
-
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/subtask-asset.png"
-  alt="An overview of subtask annotation showing how frames are labeled with intermediate subtask stages"
-  width="80%"
-/>
-
-<p>
-  <em>Figure: Overview of subtask annotation.</em>
-</p>
-
-**Reference:** _Subtask-learning based for robot self-assembly in flexible collaborative assembly in manufacturing_, Original Article, Published: 19 April 2022.
-
-## Dataset Structure
-
-Subtask information is stored in the dataset metadata:
-
-```
-my-dataset/
-├── data/
-│   └── ...
-├── meta/
-│   ├── info.json
-│   ├── stats.json
-│   ├── tasks.parquet
-│   ├── subtasks.parquet      # Subtask index → subtask string mapping
-│   └── episodes/
-│       └── ...
-└── videos/
-    └── ...
-```
-
-### Subtasks Parquet File
-
-The `meta/subtasks.parquet` file maps subtask indices to their natural language descriptions:
-
-| subtask_index | subtask (index column) |
-| ------------- | ---------------------- |
-| 0             | "Approach the apple"   |
-| 1             | "Grasp the apple"      |
-| 2             | "Lift the apple"       |
-| ...           | ...                    |
-
-### Frame-Level Annotations
-
-Each frame in the dataset can include a `subtask_index` field that references the subtasks parquet file:
-
-```python
-# Example frame data in the parquet file
-{
-    "index": 42,
-    "timestamp": 1.4,
-    "episode_index": 0,
-    "task_index": 0,
-    "subtask_index": 2,  # References "Lift the apple"
-    "observation.state": [...],
-    "action": [...],
-}
-```
-
-## Annotating Datasets with Subtasks
-
-We provide a HuggingFace Space for easily annotating any LeRobotDataset with subtasks:
-
-**[https://huggingface.co/spaces/lerobot/annotate](https://huggingface.co/spaces/lerobot/annotate)**
-
-After completing your annotation:
-
-1. Click "Push to Hub" to upload your annotated dataset
-2. You can also run the annotation space locally by following the instructions at [github.com/huggingface/lerobot-annotate](https://github.com/huggingface/lerobot-annotate)
-
-## Loading Datasets with Subtasks
-
-When you load a dataset with subtask annotations, the subtask information is automatically available:
-
-```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-# Load a dataset with subtask annotations
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-# Access a sample
-sample = dataset[100]
-
-# The sample includes both task and subtask information
-print(sample["task"])        # "Collect the fruit"
-print(sample["subtask"])     # "Grasp the apple"
-print(sample["task_index"])  # tensor(0)
-print(sample["subtask_index"])  # tensor(2)
-```
-
-### Checking for Subtask Support
-
-You can check if a dataset has subtask annotations:
-
-```python
-# Check if subtasks are available
-has_subtasks = (
-    "subtask_index" in dataset.features
-    and dataset.meta.subtasks is not None
-)
-
-if has_subtasks:
-    print(f"Dataset has {len(dataset.meta.subtasks)} unique subtasks")
-    print("Subtasks:", list(dataset.meta.subtasks.index))
-```
-
-## Using Subtasks for Training
-
-### With the Tokenizer Processor
-
-The `TokenizerProcessor` automatically handles subtask tokenization for Vision-Language Action (VLA) models:
-
-```python
-from lerobot.processor.tokenizer_processor import TokenizerProcessor
-from lerobot.processor.pipeline import ProcessorPipeline
-
-# Create a tokenizer processor
-tokenizer_processor = TokenizerProcessor(
-    tokenizer_name_or_path="google/paligemma-3b-pt-224",
-    padding="max_length",
-    max_length=64,
-)
-
-# The processor will automatically tokenize subtasks if present in the batch
-# and add them to the observation under:
-# - "observation.subtask.tokens"
-# - "observation.subtask.attention_mask"
-```
-
-When subtasks are available in the batch, the tokenizer processor adds:
-
-- `observation.subtask.tokens`: Tokenized subtask text
-- `observation.subtask.attention_mask`: Attention mask for the subtask tokens
-
-### DataLoader with Subtasks
-
-```python
-import torch
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-dataloader = torch.utils.data.DataLoader(
-    dataset,
-    batch_size=16,
-    shuffle=True,
-)
-
-for batch in dataloader:
-    # Access subtask information in the batch
-    subtasks = batch["subtask"]  # List of subtask strings
-    subtask_indices = batch["subtask_index"]  # Tensor of subtask indices
-
-    # Use for training hierarchical policies or reward models
-    print(f"Batch subtasks: {set(subtasks)}")
-```
-
-## Example Datasets with Subtask Annotations
-
-Try loading a dataset with subtask annotations:
-
-```python
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-
-# Example dataset with subtask annotations
-dataset = LeRobotDataset("jadechoghari/collect-fruit-annotated")
-
-# Explore the subtasks
-print("Available subtasks:")
-for subtask_name in dataset.meta.subtasks.index:
-    print(f"  - {subtask_name}")
-
-# Get subtask distribution
-subtask_counts = {}
-for i in range(len(dataset)):
-    sample = dataset[i]
-    subtask = sample["subtask"]
-    subtask_counts[subtask] = subtask_counts.get(subtask, 0) + 1
-
-print("\nSubtask distribution:")
-for subtask, count in sorted(subtask_counts.items(), key=lambda x: -x[1]):
-    print(f"  {subtask}: {count} frames")
-```
-
-## Use Cases
-
-### 1. Hierarchical Policy Training
-
-Train policies that predict both actions and current subtask:
-
-```python
-class HierarchicalPolicy(nn.Module):
-    def __init__(self, num_subtasks):
-        super().__init__()
-        self.action_head = nn.Linear(hidden_dim, action_dim)
-        self.subtask_head = nn.Linear(hidden_dim, num_subtasks)
-
-    def forward(self, observations):
-        features = self.encoder(observations)
-        actions = self.action_head(features)
-        subtask_logits = self.subtask_head(features)
-        return actions, subtask_logits
-```
-
-### 2. Stage-Aware Reward Modeling (SARM)
-
-Build reward models that understand task progression:
-
-```python
-# SARM predicts:
-# - Stage: Which subtask is being executed (discrete)
-# - Progress: How far along the subtask (continuous 0-1)
-
-class SARMRewardModel(nn.Module):
-    def forward(self, observations):
-        features = self.encoder(observations)
-        stage_logits = self.stage_classifier(features)
-        progress = self.progress_regressor(features)
-        return stage_logits, progress
-```
-
-### 3. Progress Visualization
-
-Monitor robot execution by tracking subtask progression:
-
-```python
-def visualize_execution(model, observations):
-    for t, obs in enumerate(observations):
-        action, subtask_logits = model(obs)
-        predicted_subtask = subtask_names[subtask_logits.argmax()]
-        print(f"t={t}: Executing '{predicted_subtask}'")
-```
-
-## API Reference
-
-### LeRobotDataset Properties
-
-| Property                    | Type                   | Description                                |
-| --------------------------- | ---------------------- | ------------------------------------------ |
-| `meta.subtasks`             | `pd.DataFrame \| None` | DataFrame mapping subtask names to indices |
-| `features["subtask_index"]` | `dict`                 | Feature spec for subtask_index if present  |
-
-### Sample Keys
-
-When subtasks are available, each sample includes:
-
-| Key             | Type           | Description                          |
-| --------------- | -------------- | ------------------------------------ |
-| `subtask_index` | `torch.Tensor` | Integer index of the current subtask |
-| `subtask`       | `str`          | Natural language subtask description |
-
-## Related Resources
-
-- [SARM Paper](https://arxiv.org/pdf/2509.25358) - Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
-- [LeRobot Annotate Space](https://huggingface.co/spaces/lerobot/annotate) - Interactive annotation tool
-- [LeRobotDataset v3.0](./lerobot-dataset-v3) - Dataset format documentation

docs/source/earthrover_mini_plus.mdx

@@ -1,11 +1,5 @@
 # EarthRover Mini Plus
 
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/Earth_Rover_Mini_5_240c9adc-4f9e-44b7-982f-5d1dc24af1d8.png.webp"
-  alt="EarthRover Mini Plus"
-  width="70%"
-/>
-
 The EarthRover Mini Plus is a fully open source mobile robot that connects through the cloud using the Frodobots SDK. This lets you control the robot and record datasets for training AI models.
 
 ## What You Need

docs/source/installation.mdx

@@ -1,15 +1,13 @@
 # Installation
 
-This guide uses conda (via miniforge) to manage environments. If you prefer another environment manager (e.g. `uv`, `venv`), ensure you have Python >=3.10 and ffmpeg installed with the `libsvtav1` encoder, then skip ahead to [Install LeRobot](#step-3-install-lerobot-).
-
-## Step 1: Install [`miniforge`](https://conda-forge.org/download/)
+## Install [`miniforge`](https://conda-forge.org/download/)
 
 ```bash
 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
 bash Miniforge3-$(uname)-$(uname -m).sh
 ```
 
-## Step 2: Environment Setup
+## Environment Setup
 
 Create a virtual environment with Python 3.10, using conda:
 
@@ -40,7 +38,7 @@ conda install ffmpeg -c conda-forge
 >
 > - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
 
-## Step 3: Install LeRobot 🤗
+## Install LeRobot 🤗
 
 ### From Source
 

docs/source/lekiwi.mdx

@@ -1,11 +1,5 @@
 # LeKiwi
 
-<img
-  src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/1740517739083.jpeg"
-  alt="LeKiwi"
-  width="70%"
-/>
-
 In the steps below, we explain how to assemble the LeKiwi mobile robot.
 
 ## Source the parts

docs/source/libero.mdx

@@ -42,7 +42,6 @@ lerobot-eval \
 ```
 
 - `--env.task` picks the suite (`libero_object`, `libero_spatial`, etc.).
-- `--env.task_ids` picks task ids to run (`[0]`, `[1,2,3]`, etc.). Omit this flag (or set it to `null`) to run all tasks in the suite.
 - `--eval.batch_size` controls how many environments run in parallel.
 - `--eval.n_episodes` sets how many episodes to run in total.
 

docs/source/openarm.mdx

@@ -1,276 +0,0 @@
-# OpenArm
-
-[OpenArm](https://openarm.dev) is an open-source 7DOF humanoid arm designed for physical AI research and deployment.
-
-To get your OpenArm, assembled or DIY, and join the global community, browse verified and certified manufacturers worldwide at [openarm.dev](https://openarm.dev).
-
-## What's Unique?
-
-- **Human-Scale Design**: OpenArm is designed with human-like proportions, scaled for a person around 160-165cm tall. This provides an optimal balance between practical reach and manageable inertia for safe, responsive operation.
-
-- **Safety-First Architecture**: Built with QDD backdrivable motors and high compliance, OpenArm prioritizes safe human-robot interaction while maintaining practical payload capabilities (6.0kg peak / 4.1kg nominal) for real-world tasks.
-
-- **Built for Durability**: Critical structural components use aluminum and stainless steel construction, ensuring robust performance for repetitive data collection and continuous research use.
-
-- **Fully Accessible & Buildable**: Every component, from CNC parts and 3D-printed casings to electrical wiring is designed to be purchasable and buildable by individual researchers and labs, with complete fabrication data provided.
-
-- **Practical & Affordable**: At $6,500 USD for a complete bimanual system, OpenArm delivers research-grade capabilities at a fraction of traditional humanoid robot costs.
-
-## Platform Requirements
-
-<Tip warning={true}>
-  **Linux Only**: OpenArm currently only works on Linux. The CAN bus USB adapter
-  does not have macOS drivers and has not been tested on Windows.
-</Tip>
-
-## Safety Guide
-
-Before operating OpenArm, please read the [official safety guide](https://docs.openarm.dev/getting-started/safety-guide). Key points:
-
-- **Secure installation**: Fasten the arm to a flat, stable surface with screws or clamps
-- **Safe distance**: Keep body parts and objects outside the range of motion during operation
-- **Protective equipment**: Always wear safety goggles; use additional PPE as needed
-- **Payload limits**: Do not exceed specified payload limits (6.0kg peak / 4.1kg nominal per arm)
-- **Emergency stop**: Know the location and operation of the emergency stop device
-- **Regular inspection**: Check for loose screws, damaged mechanical limits, unusual noises, and wiring damage
-
-## Hardware Setup
-
-Follow the official [OpenArm hardware documentation](https://docs.openarm.dev) for:
-
-- Bill of materials and sourcing
-- 3D printing instructions
-- Mechanical assembly
-- Electrical wiring
-
-The hardware repositories are available at [github.com/enactic/openarm](https://github.com/enactic/openarm).
-
-## CAN Bus Setup
-
-OpenArm uses CAN bus communication with Damiao motors. Once you have the CAN bus USB adapter plugged into your Linux PC, follow the [Damiao Motors and CAN Bus guide](./damiao) to configure the interface.
-
-Quick setup:
-
-```bash
-# Setup CAN interfaces
-lerobot-setup-can --mode=setup --interfaces=can0,can1
-
-# Test motor communication
-lerobot-setup-can --mode=test --interfaces=can0,can1
-```
-
-## Install LeRobot 🤗
-
-Follow our [Installation Guide](./installation), then install the Damiao motor support:
-
-```bash
-pip install -e ".[damiao]"
-```
-
-## Usage
-
-### Follower Arm (Robot)
-
-<hfoptions id="follower">
-<hfoption id="Command">
-
-```bash
-lerobot-calibrate \
-    --robot.type=openarm_follower \
-    --robot.port=can0 \
-    --robot.side=right \
-    --robot.id=my_openarm_follower
-```
-
-</hfoption>
-<hfoption id="API example">
-
-```python
-from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
-
-config = OpenArmFollowerConfig(
-    port="can0",
-    side="right",  # or "left" for left arm
-    id="my_openarm_follower",
-)
-
-follower = OpenArmFollower(config)
-follower.connect()
-
-# Read current state
-obs = follower.get_observation()
-print(obs)
-
-# Send action (position in degrees)
-action = {
-    "joint_1.pos": 0.0,
-    "joint_2.pos": 0.0,
-    "joint_3.pos": 0.0,
-    "joint_4.pos": 45.0,
-    "joint_5.pos": 0.0,
-    "joint_6.pos": 0.0,
-    "joint_7.pos": 0.0,
-    "gripper.pos": 0.0,
-}
-follower.send_action(action)
-
-follower.disconnect()
-```
-
-</hfoption>
-</hfoptions>
-
-### Leader Arm (Teleoperator)
-
-The leader arm is used for teleoperation - manually moving it to control the follower arm.
-
-<hfoptions id="leader">
-<hfoption id="Command">
-
-```bash
-lerobot-calibrate \
-    --teleop.type=openarm_leader \
-    --teleop.port=can1 \
-    --teleop.id=my_openarm_leader
-```
-
-</hfoption>
-<hfoption id="API example">
-
-```python
-from lerobot.teleoperators.openarm_leader import OpenArmLeader, OpenArmLeaderConfig
-
-config = OpenArmLeaderConfig(
-    port="can1",
-    id="my_openarm_leader",
-    manual_control=True,  # Disable torque for manual movement
-)
-
-leader = OpenArmLeader(config)
-leader.connect()
-
-# Read current position (as action to send to follower)
-action = leader.get_action()
-print(action)
-
-leader.disconnect()
-```
-
-</hfoption>
-</hfoptions>
-
-### Teleoperation
-
-To teleoperate OpenArm with leader-follower control:
-
-```bash
-lerobot-teleoperate \
-    --robot.type=openarm_follower \
-    --robot.port=can0 \
-    --robot.side=right \
-    --robot.id=my_follower \
-    --teleop.type=openarm_leader \
-    --teleop.port=can1 \
-    --teleop.id=my_leader
-```
-
-### Bimanual Teleoperation
-
-To teleoperate a bimanual OpenArm setup with two leader and two follower arms:
-
-```bash
-lerobot-teleoperate \
-    --robot.type=bi_openarm_follower \
-    --robot.left_arm_config.port=can0 \
-    --robot.left_arm_config.side=left \
-    --robot.right_arm_config.port=can1 \
-    --robot.right_arm_config.side=right \
-    --robot.id=my_bimanual_follower \
-    --teleop.type=bi_openarm_leader \
-    --teleop.left_arm_config.port=can2 \
-    --teleop.right_arm_config.port=can3 \
-    --teleop.id=my_bimanual_leader
-```
-
-### Recording Data
-
-To record a dataset during teleoperation:
-
-```bash
-lerobot-record \
-    --robot.type=openarm_follower \
-    --robot.port=can0 \
-    --robot.side=right \
-    --robot.id=my_follower \
-    --teleop.type=openarm_leader \
-    --teleop.port=can1 \
-    --teleop.id=my_leader \
-    --repo-id=my_hf_username/my_openarm_dataset \
-    --fps=30 \
-    --num-episodes=10
-```
-
-## Configuration Options
-
-### Follower Configuration
-
-| Parameter             | Default   | Description                                                |
-| --------------------- | --------- | ---------------------------------------------------------- |
-| `port`                | -         | CAN interface (e.g., `can0`)                               |
-| `side`                | `None`    | Arm side: `"left"`, `"right"`, or `None` for custom limits |
-| `use_can_fd`          | `True`    | Enable CAN FD for higher data rates                        |
-| `can_bitrate`         | `1000000` | Nominal bitrate (1 Mbps)                                   |
-| `can_data_bitrate`    | `5000000` | CAN FD data bitrate (5 Mbps)                               |
-| `max_relative_target` | `None`    | Safety limit for relative target positions                 |
-| `position_kp`         | Per-joint | Position control proportional gains                        |
-| `position_kd`         | Per-joint | Position control derivative gains                          |
-
-### Leader Configuration
-
-| Parameter          | Default   | Description                         |
-| ------------------ | --------- | ----------------------------------- |
-| `port`             | -         | CAN interface (e.g., `can1`)        |
-| `manual_control`   | `True`    | Disable torque for manual movement  |
-| `use_can_fd`       | `True`    | Enable CAN FD for higher data rates |
-| `can_bitrate`      | `1000000` | Nominal bitrate (1 Mbps)            |
-| `can_data_bitrate` | `5000000` | CAN FD data bitrate (5 Mbps)        |
-
-## Motor Configuration
-
-OpenArm uses Damiao motors with the following default configuration:
-
-| Joint                       | Motor Type | Send ID | Recv ID |
-| --------------------------- | ---------- | ------- | ------- |
-| joint_1 (Shoulder pan)      | DM8009     | 0x01    | 0x11    |
-| joint_2 (Shoulder lift)     | DM8009     | 0x02    | 0x12    |
-| joint_3 (Shoulder rotation) | DM4340     | 0x03    | 0x13    |
-| joint_4 (Elbow flex)        | DM4340     | 0x04    | 0x14    |
-| joint_5 (Wrist roll)        | DM4310     | 0x05    | 0x15    |
-| joint_6 (Wrist pitch)       | DM4310     | 0x06    | 0x16    |
-| joint_7 (Wrist rotation)    | DM4310     | 0x07    | 0x17    |
-| gripper                     | DM4310     | 0x08    | 0x18    |
-
-## Troubleshooting
-
-### No Response from Motors
-
-1. Check power supply connections
-2. Verify CAN wiring (CAN-H, CAN-L, GND)
-3. Run diagnostics: `lerobot-setup-can --mode=test --interfaces=can0`
-4. See the [Damiao troubleshooting guide](./damiao#troubleshooting) for more details
-
-### CAN Interface Not Found
-
-Ensure the CAN interface is configured:
-
-```bash
-ip link show can0
-```
-
-## Resources
-
-- [OpenArm Website](https://openarm.dev)
-- [OpenArm Documentation](https://docs.openarm.dev)
-- [OpenArm GitHub](https://github.com/enactic/openarm)
-- [Safety Guide](https://docs.openarm.dev/getting-started/safety-guide)
-- [Damiao Motors and CAN Bus](./damiao)

docs/source/sarm.mdx

@@ -465,15 +465,15 @@ This script:
 
 ### Step 5b: Train Policy with RA-BC
 
-Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`). Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
+Once you have the progress file, train your policy with RA-BC weighting. The progress file is auto-detected from the dataset path (`sarm_progress.parquet`) if not explicitly provided. Currently PI0, PI0.5 and SmolVLA are supported with RA-BC:
 
 ```bash
 python src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
-  --use_rabc=true \
-  --rabc_head_mode=sparse \
-  --rabc_kappa=0.01 \
+  --sample_weighting.type=rabc \
+  --sample_weighting.head_mode=sparse \
+  --sample_weighting.kappa=0.01 \
   --output_dir=outputs/train/policy_rabc \
   --batch_size=32 \
   --steps=40000
@@ -488,12 +488,13 @@ The training script automatically:
 
 **RA-BC Arguments:**
 
-| Argument               | Description                                                | Default                            |
-| ---------------------- | ---------------------------------------------------------- | ---------------------------------- |
-| `--use_rabc`           | Enable RA-BC sample weighting                              | `false`                            |
-| `--rabc_progress_path` | Path to progress parquet file (auto-detected from dataset) | `sarm_progress.parquet` in dataset |
-| `--rabc_head_mode`     | Which SARM head's progress to use: `sparse` or `dense`     | `sparse`                           |
-| `--rabc_kappa`         | Threshold κ for high-quality samples                       | `0.01`                             |
+| Argument                           | Description                                            | Default                 |
+| ---------------------------------- | ------------------------------------------------------ | ----------------------- |
+| `--sample_weighting.type`          | Weighting strategy type (`rabc` or `uniform`)          | `rabc`                  |
+| `--sample_weighting.progress_path` | Path to progress parquet file                          | `sarm_progress.parquet` |
+| `--sample_weighting.head_mode`     | Which SARM head's progress to use: `sparse` or `dense` | `sparse`                |
+| `--sample_weighting.kappa`         | Threshold κ for high-quality samples                   | `0.01`                  |
+| `--sample_weighting.epsilon`       | Small constant for numerical stability                 | `1e-6`                  |
 
 ### Tuning RA-BC Kappa
 
@@ -511,30 +512,30 @@ The `kappa` parameter is the threshold that determines which samples get full we
 
 Monitor these WandB metrics during training:
 
-| Metric             | Healthy Range | Problem Indicator         |
-| ------------------ | ------------- | ------------------------- |
-| `rabc_mean_weight` | 0.3 - 0.8     | ≈ 1.0 means kappa too low |
-| `rabc_delta_mean`  | > 0           | Should be positive        |
-| `rabc_delta_std`   | > 0           | Variance in data quality  |
+| Metric                        | Healthy Range | Problem Indicator         |
+| ----------------------------- | ------------- | ------------------------- |
+| `sample_weight_mean_weight`   | 0.3 - 0.8     | ≈ 1.0 means kappa too low |
+| `sample_weighting/delta_mean` | > 0           | Should be positive        |
+| `sample_weighting/delta_std`  | > 0           | Variance in data quality  |
 
-**If `rabc_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.
+**If `sample_weight_mean_weight ≈ 1.0`:** Your kappa is too low. Most samples have `delta > kappa` and bypass the soft-weighting entirely. RA-BC becomes equivalent to vanilla BC.
 
 **Setting kappa based on your data:**
 
-The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `rabc_delta_mean` and `rabc_delta_std`:
+The default `kappa=0.01` was tuned for the paper's T-shirt folding task (~90s episodes at 30fps). For your dataset, check the logged `sample_weighting/delta_mean` and `sample_weighting/delta_std`:
 
 ```
 # If delta_mean ≈ 0.03 and delta_std ≈ 0.02:
 # Most deltas fall in range [0.01, 0.05]
 
 # Option 1: Set kappa = delta_mean (medium selectivity)
---rabc_kappa=0.03
+--sample_weighting.kappa=0.03
 
 # Option 2: Set kappa = delta_mean + delta_std (high selectivity)
---rabc_kappa=0.05
+--sample_weighting.kappa=0.05
 
 # Option 3: Set kappa = delta_mean + 2*delta_std (very selective)
---rabc_kappa=0.07
+--sample_weighting.kappa=0.07
 ```
 
 **When RA-BC may not help:**
@@ -550,8 +551,8 @@ accelerate launch \
   src/lerobot/scripts/lerobot_train.py \
   --dataset.repo_id=your-username/your-dataset \
   --policy.type=pi0 \
-  --use_rabc=true \
-  --rabc_kappa=0.01 \
+  --sample_weighting.type=rabc \
+  --sample_weighting.kappa=0.01 \
   --output_dir=outputs/train/policy_rabc \
   --batch_size=32 \
   --steps=40000
@@ -576,7 +577,7 @@ accelerate launch \
 ### RA-BC
 
 1. **Train SARM first**: RA-BC quality depends entirely on SARM quality
-2. **Monitor `rabc_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
+2. **Monitor `sample_weight_mean_weight`**: If it's ≈ 1.0, increase kappa (see [Tuning RA-BC Kappa](#tuning-ra-bc-kappa))
 
 ---
 

docs/source/so101.mdx

@@ -1,18 +1,5 @@
 # SO-101
 
-<div style="display: flex; align-items: center; gap: 10px;">
-  <img
-    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Follower.webp"
-    alt="SO-101"
-    width="60%"
-  />
-  <img
-    src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/SO101_Leader.webp"
-    alt="SO-101"
-    width="60%"
-  />
-</div>
-
 In the steps below, we explain how to assemble our flagship robot, the SO-101.
 
 ## Source the parts

docs/source/unitree_g1.mdx

@@ -188,105 +188,7 @@ Press `Ctrl+C` to stop the policy.
 
 ## Running in Simulation Mode (MuJoCo)
 
-You can test policies before deploying on the physical robot using MuJoCo simulation. Set `is_simulation=True` in config or pass `--robot.is_simulation=true` via CLI.
-
-### Calibrate Exoskeleton Teleoperator
-
-```bash
-lerobot-calibrate \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo
-```
-
-### Teleoperate in Simulation
-
-```bash
-lerobot-teleoperate \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=true \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --fps=100
-```
-
-### Record Dataset in Simulation
-
-```bash
-python -m lerobot.scripts.lerobot_record \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=true \
-    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "localhost", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --dataset.repo_id=your-username/dataset-name \
-    --dataset.single_task="Test" \
-    --dataset.num_episodes=2 \
-    --dataset.episode_time_s=5 \
-    --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true
-```
-
-Example simulation dataset: [nepyope/teleop_test_sim](https://huggingface.co/datasets/nepyope/teleop_test_sim)
-
----
-
-## Running on Real Robot
-
-Once the robot server is running on the G1 (see Part 3), you can teleoperate and record on the real robot.
-
-### Start the Camera Server
-
-On the robot, start the ZMQ image server:
-
-```bash
-python src/lerobot/cameras/zmq/image_server.py
-```
-
-Keep this running in a separate terminal for camera streaming during recording.
-
-### Teleoperate Real Robot
-
-```bash
-lerobot-teleoperate \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=false \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --fps=100
-```
-
-### Record Dataset on Real Robot
-
-```bash
-python -m lerobot.scripts.lerobot_record \
-    --robot.type=unitree_g1 \
-    --robot.is_simulation=false \
-    --robot.cameras='{"global_view": {"type": "zmq", "server_address": "172.18.129.215", "port": 5555, "camera_name": "head_camera", "width": 640, "height": 480, "fps": 30}}' \
-    --teleop.type=unitree_g1 \
-    --teleop.left_arm_config.port=/dev/ttyACM1 \
-    --teleop.right_arm_config.port=/dev/ttyACM0 \
-    --teleop.id=exo \
-    --dataset.repo_id=your-username/dataset-name \
-    --dataset.single_task="Test" \
-    --dataset.num_episodes=2 \
-    --dataset.episode_time_s=5 \
-    --dataset.reset_time_s=5 \
-    --dataset.push_to_hub=true
-```
-
-**Note**: Update `server_address` to match your robot's camera server IP.
-
-Example real robot dataset: [nepyope/teleop_test_real](https://huggingface.co/datasets/nepyope/teleop_test_real)
-
----
+You can now test policies before unleashing them on the physical robot using MuJoCo. To do so simply set `is_simulation=True` in config.
 
 ## Additional Resources
 

examples/backward_compatibility/replay.py

@@ -81,25 +81,24 @@ def replay(cfg: ReplayConfig):
     actions = dataset.hf_dataset.select_columns(ACTION)
     robot.connect()
 
-    try:
-        log_say("Replaying episode", cfg.play_sounds, blocking=True)
-        for idx in range(dataset.num_frames):
-            start_episode_t = time.perf_counter()
+    log_say("Replaying episode", cfg.play_sounds, blocking=True)
+    for idx in range(dataset.num_frames):
+        start_episode_t = time.perf_counter()
 
-            action_array = actions[idx][ACTION]
-            action = {}
-            for i, name in enumerate(dataset.features[ACTION]["names"]):
-                key = f"{name.removeprefix('main_')}.pos"
-                action[key] = action_array[i].item()
+        action_array = actions[idx][ACTION]
+        action = {}
+        for i, name in enumerate(dataset.features[ACTION]["names"]):
+            key = f"{name.removeprefix('main_')}.pos"
+            action[key] = action_array[i].item()
 
-            action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
-            action["elbow_flex.pos"] -= 90
-            robot.send_action(action)
+        action["shoulder_lift.pos"] = -(action["shoulder_lift.pos"] - 90)
+        action["elbow_flex.pos"] -= 90
+        robot.send_action(action)
 
-            dt_s = time.perf_counter() - start_episode_t
-            precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
-    finally:
-        robot.disconnect()
+        dt_s = time.perf_counter() - start_episode_t
+        precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
+
+    robot.disconnect()
 
 
 if __name__ == "__main__":

examples/lekiwi/evaluate.py

@@ -78,24 +78,40 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="lekiwi_evaluate")
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    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}")
+    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
+        # 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=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            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,
-                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,
@@ -104,42 +120,24 @@ 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()
 
-    finally:
-        # Clean up
-        log_say("Stop recording")
-        robot.disconnect()
-        listener.stop()
-
-        dataset.finalize()
-        dataset.push_to_hub()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/lekiwi/record.py

@@ -74,23 +74,40 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="lekiwi_record")
 
-    try:
-        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 record loop...")
-        recorded_episodes = 0
-        while recorded_episodes < NUM_EPISODES and not events["stop_recording"]:
-            log_say(f"Recording episode {recorded_episodes}")
+    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
+        # Main record loop
+        record_loop(
+            robot=robot,
+            events=events,
+            fps=FPS,
+            dataset=dataset,
+            teleop=[leader_arm, keyboard],
+            control_time_s=EPISODE_TIME_SEC,
+            single_task=TASK_DESCRIPTION,
+            display_data=True,
+            teleop_action_processor=teleop_action_processor,
+            robot_action_processor=robot_action_processor,
+            robot_observation_processor=robot_observation_processor,
+        )
+
+        # 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,
@@ -98,44 +115,26 @@ 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
-    finally:
-        # 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()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/lekiwi/replay.py

@@ -42,27 +42,25 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    try:
-        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)
 
-            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-    finally:
-        robot.disconnect()
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+
+    robot.disconnect()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/evaluate.py

@@ -142,24 +142,38 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="phone_so100_evaluate")
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    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}")
+    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
+        # 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,
+        )
+
+        # 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,
@@ -168,41 +182,24 @@ 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
-    finally:
-        # 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()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/record.py

@@ -149,23 +149,38 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="phone_so100_record")
 
-    try:
-        if not robot.is_connected or not phone.is_connected:
-            raise ValueError("Robot or teleop is not connected!")
+    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}")
+    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
+        # 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,
+        )
+
+        # 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,
@@ -173,43 +188,25 @@ 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
-    finally:
-        # 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()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/phone_to_so100/replay.py

@@ -73,34 +73,32 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    try:
-        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)
 
-            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
-    finally:
-        # Clean up
-        robot.disconnect()
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+
+    # Clean up
+    robot.disconnect()
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/evaluate.py

@@ -142,24 +142,38 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="so100_so100_evaluate")
 
-    try:
-        if not robot.is_connected:
-            raise ValueError("Robot is not connected!")
+    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}")
+    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
+        # 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,
+        )
+
+        # 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,
@@ -168,41 +182,24 @@ 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
-    finally:
-        # 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()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/record.py

@@ -146,23 +146,38 @@ def main():
     listener, events = init_keyboard_listener()
     init_rerun(session_name="recording_phone")
 
-    try:
-        if not leader.is_connected or not follower.is_connected:
-            raise ValueError("Robot or teleop is not connected!")
+    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}")
+    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
+        # 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,
+        )
+
+        # 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,
@@ -170,44 +185,25 @@ 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()
 
-    finally:
-        # Clean up
-        log_say("Stop recording")
-        leader.disconnect()
-        follower.disconnect()
-        listener.stop()
-
-        dataset.finalize()
-        dataset.push_to_hub()
+    dataset.finalize()
+    dataset.push_to_hub()
 
 
 if __name__ == "__main__":

examples/so100_to_so100_EE/replay.py

@@ -74,35 +74,32 @@ def main():
     # Connect to the robot
     robot.connect()
 
-    try:
-        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)
 
-            precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
+        precise_sleep(max(1.0 / dataset.fps - (time.perf_counter() - t0), 0.0))
 
-    finally:
-        # Clean up
-        robot.disconnect()
+    # Clean up
+    robot.disconnect()
 
 
 if __name__ == "__main__":

pyproject.toml

@@ -102,20 +102,14 @@ grpcio-dep = ["grpcio==1.73.1", "protobuf>=6.31.1,<6.32.0"]
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0"]
 dynamixel = ["dynamixel-sdk>=3.7.31,<3.9.0"]
-damiao = ["python-can>=4.2.0,<5.0.0"]
 
 # Robots
-openarms = ["lerobot[damiao]"]
 gamepad = ["lerobot[pygame-dep]", "hidapi>=0.14.0,<0.15.0"]
 hopejr = ["lerobot[feetech]", "lerobot[pygame-dep]"]
 lekiwi = ["lerobot[feetech]", "pyzmq>=26.2.1,<28.0.0"]
 unitree_g1 = [
     "pyzmq>=26.2.1,<28.0.0",
-    "onnxruntime>=1.16.0,<2.0.0",
-    "pin>=3.0.0,<4.0.0",
-    "meshcat>=0.3.0,<0.4.0",
-    "matplotlib>=3.9.0,<4.0.0",
-    "casadi>=3.6.0,<4.0.0",
+    "onnxruntime>=1.16.0,<2.0.0"
 ]
 reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
 kinematics = ["lerobot[placo-dep]"]
@@ -209,7 +203,6 @@ lerobot-info="lerobot.scripts.lerobot_info:main"
 lerobot-find-joint-limits="lerobot.scripts.lerobot_find_joint_limits:main"
 lerobot-imgtransform-viz="lerobot.scripts.lerobot_imgtransform_viz:main"
 lerobot-edit-dataset="lerobot.scripts.lerobot_edit_dataset:main"
-lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 
 # ---------------- Tool Configurations ----------------
 [tool.setuptools.packages.find]
@@ -285,7 +278,6 @@ default.extend-ignore-identifiers-re = [
     "thw",
     "inpt",
     "ROBOTIS",
-    "OT_VALUE"
 ]
 
 # TODO: Uncomment when ready to use
@@ -360,9 +352,9 @@ ignore_errors = false
 module = "lerobot.cameras.*"
 ignore_errors = false
 
-[[tool.mypy.overrides]]
-module = "lerobot.motors.*"
-ignore_errors = false
+# [[tool.mypy.overrides]]
+# module = "lerobot.motors.*"
+# ignore_errors = false
 
 # [[tool.mypy.overrides]]
 # module = "lerobot.robots.*"

src/lerobot/cameras/__init__.py

@@ -13,5 +13,5 @@
 # limitations under the License.
 
 from .camera import Camera
-from .configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
+from .configs import CameraConfig, ColorMode, Cv2Rotation
 from .utils import make_cameras_from_configs

src/lerobot/cameras/camera.py

@@ -15,12 +15,11 @@
 # limitations under the License.
 
 import abc
-import warnings
 from typing import Any
 
 from numpy.typing import NDArray  # type: ignore  # TODO: add type stubs for numpy.typing
 
-from .configs import CameraConfig
+from .configs import CameraConfig, ColorMode
 
 
 class Camera(abc.ABC):
@@ -31,12 +30,20 @@ class Camera(abc.ABC):
 
     Manages basic camera properties (FPS, resolution) and core operations:
     - Connection/disconnection
-    - Frame capture (sync/async/latest)
+    - Frame capture (sync/async)
 
     Attributes:
         fps (int | None): Configured frames per second
         width (int | None): Frame width in pixels
         height (int | None): Frame height in pixels
+
+    Example:
+        class MyCamera(Camera):
+            def __init__(self, config): ...
+            @property
+            def is_connected(self) -> bool: ...
+            def connect(self, warmup=True): ...
+            # Plus other required methods
     """
 
     def __init__(self, config: CameraConfig):
@@ -49,32 +56,6 @@ class Camera(abc.ABC):
         self.width: int | None = config.width
         self.height: int | None = config.height
 
-    def __enter__(self):
-        """
-        Context manager entry.
-        Automatically connects to the camera.
-        """
-        self.connect()
-        return self
-
-    def __exit__(self, exc_type, exc_value, traceback) -> None:
-        """
-        Context manager exit.
-        Automatically disconnects, ensuring resources are released even on error.
-        """
-        self.disconnect()
-
-    def __del__(self) -> None:
-        """
-        Destructor safety net.
-        Attempts to disconnect if the object is garbage collected without cleanup.
-        """
-        try:
-            if self.is_connected:
-                self.disconnect()
-        except Exception:  # nosec B110
-            pass
-
     @property
     @abc.abstractmethod
     def is_connected(self) -> bool:
@@ -108,10 +89,12 @@ class Camera(abc.ABC):
         pass
 
     @abc.abstractmethod
-    def read(self) -> NDArray[Any]:
-        """Capture and return a single frame from the camera synchronously.
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
+        """Capture and return a single frame from the camera.
 
-        This is a blocking call that will wait for the hardware and its SDK.
+        Args:
+            color_mode: Desired color mode for the output frame. If None,
+                        uses the camera's default color mode.
 
         Returns:
             np.ndarray: Captured frame as a numpy array.
@@ -120,64 +103,17 @@ class Camera(abc.ABC):
 
     @abc.abstractmethod
     def async_read(self, timeout_ms: float = ...) -> NDArray[Any]:
-        """Return the most recent new frame.
-
-        This method retrieves the latest frame captured by the background thread.
-        If a new frame is already available in the buffer (captured since the last call),
-        it returns it immediately.
-
-        It blocks up to `timeout_ms` only if the buffer is empty or if the latest frame
-        was already consumed by a previous `async_read` call.
-
-        Essentially, this method return the latest unconsumed frame, waiting if necessary
-        for a new one to arrive within the specified timeout.
-
-        Usage:
-            - Ideal for control loops where you want to ensure every processed frame
-            is fresh, effectively synchronizing your loop to the camera's FPS.
-            - Causes of a timeout usually include: very low camera FPS, heavy processing load,
-            or if the camera is disconnected.
+        """Asynchronously capture and return a single frame from the camera.
 
         Args:
-            timeout_ms: Maximum time to wait for a new frame in milliseconds.
-                        Defaults to 200ms (0.2s).
+            timeout_ms: Maximum time to wait for a frame in milliseconds.
+                        Defaults to implementation-specific timeout.
 
         Returns:
             np.ndarray: Captured frame as a numpy array.
-
-        Raises:
-            TimeoutError: If no new frame arrives within `timeout_ms`.
         """
         pass
 
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-        """Return the most recent frame captured immediately (Peeking).
-
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
-
-        Usage:
-            Ideal for scenarios requiring zero latency or decoupled frequencies & when
-            we want a guaranteed frame, such as UI visualization, logging, or
-            non-critical monitoring.
-
-        Returns:
-            NDArray[Any]: The frame image (numpy array).
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            NotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-        warnings.warn(
-            f"{self.__class__.__name__}.read_latest() is not implemented. "
-            "Please override read_latest(); it will be required in future releases.",
-            FutureWarning,
-            stacklevel=2,
-        )
-        return self.async_read()
-
     @abc.abstractmethod
     def disconnect(self) -> None:
         """Disconnect from the camera and release resources."""

src/lerobot/cameras/configs.py

@@ -25,10 +25,6 @@ class ColorMode(str, Enum):
     RGB = "rgb"
     BGR = "bgr"
 
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`color_mode` is expected to be in {list(cls)}, but {value} is provided.")
-
 
 class Cv2Rotation(int, Enum):
     NO_ROTATION = 0
@@ -36,25 +32,6 @@ class Cv2Rotation(int, Enum):
     ROTATE_180 = 180
     ROTATE_270 = -90
 
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`rotation` is expected to be in {list(cls)}, but {value} is provided.")
-
-
-# Subset from https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html
-class Cv2Backends(int, Enum):
-    ANY = 0
-    V4L2 = 200
-    DSHOW = 700
-    PVAPI = 800
-    ANDROID = 1000
-    AVFOUNDATION = 1200
-    MSMF = 1400
-
-    @classmethod
-    def _missing_(cls, value: object) -> None:
-        raise ValueError(f"`backend` is expected to be in {list(cls)}, but {value} is provided.")
-
 
 @dataclass(kw_only=True)
 class CameraConfig(draccus.ChoiceRegistry, abc.ABC):  # type: ignore  # TODO: add type stubs for draccus

src/lerobot/cameras/opencv/camera_opencv.py

@@ -32,11 +32,10 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
     os.environ["OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"] = "0"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
-from ..utils import get_cv2_rotation
+from ..utils import get_cv2_backend, get_cv2_rotation
 from .configuration_opencv import ColorMode, OpenCVCameraConfig
 
 # NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
@@ -71,24 +70,34 @@ class OpenCVCamera(Camera):
     Example:
         ```python
         from lerobot.cameras.opencv import OpenCVCamera
-        from lerobot.cameras.configuration_opencv import OpenCVCameraConfig
+        from lerobot.cameras.configuration_opencv import OpenCVCameraConfig, ColorMode, Cv2Rotation
 
         # Basic usage with camera index 0
         config = OpenCVCameraConfig(index_or_path=0)
         camera = OpenCVCamera(config)
         camera.connect()
 
-        # Read 1 frame synchronously (blocking)
+        # Read 1 frame synchronously
         color_image = camera.read()
+        print(color_image.shape)
 
-        # Read 1 frame asynchronously (waits for new frame with a timeout)
+        # Read 1 frame asynchronously
         async_image = camera.async_read()
 
-        # Get the latest frame immediately (no wait, returns timestamp)
-        latest_image, timestamp = camera.read_latest()
-
         # When done, properly disconnect the camera using
         camera.disconnect()
+
+        # Example with custom settings
+        custom_config = OpenCVCameraConfig(
+            index_or_path='/dev/video0', # Or use an index
+            fps=30,
+            width=1280,
+            height=720,
+            color_mode=ColorMode.RGB,
+            rotation=Cv2Rotation.ROTATE_90
+        )
+        custom_camera = OpenCVCamera(custom_config)
+        # ... connect, read, disconnect ...
         ```
     """
 
@@ -114,11 +123,10 @@ class OpenCVCamera(Camera):
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
         self.latest_frame: NDArray[Any] | None = None
-        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = config.backend
+        self.backend: int = get_cv2_backend()
 
         if self.height and self.width:
             self.capture_width, self.capture_height = self.width, self.height
@@ -133,23 +141,20 @@ class OpenCVCamera(Camera):
         """Checks if the camera is currently connected and opened."""
         return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the OpenCV camera specified in the configuration.
 
         Initializes the OpenCV VideoCapture object, sets desired camera properties
-        (FPS, width, height), starts the background reading thread and performs initial checks.
-
-        Args:
-            warmup (bool): If True, waits at connect() time until at least one valid frame
-                           has been captured by the background thread. Defaults to True.
+        (FPS, width, height), and performs initial checks.
 
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
-            ConnectionError: If the specified camera index/path is not found or fails to open.
-            RuntimeError: If the camera opens but fails to apply requested settings.
+            ConnectionError: If the specified camera index/path is not found or the camera is found but fails to open.
+            RuntimeError: If the camera opens but fails to apply requested FPS/resolution settings.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         # Use 1 thread for OpenCV operations to avoid potential conflicts or
         # blocking in multi-threaded applications, especially during data collection.
@@ -165,20 +170,15 @@ class OpenCVCamera(Camera):
             )
 
         self._configure_capture_settings()
-        self._start_read_thread()
 
-        if warmup and self.warmup_s > 0:
+        if warmup:
             start_time = time.time()
             while time.time() - start_time < self.warmup_s:
-                self.async_read(timeout_ms=self.warmup_s * 1000)
+                self.read()
                 time.sleep(0.1)
-            with self.frame_lock:
-                if self.latest_frame is None:
-                    raise ConnectionError(f"{self} failed to capture frames during warmup.")
 
         logger.info(f"{self} connected.")
 
-    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """
         Applies the specified FOURCC, FPS, width, and height settings to the connected camera.
@@ -196,8 +196,11 @@ class OpenCVCamera(Camera):
         Raises:
             RuntimeError: If the camera fails to set any of the specified properties
                           to the requested value.
-            DeviceNotConnectedError: If the camera is not connected.
+            DeviceNotConnectedError: If the camera is not connected when attempting
+                                     to configure settings.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
 
         # Set FOURCC first (if specified) as it can affect available FPS/resolution options
         if self.config.fourcc is not None:
@@ -336,18 +339,6 @@ class OpenCVCamera(Camera):
 
         return found_cameras_info
 
-    def _read_from_hardware(self) -> NDArray[Any]:
-        if self.videocapture is None:
-            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
-
-        ret, frame = self.videocapture.read()
-
-        if not ret:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
-
-        return frame
-
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
@@ -355,6 +346,11 @@ class OpenCVCamera(Camera):
         This is a blocking call. It waits for the next available frame from the
         camera hardware via OpenCV.
 
+        Args:
+            color_mode (Optional[ColorMode]): If specified, overrides the default
+                color mode (`self.color_mode`) for this read operation (e.g.,
+                request RGB even if default is BGR).
+
         Returns:
             np.ndarray: The captured frame as a NumPy array in the format
                        (height, width, channels), using the specified or default
@@ -366,31 +362,34 @@ class OpenCVCamera(Camera):
                           received frame dimensions don't match expectations before rotation.
             ValueError: If an invalid `color_mode` is requested.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         start_time = time.perf_counter()
 
-        if color_mode is not None:
-            logger.warning(
-                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
-            )
+        if self.videocapture is None:
+            raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        ret, frame = self.videocapture.read()
 
-        self.new_frame_event.clear()
-        frame = self.async_read(timeout_ms=10000)
+        if not ret or frame is None:
+            raise RuntimeError(f"{self} read failed (status={ret}).")
+
+        processed_frame = self._postprocess_image(frame, color_mode)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
-        return frame
+        return processed_frame
 
-    def _postprocess_image(self, image: NDArray[Any]) -> NDArray[Any]:
+    def _postprocess_image(self, image: NDArray[Any], color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw frame.
 
         Args:
             image (np.ndarray): The raw image frame (expected BGR format from OpenCV).
+            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
+                                             uses the instance's default `self.color_mode`.
 
         Returns:
             np.ndarray: The processed image frame.
@@ -400,10 +399,11 @@ class OpenCVCamera(Camera):
             RuntimeError: If the raw frame dimensions do not match the configured
                           `width` and `height`.
         """
+        requested_color_mode = self.color_mode if color_mode is None else color_mode
 
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+        if requested_color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
-                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+                f"Invalid color mode '{requested_color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
 
         h, w, c = image.shape
@@ -417,7 +417,7 @@ class OpenCVCamera(Camera):
             raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
 
         processed_image = image
-        if self.color_mode == ColorMode.RGB:
+        if requested_color_mode == ColorMode.RGB:
             processed_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
 
         if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE, cv2.ROTATE_180]:
@@ -431,7 +431,7 @@ class OpenCVCamera(Camera):
 
         On each iteration:
         1. Reads a color frame
-        2. Stores result in latest_frame and updates timestamp (thread-safe)
+        2. Stores result in latest_frame (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -439,37 +439,30 @@ class OpenCVCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
-        failure_count = 0
         while not self.stop_event.is_set():
             try:
-                raw_frame = self._read_from_hardware()
-                processed_frame = self._postprocess_image(raw_frame)
-                capture_time = time.perf_counter()
+                color_image = self.read()
 
                 with self.frame_lock:
-                    self.latest_frame = processed_frame
-                    self.latest_timestamp = capture_time
+                    self.latest_frame = color_image
                 self.new_frame_event.set()
-                failure_count = 0
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
-                if failure_count <= 10:
-                    failure_count += 1
-                    logger.warning(f"Error reading frame in background thread for {self}: {e}")
-                else:
-                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
+                logger.warning(f"Error reading frame in background thread for {self}: {e}")
 
     def _start_read_thread(self) -> None:
         """Starts or restarts the background read thread if it's not running."""
-        self._stop_read_thread()
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=0.1)
+        if self.stop_event is not None:
+            self.stop_event.set()
 
         self.stop_event = Event()
         self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
         self.thread.daemon = True
         self.thread.start()
-        time.sleep(0.1)
 
     def _stop_read_thread(self) -> None:
         """Signals the background read thread to stop and waits for it to join."""
@@ -482,12 +475,6 @@ class OpenCVCamera(Camera):
         self.thread = None
         self.stop_event = None
 
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame asynchronously.
@@ -495,7 +482,6 @@ class OpenCVCamera(Camera):
         This method retrieves the most recent frame captured by the background
         read thread. It does not block waiting for the camera hardware directly,
         but may wait up to timeout_ms for the background thread to provide a frame.
-        It is “best effort” under high FPS.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -510,14 +496,17 @@ class OpenCVCamera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+            self._start_read_thread()
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            thread_alive = self.thread is not None and self.thread.is_alive()
             raise TimeoutError(
                 f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {self.thread.is_alive()}."
+                f"Read thread alive: {thread_alive}."
             )
 
         with self.frame_lock:
@@ -529,41 +518,6 @@ class OpenCVCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-        """Return the most recent frame captured immediately (Peeking).
-
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
-
-        Returns:
-            NDArray[Any]: The frame image (numpy array).
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        with self.frame_lock:
-            frame = self.latest_frame
-            timestamp = self.latest_timestamp
-
-        if frame is None or timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
-
-        age_ms = (time.perf_counter() - timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
-
-        return frame
-
     def disconnect(self) -> None:
         """
         Disconnects from the camera and cleans up resources.
@@ -584,9 +538,4 @@ class OpenCVCamera(Camera):
             self.videocapture.release()
             self.videocapture = None
 
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/opencv/configuration_opencv.py

@@ -15,9 +15,9 @@
 from dataclasses import dataclass
 from pathlib import Path
 
-from ..configs import CameraConfig, ColorMode, Cv2Backends, Cv2Rotation
+from ..configs import CameraConfig, ColorMode, Cv2Rotation
 
-__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation", "Cv2Backends"]
+__all__ = ["OpenCVCameraConfig", "ColorMode", "Cv2Rotation"]
 
 
 @CameraConfig.register_subclass("opencv")
@@ -50,7 +50,6 @@ class OpenCVCameraConfig(CameraConfig):
         rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
         warmup_s: Time reading frames before returning from connect (in seconds)
         fourcc: FOURCC code for video format (e.g., "MJPG", "YUYV", "I420"). Defaults to None (auto-detect).
-        backend: OpenCV backend identifier (https://docs.opencv.org/3.4/d4/d15/group__videoio__flags__base.html). Defaults to ANY.
 
     Note:
         - Only 3-channel color output (RGB/BGR) is currently supported.
@@ -63,12 +62,22 @@ class OpenCVCameraConfig(CameraConfig):
     rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
     warmup_s: int = 1
     fourcc: str | None = None
-    backend: Cv2Backends = Cv2Backends.ANY
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
-        self.rotation = Cv2Rotation(self.rotation)
-        self.backend = Cv2Backends(self.backend)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            raise ValueError(
+                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
+            )
 
         if self.fourcc is not None and (not isinstance(self.fourcc, str) or len(self.fourcc) != 4):
             raise ValueError(

src/lerobot/cameras/reachy2_camera/configuration_reachy2_camera.py

@@ -74,4 +74,7 @@ class Reachy2CameraConfig(CameraConfig):
                 f"`image_type` is expected to be 'left' or 'right' for teleop camera, and 'rgb' or 'depth' for depth camera, but {self.image_type} is provided."
             )
 
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in ["rgb", "bgr"]:
+            raise ValueError(
+                f"`color_mode` is expected to be 'rgb' or 'bgr', but {self.color_mode} is provided."
+            )

src/lerobot/cameras/reachy2_camera/reachy2_camera.py

@@ -32,7 +32,6 @@ if platform.system() == "Windows" and "OPENCV_VIDEOIO_MSMF_ENABLE_HW_TRANSFORMS"
 import cv2  # type: ignore  # TODO: add type stubs for OpenCV
 import numpy as np  # type: ignore  # TODO: add type stubs for numpy
 
-from lerobot.utils.decorators import check_if_not_connected
 from lerobot.utils.import_utils import _reachy2_sdk_available
 
 if TYPE_CHECKING or _reachy2_sdk_available:
@@ -81,8 +80,6 @@ class Reachy2Camera(Camera):
         self.config = config
 
         self.color_mode = config.color_mode
-        self.latest_frame: NDArray[Any] | None = None
-        self.latest_timestamp: float | None = None
 
         self.cam_manager: CameraManager | None = None
 
@@ -124,12 +121,16 @@ class Reachy2Camera(Camera):
         """
         raise NotImplementedError("Camera detection is not implemented for Reachy2 cameras.")
 
-    @check_if_not_connected
     def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
         Reads a single frame synchronously from the camera.
 
-        This method retrieves the most recent frame available in Reachy 2's low-level software.
+        This is a blocking call.
+
+        Args:
+            color_mode (Optional[ColorMode]): If specified, overrides the default
+                color mode (`self.color_mode`) for this read operation (e.g.,
+                request RGB even if default is BGR).
 
         Returns:
             np.ndarray: The captured frame as a NumPy array in the format
@@ -138,13 +139,11 @@ class Reachy2Camera(Camera):
         """
         start_time = time.perf_counter()
 
-        if self.cam_manager is None:
+        if not self.is_connected:
             raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        if color_mode is not None:
-            logger.warning(
-                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
-            )
+        if self.cam_manager is None:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         frame: NDArray[Any] = np.empty((0, 0, 3), dtype=np.uint8)
 
@@ -166,27 +165,25 @@ class Reachy2Camera(Camera):
             raise ValueError(f"Invalid camera name '{self.config.name}'. Expected 'teleop' or 'depth'.")
 
         if frame is None:
-            raise RuntimeError(f"Internal error: No frame available for {self}.")
+            return np.empty((0, 0, 3), dtype=np.uint8)
 
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"Invalid color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
-            )
-        if self.color_mode == ColorMode.RGB:
+        if self.config.color_mode == "rgb":
             frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
 
-        self.latest_frame = frame
-        self.latest_timestamp = time.perf_counter()
-
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
         return frame
 
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
-        Same as read()
+        Reads the latest available frame.
+
+        This method retrieves the most recent frame available in Reachy 2's low-level software.
+
+        Args:
+            timeout_ms (float): Maximum time in milliseconds to wait for a frame
+                to become available. Defaults to 200ms (0.2 seconds).
 
         Returns:
             np.ndarray: The latest captured frame as a NumPy array in the format
@@ -197,40 +194,16 @@ class Reachy2Camera(Camera):
             TimeoutError: If no frame becomes available within the specified timeout.
             RuntimeError: If an unexpected error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
-        return self.read()
+        frame = self.read()
 
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-        """Return the most recent frame captured immediately (Peeking).
+        if frame is None:
+            raise RuntimeError(f"Internal error: No frame available for {self}.")
 
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
+        return frame
 
-        Returns:
-            tuple[NDArray, float]:
-                - The frame image (numpy array).
-                - The timestamp (time.perf_counter) when this frame was captured.
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-
-        if self.latest_frame is None or self.latest_timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
-
-        age_ms = (time.perf_counter() - self.latest_timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
-
-        return self.latest_frame
-
-    @check_if_not_connected
     def disconnect(self) -> None:
         """
         Stops the background read thread (if running).
@@ -238,6 +211,8 @@ class Reachy2Camera(Camera):
         Raises:
             DeviceNotConnectedError: If the camera is already disconnected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} not connected.")
 
         if self.cam_manager is not None:
             self.cam_manager.disconnect()

src/lerobot/cameras/realsense/camera_realsense.py

@@ -30,8 +30,7 @@ try:
 except Exception as e:
     logging.info(f"Could not import realsense: {e}")
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -73,14 +72,15 @@ class RealSenseCamera(Camera):
         camera = RealSenseCamera(config)
         camera.connect()
 
-        # Read 1 frame synchronously (blocking)
+        # Read 1 frame synchronously
         color_image = camera.read()
+        print(color_image.shape)
 
-        # Read 1 frame asynchronously (waits for new frame with a timeout)
+        # Read 1 frame asynchronously
         async_image = camera.async_read()
 
-        # Get the latest frame immediately (no wait, returns timestamp)
-        latest_image, timestamp = camera.read_latest()
+        # When done, properly disconnect the camera using
+        camera.disconnect()
 
         # Example with depth capture and custom settings
         custom_config = RealSenseCameraConfig(
@@ -133,9 +133,7 @@ class RealSenseCamera(Camera):
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
-        self.latest_color_frame: NDArray[Any] | None = None
-        self.latest_depth_frame: NDArray[Any] | None = None
-        self.latest_timestamp: float | None = None
+        self.latest_frame: NDArray[Any] | None = None
         self.new_frame_event: Event = Event()
 
         self.rotation: int | None = get_cv2_rotation(config.rotation)
@@ -153,7 +151,6 @@ class RealSenseCamera(Camera):
         """Checks if the camera pipeline is started and streams are active."""
         return self.rs_pipeline is not None and self.rs_profile is not None
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
         """
         Connects to the RealSense camera specified in the configuration.
@@ -161,16 +158,14 @@ class RealSenseCamera(Camera):
         Initializes the RealSense pipeline, configures the required streams (color
         and optionally depth), starts the pipeline, and validates the actual stream settings.
 
-        Args:
-            warmup (bool): If True, waits at connect() time until at least one valid frame
-                           has been captured by the background thread. Defaults to True.
-
         Raises:
             DeviceAlreadyConnectedError: If the camera is already connected.
             ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
             ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
             RuntimeError: If the pipeline starts but fails to apply requested settings.
         """
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         self.rs_pipeline = rs.pipeline()
         rs_config = rs.config()
@@ -186,18 +181,15 @@ class RealSenseCamera(Camera):
             ) from e
 
         self._configure_capture_settings()
-        self._start_read_thread()
 
-        # NOTE(Steven/Caroline): Enforcing at least one second of warmup as RS cameras need a bit of time before the first read. If we don't wait, the first read from the warmup will raise.
-        self.warmup_s = max(self.warmup_s, 1)
-
-        start_time = time.time()
-        while time.time() - start_time < self.warmup_s:
-            self.async_read(timeout_ms=self.warmup_s * 1000)
-            time.sleep(0.1)
-        with self.frame_lock:
-            if self.latest_color_frame is None or self.use_depth and self.latest_depth_frame is None:
-                raise ConnectionError(f"{self} failed to capture frames during warmup.")
+        if warmup:
+            time.sleep(
+                1
+            )  # NOTE(Steven): RS cameras need a bit of time to warm up before the first read. If we don't wait, the first read from the warmup will raise.
+            start_time = time.time()
+            while time.time() - start_time < self.warmup_s:
+                self.read()
+                time.sleep(0.1)
 
         logger.info(f"{self} connected.")
 
@@ -290,7 +282,6 @@ class RealSenseCamera(Camera):
             if self.use_depth:
                 rs_config.enable_stream(rs.stream.depth)
 
-    @check_if_not_connected
     def _configure_capture_settings(self) -> None:
         """Sets fps, width, and height from device stream if not already configured.
 
@@ -300,6 +291,8 @@ class RealSenseCamera(Camera):
         Raises:
             DeviceNotConnectedError: If device is not connected.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
 
         if self.rs_profile is None:
             raise RuntimeError(f"{self}: rs_profile must be initialized before use.")
@@ -319,7 +312,6 @@ class RealSenseCamera(Camera):
                 self.width, self.height = actual_width, actual_height
                 self.capture_width, self.capture_height = actual_width, actual_height
 
-    @check_if_not_connected
     def read_depth(self, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (depth) synchronously from the camera.
@@ -327,6 +319,9 @@ class RealSenseCamera(Camera):
         This is a blocking call. It waits for a coherent set of frames (depth)
         from the camera hardware via the RealSense pipeline.
 
+        Args:
+            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
+
         Returns:
             np.ndarray: The depth map as a NumPy array (height, width)
                   of type `np.uint16` (raw depth values in millimeters) and rotation.
@@ -335,50 +330,44 @@ class RealSenseCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
             RuntimeError: If reading frames from the pipeline fails or frames are invalid.
         """
-        if timeout_ms:
-            logger.warning(
-                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
-            )
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
         if not self.use_depth:
             raise RuntimeError(
                 f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
             )
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        start_time = time.perf_counter()
 
-        self.new_frame_event.clear()
-
-        _ = self.async_read(timeout_ms=10000)
-
-        with self.frame_lock:
-            depth_map = self.latest_depth_frame
-
-        if depth_map is None:
-            raise RuntimeError("No depth frame available. Ensure camera is streaming.")
-
-        return depth_map
-
-    def _read_from_hardware(self):
         if self.rs_pipeline is None:
             raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
 
-        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=10000)
+        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
         if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
+            raise RuntimeError(f"{self} read_depth failed (status={ret}).")
 
-        return frame
+        depth_frame = frame.get_depth_frame()
+        depth_map = np.asanyarray(depth_frame.get_data())
 
-    @check_if_not_connected
-    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 0) -> NDArray[Any]:
+        depth_map_processed = self._postprocess_image(depth_map, depth_frame=True)
+
+        read_duration_ms = (time.perf_counter() - start_time) * 1e3
+        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
+
+        return depth_map_processed
+
+    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> NDArray[Any]:
         """
         Reads a single frame (color) synchronously from the camera.
 
         This is a blocking call. It waits for a coherent set of frames (color)
         from the camera hardware via the RealSense pipeline.
 
+        Args:
+            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
+
         Returns:
             np.ndarray: The captured color frame as a NumPy array
               (height, width, channels), processed according to `color_mode` and rotation.
@@ -389,36 +378,39 @@ class RealSenseCamera(Camera):
             ValueError: If an invalid `color_mode` is requested.
         """
 
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
+
         start_time = time.perf_counter()
 
-        if color_mode is not None:
-            logger.warning(
-                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
-            )
+        if self.rs_pipeline is None:
+            raise RuntimeError(f"{self}: rs_pipeline must be initialized before use.")
 
-        if timeout_ms:
-            logger.warning(
-                f"{self} read() timeout_ms parameter is deprecated and will be removed in future versions."
-            )
+        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
 
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        if not ret or frame is None:
+            raise RuntimeError(f"{self} read failed (status={ret}).")
 
-        self.new_frame_event.clear()
+        color_frame = frame.get_color_frame()
+        color_image_raw = np.asanyarray(color_frame.get_data())
 
-        frame = self.async_read(timeout_ms=10000)
+        color_image_processed = self._postprocess_image(color_image_raw, color_mode)
 
         read_duration_ms = (time.perf_counter() - start_time) * 1e3
         logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
 
-        return frame
+        return color_image_processed
 
-    def _postprocess_image(self, image: NDArray[Any], depth_frame: bool = False) -> NDArray[Any]:
+    def _postprocess_image(
+        self, image: NDArray[Any], color_mode: ColorMode | None = None, depth_frame: bool = False
+    ) -> NDArray[Any]:
         """
         Applies color conversion, dimension validation, and rotation to a raw color frame.
 
         Args:
             image (np.ndarray): The raw image frame (expected RGB format from RealSense).
+            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
+                                             uses the instance's default `self.color_mode`.
 
         Returns:
             np.ndarray: The processed image frame according to `self.color_mode` and `self.rotation`.
@@ -429,9 +421,9 @@ class RealSenseCamera(Camera):
                           `width` and `height`.
         """
 
-        if self.color_mode and self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+        if color_mode and color_mode not in (ColorMode.RGB, ColorMode.BGR):
             raise ValueError(
-                f"Invalid requested color mode '{self.color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
+                f"Invalid requested color mode '{color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
             )
 
         if depth_frame:
@@ -462,7 +454,7 @@ class RealSenseCamera(Camera):
 
         On each iteration:
         1. Reads a color frame with 500ms timeout
-        2. Stores result in latest_frame and updates timestamp (thread-safe)
+        2. Stores result in latest_frame (thread-safe)
         3. Sets new_frame_event to notify listeners
 
         Stops on DeviceNotConnectedError, logs other errors and continues.
@@ -470,41 +462,25 @@ class RealSenseCamera(Camera):
         if self.stop_event is None:
             raise RuntimeError(f"{self}: stop_event is not initialized before starting read loop.")
 
-        failure_count = 0
         while not self.stop_event.is_set():
             try:
-                frame = self._read_from_hardware()
-                color_frame_raw = frame.get_color_frame()
-                color_frame = np.asanyarray(color_frame_raw.get_data())
-                processed_color_frame = self._postprocess_image(color_frame)
-
-                if self.use_depth:
-                    depth_frame_raw = frame.get_depth_frame()
-                    depth_frame = np.asanyarray(depth_frame_raw.get_data())
-                    processed_depth_frame = self._postprocess_image(depth_frame, depth_frame=True)
-
-                capture_time = time.perf_counter()
+                color_image = self.read(timeout_ms=500)
 
                 with self.frame_lock:
-                    self.latest_color_frame = processed_color_frame
-                    if self.use_depth:
-                        self.latest_depth_frame = processed_depth_frame
-                    self.latest_timestamp = capture_time
+                    self.latest_frame = color_image
                 self.new_frame_event.set()
-                failure_count = 0
 
             except DeviceNotConnectedError:
                 break
             except Exception as e:
-                if failure_count <= 10:
-                    failure_count += 1
-                    logger.warning(f"Error reading frame in background thread for {self}: {e}")
-                else:
-                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
+                logger.warning(f"Error reading frame in background thread for {self}: {e}")
 
     def _start_read_thread(self) -> None:
         """Starts or restarts the background read thread if it's not running."""
-        self._stop_read_thread()
+        if self.thread is not None and self.thread.is_alive():
+            self.thread.join(timeout=0.1)
+        if self.stop_event is not None:
+            self.stop_event.set()
 
         self.stop_event = Event()
         self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
@@ -522,14 +498,7 @@ class RealSenseCamera(Camera):
         self.thread = None
         self.stop_event = None
 
-        with self.frame_lock:
-            self.latest_color_frame = None
-            self.latest_depth_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
     # NOTE(Steven): Missing implementation for depth for now
-    @check_if_not_connected
     def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
         """
         Reads the latest available frame data (color) asynchronously.
@@ -537,7 +506,6 @@ class RealSenseCamera(Camera):
         This method retrieves the most recent color frame captured by the background
         read thread. It does not block waiting for the camera hardware directly,
         but may wait up to timeout_ms for the background thread to provide a frame.
-        It is “best effort” under high FPS.
 
         Args:
             timeout_ms (float): Maximum time in milliseconds to wait for a frame
@@ -552,18 +520,21 @@ class RealSenseCamera(Camera):
             TimeoutError: If no frame data becomes available within the specified timeout.
             RuntimeError: If the background thread died unexpectedly or another error occurs.
         """
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
         if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+            self._start_read_thread()
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
+            thread_alive = self.thread is not None and self.thread.is_alive()
             raise TimeoutError(
                 f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {self.thread.is_alive()}."
+                f"Read thread alive: {thread_alive}."
             )
 
         with self.frame_lock:
-            frame = self.latest_color_frame
+            frame = self.latest_frame
             self.new_frame_event.clear()
 
         if frame is None:
@@ -571,42 +542,6 @@ class RealSenseCamera(Camera):
 
         return frame
 
-    # NOTE(Steven): Missing implementation for depth for now
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-        """Return the most recent (color) frame captured immediately (Peeking).
-
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
-
-        Returns:
-            NDArray[Any]: The frame image (numpy array).
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        with self.frame_lock:
-            frame = self.latest_color_frame
-            timestamp = self.latest_timestamp
-
-        if frame is None or timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
-
-        age_ms = (time.perf_counter() - timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
-
-        return frame
-
     def disconnect(self) -> None:
         """
         Disconnects from the camera, stops the pipeline, and cleans up resources.
@@ -630,10 +565,4 @@ class RealSenseCamera(Camera):
             self.rs_pipeline = None
             self.rs_profile = None
 
-        with self.frame_lock:
-            self.latest_color_frame = None
-            self.latest_depth_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/realsense/configuration_realsense.py

@@ -60,8 +60,20 @@ class RealSenseCameraConfig(CameraConfig):
     warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
-        self.rotation = Cv2Rotation(self.rotation)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
+
+        if self.rotation not in (
+            Cv2Rotation.NO_ROTATION,
+            Cv2Rotation.ROTATE_90,
+            Cv2Rotation.ROTATE_180,
+            Cv2Rotation.ROTATE_270,
+        ):
+            raise ValueError(
+                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
+            )
 
         values = (self.fps, self.width, self.height)
         if any(v is not None for v in values) and any(v is None for v in values):

src/lerobot/cameras/utils.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import platform
 from typing import cast
 
 from lerobot.utils.import_utils import make_device_from_device_class
@@ -67,3 +68,14 @@ def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
         return int(cv2.ROTATE_90_COUNTERCLOCKWISE)
     else:
         return None
+
+
+def get_cv2_backend() -> int:
+    import cv2
+
+    if platform.system() == "Windows":
+        return int(cv2.CAP_MSMF)  # Use MSMF for Windows instead of AVFOUNDATION
+    # elif platform.system() == "Darwin":  # macOS
+    #     return cv2.CAP_AVFOUNDATION
+    else:  # Linux and others
+        return int(cv2.CAP_ANY)

src/lerobot/cameras/zmq/camera_zmq.py

@@ -34,8 +34,7 @@ import cv2
 import numpy as np
 from numpy.typing import NDArray
 
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.errors import DeviceNotConnectedError
+from lerobot.utils.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
 
 from ..camera import Camera
 from ..configs import ColorMode
@@ -46,12 +45,6 @@ logger = logging.getLogger(__name__)
 
 class ZMQCamera(Camera):
     """
-    Manages camera interactions via ZeroMQ for receiving frames from a remote server.
-
-    This class connects to a ZMQ Publisher, subscribes to frame topics, and decodes
-    incoming JSON messages containing Base64 encoded images. It supports both
-    synchronous and asynchronous frame reading patterns.
-
     Example usage:
         ```python
         from lerobot.cameras.zmq import ZMQCamera, ZMQCameraConfig
@@ -59,16 +52,7 @@ class ZMQCamera(Camera):
         config = ZMQCameraConfig(server_address="192.168.123.164", port=5555, camera_name="head_camera")
         camera = ZMQCamera(config)
         camera.connect()
-
-        # Read 1 frame synchronously (blocking)
-        color_image = camera.read()
-
-        # Read 1 frame asynchronously (waits for new frame with a timeout)
-        async_image = camera.async_read()
-
-        # Get the latest frame immediately (no wait, returns timestamp)
-        latest_image, timestamp = camera.read_latest()
-
+        frame = camera.read()
         camera.disconnect()
         ```
     """
@@ -84,17 +68,14 @@ class ZMQCamera(Camera):
         self.color_mode = config.color_mode
         self.timeout_ms = config.timeout_ms
 
-        # ZMQ Context and Socket
         self.context: zmq.Context | None = None
         self.socket: zmq.Socket | None = None
         self._connected = False
 
-        # Threading resources
         self.thread: Thread | None = None
         self.stop_event: Event | None = None
         self.frame_lock: Lock = Lock()
         self.latest_frame: NDArray[Any] | None = None
-        self.latest_timestamp: float | None = None
         self.new_frame_event: Event = Event()
 
     def __str__(self) -> str:
@@ -102,17 +83,12 @@ class ZMQCamera(Camera):
 
     @property
     def is_connected(self) -> bool:
-        """Checks if the ZMQ socket is initialized and connected."""
         return self._connected and self.context is not None and self.socket is not None
 
-    @check_if_already_connected
     def connect(self, warmup: bool = True) -> None:
-        """Connect to ZMQ camera server.
-
-        Args:
-            warmup (bool): If True, waits for the camera to provide at least one
-                           valid frame before returning. Defaults to True.
-        """
+        """Connect to ZMQ camera server."""
+        if self.is_connected:
+            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
 
         logger.info(f"Connecting to {self}...")
 
@@ -127,28 +103,17 @@ class ZMQCamera(Camera):
             self.socket.connect(f"tcp://{self.server_address}:{self.port}")
             self._connected = True
 
-            # Auto-detect resolution if not provided
+            # Auto-detect resolution
             if self.width is None or self.height is None:
-                # Read directly from hardware because the thread isn't running yet
-                temp_frame = self._read_from_hardware()
-                h, w = temp_frame.shape[:2]
+                h, w = self.read().shape[:2]
                 self.height = h
                 self.width = w
-                logger.info(f"{self} resolution detected: {w}x{h}")
+                logger.info(f"{self} resolution: {w}x{h}")
 
-            self._start_read_thread()
             logger.info(f"{self} connected.")
 
             if warmup:
-                # Ensure we have captured at least one frame via the thread
-                start_time = time.time()
-                while time.time() - start_time < (self.config.warmup_s):  # Wait a bit more than timeout
-                    self.async_read(timeout_ms=self.config.warmup_s * 1000)
-                    time.sleep(0.1)
-
-                with self.frame_lock:
-                    if self.latest_frame is None:
-                        raise ConnectionError(f"{self} failed to capture frames during warmup.")
+                time.sleep(0.1)
 
         except Exception as e:
             self._cleanup()
@@ -166,14 +131,15 @@ class ZMQCamera(Camera):
 
     @staticmethod
     def find_cameras() -> list[dict[str, Any]]:
-        """
-        Detection not implemented for ZMQ cameras. These cameras require manual configuration (server address/port).
-        """
-        raise NotImplementedError("Camera detection is not implemented for ZMQ cameras.")
+        """ZMQ cameras require manual configuration (server address/port)."""
+        return []
 
-    def _read_from_hardware(self) -> NDArray[Any]:
+    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
         """
-        Reads a single frame directly from the ZMQ socket.
+        Read a single frame from the ZMQ camera.
+
+        Returns:
+            np.ndarray: Decoded frame (height, width, 3)
         """
         if not self.is_connected or self.socket is None:
             raise DeviceNotConnectedError(f"{self} is not connected.")
@@ -181,7 +147,6 @@ class ZMQCamera(Camera):
         try:
             message = self.socket.recv_string()
         except Exception as e:
-            # Check for ZMQ timeout (EAGAIN/Again) without requiring global zmq import
             if type(e).__name__ == "Again":
                 raise TimeoutError(f"{self} timeout after {self.timeout_ms}ms") from e
             raise
@@ -211,114 +176,42 @@ class ZMQCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
-    def read(self, color_mode: ColorMode | None = None) -> NDArray[Any]:
-        """
-        Reads a single frame synchronously from the camera.
-
-        This is a blocking call. It waits for the next available frame from the
-        camera background thread.
-
-        Returns:
-            np.ndarray: Decoded frame (height, width, 3)
-        """
-        start_time = time.perf_counter()
-
-        if color_mode is not None:
-            logger.warning(
-                f"{self} read() color_mode parameter is deprecated and will be removed in future versions."
-            )
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        self.new_frame_event.clear()
-        frame = self.async_read(timeout_ms=10000)
-
-        read_duration_ms = (time.perf_counter() - start_time) * 1e3
-        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
-
-        return frame
-
     def _read_loop(self) -> None:
-        """
-        Internal loop run by the background thread for asynchronous reading.
-        """
-        if self.stop_event is None:
-            raise RuntimeError(f"{self}: stop_event is not initialized.")
-
-        failure_count = 0
-        while not self.stop_event.is_set():
+        while self.stop_event and not self.stop_event.is_set():
             try:
-                frame = self._read_from_hardware()
-                capture_time = time.perf_counter()
-
+                frame = self.read()
                 with self.frame_lock:
                     self.latest_frame = frame
-                    self.latest_timestamp = capture_time
                 self.new_frame_event.set()
-                failure_count = 0
-
             except DeviceNotConnectedError:
                 break
-            except (TimeoutError, Exception) as e:
-                if failure_count <= 10:
-                    failure_count += 1
-                    logger.warning(f"Read error: {e}")
-                else:
-                    raise RuntimeError(f"{self} exceeded maximum consecutive read failures.") from e
+            except TimeoutError:
+                pass
+            except Exception as e:
+                logger.warning(f"Read error: {e}")
 
     def _start_read_thread(self) -> None:
-        if self.stop_event is not None:
-            self.stop_event.set()
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=2.0)
-
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
+        if self.thread and self.thread.is_alive():
+            return
         self.stop_event = Event()
-        self.thread = Thread(target=self._read_loop, daemon=True, name=f"{self}_read_loop")
+        self.thread = Thread(target=self._read_loop, daemon=True)
         self.thread.start()
-        time.sleep(0.1)
 
     def _stop_read_thread(self) -> None:
-        if self.stop_event is not None:
+        if self.stop_event:
             self.stop_event.set()
-
-        if self.thread is not None and self.thread.is_alive():
+        if self.thread and self.thread.is_alive():
             self.thread.join(timeout=2.0)
-
         self.thread = None
         self.stop_event = None
 
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
+    def async_read(self, timeout_ms: float = 10000) -> NDArray[Any]:
+        """Read latest frame asynchronously (non-blocking)."""
+        if not self.is_connected:
+            raise DeviceNotConnectedError(f"{self} is not connected.")
 
-    @check_if_not_connected
-    def async_read(self, timeout_ms: float = 200) -> NDArray[Any]:
-        """
-        Reads the latest available frame asynchronously.
-
-        Args:
-            timeout_ms (float): Maximum time in milliseconds to wait for a frame
-                to become available. Defaults to 200ms.
-
-        Returns:
-            np.ndarray: The latest captured frame.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is not connected.
-            TimeoutError: If no frame data becomes available within the specified timeout.
-            RuntimeError: If the background thread is not running.
-        """
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
+        if not self.thread or not self.thread.is_alive():
+            self._start_read_thread()
 
         if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
             raise TimeoutError(f"{self} async_read timeout after {timeout_ms}ms")
@@ -332,54 +225,11 @@ class ZMQCamera(Camera):
 
         return frame
 
-    @check_if_not_connected
-    def read_latest(self, max_age_ms: int = 1000) -> NDArray[Any]:
-        """Return the most recent frame captured immediately (Peeking).
-
-        This method is non-blocking and returns whatever is currently in the
-        memory buffer. The frame may be stale,
-        meaning it could have been captured a while ago (hanging camera scenario e.g.).
-
-        Returns:
-            NDArray[Any]: The frame image (numpy array).
-
-        Raises:
-            TimeoutError: If the latest frame is older than `max_age_ms`.
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If the camera is connected but has not captured any frames yet.
-        """
-
-        if self.thread is None or not self.thread.is_alive():
-            raise RuntimeError(f"{self} read thread is not running.")
-
-        with self.frame_lock:
-            frame = self.latest_frame
-            timestamp = self.latest_timestamp
-
-        if frame is None or timestamp is None:
-            raise RuntimeError(f"{self} has not captured any frames yet.")
-
-        age_ms = (time.perf_counter() - timestamp) * 1e3
-        if age_ms > max_age_ms:
-            raise TimeoutError(
-                f"{self} latest frame is too old: {age_ms:.1f} ms (max allowed: {max_age_ms} ms)."
-            )
-
-        return frame
-
     def disconnect(self) -> None:
         """Disconnect from ZMQ camera."""
-        if not self.is_connected and self.thread is None:
+        if not self.is_connected and not self.thread:
             raise DeviceNotConnectedError(f"{self} not connected.")
 
-        if self.thread is not None:
-            self._stop_read_thread()
-
+        self._stop_read_thread()
         self._cleanup()
-
-        with self.frame_lock:
-            self.latest_frame = None
-            self.latest_timestamp = None
-            self.new_frame_event.clear()
-
         logger.info(f"{self} disconnected.")

src/lerobot/cameras/zmq/configuration_zmq.py

@@ -29,10 +29,12 @@ class ZMQCameraConfig(CameraConfig):
     camera_name: str = "zmq_camera"
     color_mode: ColorMode = ColorMode.RGB
     timeout_ms: int = 5000
-    warmup_s: int = 1
 
     def __post_init__(self) -> None:
-        self.color_mode = ColorMode(self.color_mode)
+        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
+            raise ValueError(
+                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
+            )
 
         if self.timeout_ms <= 0:
             raise ValueError(f"`timeout_ms` must be positive, but {self.timeout_ms} is provided.")

src/lerobot/configs/policies.py

@@ -45,12 +45,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy.
+        output_shapes: A dictionary defining the shapes of the output data for the policy.
+        input_normalization_modes: A dictionary with key representing the modality and the value specifies the
+            normalization mode to apply.
+        output_normalization_modes: Similar dictionary as `input_normalization_modes`, but to unnormalize to
+            the original scale.
     """
 
     n_obs_steps: int = 1

src/lerobot/configs/train.py

@@ -29,6 +29,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.optim import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
 from lerobot.utils.hub import HubMixin
+from lerobot.utils.sample_weighting import SampleWeightingConfig
 
 TRAIN_CONFIG_NAME = "train_config.json"
 
@@ -67,12 +68,8 @@ class TrainPipelineConfig(HubMixin):
     wandb: WandBConfig = field(default_factory=WandBConfig)
     peft: PeftConfig | None = None
 
-    # RA-BC (Reward-Aligned Behavior Cloning) parameters
-    use_rabc: bool = False  # Enable reward-weighted training
-    rabc_progress_path: str | None = None  # Path to precomputed SARM progress parquet file
-    rabc_kappa: float = 0.01  # Hard threshold for high-quality samples
-    rabc_epsilon: float = 1e-6  # Small constant for numerical stability
-    rabc_head_mode: str | None = "sparse"  # For dual-head models: "sparse" or "dense"
+    # Sample weighting configuration (e.g., for RA-BC training)
+    sample_weighting: SampleWeightingConfig | None = None
 
     # Rename map for the observation to override the image and state keys
     rename_map: dict[str, str] = field(default_factory=dict)
@@ -140,14 +137,6 @@ class TrainPipelineConfig(HubMixin):
                 "'policy.repo_id' argument missing. Please specify it to push the model to the hub."
             )
 
-        if self.use_rabc and not self.rabc_progress_path:
-            # Auto-detect from dataset path
-            repo_id = self.dataset.repo_id
-            if self.dataset.root:
-                self.rabc_progress_path = str(Path(self.dataset.root) / "sarm_progress.parquet")
-            else:
-                self.rabc_progress_path = f"hf://datasets/{repo_id}/sarm_progress.parquet"
-
     @classmethod
     def __get_path_fields__(cls) -> list[str]:
         """This enables the parser to load config from the policy using `--policy.path=local/dir`"""

src/lerobot/datasets/aggregate.py

@@ -116,9 +116,6 @@ def update_meta_data(
     Adjusts all indices and timestamps to account for previously aggregated
     data and videos in the destination dataset.
 
-    For data file indices, uses the 'src_to_dst' mapping from aggregate_data()
-    to correctly map source file indices to their destination locations.
-
     Args:
         df: DataFrame containing the metadata to be updated.
         dst_meta: Destination dataset metadata.
@@ -132,50 +129,8 @@ def update_meta_data(
 
     df["meta/episodes/chunk_index"] = df["meta/episodes/chunk_index"] + meta_idx["chunk"]
     df["meta/episodes/file_index"] = df["meta/episodes/file_index"] + meta_idx["file"]
-
-    # Update data file indices using source-to-destination mapping
-    # This is critical for handling datasets that are already results of a merge
-    data_src_to_dst = data_idx.get("src_to_dst", {})
-    if data_src_to_dst:
-        # Store original indices for lookup
-        df["_orig_data_chunk"] = df["data/chunk_index"].copy()
-        df["_orig_data_file"] = df["data/file_index"].copy()
-
-        # Vectorized mapping from (src_chunk, src_file) to (dst_chunk, dst_file)
-        # This is much faster than per-row iteration for large metadata tables
-        mapping_index = pd.MultiIndex.from_tuples(
-            list(data_src_to_dst.keys()),
-            names=["chunk_index", "file_index"],
-        )
-        mapping_values = list(data_src_to_dst.values())
-        mapping_df = pd.DataFrame(
-            mapping_values,
-            index=mapping_index,
-            columns=["dst_chunk", "dst_file"],
-        )
-
-        # Construct a MultiIndex for each row based on original data indices
-        row_index = pd.MultiIndex.from_arrays(
-            [df["_orig_data_chunk"], df["_orig_data_file"]],
-            names=["chunk_index", "file_index"],
-        )
-
-        # Align mapping to rows; missing keys fall back to the default destination
-        reindexed = mapping_df.reindex(row_index)
-        reindexed[["dst_chunk", "dst_file"]] = reindexed[["dst_chunk", "dst_file"]].fillna(
-            {"dst_chunk": data_idx["chunk"], "dst_file": data_idx["file"]}
-        )
-
-        # Assign mapped destination indices back to the DataFrame
-        df["data/chunk_index"] = reindexed["dst_chunk"].to_numpy()
-        df["data/file_index"] = reindexed["dst_file"].to_numpy()
-
-        # Clean up temporary columns
-        df = df.drop(columns=["_orig_data_chunk", "_orig_data_file"])
-    else:
-        # Fallback to simple offset (backward compatibility for single-file sources)
-        df["data/chunk_index"] = df["data/chunk_index"] + data_idx["chunk"]
-        df["data/file_index"] = df["data/file_index"] + data_idx["file"]
+    df["data/chunk_index"] = df["data/chunk_index"] + data_idx["chunk"]
+    df["data/file_index"] = df["data/file_index"] + data_idx["file"]
     for key, video_idx in videos_idx.items():
         # Store original video file indices before updating
         orig_chunk_col = f"videos/{key}/chunk_index"
@@ -191,7 +146,8 @@ def update_meta_data(
         if src_to_dst:
             # Map each episode to its correct destination file and apply offset
             for idx in df.index:
-                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
+                # Convert to Python int to avoid numpy type mismatch in dict lookup
+                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
 
                 # Get destination chunk/file for this source file
                 dst_chunk, dst_file = src_to_dst.get(src_key, (video_idx["chunk"], video_idx["file"]))
@@ -207,7 +163,8 @@ def update_meta_data(
             df[orig_chunk_col] = video_idx["chunk"]
             df[orig_file_col] = video_idx["file"]
             for idx in df.index:
-                src_key = (df.at[idx, "_orig_chunk"], df.at[idx, "_orig_file"])
+                # Convert to Python int to avoid numpy type mismatch in dict lookup
+                src_key = (int(df.at[idx, "_orig_chunk"]), int(df.at[idx, "_orig_file"]))
                 offset = src_to_offset.get(src_key, 0)
                 df.at[idx, f"videos/{key}/from_timestamp"] += offset
                 df.at[idx, f"videos/{key}/to_timestamp"] += offset
@@ -305,10 +262,6 @@ def aggregate_datasets(
 
         meta_idx = aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx)
 
-        # Clear the src_to_dst mapping after processing each source dataset
-        # to avoid interference between different source datasets
-        data_idx.pop("src_to_dst", None)
-
         dst_meta.info["total_episodes"] += src_meta.total_episodes
         dst_meta.info["total_frames"] += src_meta.total_frames
 
@@ -359,6 +312,10 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
         dst_file_durations = video_idx["dst_file_durations"]
 
         for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
+            # Convert to Python int to ensure consistent dict keys
+            src_chunk_idx = int(src_chunk_idx)
+            src_file_idx = int(src_file_idx)
+
             src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
                 video_key=key,
                 chunk_index=src_chunk_idx,
@@ -431,16 +388,10 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
     Reads source data files, updates indices to match the aggregated dataset,
     and writes them to the destination with proper file rotation.
 
-    Tracks a `src_to_dst` mapping from source (chunk, file) to destination (chunk, file)
-    which is critical for correctly updating episode metadata when source datasets
-    have multiple data files (e.g., from a previous merge operation).
-
     Args:
         src_meta: Source dataset metadata.
         dst_meta: Destination dataset metadata.
         data_idx: Dictionary tracking data chunk and file indices.
-        data_files_size_in_mb: Maximum size for data files in MB.
-        chunk_size: Maximum number of files per chunk.
 
     Returns:
         dict: Updated data_idx with current chunk and file indices.
@@ -458,10 +409,6 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
     # retrieve features schema for proper image typing in parquet
     hf_features = get_hf_features_from_features(dst_meta.features) if contains_images else None
 
-    # Track source to destination file mapping for metadata update
-    # This is critical for handling datasets that are already results of a merge
-    src_to_dst: dict[tuple[int, int], tuple[int, int]] = {}
-
     for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
         src_path = src_meta.root / DEFAULT_DATA_PATH.format(
             chunk_index=src_chunk_idx, file_index=src_file_idx
@@ -474,9 +421,7 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             df = pd.read_parquet(src_path)
         df = update_data_df(df, src_meta, dst_meta)
 
-        # Write data and get the actual destination file it was written to
-        # This avoids duplicating the rotation logic here
-        data_idx, (dst_chunk, dst_file) = append_or_create_parquet_file(
+        data_idx = append_or_create_parquet_file(
             df,
             src_path,
             data_idx,
@@ -488,12 +433,6 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             hf_features=hf_features,
         )
 
-        # Record the mapping from source to actual destination
-        src_to_dst[(src_chunk_idx, src_file_idx)] = (dst_chunk, dst_file)
-
-    # Add the mapping to data_idx for use in metadata update
-    data_idx["src_to_dst"] = src_to_dst
-
     return data_idx
 
 
@@ -534,7 +473,7 @@ def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
             videos_idx,
         )
 
-        meta_idx, _ = append_or_create_parquet_file(
+        meta_idx = append_or_create_parquet_file(
             df,
             src_path,
             meta_idx,
@@ -562,7 +501,7 @@ def append_or_create_parquet_file(
     contains_images: bool = False,
     aggr_root: Path = None,
     hf_features: datasets.Features | None = None,
-) -> tuple[dict[str, int], tuple[int, int]]:
+):
     """Appends data to an existing parquet file or creates a new one based on size constraints.
 
     Manages file rotation when size limits are exceeded to prevent individual files
@@ -580,11 +519,9 @@ def append_or_create_parquet_file(
         hf_features: Optional HuggingFace Features schema for proper image typing.
 
     Returns:
-        tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
-               and (dst_chunk, dst_file) is the actual destination file the data was written to.
+        dict: Updated index dictionary with current chunk and file indices.
     """
-    dst_chunk, dst_file = idx["chunk"], idx["file"]
-    dst_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
+    dst_path = aggr_root / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
 
     if not dst_path.exists():
         dst_path.parent.mkdir(parents=True, exist_ok=True)
@@ -592,15 +529,14 @@ def append_or_create_parquet_file(
             to_parquet_with_hf_images(df, dst_path, features=hf_features)
         else:
             df.to_parquet(dst_path)
-        return idx, (dst_chunk, dst_file)
+        return idx
 
     src_size = get_parquet_file_size_in_mb(src_path)
     dst_size = get_parquet_file_size_in_mb(dst_path)
 
     if dst_size + src_size >= max_mb:
         idx["chunk"], idx["file"] = update_chunk_file_indices(idx["chunk"], idx["file"], chunk_size)
-        dst_chunk, dst_file = idx["chunk"], idx["file"]
-        new_path = aggr_root / default_path.format(chunk_index=dst_chunk, file_index=dst_file)
+        new_path = aggr_root / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
         new_path.parent.mkdir(parents=True, exist_ok=True)
         final_df = df
         target_path = new_path
@@ -619,7 +555,7 @@ def append_or_create_parquet_file(
     else:
         final_df.to_parquet(target_path)
 
-    return idx, (dst_chunk, dst_file)
+    return idx
 
 
 def finalize_aggregation(aggr_meta, all_metadata):

src/lerobot/datasets/dataset_tools.py

@@ -1396,132 +1396,6 @@ BYTES_PER_KIB = 1024
 BYTES_PER_MIB = BYTES_PER_KIB * BYTES_PER_KIB
 
 
-def modify_tasks(
-    dataset: LeRobotDataset,
-    new_task: str | None = None,
-    episode_tasks: dict[int, str] | None = None,
-) -> LeRobotDataset:
-    """Modify tasks in a LeRobotDataset.
-
-    This function allows you to either:
-    1. Set a single task for the entire dataset (using `new_task`)
-    2. Set specific tasks for specific episodes (using `episode_tasks`)
-
-    You can combine both: `new_task` sets the default, and `episode_tasks` overrides
-    specific episodes.
-
-    The dataset is modified in-place, updating only the task-related files:
-    - meta/tasks.parquet
-    - data/**/*.parquet (task_index column)
-    - meta/episodes/**/*.parquet (tasks column)
-    - meta/info.json (total_tasks)
-
-    Args:
-        dataset: The source LeRobotDataset to modify.
-        new_task: A single task string to apply to all episodes. If None and episode_tasks
-            is also None, raises an error.
-        episode_tasks: Optional dict mapping episode indices to their task strings.
-            Overrides `new_task` for specific episodes.
-
-
-    Examples:
-        Set a single task for all episodes:
-            dataset = modify_tasks(dataset, new_task="Pick up the cube")
-
-        Set different tasks for specific episodes:
-            dataset = modify_tasks(
-                dataset,
-                episode_tasks={0: "Task A", 1: "Task B", 2: "Task A"}
-            )
-
-        Set a default task with overrides:
-            dataset = modify_tasks(
-                dataset,
-                new_task="Default task",
-                episode_tasks={5: "Special task for episode 5"}
-            )
-    """
-    if new_task is None and episode_tasks is None:
-        raise ValueError("Must specify at least one of new_task or episode_tasks")
-
-    if episode_tasks is not None:
-        valid_indices = set(range(dataset.meta.total_episodes))
-        invalid = set(episode_tasks.keys()) - valid_indices
-        if invalid:
-            raise ValueError(f"Invalid episode indices: {invalid}")
-
-    # Ensure episodes metadata is loaded
-    if dataset.meta.episodes is None:
-        dataset.meta.episodes = load_episodes(dataset.root)
-
-    # Build the mapping from episode index to task string
-    episode_to_task: dict[int, str] = {}
-    for ep_idx in range(dataset.meta.total_episodes):
-        if episode_tasks and ep_idx in episode_tasks:
-            episode_to_task[ep_idx] = episode_tasks[ep_idx]
-        elif new_task is not None:
-            episode_to_task[ep_idx] = new_task
-        else:
-            # Keep original task if not overridden and no default provided
-            original_tasks = dataset.meta.episodes[ep_idx]["tasks"]
-            if not original_tasks:
-                raise ValueError(f"Episode {ep_idx} has no tasks and no default task was provided")
-            episode_to_task[ep_idx] = original_tasks[0]
-
-    # Collect all unique tasks and create new task mapping
-    unique_tasks = sorted(set(episode_to_task.values()))
-    new_task_df = pd.DataFrame({"task_index": list(range(len(unique_tasks)))}, index=unique_tasks)
-    task_to_index = {task: idx for idx, task in enumerate(unique_tasks)}
-
-    logging.info(f"Modifying tasks in {dataset.repo_id}")
-    logging.info(f"New tasks: {unique_tasks}")
-
-    root = dataset.root
-
-    # Update data files - modify task_index column
-    logging.info("Updating data files...")
-    data_dir = root / DATA_DIR
-
-    for parquet_path in tqdm(sorted(data_dir.rglob("*.parquet")), desc="Updating data"):
-        df = pd.read_parquet(parquet_path)
-
-        # Build a mapping from episode_index to new task_index for rows in this file
-        episode_indices_in_file = df["episode_index"].unique()
-        ep_to_new_task_idx = {
-            ep_idx: task_to_index[episode_to_task[ep_idx]] for ep_idx in episode_indices_in_file
-        }
-
-        # Update task_index column
-        df["task_index"] = df["episode_index"].map(ep_to_new_task_idx)
-        df.to_parquet(parquet_path, index=False)
-
-    # Update episodes metadata - modify tasks column
-    logging.info("Updating episodes metadata...")
-    episodes_dir = root / "meta" / "episodes"
-
-    for parquet_path in tqdm(sorted(episodes_dir.rglob("*.parquet")), desc="Updating episodes"):
-        df = pd.read_parquet(parquet_path)
-
-        # Update tasks column
-        df["tasks"] = df["episode_index"].apply(lambda ep_idx: [episode_to_task[ep_idx]])
-        df.to_parquet(parquet_path, index=False)
-
-    # Write new tasks.parquet
-    write_tasks(new_task_df, root)
-
-    # Update info.json
-    dataset.meta.info["total_tasks"] = len(unique_tasks)
-    write_info(dataset.meta.info, root)
-
-    # Reload metadata to reflect changes
-    dataset.meta.tasks = new_task_df
-    dataset.meta.episodes = load_episodes(root)
-
-    logging.info(f"Tasks: {unique_tasks}")
-
-    return dataset
-
-
 def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
     output_dir: Path,

src/lerobot/datasets/lerobot_dataset.py

@@ -57,7 +57,6 @@ from lerobot.datasets.utils import (
     load_info,
     load_nested_dataset,
     load_stats,
-    load_subtasks,
     load_tasks,
     update_chunk_file_indices,
     validate_episode_buffer,
@@ -163,7 +162,6 @@ class LeRobotDatasetMetadata:
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
         self.tasks = load_tasks(self.root)
-        self.subtasks = load_subtasks(self.root)
         self.episodes = load_episodes(self.root)
         self.stats = load_stats(self.root)
 
@@ -520,7 +518,6 @@ class LeRobotDatasetMetadata:
         _validate_feature_names(features)
 
         obj.tasks = None
-        obj.subtasks = None
         obj.episodes = None
         obj.stats = None
         obj.info = create_empty_dataset_info(
@@ -1078,12 +1075,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         # Add task as a string
         task_idx = item["task_index"].item()
         item["task"] = self.meta.tasks.iloc[task_idx].name
-
-        # add subtask information if available
-        if "subtask_index" in self.features and self.meta.subtasks is not None:
-            subtask_idx = item["subtask_index"].item()
-            item["subtask"] = self.meta.subtasks.iloc[subtask_idx].name
-
         return item
 
     def __repr__(self):

src/lerobot/datasets/transforms.py

@@ -216,17 +216,16 @@ class ImageTransformsConfig:
 
 
 def make_transform_from_config(cfg: ImageTransformConfig):
-    if cfg.type == "SharpnessJitter":
+    if cfg.type == "Identity":
+        return v2.Identity(**cfg.kwargs)
+    elif cfg.type == "ColorJitter":
+        return v2.ColorJitter(**cfg.kwargs)
+    elif cfg.type == "SharpnessJitter":
         return SharpnessJitter(**cfg.kwargs)
-
-    transform_cls = getattr(v2, cfg.type, None)
-    if isinstance(transform_cls, type) and issubclass(transform_cls, Transform):
-        return transform_cls(**cfg.kwargs)
-
-    raise ValueError(
-        f"Transform '{cfg.type}' is not valid. It must be a class in "
-        f"torchvision.transforms.v2 or 'SharpnessJitter'."
-    )
+    elif cfg.type == "RandomAffine":
+        return v2.RandomAffine(**cfg.kwargs)
+    else:
+        raise ValueError(f"Transform '{cfg.type}' is not valid.")
 
 
 class ImageTransforms(Transform):

src/lerobot/datasets/utils.py

@@ -60,7 +60,6 @@ VIDEO_DIR = "videos"
 
 CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
 DEFAULT_TASKS_PATH = "meta/tasks.parquet"
-DEFAULT_SUBTASKS_PATH = "meta/subtasks.parquet"
 DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
 DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
@@ -354,14 +353,6 @@ def load_tasks(local_dir: Path) -> pandas.DataFrame:
     return tasks
 
 
-def load_subtasks(local_dir: Path) -> pandas.DataFrame | None:
-    """Load subtasks from subtasks.parquet if it exists."""
-    subtasks_path = local_dir / DEFAULT_SUBTASKS_PATH
-    if subtasks_path.exists():
-        return pd.read_parquet(subtasks_path)
-    return None
-
-
 def write_episodes(episodes: Dataset, local_dir: Path) -> None:
     """Write episode metadata to a parquet file in the LeRobot v3.0 format.
     This function writes episode-level metadata to a single parquet file.

src/lerobot/envs/configs.py

@@ -205,7 +205,6 @@ class ObservationConfig:
 
     add_joint_velocity_to_observation: bool = False
     add_current_to_observation: bool = False
-    add_ee_pose_to_observation: bool = False
     display_cameras: bool = False
 
 
@@ -261,7 +260,6 @@ class HILSerlRobotEnvConfig(EnvConfig):
 @dataclass
 class LiberoEnv(EnvConfig):
     task: str = "libero_10"  # can also choose libero_spatial, libero_object, etc.
-    task_ids: list[int] | None = None
     fps: int = 30
     episode_length: int | None = None
     obs_type: str = "pixels_agent_pos"
@@ -340,10 +338,10 @@ class LiberoEnv(EnvConfig):
 
     @property
     def gym_kwargs(self) -> dict:
-        kwargs: dict[str, Any] = {"obs_type": self.obs_type, "render_mode": self.render_mode}
-        if self.task_ids is not None:
-            kwargs["task_ids"] = self.task_ids
-        return kwargs
+        return {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+        }
 
 
 @EnvConfig.register_subclass("metaworld")

src/lerobot/envs/libero.py

@@ -112,7 +112,6 @@ class LiberoEnv(gym.Env):
         visualization_height: int = 480,
         init_states: bool = True,
         episode_index: int = 0,
-        n_envs: int = 1,
         camera_name_mapping: dict[str, str] | None = None,
         num_steps_wait: int = 10,
         control_mode: str = "relative",
@@ -146,9 +145,7 @@ class LiberoEnv(gym.Env):
         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._reset_stride = n_envs  # when performing a reset, append `_reset_stride` to `init_state_id`.
-
-        self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state
+        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
@@ -298,8 +295,7 @@ class LiberoEnv(gym.Env):
         self._env.seed(seed)
         raw_obs = self._env.reset()
         if self.init_states and self._init_states is not None:
-            raw_obs = self._env.set_init_state(self._init_states[self.init_state_id % len(self._init_states)])
-            self.init_state_id += self._reset_stride  # Change init_state_id when reset
+            raw_obs = self._env.set_init_state(self._init_states[self._init_state_id])
 
         # After reset, objects may be unstable (slightly floating, intersecting, etc.).
         # Step the simulator with a no-op action for a few frames so everything settles.
@@ -377,7 +373,6 @@ def _make_env_fns(
             init_states=init_states,
             episode_length=episode_length,
             episode_index=episode_index,
-            n_envs=n_envs,
             control_mode=control_mode,
             **local_kwargs,
         )

src/lerobot/motors/__init__.py

@@ -14,8 +14,4 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .motors_bus import (
-    Motor,
-    MotorCalibration,
-    MotorNormMode,
-)
+from .motors_bus import Motor, MotorCalibration, MotorNormMode, MotorsBus

src/lerobot/motors/calibration_gui.py

@@ -18,7 +18,7 @@ from dataclasses import dataclass
 
 os.environ["PYGAME_HIDE_SUPPORT_PROMPT"] = "1"
 
-from .motors_bus import MotorCalibration, MotorsBus
+from lerobot.motors import MotorCalibration, MotorsBus
 
 BAR_LEN, BAR_THICKNESS = 450, 8
 HANDLE_R = 10
@@ -221,7 +221,7 @@ class RangeFinderGUI:
 
         self.bus = bus
         self.groups = groups if groups is not None else {"all": list(bus.motors)}
-        self.group_names = list(self.groups)
+        self.group_names = list(groups)
         self.current_group = self.group_names[0]
 
         if not bus.is_connected:
@@ -230,20 +230,18 @@ class RangeFinderGUI:
         self.calibration = bus.read_calibration()
         self.res_table = bus.model_resolution_table
         self.present_cache = {
-            m: bus.read("Present_Position", m, normalize=False)
-            for motors in self.groups.values()
-            for m in motors
+            m: bus.read("Present_Position", m, normalize=False) for motors in groups.values() for m in motors
         }
 
         pygame.init()
         self.font = pygame.font.Font(None, FONT_SIZE)
 
-        label_pad = max(self.font.size(m)[0] for ms in self.groups.values() for m in ms)
+        label_pad = max(self.font.size(m)[0] for ms in groups.values() for m in ms)
         self.label_pad = label_pad
         width = 40 + label_pad + BAR_LEN + 6 + BTN_W + 10 + SAVE_W + 10
         self.controls_bottom = 10 + SAVE_H
         self.base_y = self.controls_bottom + TOP_GAP
-        height = self.base_y + PADDING_Y * len(self.groups[self.current_group]) + 40
+        height = self.base_y + PADDING_Y * len(groups[self.current_group]) + 40
 
         self.screen = pygame.display.set_mode((width, height))
         pygame.display.set_caption("Motors range finder")

src/lerobot/motors/damiao/__init__.py

@@ -1,18 +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 .damiao import DamiaoMotorsBus
-from .tables import *

src/lerobot/motors/damiao/damiao.py

@@ -1,859 +0,0 @@
-# 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.
-
-# Portions of this file are derived from DM_Control_Python by cmjang.
-# Licensed under the MIT License; see `LICENSE` for the full text:
-# https://github.com/cmjang/DM_Control_Python
-
-import logging
-import time
-from contextlib import contextmanager
-from copy import deepcopy
-from functools import cached_property
-from typing import TYPE_CHECKING, Any, TypedDict
-
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-from lerobot.utils.import_utils import _can_available
-
-if TYPE_CHECKING or _can_available:
-    import can
-else:
-
-    class can:  # noqa: N801
-        Message = object
-        interface = None
-
-
-import numpy as np
-
-from lerobot.utils.robot_utils import precise_sleep
-from lerobot.utils.utils import enter_pressed, move_cursor_up
-
-from ..motors_bus import Motor, MotorCalibration, MotorsBusBase, NameOrID, Value
-from .tables import (
-    AVAILABLE_BAUDRATES,
-    CAN_CMD_DISABLE,
-    CAN_CMD_ENABLE,
-    CAN_CMD_REFRESH,
-    CAN_CMD_SET_ZERO,
-    CAN_PARAM_ID,
-    DEFAULT_BAUDRATE,
-    DEFAULT_TIMEOUT_MS,
-    MIT_KD_RANGE,
-    MIT_KP_RANGE,
-    MOTOR_LIMIT_PARAMS,
-    MotorType,
-)
-
-logger = logging.getLogger(__name__)
-
-
-LONG_TIMEOUT_SEC = 0.1
-MEDIUM_TIMEOUT_SEC = 0.01
-SHORT_TIMEOUT_SEC = 0.001
-PRECISE_TIMEOUT_SEC = 0.0001
-
-
-class MotorState(TypedDict):
-    position: float
-    velocity: float
-    torque: float
-    temp_mos: float
-    temp_rotor: float
-
-
-class DamiaoMotorsBus(MotorsBusBase):
-    """
-    The Damiao implementation for a MotorsBus using CAN bus communication.
-
-    This class uses python-can for CAN bus communication with Damiao motors.
-    For more info, see:
-    - python-can documentation: https://python-can.readthedocs.io/en/stable/
-    - Seedstudio documentation: https://wiki.seeedstudio.com/damiao_series/
-    - DM_Control_Python repo: https://github.com/cmjang/DM_Control_Python
-    """
-
-    # CAN-specific settings
-    available_baudrates = deepcopy(AVAILABLE_BAUDRATES)
-    default_baudrate = DEFAULT_BAUDRATE
-    default_timeout = DEFAULT_TIMEOUT_MS
-
-    def __init__(
-        self,
-        port: str,
-        motors: dict[str, Motor],
-        calibration: dict[str, MotorCalibration] | None = None,
-        can_interface: str = "auto",
-        use_can_fd: bool = True,
-        bitrate: int = 1000000,
-        data_bitrate: int | None = 5000000,
-    ):
-        """
-        Initialize the Damiao motors bus.
-
-        Args:
-            port: CAN interface name (e.g., "can0" for Linux, "/dev/cu.usbmodem*" for macOS)
-            motors: Dictionary mapping motor names to Motor objects
-            calibration: Optional calibration data
-            can_interface: CAN interface type - "auto" (default), "socketcan" (Linux), or "slcan" (macOS/serial)
-            use_can_fd: Whether to use CAN FD mode (default: True for OpenArms)
-            bitrate: Nominal bitrate in bps (default: 1000000 = 1 Mbps)
-            data_bitrate: Data bitrate for CAN FD in bps (default: 5000000 = 5 Mbps), ignored if use_can_fd is False
-        """
-        super().__init__(port, motors, calibration)
-        self.port = port
-        self.can_interface = can_interface
-        self.use_can_fd = use_can_fd
-        self.bitrate = bitrate
-        self.data_bitrate = data_bitrate
-        self.canbus: can.interface.Bus | None = None
-        self._is_connected = False
-
-        # Map motor names to CAN IDs
-        self._motor_can_ids: dict[str, int] = {}
-        self._recv_id_to_motor: dict[int, str] = {}
-        self._motor_types: dict[str, MotorType] = {}
-
-        for name, motor in self.motors.items():
-            if motor.motor_type_str is None:
-                raise ValueError(f"Motor '{name}' is missing required 'motor_type'")
-            self._motor_types[name] = getattr(MotorType, motor.motor_type_str.upper().replace("-", "_"))
-
-            # Map recv_id to motor name for filtering responses
-            if motor.recv_id is not None:
-                self._recv_id_to_motor[motor.recv_id] = name
-
-        # State cache for handling packet drops safely
-        self._last_known_states: dict[str, MotorState] = {
-            name: {
-                "position": 0.0,
-                "velocity": 0.0,
-                "torque": 0.0,
-                "temp_mos": 0.0,
-                "temp_rotor": 0.0,
-            }
-            for name in self.motors
-        }
-
-        # Dynamic gains storage
-        # Defaults: Kp=10.0 (Stiffness), Kd=0.5 (Damping)
-        self._gains: dict[str, dict[str, float]] = {name: {"kp": 10.0, "kd": 0.5} for name in self.motors}
-
-    @property
-    def is_connected(self) -> bool:
-        """Check if the CAN bus is connected."""
-        return self._is_connected and self.canbus is not None
-
-    @check_if_already_connected
-    def connect(self, handshake: bool = True) -> None:
-        """
-        Open the CAN bus and initialize communication.
-
-        Args:
-            handshake: If True, ping all motors to verify they're present
-        """
-
-        try:
-            # Auto-detect interface type based on port name
-            if self.can_interface == "auto":
-                if self.port.startswith("/dev/"):
-                    self.can_interface = "slcan"
-                    logger.info(f"Auto-detected slcan interface for port {self.port}")
-                else:
-                    self.can_interface = "socketcan"
-                    logger.info(f"Auto-detected socketcan interface for port {self.port}")
-
-            # Connect to CAN bus
-            kwargs = {
-                "channel": self.port,
-                "bitrate": self.bitrate,
-                "interface": self.can_interface,
-            }
-
-            if self.can_interface == "socketcan" and self.use_can_fd and self.data_bitrate is not None:
-                kwargs.update({"data_bitrate": self.data_bitrate, "fd": True})
-                logger.info(
-                    f"Connected to {self.port} with CAN FD (bitrate={self.bitrate}, data_bitrate={self.data_bitrate})"
-                )
-            else:
-                logger.info(f"Connected to {self.port} with {self.can_interface} (bitrate={self.bitrate})")
-
-            self.canbus = can.interface.Bus(**kwargs)
-            self._is_connected = True
-
-            if handshake:
-                self._handshake()
-
-            logger.debug(f"{self.__class__.__name__} connected via {self.can_interface}.")
-        except Exception as e:
-            self._is_connected = False
-            raise ConnectionError(f"Failed to connect to CAN bus: {e}") from e
-
-    def _handshake(self) -> None:
-        """
-        Verify all motors are present and populate initial state cache.
-        Raises ConnectionError if any motor fails to respond.
-        """
-        logger.info("Starting handshake with motors...")
-
-        # Drain any pending messages
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
-        while self.canbus.recv(timeout=0.01):
-            pass
-
-        missing_motors = []
-        for motor_name in self.motors:
-            motor_id = self._get_motor_id(motor_name)
-            recv_id = self._get_motor_recv_id(motor_name)
-
-            # Send enable command
-            data = [0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, CAN_CMD_ENABLE]
-            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
-            self.canbus.send(msg)
-
-            # Wait for response with longer timeout
-            response = None
-            start_time = time.time()
-            while time.time() - start_time < 0.1:
-                response = self.canbus.recv(timeout=0.1)
-                if response and response.arbitration_id == recv_id:
-                    break
-                response = None
-
-            if response is None:
-                missing_motors.append(motor_name)
-            else:
-                self._process_response(motor_name, msg)
-            time.sleep(MEDIUM_TIMEOUT_SEC)
-
-        if missing_motors:
-            raise ConnectionError(
-                f"Handshake failed. The following motors did not respond: {missing_motors}. "
-                "Check power (24V) and CAN wiring."
-            )
-        logger.info("Handshake successful. All motors ready.")
-
-    @check_if_not_connected
-    def disconnect(self, disable_torque: bool = True) -> None:
-        """
-        Close the CAN bus connection.
-
-        Args:
-            disable_torque: If True, disable torque on all motors before disconnecting
-        """
-
-        if disable_torque:
-            try:
-                self.disable_torque()
-            except Exception as e:
-                logger.warning(f"Failed to disable torque during disconnect: {e}")
-
-        if self.canbus:
-            self.canbus.shutdown()
-            self.canbus = None
-        self._is_connected = False
-        logger.debug(f"{self.__class__.__name__} disconnected.")
-
-    def configure_motors(self) -> None:
-        """Configure all motors with default settings."""
-        # Damiao motors don't require much configuration in MIT mode
-        # Just ensure they're enabled
-        for motor in self.motors:
-            self._send_simple_command(motor, CAN_CMD_ENABLE)
-            time.sleep(MEDIUM_TIMEOUT_SEC)
-
-    def _send_simple_command(self, motor: NameOrID, command_byte: int) -> None:
-        """Helper to send simple 8-byte commands (Enable, Disable, Zero)."""
-        motor_id = self._get_motor_id(motor)
-        motor_name = self._get_motor_name(motor)
-        recv_id = self._get_motor_recv_id(motor)
-        data = [0xFF] * 7 + [command_byte]
-        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
-        self.canbus.send(msg)
-        if msg := self._recv_motor_response(expected_recv_id=recv_id):
-            self._process_response(motor_name, msg)
-        else:
-            logger.debug(f"No response from {motor_name} after command 0x{command_byte:02X}")
-
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
-        """Enable torque on selected motors."""
-        target_motors = self._get_motors_list(motors)
-        for motor in target_motors:
-            for _ in range(num_retry + 1):
-                try:
-                    self._send_simple_command(motor, CAN_CMD_ENABLE)
-                    break
-                except Exception as e:
-                    if _ == num_retry:
-                        raise e
-                    time.sleep(MEDIUM_TIMEOUT_SEC)
-
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
-        """Disable torque on selected motors."""
-        target_motors = self._get_motors_list(motors)
-        for motor in target_motors:
-            for _ in range(num_retry + 1):
-                try:
-                    self._send_simple_command(motor, CAN_CMD_DISABLE)
-                    break
-                except Exception as e:
-                    if _ == num_retry:
-                        raise e
-                    time.sleep(MEDIUM_TIMEOUT_SEC)
-
-    @contextmanager
-    def torque_disabled(self, motors: str | list[str] | None = None):
-        """
-        Context manager that guarantees torque is re-enabled.
-
-        This helper is useful to temporarily disable torque when configuring motors.
-        """
-        self.disable_torque(motors)
-        try:
-            yield
-        finally:
-            self.enable_torque(motors)
-
-    def set_zero_position(self, motors: str | list[str] | None = None) -> None:
-        """Set current position as zero for selected motors."""
-        target_motors = self._get_motors_list(motors)
-        for motor in target_motors:
-            self._send_simple_command(motor, CAN_CMD_SET_ZERO)
-            time.sleep(MEDIUM_TIMEOUT_SEC)
-
-    def _refresh_motor(self, motor: NameOrID) -> can.Message | None:
-        """Refresh motor status and return the response."""
-        motor_id = self._get_motor_id(motor)
-        recv_id = self._get_motor_recv_id(motor)
-        data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
-        msg = can.Message(arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd)
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
-        self.canbus.send(msg)
-        return self._recv_motor_response(expected_recv_id=recv_id)
-
-    def _recv_motor_response(
-        self, expected_recv_id: int | None = None, timeout: float = 0.001
-    ) -> can.Message | None:
-        """
-        Receive a response from a motor.
-
-        Args:
-            expected_recv_id: If provided, only return messages from this CAN ID
-            timeout: Timeout in seconds (default: 1ms for high-speed operation)
-        Returns:
-            CAN message if received, None otherwise
-        """
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
-        try:
-            start_time = time.time()
-            messages_seen = []
-            while time.time() - start_time < timeout:
-                msg = self.canbus.recv(timeout=PRECISE_TIMEOUT_SEC)
-                if msg:
-                    messages_seen.append(f"0x{msg.arbitration_id:02X}")
-                    if expected_recv_id is None or msg.arbitration_id == expected_recv_id:
-                        return msg
-                    logger.debug(
-                        f"Ignoring message from 0x{msg.arbitration_id:02X}, expected 0x{expected_recv_id:02X}"
-                    )
-
-            if logger.isEnabledFor(logging.DEBUG):
-                if messages_seen:
-                    logger.debug(
-                        f"Received {len(messages_seen)} msgs from {set(messages_seen)}, expected 0x{expected_recv_id:02X}"
-                    )
-                else:
-                    logger.debug(f"No CAN messages received (expected 0x{expected_recv_id:02X})")
-        except Exception as e:
-            logger.debug(f"Failed to receive CAN message: {e}")
-        return None
-
-    def _recv_all_responses(
-        self, expected_recv_ids: list[int], timeout: float = 0.002
-    ) -> dict[int, can.Message]:
-        """
-        Efficiently receive responses from multiple motors at once.
-        Uses the OpenArms pattern: collect all available messages within timeout.
-
-        Args:
-            expected_recv_ids: List of CAN IDs we expect responses from
-            timeout: Total timeout in seconds (default: 2ms)
-
-        Returns:
-            Dictionary mapping recv_id to CAN message
-        """
-        responses: dict[int, can.Message] = {}
-        expected_set = set(expected_recv_ids)
-        start_time = time.time()
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
-        try:
-            while len(responses) < len(expected_recv_ids) and (time.time() - start_time) < timeout:
-                # 100us poll timeout
-                msg = self.canbus.recv(timeout=PRECISE_TIMEOUT_SEC)
-                if msg and msg.arbitration_id in expected_set:
-                    responses[msg.arbitration_id] = msg
-                    if len(responses) == len(expected_recv_ids):
-                        break
-        except Exception as e:
-            logger.debug(f"Error receiving responses: {e}")
-
-        return responses
-
-    def _encode_mit_packet(
-        self,
-        motor_type: MotorType,
-        kp: float,
-        kd: float,
-        position_degrees: float,
-        velocity_deg_per_sec: float,
-        torque: float,
-    ) -> list[int]:
-        """Helper to encode control parameters into 8 bytes for MIT mode."""
-        # Convert degrees to radians
-        position_rad = np.radians(position_degrees)
-        velocity_rad_per_sec = np.radians(velocity_deg_per_sec)
-
-        # Get motor limits
-        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
-
-        # Encode parameters
-        kp_uint = self._float_to_uint(kp, *MIT_KP_RANGE, 12)
-        kd_uint = self._float_to_uint(kd, *MIT_KD_RANGE, 12)
-        q_uint = self._float_to_uint(position_rad, -pmax, pmax, 16)
-        dq_uint = self._float_to_uint(velocity_rad_per_sec, -vmax, vmax, 12)
-        tau_uint = self._float_to_uint(torque, -tmax, tmax, 12)
-
-        # Pack data
-        data = [0] * 8
-        data[0] = (q_uint >> 8) & 0xFF
-        data[1] = q_uint & 0xFF
-        data[2] = dq_uint >> 4
-        data[3] = ((dq_uint & 0xF) << 4) | ((kp_uint >> 8) & 0xF)
-        data[4] = kp_uint & 0xFF
-        data[5] = kd_uint >> 4
-        data[6] = ((kd_uint & 0xF) << 4) | ((tau_uint >> 8) & 0xF)
-        data[7] = tau_uint & 0xFF
-        return data
-
-    def _mit_control(
-        self,
-        motor: NameOrID,
-        kp: float,
-        kd: float,
-        position_degrees: float,
-        velocity_deg_per_sec: float,
-        torque: float,
-    ) -> None:
-        """Send MIT control command to a motor."""
-        motor_id = self._get_motor_id(motor)
-        motor_name = self._get_motor_name(motor)
-        motor_type = self._motor_types[motor_name]
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
-        data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
-        msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
-        self.canbus.send(msg)
-
-        recv_id = self._get_motor_recv_id(motor)
-        if msg := self._recv_motor_response(expected_recv_id=recv_id):
-            self._process_response(motor_name, msg)
-        else:
-            logger.debug(f"No response from {motor_name} after MIT control command")
-
-    def _mit_control_batch(
-        self,
-        commands: dict[NameOrID, tuple[float, float, float, float, float]],
-    ) -> None:
-        """
-        Send MIT control commands to multiple motors in batch.
-        Sends all commands first, then collects responses.
-
-        Args:
-            commands: Dict mapping motor name/ID to (kp, kd, position_deg, velocity_deg/s, torque)
-                     Example: {'joint_1': (10.0, 0.5, 45.0, 0.0, 0.0), ...}
-        """
-        if not commands:
-            return
-
-        recv_id_to_motor: dict[int, str] = {}
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
-        # Step 1: Send all MIT control commands
-        for motor, (kp, kd, position_degrees, velocity_deg_per_sec, torque) in commands.items():
-            motor_id = self._get_motor_id(motor)
-            motor_name = self._get_motor_name(motor)
-            motor_type = self._motor_types[motor_name]
-
-            data = self._encode_mit_packet(motor_type, kp, kd, position_degrees, velocity_deg_per_sec, torque)
-            msg = can.Message(arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd)
-            self.canbus.send(msg)
-
-            recv_id_to_motor[self._get_motor_recv_id(motor)] = motor_name
-
-        # Step 2: Collect responses and update state cache
-        responses = self._recv_all_responses(list(recv_id_to_motor.keys()), timeout=SHORT_TIMEOUT_SEC)
-        for recv_id, motor_name in recv_id_to_motor.items():
-            if msg := responses.get(recv_id):
-                self._process_response(motor_name, msg)
-
-    def _float_to_uint(self, x: float, x_min: float, x_max: float, bits: int) -> int:
-        """Convert float to unsigned integer for CAN transmission."""
-        x = max(x_min, min(x_max, x))  # Clamp to range
-        span = x_max - x_min
-        data_norm = (x - x_min) / span
-        return int(data_norm * ((1 << bits) - 1))
-
-    def _uint_to_float(self, x: int, x_min: float, x_max: float, bits: int) -> float:
-        """Convert unsigned integer from CAN to float."""
-        span = x_max - x_min
-        data_norm = float(x) / ((1 << bits) - 1)
-        return data_norm * span + x_min
-
-    def _decode_motor_state(
-        self, data: bytearray | bytes, motor_type: MotorType
-    ) -> tuple[float, float, float, int, int]:
-        """
-        Decode motor state from CAN data.
-        Returns: (position_deg, velocity_deg_s, torque, temp_mos, temp_rotor)
-        """
-        if len(data) < 8:
-            raise ValueError("Invalid motor state data")
-
-        # Extract encoded values
-        q_uint = (data[1] << 8) | data[2]
-        dq_uint = (data[3] << 4) | (data[4] >> 4)
-        tau_uint = ((data[4] & 0x0F) << 8) | data[5]
-        t_mos = data[6]
-        t_rotor = data[7]
-
-        # Get motor limits
-        pmax, vmax, tmax = MOTOR_LIMIT_PARAMS[motor_type]
-
-        # Decode to physical values
-        position_rad = self._uint_to_float(q_uint, -pmax, pmax, 16)
-        velocity_rad_per_sec = self._uint_to_float(dq_uint, -vmax, vmax, 12)
-        torque = self._uint_to_float(tau_uint, -tmax, tmax, 12)
-
-        return np.degrees(position_rad), np.degrees(velocity_rad_per_sec), torque, t_mos, t_rotor
-
-    def _process_response(self, motor: str, msg: can.Message) -> None:
-        """Decode a message and update the motor state cache."""
-        try:
-            motor_type = self._motor_types[motor]
-            pos, vel, torque, t_mos, t_rotor = self._decode_motor_state(msg.data, motor_type)
-
-            self._last_known_states[motor] = {
-                "position": pos,
-                "velocity": vel,
-                "torque": torque,
-                "temp_mos": float(t_mos),
-                "temp_rotor": float(t_rotor),
-            }
-        except Exception as e:
-            logger.warning(f"Failed to decode response from {motor}: {e}")
-
-    @check_if_not_connected
-    def read(self, data_name: str, motor: str) -> Value:
-        """Read a value from a single motor. Positions are always in degrees."""
-
-        # Refresh motor to get latest state
-        msg = self._refresh_motor(motor)
-        if msg is None:
-            motor_id = self._get_motor_id(motor)
-            recv_id = self._get_motor_recv_id(motor)
-            raise ConnectionError(
-                f"No response from motor '{motor}' (send ID: 0x{motor_id:02X}, recv ID: 0x{recv_id:02X}). "
-                f"Check that: 1) Motor is powered (24V), 2) CAN wiring is correct, "
-                f"3) Motor IDs are configured correctly using Damiao Debugging Tools"
-            )
-
-        self._process_response(motor, msg)
-        return self._get_cached_value(motor, data_name)
-
-    def _get_cached_value(self, motor: str, data_name: str) -> Value:
-        """Retrieve a specific value from the cache."""
-        state = self._last_known_states[motor]
-        mapping: dict[str, Any] = {
-            "Present_Position": state["position"],
-            "Present_Velocity": state["velocity"],
-            "Present_Torque": state["torque"],
-            "Temperature_MOS": state["temp_mos"],
-            "Temperature_Rotor": state["temp_rotor"],
-        }
-        if data_name not in mapping:
-            raise ValueError(f"Unknown data_name: {data_name}")
-        return mapping[data_name]
-
-    @check_if_not_connected
-    def write(
-        self,
-        data_name: str,
-        motor: str,
-        value: Value,
-    ) -> None:
-        """
-        Write a value to a single motor. Positions are always in degrees.
-        Can write 'Goal_Position', 'Kp', or 'Kd'.
-        """
-
-        if data_name in ("Kp", "Kd"):
-            self._gains[motor][data_name.lower()] = float(value)
-        elif data_name == "Goal_Position":
-            kp = self._gains[motor]["kp"]
-            kd = self._gains[motor]["kd"]
-            self._mit_control(motor, kp, kd, float(value), 0.0, 0.0)
-        else:
-            raise ValueError(f"Writing {data_name} not supported in MIT mode")
-
-    def sync_read(
-        self,
-        data_name: str,
-        motors: str | list[str] | None = None,
-    ) -> dict[str, Value]:
-        """
-        Read the same value from multiple motors simultaneously.
-        """
-        target_motors = self._get_motors_list(motors)
-        self._batch_refresh(target_motors)
-
-        result = {}
-        for motor in target_motors:
-            result[motor] = self._get_cached_value(motor, data_name)
-        return result
-
-    def sync_read_all_states(
-        self,
-        motors: str | list[str] | None = None,
-        *,
-        num_retry: int = 0,
-    ) -> dict[str, MotorState]:
-        """
-        Read ALL motor states (position, velocity, torque) from multiple motors in ONE refresh cycle.
-
-        Returns:
-            Dictionary mapping motor names to state dicts with keys: 'position', 'velocity', 'torque'
-            Example: {'joint_1': {'position': 45.2, 'velocity': 1.3, 'torque': 0.5}, ...}
-        """
-        target_motors = self._get_motors_list(motors)
-        self._batch_refresh(target_motors)
-
-        result = {}
-        for motor in target_motors:
-            result[motor] = self._last_known_states[motor].copy()
-        return result
-
-    def _batch_refresh(self, motors: list[str]) -> None:
-        """Internal helper to refresh a list of motors and update cache."""
-
-        if self.canbus is None:
-            raise RuntimeError("CAN bus is not initialized.")
-
-        # Send refresh commands
-        for motor in motors:
-            motor_id = self._get_motor_id(motor)
-            data = [motor_id & 0xFF, (motor_id >> 8) & 0xFF, CAN_CMD_REFRESH, 0, 0, 0, 0, 0]
-            msg = can.Message(
-                arbitration_id=CAN_PARAM_ID, data=data, is_extended_id=False, is_fd=self.use_can_fd
-            )
-            self.canbus.send(msg)
-
-        # Collect responses
-        expected_recv_ids = [self._get_motor_recv_id(m) for m in motors]
-        responses = self._recv_all_responses(expected_recv_ids, timeout=MEDIUM_TIMEOUT_SEC)
-
-        # Update cache
-        for motor in motors:
-            recv_id = self._get_motor_recv_id(motor)
-            msg = responses.get(recv_id)
-            if msg:
-                self._process_response(motor, msg)
-            else:
-                logger.warning(f"Packet drop: {motor} (ID: 0x{recv_id:02X}). Using last known state.")
-
-    @check_if_not_connected
-    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
-        """
-        Write values to multiple motors simultaneously. Positions are always in degrees.
-        """
-
-        if data_name in ("Kp", "Kd"):
-            key = data_name.lower()
-            for motor, val in values.items():
-                self._gains[motor][key] = float(val)
-
-        elif data_name == "Goal_Position":
-            # Step 1: Send all MIT control commands
-            recv_id_to_motor: dict[int, str] = {}
-            if self.canbus is None:
-                raise RuntimeError("CAN bus is not initialized.")
-            for motor, value_degrees in values.items():
-                motor_id = self._get_motor_id(motor)
-                motor_name = self._get_motor_name(motor)
-                motor_type = self._motor_types[motor_name]
-
-                kp = self._gains[motor]["kp"]
-                kd = self._gains[motor]["kd"]
-
-                data = self._encode_mit_packet(motor_type, kp, kd, float(value_degrees), 0.0, 0.0)
-                msg = can.Message(
-                    arbitration_id=motor_id, data=data, is_extended_id=False, is_fd=self.use_can_fd
-                )
-                self.canbus.send(msg)
-                precise_sleep(PRECISE_TIMEOUT_SEC)
-
-                recv_id_to_motor[self._get_motor_recv_id(motor)] = motor_name
-
-            # Step 2: Collect responses and update state cache
-            responses = self._recv_all_responses(list(recv_id_to_motor.keys()), timeout=MEDIUM_TIMEOUT_SEC)
-            for recv_id, motor_name in recv_id_to_motor.items():
-                if msg := responses.get(recv_id):
-                    self._process_response(motor_name, msg)
-        else:
-            # Fall back to individual writes
-            for motor, value in values.items():
-                self.write(data_name, motor, value)
-
-    def read_calibration(self) -> dict[str, MotorCalibration]:
-        """Read calibration data from motors."""
-        # Damiao motors don't store calibration internally
-        # Return existing calibration or empty dict
-        return self.calibration if self.calibration else {}
-
-    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
-        """Write calibration data to motors."""
-        # Damiao motors don't store calibration internally
-        # Just cache it in memory
-        if cache:
-            self.calibration = calibration_dict
-
-    def record_ranges_of_motion(
-        self,
-        motors: str | list[str] | None = None,
-        display_values: bool = True,
-    ) -> tuple[dict[str, Value], dict[str, Value]]:
-        """
-        Interactively record the min/max values of each motor in degrees.
-
-        Move the joints by hand (with torque disabled) while the method streams live positions.
-        Press Enter to finish.
-        """
-        target_motors = self._get_motors_list(motors)
-
-        self.disable_torque(target_motors)
-        time.sleep(LONG_TIMEOUT_SEC)
-
-        start_positions = self.sync_read("Present_Position", target_motors)
-        mins = start_positions.copy()
-        maxes = start_positions.copy()
-
-        print("\nMove joints through their full range of motion. Press ENTER when done.")
-        user_pressed_enter = False
-
-        while not user_pressed_enter:
-            positions = self.sync_read("Present_Position", target_motors)
-
-            for motor in target_motors:
-                if motor in positions:
-                    mins[motor] = min(positions[motor], mins.get(motor, positions[motor]))
-                    maxes[motor] = max(positions[motor], maxes.get(motor, positions[motor]))
-
-            if display_values:
-                print("\n" + "=" * 50)
-                print(f"{'MOTOR':<20} | {'MIN (deg)':>12} | {'POS (deg)':>12} | {'MAX (deg)':>12}")
-                print("-" * 50)
-                for motor in target_motors:
-                    if motor in positions:
-                        print(
-                            f"{motor:<20} | {mins[motor]:>12.1f} | {positions[motor]:>12.1f} | {maxes[motor]:>12.1f}"
-                        )
-
-            if enter_pressed():
-                user_pressed_enter = True
-
-            if display_values and not user_pressed_enter:
-                move_cursor_up(len(target_motors) + 4)
-
-            time.sleep(LONG_TIMEOUT_SEC)
-
-        self.enable_torque(target_motors)
-
-        for motor in target_motors:
-            if (motor in mins) and (motor in maxes) and (int(abs(maxes[motor] - mins[motor])) < 5):
-                raise ValueError(f"Motor {motor} has insufficient range of motion (< 5 degrees)")
-
-        return mins, maxes
-
-    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
-        """Convert motor specification to list of motor names."""
-        if motors is None:
-            return list(self.motors.keys())
-        elif isinstance(motors, str):
-            return [motors]
-        elif isinstance(motors, list):
-            return motors
-        else:
-            raise TypeError(f"Invalid motors type: {type(motors)}")
-
-    def _get_motor_id(self, motor: NameOrID) -> int:
-        """Get CAN ID for a motor."""
-        if isinstance(motor, str):
-            if motor in self.motors:
-                return self.motors[motor].id
-            else:
-                raise ValueError(f"Unknown motor: {motor}")
-        else:
-            return motor
-
-    def _get_motor_name(self, motor: NameOrID) -> str:
-        """Get motor name from name or ID."""
-        if isinstance(motor, str):
-            return motor
-        else:
-            for name, m in self.motors.items():
-                if m.id == motor:
-                    return name
-            raise ValueError(f"Unknown motor ID: {motor}")
-
-    def _get_motor_recv_id(self, motor: NameOrID) -> int:
-        """Get motor recv_id from name or ID."""
-        motor_name = self._get_motor_name(motor)
-        motor_obj = self.motors.get(motor_name)
-        if motor_obj and motor_obj.recv_id is not None:
-            return motor_obj.recv_id
-        else:
-            raise ValueError(f"Motor {motor_obj} doesn't have a valid recv_id (None).")
-
-    @cached_property
-    def is_calibrated(self) -> bool:
-        """Check if motors are calibrated."""
-        return bool(self.calibration)

src/lerobot/motors/damiao/tables.py

@@ -1,209 +0,0 @@
-# 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.
-
-"""Configuration tables for Damiao motors."""
-
-from enum import IntEnum
-
-
-# Motor type definitions
-class MotorType(IntEnum):
-    DM3507 = 0
-    DM4310 = 1
-    DM4310_48V = 2
-    DM4340 = 3
-    DM4340_48V = 4
-    DM6006 = 5
-    DM8006 = 6
-    DM8009 = 7
-    DM10010L = 8
-    DM10010 = 9
-    DMH3510 = 10
-    DMH6215 = 11
-    DMG6220 = 12
-
-
-# Control modes
-class ControlMode(IntEnum):
-    MIT = 1
-    POS_VEL = 2
-    VEL = 3
-    TORQUE_POS = 4
-
-
-# Motor variable IDs (RID)
-class MotorVariable(IntEnum):
-    UV_VALUE = 0
-    KT_VALUE = 1
-    OT_VALUE = 2
-    OC_VALUE = 3
-    ACC = 4
-    DEC = 5
-    MAX_SPD = 6
-    MST_ID = 7
-    ESC_ID = 8
-    TIMEOUT = 9
-    CTRL_MODE = 10
-    DAMP = 11
-    INERTIA = 12
-    HW_VER = 13
-    SW_VER = 14
-    SN = 15
-    NPP = 16
-    RS = 17
-    LS = 18
-    FLUX = 19
-    GR = 20
-    PMAX = 21
-    VMAX = 22
-    TMAX = 23
-    I_BW = 24
-    KP_ASR = 25
-    KI_ASR = 26
-    KP_APR = 27
-    KI_APR = 28
-    OV_VALUE = 29
-    GREF = 30
-    DETA = 31
-    V_BW = 32
-    IQ_C1 = 33
-    VL_C1 = 34
-    CAN_BR = 35
-    SUB_VER = 36
-    U_OFF = 50
-    V_OFF = 51
-    K1 = 52
-    K2 = 53
-    M_OFF = 54
-    DIR = 55
-    P_M = 80
-    XOUT = 81
-
-
-# Motor limit parameters [PMAX, VMAX, TMAX]
-# PMAX: Maximum position (rad)
-# VMAX: Maximum velocity (rad/s)
-# TMAX: Maximum torque (N·m)
-MOTOR_LIMIT_PARAMS = {
-    MotorType.DM3507: (12.5, 30, 10),
-    MotorType.DM4310: (12.5, 30, 10),
-    MotorType.DM4310_48V: (12.5, 50, 10),
-    MotorType.DM4340: (12.5, 8, 28),
-    MotorType.DM4340_48V: (12.5, 10, 28),
-    MotorType.DM6006: (12.5, 45, 20),
-    MotorType.DM8006: (12.5, 45, 40),
-    MotorType.DM8009: (12.5, 45, 54),
-    MotorType.DM10010L: (12.5, 25, 200),
-    MotorType.DM10010: (12.5, 20, 200),
-    MotorType.DMH3510: (12.5, 280, 1),
-    MotorType.DMH6215: (12.5, 45, 10),
-    MotorType.DMG6220: (12.5, 45, 10),
-}
-
-# Motor model names
-MODEL_NAMES = {
-    MotorType.DM3507: "dm3507",
-    MotorType.DM4310: "dm4310",
-    MotorType.DM4310_48V: "dm4310_48v",
-    MotorType.DM4340: "dm4340",
-    MotorType.DM4340_48V: "dm4340_48v",
-    MotorType.DM6006: "dm6006",
-    MotorType.DM8006: "dm8006",
-    MotorType.DM8009: "dm8009",
-    MotorType.DM10010L: "dm10010l",
-    MotorType.DM10010: "dm10010",
-    MotorType.DMH3510: "dmh3510",
-    MotorType.DMH6215: "dmh6215",
-    MotorType.DMG6220: "dmg6220",
-}
-
-# Motor resolution table (encoder counts per revolution)
-MODEL_RESOLUTION = {
-    "dm3507": 65536,
-    "dm4310": 65536,
-    "dm4310_48v": 65536,
-    "dm4340": 65536,
-    "dm4340_48v": 65536,
-    "dm6006": 65536,
-    "dm8006": 65536,
-    "dm8009": 65536,
-    "dm10010l": 65536,
-    "dm10010": 65536,
-    "dmh3510": 65536,
-    "dmh6215": 65536,
-    "dmg6220": 65536,
-}
-
-# CAN baudrates supported by Damiao motors
-AVAILABLE_BAUDRATES = [
-    125000,  # 0: 125 kbps
-    200000,  # 1: 200 kbps
-    250000,  # 2: 250 kbps
-    500000,  # 3: 500 kbps
-    1000000,  # 4: 1 mbps (default for OpenArms)
-    2000000,  # 5: 2 mbps
-    2500000,  # 6: 2.5 mbps
-    3200000,  # 7: 3.2 mbps
-    4000000,  # 8: 4 mbps
-    5000000,  # 9: 5 mbps
-]
-DEFAULT_BAUDRATE = 1000000  # 1 Mbps is standard for OpenArms
-
-# Default timeout in milliseconds
-DEFAULT_TIMEOUT_MS = 1000
-
-# OpenArms specific configurations
-# Based on: https://docs.openarm.dev/software/setup/configure-test
-# OpenArms has 7 DOF per arm (14 total for dual arm)
-OPENARMS_ARM_MOTOR_IDS = {
-    "joint_1": {"send": 0x01, "recv": 0x11},  # J1 - Shoulder pan
-    "joint_2": {"send": 0x02, "recv": 0x12},  # J2 - Shoulder lift
-    "joint_3": {"send": 0x03, "recv": 0x13},  # J3 - Elbow flex
-    "joint_4": {"send": 0x04, "recv": 0x14},  # J4 - Wrist flex
-    "joint_5": {"send": 0x05, "recv": 0x15},  # J5 - Wrist roll
-    "joint_6": {"send": 0x06, "recv": 0x16},  # J6 - Wrist pitch
-    "joint_7": {"send": 0x07, "recv": 0x17},  # J7 - Wrist rotation
-}
-
-OPENARMS_GRIPPER_MOTOR_IDS = {
-    "gripper": {"send": 0x08, "recv": 0x18},  # J8 - Gripper
-}
-
-# Default motor types for OpenArms
-OPENARMS_DEFAULT_MOTOR_TYPES = {
-    "joint_1": MotorType.DM8009,  # Shoulder pan - high torque
-    "joint_2": MotorType.DM8009,  # Shoulder lift - high torque
-    "joint_3": MotorType.DM4340,  # Shoulder rotation
-    "joint_4": MotorType.DM4340,  # Elbow flex
-    "joint_5": MotorType.DM4310,  # Wrist roll
-    "joint_6": MotorType.DM4310,  # Wrist pitch
-    "joint_7": MotorType.DM4310,  # Wrist rotation
-    "gripper": MotorType.DM4310,  # Gripper
-}
-
-# MIT control parameter ranges
-MIT_KP_RANGE = (0.0, 500.0)
-MIT_KD_RANGE = (0.0, 5.0)
-
-# CAN frame command IDs
-CAN_CMD_ENABLE = 0xFC
-CAN_CMD_DISABLE = 0xFD
-CAN_CMD_SET_ZERO = 0xFE
-CAN_CMD_REFRESH = 0xCC
-CAN_CMD_QUERY_PARAM = 0x33
-CAN_CMD_WRITE_PARAM = 0x55
-CAN_CMD_SAVE_PARAM = 0xAA
-
-# CAN ID for parameter operations
-CAN_PARAM_ID = 0x7FF

src/lerobot/motors/dynamixel/dynamixel.py

@@ -22,8 +22,9 @@ import logging
 from copy import deepcopy
 from enum import Enum
 
-from ..encoding_utils import decode_twos_complement, encode_twos_complement
-from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
+from lerobot.motors.encoding_utils import decode_twos_complement, encode_twos_complement
+
+from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
 from .tables import (
     AVAILABLE_BAUDRATES,
     MODEL_BAUDRATE_TABLE,
@@ -99,7 +100,7 @@ def _split_into_byte_chunks(value: int, length: int) -> list[int]:
     return data
 
 
-class DynamixelMotorsBus(SerialMotorsBus):
+class DynamixelMotorsBus(MotorsBus):
     """
     The Dynamixel implementation for a MotorsBus. It relies on the python dynamixel sdk to communicate with
     the motors. For more info, see the Dynamixel SDK Documentation:
@@ -181,10 +182,10 @@ class DynamixelMotorsBus(SerialMotorsBus):
         for motor, m in self.motors.items():
             calibration[motor] = MotorCalibration(
                 id=m.id,
-                drive_mode=int(drive_modes[motor]),
-                homing_offset=int(offsets[motor]),
-                range_min=int(mins[motor]),
-                range_max=int(maxes[motor]),
+                drive_mode=drive_modes[motor],
+                homing_offset=offsets[motor],
+                range_min=mins[motor],
+                range_max=maxes[motor],
             )
 
         return calibration
@@ -198,15 +199,15 @@ class DynamixelMotorsBus(SerialMotorsBus):
         if cache:
             self.calibration = calibration_dict
 
-    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
 
-    def _disable_torque(self, motor: int, model: str, num_retry: int = 0) -> None:
+    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
+        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
 
-    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
 
@@ -235,7 +236,7 @@ class DynamixelMotorsBus(SerialMotorsBus):
         On Dynamixel Motors:
         Present_Position = Actual_Position + Homing_Offset
         """
-        half_turn_homings: dict[NameOrID, Value] = {}
+        half_turn_homings = {}
         for motor, pos in positions.items():
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
@@ -258,6 +259,6 @@ class DynamixelMotorsBus(SerialMotorsBus):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
 
-            return None
+            return
 
         return {id_: data[0] for id_, data in data_list.items()}

src/lerobot/motors/feetech/feetech.py

@@ -17,8 +17,9 @@ from copy import deepcopy
 from enum import Enum
 from pprint import pformat
 
-from ..encoding_utils import decode_sign_magnitude, encode_sign_magnitude
-from ..motors_bus import Motor, MotorCalibration, NameOrID, SerialMotorsBus, Value, get_address
+from lerobot.motors.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
+
+from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
 from .tables import (
     FIRMWARE_MAJOR_VERSION,
     FIRMWARE_MINOR_VERSION,
@@ -95,7 +96,7 @@ def patch_setPacketTimeout(self, packet_length):  # noqa: N802
     self.packet_timeout = (self.tx_time_per_byte * packet_length) + (self.tx_time_per_byte * 3.0) + 50
 
 
-class FeetechMotorsBus(SerialMotorsBus):
+class FeetechMotorsBus(MotorsBus):
     """
     The FeetechMotorsBus class allows to efficiently read and write to the attached motors. It relies on the
     python feetech sdk to communicate with the motors, which is itself based on the dynamixel sdk.
@@ -126,7 +127,7 @@ class FeetechMotorsBus(SerialMotorsBus):
 
         self.port_handler = scs.PortHandler(self.port)
         # HACK: monkeypatch
-        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(  # type: ignore[method-assign]
+        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(
             self.port_handler, scs.PortHandler
         )
         self.packet_handler = scs.PacketHandler(protocol_version)
@@ -262,9 +263,9 @@ class FeetechMotorsBus(SerialMotorsBus):
             calibration[motor] = MotorCalibration(
                 id=m.id,
                 drive_mode=0,
-                homing_offset=int(offsets[motor]),
-                range_min=int(mins[motor]),
-                range_max=int(maxes[motor]),
+                homing_offset=offsets[motor],
+                range_min=mins[motor],
+                range_max=maxes[motor],
             )
 
         return calibration
@@ -284,7 +285,7 @@ class FeetechMotorsBus(SerialMotorsBus):
         On Feetech Motors:
         Present_Position = Actual_Position - Homing_Offset
         """
-        half_turn_homings: dict[NameOrID, Value] = {}
+        half_turn_homings = {}
         for motor, pos in positions.items():
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
@@ -292,18 +293,18 @@ class FeetechMotorsBus(SerialMotorsBus):
 
         return half_turn_homings
 
-    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 0, num_retry=num_retry)
 
-    def _disable_torque(self, motor: int, model: str, num_retry: int = 0) -> None:
+    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
         addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor, TorqueMode.DISABLED.value, num_retry=num_retry)
+        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
         addr, length = get_address(self.model_ctrl_table, model, "Lock")
-        self._write(addr, length, motor, 0, num_retry=num_retry)
+        self._write(addr, length, motor_id, 0, num_retry=num_retry)
 
-    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         for motor in self._get_motors_list(motors):
             self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
             self.write("Lock", motor, 1, num_retry=num_retry)
@@ -334,7 +335,7 @@ class FeetechMotorsBus(SerialMotorsBus):
     def _broadcast_ping(self) -> tuple[dict[int, int], int]:
         import scservo_sdk as scs
 
-        data_list: dict[int, int] = {}
+        data_list = {}
 
         status_length = 6
 
@@ -414,7 +415,7 @@ class FeetechMotorsBus(SerialMotorsBus):
         if not self._is_comm_success(comm):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
-            return None
+            return
 
         ids_errors = {id_: status for id_, status in ids_status.items() if self._is_error(status)}
         if ids_errors:

src/lerobot/motors/motors_bus.py

@@ -19,11 +19,8 @@
 # TODO(aliberts): Add block noqa when feature below is available
 # https://github.com/astral-sh/ruff/issues/3711
 
-from __future__ import annotations
-
 import abc
 import logging
-from collections.abc import Sequence
 from contextlib import contextmanager
 from dataclasses import dataclass
 from enum import Enum
@@ -44,81 +41,6 @@ Value: TypeAlias = int | float
 logger = logging.getLogger(__name__)
 
 
-class MotorsBusBase(abc.ABC):
-    """
-    Base class for all motor bus implementations.
-
-    This is a minimal interface that all motor buses must implement, regardless of their
-    communication protocol (serial, CAN, etc.).
-    """
-
-    def __init__(
-        self,
-        port: str,
-        motors: dict[str, Motor],
-        calibration: dict[str, MotorCalibration] | None = None,
-    ):
-        self.port = port
-        self.motors = motors
-        self.calibration = calibration if calibration else {}
-
-    @abc.abstractmethod
-    def connect(self, handshake: bool = True) -> None:
-        """Establish connection to the motors."""
-        pass
-
-    @abc.abstractmethod
-    def disconnect(self, disable_torque: bool = True) -> None:
-        """Disconnect from the motors."""
-        pass
-
-    @property
-    @abc.abstractmethod
-    def is_connected(self) -> bool:
-        """Check if connected to the motors."""
-        pass
-
-    @abc.abstractmethod
-    def read(self, data_name: str, motor: str) -> Value:
-        """Read a value from a single motor."""
-        pass
-
-    @abc.abstractmethod
-    def write(self, data_name: str, motor: str, value: Value) -> None:
-        """Write a value to a single motor."""
-        pass
-
-    @abc.abstractmethod
-    def sync_read(self, data_name: str, motors: str | list[str] | None = None) -> dict[str, Value]:
-        """Read a value from multiple motors."""
-        pass
-
-    @abc.abstractmethod
-    def sync_write(self, data_name: str, values: dict[str, Value]) -> None:
-        """Write values to multiple motors."""
-        pass
-
-    @abc.abstractmethod
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
-        """Enable torque on selected motors."""
-        pass
-
-    @abc.abstractmethod
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
-        """Disable torque on selected motors."""
-        pass
-
-    @abc.abstractmethod
-    def read_calibration(self) -> dict[str, MotorCalibration]:
-        """Read calibration parameters from the motors."""
-        pass
-
-    @abc.abstractmethod
-    def write_calibration(self, calibration_dict: dict[str, MotorCalibration], cache: bool = True) -> None:
-        """Write calibration parameters to the motors."""
-        pass
-
-
 def get_ctrl_table(model_ctrl_table: dict[str, dict], model: str) -> dict[str, tuple[int, int]]:
     ctrl_table = model_ctrl_table.get(model)
     if ctrl_table is None:
@@ -175,21 +97,18 @@ class Motor:
     id: int
     model: str
     norm_mode: MotorNormMode
-    motor_type_str: str | None = None
-    recv_id: int | None = None
 
 
 class PortHandler(Protocol):
-    is_open: bool
-    baudrate: int
-    packet_start_time: float
-    packet_timeout: float
-    tx_time_per_byte: float
-    is_using: bool
-    port_name: str
-    ser: serial.Serial
-
-    def __init__(self, port_name: str) -> None: ...
+    def __init__(self, port_name):
+        self.is_open: bool
+        self.baudrate: int
+        self.packet_start_time: float
+        self.packet_timeout: float
+        self.tx_time_per_byte: float
+        self.is_using: bool
+        self.port_name: str
+        self.ser: serial.Serial
 
     def openPort(self): ...
     def closePort(self): ...
@@ -242,22 +161,19 @@ class PacketHandler(Protocol):
     def regWriteTxRx(self, port, id, address, length, data): ...
     def syncReadTx(self, port, start_address, data_length, param, param_length): ...
     def syncWriteTxOnly(self, port, start_address, data_length, param, param_length): ...
-    def broadcastPing(self, port): ...
 
 
 class GroupSyncRead(Protocol):
-    port: str
-    ph: PortHandler
-    start_address: int
-    data_length: int
-    last_result: bool
-    is_param_changed: bool
-    param: list
-    data_dict: dict
+    def __init__(self, port, ph, start_address, data_length):
+        self.port: str
+        self.ph: PortHandler
+        self.start_address: int
+        self.data_length: int
+        self.last_result: bool
+        self.is_param_changed: bool
+        self.param: list
+        self.data_dict: dict
 
-    def __init__(
-        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
-    ) -> None: ...
     def makeParam(self): ...
     def addParam(self, id): ...
     def removeParam(self, id): ...
@@ -270,17 +186,15 @@ class GroupSyncRead(Protocol):
 
 
 class GroupSyncWrite(Protocol):
-    port: str
-    ph: PortHandler
-    start_address: int
-    data_length: int
-    is_param_changed: bool
-    param: list
-    data_dict: dict
+    def __init__(self, port, ph, start_address, data_length):
+        self.port: str
+        self.ph: PortHandler
+        self.start_address: int
+        self.data_length: int
+        self.is_param_changed: bool
+        self.param: list
+        self.data_dict: dict
 
-    def __init__(
-        self, port: PortHandler, ph: PacketHandler, start_address: int, data_length: int
-    ) -> None: ...
     def makeParam(self): ...
     def addParam(self, id, data): ...
     def removeParam(self, id): ...
@@ -289,15 +203,15 @@ class GroupSyncWrite(Protocol):
     def txPacket(self): ...
 
 
-class SerialMotorsBus(MotorsBusBase):
+class MotorsBus(abc.ABC):
     """
-    A SerialMotorsBus allows to efficiently read and write to motors connected via serial communication.
+    A MotorsBus allows to efficiently read and write to the attached motors.
     It represents several motors daisy-chained together and connected through a serial port.
-    There are currently two implementations of this class:
+    There are currently two implementations of this abstract class:
         - DynamixelMotorsBus
         - FeetechMotorsBus
 
-    This class is specifically for serial-based motor protocols (Dynamixel, Feetech, etc.).
+    Note: This class may evolve in the future should we add support for other types of bus.
 
     A MotorsBus subclass instance requires a port (e.g. `FeetechMotorsBus(port="/dev/tty.usbmodem575E0031751"`)).
     To find the port, you can run our utility script:
@@ -346,7 +260,9 @@ class SerialMotorsBus(MotorsBusBase):
         motors: dict[str, Motor],
         calibration: dict[str, MotorCalibration] | None = None,
     ):
-        super().__init__(port, motors, calibration)
+        self.port = port
+        self.motors = motors
+        self.calibration = calibration if calibration else {}
 
         self.port_handler: PortHandler
         self.packet_handler: PacketHandler
@@ -407,7 +323,7 @@ class SerialMotorsBus(MotorsBusBase):
         else:
             raise TypeError(f"'{motor}' should be int, str.")
 
-    def _get_motor_model(self, motor: NameOrID) -> str:
+    def _get_motor_model(self, motor: NameOrID) -> int:
         if isinstance(motor, str):
             return self.motors[motor].model
         elif isinstance(motor, int):
@@ -415,19 +331,17 @@ class SerialMotorsBus(MotorsBusBase):
         else:
             raise TypeError(f"'{motor}' should be int, str.")
 
-    def _get_motors_list(self, motors: NameOrID | Sequence[NameOrID] | None) -> list[str]:
+    def _get_motors_list(self, motors: str | list[str] | None) -> list[str]:
         if motors is None:
             return list(self.motors)
         elif isinstance(motors, str):
             return [motors]
-        elif isinstance(motors, int):
-            return [self._id_to_name(motors)]
-        elif isinstance(motors, Sequence):
-            return [m if isinstance(m, str) else self._id_to_name(m) for m in motors]
+        elif isinstance(motors, list):
+            return motors.copy()
         else:
             raise TypeError(motors)
 
-    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> dict[int, Value]:
+    def _get_ids_values_dict(self, values: Value | dict[str, Value] | None) -> list[str]:
         if isinstance(values, (int | float)):
             return dict.fromkeys(self.ids, values)
         elif isinstance(values, dict):
@@ -618,7 +532,7 @@ class SerialMotorsBus(MotorsBusBase):
         self.set_baudrate(self.default_baudrate)
 
     @abc.abstractmethod
-    def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
+    def _find_single_motor(self, motor: str, initial_baudrate: int | None) -> tuple[int, int]:
         pass
 
     @abc.abstractmethod
@@ -631,13 +545,13 @@ class SerialMotorsBus(MotorsBusBase):
         pass
 
     @abc.abstractmethod
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
+    def disable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
         """Disable torque on selected motors.
 
         Disabling Torque allows to write to the motors' permanent memory area (EPROM/EEPROM).
 
         Args:
-            motors ( str | list[str] | None, optional): Target motors.  Accepts a motor name, an ID, a
+            motors (int | str | list[str] | None, optional): Target motors.  Accepts a motor name, an ID, a
                 list of names or `None` to affect every registered motor.  Defaults to `None`.
             num_retry (int, optional): Number of additional retry attempts on communication failure.
                 Defaults to 0.
@@ -649,19 +563,18 @@ class SerialMotorsBus(MotorsBusBase):
         pass
 
     @abc.abstractmethod
-    def enable_torque(self, motors: int | str | list[str] | None = None, num_retry: int = 0) -> None:
+    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
         """Enable torque on selected motors.
 
         Args:
-            motors (int | str | list[str] | None, optional): Same semantics as :pymeth:`disable_torque`.
-                Defaults to `None`.
+            motor (int): Same semantics as :pymeth:`disable_torque`. Defaults to `None`.
             num_retry (int, optional): Number of additional retry attempts on communication failure.
                 Defaults to 0.
         """
         pass
 
     @contextmanager
-    def torque_disabled(self, motors: str | list[str] | None = None):
+    def torque_disabled(self, motors: int | str | list[str] | None = None):
         """Context-manager that guarantees torque is re-enabled.
 
         This helper is useful to temporarily disable torque when configuring motors.
@@ -738,19 +651,24 @@ class SerialMotorsBus(MotorsBusBase):
         """
         pass
 
-    def reset_calibration(self, motors: NameOrID | Sequence[NameOrID] | None = None) -> None:
+    def reset_calibration(self, motors: NameOrID | list[NameOrID] | None = None) -> None:
         """Restore factory calibration for the selected motors.
 
         Homing offset is set to ``0`` and min/max position limits are set to the full usable range.
         The in-memory :pyattr:`calibration` is cleared.
 
         Args:
-            motors (NameOrID | Sequence[NameOrID] | None, optional): Selection of motors. `None` (default)
+            motors (NameOrID | list[NameOrID] | None, optional): Selection of motors. `None` (default)
                 resets every motor.
         """
-        motor_names = self._get_motors_list(motors)
+        if motors is None:
+            motors = list(self.motors)
+        elif isinstance(motors, (str | int)):
+            motors = [motors]
+        elif not isinstance(motors, list):
+            raise TypeError(motors)
 
-        for motor in motor_names:
+        for motor in motors:
             model = self._get_motor_model(motor)
             max_res = self.model_resolution_table[model] - 1
             self.write("Homing_Offset", motor, 0, normalize=False)
@@ -759,9 +677,7 @@ class SerialMotorsBus(MotorsBusBase):
 
         self.calibration = {}
 
-    def set_half_turn_homings(
-        self, motors: NameOrID | Sequence[NameOrID] | None = None
-    ) -> dict[NameOrID, Value]:
+    def set_half_turn_homings(self, motors: NameOrID | list[NameOrID] | None = None) -> dict[NameOrID, Value]:
         """Centre each motor range around its current position.
 
         The function computes and writes a homing offset such that the present position becomes exactly one
@@ -771,12 +687,17 @@ class SerialMotorsBus(MotorsBusBase):
             motors (NameOrID | list[NameOrID] | None, optional): Motors to adjust. Defaults to all motors (`None`).
 
         Returns:
-            dict[str, Value]: Mapping *motor name → written homing offset*.
+            dict[NameOrID, Value]: Mapping *motor → written homing offset*.
         """
-        motor_names = self._get_motors_list(motors)
+        if motors is None:
+            motors = list(self.motors)
+        elif isinstance(motors, (str | int)):
+            motors = [motors]
+        elif not isinstance(motors, list):
+            raise TypeError(motors)
 
-        self.reset_calibration(motor_names)
-        actual_positions = self.sync_read("Present_Position", motor_names, normalize=False)
+        self.reset_calibration(motors)
+        actual_positions = self.sync_read("Present_Position", motors, normalize=False)
         homing_offsets = self._get_half_turn_homings(actual_positions)
         for motor, offset in homing_offsets.items():
             self.write("Homing_Offset", motor, offset)
@@ -788,8 +709,8 @@ class SerialMotorsBus(MotorsBusBase):
         pass
 
     def record_ranges_of_motion(
-        self, motors: NameOrID | Sequence[NameOrID] | None = None, display_values: bool = True
-    ) -> tuple[dict[str, Value], dict[str, Value]]:
+        self, motors: NameOrID | list[NameOrID] | None = None, display_values: bool = True
+    ) -> tuple[dict[NameOrID, Value], dict[NameOrID, Value]]:
         """Interactively record the min/max encoder values of each motor.
 
         Move the joints by hand (with torque disabled) while the method streams live positions. Press
@@ -801,25 +722,30 @@ class SerialMotorsBus(MotorsBusBase):
             display_values (bool, optional): When `True` (default) a live table is printed to the console.
 
         Returns:
-            tuple[dict[str, Value], dict[str, Value]]: Two dictionaries *mins* and *maxes* with the
+            tuple[dict[NameOrID, Value], dict[NameOrID, Value]]: Two dictionaries *mins* and *maxes* with the
                 extreme values observed for each motor.
         """
-        motor_names = self._get_motors_list(motors)
+        if motors is None:
+            motors = list(self.motors)
+        elif isinstance(motors, (str | int)):
+            motors = [motors]
+        elif not isinstance(motors, list):
+            raise TypeError(motors)
 
-        start_positions = self.sync_read("Present_Position", motor_names, normalize=False)
+        start_positions = self.sync_read("Present_Position", motors, normalize=False)
         mins = start_positions.copy()
         maxes = start_positions.copy()
 
         user_pressed_enter = False
         while not user_pressed_enter:
-            positions = self.sync_read("Present_Position", motor_names, normalize=False)
+            positions = self.sync_read("Present_Position", motors, normalize=False)
             mins = {motor: min(positions[motor], min_) for motor, min_ in mins.items()}
             maxes = {motor: max(positions[motor], max_) for motor, max_ in maxes.items()}
 
             if display_values:
                 print("\n-------------------------------------------")
                 print(f"{'NAME':<15} | {'MIN':>6} | {'POS':>6} | {'MAX':>6}")
-                for motor in motor_names:
+                for motor in motors:
                     print(f"{motor:<15} | {mins[motor]:>6} | {positions[motor]:>6} | {maxes[motor]:>6}")
 
             if enter_pressed():
@@ -827,9 +753,9 @@ class SerialMotorsBus(MotorsBusBase):
 
             if display_values and not user_pressed_enter:
                 # Move cursor up to overwrite the previous output
-                move_cursor_up(len(motor_names) + 3)
+                move_cursor_up(len(motors) + 3)
 
-        same_min_max = [motor for motor in motor_names if mins[motor] == maxes[motor]]
+        same_min_max = [motor for motor in motors if mins[motor] == maxes[motor]]
         if same_min_max:
             raise ValueError(f"Some motors have the same min and max values:\n{pformat(same_min_max)}")
 
@@ -952,12 +878,12 @@ class SerialMotorsBus(MotorsBusBase):
             if raise_on_error:
                 raise ConnectionError(self.packet_handler.getTxRxResult(comm))
             else:
-                return None
+                return
         if self._is_error(error):
             if raise_on_error:
                 raise RuntimeError(self.packet_handler.getRxPacketError(error))
             else:
-                return None
+                return
 
         return model_number
 
@@ -1004,13 +930,12 @@ class SerialMotorsBus(MotorsBusBase):
         err_msg = f"Failed to read '{data_name}' on {id_=} after {num_retry + 1} tries."
         value, _, _ = self._read(addr, length, id_, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
-        decoded = self._decode_sign(data_name, {id_: value})
+        id_value = self._decode_sign(data_name, {id_: value})
 
         if normalize and data_name in self.normalized_data:
-            normalized = self._normalize(decoded)
-            return normalized[id_]
+            id_value = self._normalize(id_value)
 
-        return decoded[id_]
+        return id_value[id_]
 
     def _read(
         self,
@@ -1021,7 +946,7 @@ class SerialMotorsBus(MotorsBusBase):
         num_retry: int = 0,
         raise_on_error: bool = True,
         err_msg: str = "",
-    ) -> tuple[int, int, int]:
+    ) -> tuple[int, int]:
         if length == 1:
             read_fn = self.packet_handler.read1ByteTxRx
         elif length == 2:
@@ -1071,14 +996,13 @@ class SerialMotorsBus(MotorsBusBase):
         model = self.motors[motor].model
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
-        int_value = int(value)
         if normalize and data_name in self.normalized_data:
-            int_value = self._unnormalize({id_: value})[id_]
+            value = self._unnormalize({id_: value})[id_]
 
-        int_value = self._encode_sign(data_name, {id_: int_value})[id_]
+        value = self._encode_sign(data_name, {id_: value})[id_]
 
-        err_msg = f"Failed to write '{data_name}' on {id_=} with '{int_value}' after {num_retry + 1} tries."
-        self._write(addr, length, id_, int_value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
+        err_msg = f"Failed to write '{data_name}' on {id_=} with '{value}' after {num_retry + 1} tries."
+        self._write(addr, length, id_, value, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
     def _write(
         self,
@@ -1112,7 +1036,7 @@ class SerialMotorsBus(MotorsBusBase):
     def sync_read(
         self,
         data_name: str,
-        motors: NameOrID | Sequence[NameOrID] | None = None,
+        motors: str | list[str] | None = None,
         *,
         normalize: bool = True,
         num_retry: int = 0,
@@ -1121,7 +1045,7 @@ class SerialMotorsBus(MotorsBusBase):
 
         Args:
             data_name (str): Register name.
-            motors (NameOrID | Sequence[NameOrID] | None, optional): Motors to query. `None` (default) reads every motor.
+            motors (str | list[str] | None, optional): Motors to query. `None` (default) reads every motor.
             normalize (bool, optional): Normalisation flag.  Defaults to `True`.
             num_retry (int, optional): Retry attempts.  Defaults to `0`.
 
@@ -1142,17 +1066,16 @@ class SerialMotorsBus(MotorsBusBase):
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
         err_msg = f"Failed to sync read '{data_name}' on {ids=} after {num_retry + 1} tries."
-        raw_ids_values, _ = self._sync_read(
+        ids_values, _ = self._sync_read(
             addr, length, ids, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
         )
 
-        decoded = self._decode_sign(data_name, raw_ids_values)
+        ids_values = self._decode_sign(data_name, ids_values)
 
         if normalize and data_name in self.normalized_data:
-            normalized = self._normalize(decoded)
-            return {self._id_to_name(id_): value for id_, value in normalized.items()}
+            ids_values = self._normalize(ids_values)
 
-        return {self._id_to_name(id_): value for id_, value in decoded.items()}
+        return {self._id_to_name(id_): value for id_, value in ids_values.items()}
 
     def _sync_read(
         self,
@@ -1224,24 +1147,21 @@ class SerialMotorsBus(MotorsBusBase):
             num_retry (int, optional): Retry attempts.  Defaults to `0`.
         """
 
-        raw_ids_values = self._get_ids_values_dict(values)
-        models = [self._id_to_model(id_) for id_ in raw_ids_values]
+        ids_values = self._get_ids_values_dict(values)
+        models = [self._id_to_model(id_) for id_ in ids_values]
         if self._has_different_ctrl_tables:
             assert_same_address(self.model_ctrl_table, models, data_name)
 
         model = next(iter(models))
         addr, length = get_address(self.model_ctrl_table, model, data_name)
 
-        int_ids_values = {id_: int(val) for id_, val in raw_ids_values.items()}
         if normalize and data_name in self.normalized_data:
-            int_ids_values = self._unnormalize(raw_ids_values)
+            ids_values = self._unnormalize(ids_values)
 
-        int_ids_values = self._encode_sign(data_name, int_ids_values)
+        ids_values = self._encode_sign(data_name, ids_values)
 
-        err_msg = f"Failed to sync write '{data_name}' with ids_values={int_ids_values} after {num_retry + 1} tries."
-        self._sync_write(
-            addr, length, int_ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg
-        )
+        err_msg = f"Failed to sync write '{data_name}' with {ids_values=} after {num_retry + 1} tries."
+        self._sync_write(addr, length, ids_values, num_retry=num_retry, raise_on_error=True, err_msg=err_msg)
 
     def _sync_write(
         self,
@@ -1274,7 +1194,3 @@ class SerialMotorsBus(MotorsBusBase):
         for id_, value in ids_values.items():
             data = self._serialize_data(value, length)
             self.sync_writer.addParam(id_, data)
-
-
-# Backward compatibility alias
-MotorsBus: TypeAlias = SerialMotorsBus

src/lerobot/policies/act/configuration_act.py

@@ -28,7 +28,7 @@ class ACTConfig(PreTrainedConfig):
     Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_features` and `output_features`.
+    Those are: `input_shapes` and 'output_shapes`.
 
     Notes on the inputs and outputs:
         - Either:
@@ -48,12 +48,21 @@ class ACTConfig(PreTrainedConfig):
             This should be no greater than the chunk size. For example, if the chunk size size 100, you may
             set this to 50. This would mean that the model predicts 100 steps worth of actions, runs 50 in the
             environment, and throws the other 50 out.
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
             `None` means no pretrained weights.

src/lerobot/policies/diffusion/configuration_diffusion.py

@@ -30,7 +30,7 @@ class DiffusionConfig(PreTrainedConfig):
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_features` and `output_features`.
+    Those are: `input_shapes` and `output_shapes`.
 
     Notes on the inputs and outputs:
         - "observation.state" is required as an input key.
@@ -48,12 +48,21 @@ class DiffusionConfig(PreTrainedConfig):
         horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
         n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
             See `DiffusionPolicy.select_action` for more details.
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
             within the image size. If None, no cropping is done.
@@ -64,7 +73,7 @@ class DiffusionConfig(PreTrainedConfig):
         use_group_norm: Whether to replace batch normalization with group normalization in the backbone.
             The group sizes are set to be about 16 (to be precise, feature_dim // 16).
         spatial_softmax_num_keypoints: Number of keypoints for SpatialSoftmax.
-        use_separate_rgb_encoder_per_camera: Whether to use a separate RGB encoder for each camera view.
+        use_separate_rgb_encoders_per_camera: Whether to use a separate RGB encoder for each camera view.
         down_dims: Feature dimension for each stage of temporal downsampling in the diffusion modeling Unet.
             You may provide a variable number of dimensions, therefore also controlling the degree of
             downsampling.

src/lerobot/policies/sac/modeling_sac.py

@@ -239,10 +239,8 @@ class SACPolicy(
                     + target_param.data * (1.0 - self.config.critic_target_update_weight)
                 )
 
-    @property
-    def temperature(self) -> float:
-        """Return the current temperature value, always in sync with log_alpha."""
-        return self.log_alpha.exp().item()
+    def update_temperature(self):
+        self.temperature = self.log_alpha.exp().item()
 
     def compute_loss_critic(
         self,
@@ -459,10 +457,11 @@ class SACPolicy(
             dim = continuous_action_dim + (1 if self.config.num_discrete_actions is not None else 0)
             self.target_entropy = -np.prod(dim) / 2
 
-    def _init_temperature(self) -> None:
-        """Set up temperature parameter (log_alpha)."""
+    def _init_temperature(self):
+        """Set up temperature parameter and initial log_alpha."""
         temp_init = self.config.temperature_init
         self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
+        self.temperature = self.log_alpha.exp().item()
 
 
 class SACObservationEncoder(nn.Module):

src/lerobot/policies/sarm/rabc.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2025 The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -14,6 +12,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+"""
+RA-BC (Reward-Aligned Behavior Cloning) sample weighting implementation.
+
+This module implements the SampleWeighter protocol for RA-BC training,
+which weights training samples based on their task progress as measured
+by the SARM reward model.
+
+The weights are computed based on progress deltas:
+    delta = progress[t + chunk_size] - progress[t]
+
+High-quality samples (positive progress) get higher weights, while
+samples with negative progress (going backwards) get zero weight.
+
+See: https://arxiv.org/abs/2509.25358 for the SARM paper.
+"""
+
 import logging
 from pathlib import Path
 
@@ -22,6 +36,8 @@ import pandas as pd
 import torch
 from huggingface_hub import hf_hub_download
 
+from lerobot.utils.sample_weighting import SampleWeighter
+
 
 def resolve_hf_path(path: str | Path) -> Path:
     """Resolve a path that may be a HuggingFace URL (hf://datasets/...) to a local path."""
@@ -34,23 +50,27 @@ def resolve_hf_path(path: str | Path) -> Path:
     return Path(path)
 
 
-class RABCWeights:
+class RABCWeights(SampleWeighter):
     """
     Load precomputed SARM progress values and compute RA-BC weights during training.
 
+    This class implements the SampleWeighter ABC for use with the generic
+    sample weighting infrastructure in lerobot.
+
     Progress values are loaded from a parquet file (generated by compute_rabc_weights.py).
     During training, computes:
         - progress_delta = progress[t + chunk_size] - progress[t]
         - rabc_weight based on the delta (paper Eq. 8-9)
 
     Args:
-        progress_path: Path to parquet file with precomputed progress values
-        chunk_size: Number of frames ahead for computing progress delta
-        head_mode: Which SARM head to use ("sparse" or "dense")
-        kappa: Hard threshold for high-quality samples (default: 0.01)
-        epsilon: Small constant for numerical stability (default: 1e-6)
-        fallback_weight: Weight to use for frames without valid delta (default: 1.0)
-        device: Device to return tensors on
+        progress_path: Path to parquet file with precomputed progress values.
+                      Supports HuggingFace URLs (hf://datasets/...).
+        chunk_size: Number of frames ahead for computing progress delta.
+        head_mode: Which SARM head to use ("sparse" or "dense").
+        kappa: Hard threshold for high-quality samples (default: 0.01).
+        epsilon: Small constant for numerical stability (default: 1e-6).
+        fallback_weight: Weight to use for frames without valid delta (default: 1.0).
+        device: Device to return tensors on.
     """
 
     def __init__(
@@ -61,7 +81,7 @@ class RABCWeights:
         kappa: float = 0.01,
         epsilon: float = 1e-6,
         fallback_weight: float = 1.0,
-        device: torch.device = None,
+        device: torch.device | None = None,
     ):
         self.progress_path = resolve_hf_path(progress_path)
         self.chunk_size = chunk_size
@@ -87,8 +107,8 @@ class RABCWeights:
 
         logging.info(f"Using progress column: {self.progress_column}")
 
-        self.progress_lookup = {}
-        self.episode_lookup = {}
+        self.progress_lookup: dict[int, float] = {}
+        self.episode_lookup: dict[int, int] = {}
 
         for _, row in self.df.iterrows():
             global_idx = int(row["index"])
@@ -100,7 +120,7 @@ class RABCWeights:
             self.episode_lookup[global_idx] = episode_idx
 
         # Build episode boundaries for delta computation
-        self.episode_boundaries = {}
+        self.episode_boundaries: dict[int, dict[str, int]] = {}
         for episode_idx in self.df["episode_index"].unique():
             ep_df = self.df[self.df["episode_index"] == episode_idx]
             self.episode_boundaries[int(episode_idx)] = {
@@ -114,7 +134,7 @@ class RABCWeights:
         # Compute global statistics for weight computation
         self._compute_global_stats()
 
-    def _compute_global_stats(self):
+    def _compute_global_stats(self) -> None:
         """Compute global mean and std of progress deltas for weight calculation."""
         all_deltas = []
 
@@ -138,8 +158,8 @@ class RABCWeights:
                 all_deltas.append(delta)
 
         if all_deltas:
-            self.delta_mean = max(np.mean(all_deltas), 0.0)
-            self.delta_std = max(np.std(all_deltas), self.epsilon)
+            self.delta_mean = max(float(np.mean(all_deltas)), 0.0)
+            self.delta_std = max(float(np.std(all_deltas)), self.epsilon)
             logging.info(f"Progress delta stats: mean={self.delta_mean:.4f}, std={self.delta_std:.4f}")
         else:
             self.delta_mean = 0.0
@@ -157,18 +177,19 @@ class RABCWeights:
         4. Compute weight using paper Eq. 8-9
 
         Args:
-            batch: Training batch containing "index" key with global frame indices
+            batch: Training batch containing "index" key with global frame indices.
 
         Returns:
             Tuple of:
-            - Weights tensor (batch_size,) normalized to sum to batch_size
-            - Stats dict with raw_mean_weight, num_zero_weight, num_full_weight
+            - Weights tensor (batch_size,) normalized to sum to batch_size.
+            - Stats dict with weighting statistics for logging.
         """
         indices = batch.get("index")
         if indices is None:
             logging.warning("RA-BC: Batch missing 'index' key, using uniform weights")
             batch_size = self._get_batch_size(batch)
-            return torch.ones(batch_size, device=self.device), {"raw_mean_weight": 1.0}
+            stats = {"mean_weight": 1.0, "num_zero_weight": 0, "num_full_weight": batch_size}
+            return torch.ones(batch_size, device=self.device), stats
 
         # Convert to list of ints
         if isinstance(indices, torch.Tensor):
@@ -183,29 +204,29 @@ class RABCWeights:
             delta = self._compute_delta(idx)
             deltas.append(delta)
 
-        deltas = np.array(deltas, dtype=np.float32)
+        deltas_array = np.array(deltas, dtype=np.float32)
 
         # Compute weights from deltas
-        weights = self._compute_weights(deltas)
+        weights = self._compute_weights(deltas_array)
 
         # Compute stats before normalization for logging
         raw_mean_weight = float(np.nanmean(weights))
         num_zero_weight = int(np.sum(weights == 0))
         num_full_weight = int(np.sum(weights == 1.0))
         batch_stats = {
-            "raw_mean_weight": raw_mean_weight,
+            "mean_weight": raw_mean_weight,
             "num_zero_weight": num_zero_weight,
             "num_full_weight": num_full_weight,
         }
 
-        weights = torch.tensor(weights, device=self.device, dtype=torch.float32)
+        weights_tensor = torch.tensor(weights, device=self.device, dtype=torch.float32)
 
         # Normalize to sum to batch_size
-        batch_size = len(weights)
-        weight_sum = weights.sum() + self.epsilon
-        weights = weights * batch_size / weight_sum
+        batch_size = len(weights_tensor)
+        weight_sum = weights_tensor.sum() + self.epsilon
+        weights_tensor = weights_tensor * batch_size / weight_sum
 
-        return weights, batch_stats
+        return weights_tensor, batch_stats
 
     def _compute_delta(self, global_idx: int) -> float:
         """Compute progress delta for a single frame."""
@@ -241,7 +262,7 @@ class RABCWeights:
         - Final weight: wi = 1{ri > κ} + 1{0 ≤ ri ≤ κ}˜wi
 
         Returns:
-            Array of weights
+            Array of weights.
         """
         valid_mask = ~np.isnan(deltas)
 
@@ -273,12 +294,13 @@ class RABCWeights:
             if key in batch:
                 val = batch[key]
                 if isinstance(val, (torch.Tensor, np.ndarray)):
-                    return val.shape[0]
+                    return int(val.shape[0])
         return 1
 
     def get_stats(self) -> dict:
-        """Get statistics."""
+        """Get global statistics about the RA-BC weighting."""
         return {
+            "type": "rabc",
             "num_frames": len(self.progress_lookup),
             "chunk_size": self.chunk_size,
             "head_mode": self.head_mode,

src/lerobot/policies/smolvla/modeling_smolvla.py

@@ -378,16 +378,16 @@ class SmolVLAPolicy(PreTrainedPolicy):
         actions_is_pad = batch.get("actions_id_pad")
         loss_dict = {}
         losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
-        loss_dict["losses_after_forward"] = losses.clone().mean().item()
+        loss_dict["losses_after_forward"] = losses.clone()
 
         if actions_is_pad is not None:
             in_episode_bound = ~actions_is_pad
             losses = losses * in_episode_bound.unsqueeze(-1)
-            loss_dict["losses_after_in_ep_bound"] = losses.clone().mean().item()
+            loss_dict["losses_after_in_ep_bound"] = losses.clone()
 
         # Remove padding
         losses = losses[:, :, : self.config.max_action_dim]
-        loss_dict["losses_after_rm_padding"] = losses.clone().mean().item()
+        loss_dict["losses_after_rm_padding"] = losses.clone()
 
         if reduction == "none":
             # Return per-sample losses (B,) by averaging over time and action dims

src/lerobot/policies/tdmpc/configuration_tdmpc.py

@@ -30,7 +30,7 @@ class TDMPCConfig(PreTrainedConfig):
     camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_features`, `output_features`, and perhaps `max_random_shift_ratio`.
+    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift_ratio`.
 
     Args:
         n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
@@ -40,12 +40,24 @@ class TDMPCConfig(PreTrainedConfig):
             is an alternative to using action repeats. If this is set to more than 1, then we require
             `n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
             approach of using multiple steps from the plan is not in the original implementation.
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
+            include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
+            the output data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
+            Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range. Note that here this defaults to None meaning inputs are not normalized. This is to
+            match the original implementation.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets. NOTE: Clipping
+            to [-1, +1] is used during MPPI/CEM. Therefore, it is recommended that you stick with "min_max"
+            normalization mode here.
         image_encoder_hidden_dim: Number of channels for the convolutional layers used for image encoding.
         state_encoder_hidden_dim: Hidden dimension for MLP used for state vector encoding.
         latent_dim: Observation's latent embedding dimension.

src/lerobot/policies/vqbet/configuration_vqbet.py

@@ -32,7 +32,7 @@ class VQBeTConfig(PreTrainedConfig):
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_features` and `output_features`.
+    Those are: `input_shapes` and `output_shapes`.
 
     Notes on the inputs and outputs:
         - "observation.state" is required as an input key.
@@ -46,12 +46,21 @@ class VQBeTConfig(PreTrainedConfig):
             current step and additional steps going back).
         n_action_pred_token: Total number of current token and future tokens that VQ-BeT predicts.
         action_chunk_size: Action chunk size of each action prediction token.
-        input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
-            the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
-            the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
-        normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
-            a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
+        input_shapes: A dictionary defining the shapes of the input data for the policy.
+            The key represents the input data name, and the value is a list indicating the dimensions
+            of the corresponding data. For example, "observation.image" refers to an input from
+            a camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
+            Importantly, shapes doesnt include batch dimension or temporal dimension.
+        output_shapes: A dictionary defining the shapes of the output data for the policy.
+            The key represents the output data name, and the value is a list indicating the dimensions
+            of the corresponding data. For example, "action" refers to an output shape of [14], indicating
+            14-dimensional actions. Importantly, shapes doesnt include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
             within the image size. If None, no cropping is done.

src/lerobot/processor/converters.py

@@ -168,12 +168,11 @@ def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
     """
     pad_keys = {k: v for k, v in batch.items() if "_is_pad" in k}
     task_key = {"task": batch["task"]} if "task" in batch else {}
-    subtask_key = {"subtask": batch["subtask"]} if "subtask" in batch else {}
     index_key = {"index": batch["index"]} if "index" in batch else {}
     task_index_key = {"task_index": batch["task_index"]} if "task_index" in batch else {}
     episode_index_key = {"episode_index": batch["episode_index"]} if "episode_index" in batch else {}
 
-    return {**pad_keys, **task_key, **subtask_key, **index_key, **task_index_key, **episode_index_key}
+    return {**pad_keys, **task_key, **index_key, **task_index_key, **episode_index_key}
 
 
 def create_transition(

src/lerobot/processor/env_processor.py

@@ -17,7 +17,7 @@ from dataclasses import dataclass
 
 import torch
 
-from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
+from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 from lerobot.utils.constants import OBS_IMAGES, OBS_PREFIX, OBS_STATE, OBS_STR
 
 from .pipeline import ObservationProcessorStep, ProcessorStepRegistry
@@ -92,7 +92,7 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
         # copy over non-STATE features
         for ft, feats in features.items():
-            if ft != FeatureType.STATE:
+            if ft != PipelineFeatureType.STATE:
                 new_features[ft] = feats.copy()
 
         # rebuild STATE features
@@ -100,11 +100,13 @@ class LiberoProcessorStep(ObservationProcessorStep):
 
         # add our new flattened state
         state_feats[OBS_STATE] = PolicyFeature(
-            type=FeatureType.STATE,
+            key=OBS_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[FeatureType.STATE] = state_feats
+        new_features[PipelineFeatureType.STATE] = state_feats
 
         return new_features
 

src/lerobot/processor/hil_processor.py

@@ -18,18 +18,16 @@
 import math
 import time
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Protocol, TypeVar, runtime_checkable
+from typing import Any, Protocol, TypeVar, runtime_checkable
 
 import numpy as np
 import torch
 import torchvision.transforms.functional as F  # noqa: N812
 
 from lerobot.configs.types import PipelineFeatureType, PolicyFeature
+from lerobot.teleoperators.teleoperator import Teleoperator
 from lerobot.teleoperators.utils import TeleopEvents
 
-if TYPE_CHECKING:
-    from lerobot.teleoperators.teleoperator import Teleoperator
-
 from .core import EnvTransition, PolicyAction, TransitionKey
 from .pipeline import (
     ComplementaryDataProcessorStep,
@@ -71,10 +69,10 @@ class HasTeleopEvents(Protocol):
 
 
 # Type variable constrained to Teleoperator subclasses that also implement events
-TeleopWithEvents = TypeVar("TeleopWithEvents", bound="Teleoperator")
+TeleopWithEvents = TypeVar("TeleopWithEvents", bound=Teleoperator)
 
 
-def _check_teleop_with_events(teleop: "Teleoperator") -> None:
+def _check_teleop_with_events(teleop: Teleoperator) -> None:
     """
     Runtime check that a teleoperator implements the `HasTeleopEvents` protocol.
 
@@ -105,7 +103,7 @@ class AddTeleopActionAsComplimentaryDataStep(ComplementaryDataProcessorStep):
         teleop_device: The teleoperator instance to get the action from.
     """
 
-    teleop_device: "Teleoperator"
+    teleop_device: Teleoperator
 
     def complementary_data(self, complementary_data: dict) -> dict:
         """
@@ -314,7 +312,7 @@ class TimeLimitProcessorStep(TruncatedProcessorStep):
 
 @dataclass
 @ProcessorStepRegistry.register("gripper_penalty_processor")
-class GripperPenaltyProcessorStep(ProcessorStep):
+class GripperPenaltyProcessorStep(ComplementaryDataProcessorStep):
     """
     Applies a penalty for inefficient gripper usage.
 
@@ -329,27 +327,26 @@ class GripperPenaltyProcessorStep(ProcessorStep):
     penalty: float = -0.01
     max_gripper_pos: float = 30.0
 
-    def __call__(self, transition: EnvTransition) -> EnvTransition:
+    def complementary_data(self, complementary_data: dict) -> dict:
         """
         Calculates the gripper penalty and adds it to the complementary data.
 
         Args:
-            transition: The incoming environment transition.
+            complementary_data: The incoming complementary data, which should contain
+                                raw joint positions.
 
         Returns:
-            The modified transition with the penalty added to complementary data.
+            A new complementary data dictionary with the `discrete_penalty` key added.
         """
-        new_transition = transition.copy()
-        action = new_transition.get(TransitionKey.ACTION)
-        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        action = self.transition.get(TransitionKey.ACTION)
 
         raw_joint_positions = complementary_data.get("raw_joint_positions")
         if raw_joint_positions is None:
-            return new_transition
+            return complementary_data
 
         current_gripper_pos = raw_joint_positions.get(GRIPPER_KEY, None)
         if current_gripper_pos is None:
-            return new_transition
+            return complementary_data
 
         # Gripper action is a PolicyAction at this stage
         gripper_action = action[-1].item()
@@ -365,12 +362,11 @@ class GripperPenaltyProcessorStep(ProcessorStep):
 
         gripper_penalty = self.penalty * int(gripper_penalty_bool)
 
-        # Update complementary data with penalty info
+        # Create new complementary data with penalty info
         new_complementary_data = dict(complementary_data)
         new_complementary_data[DISCRETE_PENALTY_KEY] = gripper_penalty
-        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
 
-        return new_transition
+        return new_complementary_data
 
     def get_config(self) -> dict[str, Any]:
         """

src/lerobot/processor/tokenizer_processor.py

@@ -34,8 +34,6 @@ from lerobot.utils.constants import (
     ACTION_TOKEN_MASK,
     ACTION_TOKENS,
     OBS_LANGUAGE_ATTENTION_MASK,
-    OBS_LANGUAGE_SUBTASK_ATTENTION_MASK,
-    OBS_LANGUAGE_SUBTASK_TOKENS,
     OBS_LANGUAGE_TOKENS,
 )
 from lerobot.utils.import_utils import _transformers_available
@@ -141,32 +139,6 @@ class TokenizerProcessorStep(ObservationProcessorStep):
 
         return None
 
-    def get_subtask(self, transition: EnvTransition) -> list[str] | None:
-        """
-        Extracts the subtask from the transition's complementary data.
-
-        Args:
-            transition: The environment transition.
-
-        Returns:
-            A list of subtask strings, or None if the subtask key is not found or the value is None.
-        """
-        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
-        if complementary_data is None:
-            return None
-
-        subtask = complementary_data.get("subtask")
-        if subtask is None:
-            return None
-
-        # Standardize to a list of strings for the tokenizer
-        if isinstance(subtask, str):
-            return [subtask]
-        elif isinstance(subtask, list) and all(isinstance(t, str) for t in subtask):
-            return subtask
-
-        return None
-
     def observation(self, observation: RobotObservation) -> RobotObservation:
         """
         Tokenizes the task description and adds it to the observation dictionary.
@@ -204,24 +176,6 @@ class TokenizerProcessorStep(ObservationProcessorStep):
         new_observation[OBS_LANGUAGE_TOKENS] = tokenized_prompt["input_ids"]
         new_observation[OBS_LANGUAGE_ATTENTION_MASK] = tokenized_prompt["attention_mask"].to(dtype=torch.bool)
 
-        # Tokenize subtask if available
-        subtask = self.get_subtask(self.transition)
-        if subtask is not None:
-            tokenized_subtask = self._tokenize_text(subtask)
-
-            # Move new tokenized tensors to the detected device
-            if target_device is not None:
-                tokenized_subtask = {
-                    k: v.to(target_device) if isinstance(v, torch.Tensor) else v
-                    for k, v in tokenized_subtask.items()
-                }
-
-            # Add tokenized subtask to the observation
-            new_observation[OBS_LANGUAGE_SUBTASK_TOKENS] = tokenized_subtask["input_ids"]
-            new_observation[OBS_LANGUAGE_SUBTASK_ATTENTION_MASK] = tokenized_subtask["attention_mask"].to(
-                dtype=torch.bool
-            )
-
         return new_observation
 
     def _detect_device(self, transition: EnvTransition) -> torch.device | None:

src/lerobot/rl/gym_manipulator.py

@@ -412,10 +412,7 @@ def make_processors(
         if cfg.processor.observation.add_current_to_observation:
             env_pipeline_steps.append(MotorCurrentProcessorStep(robot=env.robot))
 
-    add_ee_pose = (
-        cfg.processor.observation is not None and cfg.processor.observation.add_ee_pose_to_observation
-    )
-    if kinematics_solver is not None and add_ee_pose:
+    if kinematics_solver is not None:
         env_pipeline_steps.append(
             ForwardKinematicsJointsToEEObservation(
                 kinematics=kinematics_solver,
@@ -438,12 +435,7 @@ def make_processors(
         )
 
     # Add gripper penalty processor if gripper config exists and enabled
-    # Only add if max_gripper_pos is explicitly configured (required for normalization)
-    if (
-        cfg.processor.gripper is not None
-        and cfg.processor.gripper.use_gripper
-        and cfg.processor.max_gripper_pos is not None
-    ):
+    if cfg.processor.gripper is not None and cfg.processor.gripper.use_gripper:
         env_pipeline_steps.append(
             GripperPenaltyProcessorStep(
                 penalty=cfg.processor.gripper.gripper_penalty,

src/lerobot/rl/learner.py

@@ -545,6 +545,9 @@ def add_actor_information_and_train(
                 training_infos["temperature_grad_norm"] = temp_grad_norm
                 training_infos["temperature"] = policy.temperature
 
+                # Update temperature
+                policy.update_temperature()
+
         # Push policy to actors if needed
         if time.time() - last_time_policy_pushed > policy_parameters_push_frequency:
             push_actor_policy_to_queue(parameters_queue=parameters_queue, policy=policy)

src/lerobot/rl/wandb_utils.py

@@ -26,21 +26,8 @@ from lerobot.configs.train import TrainPipelineConfig
 from lerobot.utils.constants import PRETRAINED_MODEL_DIR
 
 
-def cfg_to_group(
-    cfg: TrainPipelineConfig, return_list: bool = False, truncate_tags: bool = False, max_tag_length: int = 64
-) -> list[str] | str:
+def cfg_to_group(cfg: TrainPipelineConfig, return_list: bool = False) -> list[str] | str:
     """Return a group name for logging. Optionally returns group name as list."""
-
-    def _maybe_truncate(tag: str) -> str:
-        """Truncate tag to max_tag_length characters if required.
-
-        wandb rejects tags longer than 64 characters.
-        See: https://github.com/wandb/wandb/blob/main/wandb/sdk/wandb_settings.py
-        """
-        if len(tag) <= max_tag_length:
-            return tag
-        return tag[:max_tag_length]
-
     lst = [
         f"policy:{cfg.policy.type}",
         f"seed:{cfg.seed}",
@@ -49,8 +36,6 @@ def cfg_to_group(
         lst.append(f"dataset:{cfg.dataset.repo_id}")
     if cfg.env is not None:
         lst.append(f"env:{cfg.env.type}")
-    if truncate_tags:
-        lst = [_maybe_truncate(tag) for tag in lst]
     return lst if return_list else "-".join(lst)
 
 
@@ -98,7 +83,7 @@ class WandBLogger:
             entity=self.cfg.entity,
             name=self.job_name,
             notes=self.cfg.notes,
-            tags=cfg_to_group(cfg, return_list=True, truncate_tags=True),
+            tags=cfg_to_group(cfg, return_list=True),
             dir=self.log_dir,
             config=cfg.to_dict(),
             # TODO(rcadene): try set to True

src/lerobot/robots/bi_openarm_follower/__init__.py

@@ -1,20 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 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 .bi_openarm_follower import BiOpenArmFollower
-from .config_bi_openarm_follower import BiOpenArmFollowerConfig
-
-__all__ = ["BiOpenArmFollower", "BiOpenArmFollowerConfig"]

src/lerobot/robots/bi_openarm_follower/bi_openarm_follower.py

@@ -1,180 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import logging
-from functools import cached_property
-
-from lerobot.processor import RobotAction, RobotObservation
-from lerobot.robots.openarm_follower import OpenArmFollower, OpenArmFollowerConfig
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-
-from ..robot import Robot
-from .config_bi_openarm_follower import BiOpenArmFollowerConfig
-
-logger = logging.getLogger(__name__)
-
-
-class BiOpenArmFollower(Robot):
-    """
-    Bimanual OpenArm Follower Arms
-    """
-
-    config_class = BiOpenArmFollowerConfig
-    name = "bi_openarm_follower"
-
-    def __init__(self, config: BiOpenArmFollowerConfig):
-        super().__init__(config)
-        self.config = config
-
-        left_arm_config = OpenArmFollowerConfig(
-            id=f"{config.id}_left" if config.id else None,
-            calibration_dir=config.calibration_dir,
-            port=config.left_arm_config.port,
-            disable_torque_on_disconnect=config.left_arm_config.disable_torque_on_disconnect,
-            max_relative_target=config.left_arm_config.max_relative_target,
-            cameras=config.left_arm_config.cameras,
-            side=config.left_arm_config.side,
-            can_interface=config.left_arm_config.can_interface,
-            use_can_fd=config.left_arm_config.use_can_fd,
-            can_bitrate=config.left_arm_config.can_bitrate,
-            can_data_bitrate=config.left_arm_config.can_data_bitrate,
-            motor_config=config.left_arm_config.motor_config,
-            position_kd=config.left_arm_config.position_kd,
-            position_kp=config.left_arm_config.position_kp,
-            joint_limits=config.left_arm_config.joint_limits,
-        )
-
-        right_arm_config = OpenArmFollowerConfig(
-            id=f"{config.id}_right" if config.id else None,
-            calibration_dir=config.calibration_dir,
-            port=config.right_arm_config.port,
-            disable_torque_on_disconnect=config.right_arm_config.disable_torque_on_disconnect,
-            max_relative_target=config.right_arm_config.max_relative_target,
-            cameras=config.right_arm_config.cameras,
-            side=config.right_arm_config.side,
-            can_interface=config.right_arm_config.can_interface,
-            use_can_fd=config.right_arm_config.use_can_fd,
-            can_bitrate=config.right_arm_config.can_bitrate,
-            can_data_bitrate=config.right_arm_config.can_data_bitrate,
-            motor_config=config.right_arm_config.motor_config,
-            position_kd=config.right_arm_config.position_kd,
-            position_kp=config.right_arm_config.position_kp,
-            joint_limits=config.right_arm_config.joint_limits,
-        )
-
-        self.left_arm = OpenArmFollower(left_arm_config)
-        self.right_arm = OpenArmFollower(right_arm_config)
-
-        # Only for compatibility with other parts of the codebase that expect a `robot.cameras` attribute
-        self.cameras = {**self.left_arm.cameras, **self.right_arm.cameras}
-
-    @property
-    def _motors_ft(self) -> dict[str, type]:
-        left_arm_motors_ft = self.left_arm._motors_ft
-        right_arm_motors_ft = self.right_arm._motors_ft
-
-        return {
-            **{f"left_{k}": v for k, v in left_arm_motors_ft.items()},
-            **{f"right_{k}": v for k, v in right_arm_motors_ft.items()},
-        }
-
-    @property
-    def _cameras_ft(self) -> dict[str, tuple]:
-        left_arm_cameras_ft = self.left_arm._cameras_ft
-        right_arm_cameras_ft = self.right_arm._cameras_ft
-
-        return {
-            **{f"left_{k}": v for k, v in left_arm_cameras_ft.items()},
-            **{f"right_{k}": v for k, v in right_arm_cameras_ft.items()},
-        }
-
-    @cached_property
-    def observation_features(self) -> dict[str, type | tuple]:
-        return {**self._motors_ft, **self._cameras_ft}
-
-    @cached_property
-    def action_features(self) -> dict[str, type]:
-        return self._motors_ft
-
-    @property
-    def is_connected(self) -> bool:
-        return self.left_arm.is_connected and self.right_arm.is_connected
-
-    @check_if_already_connected
-    def connect(self, calibrate: bool = True) -> None:
-        self.left_arm.connect(calibrate)
-        self.right_arm.connect(calibrate)
-
-    @property
-    def is_calibrated(self) -> bool:
-        return self.left_arm.is_calibrated and self.right_arm.is_calibrated
-
-    def calibrate(self) -> None:
-        self.left_arm.calibrate()
-        self.right_arm.calibrate()
-
-    def configure(self) -> None:
-        self.left_arm.configure()
-        self.right_arm.configure()
-
-    def setup_motors(self) -> None:
-        raise NotImplementedError(
-            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
-        )
-
-    @check_if_not_connected
-    def get_observation(self) -> RobotObservation:
-        obs_dict = {}
-
-        # Add "left_" prefix
-        left_obs = self.left_arm.get_observation()
-        obs_dict.update({f"left_{key}": value for key, value in left_obs.items()})
-
-        # Add "right_" prefix
-        right_obs = self.right_arm.get_observation()
-        obs_dict.update({f"right_{key}": value for key, value in right_obs.items()})
-
-        return obs_dict
-
-    @check_if_not_connected
-    def send_action(
-        self,
-        action: RobotAction,
-        custom_kp: dict[str, float] | None = None,
-        custom_kd: dict[str, float] | None = None,
-    ) -> RobotAction:
-        # Remove "left_" prefix
-        left_action = {
-            key.removeprefix("left_"): value for key, value in action.items() if key.startswith("left_")
-        }
-        # Remove "right_" prefix
-        right_action = {
-            key.removeprefix("right_"): value for key, value in action.items() if key.startswith("right_")
-        }
-
-        sent_action_left = self.left_arm.send_action(left_action, custom_kp, custom_kd)
-        sent_action_right = self.right_arm.send_action(right_action, custom_kp, custom_kd)
-
-        # Add prefixes back
-        prefixed_sent_action_left = {f"left_{key}": value for key, value in sent_action_left.items()}
-        prefixed_sent_action_right = {f"right_{key}": value for key, value in sent_action_right.items()}
-
-        return {**prefixed_sent_action_left, **prefixed_sent_action_right}
-
-    @check_if_not_connected
-    def disconnect(self):
-        self.left_arm.disconnect()
-        self.right_arm.disconnect()

src/lerobot/robots/bi_openarm_follower/config_bi_openarm_follower.py

@@ -1,30 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from dataclasses import dataclass
-
-from lerobot.robots.openarm_follower import OpenArmFollowerConfigBase
-
-from ..config import RobotConfig
-
-
-@RobotConfig.register_subclass("bi_openarm_follower")
-@dataclass
-class BiOpenArmFollowerConfig(RobotConfig):
-    """Configuration class for Bi OpenArm Follower robots."""
-
-    left_arm_config: OpenArmFollowerConfigBase
-    right_arm_config: OpenArmFollowerConfigBase

src/lerobot/robots/bi_so_follower/bi_so_follower.py

@@ -19,7 +19,6 @@ from functools import cached_property
 
 from lerobot.processor import RobotAction, RobotObservation
 from lerobot.robots.so_follower import SOFollower, SOFollowerRobotConfig
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
 
 from ..robot import Robot
 from .config_bi_so_follower import BiSOFollowerConfig
@@ -97,7 +96,6 @@ class BiSOFollower(Robot):
     def is_connected(self) -> bool:
         return self.left_arm.is_connected and self.right_arm.is_connected
 
-    @check_if_already_connected
     def connect(self, calibrate: bool = True) -> None:
         self.left_arm.connect(calibrate)
         self.right_arm.connect(calibrate)
@@ -118,7 +116,6 @@ class BiSOFollower(Robot):
         self.left_arm.setup_motors()
         self.right_arm.setup_motors()
 
-    @check_if_not_connected
     def get_observation(self) -> RobotObservation:
         obs_dict = {}
 
@@ -132,7 +129,6 @@ class BiSOFollower(Robot):
 
         return obs_dict
 
-    @check_if_not_connected
     def send_action(self, action: RobotAction) -> RobotAction:
         # Remove "left_" prefix
         left_action = {
@@ -152,7 +148,6 @@ class BiSOFollower(Robot):
 
         return {**prefixed_sent_action_left, **prefixed_sent_action_right}
 
-    @check_if_not_connected
     def disconnect(self):
         self.left_arm.disconnect()
         self.right_arm.disconnect()

src/lerobot/robots/openarm_follower/__init__.py

@@ -1,20 +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 .config_openarm_follower import OpenArmFollowerConfig, OpenArmFollowerConfigBase
-from .openarm_follower import OpenArmFollower
-
-__all__ = ["OpenArmFollower", "OpenArmFollowerConfig", "OpenArmFollowerConfigBase"]

src/lerobot/robots/openarm_follower/config_openarm_follower.py

@@ -1,122 +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, field
-
-from lerobot.cameras import CameraConfig
-
-from ..config import RobotConfig
-
-LEFT_DEFAULT_JOINTS_LIMITS: dict[str, tuple[float, float]] = {
-    "joint_1": (-75.0, 75.0),
-    "joint_2": (-90.0, 9.0),
-    "joint_3": (-85.0, 85.0),
-    "joint_4": (0.0, 135.0),
-    "joint_5": (-85.0, 85.0),
-    "joint_6": (-40.0, 40.0),
-    "joint_7": (-80.0, 80.0),
-    "gripper": (-65.0, 0.0),
-}
-
-RIGHT_DEFAULT_JOINTS_LIMITS: dict[str, tuple[float, float]] = {
-    "joint_1": (-75.0, 75.0),
-    "joint_2": (-9.0, 90.0),
-    "joint_3": (-85.0, 85.0),
-    "joint_4": (0.0, 135.0),
-    "joint_5": (-85.0, 85.0),
-    "joint_6": (-40.0, 40.0),
-    "joint_7": (-80.0, 80.0),
-    "gripper": (-65.0, 0.0),
-}
-
-
-@dataclass
-class OpenArmFollowerConfigBase:
-    """Base configuration for the OpenArms follower robot with Damiao motors."""
-
-    # CAN interfaces - one per arm
-    # arm CAN interface (e.g., "can1")
-    # Linux: "can0", "can1", etc.
-    port: str
-
-    # side of the arm: "left" or "right". If "None" default values will be used
-    side: str | None = None
-
-    # CAN interface type: "socketcan" (Linux), "slcan" (serial), or "auto" (auto-detect)
-    can_interface: str = "socketcan"
-
-    # CAN FD settings (OpenArms uses CAN FD by default)
-    use_can_fd: bool = True
-    can_bitrate: int = 1000000  # Nominal bitrate (1 Mbps)
-    can_data_bitrate: int = 5000000  # Data bitrate for CAN FD (5 Mbps)
-
-    # Whether to disable torque when disconnecting
-    disable_torque_on_disconnect: bool = True
-
-    # Safety limit for relative target positions
-    # Set to a positive scalar for all motors, or a dict mapping motor names to limits
-    max_relative_target: float | dict[str, float] | None = None
-
-    # Camera configurations
-    cameras: dict[str, CameraConfig] = field(default_factory=dict)
-
-    # Motor configuration for OpenArms (7 DOF per arm)
-    # Maps motor names to (send_can_id, recv_can_id, motor_type)
-    # Based on: https://docs.openarm.dev/software/setup/configure-test
-    # OpenArms uses 4 types of motors:
-    # - DM8009 (DM-J8009P-2EC) for shoulders (high torque)
-    # - DM4340P and DM4340 for shoulder rotation and elbow
-    # - DM4310 (DM-J4310-2EC V1.1) for wrist and gripper
-    motor_config: dict[str, tuple[int, int, str]] = field(
-        default_factory=lambda: {
-            "joint_1": (0x01, 0x11, "dm8009"),  # J1 - Shoulder pan (DM8009)
-            "joint_2": (0x02, 0x12, "dm8009"),  # J2 - Shoulder lift (DM8009)
-            "joint_3": (0x03, 0x13, "dm4340"),  # J3 - Shoulder rotation (DM4340)
-            "joint_4": (0x04, 0x14, "dm4340"),  # J4 - Elbow flex (DM4340)
-            "joint_5": (0x05, 0x15, "dm4310"),  # J5 - Wrist roll (DM4310)
-            "joint_6": (0x06, 0x16, "dm4310"),  # J6 - Wrist pitch (DM4310)
-            "joint_7": (0x07, 0x17, "dm4310"),  # J7 - Wrist rotation (DM4310)
-            "gripper": (0x08, 0x18, "dm4310"),  # J8 - Gripper (DM4310)
-        }
-    )
-
-    # MIT control parameters for position control (used in send_action)
-    # List of 8 values: [joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, joint_7, gripper]
-    position_kp: list[float] = field(
-        default_factory=lambda: [240.0, 240.0, 240.0, 240.0, 24.0, 31.0, 25.0, 25.0]
-    )
-    position_kd: list[float] = field(default_factory=lambda: [5.0, 5.0, 3.0, 5.0, 0.3, 0.3, 0.3, 0.3])
-
-    # Values for joint limits. Can be overridden via CLI (for custom values) or by setting config.side to either 'left' or 'right'.
-    # If config.side is left set to None and no CLI values are passed, the default joint limit values are small for safety.
-    joint_limits: dict[str, tuple[float, float]] = field(
-        default_factory=lambda: {
-            "joint_1": (-5.0, 5.0),
-            "joint_2": (-5.0, 5.0),
-            "joint_3": (-5.0, 5.0),
-            "joint_4": (0.0, 5.0),
-            "joint_5": (-5.0, 5.0),
-            "joint_6": (-5.0, 5.0),
-            "joint_7": (-5.0, 5.0),
-            "gripper": (-5.0, 0.0),
-        }
-    )
-
-
-@RobotConfig.register_subclass("openarm_follower")
-@dataclass
-class OpenArmFollowerConfig(RobotConfig, OpenArmFollowerConfigBase):
-    pass

src/lerobot/robots/openarm_follower/openarm_follower.py

@@ -1,343 +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 logging
-import time
-from functools import cached_property
-from typing import Any
-
-from lerobot.cameras.utils import make_cameras_from_configs
-from lerobot.motors import Motor, MotorCalibration, MotorNormMode
-from lerobot.motors.damiao import DamiaoMotorsBus
-from lerobot.processor import RobotAction, RobotObservation
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-
-from ..robot import Robot
-from ..utils import ensure_safe_goal_position
-from .config_openarm_follower import (
-    LEFT_DEFAULT_JOINTS_LIMITS,
-    RIGHT_DEFAULT_JOINTS_LIMITS,
-    OpenArmFollowerConfig,
-)
-
-logger = logging.getLogger(__name__)
-
-
-class OpenArmFollower(Robot):
-    """
-    OpenArms Follower Robot which uses CAN bus communication to control 7 DOF arm with a gripper.
-    The arm uses Damiao motors in MIT control mode.
-    """
-
-    config_class = OpenArmFollowerConfig
-    name = "openarm_follower"
-
-    def __init__(self, config: OpenArmFollowerConfig):
-        super().__init__(config)
-        self.config = config
-
-        # Arm motors
-        motors: dict[str, Motor] = {}
-        for motor_name, (send_id, recv_id, motor_type_str) in config.motor_config.items():
-            motor = Motor(
-                send_id, motor_type_str, MotorNormMode.DEGREES
-            )  # Always use degrees for Damiao motors
-            motor.recv_id = recv_id
-            motor.motor_type_str = motor_type_str
-            motors[motor_name] = motor
-
-        self.bus = DamiaoMotorsBus(
-            port=self.config.port,
-            motors=motors,
-            calibration=self.calibration,
-            can_interface=self.config.can_interface,
-            use_can_fd=self.config.use_can_fd,
-            bitrate=self.config.can_bitrate,
-            data_bitrate=self.config.can_data_bitrate if self.config.use_can_fd else None,
-        )
-
-        if config.side is not None:
-            if config.side == "left":
-                config.joint_limits = LEFT_DEFAULT_JOINTS_LIMITS
-            elif config.side == "right":
-                config.joint_limits = RIGHT_DEFAULT_JOINTS_LIMITS
-            else:
-                raise ValueError(
-                    "config.side must be either 'left', 'right' (for default values) or 'None' (for CLI values)"
-                )
-        else:
-            logger.info(
-                "Set config.side to either 'left' or 'right' to use pre-configured values for joint limits."
-            )
-        logger.info(f"Values used for joint limits: {config.joint_limits}.")
-
-        # Initialize cameras
-        self.cameras = make_cameras_from_configs(config.cameras)
-
-    @property
-    def _motors_ft(self) -> dict[str, type]:
-        """Motor features for observation and action spaces."""
-        features: dict[str, type] = {}
-        for motor in self.bus.motors:
-            features[f"{motor}.pos"] = float
-            features[f"{motor}.vel"] = float  # Add this
-            features[f"{motor}.torque"] = float  # Add this
-        return features
-
-    @property
-    def _cameras_ft(self) -> dict[str, tuple]:
-        """Camera features for observation space."""
-        return {
-            cam: (self.config.cameras[cam].height, self.config.cameras[cam].width, 3) for cam in self.cameras
-        }
-
-    @cached_property
-    def observation_features(self) -> dict[str, type | tuple]:
-        """Combined observation features from motors and cameras."""
-        return {**self._motors_ft, **self._cameras_ft}
-
-    @cached_property
-    def action_features(self) -> dict[str, type]:
-        """Action features."""
-        return self._motors_ft
-
-    @property
-    def is_connected(self) -> bool:
-        """Check if robot is connected."""
-        return self.bus.is_connected and all(cam.is_connected for cam in self.cameras.values())
-
-    @check_if_already_connected
-    def connect(self, calibrate: bool = True) -> None:
-        """
-        Connect to the robot and optionally calibrate.
-
-        We assume that at connection time, the arms are in a safe rest position,
-        and torque can be safely disabled to run calibration if needed.
-        """
-
-        # Connect to CAN bus
-        logger.info(f"Connecting arm on {self.config.port}...")
-        self.bus.connect()
-
-        # Run calibration if needed
-        if not self.is_calibrated and calibrate:
-            logger.info(
-                "Mismatch between calibration values in the motor and the calibration file or no calibration file found"
-            )
-            self.calibrate()
-
-        for cam in self.cameras.values():
-            cam.connect()
-
-        self.configure()
-
-        if self.is_calibrated:
-            self.bus.set_zero_position()
-
-        self.bus.enable_torque()
-
-        logger.info(f"{self} connected.")
-
-    @property
-    def is_calibrated(self) -> bool:
-        """Check if robot is calibrated."""
-        return self.bus.is_calibrated
-
-    def calibrate(self) -> None:
-        """
-        Run calibration procedure for OpenArms robot.
-
-        The calibration procedure:
-        1. Disable torque
-        2. Ask user to position arms in hanging position with grippers closed
-        3. Set this as zero position
-        4. Record range of motion for each joint
-        5. Save calibration
-        """
-        if self.calibration:
-            # Calibration file exists, ask user whether to use it or run new calibration
-            user_input = input(
-                f"Press ENTER to use provided calibration file associated with the id {self.id}, or type 'c' and press ENTER to run calibration: "
-            )
-            if user_input.strip().lower() != "c":
-                logger.info(f"Writing calibration file associated with the id {self.id} to the motors")
-                self.bus.write_calibration(self.calibration)
-                return
-
-        logger.info(f"\nRunning calibration for {self}")
-        self.bus.disable_torque()
-
-        # Step 1: Set zero position
-        input(
-            "\nCalibration: Set Zero Position)\n"
-            "Position the arm in the following configuration:\n"
-            "  - Arm hanging straight down\n"
-            "  - Gripper closed\n"
-            "Press ENTER when ready..."
-        )
-
-        # Set current position as zero for all motors
-        self.bus.set_zero_position()
-        logger.info("Arm zero position set.")
-
-        logger.info("Setting range: -90° to +90° for safety by default for all joints")
-        for motor_name, motor in self.bus.motors.items():
-            self.calibration[motor_name] = MotorCalibration(
-                id=motor.id,
-                drive_mode=0,
-                homing_offset=0,
-                range_min=-90,
-                range_max=90,
-            )
-
-        self.bus.write_calibration(self.calibration)
-        self._save_calibration()
-        print(f"Calibration saved to {self.calibration_fpath}")
-
-    def configure(self) -> None:
-        """Configure motors with appropriate settings."""
-        # TODO(Steven, Pepijn): Slightly different from what it is happening in the leader
-        with self.bus.torque_disabled():
-            self.bus.configure_motors()
-
-    def setup_motors(self) -> None:
-        raise NotImplementedError(
-            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
-        )
-
-    @check_if_not_connected
-    def get_observation(self) -> RobotObservation:
-        """
-        Get current observation from robot including position, velocity, and torque.
-
-        Reads all motor states (pos/vel/torque) in one CAN refresh cycle
-        instead of 3 separate reads.
-        """
-        start = time.perf_counter()
-
-        obs_dict: dict[str, Any] = {}
-
-        states = self.bus.sync_read_all_states()
-
-        for motor in self.bus.motors:
-            state = states.get(motor, {})
-            obs_dict[f"{motor}.pos"] = state.get("position", 0.0)
-            obs_dict[f"{motor}.vel"] = state.get("velocity", 0.0)
-            obs_dict[f"{motor}.torque"] = state.get("torque", 0.0)
-
-        # Capture images from cameras
-        for cam_key, cam in self.cameras.items():
-            start = time.perf_counter()
-            obs_dict[cam_key] = cam.async_read()
-            dt_ms = (time.perf_counter() - start) * 1e3
-            logger.debug(f"{self} read {cam_key}: {dt_ms:.1f}ms")
-
-        dt_ms = (time.perf_counter() - start) * 1e3
-        logger.debug(f"{self} get_observation took: {dt_ms:.1f}ms")
-
-        return obs_dict
-
-    @check_if_not_connected
-    def send_action(
-        self,
-        action: RobotAction,
-        custom_kp: dict[str, float] | None = None,
-        custom_kd: dict[str, float] | None = None,
-    ) -> RobotAction:
-        """
-        Send action command to robot.
-
-        The action magnitude may be clipped based on safety limits.
-
-        Args:
-            action: Dictionary with motor positions (e.g., "joint_1.pos", "joint_2.pos")
-            custom_kp: Optional custom kp gains per motor (e.g., {"joint_1": 120.0, "joint_2": 150.0})
-            custom_kd: Optional custom kd gains per motor (e.g., {"joint_1": 1.5, "joint_2": 2.0})
-
-        Returns:
-            The action actually sent (potentially clipped)
-        """
-
-        goal_pos = {key.removesuffix(".pos"): val for key, val in action.items() if key.endswith(".pos")}
-
-        # Apply joint limit clipping to arm
-        for motor_name, position in goal_pos.items():
-            if motor_name in self.config.joint_limits:
-                min_limit, max_limit = self.config.joint_limits[motor_name]
-                clipped_position = max(min_limit, min(max_limit, position))
-                if clipped_position != position:
-                    logger.debug(f"Clipped {motor_name} from {position:.2f}° to {clipped_position:.2f}°")
-                goal_pos[motor_name] = clipped_position
-
-        # Cap goal position when too far away from present position.
-        # /!\ Slower fps expected due to reading from the follower.
-        if self.config.max_relative_target is not None:
-            present_pos = self.bus.sync_read("Present_Position")
-            goal_present_pos = {key: (g_pos, present_pos[key]) for key, g_pos in goal_pos.items()}
-            goal_pos = ensure_safe_goal_position(goal_present_pos, self.config.max_relative_target)
-
-        # TODO(Steven, Pepijn): Refactor writing
-        # Motor name to index mapping for gains
-        motor_index = {
-            "joint_1": 0,
-            "joint_2": 1,
-            "joint_3": 2,
-            "joint_4": 3,
-            "joint_5": 4,
-            "joint_6": 5,
-            "joint_7": 6,
-            "gripper": 7,
-        }
-
-        # Use batch MIT control for arm (sends all commands, then collects responses)
-        commands = {}
-        for motor_name, position_degrees in goal_pos.items():
-            idx = motor_index.get(motor_name, 0)
-            # Use custom gains if provided, otherwise use config defaults
-            if custom_kp is not None and motor_name in custom_kp:
-                kp = custom_kp[motor_name]
-            else:
-                kp = (
-                    self.config.position_kp[idx]
-                    if isinstance(self.config.position_kp, list)
-                    else self.config.position_kp
-                )
-            if custom_kd is not None and motor_name in custom_kd:
-                kd = custom_kd[motor_name]
-            else:
-                kd = (
-                    self.config.position_kd[idx]
-                    if isinstance(self.config.position_kd, list)
-                    else self.config.position_kd
-                )
-            commands[motor_name] = (kp, kd, position_degrees, 0.0, 0.0)
-
-        self.bus._mit_control_batch(commands)
-
-        return {f"{motor}.pos": val for motor, val in goal_pos.items()}
-
-    @check_if_not_connected
-    def disconnect(self):
-        """Disconnect from robot."""
-
-        # Disconnect CAN bus
-        self.bus.disconnect(self.config.disable_torque_on_disconnect)
-
-        # Disconnect cameras
-        for cam in self.cameras.values():
-            cam.disconnect()
-
-        logger.info(f"{self} disconnected.")

src/lerobot/robots/unitree_g1/config_unitree_g1.py

@@ -65,6 +65,3 @@ class UnitreeG1Config(RobotConfig):
 
     # Cameras (ZMQ-based remote cameras)
     cameras: dict[str, CameraConfig] = field(default_factory=dict)
-
-    # Compensates for gravity on the unitree's arms using the arm ik solver
-    gravity_compensation: bool = False

src/lerobot/robots/unitree_g1/g1_utils.py

@@ -18,7 +18,7 @@ from enum import IntEnum
 
 # ruff: noqa: N801, N815
 
-NUM_MOTORS = 29
+NUM_MOTORS = 35
 
 
 class G1_29_JointArmIndex(IntEnum):

src/lerobot/robots/unitree_g1/robot_kinematic_processor.py

@@ -1,313 +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 logging
-import os
-import sys
-
-import numpy as np
-
-logger = logging.getLogger(__name__)
-parent2_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-sys.path.append(parent2_dir)
-
-
-class WeightedMovingFilter:
-    def __init__(self, weights, data_size=14):
-        self._window_size = len(weights)
-        self._weights = np.array(weights)
-        self._data_size = data_size
-        self._filtered_data = np.zeros(self._data_size)
-        self._data_queue = []
-
-    def _apply_filter(self):
-        if len(self._data_queue) < self._window_size:
-            return self._data_queue[-1]
-
-        data_array = np.array(self._data_queue)
-        temp_filtered_data = np.zeros(self._data_size)
-        for i in range(self._data_size):
-            temp_filtered_data[i] = np.convolve(data_array[:, i], self._weights, mode="valid")[-1]
-
-        return temp_filtered_data
-
-    def add_data(self, new_data):
-        assert len(new_data) == self._data_size
-
-        if len(self._data_queue) > 0 and np.array_equal(
-            new_data, self._data_queue[-1]
-        ):  # skip duplicate data
-            return
-
-        if len(self._data_queue) >= self._window_size:
-            self._data_queue.pop(0)
-
-        self._data_queue.append(new_data)
-        self._filtered_data = self._apply_filter()
-
-    @property
-    def filtered_data(self):
-        return self._filtered_data
-
-
-class G1_29_ArmIK:  # noqa: N801
-    def __init__(self, unit_test=False):
-        import casadi
-        import pinocchio as pin
-        from huggingface_hub import snapshot_download
-        from pinocchio import casadi as cpin
-
-        self._pin = pin
-        np.set_printoptions(precision=5, suppress=True, linewidth=200)
-
-        self.unit_test = unit_test
-
-        self.repo_path = snapshot_download("lerobot/unitree-g1-mujoco")
-        urdf_path = os.path.join(self.repo_path, "assets", "g1_body29_hand14.urdf")
-        mesh_dir = os.path.join(self.repo_path, "assets")
-
-        self.robot = self._pin.RobotWrapper.BuildFromURDF(urdf_path, mesh_dir)
-
-        self.mixed_jointsToLockIDs = [
-            "left_hip_pitch_joint",
-            "left_hip_roll_joint",
-            "left_hip_yaw_joint",
-            "left_knee_joint",
-            "left_ankle_pitch_joint",
-            "left_ankle_roll_joint",
-            "right_hip_pitch_joint",
-            "right_hip_roll_joint",
-            "right_hip_yaw_joint",
-            "right_knee_joint",
-            "right_ankle_pitch_joint",
-            "right_ankle_roll_joint",
-            "waist_yaw_joint",
-            "waist_roll_joint",
-            "waist_pitch_joint",
-            "left_hand_thumb_0_joint",
-            "left_hand_thumb_1_joint",
-            "left_hand_thumb_2_joint",
-            "left_hand_middle_0_joint",
-            "left_hand_middle_1_joint",
-            "left_hand_index_0_joint",
-            "left_hand_index_1_joint",
-            "right_hand_thumb_0_joint",
-            "right_hand_thumb_1_joint",
-            "right_hand_thumb_2_joint",
-            "right_hand_index_0_joint",
-            "right_hand_index_1_joint",
-            "right_hand_middle_0_joint",
-            "right_hand_middle_1_joint",
-        ]
-
-        self.reduced_robot = self.robot.buildReducedRobot(
-            list_of_joints_to_lock=self.mixed_jointsToLockIDs,
-            reference_configuration=np.array([0.0] * self.robot.model.nq),
-        )
-
-        # Arm joint names in G1 motor order (G1_29_JointArmIndex)
-        self._arm_joint_names_g1 = [
-            "left_shoulder_pitch_joint",
-            "left_shoulder_roll_joint",
-            "left_shoulder_yaw_joint",
-            "left_elbow_joint",
-            "left_wrist_roll_joint",
-            "left_wrist_pitch_joint",
-            "left_wrist_yaw_joint",
-            "right_shoulder_pitch_joint",
-            "right_shoulder_roll_joint",
-            "right_shoulder_yaw_joint",
-            "right_elbow_joint",
-            "right_wrist_roll_joint",
-            "right_wrist_pitch_joint",
-            "right_wrist_yaw_joint",
-        ]
-        # Pinocchio uses its own joint order in q; build index mapping.
-        self._arm_joint_names_pin = sorted(
-            self._arm_joint_names_g1,
-            key=lambda name: self.reduced_robot.model.idx_qs[self.reduced_robot.model.getJointId(name)],
-        )
-        logger.info(f"Pinocchio arm joint order: {self._arm_joint_names_pin}")
-        self._arm_reorder_g1_to_pin = [
-            self._arm_joint_names_g1.index(name) for name in self._arm_joint_names_pin
-        ]
-        # Inverse mapping to return tau in G1 motor order.
-        self._arm_reorder_pin_to_g1 = np.argsort(self._arm_reorder_g1_to_pin)
-
-        self.reduced_robot.model.addFrame(
-            self._pin.Frame(
-                "L_ee",
-                self.reduced_robot.model.getJointId("left_wrist_yaw_joint"),
-                self._pin.SE3(np.eye(3), np.array([0.05, 0, 0]).T),
-                self._pin.FrameType.OP_FRAME,
-            )
-        )
-
-        self.reduced_robot.model.addFrame(
-            self._pin.Frame(
-                "R_ee",
-                self.reduced_robot.model.getJointId("right_wrist_yaw_joint"),
-                self._pin.SE3(np.eye(3), np.array([0.05, 0, 0]).T),
-                self._pin.FrameType.OP_FRAME,
-            )
-        )
-
-        # Creating Casadi models and data for symbolic computing
-        self.cmodel = cpin.Model(self.reduced_robot.model)
-        self.cdata = self.cmodel.createData()
-
-        # Creating symbolic variables
-        self.cq = casadi.SX.sym("q", self.reduced_robot.model.nq, 1)
-        self.cTf_l = casadi.SX.sym("tf_l", 4, 4)
-        self.cTf_r = casadi.SX.sym("tf_r", 4, 4)
-        cpin.framesForwardKinematics(self.cmodel, self.cdata, self.cq)
-
-        # Get the hand joint ID and define the error function
-        self.L_hand_id = self.reduced_robot.model.getFrameId("L_ee")
-        self.R_hand_id = self.reduced_robot.model.getFrameId("R_ee")
-
-        self.translational_error = casadi.Function(
-            "translational_error",
-            [self.cq, self.cTf_l, self.cTf_r],
-            [
-                casadi.vertcat(
-                    self.cdata.oMf[self.L_hand_id].translation - self.cTf_l[:3, 3],
-                    self.cdata.oMf[self.R_hand_id].translation - self.cTf_r[:3, 3],
-                )
-            ],
-        )
-        self.rotational_error = casadi.Function(
-            "rotational_error",
-            [self.cq, self.cTf_l, self.cTf_r],
-            [
-                casadi.vertcat(
-                    cpin.log3(self.cdata.oMf[self.L_hand_id].rotation @ self.cTf_l[:3, :3].T),
-                    cpin.log3(self.cdata.oMf[self.R_hand_id].rotation @ self.cTf_r[:3, :3].T),
-                )
-            ],
-        )
-
-        # Defining the optimization problem
-        self.opti = casadi.Opti()
-        self.var_q = self.opti.variable(self.reduced_robot.model.nq)
-        self.var_q_last = self.opti.parameter(self.reduced_robot.model.nq)  # for smooth
-        self.param_tf_l = self.opti.parameter(4, 4)
-        self.param_tf_r = self.opti.parameter(4, 4)
-        self.translational_cost = casadi.sumsqr(
-            self.translational_error(self.var_q, self.param_tf_l, self.param_tf_r)
-        )
-        self.rotation_cost = casadi.sumsqr(
-            self.rotational_error(self.var_q, self.param_tf_l, self.param_tf_r)
-        )
-        self.regularization_cost = casadi.sumsqr(self.var_q)
-        self.smooth_cost = casadi.sumsqr(self.var_q - self.var_q_last)
-
-        # Setting optimization constraints and goals
-        self.opti.subject_to(
-            self.opti.bounded(
-                self.reduced_robot.model.lowerPositionLimit,
-                self.var_q,
-                self.reduced_robot.model.upperPositionLimit,
-            )
-        )
-        self.opti.minimize(
-            50 * self.translational_cost
-            + self.rotation_cost
-            + 0.02 * self.regularization_cost
-            + 0.1 * self.smooth_cost
-        )
-
-        opts = {
-            "ipopt": {"print_level": 0, "max_iter": 50, "tol": 1e-6},
-            "print_time": False,  # print or not
-            "calc_lam_p": False,  # https://github.com/casadi/casadi/wiki/FAQ:-Why-am-I-getting-%22NaN-detected%22in-my-optimization%3F
-        }
-        self.opti.solver("ipopt", opts)
-
-        self.init_data = np.zeros(self.reduced_robot.model.nq)
-        self.smooth_filter = WeightedMovingFilter(np.array([0.4, 0.3, 0.2, 0.1]), 14)
-
-    def solve_ik(self, left_wrist, right_wrist, current_lr_arm_motor_q=None, current_lr_arm_motor_dq=None):
-        if current_lr_arm_motor_q is not None:
-            self.init_data = current_lr_arm_motor_q
-        self.opti.set_initial(self.var_q, self.init_data)
-
-        self.opti.set_value(self.param_tf_l, left_wrist)
-        self.opti.set_value(self.param_tf_r, right_wrist)
-        self.opti.set_value(self.var_q_last, self.init_data)  # for smooth
-
-        try:
-            self.opti.solve()
-
-            sol_q = self.opti.value(self.var_q)
-            self.smooth_filter.add_data(sol_q)
-            sol_q = self.smooth_filter.filtered_data
-
-            if current_lr_arm_motor_dq is not None:
-                v = current_lr_arm_motor_dq * 0.0
-            else:
-                v = (sol_q - self.init_data) * 0.0
-
-            self.init_data = sol_q
-
-            sol_tauff = self._pin.rnea(
-                self.reduced_robot.model,
-                self.reduced_robot.data,
-                sol_q,
-                v,
-                np.zeros(self.reduced_robot.model.nv),
-            )
-
-            return sol_q, sol_tauff
-
-        except Exception as e:
-            logger.error(f"ERROR in convergence, plotting debug info.{e}")
-
-            sol_q = self.opti.debug.value(self.var_q)
-            self.smooth_filter.add_data(sol_q)
-            sol_q = self.smooth_filter.filtered_data
-
-            if current_lr_arm_motor_dq is not None:
-                v = current_lr_arm_motor_dq * 0.0
-            else:
-                v = (sol_q - self.init_data) * 0.0
-
-            self.init_data = sol_q
-
-            logger.error(
-                f"sol_q:{sol_q} \nmotorstate: \n{current_lr_arm_motor_q} \nleft_pose: \n{left_wrist} \nright_pose: \n{right_wrist}"
-            )
-
-            return current_lr_arm_motor_q, np.zeros(self.reduced_robot.model.nv)
-
-    def solve_tau(self, current_lr_arm_motor_q=None, current_lr_arm_motor_dq=None):
-        try:
-            q_g1 = np.array(current_lr_arm_motor_q, dtype=float)
-            if q_g1.shape[0] != len(self._arm_joint_names_g1):
-                raise ValueError(f"Expected {len(self._arm_joint_names_g1)} arm joints, got {q_g1.shape[0]}")
-            q_pin = q_g1[self._arm_reorder_g1_to_pin]
-            sol_tauff = self._pin.rnea(
-                self.reduced_robot.model,
-                self.reduced_robot.data,
-                q_pin,
-                np.zeros(self.reduced_robot.model.nv),
-                np.zeros(self.reduced_robot.model.nv),
-            )
-            return sol_tauff[self._arm_reorder_pin_to_g1]
-
-        except Exception as e:
-            logger.error(f"ERROR in convergence, plotting debug info.{e}")
-            return np.zeros(self.reduced_robot.model.nv)

src/lerobot/robots/unitree_g1/unitree_g1.py

@@ -27,8 +27,7 @@ import numpy as np
 from lerobot.cameras.utils import make_cameras_from_configs
 from lerobot.envs.factory import make_env
 from lerobot.processor import RobotAction, RobotObservation
-from lerobot.robots.unitree_g1.g1_utils import G1_29_JointArmIndex, G1_29_JointIndex
-from lerobot.robots.unitree_g1.robot_kinematic_processor import G1_29_ArmIK
+from lerobot.robots.unitree_g1.g1_utils import G1_29_JointIndex
 
 from ..robot import Robot
 from .config_unitree_g1 import UnitreeG1Config
@@ -128,8 +127,6 @@ class UnitreeG1(Robot):
         self.subscribe_thread = None
         self.remote_controller = self.RemoteController()
 
-        self.arm_ik = G1_29_ArmIK()
-
     def _subscribe_motor_state(self):  # polls robot state @ 250Hz
         while not self._shutdown_event.is_set():
             start_time = time.time()
@@ -364,20 +361,6 @@ class UnitreeG1(Robot):
                 self.msg.motor_cmd[motor.value].kd = self.kd[motor.value]
                 self.msg.motor_cmd[motor.value].tau = 0
 
-        if self.config.gravity_compensation:
-            # Build action_np from motor commands (arm joints are indices 15-28, local indices 0-13)
-            action_np = np.zeros(14)
-            arm_start_idx = G1_29_JointArmIndex.kLeftShoulderPitch.value  # 15
-            for joint in G1_29_JointArmIndex:
-                local_idx = joint.value - arm_start_idx
-                action_np[local_idx] = self.msg.motor_cmd[joint.value].q
-            tau = self.arm_ik.solve_tau(action_np)
-
-            # Apply tau back to motor commands
-            for joint in G1_29_JointArmIndex:
-                local_idx = joint.value - arm_start_idx
-                self.msg.motor_cmd[joint.value].tau = tau[local_idx]
-
         self.msg.crc = self.crc.Crc(self.msg)
         self.lowcmd_publisher.Write(self.msg)
         return action

src/lerobot/robots/utils.py

@@ -60,14 +60,6 @@ def make_robot_from_config(config: RobotConfig) -> Robot:
         from .reachy2 import Reachy2Robot
 
         return Reachy2Robot(config)
-    elif config.type == "openarm_follower":
-        from .openarm_follower import OpenArmFollower
-
-        return OpenArmFollower(config)
-    elif config.type == "bi_openarm_follower":
-        from .bi_openarm_follower import BiOpenArmFollower
-
-        return BiOpenArmFollower(config)
     elif config.type == "mock_robot":
         from tests.mocks.mock_robot import MockRobot
 

src/lerobot/scripts/lerobot_calibrate.py

@@ -36,28 +36,23 @@ from lerobot.cameras.realsense.configuration_realsense import RealSenseCameraCon
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
-    bi_openarm_follower,
     bi_so_follower,
     hope_jr,
     koch_follower,
     lekiwi,
     make_robot_from_config,
     omx_follower,
-    openarm_follower,
     so_follower,
 )
 from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
-    bi_openarm_leader,
     bi_so_leader,
     homunculus,
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
-    openarm_leader,
     so_leader,
-    unitree_g1,
 )
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.utils import init_logging
@@ -86,11 +81,8 @@ def calibrate(cfg: CalibrateConfig):
         device = make_teleoperator_from_config(cfg.device)
 
     device.connect(calibrate=False)
-
-    try:
-        device.calibrate()
-    finally:
-        device.disconnect()
+    device.calibrate()
+    device.disconnect()
 
 
 def main():

src/lerobot/scripts/lerobot_edit_dataset.py

@@ -18,7 +18,7 @@
 Edit LeRobot datasets using various transformation tools.
 
 This script allows you to delete episodes, split datasets, merge datasets,
-remove features, modify tasks, and convert image datasets to video format.
+remove features, and convert image datasets to video format.
 When new_repo_id is specified, creates a new dataset.
 
 Usage Examples:
@@ -66,25 +66,6 @@ Remove camera feature:
         --operation.type remove_feature \
         --operation.feature_names "['observation.images.top']"
 
-Modify tasks - set a single task for all episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
-        --repo_id lerobot/pusht \
-        --operation.type modify_tasks \
-        --operation.new_task "Pick up the cube and place it"
-
-Modify tasks - set different tasks for specific episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
-        --repo_id lerobot/pusht \
-        --operation.type modify_tasks \
-        --operation.episode_tasks '{"0": "Task A", "1": "Task B", "2": "Task A"}'
-
-Modify tasks - set default task with overrides for specific episodes (WARNING: modifies in-place):
-    python -m lerobot.scripts.lerobot_edit_dataset \
-        --repo_id lerobot/pusht \
-        --operation.type modify_tasks \
-        --operation.new_task "Default task" \
-        --operation.episode_tasks '{"5": "Special task for episode 5"}'
-
 Convert image dataset to video format and save locally:
     python -m lerobot.scripts.lerobot_edit_dataset \
         --repo_id lerobot/pusht_image \
@@ -119,7 +100,6 @@ from lerobot.datasets.dataset_tools import (
     convert_image_to_video_dataset,
     delete_episodes,
     merge_datasets,
-    modify_tasks,
     remove_feature,
     split_dataset,
 )
@@ -152,13 +132,6 @@ class RemoveFeatureConfig:
     feature_names: list[str] | None = None
 
 
-@dataclass
-class ModifyTasksConfig:
-    type: str = "modify_tasks"
-    new_task: str | None = None
-    episode_tasks: dict[str, str] | None = None
-
-
 @dataclass
 class ConvertImageToVideoConfig:
     type: str = "convert_image_to_video"
@@ -178,12 +151,7 @@ class ConvertImageToVideoConfig:
 class EditDatasetConfig:
     repo_id: str
     operation: (
-        DeleteEpisodesConfig
-        | SplitConfig
-        | MergeConfig
-        | RemoveFeatureConfig
-        | ModifyTasksConfig
-        | ConvertImageToVideoConfig
+        DeleteEpisodesConfig | SplitConfig | MergeConfig | RemoveFeatureConfig | ConvertImageToVideoConfig
     )
     root: str | None = None
     new_repo_id: str | None = None
@@ -328,48 +296,6 @@ def handle_remove_feature(cfg: EditDatasetConfig) -> None:
         LeRobotDataset(output_repo_id, root=output_dir).push_to_hub()
 
 
-def handle_modify_tasks(cfg: EditDatasetConfig) -> None:
-    if not isinstance(cfg.operation, ModifyTasksConfig):
-        raise ValueError("Operation config must be ModifyTasksConfig")
-
-    new_task = cfg.operation.new_task
-    episode_tasks_raw = cfg.operation.episode_tasks
-
-    if new_task is None and episode_tasks_raw is None:
-        raise ValueError("Must specify at least one of new_task or episode_tasks for modify_tasks operation")
-
-    # Warn about in-place modification behavior
-    if cfg.new_repo_id is not None:
-        logging.warning("modify_tasks modifies datasets in-place. The --new_repo_id parameter is ignored.")
-
-    dataset = LeRobotDataset(cfg.repo_id, root=cfg.root)
-    logging.warning(f"Modifying dataset in-place at {dataset.root}. Original data will be overwritten.")
-
-    # Convert episode_tasks keys from string to int if needed (CLI passes strings)
-    episode_tasks: dict[int, str] | None = None
-    if episode_tasks_raw is not None:
-        episode_tasks = {int(k): v for k, v in episode_tasks_raw.items()}
-
-    logging.info(f"Modifying tasks in {cfg.repo_id}")
-    if new_task:
-        logging.info(f"  Default task: '{new_task}'")
-    if episode_tasks:
-        logging.info(f"  Episode-specific tasks: {episode_tasks}")
-
-    modified_dataset = modify_tasks(
-        dataset,
-        new_task=new_task,
-        episode_tasks=episode_tasks,
-    )
-
-    logging.info(f"Dataset modified at {dataset.root}")
-    logging.info(f"Tasks: {list(modified_dataset.meta.tasks.index)}")
-
-    if cfg.push_to_hub:
-        logging.info(f"Pushing to hub as {cfg.repo_id}")
-        modified_dataset.push_to_hub()
-
-
 def handle_convert_image_to_video(cfg: EditDatasetConfig) -> None:
     # Note: Parser may create any config type with the right fields, so we access fields directly
     # instead of checking isinstance()
@@ -445,14 +371,12 @@ def edit_dataset(cfg: EditDatasetConfig) -> None:
         handle_merge(cfg)
     elif operation_type == "remove_feature":
         handle_remove_feature(cfg)
-    elif operation_type == "modify_tasks":
-        handle_modify_tasks(cfg)
     elif operation_type == "convert_image_to_video":
         handle_convert_image_to_video(cfg)
     else:
         raise ValueError(
             f"Unknown operation type: {operation_type}\n"
-            f"Available operations: delete_episodes, split, merge, remove_feature, modify_tasks, convert_image_to_video"
+            f"Available operations: delete_episodes, split, merge, remove_feature, convert_to_video"
         )
 
 

src/lerobot/scripts/lerobot_find_joint_limits.py

@@ -44,23 +44,19 @@ import numpy as np
 from lerobot.model.kinematics import RobotKinematics
 from lerobot.robots import (  # noqa: F401
     RobotConfig,
-    bi_openarm_follower,
     bi_so_follower,
     koch_follower,
     make_robot_from_config,
     omx_follower,
-    openarm_follower,
     so_follower,
 )
 from lerobot.teleoperators import (  # noqa: F401
     TeleoperatorConfig,
-    bi_openarm_leader,
     bi_so_leader,
     gamepad,
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
-    openarm_leader,
     so_leader,
 )
 from lerobot.utils.robot_utils import precise_sleep

src/lerobot/scripts/lerobot_record.py

@@ -98,31 +98,26 @@ from lerobot.processor.rename_processor import rename_stats
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
-    bi_openarm_follower,
     bi_so_follower,
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
     make_robot_from_config,
     omx_follower,
-    openarm_follower,
     reachy2,
     so_follower,
-    unitree_g1 as unitree_g1_robot,
+    unitree_g1,
 )
 from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
-    bi_openarm_leader,
     bi_so_leader,
     homunculus,
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
-    openarm_leader,
     reachy2_teleoperator,
     so_leader,
-    unitree_g1,
 )
 from lerobot.teleoperators.keyboard.teleop_keyboard import KeyboardTeleop
 from lerobot.utils.constants import ACTION, OBS_STR

src/lerobot/scripts/lerobot_replay.py

@@ -53,14 +53,12 @@ from lerobot.processor import (
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
-    bi_openarm_follower,
     bi_so_follower,
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
     make_robot_from_config,
     omx_follower,
-    openarm_follower,
     reachy2,
     so_follower,
     unitree_g1,
@@ -110,26 +108,25 @@ def replay(cfg: ReplayConfig):
 
     robot.connect()
 
-    try:
-        log_say("Replaying episode", cfg.play_sounds, blocking=True)
-        for idx in range(len(episode_frames)):
-            start_episode_t = time.perf_counter()
+    log_say("Replaying episode", cfg.play_sounds, blocking=True)
+    for idx in range(len(episode_frames)):
+        start_episode_t = time.perf_counter()
 
-            action_array = actions[idx][ACTION]
-            action = {}
-            for i, name in enumerate(dataset.features[ACTION]["names"]):
-                action[name] = action_array[i]
+        action_array = actions[idx][ACTION]
+        action = {}
+        for i, name in enumerate(dataset.features[ACTION]["names"]):
+            action[name] = action_array[i]
 
-            robot_obs = robot.get_observation()
+        robot_obs = robot.get_observation()
 
-            processed_action = robot_action_processor((action, robot_obs))
+        processed_action = robot_action_processor((action, robot_obs))
 
-            _ = robot.send_action(processed_action)
+        _ = robot.send_action(processed_action)
 
-            dt_s = time.perf_counter() - start_episode_t
-            precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
-    finally:
-        robot.disconnect()
+        dt_s = time.perf_counter() - start_episode_t
+        precise_sleep(max(1 / dataset.fps - dt_s, 0.0))
+
+    robot.disconnect()
 
 
 def main():

src/lerobot/scripts/lerobot_setup_can.py

@@ -1,360 +0,0 @@
-# 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.
-
-"""
-Setup and debug CAN interfaces for Damiao motors (e.g., OpenArms).
-
-Examples:
-
-Setup CAN interfaces with CAN FD:
-```shell
-lerobot-setup-can --mode=setup --interfaces=can0,can1,can2,can3
-```
-
-Test motors on a single interface:
-```shell
-lerobot-setup-can --mode=test --interfaces=can0
-```
-
-Test motors on all interfaces:
-```shell
-lerobot-setup-can --mode=test --interfaces=can0,can1,can2,can3
-```
-
-Speed test:
-```shell
-lerobot-setup-can --mode=speed --interfaces=can0
-```
-"""
-
-import subprocess
-import sys
-import time
-from dataclasses import dataclass, field
-
-import draccus
-
-from lerobot.utils.import_utils import _can_available
-
-MOTOR_NAMES = {
-    0x01: "joint_1",
-    0x02: "joint_2",
-    0x03: "joint_3",
-    0x04: "joint_4",
-    0x05: "joint_5",
-    0x06: "joint_6",
-    0x07: "joint_7",
-    0x08: "gripper",
-}
-
-
-@dataclass
-class CANSetupConfig:
-    mode: str = "test"
-    interfaces: str = "can0"  # Comma-separated, e.g. "can0,can1,can2,can3"
-    bitrate: int = 1000000
-    data_bitrate: int = 5000000
-    use_fd: bool = True
-    motor_ids: list[int] = field(default_factory=lambda: list(range(0x01, 0x09)))
-    timeout: float = 1.0
-    speed_iterations: int = 100
-
-    def get_interfaces(self) -> list[str]:
-        return [i.strip() for i in self.interfaces.split(",") if i.strip()]
-
-
-def check_interface_status(interface: str) -> tuple[bool, str, bool]:
-    """Check if CAN interface is UP and configured."""
-    try:
-        result = subprocess.run(["ip", "link", "show", interface], capture_output=True, text=True)  # nosec B607
-        if result.returncode != 0:
-            return False, "Interface not found", False
-
-        output = result.stdout
-        is_up = "UP" in output
-        is_fd = "fd on" in output.lower() or "canfd" in output.lower()
-        status = "UP" if is_up else "DOWN"
-        if is_fd:
-            status += " (CAN FD)"
-
-        return is_up, status, is_fd
-    except FileNotFoundError:
-        return False, "ip command not found", False
-
-
-def setup_interface(interface: str, bitrate: int, data_bitrate: int, use_fd: bool) -> bool:
-    """Configure a CAN interface."""
-    try:
-        subprocess.run(["sudo", "ip", "link", "set", interface, "down"], check=False, capture_output=True)  # nosec B607
-
-        cmd = ["sudo", "ip", "link", "set", interface, "type", "can", "bitrate", str(bitrate)]
-        if use_fd:
-            cmd.extend(["dbitrate", str(data_bitrate), "fd", "on"])
-
-        result = subprocess.run(cmd, capture_output=True, text=True)  # nosec B607
-        if result.returncode != 0:
-            print(f"  ✗ Failed to configure: {result.stderr}")
-            return False
-
-        result = subprocess.run(  # nosec B607
-            ["sudo", "ip", "link", "set", interface, "up"], capture_output=True, text=True
-        )
-        if result.returncode != 0:
-            print(f"  ✗ Failed to bring up: {result.stderr}")
-            return False
-
-        return True
-    except Exception as e:
-        print(f"  ✗ Error: {e}")
-        return False
-
-
-def test_motor(bus, motor_id: int, timeout: float, use_fd: bool):
-    """Test a single motor and return responses."""
-    import can
-
-    enable_msg = can.Message(
-        arbitration_id=motor_id,
-        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
-        is_extended_id=False,
-        is_fd=use_fd,
-    )
-
-    try:
-        bus.send(enable_msg)
-    except Exception as e:
-        return None, f"Send error: {e}"
-
-    responses = []
-    start_time = time.time()
-
-    while time.time() - start_time < timeout:
-        msg = bus.recv(timeout=0.1)
-        if msg:
-            responses.append((msg.arbitration_id, msg.data.hex(), getattr(msg, "is_fd", False)))
-
-    disable_msg = can.Message(
-        arbitration_id=motor_id,
-        data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFD],
-        is_extended_id=False,
-        is_fd=use_fd,
-    )
-    try:
-        bus.send(disable_msg)
-    except Exception:
-        print(f"Error sending message to motor 0x{motor_id:02X}")
-
-    return responses, None
-
-
-def test_interface(cfg: CANSetupConfig, interface: str):
-    """Test all motors on a CAN interface."""
-    import can
-
-    is_up, status, _ = check_interface_status(interface)
-    print(f"\n{interface}: {status}")
-
-    if not is_up:
-        print(f"  ⚠ Interface is not UP. Run: lerobot-setup-can --mode=setup --interfaces {interface}")
-        return {}
-
-    try:
-        kwargs = {"channel": interface, "interface": "socketcan", "bitrate": cfg.bitrate}
-        if cfg.use_fd:
-            kwargs.update({"data_bitrate": cfg.data_bitrate, "fd": True})
-        bus = can.interface.Bus(**kwargs)
-    except Exception as e:
-        print(f"  ✗ Connection failed: {e}")
-        return {}
-
-    results = {}
-    try:
-        while bus.recv(timeout=0.01):
-            pass
-
-        for motor_id in cfg.motor_ids:
-            motor_name = MOTOR_NAMES.get(motor_id, f"motor_0x{motor_id:02X}")
-            responses, error = test_motor(bus, motor_id, cfg.timeout, cfg.use_fd)
-
-            if error:
-                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ {error}")
-                results[motor_id] = {"found": False, "error": error}
-            elif responses:
-                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✓ FOUND")
-                for resp_id, data, is_fd in responses:
-                    fd_flag = " [FD]" if is_fd else ""
-                    print(f"    → Response 0x{resp_id:02X}{fd_flag}: {data}")
-                results[motor_id] = {"found": True, "responses": responses}
-            else:
-                print(f"  Motor 0x{motor_id:02X} ({motor_name}): ✗ No response")
-                results[motor_id] = {"found": False}
-
-            time.sleep(0.05)
-    finally:
-        bus.shutdown()
-
-    found = sum(1 for r in results.values() if r.get("found"))
-    print(f"\n  Summary: {found}/{len(cfg.motor_ids)} motors found")
-    return results
-
-
-def speed_test(cfg: CANSetupConfig, interface: str):
-    """Test communication speed with motors."""
-    import can
-
-    is_up, status, _ = check_interface_status(interface)
-    if not is_up:
-        print(f"{interface}: {status} - skipping")
-        return
-
-    print(f"\n{interface}: Running speed test ({cfg.speed_iterations} iterations)...")
-
-    try:
-        kwargs = {"channel": interface, "interface": "socketcan", "bitrate": cfg.bitrate}
-        if cfg.use_fd:
-            kwargs.update({"data_bitrate": cfg.data_bitrate, "fd": True})
-        bus = can.interface.Bus(**kwargs)
-    except Exception as e:
-        print(f"  ✗ Connection failed: {e}")
-        return
-
-    responding_motor = None
-    for motor_id in cfg.motor_ids:
-        responses, _ = test_motor(bus, motor_id, 0.5, cfg.use_fd)
-        if responses:
-            responding_motor = motor_id
-            break
-
-    if not responding_motor:
-        print("  ✗ No responding motors found")
-        bus.shutdown()
-        return
-
-    print(f"  Testing with motor 0x{responding_motor:02X}...")
-    latencies = []
-
-    for _ in range(cfg.speed_iterations):
-        start = time.perf_counter()
-        msg = can.Message(
-            arbitration_id=responding_motor,
-            data=[0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFC],
-            is_extended_id=False,
-            is_fd=cfg.use_fd,
-        )
-        bus.send(msg)
-        resp = bus.recv(timeout=0.1)
-        if resp:
-            latencies.append((time.perf_counter() - start) * 1000)
-
-    bus.shutdown()
-
-    if latencies:
-        avg_latency = sum(latencies) / len(latencies)
-        hz = 1000.0 / avg_latency if avg_latency > 0 else 0
-        print(f"  ✓ Success rate: {len(latencies)}/{cfg.speed_iterations}")
-        print(f"  ✓ Avg latency: {avg_latency:.2f} ms")
-        print(f"  ✓ Max frequency: {hz:.1f} Hz")
-    else:
-        print("  ✗ No successful responses")
-
-
-def run_setup(cfg: CANSetupConfig):
-    """Setup CAN interfaces."""
-    print("=" * 50)
-    print("CAN Interface Setup")
-    print("=" * 50)
-    print(f"Mode: {'CAN FD' if cfg.use_fd else 'CAN 2.0'}")
-    print(f"Bitrate: {cfg.bitrate / 1_000_000:.1f} Mbps")
-    if cfg.use_fd:
-        print(f"Data bitrate: {cfg.data_bitrate / 1_000_000:.1f} Mbps")
-    print()
-
-    interfaces = cfg.get_interfaces()
-    for interface in interfaces:
-        print(f"Configuring {interface}...")
-        if setup_interface(interface, cfg.bitrate, cfg.data_bitrate, cfg.use_fd):
-            is_up, status, _ = check_interface_status(interface)
-            print(f"  ✓ {interface}: {status}")
-        else:
-            print(f"  ✗ {interface}: Failed")
-
-    print("\nSetup complete!")
-    print("\nNext: Test motors with:")
-    print(f"  lerobot-setup-can --mode=test --interfaces {','.join(interfaces)}")
-
-
-def run_test(cfg: CANSetupConfig):
-    """Test motors on CAN interfaces."""
-    print("=" * 50)
-    print("CAN Motor Test")
-    print("=" * 50)
-    print(f"Testing motors 0x{min(cfg.motor_ids):02X}-0x{max(cfg.motor_ids):02X}")
-    print(f"Mode: {'CAN FD' if cfg.use_fd else 'CAN 2.0'}")
-    print()
-
-    interfaces = cfg.get_interfaces()
-    all_results = {}
-    for interface in interfaces:
-        all_results[interface] = test_interface(cfg, interface)
-
-    total_found = sum(sum(1 for r in res.values() if r.get("found")) for res in all_results.values())
-
-    print("\n" + "=" * 50)
-    print("Summary")
-    print("=" * 50)
-    print(f"Total motors found: {total_found}")
-
-    if total_found == 0:
-        print("\n⚠ No motors found! Check:")
-        print("  1. Motors are powered (24V)")
-        print("  2. CAN wiring (CANH, CANL, GND)")
-        print("  3. Motor timeout parameter > 0 (use Damiao tools)")
-        print("  4. 120Ω termination at both cable ends")
-        print(f"  5. Interface configured: lerobot-setup-can --mode=setup --interfaces {interfaces[0]}")
-
-
-def run_speed(cfg: CANSetupConfig):
-    """Run speed tests on CAN interfaces."""
-    print("=" * 50)
-    print("CAN Speed Test")
-    print("=" * 50)
-
-    for interface in cfg.get_interfaces():
-        speed_test(cfg, interface)
-
-
-@draccus.wrap()
-def setup_can(cfg: CANSetupConfig):
-    if not _can_available:
-        print("Error: python-can not installed. Install with: pip install python-can")
-        sys.exit(1)
-
-    if cfg.mode == "setup":
-        run_setup(cfg)
-    elif cfg.mode == "test":
-        run_test(cfg)
-    elif cfg.mode == "speed":
-        run_speed(cfg)
-    else:
-        print(f"Unknown mode: {cfg.mode}")
-        print("Available modes: setup, test, speed")
-        sys.exit(1)
-
-
-def main():
-    setup_can()
-
-
-if __name__ == "__main__":
-    main()

src/lerobot/scripts/lerobot_teleoperate.py

@@ -70,22 +70,18 @@ from lerobot.processor import (
 from lerobot.robots import (  # noqa: F401
     Robot,
     RobotConfig,
-    bi_openarm_follower,
     bi_so_follower,
     earthrover_mini_plus,
     hope_jr,
     koch_follower,
     make_robot_from_config,
     omx_follower,
-    openarm_follower,
     reachy2,
     so_follower,
-    unitree_g1 as unitree_g1_robot,
 )
 from lerobot.teleoperators import (  # noqa: F401
     Teleoperator,
     TeleoperatorConfig,
-    bi_openarm_leader,
     bi_so_leader,
     gamepad,
     homunculus,
@@ -93,10 +89,8 @@ from lerobot.teleoperators import (  # noqa: F401
     koch_leader,
     make_teleoperator_from_config,
     omx_leader,
-    openarm_leader,
     reachy2_teleoperator,
     so_leader,
-    unitree_g1,
 )
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.robot_utils import precise_sleep

src/lerobot/scripts/lerobot_train.py

@@ -63,8 +63,8 @@ def update_policy(
     accelerator: Accelerator,
     lr_scheduler=None,
     lock=None,
-    rabc_weights_provider=None,
-) -> tuple[MetricsTracker, dict]:
+    sample_weighter=None,
+) -> tuple[MetricsTracker, dict | None]:
     """
     Performs a single training step to update the policy's weights.
 
@@ -80,7 +80,7 @@ def update_policy(
         accelerator: The Accelerator instance for distributed training and mixed precision.
         lr_scheduler: An optional learning rate scheduler.
         lock: An optional lock for thread-safe optimizer updates.
-        rabc_weights_provider: Optional RABCWeights instance for sample weighting.
+        sample_weighter: Optional SampleWeighter instance for per-sample loss weighting.
 
     Returns:
         A tuple containing:
@@ -90,27 +90,31 @@ def update_policy(
     start_time = time.perf_counter()
     policy.train()
 
-    # Get RA-BC weights if enabled
-    rabc_batch_weights = None
-    rabc_batch_stats = None
-    if rabc_weights_provider is not None:
-        rabc_batch_weights, rabc_batch_stats = rabc_weights_provider.compute_batch_weights(batch)
+    # Compute sample weights if a weighter is provided
+    sample_weights = None
+    weight_stats = None
+    if sample_weighter is not None:
+        sample_weights, weight_stats = sample_weighter.compute_batch_weights(batch)
 
     # Let accelerator handle mixed precision
     with accelerator.autocast():
-        # Use per-sample loss when RA-BC is enabled for proper weighting
-        if rabc_batch_weights is not None:
-            # Get per-sample losses
+        if sample_weights is not None:
+            # Use per-sample loss for weighted training
+            # Note: Policies supporting sample weighting must implement forward(batch, reduction="none")
             per_sample_loss, output_dict = policy.forward(batch, reduction="none")
 
-            # Apply RA-BC weights: L_RA-BC = Σ(w_i * l_i) / (Σw_i + ε)
-            # rabc_batch_weights is already normalized to sum to batch_size
+            # Weighted loss: each sample's contribution is scaled by its weight.
+            # We divide by weight sum (not batch size) so that if some weights are zero,
+            # the remaining samples contribute proportionally more, preserving gradient scale.
+            # Weights are pre-normalized to sum to batch_size for stable training dynamics.
             epsilon = 1e-6
-            loss = (per_sample_loss * rabc_batch_weights).sum() / (rabc_batch_weights.sum() + epsilon)
-            # Log raw mean weight (before normalization) - this is the meaningful metric
-            output_dict["rabc_mean_weight"] = rabc_batch_stats["raw_mean_weight"]
-            output_dict["rabc_num_zero_weight"] = rabc_batch_stats["num_zero_weight"]
-            output_dict["rabc_num_full_weight"] = rabc_batch_stats["num_full_weight"]
+            loss = (per_sample_loss * sample_weights).sum() / (sample_weights.sum() + epsilon)
+
+            # Log weighting statistics
+            if output_dict is None:
+                output_dict = {}
+            for key, value in weight_stats.items():
+                output_dict[f"sample_weight_{key}"] = value
         else:
             loss, output_dict = policy.forward(batch)
 
@@ -288,27 +292,19 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         logging.info("Creating optimizer and scheduler")
     optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
 
-    # Load precomputed SARM progress for RA-BC if enabled
-    # Generate progress using: src/lerobot/policies/sarm/compute_rabc_weights.py
-    rabc_weights = None
-    if cfg.use_rabc:
-        from lerobot.utils.rabc import RABCWeights
+    # Create sample weighter if configured (e.g., for RA-BC training)
+    sample_weighter = None
+    if cfg.sample_weighting is not None:
+        from lerobot.utils.sample_weighting import make_sample_weighter
 
-        # Get chunk_size from policy config
-        chunk_size = getattr(policy.config, "chunk_size", None)
-        if chunk_size is None:
-            raise ValueError("Chunk size is not found in policy config")
-
-        head_mode = getattr(cfg, "rabc_head_mode", "sparse")
-        logging.info(f"Loading SARM progress for RA-BC from {cfg.rabc_progress_path}")
-        logging.info(f"Using chunk_size={chunk_size} from policy config, head_mode={head_mode}")
-        rabc_weights = RABCWeights(
-            progress_path=cfg.rabc_progress_path,
-            chunk_size=chunk_size,
-            head_mode=head_mode,
-            kappa=getattr(cfg, "rabc_kappa", 0.01),
-            epsilon=getattr(cfg, "rabc_epsilon", 1e-6),
-            device=device,
+        if is_main_process:
+            logging.info(f"Creating sample weighter: {cfg.sample_weighting.type}")
+        sample_weighter = make_sample_weighter(
+            cfg.sample_weighting,
+            policy,
+            device,
+            dataset_root=cfg.dataset.root,
+            dataset_repo_id=cfg.dataset.repo_id,
         )
 
     step = 0  # number of policy updates (forward + backward + optim)
@@ -408,7 +404,7 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
             cfg.optimizer.grad_clip_norm,
             accelerator=accelerator,
             lr_scheduler=lr_scheduler,
-            rabc_weights_provider=rabc_weights,
+            sample_weighter=sample_weighter,
         )
 
         # Note: eval and checkpoint happens *after* the `step`th training update has completed, so we
@@ -425,16 +421,10 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
                 wandb_log_dict = train_tracker.to_dict()
                 if output_dict:
                     wandb_log_dict.update(output_dict)
-                # Log RA-BC statistics if enabled
-                if rabc_weights is not None:
-                    rabc_stats = rabc_weights.get_stats()
-                    wandb_log_dict.update(
-                        {
-                            "rabc_delta_mean": rabc_stats["delta_mean"],
-                            "rabc_delta_std": rabc_stats["delta_std"],
-                            "rabc_num_frames": rabc_stats["num_frames"],
-                        }
-                    )
+                # Log sample weighting statistics if enabled
+                if sample_weighter is not None:
+                    weighter_stats = sample_weighter.get_stats()
+                    wandb_log_dict.update({f"sample_weighting/{k}": v for k, v in weighter_stats.items()})
                 wandb_logger.log_dict(wandb_log_dict, step)
             train_tracker.reset_averages()
 

src/lerobot/scripts/lerobot_train_tokenizer.py

@@ -166,9 +166,9 @@ def apply_normalization(
         if q01 is None or q99 is None:
             raise ValueError("QUANTILES mode requires 'q01' and 'q99' in stats")
         denom = np.maximum(q99 - q01, eps)
-        # No clipping: match training pipeline NormalizerProcessorStep so tokenizer
-        # is fit on the full range of normalized values (including tails outside [-1, 1]).
-        return 2.0 * (data - q01) / denom - 1.0
+        # Clip to quantile range then normalize to [-1, 1]
+        clipped = np.clip(data, q01, q99)
+        return 2.0 * (clipped - q01) / denom - 1.0
 
     if mode == NormalizationMode.QUANTILE10:
         q10 = stats.get("q10")
@@ -176,8 +176,9 @@ def apply_normalization(
         if q10 is None or q90 is None:
             raise ValueError("QUANTILE10 mode requires 'q10' and 'q90' in stats")
         denom = np.maximum(q90 - q10, eps)
-        # No clipping: match training pipeline NormalizerProcessorStep.
-        return 2.0 * (data - q10) / denom - 1.0
+        # Clip to quantile range then normalize to [-1, 1]
+        clipped = np.clip(data, q10, q90)
+        return 2.0 * (clipped - q10) / denom - 1.0
 
     raise ValueError(f"Unsupported normalization mode: {mode}")
 
@@ -305,7 +306,7 @@ def train_fast_tokenizer(
 
     # download the tokenizer source code (not pretrained weights)
     # we'll train a new tokenizer on our own data
-    base_tokenizer = AutoProcessor.from_pretrained("/fsx/jade_choghari/outputs/libero_tokenizer_wavetoken1", trust_remote_code=True)
+    base_tokenizer = AutoProcessor.from_pretrained("physical-intelligence/fast", trust_remote_code=True)
 
     # convert action_chunks array to list of arrays (expected by .fit())
     action_data_list = [action_chunks[i] for i in range(len(action_chunks))]
@@ -319,8 +320,6 @@ def train_fast_tokenizer(
         vocab_size=vocab_size,
         time_horizon=action_chunks.shape[1],  # action_horizon
         action_dim=action_chunks.shape[2],  # encoded dimensions
-        wavelet="dmey",
-        level=1,
     )
     print("✓ Tokenizer training complete!")
 

src/lerobot/teleoperators/bi_openarm_leader/__init__.py

@@ -1,20 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 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 .bi_openarm_leader import BiOpenArmLeader
-from .config_bi_openarm_leader import BiOpenArmLeaderConfig
-
-__all__ = ["BiOpenArmLeader", "BiOpenArmLeaderConfig"]

src/lerobot/teleoperators/bi_openarm_leader/bi_openarm_leader.py

@@ -1,135 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import logging
-from functools import cached_property
-
-from lerobot.processor import RobotAction
-from lerobot.teleoperators.openarm_leader import OpenArmLeaderConfig
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-
-from ..openarm_leader import OpenArmLeader
-from ..teleoperator import Teleoperator
-from .config_bi_openarm_leader import BiOpenArmLeaderConfig
-
-logger = logging.getLogger(__name__)
-
-
-class BiOpenArmLeader(Teleoperator):
-    """
-    Bimanual OpenArm Leader Arms
-    """
-
-    config_class = BiOpenArmLeaderConfig
-    name = "bi_openarm_leader"
-
-    def __init__(self, config: BiOpenArmLeaderConfig):
-        super().__init__(config)
-        self.config = config
-
-        left_arm_config = OpenArmLeaderConfig(
-            id=f"{config.id}_left" if config.id else None,
-            calibration_dir=config.calibration_dir,
-            port=config.left_arm_config.port,
-            can_interface=config.left_arm_config.can_interface,
-            use_can_fd=config.left_arm_config.use_can_fd,
-            can_bitrate=config.left_arm_config.can_bitrate,
-            can_data_bitrate=config.left_arm_config.can_data_bitrate,
-            motor_config=config.left_arm_config.motor_config,
-            manual_control=config.left_arm_config.manual_control,
-            position_kd=config.left_arm_config.position_kd,
-            position_kp=config.left_arm_config.position_kp,
-        )
-
-        right_arm_config = OpenArmLeaderConfig(
-            id=f"{config.id}_right" if config.id else None,
-            calibration_dir=config.calibration_dir,
-            port=config.right_arm_config.port,
-            can_interface=config.right_arm_config.can_interface,
-            use_can_fd=config.right_arm_config.use_can_fd,
-            can_bitrate=config.right_arm_config.can_bitrate,
-            can_data_bitrate=config.right_arm_config.can_data_bitrate,
-            motor_config=config.right_arm_config.motor_config,
-            manual_control=config.right_arm_config.manual_control,
-            position_kd=config.right_arm_config.position_kd,
-            position_kp=config.right_arm_config.position_kp,
-        )
-
-        self.left_arm = OpenArmLeader(left_arm_config)
-        self.right_arm = OpenArmLeader(right_arm_config)
-
-    @cached_property
-    def action_features(self) -> dict[str, type]:
-        left_arm_features = self.left_arm.action_features
-        right_arm_features = self.right_arm.action_features
-
-        return {
-            **{f"left_{k}": v for k, v in left_arm_features.items()},
-            **{f"right_{k}": v for k, v in right_arm_features.items()},
-        }
-
-    @cached_property
-    def feedback_features(self) -> dict[str, type]:
-        return {}
-
-    @property
-    def is_connected(self) -> bool:
-        return self.left_arm.is_connected and self.right_arm.is_connected
-
-    @check_if_already_connected
-    def connect(self, calibrate: bool = True) -> None:
-        self.left_arm.connect(calibrate)
-        self.right_arm.connect(calibrate)
-
-    @property
-    def is_calibrated(self) -> bool:
-        return self.left_arm.is_calibrated and self.right_arm.is_calibrated
-
-    def calibrate(self) -> None:
-        self.left_arm.calibrate()
-        self.right_arm.calibrate()
-
-    def configure(self) -> None:
-        self.left_arm.configure()
-        self.right_arm.configure()
-
-    def setup_motors(self) -> None:
-        raise NotImplementedError(
-            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
-        )
-
-    @check_if_not_connected
-    def get_action(self) -> RobotAction:
-        action_dict = {}
-
-        # Add "left_" prefix
-        left_action = self.left_arm.get_action()
-        action_dict.update({f"left_{key}": value for key, value in left_action.items()})
-
-        # Add "right_" prefix
-        right_action = self.right_arm.get_action()
-        action_dict.update({f"right_{key}": value for key, value in right_action.items()})
-
-        return action_dict
-
-    def send_feedback(self, feedback: dict[str, float]) -> None:
-        # TODO: Implement force feedback
-        raise NotImplementedError
-
-    @check_if_not_connected
-    def disconnect(self) -> None:
-        self.left_arm.disconnect()
-        self.right_arm.disconnect()

src/lerobot/teleoperators/bi_openarm_leader/config_bi_openarm_leader.py

@@ -1,30 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from dataclasses import dataclass
-
-from lerobot.teleoperators.openarm_leader import OpenArmLeaderConfigBase
-
-from ..config import TeleoperatorConfig
-
-
-@TeleoperatorConfig.register_subclass("bi_openarm_leader")
-@dataclass
-class BiOpenArmLeaderConfig(TeleoperatorConfig):
-    """Configuration class for Bi OpenArm Follower robots."""
-
-    left_arm_config: OpenArmLeaderConfigBase
-    right_arm_config: OpenArmLeaderConfigBase

src/lerobot/teleoperators/bi_so_leader/bi_so_leader.py

@@ -18,7 +18,7 @@ import logging
 from functools import cached_property
 
 from lerobot.teleoperators.so_leader import SOLeaderTeleopConfig
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
+from lerobot.utils.decorators import check_if_not_connected
 
 from ..so_leader import SOLeader
 from ..teleoperator import Teleoperator
@@ -72,7 +72,6 @@ class BiSOLeader(Teleoperator):
     def is_connected(self) -> bool:
         return self.left_arm.is_connected and self.right_arm.is_connected
 
-    @check_if_already_connected
     def connect(self, calibrate: bool = True) -> None:
         self.left_arm.connect(calibrate)
         self.right_arm.connect(calibrate)
@@ -111,7 +110,6 @@ class BiSOLeader(Teleoperator):
         # TODO: Implement force feedback
         raise NotImplementedError
 
-    @check_if_not_connected
     def disconnect(self) -> None:
         self.left_arm.disconnect()
         self.right_arm.disconnect()

src/lerobot/teleoperators/openarm_leader/__init__.py

@@ -1,20 +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 .config_openarm_leader import OpenArmLeaderConfig, OpenArmLeaderConfigBase
-from .openarm_leader import OpenArmLeader
-
-__all__ = ["OpenArmLeader", "OpenArmLeaderConfig", "OpenArmLeaderConfigBase"]

src/lerobot/teleoperators/openarm_leader/config_openarm_leader.py

@@ -1,75 +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, field
-
-from ..config import TeleoperatorConfig
-
-
-@dataclass
-class OpenArmLeaderConfigBase:
-    """Base configuration for the OpenArms leader/teleoperator with Damiao motors."""
-
-    # CAN interfaces - one per arm
-    # Arm CAN interface (e.g., "can3")
-    # Linux: "can0", "can1", etc.
-    port: str
-
-    # CAN interface type: "socketcan" (Linux), "slcan" (serial), or "auto" (auto-detect)
-    can_interface: str = "socketcan"
-
-    # CAN FD settings (OpenArms uses CAN FD by default)
-    use_can_fd: bool = True
-    can_bitrate: int = 1000000  # Nominal bitrate (1 Mbps)
-    can_data_bitrate: int = 5000000  # Data bitrate for CAN FD (5 Mbps)
-
-    # Motor configuration for OpenArms (7 DOF per arm)
-    # Maps motor names to (send_can_id, recv_can_id, motor_type)
-    # Based on: https://docs.openarm.dev/software/setup/configure-test
-    # OpenArms uses 4 types of motors:
-    # - DM8009 (DM-J8009P-2EC) for shoulders (high torque)
-    # - DM4340P and DM4340 for shoulder rotation and elbow
-    # - DM4310 (DM-J4310-2EC V1.1) for wrist and gripper
-    motor_config: dict[str, tuple[int, int, str]] = field(
-        default_factory=lambda: {
-            "joint_1": (0x01, 0x11, "dm8009"),  # J1 - Shoulder pan (DM8009)
-            "joint_2": (0x02, 0x12, "dm8009"),  # J2 - Shoulder lift (DM8009)
-            "joint_3": (0x03, 0x13, "dm4340"),  # J3 - Shoulder rotation (DM4340)
-            "joint_4": (0x04, 0x14, "dm4340"),  # J4 - Elbow flex (DM4340)
-            "joint_5": (0x05, 0x15, "dm4310"),  # J5 - Wrist roll (DM4310)
-            "joint_6": (0x06, 0x16, "dm4310"),  # J6 - Wrist pitch (DM4310)
-            "joint_7": (0x07, 0x17, "dm4310"),  # J7 - Wrist rotation (DM4310)
-            "gripper": (0x08, 0x18, "dm4310"),  # J8 - Gripper (DM4310)
-        }
-    )
-
-    # Torque mode settings for manual control
-    # When enabled, motors have torque disabled for manual movement
-    manual_control: bool = True
-
-    # TODO(Steven, Pepijn): Not used ... ?
-    # MIT control parameters (used when manual_control=False for torque control)
-    # List of 8 values: [joint_1, joint_2, joint_3, joint_4, joint_5, joint_6, joint_7, gripper]
-    position_kp: list[float] = field(
-        default_factory=lambda: [240.0, 240.0, 240.0, 240.0, 24.0, 31.0, 25.0, 16.0]
-    )
-    position_kd: list[float] = field(default_factory=lambda: [3.0, 3.0, 3.0, 3.0, 0.2, 0.2, 0.2, 0.2])
-
-
-@TeleoperatorConfig.register_subclass("openarm_leader")
-@dataclass
-class OpenArmLeaderConfig(TeleoperatorConfig, OpenArmLeaderConfigBase):
-    pass

src/lerobot/teleoperators/openarm_leader/openarm_leader.py

@@ -1,222 +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 logging
-import time
-from typing import Any
-
-from lerobot.motors import Motor, MotorCalibration, MotorNormMode
-from lerobot.motors.damiao import DamiaoMotorsBus
-from lerobot.processor import RobotAction
-from lerobot.utils.decorators import check_if_already_connected, check_if_not_connected
-
-from ..teleoperator import Teleoperator
-from .config_openarm_leader import OpenArmLeaderConfig
-
-logger = logging.getLogger(__name__)
-
-
-class OpenArmLeader(Teleoperator):
-    """
-    OpenArm Leader/Teleoperator Arm with Damiao motors.
-
-    This teleoperator uses CAN bus communication to read positions from
-    Damiao motors that are manually moved (torque disabled).
-    """
-
-    config_class = OpenArmLeaderConfig
-    name = "openarm_leader"
-
-    def __init__(self, config: OpenArmLeaderConfig):
-        super().__init__(config)
-        self.config = config
-
-        # Arm motors
-        motors: dict[str, Motor] = {}
-        for motor_name, (send_id, recv_id, motor_type_str) in config.motor_config.items():
-            motor = Motor(
-                send_id, motor_type_str, MotorNormMode.DEGREES
-            )  # Always use degrees for Damiao motors
-            motor.recv_id = recv_id
-            motor.motor_type_str = motor_type_str
-            motors[motor_name] = motor
-
-        self.bus = DamiaoMotorsBus(
-            port=self.config.port,
-            motors=motors,
-            calibration=self.calibration,
-            can_interface=self.config.can_interface,
-            use_can_fd=self.config.use_can_fd,
-            bitrate=self.config.can_bitrate,
-            data_bitrate=self.config.can_data_bitrate if self.config.use_can_fd else None,
-        )
-
-    @property
-    def action_features(self) -> dict[str, type]:
-        """Features produced by this teleoperator."""
-        features: dict[str, type] = {}
-        for motor in self.bus.motors:
-            features[f"{motor}.pos"] = float
-            features[f"{motor}.vel"] = float
-            features[f"{motor}.torque"] = float
-        return features
-
-    @property
-    def feedback_features(self) -> dict[str, type]:
-        """Feedback features (not implemented for OpenArms)."""
-        return {}
-
-    @property
-    def is_connected(self) -> bool:
-        """Check if teleoperator is connected."""
-        return self.bus.is_connected
-
-    @check_if_already_connected
-    def connect(self, calibrate: bool = True) -> None:
-        """
-        Connect to the teleoperator.
-
-        For manual control, we disable torque after connecting so the
-        arm can be moved by hand.
-        """
-
-        # Connect to CAN bus
-        logger.info(f"Connecting arm on {self.config.port}...")
-        self.bus.connect()
-
-        # Run calibration if needed
-        if not self.is_calibrated and calibrate:
-            logger.info(
-                "Mismatch between calibration values in the motor and the calibration file or no calibration file found"
-            )
-            self.calibrate()
-
-        self.configure()
-
-        if self.is_calibrated:
-            self.bus.set_zero_position()
-
-        logger.info(f"{self} connected.")
-
-    @property
-    def is_calibrated(self) -> bool:
-        """Check if teleoperator is calibrated."""
-        return self.bus.is_calibrated
-
-    def calibrate(self) -> None:
-        """
-        Run calibration procedure for OpenArms leader.
-
-        The calibration procedure:
-        1. Disable torque (if not already disabled)
-        2. Ask user to position arm in zero position (hanging with gripper closed)
-        3. Set this as zero position
-        4. Record range of motion for each joint
-        5. Save calibration
-        """
-        if self.calibration:
-            # Calibration file exists, ask user whether to use it or run new calibration
-            user_input = input(
-                f"Press ENTER to use provided calibration file associated with the id {self.id}, or type 'c' and press ENTER to run calibration: "
-            )
-            if user_input.strip().lower() != "c":
-                logger.info(f"Writing calibration file associated with the id {self.id} to the motors")
-                self.bus.write_calibration(self.calibration)
-                return
-
-        logger.info(f"\nRunning calibration for {self}")
-        self.bus.disable_torque()
-
-        # Step 1: Set zero position
-        input(
-            "\nCalibration: Set Zero Position)\n"
-            "Position the arm in the following configuration:\n"
-            "  - Arm hanging straight down\n"
-            "  - Gripper closed\n"
-            "Press ENTER when ready..."
-        )
-
-        # Set current position as zero for all motors
-        self.bus.set_zero_position()
-        logger.info("Arm zero position set.")
-
-        logger.info("Setting range: -90° to +90° by default for all joints")
-        # TODO(Steven, Pepijn): Check if MotorCalibration is actually needed here given that we only use Degrees
-        for motor_name, motor in self.bus.motors.items():
-            self.calibration[motor_name] = MotorCalibration(
-                id=motor.id,
-                drive_mode=0,
-                homing_offset=0,
-                range_min=-90,
-                range_max=90,
-            )
-
-        self.bus.write_calibration(self.calibration)
-        self._save_calibration()
-        print(f"Calibration saved to {self.calibration_fpath}")
-
-    def configure(self) -> None:
-        """
-        Configure motors for manual teleoperation.
-
-        For manual control, we disable torque so the arm can be moved by hand.
-        """
-
-        return self.bus.disable_torque() if self.config.manual_control else self.bus.configure_motors()
-
-    def setup_motors(self) -> None:
-        raise NotImplementedError(
-            "Motor ID configuration is typically done via manufacturer tools for CAN motors."
-        )
-
-    @check_if_not_connected
-    def get_action(self) -> RobotAction:
-        """
-        Get current action from the leader arm.
-
-        This is the main method for teleoperators - it reads the current state
-        of the leader arm and returns it as an action that can be sent to a follower.
-
-        Reads all motor states (pos/vel/torque) in one CAN refresh cycle.
-        """
-        start = time.perf_counter()
-
-        action_dict: dict[str, Any] = {}
-
-        # Use sync_read_all_states to get pos/vel/torque in one go
-        states = self.bus.sync_read_all_states()
-        for motor in self.bus.motors:
-            state = states.get(motor, {})
-            action_dict[f"{motor}.pos"] = state.get("position")
-            action_dict[f"{motor}.vel"] = state.get("velocity")
-            action_dict[f"{motor}.torque"] = state.get("torque")
-
-        dt_ms = (time.perf_counter() - start) * 1e3
-        logger.debug(f"{self} read state: {dt_ms:.1f}ms")
-
-        return action_dict
-
-    def send_feedback(self, feedback: dict[str, float]) -> None:
-        raise NotImplementedError("Feedback is not yet implemented for OpenArm leader.")
-
-    @check_if_not_connected
-    def disconnect(self) -> None:
-        """Disconnect from teleoperator."""
-
-        # Disconnect CAN bus
-        # For manual control, ensure torque is disabled before disconnecting
-        self.bus.disconnect(disable_torque=self.config.manual_control)
-        logger.info(f"{self} disconnected.")

src/lerobot/teleoperators/unitree_g1/__init__.py

@@ -1,21 +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 .config_unitree_g1 import ExoskeletonArmPortConfig, UnitreeG1TeleoperatorConfig
-from .exo_calib import ExoskeletonCalibration, ExoskeletonJointCalibration
-from .exo_ik import ExoskeletonIKHelper
-from .exo_serial import ExoskeletonArm
-from .unitree_g1 import UnitreeG1Teleoperator

src/lerobot/teleoperators/unitree_g1/config_unitree_g1.py

@@ -1,37 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from dataclasses import dataclass, field
-
-from ..config import TeleoperatorConfig
-
-
-@dataclass
-class ExoskeletonArmPortConfig:
-    """Serial port configuration for individual exoskeleton arm."""
-
-    port: str = ""
-    baud_rate: int = 115200
-
-
-@TeleoperatorConfig.register_subclass("unitree_g1")
-@dataclass
-class UnitreeG1TeleoperatorConfig(TeleoperatorConfig):
-    left_arm_config: ExoskeletonArmPortConfig = field(default_factory=ExoskeletonArmPortConfig)
-    right_arm_config: ExoskeletonArmPortConfig = field(default_factory=ExoskeletonArmPortConfig)
-
-    # Frozen joints (comma-separated joint names that won't be moved by IK)
-    frozen_joints: str = ""