fix(rewards/topreward): Remove TOPReward README symlink, use docs page only

Add lerobot[topreward] extra to all in
pyproject.toml, drop the redundant labels arg in scoring, and
collapse the dead-branch shape check in the encoder processor.

.github/workflows/release.yml

@@ -152,13 +152,14 @@ jobs:
             BASE_VERSION="${VERSION%%-*}"
             echo "Installing pre-release version $BASE_VERSION from TestPyPI..."
             uv pip install \
+              --torch-backend cpu \
               --index-url https://test.pypi.org/simple/ \
               --extra-index-url https://pypi.org/simple \
               --index-strategy unsafe-best-match \
                "lerobot[all]==$BASE_VERSION"
           else
             echo "Installing release version $VERSION from PyPI..."
-            uv pip install "lerobot[all]==$VERSION"
+            uv pip install --torch-backend cpu "lerobot[all]==$VERSION"
           fi
       - name: Check lerobot version
         run: uv run python -c "import lerobot; print(lerobot.__version__)"

AGENT_GUIDE.md

@@ -232,6 +232,8 @@ Match the policy to the user's **GPU memory** and **time budget**. Numbers below
 
 All policies typically train for **5–10 epochs** (see §7).
 
+> **Human-facing version:** the [Compute Hardware Guide](./docs/source/hardware_guide.mdx) reuses the table below and adds a cloud-GPU tier guide and a Hugging Face Jobs pointer.
+
 | Policy      | Batch | Update (ms) | Peak GPU mem (GB) | Best for                                                                                         |
 | ----------- | ----: | ----------: | ----------------: | ------------------------------------------------------------------------------------------------ |
 | `act`       |     4 |    **83.9** |          **0.94** | First-time users, laptops, single-task. Fast and reliable.                                       |

README.md

@@ -109,7 +109,7 @@ lerobot-train \
 
 Similarly to the hardware, you can easily implement your own policy & leverage LeRobot's data collection, training, and visualization tools, and share your model to the HF Hub
 
-For detailed policy setup guides, see the [Policy Documentation](https://huggingface.co/docs/lerobot/bring_your_own_policies).
+For detailed policy setup guides, see the [Policy Documentation](https://huggingface.co/docs/lerobot/bring_your_own_policies). For GPU/RAM requirements and expected training time per policy, see the [Compute Hardware Guide](https://huggingface.co/docs/lerobot/hardware_guide).
 
 ## Inference & Evaluation
 

benchmarks/video/README.md

@@ -1,288 +0,0 @@
-# Video benchmark
-
-## Questions
-
-What is the optimal trade-off between:
-
-- maximizing loading time with random access,
-- minimizing memory space on disk,
-- maximizing success rate of policies,
-- compatibility across devices/platforms for decoding videos (e.g. video players, web browsers).
-
-How to encode videos?
-
-- Which video codec (`-vcodec`) to use? h264, h265, AV1?
-- What pixel format to use (`-pix_fmt`)? `yuv444p` or `yuv420p`?
-- How much compression (`-crf`)? No compression with `0`, intermediate compression with `25` or extreme with `50+`?
-- Which frequency to chose for key frames (`-g`)? A key frame every `10` frames?
-
-How to decode videos?
-
-- Which `decoder`? `torchvision`, `torchaudio`, `ffmpegio`, `decord`, or `nvc`?
-- What scenarios to use for the requesting timestamps during benchmark? (`timestamps_mode`)
-
-## Variables
-
-**Image content & size**
-We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an apartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
-For these reasons, we run this benchmark on four representative datasets:
-
-- `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
-- `lerobot/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
-- `lerobot/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
-- `lerobot/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
-
-Note: The datasets used for this benchmark need to be image datasets, not video datasets.
-
-**Data augmentations**
-We might revisit this benchmark and find better settings if we train our policies with various data augmentations to make them more robust (e.g. robust to color changes, compression, etc.).
-
-### Encoding parameters
-
-| parameter   | values                                                       |
-| ----------- | ------------------------------------------------------------ |
-| **vcodec**  | `libx264`, `libx265`, `libsvtav1`                            |
-| **pix_fmt** | `yuv444p`, `yuv420p`                                         |
-| **g**       | `1`, `2`, `3`, `4`, `5`, `6`, `10`, `15`, `20`, `40`, `None` |
-| **crf**     | `0`, `5`, `10`, `15`, `20`, `25`, `30`, `40`, `50`, `None`   |
-
-Note that `crf` value might be interpreted differently by various video codecs. In other words, the same value used with one codec doesn't necessarily translate into the same compression level with another codec. In fact, the default value (`None`) isn't the same amongst the different video codecs. Importantly, it is also the case for many other ffmpeg arguments like `g` which specifies the frequency of the key frames.
-
-For a comprehensive list and documentation of these parameters, see the ffmpeg documentation depending on the video codec used:
-
-- h264: https://trac.ffmpeg.org/wiki/Encode/H.264
-- h265: https://trac.ffmpeg.org/wiki/Encode/H.265
-- AV1: https://trac.ffmpeg.org/wiki/Encode/AV1
-
-### Decoding parameters
-
-**Decoder**
-We tested two video decoding backends from torchvision:
-
-- `pyav`
-- `video_reader` (requires to build torchvision from source)
-
-**Requested timestamps**
-Given the way video decoding works, once a keyframe has been loaded, the decoding of subsequent frames is fast.
-This of course is affected by the `-g` parameter during encoding, which specifies the frequency of the keyframes. Given our typical use cases in robotics policies which might request a few timestamps in different random places, we want to replicate these use cases with the following scenarios:
-
-- `1_frame`: 1 frame,
-- `2_frames`: 2 consecutive frames (e.g. `[t, t + 1 / fps]`),
-- `6_frames`: 6 consecutive frames (e.g. `[t + i / fps for i in range(6)]`)
-
-Note that this differs significantly from a typical use case like watching a movie, in which every frame is loaded sequentially from the beginning to the end and it's acceptable to have big values for `-g`.
-
-Additionally, because some policies might request single timestamps that are a few frames apart, we also have the following scenario:
-
-- `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
-
-However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
-
-## Metrics
-
-**Data compression ratio (lower is better)**
-`video_images_size_ratio` is the ratio of the memory space on disk taken by the encoded video over the memory space taken by the original images. For instance, `video_images_size_ratio=25%` means that the video takes 4 times less memory space on disk compared to the original images.
-
-**Loading time ratio (lower is better)**
-`video_images_load_time_ratio` is the ratio of the time it takes to decode frames from the video at a given timestamps over the time it takes to load the exact same original images. Lower is better. For instance, `video_images_load_time_ratio=200%` means that decoding from video is 2 times slower than loading the original images.
-
-**Average Mean Square Error (lower is better)**
-`avg_mse` is the average mean square error between each decoded frame and its corresponding original image over all requested timestamps, and also divided by the number of pixels in the image to be comparable when switching to different image sizes.
-
-**Average Peak Signal to Noise Ratio (higher is better)**
-`avg_psnr` measures the ratio between the maximum possible power of a signal and the power of corrupting noise that affects the fidelity of its representation. Higher PSNR indicates better quality.
-
-**Average Structural Similarity Index Measure (higher is better)**
-`avg_ssim` evaluates the perceived quality of images by comparing luminance, contrast, and structure. SSIM values range from -1 to 1, where 1 indicates perfect similarity.
-
-One aspect that can't be measured here with those metrics is the compatibility of the encoding across platforms, in particular on web browser, for visualization purposes.
-h264, h265 and AV1 are all commonly used codecs and should not pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
-
-- `yuv420p` is more widely supported across various platforms, including web browsers.
-- `yuv444p` offers higher color fidelity but might not be supported as broadly.
-
-<!-- **Loss of a pretrained policy (higher is better)** (not available)
-`loss_pretrained` is the result of evaluating with the selected encoding/decoding settings a policy pretrained on original images. It is easier to understand than `avg_l2_error`.
-
-**Success rate after retraining (higher is better)** (not available)
-`success_rate` is the result of training and evaluating a policy with the selected encoding/decoding settings. It is the most difficult metric to get but also the very best. -->
-
-## How the benchmark works
-
-The benchmark evaluates both encoding and decoding of video frames on the first episode of each dataset.
-
-**Encoding:** for each `vcodec` and `pix_fmt` pair, we use a default value for `g` and `crf` upon which we change a single value (either `g` or `crf`) to one of the specified values (we don't test every combination of those as this would be computationally too heavy).
-This gives a unique set of encoding parameters which is used to encode the episode.
-
-**Decoding:** Then, for each of those unique encodings, we iterate through every combination of the decoding parameters `backend` and `timestamps_mode`. For each of them, we record the metrics of a number of samples (given by `--num-samples`). This is parallelized for efficiency and the number of processes can be controlled with `--num-workers`. Ideally, it's best to have a `--num-samples` that is divisible by `--num-workers`.
-
-Intermediate results saved for each `vcodec` and `pix_fmt` combination in csv tables.
-These are then all concatenated to a single table ready for analysis.
-
-## Caveats
-
-We tried to measure the most impactful parameters for both encoding and decoding. However, for computational reasons we can't test out every combination.
-
-Additional encoding parameters exist that are not included in this benchmark. In particular:
-
-- `-preset` which allows for selecting encoding presets. This represents a collection of options that will provide a certain encoding speed to compression ratio. By leaving this parameter unspecified, it is considered to be `medium` for libx264 and libx265 and `8` for libsvtav1.
-- `-tune` which allows to optimize the encoding for certain aspects (e.g. film quality, fast decoding, etc.).
-
-See the documentation mentioned above for more detailed info on these settings and for a more comprehensive list of other parameters.
-
-Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
-
-- `torchaudio`
-- `ffmpegio`
-- `decord`
-- `nvc`
-
-Note as well that since we are mostly interested in the performance at decoding time (also because encoding is done only once before uploading a dataset), we did not measure encoding times nor have any metrics regarding encoding.
-However, besides the necessity to build ffmpeg from source, encoding did not pose any issue and it didn't take a significant amount of time during this benchmark.
-
-## Install
-
-Building ffmpeg from source is required to include libx265 and libaom/libsvtav1 (av1) video codecs ([compilation guide](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu)).
-
-**Note:** While you still need to build torchvision with a conda-installed `ffmpeg<4.3` to use the `video_reader` decoder (as described in [#220](https://github.com/huggingface/lerobot/pull/220)), you also need another version which is custom-built with all the video codecs for encoding. For the script to then use that version, you can prepend the command above with `PATH="$HOME/bin:$PATH"`, which is where ffmpeg should be built.
-
-## Adding a video decoder
-
-Right now, we're only benchmarking the two video decoder available with torchvision: `pyav` and `video_reader`.
-You can easily add a new decoder to benchmark by adding it to this function in the script:
-
-```diff
-def decode_video_frames(
-    video_path: str,
-    timestamps: list[float],
-    tolerance_s: float,
-    backend: str,
-) -> torch.Tensor:
-    if backend in ["pyav", "video_reader"]:
-        return decode_video_frames_torchvision(
-            video_path, timestamps, tolerance_s, backend
-        )
-+    elif backend == ["your_decoder"]:
-+        return your_decoder_function(
-+            video_path, timestamps, tolerance_s, backend
-+        )
-    else:
-        raise NotImplementedError(backend)
-```
-
-## Example
-
-For a quick run, you can try these parameters:
-
-```bash
-python benchmark/video/run_video_benchmark.py \
-    --output-dir outputs/video_benchmark \
-    --repo-ids \
-        lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-    --vcodec libx264 libx265 \
-    --pix-fmt yuv444p yuv420p \
-    --g 2 20 None \
-    --crf 10 40 None \
-    --timestamps-modes 1_frame 2_frames \
-    --backends pyav video_reader \
-    --num-samples 5 \
-    --num-workers 5 \
-    --save-frames 0
-```
-
-## Results
-
-### Reproduce
-
-We ran the benchmark with the following parameters:
-
-```bash
-# h264 and h265 encodings
-python benchmark/video/run_video_benchmark.py \
-    --output-dir outputs/video_benchmark \
-    --repo-ids \
-        lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
-    --vcodec libx264 libx265 \
-    --pix-fmt yuv444p yuv420p \
-    --g 1 2 3 4 5 6 10 15 20 40 None \
-    --crf 0 5 10 15 20 25 30 40 50 None \
-    --timestamps-modes 1_frame 2_frames 6_frames \
-    --backends pyav video_reader \
-    --num-samples 50 \
-    --num-workers 5 \
-    --save-frames 1
-
-# av1 encoding (only compatible with yuv420p and pyav decoder)
-python benchmark/video/run_video_benchmark.py \
-    --output-dir outputs/video_benchmark \
-    --repo-ids \
-        lerobot/pusht_image \
-        lerobot/aloha_mobile_shrimp_image \
-        lerobot/paris_street \
-        lerobot/kitchen \
-    --vcodec libsvtav1 \
-    --pix-fmt yuv420p \
-    --g 1 2 3 4 5 6 10 15 20 40 None \
-    --crf 0 5 10 15 20 25 30 40 50 None \
-    --timestamps-modes 1_frame 2_frames 6_frames \
-    --backends pyav \
-    --num-samples 50 \
-    --num-workers 5 \
-    --save-frames 1
-```
-
-The full results are available [here](https://docs.google.com/spreadsheets/d/1OYJB43Qu8fC26k_OyoMFgGBBKfQRCi4BIuYitQnq3sw/edit?usp=sharing)
-
-### Parameters selected for LeRobotDataset
-
-Considering these results, we chose what we think is the best set of encoding parameter:
-
-- vcodec: `libsvtav1`
-- pix-fmt: `yuv420p`
-- g: `2`
-- crf: `30`
-
-Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_reader` does not support it (and `pyav` doesn't require a custom build of `torchvision`).
-
-### Summary
-
-These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
-
-| video_images_size_ratio           | vcodec     | pix_fmt |           |           |           |
-| --------------------------------- | ---------- | ------- | --------- | --------- | --------- |
-|                                   | libx264    |         | libx265   |           | libsvtav1 |
-| repo_id                           | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
-| lerobot/pusht_image               | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
-| lerobot/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
-| lerobot/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
-| lerobot/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
-
-| video_images_load_time_ratio      | vcodec  | pix_fmt |          |         |           |
-| --------------------------------- | ------- | ------- | -------- | ------- | --------- |
-|                                   | libx264 |         | libx265  |         | libsvtav1 |
-| repo_id                           | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
-| lerobot/pusht_image               | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
-| lerobot/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
-| lerobot/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
-| lerobot/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
-
-|                                   |          | vcodec   | pix_fmt      |          |           |              |
-| --------------------------------- | -------- | -------- | ------------ | -------- | --------- | ------------ |
-|                                   |          | libx264  |              | libx265  |           | libsvtav1    |
-| repo_id                           | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
-| lerobot/pusht_image               | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
-|                                   | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
-|                                   | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
-| lerobot/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
-|                                   | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
-|                                   | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
-| lerobot/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
-|                                   | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
-|                                   | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
-| lerobot/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
-|                                   | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
-|                                   | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

benchmarks/video/run_video_benchmark.py

@@ -1,488 +0,0 @@
-#!/usr/bin/env python
-
-# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Assess the performance of video decoding in various configurations.
-
-This script will benchmark different video encoding and decoding parameters.
-See the provided README.md or run `python benchmark/video/run_video_benchmark.py --help` for usage info.
-"""
-
-import argparse
-import datetime as dt
-import itertools
-import random
-import shutil
-from collections import OrderedDict
-from concurrent.futures import ThreadPoolExecutor, as_completed
-from pathlib import Path
-from threading import Lock
-
-import einops
-import numpy as np
-import pandas as pd
-import PIL
-import torch
-from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
-from tqdm import tqdm
-
-from lerobot.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.datasets.video_utils import (
-    decode_video_frames,
-    encode_video_frames,
-)
-from lerobot.utils.constants import OBS_IMAGE
-from lerobot.utils.utils import TimerManager
-
-BASE_ENCODING = OrderedDict(
-    [
-        ("vcodec", "libx264"),
-        ("pix_fmt", "yuv444p"),
-        ("g", 2),
-        ("crf", None),
-        # TODO(aliberts): Add fastdecode
-        # ("fastdecode", 0),
-    ]
-)
-
-
-# TODO(rcadene, aliberts): move to `utils.py` folder when we want to refactor
-def parse_int_or_none(value) -> int | None:
-    if value.lower() == "none":
-        return None
-    try:
-        return int(value)
-    except ValueError as e:
-        raise argparse.ArgumentTypeError(f"Invalid int or None: {value}") from e
-
-
-def check_datasets_formats(repo_ids: list) -> None:
-    for repo_id in repo_ids:
-        dataset = LeRobotDataset(repo_id)
-        if len(dataset.meta.video_keys) > 0:
-            raise ValueError(
-                f"Use only image dataset for running this benchmark. Video dataset provided: {repo_id}"
-            )
-
-
-def get_directory_size(directory: Path) -> int:
-    total_size = 0
-    for item in directory.rglob("*"):
-        if item.is_file():
-            total_size += item.stat().st_size
-    return total_size
-
-
-def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> torch.Tensor:
-    frames = []
-    for ts in timestamps:
-        idx = int(ts * fps)
-        frame = PIL.Image.open(imgs_dir / f"frame-{idx:06d}.png")
-        frame = torch.from_numpy(np.array(frame))
-        frame = frame.type(torch.float32) / 255
-        frame = einops.rearrange(frame, "h w c -> c h w")
-        frames.append(frame)
-    return torch.stack(frames)
-
-
-def save_decoded_frames(
-    imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
-) -> None:
-    if save_dir.exists() and len(list(save_dir.glob("frame-*.png"))) == len(timestamps):
-        return
-
-    save_dir.mkdir(parents=True, exist_ok=True)
-    for i, ts in enumerate(timestamps):
-        idx = int(ts * fps)
-        frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
-        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame-{idx:06d}_decoded.png")
-        shutil.copyfile(imgs_dir / f"frame-{idx:06d}.png", save_dir / f"frame-{idx:06d}_original.png")
-
-
-def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
-    if imgs_dir.exists() and len(list(imgs_dir.glob("frame-*.png"))) == ep_num_images:
-        return
-
-    imgs_dir.mkdir(parents=True, exist_ok=True)
-    hf_dataset = dataset.hf_dataset.with_format(None)
-
-    # We only save images from the first camera
-    img_keys = [key for key in hf_dataset.features if key.startswith(OBS_IMAGE)]
-    imgs_dataset = hf_dataset.select_columns(img_keys[0])
-
-    for i, item in enumerate(
-        tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
-    ):
-        img = item[img_keys[0]]
-        img.save(str(imgs_dir / f"frame-{i:06d}.png"), quality=100)
-
-        if i >= ep_num_images - 1:
-            break
-
-
-def sample_timestamps(timestamps_mode: str, ep_num_images: int, fps: int) -> list[float]:
-    # Start at 5 to allow for 2_frames_4_space and 6_frames
-    idx = random.randint(5, ep_num_images - 1)
-    match timestamps_mode:
-        case "1_frame":
-            frame_indexes = [idx]
-        case "2_frames":
-            frame_indexes = [idx - 1, idx]
-        case "2_frames_4_space":
-            frame_indexes = [idx - 5, idx]
-        case "6_frames":
-            frame_indexes = [idx - i for i in range(6)][::-1]
-        case _:
-            raise ValueError(timestamps_mode)
-
-    return [idx / fps for idx in frame_indexes]
-
-
-def benchmark_decoding(
-    imgs_dir: Path,
-    video_path: Path,
-    timestamps_mode: str,
-    backend: str,
-    ep_num_images: int,
-    fps: int,
-    num_samples: int = 50,
-    num_workers: int = 4,
-    save_frames: bool = False,
-) -> dict:
-    def process_sample(sample: int, lock: Lock):
-        time_benchmark = TimerManager(log=False)
-        timestamps = sample_timestamps(timestamps_mode, ep_num_images, fps)
-        num_frames = len(timestamps)
-        result = {
-            "psnr_values": [],
-            "ssim_values": [],
-            "mse_values": [],
-        }
-
-        with time_benchmark, lock:
-            frames = decode_video_frames(video_path, timestamps=timestamps, tolerance_s=5e-1, backend=backend)
-        result["load_time_video_ms"] = (time_benchmark.last * 1000) / num_frames
-
-        with time_benchmark:
-            original_frames = load_original_frames(imgs_dir, timestamps, fps)
-        result["load_time_images_ms"] = (time_benchmark.last * 1000) / num_frames
-
-        frames_np, original_frames_np = frames.numpy(), original_frames.numpy()
-        for i in range(num_frames):
-            result["mse_values"].append(mean_squared_error(original_frames_np[i], frames_np[i]))
-            result["psnr_values"].append(
-                peak_signal_noise_ratio(original_frames_np[i], frames_np[i], data_range=1.0)
-            )
-            result["ssim_values"].append(
-                structural_similarity(original_frames_np[i], frames_np[i], data_range=1.0, channel_axis=0)
-            )
-
-        if save_frames and sample == 0:
-            save_dir = video_path.with_suffix("") / f"{timestamps_mode}_{backend}"
-            save_decoded_frames(imgs_dir, save_dir, frames, timestamps, fps)
-
-        return result
-
-    load_times_video_ms = []
-    load_times_images_ms = []
-    mse_values = []
-    psnr_values = []
-    ssim_values = []
-
-    # A sample is a single set of decoded frames specified by timestamps_mode (e.g. a single frame, 2 frames, etc.).
-    # For each sample, we record metrics (loading time and quality metrics) which are then averaged over all samples.
-    # As these samples are independent, we run them in parallel threads to speed up the benchmark.
-    # Use a single shared lock for all worker threads
-    shared_lock = Lock()
-    with ThreadPoolExecutor(max_workers=num_workers) as executor:
-        futures = [executor.submit(process_sample, i, shared_lock) for i in range(num_samples)]
-        for future in tqdm(as_completed(futures), total=num_samples, desc="samples", leave=False):
-            result = future.result()
-            load_times_video_ms.append(result["load_time_video_ms"])
-            load_times_images_ms.append(result["load_time_images_ms"])
-            psnr_values.extend(result["psnr_values"])
-            ssim_values.extend(result["ssim_values"])
-            mse_values.extend(result["mse_values"])
-
-    avg_load_time_video_ms = float(np.array(load_times_video_ms).mean())
-    avg_load_time_images_ms = float(np.array(load_times_images_ms).mean())
-    video_images_load_time_ratio = avg_load_time_video_ms / avg_load_time_images_ms
-
-    return {
-        "avg_load_time_video_ms": avg_load_time_video_ms,
-        "avg_load_time_images_ms": avg_load_time_images_ms,
-        "video_images_load_time_ratio": video_images_load_time_ratio,
-        "avg_mse": float(np.mean(mse_values)),
-        "avg_psnr": float(np.mean(psnr_values)),
-        "avg_ssim": float(np.mean(ssim_values)),
-    }
-
-
-def benchmark_encoding_decoding(
-    dataset: LeRobotDataset,
-    video_path: Path,
-    imgs_dir: Path,
-    encoding_cfg: dict,
-    decoding_cfg: dict,
-    num_samples: int,
-    num_workers: int,
-    save_frames: bool,
-    overwrite: bool = False,
-    seed: int = 1337,
-) -> list[dict]:
-    fps = dataset.fps
-
-    if overwrite or not video_path.is_file():
-        tqdm.write(f"encoding {video_path}")
-        encode_video_frames(
-            imgs_dir=imgs_dir,
-            video_path=video_path,
-            fps=fps,
-            vcodec=encoding_cfg["vcodec"],
-            pix_fmt=encoding_cfg["pix_fmt"],
-            g=encoding_cfg.get("g"),
-            crf=encoding_cfg.get("crf"),
-            # fast_decode=encoding_cfg.get("fastdecode"),
-            overwrite=True,
-        )
-
-    episode_index = 0
-    ep_num_images = dataset.meta.episodes["length"][episode_index]
-    width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
-    num_pixels = width * height
-    video_size_bytes = video_path.stat().st_size
-    images_size_bytes = get_directory_size(imgs_dir)
-    video_images_size_ratio = video_size_bytes / images_size_bytes
-
-    random.seed(seed)
-    benchmark_table = []
-    for timestamps_mode in tqdm(
-        decoding_cfg["timestamps_modes"], desc="decodings (timestamps_modes)", leave=False
-    ):
-        for backend in tqdm(decoding_cfg["backends"], desc="decodings (backends)", leave=False):
-            benchmark_row = benchmark_decoding(
-                imgs_dir,
-                video_path,
-                timestamps_mode,
-                backend,
-                ep_num_images,
-                fps,
-                num_samples,
-                num_workers,
-                save_frames,
-            )
-            benchmark_row.update(
-                **{
-                    "repo_id": dataset.repo_id,
-                    "resolution": f"{width} x {height}",
-                    "num_pixels": num_pixels,
-                    "video_size_bytes": video_size_bytes,
-                    "images_size_bytes": images_size_bytes,
-                    "video_images_size_ratio": video_images_size_ratio,
-                    "timestamps_mode": timestamps_mode,
-                    "backend": backend,
-                },
-                **encoding_cfg,
-            )
-            benchmark_table.append(benchmark_row)
-
-    return benchmark_table
-
-
-def main(
-    output_dir: Path,
-    repo_ids: list[str],
-    vcodec: list[str],
-    pix_fmt: list[str],
-    g: list[int],
-    crf: list[int],
-    # fastdecode: list[int],
-    timestamps_modes: list[str],
-    backends: list[str],
-    num_samples: int,
-    num_workers: int,
-    save_frames: bool,
-):
-    check_datasets_formats(repo_ids)
-    encoding_benchmarks = {
-        "g": g,
-        "crf": crf,
-        # "fastdecode": fastdecode,
-    }
-    decoding_benchmarks = {
-        "timestamps_modes": timestamps_modes,
-        "backends": backends,
-    }
-    headers = ["repo_id", "resolution", "num_pixels"]
-    headers += list(BASE_ENCODING.keys())
-    headers += [
-        "timestamps_mode",
-        "backend",
-        "video_size_bytes",
-        "images_size_bytes",
-        "video_images_size_ratio",
-        "avg_load_time_video_ms",
-        "avg_load_time_images_ms",
-        "video_images_load_time_ratio",
-        "avg_mse",
-        "avg_psnr",
-        "avg_ssim",
-    ]
-    file_paths = []
-    for video_codec in tqdm(vcodec, desc="encodings (vcodec)"):
-        for pixel_format in tqdm(pix_fmt, desc="encodings (pix_fmt)", leave=False):
-            benchmark_table = []
-            for repo_id in tqdm(repo_ids, desc="encodings (datasets)", leave=False):
-                dataset = LeRobotDataset(repo_id)
-                imgs_dir = output_dir / "images" / dataset.repo_id.replace("/", "_")
-                # We only use the first episode
-                save_first_episode(imgs_dir, dataset)
-                for duet in [
-                    dict(zip(encoding_benchmarks.keys(), unique_combination, strict=False))
-                    for unique_combination in itertools.product(*encoding_benchmarks.values())
-                ]:
-                    encoding_cfg = BASE_ENCODING.copy()
-                    encoding_cfg["vcodec"] = video_codec
-                    encoding_cfg["pix_fmt"] = pixel_format
-                    for key, value in duet.items():
-                        encoding_cfg[key] = value
-                    args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
-                    video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
-                    benchmark_table += benchmark_encoding_decoding(
-                        dataset,
-                        video_path,
-                        imgs_dir,
-                        encoding_cfg,
-                        decoding_benchmarks,
-                        num_samples,
-                        num_workers,
-                        save_frames,
-                    )
-
-            # Save intermediate results
-            benchmark_df = pd.DataFrame(benchmark_table, columns=headers)
-            now = dt.datetime.now()
-            csv_path = (
-                output_dir
-                / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_{video_codec}_{pixel_format}_{num_samples}-samples.csv"
-            )
-            benchmark_df.to_csv(csv_path, header=True, index=False)
-            file_paths.append(csv_path)
-            del benchmark_df
-
-    # Concatenate all results
-    df_list = [pd.read_csv(csv_path) for csv_path in file_paths]
-    concatenated_df = pd.concat(df_list, ignore_index=True)
-    concatenated_path = output_dir / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_all_{num_samples}-samples.csv"
-    concatenated_df.to_csv(concatenated_path, header=True, index=False)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--output-dir",
-        type=Path,
-        default=Path("outputs/video_benchmark"),
-        help="Directory where the video benchmark outputs are written.",
-    )
-    parser.add_argument(
-        "--repo-ids",
-        type=str,
-        nargs="*",
-        default=[
-            "lerobot/pusht_image",
-            "lerobot/aloha_mobile_shrimp_image",
-            "lerobot/paris_street",
-            "lerobot/kitchen",
-        ],
-        help="Datasets repo-ids to test against. First episodes only are used. Must be images.",
-    )
-    parser.add_argument(
-        "--vcodec",
-        type=str,
-        nargs="*",
-        default=["h264", "hevc", "libsvtav1"],
-        help="Video codecs to be tested",
-    )
-    parser.add_argument(
-        "--pix-fmt",
-        type=str,
-        nargs="*",
-        default=["yuv444p", "yuv420p"],
-        help="Pixel formats (chroma subsampling) to be tested",
-    )
-    parser.add_argument(
-        "--g",
-        type=parse_int_or_none,
-        nargs="*",
-        default=[1, 2, 3, 4, 5, 6, 10, 15, 20, 40, 100, None],
-        help="Group of pictures sizes to be tested.",
-    )
-    parser.add_argument(
-        "--crf",
-        type=parse_int_or_none,
-        nargs="*",
-        default=[0, 5, 10, 15, 20, 25, 30, 40, 50, None],
-        help="Constant rate factors to be tested.",
-    )
-    # parser.add_argument(
-    #     "--fastdecode",
-    #     type=int,
-    #     nargs="*",
-    #     default=[0, 1],
-    #     help="Use the fastdecode tuning option. 0 disables it. "
-    #         "For libx264 and libx265/hevc, only 1 is possible. "
-    #         "For libsvtav1, 1, 2 or 3 are possible values with a higher number meaning a faster decoding optimization",
-    # )
-    parser.add_argument(
-        "--timestamps-modes",
-        type=str,
-        nargs="*",
-        default=[
-            "1_frame",
-            "2_frames",
-            "2_frames_4_space",
-            "6_frames",
-        ],
-        help="Timestamps scenarios to be tested.",
-    )
-    parser.add_argument(
-        "--backends",
-        type=str,
-        nargs="*",
-        default=["torchcodec", "pyav"],
-        help="Torchvision decoding backend to be tested.",
-    )
-    parser.add_argument(
-        "--num-samples",
-        type=int,
-        default=50,
-        help="Number of samples for each encoding x decoding config.",
-    )
-    parser.add_argument(
-        "--num-workers",
-        type=int,
-        default=10,
-        help="Number of processes for parallelized sample processing.",
-    )
-    parser.add_argument(
-        "--save-frames",
-        type=int,
-        default=0,
-        help="Whether to save decoded frames or not. Enter a non-zero number for true.",
-    )
-    args = parser.parse_args()
-    main(**vars(args))

docker/Dockerfile.benchmark.robotwin

@@ -35,7 +35,7 @@ USER root
 ARG ROBOTWIN_SHA=0aeea2d669c0f8516f4d5785f0aa33ba812c14b4
 RUN apt-get update \
     && apt-get install -y --no-install-recommends \
-         cuda-nvcc-12-6 cuda-cudart-dev-12-6 \
+         cuda-nvcc-12-8 cuda-cudart-dev-12-8 \
          libvulkan1 vulkan-tools \
     && mkdir -p /usr/share/vulkan/icd.d \
     && echo '{"file_format_version":"1.0.0","ICD":{"library_path":"libGLX_nvidia.so.0","api_version":"1.3.0"}}' \

docker/Dockerfile.internal

@@ -18,7 +18,7 @@
 # docker build -f docker/Dockerfile.internal -t lerobot-internal .
 
 # Configure the base image for CI with GPU access
-ARG CUDA_VERSION=12.6.3
+ARG CUDA_VERSION=12.8.1
 ARG OS_VERSION=24.04
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu${OS_VERSION}
 

docs/source/_toctree.yml

@@ -3,12 +3,14 @@
     title: LeRobot
   - local: installation
     title: Installation
+  - local: cheat-sheet
+    title: Cheat sheet
   title: Get started
 - sections:
   - local: il_robots
     title: Imitation Learning for Robots
   - local: bring_your_own_policies
-    title: Bring Your Own Policies
+    title: Adding a Policy
   - local: integrate_hardware
     title: Bring Your Own Hardware
   - local: hilserl
@@ -24,6 +26,12 @@
   - local: rename_map
     title: Using Rename Map and Empty Cameras
   title: "Tutorials"
+- sections:
+  - local: hardware_guide
+    title: Compute Hardware Guide
+  - local: torch_accelerators
+    title: PyTorch accelerators
+  title: "Compute & Hardware"
 - sections:
   - local: lerobot-dataset-v3
     title: Using LeRobotDataset
@@ -31,8 +39,12 @@
     title: Porting Large Datasets
   - local: using_dataset_tools
     title: Using the Dataset Tools
-  - local: dataset_subtask
-    title: Using Subtasks in the Dataset
+  - local: language_and_recipes
+    title: Language Columns and Recipes
+  - local: tools
+    title: Tools
+  - local: video_encoding_parameters
+    title: Video encoding parameters
   - local: streaming_video_encoding
     title: Streaming Video Encoding
   title: "Datasets"
@@ -61,6 +73,8 @@
 - sections:
   - local: sarm
     title: SARM
+  - local: topreward
+    title: TOPReward
   title: "Reward Models"
 - sections:
   - local: inference
@@ -133,6 +147,8 @@
     title: OMX
   - local: openarm
     title: OpenArm
+  - local: rebot_b601
+    title: reBot B601-DM
   title: "Robots"
 - sections:
   - local: phone_teleop
@@ -142,10 +158,6 @@
   - local: cameras
     title: Cameras
   title: "Sensors"
-- sections:
-  - local: torch_accelerators
-    title: PyTorch accelerators
-  title: "Supported Hardware"
 - sections:
   - local: notebooks
     title: Notebooks

docs/source/act.mdx

@@ -90,6 +90,6 @@ lerobot-record \
   --dataset.single_task="Your task description" \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   --policy.path=${HF_USER}/act_policy
 ```

docs/source/bring_your_own_policies.mdx

@@ -1,60 +1,37 @@
-# Bring Your Own Policies
+# Adding a Policy
 
-This tutorial explains how to integrate your own custom policy implementations into the LeRobot ecosystem, allowing you to leverage all LeRobot tools for training, evaluation, and deployment while using your own algorithms.
+This guide walks you through implementing a custom policy and getting it to work with LeRobot's training, evaluation, and deployment tools. There are two paths:
 
-## Step 1: Create a Policy Package
+- **Plugin (out-of-tree)** — ship your policy as a standalone `lerobot_policy_*` package. Faster, no PR required, easy to iterate. Right for experimentation, internal use, or when you want to publish independently.
+- **In-tree (contributed to LeRobot)** — land your policy directly in `src/lerobot/policies/`. Requires a PR, but makes your policy a first-class citizen of the library.
 
-Your custom policy should be organized as an installable Python package following LeRobot's plugin conventions.
+The plugin route is usually the right starting point — promote to in-tree once the policy has stabilized and there's clear value in shipping it with the library.
 
-### Package Structure
+Either way, the building blocks are the same: a configuration class, a policy class, and a processor factory. The first half of this guide covers those shared pieces; the second half covers the path-specific scaffolding ([Path A](#path-a-out-of-tree-plugin), [Path B](#path-b-contributing-in-tree)).
 
-Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
+A note on tone: robot-learning is an actively evolving field, and "what a policy looks like" can shift with each new architecture. The conventions described here exist because they let `lerobot-train` and `lerobot-eval` work uniformly across very different models. When a new policy genuinely doesn't fit them, raise it (in your PR, or an issue) — the conventions are not sacred.
 
-```bash
-lerobot_policy_my_custom_policy/
-├── pyproject.toml
-└── src/
-    └── lerobot_policy_my_custom_policy/
-        ├── __init__.py
-        ├── configuration_my_custom_policy.py
-        ├── modeling_my_custom_policy.py
-        └── processor_my_custom_policy.py
-```
+---
 
-### Package Configuration
+## Anatomy of a policy
 
-Set up your `pyproject.toml`:
+Three building blocks make up every policy. The names below use `my_policy` as a placeholder — replace with your policy's name. That name is load-bearing: it must match the string you pass to `@PreTrainedConfig.register_subclass`, the `MyPolicy.name` class attribute, and the `make_<name>_pre_post_processors` factory function (more on each below).
 
-```toml
-[project]
-name = "lerobot_policy_my_custom_policy"
-version = "0.1.0"
-dependencies = [
-    # your policy-specific dependencies
-]
-requires-python = ">= 3.12"
+### Configuration class
 
-[build-system]
-build-backend = # your-build-backend
-requires = # your-build-system
-```
-
-## Step 2: Define the Policy Configuration
-
-Create a configuration class that inherits from [`PreTrainedConfig`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/policies.py) and registers your policy type:
-Here is a template to get you started, customize the parameters and methods as needed for your policy's architecture and training requirements.
+Inherit from [`PreTrainedConfig`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/configs/policies.py) and register your policy type. Here is a template — customize the parameters and methods as needed for your policy's architecture and training requirements.
 
 ```python
-# configuration_my_custom_policy.py
+# configuration_my_policy.py
 from dataclasses import dataclass, field
 from lerobot.configs import PreTrainedConfig
 from lerobot.optim import AdamWConfig
 from lerobot.optim import CosineDecayWithWarmupSchedulerConfig
 
-@PreTrainedConfig.register_subclass("my_custom_policy")
+@PreTrainedConfig.register_subclass("my_policy")
 @dataclass
-class MyCustomPolicyConfig(PreTrainedConfig):
-    """Configuration class for MyCustomPolicy.
+class MyPolicyConfig(PreTrainedConfig):
+    """Configuration class for MyPolicy.
 
     Args:
         n_obs_steps: Number of observation steps to use as input
@@ -77,16 +54,20 @@ class MyCustomPolicyConfig(PreTrainedConfig):
             raise ValueError("n_action_steps cannot exceed horizon")
 
     def validate_features(self) -> None:
-        """Validate input/output feature compatibility."""
+        """Validate input/output feature compatibility.
+
+        Call this explicitly from your policy's __init__ — the base class does not.
+        """
         if not self.image_features:
-            raise ValueError("MyCustomPolicy requires at least one image feature.")
+            raise ValueError("MyPolicy requires at least one image feature.")
         if self.action_feature is None:
-            raise ValueError("MyCustomPolicy requires 'action' in output_features.")
+            raise ValueError("MyPolicy requires 'action' in output_features.")
 
     def get_optimizer_preset(self) -> AdamWConfig:
         return AdamWConfig(lr=self.optimizer_lr, weight_decay=self.optimizer_weight_decay)
 
     def get_scheduler_preset(self):
+        """Return a LRSchedulerConfig from lerobot.optim, or None."""
         return None
 
     @property
@@ -101,8 +82,7 @@ class MyCustomPolicyConfig(PreTrainedConfig):
 
     @property
     def action_delta_indices(self) -> list[int]:
-        """Relative timestep offsets for the action chunk the dataset loader returns.
-        """
+        """Relative timestep offsets for the action chunk the dataset loader returns."""
         return list(range(self.horizon))
 
     @property
@@ -110,32 +90,34 @@ class MyCustomPolicyConfig(PreTrainedConfig):
         return None
 ```
 
-## Step 3: Implement the Policy Class
+The string you pass to `@register_subclass` must match `MyPolicy.name` (next section) and is what users supply as `--policy.type` on the CLI. Default to `AdamW` from `lerobot.optim` for `get_optimizer_preset` unless you genuinely need otherwise.
 
-Create your policy implementation by inheriting from [`PreTrainedPolicy`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pretrained.py):
+### Policy class
+
+Inherit from [`PreTrainedPolicy`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pretrained.py) and set two class attributes — both are checked by `__init_subclass__`:
 
 ```python
-# modeling_my_custom_policy.py
+# modeling_my_policy.py
 import torch
 import torch.nn as nn
 from typing import Any
 
 from lerobot.policies import PreTrainedPolicy
 from lerobot.utils.constants import ACTION
-from .configuration_my_custom_policy import MyCustomPolicyConfig
+from .configuration_my_policy import MyPolicyConfig
 
-class MyCustomPolicy(PreTrainedPolicy):
-    config_class = MyCustomPolicyConfig  # must match the string in @register_subclass
-    name = "my_custom_policy"
+class MyPolicy(PreTrainedPolicy):
+    config_class = MyPolicyConfig  # must match the string in @register_subclass
+    name = "my_policy"
 
-    def __init__(self, config: MyCustomPolicyConfig, dataset_stats: dict[str, Any] = None):
+    def __init__(self, config: MyPolicyConfig, dataset_stats: dict[str, Any] = None):
         super().__init__(config, dataset_stats)
         config.validate_features()  # not called automatically by the base class
         self.config = config
         self.model = ...  # your nn.Module here
 
     def reset(self):
-        """Reset episode state."""
+        """Reset per-episode state. Called by lerobot-eval at the start of each episode."""
         ...
 
     def get_optim_params(self) -> dict:
@@ -147,35 +129,51 @@ class MyCustomPolicy(PreTrainedPolicy):
         ...
 
     def select_action(self, batch: dict[str, torch.Tensor], **kwargs) -> torch.Tensor:
-        """Return a single action for the current timestep (called at inference)."""
+        """Return a single action for the current timestep (called every step at inference)."""
         ...
 
-    def forward(self, batch: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
+    def forward(self, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, dict | None]:
         """Compute the training loss.
 
+        Returns `(loss, output_dict)`. `output_dict` may be `None`; everything in it must be
+        logging-friendly Python natives (no tensors with gradients).
+
         `batch["action_is_pad"]` is a bool mask of shape (B, horizon) that marks
-        timesteps padded because the episode ended before `horizon` steps, you
+        timesteps padded because the episode ended before `horizon` steps; you
         can exclude those from your loss.
         """
         actions = batch[ACTION]
         action_is_pad = batch.get("action_is_pad")
         ...
-        return {"loss": ...}
+        return loss, {"some_loss_component": some_loss_component.item()}
 ```
 
-## Step 4: Add Data Processors
+The methods called by the train/eval loops:
 
-Create processor functions. For a concrete reference, see [processor_act.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [processor_diffusion.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
+| Method                                                            | Used by           | What it does                                                                                                                                                                                                                                         |
+| ----------------------------------------------------------------- | ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `reset() -> None`                                                 | `lerobot-eval`    | Clear per-episode state at the start of each episode.                                                                                                                                                                                                |
+| `select_action(batch, **kwargs) -> Tensor`                        | `lerobot-eval`    | Return the next action `(B, action_dim)`. Called every step.                                                                                                                                                                                         |
+| `predict_action_chunk(batch, **kwargs) -> Tensor`                 | the policy itself | Return an action chunk `(B, chunk_size, action_dim)`. Currently abstract on the base class — raise `NotImplementedError` if your policy doesn't chunk.                                                                                               |
+| `forward(batch, reduction="mean") -> tuple[Tensor, dict \| None]` | `lerobot-train`   | Return `(loss, output_dict)`. Accept `reduction="none"` if you want to support per-sample weighting.                                                                                                                                                 |
+| `get_optim_params() -> dict`                                      | the optimizer     | Return `self.parameters()` for simple policies; return a named parameter dict for [multi-optimizer policies](https://github.com/huggingface/lerobot/blob/ecd38c50d7d15b4184cf42649ff1185ee2e11eeb/src/lerobot/policies/sac/modeling_sac.py#L61-L73). |
+| `update() -> None` _(optional)_                                   | `lerobot-train`   | Called after each optimizer step _if defined_. Use for EMA, target nets, replay buffers (TDMPC uses this).                                                                                                                                           |
+
+Batches are flat dictionaries keyed by the constants in [`lerobot.utils.constants`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/utils/constants.py): `OBS_STATE` (`observation.state.<motor>`), `OBS_IMAGES` (`observation.images.<camera>`), `OBS_LANGUAGE`, `ACTION`, etc. Reuse the constants — don't invent new prefixes.
+
+### Processor functions
+
+LeRobot uses `PolicyProcessorPipeline`s to normalize inputs and de-normalize outputs around your policy. For a concrete reference, see [`processor_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/processor_act.py) or [`processor_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/processor_diffusion.py).
 
 ```python
-# processor_my_custom_policy.py
+# processor_my_policy.py
 from typing import Any
 import torch
 
 from lerobot.processor import PolicyAction, PolicyProcessorPipeline
 
 
-def make_my_custom_policy_pre_post_processors(
+def make_my_policy_pre_post_processors(
     config,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
 ) -> tuple[
@@ -187,11 +185,48 @@ def make_my_custom_policy_pre_post_processors(
     return preprocessor, postprocessor
 ```
 
-**Important - function naming:** LeRobot discovers your processor by name. The function **must** be called `make_{policy_name}_pre_post_processors` (matching the string you passed to `@PreTrainedConfig.register_subclass`).
+**Important — function naming:** LeRobot discovers your processor by name. The function **must** be called `make_{policy_name}_pre_post_processors` (matching the string you passed to `@PreTrainedConfig.register_subclass`).
 
-## Step 5: Package Initialization
+---
 
-Expose your classes in the package's `__init__.py`:
+## Path A: Out-of-tree plugin
+
+The fastest way to ship a policy: package it as a standalone Python distribution and install it alongside LeRobot. No PR required, you own the release cycle, and you can publish to PyPI under your own namespace.
+
+### Package structure
+
+Create a package with the prefix `lerobot_policy_` (IMPORTANT!) followed by your policy name:
+
+```bash
+lerobot_policy_my_policy/
+├── pyproject.toml
+└── src/
+    └── lerobot_policy_my_policy/
+        ├── __init__.py
+        ├── configuration_my_policy.py
+        ├── modeling_my_policy.py
+        └── processor_my_policy.py
+```
+
+### `pyproject.toml`
+
+```toml
+[project]
+name = "lerobot_policy_my_policy"
+version = "0.1.0"
+dependencies = [
+    # your policy-specific dependencies
+]
+requires-python = ">= 3.12"
+
+[build-system]
+build-backend = # your-build-backend
+requires = # your-build-system
+```
+
+### Package `__init__.py`
+
+Expose your classes in the package's `__init__.py` and guard against missing `lerobot`:
 
 ```python
 # __init__.py
@@ -204,44 +239,148 @@ except ImportError:
         "lerobot is not installed. Please install lerobot to use this policy package."
     )
 
-from .configuration_my_custom_policy import MyCustomPolicyConfig
-from .modeling_my_custom_policy import MyCustomPolicy
-from .processor_my_custom_policy import make_my_custom_policy_pre_post_processors
+from .configuration_my_policy import MyPolicyConfig
+from .modeling_my_policy import MyPolicy
+from .processor_my_policy import make_my_policy_pre_post_processors
 
 __all__ = [
-    "MyCustomPolicyConfig",
-    "MyCustomPolicy",
-    "make_my_custom_policy_pre_post_processors",
+    "MyPolicyConfig",
+    "MyPolicy",
+    "make_my_policy_pre_post_processors",
 ]
 ```
 
-## Step 6: Installation and Usage
-
-### Install Your Policy Package
+### Install and use
 
 ```bash
-cd lerobot_policy_my_custom_policy
+cd lerobot_policy_my_policy
 pip install -e .
 
 # Or install from PyPI if published
-pip install lerobot_policy_my_custom_policy
+pip install lerobot_policy_my_policy
 ```
 
-### Use Your Policy
-
 Once installed, your policy automatically integrates with LeRobot's training and evaluation tools:
 
 ```bash
 lerobot-train \
-    --policy.type my_custom_policy \
+    --policy.type my_policy \
     --env.type pusht \
     --steps 200000
 ```
 
-## Examples and Community Contributions
+---
+
+## Path B: Contributing in-tree
+
+When your policy has stabilized and there's clear value in shipping it with the library, you can land it directly in LeRobot. Read the general [contribution guide](./contributing) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md) first — that's where you'll find the testing/quality expectations every PR has to meet (`pre-commit run -a`, `pytest`, the community-review rule, etc.). What's below is the policy-specific layer on top of that.
+
+### In-tree layout
+
+```
+src/lerobot/policies/my_policy/
+├── __init__.py                    # re-exports config + modeling + processor factory
+├── configuration_my_policy.py     # MyPolicyConfig + @register_subclass
+├── modeling_my_policy.py          # MyPolicy(PreTrainedPolicy)
+├── processor_my_policy.py         # make_my_policy_pre_post_processors
+└── README.md                      # symlink → ../../../../docs/source/policy_my_policy_README.md
+```
+
+Two notes:
+
+- The `README.md` next to the source is a **symlink** into `docs/source/policy_<name>_README.md` — the actual file lives under `docs/`. Existing policies (act, smolvla, diffusion, …) all do this; copy one of those symlinks. The policy README is conventionally minimal: paper link + BibTeX citation.
+- The user-facing tutorial — what to install, how to train, hyperparameters, benchmark numbers — lives separately at `docs/source/<my_policy>.mdx` and is registered in `_toctree.yml` under "Policies".
+
+The file names are load-bearing: the factory does lazy imports by name, and the processor is discovered by the `make_<policy_name>_pre_post_processors` convention.
+
+### Wiring
+
+Three places need to know about your policy. All by name.
+
+1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
+2. **`factory.py:get_policy_class`** — add a branch returning `MyPolicy` from a lazy import.
+3. **`factory.py:make_policy_config`** and **`factory.py:make_pre_post_processors`** — same idea, two more branches.
+
+Mirror an existing policy that's structurally similar to yours; the diff is small.
+
+### Heavy / optional dependencies
+
+Most policies need a heavy backbone (transformers, diffusers, a specific VLM SDK). The convention is **two-step gating**: a `TYPE_CHECKING`-guarded import at module top, and a `require_package` runtime check in the constructor. [`modeling_diffusion.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/diffusion/modeling_diffusion.py) is the canonical reference:
+
+```python
+from typing import TYPE_CHECKING
+from lerobot.utils.import_utils import _diffusers_available, require_package
+
+if TYPE_CHECKING or _diffusers_available:
+    from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+else:
+    DDIMScheduler = None  # keeps the symbol bindable at import time
+
+class DiffusionPolicy(PreTrainedPolicy):
+    def __init__(self, config):
+        require_package("diffusers", extra="diffusion")
+        super().__init__(config)
+        ...
+```
+
+This way:
+
+- `import lerobot.policies` keeps working without the extra installed (the symbol is just bound to `None`).
+- Type checkers see the real symbol.
+- Instantiating the policy without the extra raises a clear `ImportError` pointing at `pip install 'lerobot[diffusion]'`.
+
+Add a matching extra to [`pyproject.toml`](https://github.com/huggingface/lerobot/blob/main/pyproject.toml) `[project.optional-dependencies]` and include it in the `all` extra so `pip install 'lerobot[all]'` keeps installing everything.
+
+### Benchmarks and a published checkpoint
+
+A new policy is much easier to review — and far more useful — when it ships with a working checkpoint and at least one number you can reproduce.
+
+**Pick at least one in-tree benchmark.** LeRobot ships sim benchmarks with per-benchmark Docker images (LIBERO, LIBERO-plus, Meta-World, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench and more). Pick the one that matches your policy's modality — VLAs usually go to LIBERO or VLABench; image-only BC to LIBERO or Meta-World. The full list lives under [Benchmarks](./libero) in the docs sidebar.
+
+**Push the checkpoint & processors** to the Hub under `lerobot/<policy>_<benchmark>` (or your namespace if you don't have write access; a maintainer can mirror it). Use `PreTrainedPolicy.push_model_to_hub` so the repo gets `config.json`, `model.safetensors`, and a model card.
+
+**Report results in your policy's MDX**, with the exact `lerobot-eval` command and hardware so anyone can re-run:
+
+```markdown
+## Results
+
+Evaluated on LIBERO with `lerobot/<policy>_libero`:
+
+| Suite          | Success rate | n_episodes |
+| -------------- | -----------: | ---------: |
+| libero_spatial |        87.5% |         50 |
+| libero_object  |        93.0% |         50 |
+| libero_goal    |        81.5% |         50 |
+| libero_10      |        62.0% |         50 |
+| **average**    |    **81.0%** |        200 |
+
+Reproduce: `lerobot-eval --policy.path=lerobot/<policy>_libero --env.type=libero --env.task=libero_spatial --eval.n_episodes=50` (1× A100 40 GB).
+```
+
+Use `n_episodes ≥ 50` per suite for stable success-rate estimates.
+
+If your policy is real-robot-only and no sim benchmark applies, swap the sim eval for: a public training dataset on the Hub, the `lerobot-train` command, the checkpoint, and a real-robot success rate over ≥10 episodes via `lerobot-rollout --policy.path=...`.
+
+### PR checklist
+
+The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md). On top of those, reviewers will look for:
+
+- [ ] `MyPolicy` and `MyPolicyConfig` cover the surface above; `__init_subclass__` accepts the class.
+- [ ] `factory.py` and `policies/__init__.py` are wired (lazy imports for modeling).
+- [ ] `make_my_policy_pre_post_processors` follows the naming convention.
+- [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard.
+- [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests.
+- [ ] `src/lerobot/policies/<name>/README.md` symlinked into `docs/source/policy_<name>_README.md`; user-facing `docs/source/<name>.mdx` written and added to `_toctree.yml`.
+- [ ] At least one reproducible benchmark eval in the policy MDX with a published checkpoint (sim benchmark, or real-robot dataset + checkpoint).
+
+The fastest way to get a clean PR is to copy the directory of the existing policy closest to yours, rename, and replace contents method by method. Don't wait until everything is polished — open a draft PR early and iterate with us; reviewers would much rather give feedback on a half-finished branch than a fully-merged one.
+
+---
+
+## Examples and community contributions
 
 Check out these example policy implementations:
 
-- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) - Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)
+- [DiTFlow Policy](https://github.com/danielsanjosepro/lerobot_policy_ditflow) — Diffusion Transformer policy with flow-matching objective. Try it out in this example: [DiTFlow Example](https://github.com/danielsanjosepro/test_lerobot_policy_ditflow)
 
-Share your policy implementations with the community! 🤗
+Thanks for taking the time to bring a new policy into LeRobot. Every architecture that lands in `main` — and every plugin published by the community — makes the library a little more useful for the next person, and a little more representative of where robot learning is going. We're looking forward to seeing what you ship. 🤗

docs/source/cheat-sheet.mdx

@@ -0,0 +1,139 @@
+# Cheat sheet
+
+All of the LeRobot commands in one place. If you forgot how to use a specific command or want to learn about a new one you can do it here.
+
+> [!WARNING]
+> For all of the commands listed below remember to change the ports/names/ids to your own values!
+
+> [!TIP]
+> Another great way to look at all the commands and get them configured for your specific setup is to use this [Jupyter Notebook](https://github.com/huggingface/lerobot/blob/main/examples/notebooks/quickstart.ipynb).
+
+### Setup and installation
+
+For installation please look at [LeRobot Installation](https://huggingface.co/docs/lerobot/main/en/installation).
+
+### Useful tools
+
+###### Find port
+
+Use this to identify which serial ports your robots are connected to. Follow the instructions in your terminal: you will be asked to unplug the USB cable and press Enter. The script will then detect and print the correct serial port for that robot.
+
+```bash
+lerobot-find-port
+```
+
+###### Find cameras
+
+Quickly find camera indices and verify their output. This command prints camera information to the terminal and saves test frames from each detected camera to `lerobot/outputs/captured_images`
+
+```bash
+lerobot-find-cameras
+```
+
+### Calibration
+
+In most cases you will need to perform calibration just once for each robot and teleoperation device. Before performing the calibration make sure that all the joints are roughly in the middle position.
+
+```bash
+lerobot-calibrate \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=my_follower_arm
+```
+
+Make sure that you use the same IDs used during calibration later for the other scripts. That's how LeRobot finds the calibration files.
+
+### Teleoperation
+
+Teleoperating with two cameras and displaying the data with Rerun.
+
+```bash
+lerobot-teleoperate \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=my_follower_arm \
+    --robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --teleop.id=my_leader_arm \
+    --display_data=true
+```
+
+### Recording a dataset
+
+The dataset is automatically uploaded to the server and saved under repo_id, make sure you are logged in to your HF account with CLI:
+`hf auth login`
+
+You can get the token from: [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)
+
+```bash
+lerobot-record \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=my_follower_arm \
+    --robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
+    --teleop.type=so101_leader \
+    --teleop.port=/dev/ttyACM1 \
+    --teleop.id=my_leader_arm \
+    --dataset.repo_id=${HF_USER}/so101_dataset_test \
+    --dataset.num_episodes=30 \
+    --dataset.single_task="put the red brick in a bowl" \
+    --dataset.streaming_encoding=true \
+    --display_data=true
+```
+
+While collecting the dataset you can control the process with your keyboard:
+Control the data recording flow using keyboard shortcuts:
+
+- Press **Right Arrow (`→`)**: Save episode and move to the next.
+- Press **Left Arrow (`←`)**: Delete current episode and retry.
+- Press **Escape (`ESC`)**: Stop, encode videos, and upload.
+
+### Training
+
+Depending on your hardware training the policy might take a few hours. That's how you train simple `ACT` policy:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=${HF_USER}/so101_dataset_test \
+    --policy.type=act \
+    --output_dir=outputs/train/act_so101_test \
+    --job_name=act_so101_test \
+    --policy.device=cuda \
+    --wandb.enable=true \
+    --policy.repo_id=${HF_USER}/policy_test \
+    --steps=20000
+```
+
+- Policy Types: `act`, `diffusion`, `smolvla`, `pi05`
+- Devices: `cuda` (NVIDIA), `mps` (Apple Silicon), `cpu`
+
+If you want to fine-tune a specific model you can provide the path to the model. In this case path is enough and type can be skipped.
+
+```bash
+lerobot-train \
+    --dataset.repo_id=${HF_USER}/so101_dataset_test \
+    --policy.path=username/the_policy_to_finetune \
+    --policy.device=cuda \
+    --policy.repo_id=${HF_USER}/policy_test \
+    --output_dir=outputs/train/act_so101_test \
+    --steps=20000
+```
+
+### Inference
+
+Inference means running the trained policy/model on a robot. For that we use `lerobot-rollout`. You will need to provide a path to your policy. It can be a local path or a path to Hugging Face for example "lerobot/folding_latest". Your cameras configuration needs to match what was used when collecting the dataset. Duration is in seconds if unspecified, it will run forever.
+
+> [!TIP]
+> If you are using the previous release V0.5.1 instead of `lerobot-rollout` you need to use `lerobot-record`. More information [here](https://huggingface.co/docs/lerobot/v0.5.1/en/il_robots#run-inference-and-evaluate-your-policy).
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --policy.path=${HF_USER}/my_policy \
+    --robot.type=so101_follower \
+    --robot.port=/dev/ttyACM1 \
+    --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video1, width: 640, height: 480, fps: 30}, side: {type: opencv, index_or_path: /dev/video5, width: 640, height: 480, fps: 30}}" \
+    --task="Put lego brick into the transparent box" \
+    --duration=60
+```

docs/source/dataset_subtask.mdx

@@ -1,277 +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 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 import TokenizerProcessorStep
-
-# Create a tokenizer processor step
-tokenizer_processor = TokenizerProcessorStep(
-    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 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 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

@@ -194,7 +194,7 @@ lerobot-record \
     --dataset.single_task="Navigate around obstacles" \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --display_data=true
 ```
 

docs/source/groot.mdx

@@ -123,7 +123,7 @@ lerobot-record \
   --dataset.single_task="Grab and handover the red cube to the other arm" \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   --policy.path=<user>/groot-bimanual \ # your trained model
   --dataset.episode_time_s=30 \
   --dataset.reset_time_s=10

docs/source/hardware_guide.mdx

@@ -0,0 +1,98 @@
+# Compute HW Guide for LeRobot Training
+
+Rough sizing for training a LeRobot policy: how much VRAM each policy needs, what training time looks like, and where to run when local hardware isn't enough.
+
+The numbers below are **indicative** — order-of-magnitude figures for picking hardware, not exact predictions. Throughput depends heavily on dataset I/O, image resolution, batch size, and number of GPUs.
+
+## Memory by policy group
+
+Policies cluster by backbone size; the groupings below give a single VRAM envelope per group instead of repeating numbers per policy. Memory scales roughly linearly with batch size; AdamW (the LeRobot default) carries optimizer state that adds ~30–100% over a forward+backward pass alone.
+
+| Group      | Policies                                    | Peak VRAM (BS 8, AdamW) | Suitable starter GPUs             |
+| ---------- | ------------------------------------------- | ----------------------: | --------------------------------- |
+| Light BC   | `act`, `vqbet`, `tdmpc`                     |                  ~2–6GB | Laptop GPU (RTX 3060), L4, A10G   |
+| Diffusion  | `diffusion`, `multi_task_dit`               |                 ~8–14GB | RTX 4070+ / L4 / A10G             |
+| Small VLA  | `smolvla`                                   |                ~10–16GB | RTX 4080+ / L4 / A10G             |
+| Large VLA  | `pi0`, `pi0_fast`, `pi05`, `xvla`, `wall_x` |                ~24–40GB | A100 40 GB+ (24 GB tight at BS 1) |
+| Multimodal | `groot`, `eo1`                              |                ~24–40GB | A100 40 GB+                       |
+| RL         | `sac`                                       |             config-dep. | See [HIL-SERL guide](./hilserl)   |
+
+Memory-bound? Drop the batch size (~linear), use gradient accumulation to recover effective batch, or for SmolVLA leave `freeze_vision_encoder=True`.
+
+## Training time
+
+Robotics imitation learning typically converges in **5–10 epochs over the dataset**, not hundreds of thousands of raw steps. Once you know your epoch count, wall-clock is essentially:
+
+```text
+total_frames    = sum of frames over all episodes      # 50 ep × 30 fps × 30 s ≈ 45,000
+steps_per_epoch = ceil(total_frames / (num_gpus × batch_size))
+total_steps     = epochs × steps_per_epoch
+wall_clock      ≈ total_steps × per_step_time
+```
+
+Per-step time depends on the policy and the GPU. The numbers in the table below are anchors — pick the row closest to your setup and scale linearly with `total_steps` if you train longer or shorter.
+
+### Common scenarios
+
+Indicative wall-clock for **5 epochs on a ~50-episode dataset (~45k frames at 30 fps × 30 s)**, default optimizer (AdamW), 640×480 images:
+
+| Setup                                | Policy         | Batch | Wall-clock |
+| ------------------------------------ | -------------- | ----- | ---------: |
+| Single RTX 4090 / RTX 3090 (24 GB)   | `act`          | 8     |  ~30–60min |
+| Single RTX 4090 / RTX 3090 (24 GB)   | `diffusion`    | 8     |      ~2–4h |
+| Single L4 / A10G (24 GB)             | `act`          | 8     |      ~1–2h |
+| Single L4 / A10G (24 GB)             | `smolvla`      | 4     |      ~3–6h |
+| Single A100 40 GB                    | `smolvla`      | 16    |      ~1–2h |
+| Single A100 40 GB                    | `pi0` / `pi05` | 4     |      ~4–8h |
+| 4× H100 80 GB cluster (`accelerate`) | `diffusion`    | 32    |  ~30–60min |
+| 4× H100 80 GB cluster (`accelerate`) | `smolvla`      | 32    |      ~1–2h |
+| Apple Silicon M1/M2/M3 Max (MPS)     | `act`          | 4     |     ~6–14h |
+
+These are order-of-magnitude figures. Real runs deviate by ±50% depending on image resolution, dataset I/O, dataloader threading, and exact GPU SKU. They are useful as "is this run going to take an hour or a day?" intuition, not as SLAs.
+
+### Multi-GPU matters a lot
+
+`accelerate launch --num_processes=N` is the easiest way to cut training time. Each optimizer step processes `N × batch_size` samples in roughly the same wall-clock as a single-GPU step, so 4 GPUs ≈ 4× speedup for compute-bound runs. See the [Multi GPU training](./multi_gpu_training) guide for the full setup.
+
+Reference data points on a 4×H100 80 GB cluster (`accelerate launch --num_processes=4`), 5000 steps, batch 32, AdamW, dataset [`imstevenpmwork/super_poulain_draft`](https://huggingface.co/datasets/imstevenpmwork/super_poulain_draft) (~50 episodes, ~640×480 images):
+
+| Policy      | Wall-clock | `update_s` | `dataloading_s` | GPU util | Notable flags                                                                                                                  |
+| ----------- | ---------- | ---------: | --------------: | -------- | ------------------------------------------------------------------------------------------------------------------------------ |
+| `diffusion` | 16m 17s    |      0.167 |           0.015 | ~90%     | defaults (training from scratch)                                                                                               |
+| `smolvla`   | 27m 49s    |      0.312 |           0.011 | ~80%     | `--policy.path=lerobot/smolvla_base`, `freeze_vision_encoder=false`, `train_expert_only=false`                                 |
+| `pi05`      | 3h 41m     |      2.548 |           0.014 | ~95%     | `--policy.pretrained_path=lerobot/pi05_base`, `gradient_checkpointing=true`, `dtype=bfloat16`, vision encoder + expert trained |
+
+The `dataloading_s` vs. `update_s` ratio is the diagnostic that matters: when `dataloading_s` approaches `update_s`, more GPUs stop helping — your dataloader is the bottleneck and you should look at `--num_workers`, image resolution, and disk speed before adding compute.
+
+### Schedule and checkpoints
+
+If you shorten training (e.g. 5k–10k steps on a small dataset), also shorten the LR schedule with `--policy.scheduler_decay_steps≈--steps`. Otherwise the LR stays near its peak and never decays. Same for `--save_freq`.
+
+## Where to run
+
+VRAM is the first filter. Within a tier, pick by budget and availability — the `$`–`$$$$` columns are relative; check current pricing on the provider you actually use.
+
+| Class                      | VRAM  | Tier   | Comfortable for                                             |
+| -------------------------- | ----- | ------ | ----------------------------------------------------------- |
+| RTX 3090 / 4090 (consumer) | 24 GB | `$`    | Light BC, Diffusion, SmolVLA. Tight for VLAs at batch 1.    |
+| L4 / A10G (cloud)          | 24 GB | `$–$$` | Same envelope; common on Google Cloud, RunPod, AWS `g5/g6`. |
+| A100 40 GB                 | 40 GB | `$$$`  | Any policy at reasonable batch sizes.                       |
+| A100 80 GB / H100 80 GB    | 80 GB | `$$$$` | Multi-GPU clusters; large batches for VLAs.                 |
+| **CPU only**               | —     | —      | Don't train. Use Colab or rent a GPU.                       |
+
+### Hugging Face Jobs
+
+[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second. The repo publishes a ready-to-use image: **`huggingface/lerobot-gpu:latest`**, rebuilt **every night at 02:00 UTC from `main`** ([`docker_publish.yml`](https://github.com/huggingface/lerobot/blob/main/.github/workflows/docker_publish.yml)) — so it tracks the current state of the repo, not a tagged release.
+
+```bash
+hf jobs run --flavor a10g-large huggingface/lerobot-gpu:latest \
+  bash -c "nvidia-smi && lerobot-train \
+    --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
+    --policy.repo_id=<USER>/act_<task> --batch_size=8 --steps=50000"
+```
+
+Notes:
+
+- The leading `nvidia-smi` is a quick sanity check that CUDA is visible inside the container — useful to fail fast if the flavor or driver mismatched.
+- The default Job timeout is 30 minutes; pass `--timeout 4h` (or longer) for real training.
+- `--flavor` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). For the current full catalogue + pricing see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs).

docs/source/hilserl.mdx

@@ -62,7 +62,7 @@ pip install -e ".[hilserl]"
 
 ### Understanding Configuration
 
-The training process begins with proper configuration for the HILSerl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
+The training process begins with proper configuration for the HILSERl environment. The main configuration class is `GymManipulatorConfig` in `lerobot/rl/gym_manipulator.py`, which contains nested `HILSerlRobotEnvConfig` (defined in `lerobot/envs/configs.py`) and `DatasetConfig`. The configuration is organized into focused, nested sub-configs:
 
 <!-- prettier-ignore-start -->
 ```python
@@ -95,6 +95,7 @@ class HILSerlProcessorConfig:
 class ObservationConfig:
     add_joint_velocity_to_observation: bool = False    # Add joint velocities to state
     add_current_to_observation: bool = False    # Add motor currents to state
+    add_ee_pose_to_observation: bool = False    # Add end-effector pose to state
     display_cameras: bool = False    # Display camera feeds during execution
 
 class ImagePreprocessingConfig:
@@ -326,14 +327,22 @@ lerobot-find-joint-limits \
    Max joint positions [-20.0, -20.0, -20.0, -20.0, -20.0, -20.0]
    Min joint positions [50.0, 50.0, 50.0, 50.0, 50.0, 50.0]
    ```
-3. Use these values in the configuration of your teleoperation device (TeleoperatorConfig) under the `end_effector_bounds` field
+3. Use these values in your environment configuration under `env.processor.inverse_kinematics.end_effector_bounds` (see `InverseKinematicsConfig` in `lerobot/envs/configs.py`)
 
 **Example Configuration**
 
 ```json
-"end_effector_bounds": {
-    "max": [0.24, 0.20, 0.10],
-    "min": [0.16, -0.08, 0.03]
+{
+  "env": {
+    "processor": {
+      "inverse_kinematics": {
+        "end_effector_bounds": {
+          "max": [0.24, 0.2, 0.1],
+          "min": [0.16, -0.08, 0.03]
+        }
+      }
+    }
+  }
 }
 ```
 
@@ -404,30 +413,24 @@ We support using a gamepad or a keyboard or the leader arm of the robot.
 
 HIL-Serl learns actions in the end-effector space of the robot. Therefore, the teleoperation will control the end-effector's x,y,z displacements.
 
-For that we need to define a version of the robot that takes actions in the end-effector space. Check the robot class `SO100FollowerEndEffector` and its configuration `SO100FollowerEndEffectorConfig` for the default parameters related to the end-effector space.
+The end-effector transformation is applied by the processor pipeline (`InverseKinematicsRLStep`, `EEBoundsAndSafety`, `EEReferenceAndDelta`, `GripperVelocityToJoint`) configured under `env.processor.inverse_kinematics` (`InverseKinematicsConfig`) and `env.processor.gripper` / `env.processor.max_gripper_pos`. The defaults related to the end-effector space are:
 
 <!-- prettier-ignore-start -->
 ```python
-class SO100FollowerEndEffectorConfig(SO100FollowerConfig):
-    """Configuration for the SO100FollowerEndEffector robot."""
+class InverseKinematicsConfig:
+    """Configuration for inverse kinematics processing."""
 
-    # Default bounds for the end-effector position (in meters)
-    end_effector_bounds: dict[str, list[float]] = field( # bounds for the end-effector in x,y,z direction
-        default_factory=lambda: {
-            "min": [-1.0, -1.0, -1.0],  # min x, y, z
-            "max": [1.0, 1.0, 1.0],  # max x, y, z
-        }
-    )
+    urdf_path: str | None = None
+    target_frame_name: str | None = None
+    # bounds for the end-effector in x,y,z direction
+    end_effector_bounds: dict[str, list[float]] | None = None
+    # maximum step size for the end-effector in x,y,z direction
+    end_effector_step_sizes: dict[str, float] | None = None
 
-    max_gripper_pos: float = 50 # maximum gripper position that the gripper will be open at
-
-    end_effector_step_sizes: dict[str, float] = field( # maximum step size for the end-effector in x,y,z direction
-        default_factory=lambda: {
-            "x": 0.02,
-            "y": 0.02,
-            "z": 0.02,
-        }
-    )
+class HILSerlProcessorConfig:
+    ...
+    # maximum gripper position that the gripper will be open at
+    max_gripper_pos: float | None = 100.0
 ```
 <!-- prettier-ignore-end -->
 
@@ -606,11 +609,11 @@ This guide explains how to train a reward classifier for human-in-the-loop reinf
 
 **Note**: Training a reward classifier is optional. You can start the first round of RL experiments by annotating the success manually with your gamepad or keyboard device.
 
-The reward classifier implementation in `modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.
+The reward classifier implementation in `lerobot/rewards/classifier/modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.
 
 **Collecting a Dataset for the reward classifier**
 
-Before training, you need to collect a dataset with labeled examples. The `record_dataset` function in `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.
+Before training, you need to collect a dataset with labeled examples. Setting `mode: "record"` in your config and running `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.
 
 To collect a dataset, you need to modify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
 
@@ -658,7 +661,7 @@ Example configuration section for data collection:
   },
   "dataset": {
     "repo_id": "hf_username/dataset_name",
-    "dataset_root": "data/your_dataset",
+    "root": "data/your_dataset",
     "task": "reward_classifier_task",
     "num_episodes_to_record": 20,
     "replay_episode": null,
@@ -671,7 +674,7 @@ Example configuration section for data collection:
 
 **Reward Classifier Configuration**
 
-The reward classifier is configured using `configuration_classifier.py`. Here are the key parameters:
+The reward classifier is configured using `lerobot/rewards/classifier/configuration_classifier.py`. Here are the key parameters:
 
 - **model_name**: Base model architecture (e.g., we mainly use `"helper2424/resnet10"`)
 - **model_type**: `"cnn"` or `"transformer"`
@@ -689,7 +692,7 @@ Example configuration for training the [reward classifier](https://huggingface.c
     "repo_id": "hf_username/dataset_name",
     "root": null
   },
-  "policy": {
+  "reward_model": {
     "type": "reward_classifier",
     "model_name": "helper2424/resnet10",
     "model_type": "cnn",
@@ -699,7 +702,6 @@ Example configuration for training the [reward classifier](https://huggingface.c
     "dropout_rate": 0.1,
     "learning_rate": 1e-4,
     "device": "cuda",
-    "use_amp": true,
     "input_features": {
       "observation.images.front": {
         "type": "VISUAL",
@@ -818,13 +820,14 @@ The LeRobot system uses a distributed actor-learner architecture for training. T
 
 **Configuration Setup**
 
-Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/configs/train.py`.
+Create a training configuration file (example available [here](https://huggingface.co/datasets/lerobot/config_examples/resolve/main/rl/train_config.json)). The training config is based on the main `TrainRLServerPipelineConfig` class in `lerobot/rl/train_rl.py`.
 
-1. Configure the policy settings (`type="sac"`, `device`, etc.)
-2. Set `dataset` to your cropped dataset
-3. Configure environment settings with crop parameters
-4. Check the other parameters related to SAC in [configuration_sac.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/sac/configuration_sac.py#L79).
-5. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.
+1. Configure the policy settings (`type="gaussian_actor"`, `device`, etc.)
+2. Configure the algorithm settings under the top-level `algorithm` block (`type="sac"`, learning rates, discount, etc., defined in `lerobot/rl/algorithms/sac/configuration_sac.py`).
+3. Set `dataset` to your cropped dataset
+4. Configure environment settings with crop parameters
+5. Check the other parameters related to the Gaussian Actor in [configuration_gaussian_actor.py](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py#L79).
+6. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.
 
 **Starting the Learner**
 
@@ -926,7 +929,7 @@ The ideal behaviour is that your intervention rate should drop gradually during
 
 Some configuration values have a disproportionate impact on training stability and speed:
 
-- **`temperature_init`** (`policy.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
+- **`temperature_init`** (`algorithm.temperature_init`) – initial entropy temperature in SAC. Higher values encourage more exploration; lower values make the policy more deterministic early on. A good starting point is `1e-2`. We observed that setting it too high can make human interventions ineffective and slow down learning.
 - **`policy_parameters_push_frequency`** (`policy.actor_learner_config.policy_parameters_push_frequency`) – interval in _seconds_ between two weight pushes from the learner to the actor. The default is `4 s`. Decrease to **1-2 s** to provide fresher weights (at the cost of more network traffic); increase only if your connection is slow, as this will reduce sample efficiency.
 - **`storage_device`** (`policy.storage_device`) – device on which the learner keeps the policy parameters. If you have spare GPU memory, set this to `"cuda"` (instead of the default `"cpu"`). Keeping the weights on-GPU removes CPU→GPU transfer overhead and can significantly increase the number of learner updates per second.
 

docs/source/hope_jr.mdx

@@ -232,7 +232,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --display_data=true
 ```
 
@@ -278,6 +278,6 @@ lerobot-record \
   --dataset.num_episodes=10 \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   --policy.path=outputs/train/hopejr_hand/checkpoints/last/pretrained_model
 ```

docs/source/il_robots.mdx

@@ -193,7 +193,7 @@ lerobot-record \
     --dataset.num_episodes=5 \
     --dataset.single_task="Grab the black cube" \
     --dataset.streaming_encoding=true \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --dataset.encoder_threads=2
 ```
 </hfoption>

docs/source/installation.mdx

@@ -207,6 +207,56 @@ pip install 'lerobot[feetech]'        # Feetech motor support
 
 _Multiple extras can be combined (e.g., `.[core_scripts,pi,pusht]`). For a full list of available extras, refer to `pyproject.toml`._
 
+### PyTorch CUDA variant (Linux only)
+
+On Linux, the install path determines which CUDA wheel you get. macOS and Windows installs use the PyPI default (MPS / CPU / CUDA-Windows wheel respectively) and can skip this section.
+
+<!-- prettier-ignore-start -->
+
+<hfoptions id="cuda_variant">
+<hfoption id="uv-source">
+
+**Source install via `uv` (`uv sync` or `uv pip install -e .`)**
+
+`torch` and `torchvision` are pinned by the project to the **CUDA 12.8** PyTorch index (`https://download.pytorch.org/whl/cu128`, driver floor **570.86**) — covers Ampere/Ada/Hopper/Blackwell GPUs. No action needed for typical NVIDIA setups.
+
+To override for a different CUDA variant:
+
+```bash
+uv pip install --force-reinstall torch torchvision \
+    --index-url https://download.pytorch.org/whl/cu126   # older drivers; or cu130 for Blackwell on driver ≥ 580
+```
+
+</hfoption>
+<hfoption id="pip-conda">
+
+**Source install via `pip`/`conda`, or `pip install lerobot` from PyPI**
+
+PyPI default torch wheel is currently a cu130-bundled Linux wheel, driver floor **580.65**.
+
+To pick a specific CUDA variant:
+
+**Using `pip` or `conda`** — install torch first with an explicit index, then lerobot:
+
+```bash
+pip install --index-url https://download.pytorch.org/whl/cu128 torch torchvision
+pip install -e ".[all]"          # source
+# — or —
+pip install lerobot              # from PyPI
+```
+
+**Using `uv` to install from PyPI** — one-liner via `--torch-backend` (uv ≥ 0.6):
+
+```bash
+uv pip install --torch-backend cu128 lerobot
+```
+
+Supported values include `auto`, `cpu`, `cu126`, `cu128`, `cu129`, `cu130`, plus various `rocm*` and `xpu`. Swap as needed for your driver.
+
+</hfoption>
+</hfoptions>
+<!-- prettier-ignore-end -->
+
 ### Troubleshooting
 
 If you encounter build errors, you may need to install additional system dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.

docs/source/language_and_recipes.mdx

@@ -0,0 +1,147 @@
+# Language columns and recipes
+
+Most LeRobot datasets ship with a single `task` string per episode — fine for
+short, single-instruction skills, but not enough for the longer-horizon,
+multi-modal robot policies the field is moving toward (high-level planning,
+memory, interjections, VQA, tool use). To support those policies without
+forking the dataset format, LeRobot extends `LeRobotDataset` with two optional
+language columns and a small recipe layer that turns those rows into
+chat-style training samples on the fly.
+
+The design splits cleanly into three layers:
+
+1. **Data in the dataset** — language annotations stored next to frames in
+   `data/chunk-*/file-*.parquet` as two optional columns (`language_persistent`
+   and `language_events`). Datasets without these columns keep their existing
+   behavior.
+2. **Recipe** — a YAML file that declares which annotation rows to bind and
+   how to lay them out as chat turns (`role`, `content`, optional images,
+   optional tool calls). Recipes are pure config; no Python required to add a
+   new one.
+3. **Training format** — at sample time, `RenderMessagesStep` resolves the
+   recipe against the per-frame annotations and emits HF-style `messages` plus
+   LeRobot-specific sidecars (`message_streams`, `target_message_indices`)
+   that policy processors consume.
+
+This page describes each layer in turn.
+
+## Layer 1 — language columns in the dataset
+
+The two optional columns live next to frame data in
+`data/chunk-*/file-*.parquet`:
+
+- `language_persistent`: a list of rows broadcast across every frame in an episode for state that remains active, such as `subtask`, `plan`, and `memory`.
+- `language_events`: a list of rows only on the exact frame where an event was emitted, such as `interjection`, `vqa`, and speech tool calls.
+
+Both columns share the same row shape (event rows omit `timestamp` because the
+frame the row sits on already provides it):
+
+```text
+role: string
+content: string | null
+style: string | null
+timestamp: float32        # persistent rows only
+camera: string | null     # observation.images.* feature key, view-dependent rows only
+tool_calls: list[Json] | null
+```
+
+The `camera` field tags rows whose `content` is grounded in a specific camera
+view. Rows of view-dependent styles (`vqa` and `trace`) MUST set `camera` to
+the matching `observation.images.*` feature key. Rows of every other style —
+including `motion`, which describes robot-frame primitives in joint / Cartesian
+terms — MUST leave `camera` as `null`. Pipeline writers and the validator
+enforce this via `validate_camera_field(style, camera)`.
+
+`meta/tasks.parquet` remains the canonical source for the task. The special `${task}` recipe binding always reads that task string and does not depend on language annotations.
+
+### Architecture
+
+The language stack itself has three internal modules backing layer 1:
+
+1. `lerobot.datasets.language` defines the schema, style registry, and `column_for_style`.
+2. `lerobot.datasets.language_render` resolves rows and renders messages.
+3. `RenderMessagesStep` turns dataset samples into `messages`, `message_streams`, and `target_message_indices`.
+
+`LeRobotDataset` stays recipe-agnostic. It passes `language_persistent` and `language_events` through when present, and unannotated datasets keep their existing behavior.
+
+## Layer 2 — recipe anatomy
+
+Recipes are YAML files backed by `TrainingRecipe` and `MessageTurn`. They
+declare which annotation rows to pull (via `bindings`) and how to compose them
+into chat turns (`messages`).
+
+```yaml
+messages:
+  - { role: user, content: "${task}", stream: high_level }
+  - { role: assistant, content: "${subtask}", stream: low_level, target: true }
+```
+
+A recipe can also branch into a weighted **blend** of sub-recipes. At sample
+time, exactly one branch is selected deterministically from the sample index,
+so different frames train different objectives (e.g. memory updates vs.
+low-level execution vs. VQA) without any Python wiring.
+
+### Temporal semantics
+
+Persistent styles are active after emission until replaced:
+
+- `active_at(t, style=subtask)`
+- `nth_prev(style=memory, offset=1)`
+- `nth_next(style=subtask, offset=1)`
+
+Event styles only exist on their exact timestamp:
+
+- `emitted_at(t, style=interjection)`
+- `emitted_at(t, style=vqa, role=user, camera=observation.images.top)`
+- `emitted_at(t, role=assistant, tool_name=say)`
+
+Exact event matching has no tolerance window, so writers must stamp event rows with frame timestamps from the parquet data.
+
+### View-dependent resolution
+
+For view-dependent styles (`vqa` and `trace`), the resolver gains a
+`camera=` filter parallel to `role=` and `tool_name=`. Datasets with multiple
+cameras typically emit one (`vqa`, `user`) + (`vqa`, `assistant`) pair per
+camera at the same timestamp; without `camera=`, those resolvers see two
+matches and raise an ambiguity error. Recipes consume each camera through its
+own binding plus a matching image block, e.g.
+
+```yaml
+ask_vqa_top:
+  bindings:
+    vqa_query: "emitted_at(t, style=vqa, role=user, camera=observation.images.top)"
+    vqa: "emitted_at(t, style=vqa, role=assistant, camera=observation.images.top)"
+  messages:
+    - role: user
+      stream: high_level
+      if_present: vqa_query
+      content:
+        - { type: image, feature: observation.images.top }
+        - { type: text, text: "${vqa_query}" }
+    - {
+        role: assistant,
+        content: "${vqa}",
+        stream: high_level,
+        target: true,
+        if_present: vqa,
+      }
+```
+
+Add one such sub-recipe per camera the dataset records.
+
+## Layer 3 — training format
+
+Rendered samples use HF-style chat messages plus LeRobot sidecars:
+
+```python
+sample["messages"]
+sample["message_streams"]
+sample["target_message_indices"]
+```
+
+The renderer does not apply a tokenizer chat template. Policy processors decide how to serialize the messages for their backbone, which keeps the same dataset usable across SmolVLA, Pi0.5, and any future VLM that expects OpenAI-style chat messages.
+
+## Graceful absence
+
+If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.
+If an event-scoped branch is selected on a frame without the required event row, rendering returns `None`, allowing a loader to retry another sample.

docs/source/lerobot-dataset-v3.mdx

@@ -10,6 +10,7 @@ This docs will guide you to:
 - Stream datasets without downloading using `StreamingLeRobotDataset`
 - Apply image transforms for data augmentation during training
 - Migrate existing `v2.1` datasets to `v3.0`
+- Experiment with other `LeRobotDataset` formats and implementations like Lance
 
 ## What’s new in `v3`
 
@@ -43,7 +44,7 @@ lerobot-record \
   --dataset.num_episodes=5 \
   --dataset.single_task="Grab the black cube" \
   --dataset.streaming_encoding=true \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   --dataset.encoder_threads=2
 ```
 
@@ -315,3 +316,39 @@ Dataset v3.0 uses incremental parquet writing with buffered metadata for efficie
 - Ensures the dataset is valid for loading
 
 Without calling `finalize()`, your parquet files will be incomplete and the dataset won't load properly.
+
+## Other formats and implementations
+
+### Lance
+
+Lance is a useful format for multimodal AI datasets, especially for large-scale training requiring high performance IO and random access.
+
+The `lerobot-lancedb` package implements `LeRobotLanceDataset` (for JPEG images) and `LeRobotLanceVideoDataset` (for mp4 videos).
+Those two storage layouts both subclass LeRobotDataset and can provide data loading speed ups.
+
+`LeRobotLanceDataset` is a drop-in replacement for `LeRobotDataset`:
+
+```python
+from lerobot.datasets import LeRobotDatasetMetadata
+from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot_lancedb import LeRobotLanceDataset, LeRobotLanceVideoDataset
+
+cfg = DiffusionConfig(...)
+meta = LeRobotDatasetMetadata(root=local_dataset_path)  # or use repo_id=... to load metadata from the Hub
+delta_timestamps = {...}
+
+# Use LeRobotLanceDataset for image datasets
+dataset = LeRobotLanceDataset(
+    root=local_dataset_path,                            # or use repo_id=... to stream from the Hub
+    delta_timestamps=delta_timestamps,
+    return_uint8=True,
+)
+# Or use LeRobotLanceVideoDataset for video datasets:
+dataset = LeRobotLanceVideoDataset(
+    root=local_dataset_path,                            # or use repo_id=... to stream from the Hub
+    delta_timestamps=delta_timestamps,
+    return_uint8=True,
+)
+```
+
+Join the discussion on [Github](https://github.com/huggingface/lerobot/issues/3608) and explore the `lerobot-lancedb` documentation [here](https://lancedb.github.io/lerobot-lancedb/).

docs/source/peft_training.mdx

@@ -28,13 +28,15 @@ lerobot-train \
  --steps=100000 \
  --batch_size=32 \
  --peft.method_type=LORA \
- --peft.r=64
+ --peft.r=64 \
+ --peft.lora_alpha=64
 ```
 
 Note the `--peft.method_type` parameter that let's you select which PEFT method to use. Here we use
 [LoRA](https://huggingface.co/docs/peft/main/en/package_reference/lora) (Low-Rank Adapter) which is probably the most
 popular fine-tuning method to date. Low-rank adaption means that we only fine-tune a matrix with comparably low rank
-instead of the full weight matrix. This rank can be specified using the `--peft.r` parameter. The higher the rank
+instead of the full weight matrix. This rank can be specified using the `--peft.r` parameter, and the LoRA scaling factor with
+`--peft.lora_alpha` (where `scaling = lora_alpha / r`). The higher the rank
 the closer you get to full fine-tuning
 
 There are more complex methods that have more parameters. These are not yet supported, feel free to raise an issue

docs/source/reachy2.mdx

@@ -161,7 +161,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --display_data=true
 ```
 
@@ -203,7 +203,7 @@ lerobot-record \
     --dataset.private=true \
     --dataset.streaming_encoding=true \
     --dataset.encoder_threads=2 \
-    # --dataset.vcodec=auto \
+    # --dataset.camera_encoder.vcodec=auto \
     --display_data=true
 ```
 

docs/source/rebot_b601.mdx

@@ -0,0 +1,186 @@
+# reBot B601-DM
+
+[reBot B601-DM](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/) is an open-source, low-cost robot arm from Seeed Studio for embodied-AI and imitation learning. It comes as a **follower** arm (the `B601-DM`, a 6-DOF arm plus gripper driven by Damiao CAN motors) and a **leader** arm (the `StarArm102` / `reBot Arm 102`, driven by FashionStar UART smart servos) used to teleoperate it.
+
+This page covers **calibration** and **teleoperation** for both single-arm and bimanual (dual-arm) setups.
+
+<div style="display: flex; align-items: center; gap: 10px;">
+  <img
+    src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/b601dm_zeroposition.jpg"
+    alt="reBot B601-DM follower arm at its zero position"
+    width="48%"
+  />
+  <img
+    src="https://files.seeedstudio.com/wiki/robotics/projects/lerobot/102_zeroposition.jpg"
+    alt="reBot Arm 102 leader arm at its zero position"
+    width="48%"
+  />
+</div>
+
+_Left: the B601-DM follower at its zero position. Right: the reBot Arm 102 leader at its zero position. Images courtesy of [Seeed Studio](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/)._
+
+## Install LeRobot 🤗
+
+Follow our [Installation Guide](./installation), then install the reBot support:
+
+```bash
+pip install -e ".[rebot]"
+```
+
+This pulls in `motorbridge` (CAN motor control for the B601-DM follower) and `motorbridge-smart-servo` (FashionStar UART servos for the reBot Arm 102 leader).
+
+## Registered device types
+
+| Type                     | Kind                                         |
+| ------------------------ | -------------------------------------------- |
+| `rebot_b601_follower`    | single-arm B601-DM follower robot            |
+| `bi_rebot_b601_follower` | bimanual (dual-arm) follower robot           |
+| `rebot_102_leader`       | single-arm reBot Arm 102 leader teleoperator |
+| `bi_rebot_102_leader`    | bimanual (dual-arm) leader teleoperator      |
+
+The bimanual types compose two single-arm instances and namespace each arm's
+observation/action keys with a `left_` / `right_` prefix. Per-arm settings are
+passed through nested `left_arm_config.*` / `right_arm_config.*` arguments.
+
+## Find the USB ports
+
+For each device, find the USB port associated with its motor bus using:
+
+```bash
+lerobot-find-port
+```
+
+<Tip warning={true}>
+  On Linux, remove `brltty` (`sudo apt remove brltty`) so it does not hold the
+  leader's USB serial port. You may also need to grant access to the serial
+  devices: `sudo chmod 666 /dev/ttyACM* /dev/ttyUSB*`.
+</Tip>
+
+## Calibration
+
+Neither arm stores a persistent hardware calibration: every time it connects, the motors are re-zeroed against the pose the arm is physically holding. Calibration simply records that zero pose. When prompted, **manually move the arm to its zero position** (the default sit-down pose shown above, gripper fully closed) and press <kbd>ENTER</kbd>.
+
+### Follower (B601-DM)
+
+<hfoptions id="calibrate-follower">
+<hfoption id="Single arm">
+
+```bash
+lerobot-calibrate \
+    --robot.type=rebot_b601_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=follower \
+    --robot.can_adapter=damiao
+```
+
+</hfoption>
+<hfoption id="Dual arm">
+
+Connect the bimanual follower; calibration runs for the left arm, then the right arm.
+
+```bash
+lerobot-calibrate \
+    --robot.type=bi_rebot_b601_follower \
+    --robot.id=bi_follower \
+    --robot.left_arm_config.port=/dev/ttyACM0 \
+    --robot.left_arm_config.can_adapter=damiao \
+    --robot.right_arm_config.port=/dev/ttyACM1 \
+    --robot.right_arm_config.can_adapter=damiao
+```
+
+Per-arm calibration files are saved with `_left` / `_right` suffixes on the id.
+
+</hfoption>
+</hfoptions>
+
+### Leader (reBot Arm 102)
+
+<hfoptions id="calibrate-leader">
+<hfoption id="Single arm">
+
+```bash
+lerobot-calibrate \
+    --teleop.type=rebot_102_leader \
+    --teleop.port=/dev/ttyUSB0 \
+    --teleop.id=leader
+```
+
+</hfoption>
+<hfoption id="Dual arm">
+
+```bash
+lerobot-calibrate \
+    --teleop.type=bi_rebot_102_leader \
+    --teleop.id=bi_leader \
+    --teleop.left_arm_config.port=/dev/ttyUSB0 \
+    --teleop.right_arm_config.port=/dev/ttyUSB1
+```
+
+</hfoption>
+</hfoptions>
+
+## Teleoperation
+
+Once both arms are calibrated, drive the follower with the leader. The follower talks to its CAN bus through a Damiao serial bridge (`can_adapter=damiao`, the default) or a SocketCAN adapter (`can_adapter=socketcan`). See the [OpenArm page](./openarm) for more details on the SocketCAN adapter configuration.
+
+<hfoptions id="teleoperate">
+<hfoption id="Single arm">
+
+```bash
+lerobot-teleoperate \
+    --robot.type=rebot_b601_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.id=follower \
+    --robot.can_adapter=damiao \
+    --teleop.type=rebot_102_leader \
+    --teleop.port=/dev/ttyUSB0 \
+    --teleop.id=leader
+```
+
+</hfoption>
+<hfoption id="Dual arm">
+
+The bimanual leader and follower reuse the single-arm classes; each arm is
+configured through nested `left_arm_config.*` / `right_arm_config.*` arguments,
+so a bimanual reBot Arm 102 leader drives a bimanual B601-DM follower.
+
+```bash
+lerobot-teleoperate \
+    --robot.type=bi_rebot_b601_follower \
+    --robot.id=bi_follower \
+    --robot.left_arm_config.port=/dev/ttyACM0 \
+    --robot.left_arm_config.can_adapter=damiao \
+    --robot.right_arm_config.port=/dev/ttyACM1 \
+    --robot.right_arm_config.can_adapter=damiao \
+    --teleop.type=bi_rebot_102_leader \
+    --teleop.id=bi_leader \
+    --teleop.left_arm_config.port=/dev/ttyUSB0 \
+    --teleop.right_arm_config.port=/dev/ttyUSB1
+```
+
+</hfoption>
+</hfoptions>
+
+<Tip>
+  The leader and follower share the same joint names (`shoulder_pan,
+  shoulder_lift, elbow_flex, wrist_flex, wrist_yaw, wrist_roll, gripper`), so
+  leader actions map directly onto the follower.
+</Tip>
+
+If the motion of a joint is reversed, flip its sign in the leader's `joint_directions` (the gripper also carries a scale to widen its range to the follower):
+
+```bash
+lerobot-teleoperate \
+    --robot.type=rebot_b601_follower \
+    --robot.port=/dev/ttyACM0 \
+    --robot.can_adapter=damiao \
+    --teleop.type=rebot_102_leader \
+    --teleop.port=/dev/ttyUSB0 \
+    --teleop.joint_directions='{"shoulder_pan":-1,"shoulder_lift":-1,"elbow_flex":1,"wrist_flex":1,"wrist_yaw":1,"wrist_roll":-1,"gripper":-6}'
+```
+
+## Recording datasets
+
+Swap `lerobot-teleoperate` for `lerobot-record` (with the same `--robot.*` / `--teleop.*` arguments, plus `--dataset.*`) to record demonstrations for training. See [Imitation Learning for Robots](./il_robots) for the full workflow.
+
+For hardware assembly and wiring, see the [Seeed Studio reBot wiki](https://wiki.seeedstudio.com/rebot_arm_b601_dm_lerobot/).

docs/source/smolvla.mdx

@@ -108,7 +108,7 @@ lerobot-record \
   --dataset.num_episodes=10 \
   --dataset.streaming_encoding=true \
   --dataset.encoder_threads=2 \
-  # --dataset.vcodec=auto \
+  # --dataset.camera_encoder.vcodec=auto \
   # <- Teleop optional if you want to teleoperate in between episodes \
   # --teleop.type=so100_leader \
   # --teleop.port=/dev/ttyACM0 \

docs/source/streaming_video_encoding.mdx

@@ -17,9 +17,9 @@ This makes `save_episode()` near-instant (the video is already encoded by the ti
 | Parameter               | CLI Flag                          | Type          | Default       | Description                                                       |
 | ----------------------- | --------------------------------- | ------------- | ------------- | ----------------------------------------------------------------- |
 | `streaming_encoding`    | `--dataset.streaming_encoding`    | `bool`        | `True`        | Enable real-time encoding during capture                          |
-| `vcodec`                | `--dataset.vcodec`                | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
+| `vcodec`                | `--dataset.camera_encoder.vcodec` | `str`         | `"libsvtav1"` | Video codec. `"auto"` detects best HW encoder                     |
 | `encoder_threads`       | `--dataset.encoder_threads`       | `int \| None` | `None` (auto) | Threads per encoder instance. `None` will leave the vcoded decide |
-| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `60`          | Max buffered frames per camera (~2s at 30fps). Consumes RAM       |
+| `encoder_queue_maxsize` | `--dataset.encoder_queue_maxsize` | `int`         | `30`          | Max buffered frames per camera (~1s at 30fps). Consumes RAM       |
 
 ## 3. Performance Considerations
 
@@ -48,7 +48,7 @@ This parameter controls how many threads each encoder instance uses internally:
 
 ### Backpressure and Frame Dropping
 
-Each camera has a bounded queue (`encoder_queue_maxsize`, default 60 frames). When the encoder can't keep up:
+Each camera has a bounded queue (`encoder_queue_maxsize`, default 30 frames). When the encoder can't keep up:
 
 1. The queue fills up (consuming RAM)
 2. New frames are **dropped** (not blocked) — the capture loop continues uninterrupted
@@ -82,15 +82,15 @@ Use HW encoding when:
 
 ### Available HW Encoders
 
-| Encoder             | Platform      | Hardware                                                                                         | CLI Value                            |
-| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------ |
-| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=h264_videotoolbox` |
-| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.vcodec=hevc_videotoolbox` |
-| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=h264_nvenc`        |
-| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.vcodec=hevc_nvenc`        |
-| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.vcodec=h264_vaapi`        |
-| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.vcodec=h264_qsv`          |
-| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.vcodec=auto`              |
+| Encoder             | Platform      | Hardware                                                                                         | CLI Value                                           |
+| ------------------- | ------------- | ------------------------------------------------------------------------------------------------ | --------------------------------------------------- |
+| `h264_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.camera_encoder.vcodec=h264_videotoolbox` |
+| `hevc_videotoolbox` | macOS         | Apple Silicon / Intel                                                                            | `--dataset.camera_encoder.vcodec=hevc_videotoolbox` |
+| `h264_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.camera_encoder.vcodec=h264_nvenc`        |
+| `hevc_nvenc`        | Linux/Windows | NVIDIA GPU                                                                                       | `--dataset.camera_encoder.vcodec=hevc_nvenc`        |
+| `h264_vaapi`        | Linux         | Intel/AMD GPU                                                                                    | `--dataset.camera_encoder.vcodec=h264_vaapi`        |
+| `h264_qsv`          | Linux/Windows | Intel Quick Sync                                                                                 | `--dataset.camera_encoder.vcodec=h264_qsv`          |
+| `auto`              | Any           | Probes the system for available HW encoders. Falls back to `libsvtav1` if no HW encoder is found | `--dataset.camera_encoder.vcodec=auto`              |
 
 > [!NOTE]
 > In order to use the HW accelerated encoders you might need to upgrade your GPU drivers.
@@ -100,15 +100,15 @@ Use HW encoding when:
 
 ## 5. Troubleshooting
 
-| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                  |
-| ------------------------------------------------------------------ | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.vcodec=auto`) |
-| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.vcodec=auto`).                                                                                                                                                    |
-| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                     |
-| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                              |
-| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                              |
-| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.vcodec=auto`                                                                                                                                                                                                                             |
-| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                   |
+| Symptom                                                            | Likely Cause                                 | Fix                                                                                                                                                                                                                                                                                                 |
+| ------------------------------------------------------------------ | -------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| System freezes or choppy robot movement or Rerun visualization lag | CPU starved (100% load usage)                | Close other apps, reduce encoding throughput, lower `encoder_threads`, use `h264`, use `display_data=False`. If the CPU continues to be at 100% then it might be insufficient for your setup, consider `--dataset.streaming_encoding=false` or HW encoding (`--dataset.camera_encoder.vcodec=auto`) |
+| "Encoder queue full" warnings or dropped frames in dataset         | Encoder can't keep up (Queue overflow)       | If CPU is not at 100%: Increase `encoder_threads`, increase `encoder_queue_maxsize` or use HW encoding (`--dataset.camera_encoder.vcodec=auto`).                                                                                                                                                    |
+| High RAM usage                                                     | Queue filling faster than encoding           | `encoder_threads` too low or CPU insufficient. Reduce `encoder_queue_maxsize` or use HW encoding                                                                                                                                                                                                    |
+| Large video files                                                  | Using HW encoder or H.264                    | Expected trade-off. Switch to `libsvtav1` if CPU allows                                                                                                                                                                                                                                             |
+| `save_episode()` still slow                                        | `streaming_encoding` is `False`              | Set `--dataset.streaming_encoding=true`                                                                                                                                                                                                                                                             |
+| Encoder thread crash                                               | Codec not available or invalid settings      | Check `vcodec` is installed, try `--dataset.camera_encoder.vcodec=auto`                                                                                                                                                                                                                             |
+| Recorded dataset is missing frames                                 | CPU/GPU starvation or occasional load spikes | If ~5% of frames are missing, your system is likely overloaded — follow the recommendations above. If fewer frames are missing (~2%), they are probably due to occasional transient load spikes (often at startup) and can be considered expected.                                                  |
 
 ## 6. Recommended Configurations
 
@@ -146,7 +146,7 @@ On very constrained systems, streaming encoding may compete too heavily with the
 # 2camsx 640x480x3 @30fps: Requires some tuning.
 
 # Use H.264, disable streaming, consider batching encoding
-lerobot-record --dataset.vcodec=h264 --dataset.streaming_encoding=false ...
+lerobot-record --dataset.camera_encoder.vcodec=h264 --dataset.streaming_encoding=false ...
 ```
 
 ## 7. Closing note

docs/source/tools.mdx

@@ -0,0 +1,210 @@
+# Tools
+
+LeRobot v3.1 supports **tool calls** in policies — assistant messages can
+emit structured invocations like `say(text="OK, starting now")` that the
+runtime dispatches to a real implementation (TTS, controller, logger, …).
+
+This page covers:
+
+1. Where the tool catalog lives.
+2. How the annotation pipeline produces tool-call atoms.
+3. How to add your own tool.
+
+## Where tools are declared
+
+Two layers.
+
+**The catalog** — a list of OpenAI-style function schemas — lives at
+`meta/info.json["tools"]` on each dataset. Example:
+
+```json
+{
+  "features": { "...": "..." },
+  "tools": [
+    {
+      "type": "function",
+      "function": {
+        "name": "say",
+        "description": "Speak a short utterance to the user via the TTS executor.",
+        "parameters": {
+          "type": "object",
+          "properties": {
+            "text": {
+              "type": "string",
+              "description": "The verbatim text to speak."
+            }
+          },
+          "required": ["text"]
+        }
+      }
+    }
+  ]
+}
+```
+
+Read it via the dataset metadata accessor:
+
+```python
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
+
+meta = LeRobotDatasetMetadata(repo_id="pepijn/super_poulain_final_annotations")
+tools = meta.tools     # list[dict] — OpenAI tool schemas
+```
+
+If the dataset's `info.json` doesn't declare any tools, `meta.tools`
+returns `DEFAULT_TOOLS` from `lerobot.datasets.language` — currently a
+single-entry list with the canonical `say` schema. So unannotated
+datasets and chat-template consumers keep working without any
+configuration:
+
+```python
+prompt_str = tokenizer.apply_chat_template(
+    sample["messages"],
+    tools=meta.tools,                 # works either way
+    add_generation_prompt=False,
+    tokenize=False,
+)
+```
+
+**The implementations** — runnable Python — will live under
+`src/lerobot/tools/`, one file per tool. The runtime dispatcher and
+the canonical `say` implementation (wrapping Kyutai's pocket-tts) are
+not part of the catalog layer described here; today this layer ships
+only the schema storage and the `DEFAULT_TOOLS` fallback constant.
+
+## Per-row tool _invocations_
+
+The catalog above describes _what can be called_. The actual _call_ — the
+function name plus the argument values — is stored per-row, on the
+assistant atoms in `language_events`:
+
+```python
+{
+  "role": "assistant",
+  "content": null,
+  "style": null,
+  "timestamp": 12.4,
+  "camera": null,
+  "tool_calls": [
+    { "type": "function",
+      "function": { "name": "say", "arguments": { "text": "On it." } } }
+  ]
+}
+```
+
+Recipes splice these into rendered messages via `tool_calls_from`:
+
+```yaml
+user_interjection_response:
+  bindings:
+    speech: "emitted_at(t, role=assistant, tool_name=say)"
+  messages:
+    - { role: user, content: "${task}", stream: high_level }
+    - {
+        role: assistant,
+        content: "${current_plan}",
+        stream: high_level,
+        target: true,
+        tool_calls_from: speech,
+      }
+```
+
+The model's training target is one assistant turn that carries both the
+plan text _and_ the `say` tool call. At inference, the runtime parses
+the generated text back into structured `tool_calls` and dispatches to
+the matching implementation.
+
+## How to add your own tool
+
+> **Note:** Steps 2 and 3 below describe the runtime layer
+> (`src/lerobot/tools/`, the `Tool` protocol, `TOOL_REGISTRY`,
+> `get_tools(meta)`) which is not part of the catalog layer shipped
+> today — those modules don't yet exist in the tree. Step 1 alone is
+> enough to make the tool visible to the chat template via
+> `meta.tools` so the model can learn to _generate_ the call;
+> executing the call at inference requires the runtime layer.
+
+Three steps. Concrete example: a `record_observation` tool the policy
+can call to capture an extra observation outside the regular control
+loop.
+
+### Step 1 — declare the schema
+
+Add an entry under `meta/info.json["tools"]`. Either edit the file
+directly on disk _before_ running the annotation pipeline (it'll be
+preserved) or hand it to `lerobot-annotate` via a config flag.
+
+```json
+{
+  "tools": [
+    { "type": "function", "function": { "name": "say", "...": "..." } },
+    {
+      "type": "function",
+      "function": {
+        "name": "record_observation",
+        "description": "Capture a high-resolution still image for the user.",
+        "parameters": {
+          "type": "object",
+          "properties": {
+            "label": {
+              "type": "string",
+              "description": "Short label for the saved image."
+            }
+          },
+          "required": ["label"]
+        }
+      }
+    }
+  ]
+}
+```
+
+The schema follows OpenAI's function-calling convention exactly, so the
+chat template can render it natively.
+
+### Step 2 — implement the call
+
+Create `src/lerobot/tools/record_observation.py`:
+
+```python
+from .base import Tool
+from typing import Any
+
+RECORD_OBSERVATION_SCHEMA: dict[str, Any] = { "...": "..." }   # mirrors the JSON above
+
+
+class RecordObservationTool:
+    name = "record_observation"
+    schema = RECORD_OBSERVATION_SCHEMA
+
+    def __init__(self, schema: dict | None = None, output_dir: str = "."):
+        self.output_dir = output_dir
+
+    def call(self, arguments: dict) -> str:
+        label = arguments["label"]
+        # ... save the latest camera frame to <output_dir>/<label>.png ...
+        return f"saved {label}.png"
+```
+
+One file per tool keeps dependencies isolated — `record_observation`
+might pull `pillow`, while `say` pulls `pocket-tts`. Users installing
+only the tools they need avoid heavy transitive deps.
+
+### Step 3 — register it
+
+Add to `src/lerobot/tools/registry.py`:
+
+```python
+from .record_observation import RecordObservationTool
+
+TOOL_REGISTRY["record_observation"] = RecordObservationTool
+```
+
+That's it. At runtime `get_tools(meta)` looks up each schema in
+`meta.tools`, instantiates the matching registered class, and returns
+a name → instance dict the dispatcher can route into.
+
+If you want to use a tool _without_ writing an implementation (e.g. for
+training-time chat-template formatting only), step 1 alone is enough —
+the model still learns to _generate_ the call. Steps 2 and 3 are only
+needed to actually _execute_ it at inference.

docs/source/topreward.mdx

@@ -0,0 +1,177 @@
+# TOPReward
+
+TOPReward is a **zero-shot reward model** that extracts token log-probabilities from an off-the-shelf vision-language model (VLM) as a robotic reward signal. Given a video trajectory and a task instruction, it returns the VLM's log-likelihood that the instruction is true — no fine-tuning required.
+
+**Paper**: [TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics](https://arxiv.org/abs/2602.19313)
+**Project**: [topreward.github.io](https://topreward.github.io/webpage/)
+**Original code**: [github.com/TOPReward/TOPReward](https://github.com/TOPReward/TOPReward)
+**Default backbone**: [Qwen/Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct)
+
+## Overview
+
+TOPReward asks a generic VLM how likely a task instruction is, **conditioned on the video** of a robot trying to complete that task. Concretely, given:
+
+- A trajectory video (a sequence of frames).
+- A task instruction (e.g. _"open the drawer"_).
+
+it builds a chat prompt of the form
+
+```text
+<video>
+"The above video shows a robot manipulation trajectory that completes the
+ following task: <instruction> Decide whether the above statement is True
+ or not. The answer is: True"
+```
+
+forwards it through the VLM, label-masks everything except the very last token, and reads back the log-probability of that token — by default the literal `"True"` that closes the suffix template. The resulting `log P("True" | video + prompt + instruction)` is the reward.
+
+Because the method only depends on a frozen VLM, TOPReward is **zero-shot**: there are no fine-tuned weights to host. The "model" in LeRobot is a small wrapper around `transformers`' `Qwen3VLForConditionalGeneration` plus the label-masking logic. The processor owns the tokeniser and builds the full chat prompt (EO-1/Robometer pattern).
+
+## What the LeRobot integration covers
+
+- Standard `reward_model.type=topreward` configuration through LeRobot.
+- VLM loading via the `transformers` `Qwen3VLForConditionalGeneration` API.
+- Prompt assembly + tokenisation in the processor (matching upstream `QwenClient.compute_instruction_reward`).
+- `compute_reward()` returns one scalar log-prob per sample.
+- LeRobot reward-model save/load — `save_pretrained` writes only `config.json` (the VLM is identified by `vlm_name`).
+- An offline labeling script that writes a `topreward_progress.parquet` (SARM-compatible schema) for RA-BC and overlay.
+
+The current LeRobot port supports the **Qwen3-VL client only**. Other upstream clients (Gemini, OpenAI, Gemma, Molmo) can be added as follow-up extras.
+
+## Installation Requirements
+
+1. Install LeRobot following the [Installation Guide](./installation).
+2. Install the TOPReward optional extra:
+
+```bash
+pip install -e ".[topreward]"
+```
+
+or, with `uv` from a source checkout:
+
+```bash
+uv sync --extra topreward
+```
+
+This pulls in `transformers`. The first time you run TOPReward, Hugging Face will also download the VLM weights from the Hub (~16 GB for Qwen3-VL-8B-Instruct). A GPU is strongly recommended.
+
+## Model Inputs and Outputs
+
+TOPReward expects:
+
+- A trajectory video or sequence of frames.
+- A natural-language task description.
+
+In LeRobot datasets the preprocessor reads:
+
+| Config field              | Default                     | Meaning                                       |
+| ------------------------- | --------------------------- | --------------------------------------------- |
+| `reward_model.image_key`  | `observation.images.top`    | Camera observation used by TOPReward          |
+| `reward_model.task_key`   | `task`                      | Key in complementary data for the task string |
+| `reward_model.max_frames` | `16`                        | Cap on frames per sample                      |
+| `reward_model.fps`        | `2.0`                       | Metadata passed to the Qwen video processor   |
+| `reward_model.vlm_name`   | `Qwen/Qwen3-VL-8B-Instruct` | Hugging Face Hub id of the underlying VLM     |
+
+The model returns:
+
+- `compute_reward(batch)`: one log-probability per sample. Higher = better task-video alignment. When `success_threshold` is finite, returns the binary thresholded value instead.
+
+## Usage
+
+### Load the reward model directly
+
+```python
+from lerobot.rewards.topreward import TOPRewardConfig, TOPRewardModel
+
+cfg = TOPRewardConfig(
+    vlm_name="Qwen/Qwen3-VL-8B-Instruct",
+    device="cuda",
+)
+reward_model = TOPRewardModel(cfg)
+```
+
+### Use the reward factory
+
+```python
+from lerobot.rewards import make_reward_model, make_reward_model_config, make_reward_pre_post_processors
+
+cfg = make_reward_model_config(
+    "topreward",
+    vlm_name="Qwen/Qwen3-VL-8B-Instruct",
+    device="cuda",
+    image_key="observation.images.top",
+)
+reward_model = make_reward_model(cfg)
+preprocessor, postprocessor = make_reward_pre_post_processors(cfg)
+```
+
+The preprocessor tokenises the full prompt (video + prefix + instruction suffix), writes Qwen-VL tensors + `prompt_length` under `observation.topreward.*`. The model reads those tensors, label-masks based on `prompt_length`, and extracts the log-prob reward.
+
+### Offline dataset labeling
+
+Write a `topreward_progress.parquet` for RA-BC training and overlay videos:
+
+```bash
+# Sparse-dense (15 anchors per episode, matches upstream)
+uv run python -m lerobot.rewards.topreward.compute_rabc_weights \
+    --dataset-repo-id lerobot/libero_10_image \
+    --num-samples 15 \
+    --device cuda
+```
+
+Then render the progress overlay for any episode:
+
+```bash
+uv run examples/dataset/create_progress_videos.py \
+    --repo-id lerobot/libero_10_image \
+    --episode 0 \
+    --progress-file topreward_progress.parquet \
+    --gif
+```
+
+## Configuration Notes
+
+### Prompt knobs
+
+The default prompt mirrors the upstream paper:
+
+```text
+prompt_prefix = "The above video shows a robot manipulation trajectory that completes the following task: "
+prompt_suffix_template = "{instruction} Decide whether the above statement is True or not. The answer is: True"
+```
+
+Both are exposed on `TOPRewardConfig` for ablation. The suffix template **must** contain `{instruction}`.
+
+### Chat template
+
+`add_chat_template=True` wraps the full prompt (including instruction) with the tokenizer's chat template before tokenisation. Default is `False`, matching the upstream paper's main experiments.
+
+## Limitations
+
+- The current LeRobot port is **inference-only and zero-shot**; `forward()` is not overridden and `is_trainable` returns `False`.
+- Only the **Qwen3-VL family** is supported; other upstream clients are out of scope.
+- TOPReward inherits the underlying VLM's biases.
+
+## References
+
+- [TOPReward project page](https://topreward.github.io/webpage/)
+- [TOPReward paper](https://arxiv.org/abs/2602.19313)
+- [Original TOPReward code](https://github.com/TOPReward/TOPReward)
+- [Qwen3-VL-8B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct)
+
+## Citation
+
+```bibtex
+@article{chen2026topreward,
+  title={TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics},
+  author={Chen, Shirui and Harrison, Cole and Lee, Ying-Chun and Yang, Angela Jin and
+          Ren, Zhongzheng and Ratliff, Lillian J and Duan, Jiafei and Fox, Dieter and
+          Krishna, Ranjay},
+  journal={arXiv preprint arXiv:2602.19313},
+  year={2026}
+}
+```
+
+## License
+
+The original TOPReward codebase is MIT-licensed. The LeRobot port follows the LeRobot Apache 2.0 license; the wrapped Qwen3-VL weights are subject to the original Qwen license.

docs/source/using_dataset_tools.mdx

@@ -117,10 +117,10 @@ lerobot-edit-dataset \
     --repo_id lerobot/pusht_image \
     --operation.type convert_image_to_video \
     --operation.output_dir outputs/pusht_video \
-    --operation.vcodec libsvtav1 \
-    --operation.pix_fmt yuv420p \
-    --operation.g 2 \
-    --operation.crf 30
+    --operation.camera_encoder.vcodec libsvtav1 \
+    --operation.camera_encoder.pix_fmt yuv420p \
+    --operation.camera_encoder.g 2 \
+    --operation.camera_encoder.crf 30
 
 # Convert only specific episodes
 lerobot-edit-dataset \
@@ -147,11 +147,7 @@ lerobot-edit-dataset \
 **Parameters:**
 
 - `output_dir`: Custom output directory (optional - by default uses `new_repo_id` or `{repo_id}_video`)
-- `vcodec`: Video codec to use - options: `h264`, `hevc`, `libsvtav1` (default: `libsvtav1`)
-- `pix_fmt`: Pixel format - options: `yuv420p`, `yuv444p` (default: `yuv420p`)
-- `g`: Group of pictures (GOP) size - lower values give better quality but larger files (default: 2)
-- `crf`: Constant rate factor - lower values give better quality but larger files, 0 is lossless (default: 30)
-- `fast_decode`: Fast decode tuning option (default: 0)
+- `camera_encoder`: Video encoder settings — all sub-fields accessible via `--operation.camera_encoder.<field>. See [Video Encoding Parameters](./video_encoding_parameters) for more details.
 - `episode_indices`: List of specific episodes to convert (default: all episodes)
 - `num_workers`: Number of parallel workers for processing (default: 4)
 

docs/source/video_encoding_parameters.mdx

@@ -0,0 +1,117 @@
+# Video encoding parameters
+
+When video storage is enabled, LeRobot stores each camera stream as an **MP4** file instead of saving one image file per timestep. Video encoding compresses across time, which usually cuts dataset size and I/O compared to a pile of PNG, while keeping MP4 — a format every player and loader understands.
+
+Encoding frames into an MP4 is a full FFmpeg pipeline: choice of encoder, pixel format, GOP/keyframes, quality vs. speed, and optional extra encoder flags. Most of these knobs are user-tunable through `camera_encoder`, a nested `VideoEncoderConfig` (`lerobot.configs.video.VideoEncoderConfig`) passed through PyAV.
+
+You can set these parameters from the CLI with `--dataset.camera_encoder.<field>` (e.g. with `lerobot-record` or `lerobot-rollout`). The same block applies to every camera video stream in that run.
+
+<Tip>
+  Video storage must be on for `camera_encoder` to have any effect —
+  `use_videos=True` in Python APIs, or `--dataset.video=true` on the CLI (the
+  recording default). With video off, inputs stay as images and `camera_encoder`
+  is ignored.
+</Tip>
+
+For details on **when** frames are written vs. encoded (streaming vs. post-episode), queues, and other top-level `--dataset.*` switches, see [Streaming Video Encoding](./streaming_video_encoding). For an encoding-parameter comparison and experiments, see the [video-benchmark Space](https://huggingface.co/spaces/lerobot/video-benchmark).
+
+---
+
+## Example
+
+```bash
+lerobot-record \
+    --robot.type=so100_follower \
+    --robot.port=/dev/tty.usbmodem58760431541 \
+    --robot.cameras="{laptop: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
+    --robot.id=black \
+    --teleop.type=so100_leader \
+    --teleop.port=/dev/tty.usbmodem58760431551 \
+    --teleop.id=blue \
+    --dataset.repo_id=<my_username>/<my_dataset_name> \
+    --dataset.num_episodes=2 \
+    --dataset.single_task="Grab the cube" \
+    --dataset.streaming_encoding=true \
+    --dataset.encoder_threads=2 \
+    --dataset.camera_encoder.vcodec=h264 \
+    --dataset.camera_encoder.preset=fast \
+    --dataset.camera_encoder.extra_options={"tune": "film", "profile:v": "high", "bf": 2} \
+    --display_data=true
+```
+
+---
+
+## Tuning parameters
+
+<Tip warning={true}>
+The defaults are tuned to balance **compression ratio**, **visual quality**, and **decoding/seek speed** for typical robotics datasets. Changing them can affect both recording (CPU load, frame drops) and training (decoding throughput, image quality).
+
+Only override these parameters if you have a specific reason to, and measure the impact on your pipeline before relying on the new settings.
+
+</Tip>
+
+All flags below are prefixed with `--dataset.camera_encoder.` on the CLI.
+
+| Parameter       | Type             | Default       | Description                                                                                                                                                                            |
+| --------------- | ---------------- | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `vcodec`        | `str`            | `"libsvtav1"` | Video codec name. `"auto"` picks the first available hardware encoder from a fixed preference list, falling back to `libsvtav1`.                                                       |
+| `pix_fmt`       | `str`            | `"yuv420p"`   | Output pixel format. Must be supported by the chosen codec in your FFmpeg build.                                                                                                       |
+| `g`             | `int`            | `2`           | GOP size — a keyframe every `g` frames. Emitted as FFmpeg option `g`.                                                                                                                  |
+| `crf`           | `int` or `float` | `30`          | Abstract quality value, mapped per codec (see the [mapping](#mapping-videoencoderconfig--ffmpeg-options) below). Lower → higher quality / larger output where the mapping is monotone. |
+| `preset`        | `int` or `str`   | `12` \*       | Encoder speed preset; meaning depends on the codec. <br/>\* When unset and `vcodec=libsvtav1`, LeRobot defaults to `12`.                                                               |
+| `fast_decode`   | `int`            | `0`           | `libsvtav1`: `0–2`, passed via `svtav1-params`. <br/>`h264` / `hevc` (software): if `>0`, sets `tune=fastdecode`. <br/>Other codecs: usually unused.                                   |
+| `video_backend` | `str`            | `"pyav"`      | Only `"pyav"` is currently implemented for video encoding.                                                                                                                             |
+| `extra_options` | `dict`           | `{}`          | Extra FFmpeg or codec specific options merged after the structured fields above. Cannot override keys already set by those fields.                                                     |
+
+---
+
+## Persistence in dataset metadata
+
+After the first episode of a video stream is encoded, the encoder configuration is **persisted into the dataset metadata** (`meta/info.json`) under each video feature, alongside the values probed from the file itself. For a video feature `observation.images.<camera>`, the layout in `info.json` is:
+
+```json
+{
+  "features": {
+    "observation.images.laptop": {
+      "dtype": "video",
+      "shape": [480, 640, 3],
+      "info": {
+        "video.height": 480,
+        "video.width": 640,
+        "video.codec": "h264",
+        "video.pix_fmt": "yuv420p",
+        "video.fps": 30,
+        "video.channels": 3,
+        "video.is_depth_map": false,
+        "video.g": 2,
+        "video.crf": 30,
+        "video.preset": "fast",
+        "video.fast_decode": 0,
+        "video.video_backend": "pyav",
+        "video.extra_options": { "tune": "film", "profile:v": "high", "bf": 2 }
+      }
+    }
+  }
+}
+```
+
+Two sources contribute to the `info` block:
+
+- **Stream-derived** (read back from the encoded MP4 with PyAV): `video.height`, `video.width`, `video.codec`, `video.pix_fmt`, `video.fps`, `video.channels`, `video.is_depth_map`, plus `audio.*` if an audio stream is present.
+- **Encoder-derived** (taken from `VideoEncoderConfig`): `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.video_backend`, `video.extra_options`.
+
+<Tip>
+  This block is populated **once**, from the **first** episode. It assumes every
+  episode in the dataset was encoded with the same `camera_encoder`. Changing
+  encoder settings partway through a recording is not supported — the
+  `info.json` will only reflect the parameters used for the first episode.
+</Tip>
+
+---
+
+## Merging datasets
+
+When aggregating datasets with `merge_datasets`, video files are concatenated as-is (no re-encoding), and encoder fields in `info.json` are merged per-key:
+
+- **Stream-derived fields must match** across sources: `video.codec`, `video.pix_fmt`, `video.height`, `video.width`, `video.fps`. Otherwise FFmpeg's concat demuxer fails.
+- **Encoder-tuning fields are merged loosely**: `video.g`, `video.crf`, `video.preset`, `video.fast_decode`, `video.extra_options`. If every source agrees, the value is kept; if not, it's set to `null` (or `{}` for `video.extra_options`) and a warning is logged.

examples/notebooks/quickstart.ipynb

@@ -80,7 +80,7 @@
     "}\n",
     "\n",
     "# Dataset\n",
-    "HF_USER = \"your_hf_username\"  # `huggingface-cli whoami` to find your username\n",
+    "HF_USER = \"your_hf_username\"  # `hf auth whoami` to find your username\n",
     "DATASET_NAME = \"my_so101_dataset\"\n",
     "TASK_DESCRIPTION = \"pick and place the block\"\n",
     "NUM_EPISODES = 10\n",
@@ -291,7 +291,34 @@
     "\n",
     "Uses `POLICY_PATH` from the Configuration cell (defaults to the Hub repo ID). You can also put there the `LAST_CHECKPOINT_PATH`.\n",
     "\n",
-    "See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details."
+    "See the [inference docs](https://huggingface.co/docs/lerobot/il_robots#run-inference-and-evaluate-your-policy) for details.\n",
+    "\n",
+    "Recently ```lerobot-rollout``` was introduced, you can [read more about it here](https://huggingface.co/docs/lerobot/main/en/il_robots?eval=Base+mode+%28no+recording%29#run-inference-and-evaluate-your-policy)."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print_cmd(\n",
+    "    \"lerobot-rollout\",\n",
+    "    \"--strategy.type=base\",\n",
+    "    f\"--policy.path={POLICY_PATH}\",\n",
+    "    f\"--robot.type={ROBOT_TYPE}\",\n",
+    "    f\"--robot.port={ROBOT_PORT}\",\n",
+    "    CAMERAS_FLAG,\n",
+    "    f'--task=\"{TASK_DESCRIPTION}\"',\n",
+    "    \"--duration=60\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "if you are using the V0.5.1 release you should use ```lerobot-record``` instead of rollout"
    ]
   },
   {

examples/omx/README.md

@@ -0,0 +1,136 @@
+# OMX Follower — Cube Pick And Place Example
+
+This is an example of what is possible to do with LeRobot on a physical setup.
+It is a WIP and being used internally at LeRobot and specific to our setup, but we hope it can be a useful reference for how to use LeRobot APIs and CLIs.
+
+It includes an end-to-end example for the **OMX Follower** robot arm: pick and place a cube dataset, train a policy, and deploy it autonomously.
+
+## Hardware
+
+| Component | Value                                |
+| --------- | ------------------------------------ |
+| Robot     | OMX Follower                         |
+| Cameras   | 2× OpenCV cameras (wrist + top-down) |
+
+## Scripts
+
+| Script                 | Purpose                                                         |
+| ---------------------- | --------------------------------------------------------------- |
+| `reset_environment.py` | Standalone utility: sweep workspace, grab cube, place cube      |
+| `record_grab.py`       | Automated data collection: reset → place → record grab episodes |
+
+## Setup
+
+Make sure you have LeRobot installed in your env. (See [the installation guide](https://huggingface.co/docs/lerobot/installation))
+
+Next, we will declare some environment variables for convenience. Adjust the camera indices and robot port to match your system configuration.
+
+```bash
+export ROBOT_PORT=/dev/ttyACM0
+export TELEOP_PORT=/dev/ttyACM1
+export HF_USERNAME=<your_hf_username>
+export ROBOT_CAMERAS="{ wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30, fourcc: MJPG}, top: {type: opencv, index_or_path: 2, width: 640, height: 480, fps: 30, fourcc: MJPG} }"
+```
+
+## Step 1 — Collect Data
+
+```bash
+lerobot-record \
+    --robot.type=omx_follower \
+    --robot.port=$ROBOT_PORT \
+    --robot.id=omx_follower \
+    --robot.cameras="$ROBOT_CAMERAS" \
+    --teleop.type=omx_leader \
+    --teleop.port=$TELEOP_PORT \
+    --teleop.id=omx_leader \
+    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
+    --dataset.root=data/omx_pickandplace \
+    --dataset.num_episodes=50 \
+    --dataset.single_task="Pick the cube and place it in the blue square" \
+    --dataset.streaming_encoding=true \
+    --dataset.push_to_hub=true
+```
+
+### Bonus Auto-Collect script
+
+/!\ This is specific to our setup and the task of picking and placing a cube. It is not a general-purpose data collection script. As you may notice, it doesn't require a teleop.
+
+```bash
+python -m examples.omx.record_grab \
+    --robot.type=omx_follower \
+    --robot.port=$ROBOT_PORT \
+    --robot.id=omx_follower \
+    --robot.cameras="$ROBOT_CAMERAS" \
+    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
+    --dataset.root=data/omx_pickandplace \
+    --dataset.num_episodes=50 \
+    --dataset.single_task="Pick the cube and place it in the blue square" \
+    --dataset.streaming_encoding=true \
+    --dataset.push_to_hub=true
+```
+
+Each episode:
+
+1. The arm grabs the cube from the center of the workspace and places it at a random position.
+2. The arm returns to HOME.
+3. A targeted grab is recorded: HOME → approach raised → lower onto cube → grasp → lift → carry → drop → HOME.
+
+A dataset is already available here [`maximellerbach/omx_pickandplace`](https://huggingface.co/datasets/maximellerbach/omx_pickandplace), so you can skip directly to training if you want.
+
+## Step 2 — Train
+
+To train a simple `ACT` policy on the collected dataset, you can use the `lerobot-train` CLI:
+
+```bash
+lerobot-train \
+    --dataset.repo_id=$HF_USERNAME/omx_pickandplace \
+    --policy.type=act \
+    --output_dir=outputs/train/omx_pickandplace_act \
+    --policy.device=cuda \
+    --policy.repo_id=$HF_USERNAME/omx_pickandplace_act \
+    --steps=20000 \
+    --wandb.enable=true
+```
+
+A pretrained `ACT` policy is already available here [`maximellerbach/omx_pickandplace_act`](https://huggingface.co/maximellerbach/omx_pickandplace_act).
+
+## Step 3 — Rollout
+
+Use the `lerobot-rollout` CLI with base strategy:
+
+```bash
+lerobot-rollout \
+    --strategy.type=base \
+    --robot.type=omx_follower \
+    --robot.port=$ROBOT_PORT \
+    --robot.id=omx_follower \
+    --robot.cameras="$ROBOT_CAMERAS" \
+    --policy.path=$HF_USERNAME/omx_pickandplace_act \
+```
+
+For continuous recording with automatic upload (sentry mode):
+
+```bash
+lerobot-rollout \
+    --strategy.type=sentry \
+    --strategy.upload_every_n_episodes=10 \
+    --robot.type=omx_follower \
+    --robot.port=$ROBOT_PORT \
+    --robot.id=omx_follower \
+    --robot.cameras="$ROBOT_CAMERAS" \
+    --policy.path=$HF_USERNAME/omx_pickandplace_act \
+    --dataset.repo_id=$HF_USERNAME/rollout_omx_pickandplace_act \
+```
+
+## Environment Reset Utility
+
+Those are specific to this particular physical setup. Those are scripts that execute hardcoded sequences of actions on the robot to reset the environment, which is useful for data collection and evaluation. They are not general-purpose scripts.
+
+`reset_environment.py` can be run standalone to prepare the workspace:
+
+```bash
+# Grab cube + place it at a random position on the left side
+python -m examples.omx.reset_environment --port $ROBOT_PORT --mode grab_and_place
+```
+
+It also exposes `grab_cube(robot)` and `place_cube(robot)` for use in custom scripts.

examples/omx/record_grab.py

@@ -0,0 +1,422 @@
+#!/usr/bin/env python3
+"""
+Auto-record grab episodes for the OMX robot arm.
+
+Each episode cycle:
+  1. grab_and_place  — grab cube from workspace center and place at a random (pan, reach) position
+  2. HOME            — return arm to home with gripper open
+  3. record_grab     — execute a targeted grab to the stored position while recording
+                       observations + actions to a LeRobotDataset
+
+Usage (run from repo root):
+    python -m examples.omx.record_grab \\
+        --robot.type=omx_follower \\
+        --robot.port=/dev/ttyACM0 \\
+        --robot.id=omx_follower \\
+        --robot.cameras="{ wrist: {type: opencv, index_or_path: 6, width: 640, height: 480, fps: 30, fourcc: MJPG}, top: {type: opencv, index_or_path: 4, width: 640, height: 480, fps: 30, fourcc: MJPG} }" \\
+        --dataset.repo_id=<hf_username>/<dataset_name> \\
+        --dataset.root=data/omx_grab \\
+        --dataset.num_episodes=50 \\
+        --dataset.single_task="Grab the cube" \\
+        --dataset.streaming_encoding=true
+"""
+
+import logging
+from dataclasses import dataclass
+from pprint import pformat
+
+import numpy as np
+
+from lerobot.cameras import CameraConfig  # noqa: F401
+from lerobot.cameras.opencv import OpenCVCameraConfig  # noqa: F401
+from lerobot.configs import parser
+from lerobot.configs.dataset import DatasetRecordConfig
+from lerobot.datasets import (
+    LeRobotDataset,
+    VideoEncodingManager,
+    aggregate_pipeline_dataset_features,
+    create_initial_features,
+)
+from lerobot.processor import make_default_processors
+from lerobot.robots import RobotConfig, make_robot_from_config
+from lerobot.robots.omx_follower import OmxFollower
+from lerobot.utils.constants import ACTION, OBS_STR
+from lerobot.utils.feature_utils import build_dataset_frame, combine_feature_dicts
+from lerobot.utils.robot_utils import precise_sleep
+
+from .reset_environment import (
+    APPROACH_SPEED,
+    GRIPPER_CLOSE_POS,
+    HOME_POSE,
+    PUSH_END_ELBOW_FLEX,
+    PUSH_END_SHOULDER_LIFT,
+    PUSH_START_ELBOW_FLEX,
+    PUSH_START_SHOULDER_LIFT,
+    array_to_pose,
+    grab_cube,
+    horizontal_wrist_flex,
+    move_to_pose,
+    place_cube,
+    pose_to_array,
+)
+
+# ── Grab-episode motion parameters ────────────────────────────────────────────
+
+# Shoulder-lift offset for the raised approach phase (subtracted from the target sl, arm is higher).
+GRAB_RAISE_SL_OFFSET = 20.0
+GRAB_LOWER_SPEED = 20.0
+RECORD_SPEED = 30.0
+
+# Pose the arm travels to after closing the gripper (cube held).
+GRAB_CARRY_POSE = {
+    "shoulder_pan.pos": -23.0,
+    "shoulder_lift.pos": 5.0,
+    "elbow_flex.pos": 18.0,
+    "wrist_flex.pos": -14.0,
+    "wrist_roll.pos": 0.0,
+    "gripper.pos": GRIPPER_CLOSE_POS,
+}
+
+# Per-joint jitter limits (degrees) applied to transit waypoints for human-like variation.
+# Cube-approach and carry poses are never jittered to preserve precision.
+_JITTER_LIMITS: dict[str, float] = {
+    "shoulder_pan.pos": 5.0,
+    "shoulder_lift.pos": 4.0,
+    "elbow_flex.pos": 4.0,
+    "wrist_flex.pos": 3.0,
+    "wrist_roll.pos": 2.0,
+    "gripper.pos": 0.0,
+}
+
+
+def _jitter_pose(pose: dict, rng: np.random.Generator) -> dict:
+    """Return a copy of pose with independent per-joint random perturbations."""
+    return {
+        k: v + rng.uniform(-_JITTER_LIMITS.get(k, 0.0), _JITTER_LIMITS.get(k, 0.0)) for k, v in pose.items()
+    }
+
+
+def _random_stuck_pose(rng: np.random.Generator) -> dict:
+    """Return a physically plausible stuck pose (failed grasp), gripper closed.
+
+    ef bounds are piecewise-linear in sl so the arm stays in a reachable,
+    table-safe envelope across the full sl range:
+      sl=-50 → ef ∈ [  0,  50]   (arm raised, can be bent forward)
+      sl=  0 → ef ∈ [-25,  25]   (mid reach)
+      sl= 30 → ef ∈ [-20,   0]   (arm extended, little room to flex)
+    wrist_flex is randomly offset from the horizontal value.
+    """
+    pan = float(rng.uniform(-5.0, 35.0))
+    sl = float(rng.uniform(-50.0, 30.0))
+
+    if sl <= 0.0:
+        alpha = (sl + 50.0) / 50.0  # 0 at sl=-50, 1 at sl=0
+        ef_lo = alpha * -25.0  # 0 → -25
+        ef_hi = 50.0 + alpha * -25.0  # 50 → 25
+    else:
+        alpha = sl / 30.0  # 0 at sl=0, 1 at sl=30
+        ef_lo = -25.0 + alpha * 5.0  # -25 → -20
+        ef_hi = 25.0 + alpha * -25.0  # 25 → 0
+
+    ef = float(rng.uniform(ef_lo, ef_hi))
+    wf = horizontal_wrist_flex(sl, ef) + float(rng.uniform(-15.0, 15.0))
+    return {
+        "shoulder_pan.pos": pan,
+        "shoulder_lift.pos": sl,
+        "elbow_flex.pos": ef,
+        "wrist_flex.pos": wf,
+        "wrist_roll.pos": float(rng.uniform(-15.0, 15.0)),
+        "gripper.pos": GRIPPER_CLOSE_POS,
+    }
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class OmxRecordGrabConfig:
+    robot: RobotConfig
+    dataset: DatasetRecordConfig
+    # Resume recording on an existing dataset.
+    resume: bool = False
+    # Fraction of episodes that start from a random stuck pose (gripper closed) to
+    # generate recovery data.  0.0 = disabled, 1.0 = all episodes are recovery starts.
+    recovery_prob: float = 0.5
+
+
+def record_episode_spline(
+    robot: OmxFollower,
+    waypoints: list[dict],
+    speeds: list[float],
+    dataset: LeRobotDataset,
+    task: str,
+) -> None:
+    """Execute a Catmull-Rom-style spline through waypoints, recording each frame.
+
+    Segment durations are parameterized from the maximum absolute joint delta
+    between consecutive waypoints divided by the requested segment speed,
+    producing non-uniform timing in joint space. Interior tangents are derived
+    from the adjacent per-segment velocities, with clamped (zero-velocity)
+    endpoints so the arm starts and stops smoothly. Each segment is cubic
+    Hermite, giving C1 continuity at every waypoint.
+    """
+    pts = [pose_to_array(w) for w in waypoints]
+    n = len(pts)
+
+    # Steps and duration per segment
+    n_steps_list = []
+    timestamps = []
+    for i in range(n - 1):
+        max_dist = float(np.max(np.abs(pts[i + 1] - pts[i])))
+        ns = max(1, int(max_dist / speeds[i] * dataset.fps)) if max_dist >= 0.5 else 0
+        n_steps_list.append(ns)
+        timestamps.append(ns / dataset.fps)
+
+    # Velocity tangents (deg/sec) — clamped at endpoints, Catmull-Rom for interior
+    vels = [np.zeros_like(pts[0])]
+    for i in range(1, n - 1):
+        v_prev = (pts[i] - pts[i - 1]) / timestamps[i - 1] if timestamps[i - 1] > 0 else np.zeros_like(pts[0])
+        v_next = (pts[i + 1] - pts[i]) / timestamps[i] if timestamps[i] > 0 else np.zeros_like(pts[0])
+        vels.append(0.5 * (v_prev + v_next))
+    vels.append(np.zeros_like(pts[0]))
+
+    dt = 1.0 / dataset.fps
+    for seg in range(n - 1):
+        ns = n_steps_list[seg]
+        if ns == 0:
+            continue
+        p0, p1 = pts[seg], pts[seg + 1]
+        # Scale velocity (deg/sec) to t-space tangent (deg/t-unit, where t: 0→1 over ns steps)
+        m0 = vels[seg] * timestamps[seg]
+        m1 = vels[seg + 1] * timestamps[seg]
+
+        for step in range(1, ns + 1):
+            t = step / ns
+            h00 = 2 * t**3 - 3 * t**2 + 1
+            h10 = t**3 - 2 * t**2 + t
+            h01 = -2 * t**3 + 3 * t**2
+            h11 = t**3 - t**2
+            commanded = h00 * p0 + h10 * m0 + h01 * p1 + h11 * m1
+
+            action = array_to_pose(commanded)
+            robot.send_action(action)
+            obs = robot.get_observation()
+            obs_frame = build_dataset_frame(dataset.features, obs, prefix=OBS_STR)
+            action_frame = build_dataset_frame(dataset.features, action, prefix=ACTION)
+            dataset.add_frame({**obs_frame, **action_frame, "task": task})
+            precise_sleep(dt)
+
+
+def record_grab_episode(
+    robot: OmxFollower,
+    dataset: LeRobotDataset,
+    pan: float,
+    t: float,
+    task: str,
+    recovery_start: bool = False,
+) -> None:
+    """Execute a targeted grab to the stored (pan, t) position, recording every frame.
+
+    Normal sequence (initial HOME move is NOT recorded):
+      HOME → raised approach above cube → lower → close gripper
+           → raise [jittered] → retract [jittered] → GRAB_CARRY_POSE → drop → HOME
+
+    Recovery sequence (recovery_start=True): arm is moved to a random stuck pose
+    (gripper closed) without recording, then recording begins from there:
+      stuck_pose → raised approach above cube → [normal grab sequence from there]
+
+    All segments are joined by a Catmull-Rom spline (C1-continuous velocities).
+    """
+    sl = PUSH_START_SHOULDER_LIFT + t * (PUSH_END_SHOULDER_LIFT - PUSH_START_SHOULDER_LIFT)
+    ef = PUSH_START_ELBOW_FLEX + t * (PUSH_END_ELBOW_FLEX - PUSH_START_ELBOW_FLEX)
+    sl_raised = sl - GRAB_RAISE_SL_OFFSET
+    wf_horizontal = horizontal_wrist_flex(sl, ef)
+
+    rng = np.random.default_rng()
+
+    if recovery_start:
+        stuck_pose = _random_stuck_pose(rng)
+        logger.info(f"Recovery start: {stuck_pose}")
+        move_to_pose(robot, stuck_pose, APPROACH_SPEED)
+        first_waypoints = [stuck_pose]
+        first_speeds = []
+    else:
+        jittery_start = _jitter_pose(HOME_POSE, rng)
+        move_to_pose(robot, jittery_start, APPROACH_SPEED)
+        first_waypoints = [jittery_start]
+        first_speeds = []
+
+    waypoints = first_waypoints + [
+        {  # raised approach: arm above cube
+            "shoulder_pan.pos": pan,
+            "shoulder_lift.pos": sl_raised,
+            "elbow_flex.pos": ef,
+            "wrist_flex.pos": horizontal_wrist_flex(sl_raised, ef),
+            "wrist_roll.pos": 0.0,
+            "gripper.pos": 60.0,
+        },
+        {  # lower onto cube — no jitter: precision needed
+            "shoulder_pan.pos": pan,
+            "shoulder_lift.pos": sl,
+            "elbow_flex.pos": ef,
+            "wrist_flex.pos": wf_horizontal,
+            "wrist_roll.pos": 0.0,
+            "gripper.pos": 60.0,
+        },
+        {  # close gripper — no jitter: precision needed
+            "shoulder_pan.pos": pan,
+            "shoulder_lift.pos": sl,
+            "elbow_flex.pos": ef,
+            "wrist_flex.pos": wf_horizontal,
+            "wrist_roll.pos": 0.0,
+            "gripper.pos": GRIPPER_CLOSE_POS,
+        },
+        _jitter_pose(
+            {  # raise with cube
+                "shoulder_pan.pos": pan,
+                "shoulder_lift.pos": sl_raised,
+                "elbow_flex.pos": ef,
+                "wrist_flex.pos": horizontal_wrist_flex(sl_raised, ef),
+                "wrist_roll.pos": 0.0,
+                "gripper.pos": GRIPPER_CLOSE_POS,
+            },
+            rng,
+        ),
+        _jitter_pose(
+            {  # retract: fold arm toward HOME before sweeping to carry zone
+                "shoulder_pan.pos": pan * 0.25,
+                "shoulder_lift.pos": HOME_POSE["shoulder_lift.pos"] + 5.0,
+                "elbow_flex.pos": HOME_POSE["elbow_flex.pos"] - 5.0,
+                "wrist_flex.pos": 0.0,
+                "wrist_roll.pos": 0.0,
+                "gripper.pos": GRIPPER_CLOSE_POS,
+            },
+            rng,
+        ),
+        GRAB_CARRY_POSE,  # no jitter: target drop zone
+        {**GRAB_CARRY_POSE, "gripper.pos": 60.0},  # drop cube
+        HOME_POSE,
+    ]
+    speeds = first_speeds + [
+        RECORD_SPEED,  # (HOME →) raised approach
+        GRAB_LOWER_SPEED,  # raised approach → lower
+        GRAB_LOWER_SPEED,  # lower → close gripper
+        RECORD_SPEED,  # close gripper → raise
+        RECORD_SPEED,  # raise → retract
+        RECORD_SPEED,  # retract → carry pose
+        RECORD_SPEED,  # carry pose → drop
+        RECORD_SPEED,  # drop → HOME
+    ]
+
+    record_episode_spline(robot, waypoints, speeds, dataset, task)
+
+    # Dwell at HOME for ~0.5 s before next episode
+    home_action = build_dataset_frame(dataset.features, HOME_POSE, prefix=ACTION)
+    dt = 1.0 / dataset.fps
+    for _ in range(int(dataset.fps * 0.5)):
+        robot.send_action(HOME_POSE)
+        obs = robot.get_observation()
+        obs_frame = build_dataset_frame(dataset.features, obs, prefix=OBS_STR)
+        dataset.add_frame({**obs_frame, **home_action, "task": task})
+        precise_sleep(dt)
+
+
+@parser.wrap()
+def record_grab(cfg: OmxRecordGrabConfig) -> LeRobotDataset:
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+    logger.info(pformat(cfg))
+
+    robot = make_robot_from_config(cfg.robot)
+    use_videos = cfg.dataset.video
+
+    teleop_action_processor, _, robot_obs_processor = make_default_processors()
+
+    dataset_features = combine_feature_dicts(
+        aggregate_pipeline_dataset_features(
+            pipeline=teleop_action_processor,
+            initial_features=create_initial_features(action=robot.action_features),
+            use_videos=use_videos,
+        ),
+        aggregate_pipeline_dataset_features(
+            pipeline=robot_obs_processor,
+            initial_features=create_initial_features(observation=robot.observation_features),
+            use_videos=use_videos,
+        ),
+    )
+
+    num_cameras = len(robot.cameras) if hasattr(robot, "cameras") else 0
+    dataset = None
+
+    try:
+        if cfg.resume:
+            dataset = LeRobotDataset.resume(
+                cfg.dataset.repo_id,
+                root=cfg.dataset.root,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                vcodec=cfg.dataset.vcodec,
+                encoder_threads=cfg.dataset.encoder_threads,
+                image_writer_processes=cfg.dataset.num_image_writer_processes if num_cameras > 0 else 0,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * num_cameras
+                if num_cameras > 0
+                else 0,
+            )
+        else:
+            cfg.dataset.stamp_repo_id()
+            dataset = LeRobotDataset.create(
+                cfg.dataset.repo_id,
+                cfg.dataset.fps,
+                root=cfg.dataset.root,
+                robot_type=robot.name,
+                features=dataset_features,
+                use_videos=use_videos,
+                streaming_encoding=cfg.dataset.streaming_encoding,
+                batch_encoding_size=cfg.dataset.video_encoding_batch_size,
+                vcodec=cfg.dataset.vcodec,
+                encoder_threads=cfg.dataset.encoder_threads,
+                image_writer_processes=cfg.dataset.num_image_writer_processes if num_cameras > 0 else 0,
+                image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * num_cameras
+                if num_cameras > 0
+                else 0,
+            )
+
+        robot.connect(calibrate=True)
+
+        rng = np.random.default_rng()
+        with VideoEncodingManager(dataset):
+            for episode_idx in range(cfg.dataset.num_episodes):
+                logger.info(f"=== Episode {episode_idx + 1}/{cfg.dataset.num_episodes} ===")
+
+                logger.info("Step 1: grabbing and placing cube...")
+                grab_cube(robot)
+                pan, t = place_cube(robot)
+                logger.info(f"Cube placed at pan={pan:.1f}, reach={t:.2f}")
+
+                recovery_start = cfg.recovery_prob > 0 and float(rng.random()) < cfg.recovery_prob
+                logger.info(f"Step 2: recording {'recovery ' if recovery_start else ''}grab episode...")
+                record_grab_episode(
+                    robot,
+                    dataset,
+                    pan,
+                    t,
+                    cfg.dataset.single_task,
+                    recovery_start=recovery_start,
+                )
+
+                dataset.save_episode()
+                logger.info(f"Episode {episode_idx + 1} saved.")
+
+    finally:
+        if dataset:
+            dataset.finalize()
+        if robot.is_connected:
+            robot.disconnect()
+
+    if cfg.dataset.push_to_hub and dataset and dataset.num_episodes > 0:
+        dataset.push_to_hub(tags=cfg.dataset.tags, private=cfg.dataset.private)
+
+    return dataset
+
+
+if __name__ == "__main__":
+    record_grab()

examples/omx/reset_environment.py

@@ -0,0 +1,267 @@
+#!/usr/bin/env python3
+"""
+Auto-reset and cube-grab utility for the OMX robot arm.
+
+Provides:
+  - grab_cube(robot): sweep workspace, center cube, close gripper
+  - place_cube(robot): carry cube to a random position, release
+
+Standalone usage (run from repo root):
+    python -m examples.omx.reset_environment --port /dev/ttyACM1 --mode grab
+    python -m examples.omx.reset_environment --port /dev/ttyACM1 --mode grab_and_place
+
+Joint range: -100 to 100 for arm joints; gripper: 50 = closed, 80 = open.
+
+To read current joint values for calibration, add after robot.connect():
+    obs = robot.get_observation()
+    print({k: round(obs[k], 1) for k in JOINT_NAMES})
+    robot.disconnect(); raise SystemExit
+
+Parallel-to-ground IK: wrist_flex = WRIST_HORIZONTAL_OFFSET - shoulder_lift - elbow_flex.
+Linear interpolation preserves this constraint between any two poses that satisfy it.
+"""
+
+import argparse
+import logging
+
+import numpy as np
+
+from lerobot.robots.omx_follower import OmxFollower, OmxFollowerConfig
+from lerobot.robots.robot import Robot
+from lerobot.utils.robot_utils import precise_sleep
+
+logger = logging.getLogger(__name__)
+
+# ── Poses ─────────────────────────────────────────────────────────────────────
+
+HOME_POSE = {
+    "shoulder_pan.pos": 0.0,
+    "shoulder_lift.pos": -50.0,
+    "elbow_flex.pos": 50.0,
+    "wrist_flex.pos": 0.0,
+    "wrist_roll.pos": 0.0,
+    "gripper.pos": 60.0,
+}
+
+SWEEP_WAYPOINTS = [
+    {
+        "shoulder_pan.pos": -60.0,
+        "shoulder_lift.pos": 50.0,
+        "elbow_flex.pos": -60.0,
+        "wrist_flex.pos": -20.0,
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": 60.0,
+    },
+    {
+        "shoulder_pan.pos": -30.0,
+        "shoulder_lift.pos": 50.0,
+        "elbow_flex.pos": -60.0,
+        "wrist_flex.pos": -5.0,
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": 60.0,
+    },
+    {
+        "shoulder_pan.pos": 20.0,
+        "shoulder_lift.pos": 50.0,
+        "elbow_flex.pos": -55.0,
+        "wrist_flex.pos": -5.0,
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": 60.0,
+    },
+]
+
+# ── Motion parameters ─────────────────────────────────────────────────────────
+
+CONTROL_HZ = 30
+APPROACH_SPEED = 50.0
+SWEEP_SPEED = 40.0
+
+# ── Grab-sequence parameters ──────────────────────────────────────────────────
+
+GRAB_PAN = 0.0
+SWEEP_LEFT_PAN = -60.0
+SWEEP_RIGHT_PAN = 60.0
+SWEEP_END_OFFSET = 5.0  # stop before center so the cube isn't pushed past GRAB_PAN
+SWEEP_END_PAN_RANGE = (15.0, 20.0)
+
+SWEEP_LOW_SHOULDER_LIFT = 50.0
+SWEEP_LOW_ELBOW_FLEX_START = -60.0
+SWEEP_LOW_ELBOW_FLEX_END = -55.0
+
+SWEEP_HIGH_WRIST_FLEX = -20.0  # wrist tilted up during high approach to clear obstacles
+
+PUSH_START_SHOULDER_LIFT = 0.0
+PUSH_START_ELBOW_FLEX = 45.0
+PUSH_END_SHOULDER_LIFT = 50.0
+PUSH_END_ELBOW_FLEX = -50.0
+# Subtracted from shoulder_lift during the push sweep to clear the platform surface.
+# Does not affect the grab-target interpolation in record_grab.py.
+PUSH_RAISE_OFFSET = 5.0
+
+WRIST_HORIZONTAL_OFFSET = 0.0  # tune if gripper tilts during push: + tilts nose up, - down
+GRIPPER_CLOSE_POS = 50.0
+
+PLACE_LEFT_PAN_RANGE = (5.0, 30.0)  # random pan range for cube placement on the left side
+PLACE_REACH_RANGE = (0.1, 0.7)  # 0 = arm retracted (PUSH_START), 1 = fully extended (PUSH_END)
+
+JOINT_NAMES = [
+    "shoulder_pan.pos",
+    "shoulder_lift.pos",
+    "elbow_flex.pos",
+    "wrist_flex.pos",
+    "wrist_roll.pos",
+    "gripper.pos",
+]
+
+# ── Helpers ───────────────────────────────────────────────────────────────────
+
+
+def pose_to_array(pose: dict) -> np.ndarray:
+    return np.array([pose[k] for k in JOINT_NAMES])
+
+
+def array_to_pose(arr: np.ndarray) -> dict:
+    return {k: float(arr[i]) for i, k in enumerate(JOINT_NAMES)}
+
+
+def horizontal_wrist_flex(shoulder_lift: float, elbow_flex: float) -> float:
+    return WRIST_HORIZONTAL_OFFSET - shoulder_lift - elbow_flex
+
+
+def _low_sweep_pose(pan: float, elbow_flex: float, wrist_flex: float | None = None) -> dict:
+    sl = SWEEP_LOW_SHOULDER_LIFT
+    return {
+        "shoulder_pan.pos": pan,
+        "shoulder_lift.pos": sl,
+        "elbow_flex.pos": elbow_flex,
+        "wrist_flex.pos": horizontal_wrist_flex(sl, elbow_flex) if wrist_flex is None else wrist_flex,
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": 60.0,
+    }
+
+
+def _high_sweep_pose(pan: float) -> dict:
+    return {**HOME_POSE, "shoulder_pan.pos": pan, "wrist_flex.pos": SWEEP_HIGH_WRIST_FLEX}
+
+
+def _push_pose(shoulder_lift: float, elbow_flex: float, pan: float = GRAB_PAN, gripper: float = 70.0) -> dict:
+    return {
+        "shoulder_pan.pos": pan,
+        "shoulder_lift.pos": shoulder_lift,
+        "elbow_flex.pos": elbow_flex,
+        "wrist_flex.pos": horizontal_wrist_flex(shoulder_lift, elbow_flex),
+        "wrist_roll.pos": 0.0,
+        "gripper.pos": gripper,
+    }
+
+
+def move_to_pose(robot: Robot, target: dict, speed: float) -> None:
+    """Interpolate from current position to target at the given speed (units/s)."""
+    obs = robot.get_observation()
+    current = np.array([obs[k] for k in JOINT_NAMES])
+    goal = pose_to_array(target)
+
+    max_distance = float(np.max(np.abs(goal - current)))
+    if max_distance < 0.5:
+        return
+
+    n_steps = max(1, int(max_distance / speed * CONTROL_HZ))
+    dt = 1.0 / CONTROL_HZ
+    for step in range(1, n_steps + 1):
+        t = step / n_steps
+        robot.send_action(array_to_pose(current + t * (goal - current)))
+        precise_sleep(dt)
+
+
+# ── Sequences ─────────────────────────────────────────────────────────────────
+
+
+def grab_cube(robot: Robot) -> None:
+    """Left sweep → right sweep → extend arm parallel to ground → close gripper."""
+    move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
+
+    for pan, end_pan in [
+        (SWEEP_LEFT_PAN, GRAB_PAN - SWEEP_END_OFFSET),
+        (SWEEP_RIGHT_PAN, GRAB_PAN + SWEEP_END_OFFSET),
+    ]:
+        logger.info(f"Sweeping {'left' if pan < 0 else 'right'} → center...")
+        move_to_pose(robot, _high_sweep_pose(pan), APPROACH_SPEED)
+        move_to_pose(
+            robot, _low_sweep_pose(pan, SWEEP_LOW_ELBOW_FLEX_START, wrist_flex=-20.0), APPROACH_SPEED
+        )
+        move_to_pose(robot, _low_sweep_pose(end_pan, SWEEP_LOW_ELBOW_FLEX_END, wrist_flex=0.0), SWEEP_SPEED)
+        move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
+
+    logger.info("Extending to push cube into gripper...")
+    move_to_pose(
+        robot,
+        _push_pose(PUSH_START_SHOULDER_LIFT - PUSH_RAISE_OFFSET, PUSH_START_ELBOW_FLEX),
+        APPROACH_SPEED,
+    )
+    move_to_pose(
+        robot,
+        _push_pose(PUSH_END_SHOULDER_LIFT - PUSH_RAISE_OFFSET, PUSH_END_ELBOW_FLEX),
+        SWEEP_SPEED,
+    )
+
+    logger.info("Closing gripper...")
+    move_to_pose(
+        robot,
+        _push_pose(PUSH_END_SHOULDER_LIFT, PUSH_END_ELBOW_FLEX, gripper=GRIPPER_CLOSE_POS),
+        APPROACH_SPEED,
+    )
+
+    logger.info("Grab complete.")
+
+
+def place_cube(robot: Robot) -> tuple[float, float]:
+    """Carry the cube (gripper closed) to a random position on the left side, then release.
+
+    Returns:
+        (pan, t): pan angle and reach scalar [0, 1] of the placement position.
+    """
+    pan = float(np.random.uniform(*PLACE_LEFT_PAN_RANGE))
+    t = float(np.random.uniform(*PLACE_REACH_RANGE))
+    sl = PUSH_START_SHOULDER_LIFT + t * (PUSH_END_SHOULDER_LIFT - PUSH_START_SHOULDER_LIFT)
+    ef = PUSH_START_ELBOW_FLEX + t * (PUSH_END_ELBOW_FLEX - PUSH_START_ELBOW_FLEX)
+    logger.info(f"Placing cube at pan={pan:.1f}, reach={t:.2f}...")
+
+    move_to_pose(robot, {**HOME_POSE, "gripper.pos": GRIPPER_CLOSE_POS}, APPROACH_SPEED)
+    move_to_pose(
+        robot, {**HOME_POSE, "shoulder_pan.pos": pan, "gripper.pos": GRIPPER_CLOSE_POS}, APPROACH_SPEED
+    )
+    move_to_pose(robot, _push_pose(sl, ef, pan=pan, gripper=GRIPPER_CLOSE_POS), APPROACH_SPEED)
+    move_to_pose(robot, _push_pose(sl, ef, pan=pan, gripper=80.0), APPROACH_SPEED)
+    move_to_pose(robot, HOME_POSE, APPROACH_SPEED)
+    logger.info("Place complete.")
+    return pan, t
+
+
+# ── Entry point ───────────────────────────────────────────────────────────────
+
+
+def main():
+    parser = argparse.ArgumentParser(description="OMX arm reset / grab script")
+    parser.add_argument("--port", default="/dev/ttyACM1")
+    parser.add_argument("--robot_id", default="omx_follower")
+    parser.add_argument("--mode", choices=["grab", "grab_and_place"], default="grab_and_place")
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+
+    robot = OmxFollower(OmxFollowerConfig(port=args.port, id=args.robot_id))
+    robot.connect(calibrate=True)
+
+    try:
+        if args.mode == "grab":
+            grab_cube(robot)
+        elif args.mode == "grab_and_place":
+            grab_cube(robot)
+            place_cube(robot)
+
+    finally:
+        robot.disconnect()
+
+
+if __name__ == "__main__":
+    main()

examples/tutorial/rl/hilserl_example.py

@@ -4,13 +4,13 @@ from pathlib import Path
 from queue import Empty, Full
 
 import torch
-import torch.optim as optim
 
 from lerobot.datasets import LeRobotDataset
 from lerobot.envs.configs import HILSerlProcessorConfig, HILSerlRobotEnvConfig
-from lerobot.policies import SACConfig
-from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies import GaussianActorConfig
+from lerobot.policies.gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
 from lerobot.rewards.classifier.modeling_classifier import Classifier
+from lerobot.rl.algorithms.sac import SACAlgorithm, SACAlgorithmConfig
 from lerobot.rl.buffer import ReplayBuffer
 from lerobot.rl.gym_manipulator import make_robot_env
 from lerobot.robots.so_follower import SO100FollowerConfig
@@ -28,7 +28,7 @@ def run_learner(
     transitions_queue: mp.Queue,
     parameters_queue: mp.Queue,
     shutdown_event: mp.Event,
-    policy_learner: SACPolicy,
+    policy_learner: GaussianActorPolicy,
     online_buffer: ReplayBuffer,
     offline_buffer: ReplayBuffer,
     lr: float = 3e-4,
@@ -40,8 +40,9 @@ def run_learner(
     policy_learner.train()
     policy_learner.to(device)
 
-    # Create Adam optimizer from scratch - simple and clean
-    optimizer = optim.Adam(policy_learner.parameters(), lr=lr)
+    algo_config = SACAlgorithmConfig.from_policy_config(policy_learner.config)
+    algorithm = SACAlgorithm(policy=policy_learner, config=algo_config)
+    algorithm.make_optimizers_and_scheduler()
 
     print(f"[LEARNER] Online buffer capacity: {online_buffer.capacity}")
     print(f"[LEARNER] Offline buffer capacity: {offline_buffer.capacity}")
@@ -83,24 +84,26 @@ def run_learner(
                 else:
                     batch[key] = online_batch[key]
 
-            loss, _ = policy_learner.forward(batch)
+            def batch_iter(b=batch):
+                while True:
+                    yield b
 
-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
+            stats = algorithm.update(batch_iter())
             training_step += 1
 
             if training_step % LOG_EVERY == 0:
+                log_dict = stats.to_log_dict()
                 print(
-                    f"[LEARNER] Training step {training_step}, Loss: {loss.item():.4f}, "
+                    f"[LEARNER] Training step {training_step}, "
+                    f"critic_loss: {log_dict.get('critic', 'N/A'):.4f}, "
                     f"Buffers: Online={len(online_buffer)}, Offline={len(offline_buffer)}"
                 )
 
             # Send updated parameters to actor every 10 training steps
             if training_step % SEND_EVERY == 0:
                 try:
-                    state_dict = {k: v.cpu() for k, v in policy_learner.state_dict().items()}
-                    parameters_queue.put_nowait(state_dict)
+                    weights = algorithm.get_weights()
+                    parameters_queue.put_nowait(weights)
                     print("[LEARNER] Sent updated parameters to actor")
                 except Full:
                     # Missing write due to queue not being consumed (should happen rarely)
@@ -113,7 +116,7 @@ def run_actor(
     transitions_queue: mp.Queue,
     parameters_queue: mp.Queue,
     shutdown_event: mp.Event,
-    policy_actor: SACPolicy,
+    policy_actor: GaussianActorPolicy,
     reward_classifier: Classifier,
     env_cfg: HILSerlRobotEnvConfig,
     device: torch.device = "mps",
@@ -144,15 +147,15 @@ def run_actor(
 
             while step < MAX_STEPS_PER_EPISODE and not shutdown_event.is_set():
                 try:
-                    new_params = parameters_queue.get_nowait()
-                    policy_actor.load_state_dict(new_params)
+                    new_weights = parameters_queue.get_nowait()
+                    policy_actor.load_state_dict(new_weights)
                     print("[ACTOR] Updated policy parameters from learner")
                 except Empty:  # No new updated parameters available from learner, waiting
                     pass
 
-                # Get action from policy
+                # Get action from policy (returns full action: continuous + discrete)
                 policy_obs = make_policy_obs(obs, device=device)
-                action_tensor = policy_actor.select_action(policy_obs)  # predicts a single action
+                action_tensor = policy_actor.select_action(policy_obs)
                 action = action_tensor.squeeze(0).cpu().numpy()
 
                 # Step environment
@@ -261,14 +264,14 @@ def main():
     action_features = hw_to_dataset_features(env.robot.action_features, "action")
 
     # Create SAC policy for action selection
-    policy_cfg = SACConfig(
+    policy_cfg = GaussianActorConfig(
         device=device,
         input_features=obs_features,
         output_features=action_features,
     )
 
-    policy_actor = SACPolicy(policy_cfg)
-    policy_learner = SACPolicy(policy_cfg)
+    policy_actor = GaussianActorPolicy(policy_cfg)
+    policy_learner = GaussianActorPolicy(policy_cfg)
 
     demonstrations_repo_id = "lerobot/example_hil_serl_dataset"
     offline_dataset = LeRobotDataset(repo_id=demonstrations_repo_id)

pyproject.toml

@@ -59,8 +59,8 @@ keywords = ["lerobot", "huggingface", "robotics",  "machine learning", "artifici
 
 dependencies = [
     # Core ML
-    "torch>=2.7,<2.13.0",
-    "torchvision>=0.22.0,<0.28.0",
+    "torch>=2.7,<2.12.0",
+    "torchvision>=0.22.0,<0.27.0",
     "numpy>=2.0.0,<2.3.0", # NOTE: Explicitly listing numpy helps the resolver converge faster. Upper bound imposed by opencv-python-headless.
     "opencv-python-headless>=4.9.0,<4.14.0",
     "Pillow>=10.0.0,<13.0.0",
@@ -95,11 +95,22 @@ dependencies = [
 
 # ── Feature-scoped extras ──────────────────────────────────
 dataset = [
-    "datasets>=4.0.0,<5.0.0",
+    "datasets>=4.7.0,<5.0.0",
     "pandas>=2.0.0,<3.0.0", # NOTE: Transitive dependency of datasets
     "pyarrow>=21.0.0,<30.0.0", # NOTE: Transitive dependency of datasets
     "lerobot[av-dep]",
-    "torchcodec>=0.3.0,<0.13.0; sys_platform != 'win32' and (sys_platform != 'linux' or (platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')) and (sys_platform != 'darwin' or platform_machine != 'x86_64')", # NOTE: Windows support starts at version 0.7 (needs torch==2.8), ffmpeg>=8 support starts at version 0.8.1 (needs torch==2.9), system-wide ffmpeg support starts at version 0.10 (needs torch==2.10), 0.11 needs torch==2.11, 0.12 needs torch==2.12.
+
+    # NOTE: torchcodec wheel availability matrix (PyPI):
+    #   - linux x86_64/amd64 + macOS arm64 : wheels since 0.3.0 (the historic supported set).
+    #   - win32 x86_64                     : wheels since 0.7.0  (needs torch>=2.8).
+    #   - linux aarch64/arm64              : wheels since 0.11.0 (needs torch>=2.11).
+    #   - macOS x86_64 (Intel) and linux armv7l: no wheels in any released version -> fall through to the PyAV decoder.
+    # Each platform gets its own line so the resolver picks the minimum version that has a wheel for it.
+
+    # Other torch/torchcodec pairings (informational): 0.8.1 = ffmpeg>=8 support, 0.10 = system-wide ffmpeg support, 0.12 needs torch==2.12.
+    "torchcodec>=0.3.0,<0.12.0; (sys_platform == 'linux' and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64')",
+    "torchcodec>=0.7.0,<0.12.0; sys_platform == 'win32'",
+    "torchcodec>=0.11.0,<0.12.0; sys_platform == 'linux' and (platform_machine == 'aarch64' or platform_machine == 'arm64')",
     "jsonlines>=4.0.0,<5.0.0",
 ]
 training = [
@@ -140,6 +151,8 @@ pyserial-dep = ["pyserial>=3.5,<4.0"]
 deepdiff-dep = ["deepdiff>=7.0.1,<9.0.0"]
 pynput-dep = ["pynput>=1.7.8,<1.9.0"]
 pyzmq-dep = ["pyzmq>=26.2.1,<28.0.0"]
+motorbridge-dep = ["motorbridge>=0.3.2,<0.4.0"]
+motorbridge-smart-servo-dep = ["motorbridge-smart-servo>=0.0.4,<0.1.0"]
 
 # Motors
 feetech = ["feetech-servo-sdk>=1.0.0,<2.0.0", "lerobot[pyserial-dep]", "lerobot[deepdiff-dep]"]
@@ -163,6 +176,9 @@ unitree_g1 = [
     "lerobot[pygame-dep]",
 ]
 reachy2 = ["reachy2_sdk>=1.0.15,<1.1.0"]
+# Seeed Studio reBot B601-DM follower (motorbridge / CAN) + StarArm102 / reBot Arm 102
+# leader (motorbridge-smart-servo / FashionStar UART servos).
+rebot = ["lerobot[motorbridge-dep]", "lerobot[motorbridge-smart-servo-dep]"]
 kinematics = ["lerobot[placo-dep]"]
 intelrealsense = [
     "pyrealsense2>=2.55.1.6486,<2.57.0 ; sys_platform != 'darwin'",
@@ -193,9 +209,10 @@ groot = [
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
 sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
+topreward = ["lerobot[transformers-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 eo1 = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"]
-hilserl = ["lerobot[transformers-dep]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
+hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]
 
 # Features
 async = ["lerobot[grpcio-dep]", "lerobot[matplotlib-dep]"]
@@ -249,6 +266,7 @@ all = [
     "lerobot[lekiwi]",
     "lerobot[openarms]",
     "lerobot[reachy2]",
+    "lerobot[rebot]",
     "lerobot[kinematics]",
     "lerobot[intelrealsense]",
     "lerobot[diffusion]",
@@ -269,6 +287,7 @@ all = [
     "lerobot[libero]; sys_platform == 'linux'",
     "lerobot[metaworld]",
     "lerobot[sarm]",
+    "lerobot[topreward]",
     "lerobot[peft]",
     # "lerobot[unitree_g1]", TODO: Unitree requires specific installation instructions for unitree_sdk2
 ]
@@ -293,6 +312,20 @@ lerobot-setup-can="lerobot.scripts.lerobot_setup_can:main"
 lerobot-rollout="lerobot.scripts.lerobot_rollout:main"
 
 # ---------------- Tool Configurations ----------------
+
+# cu128 wheels keep broad hardware reach; the driver floor is 570.86.
+# To use a different CUDA variant, reinstall torch with an explicit index, e.g.:
+#   uv pip install --force-reinstall torch torchvision \
+#       --index-url https://download.pytorch.org/whl/cu130
+[[tool.uv.index]]
+name = "pytorch-cu128"
+url = "https://download.pytorch.org/whl/cu128"
+explicit = true
+
+[tool.uv.sources]
+torch = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
+torchvision = [{ index = "pytorch-cu128", marker = "sys_platform == 'linux'" }]
+
 [tool.setuptools.package-data]
 lerobot = ["envs/*.json"]
 

src/lerobot/cameras/opencv/camera_opencv.py

@@ -199,12 +199,13 @@ class OpenCVCamera(Camera):
             DeviceNotConnectedError: If the camera is not connected.
         """
 
-        # Set FOURCC first (if specified) as it can affect available FPS/resolution options
-        if self.config.fourcc is not None:
-            self._validate_fourcc()
         if self.videocapture is None:
             raise DeviceNotConnectedError(f"{self} videocapture is not initialized")
 
+        set_fourcc_after_size_and_fps = platform.system() == "Windows"
+        if self.config.fourcc is not None and not set_fourcc_after_size_and_fps:
+            self._validate_fourcc()
+
         default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
         default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
 
@@ -222,6 +223,11 @@ class OpenCVCamera(Camera):
         else:
             self._validate_fps()
 
+        if self.config.fourcc is not None and set_fourcc_after_size_and_fps:
+            # On Windows with DSHOW, changing the resolution can silently override the FOURCC setting.
+            # Set FOURCC last to make sure the requested pixel format is actually enforced.
+            self._validate_fourcc()
+
     def _validate_fps(self) -> None:
         """Validates and sets the camera's frames per second (FPS)."""
 

src/lerobot/common/train_utils.py

@@ -99,6 +99,7 @@ def save_checkpoint(
         optimizer (Optimizer | None, optional): The optimizer to save the state from. Defaults to None.
         scheduler (LRScheduler | None, optional): The scheduler to save the state from. Defaults to None.
         preprocessor: The preprocessor/pipeline to save. Defaults to None.
+        postprocessor: The postprocessor/pipeline to save. Defaults to None.
     """
     pretrained_dir = checkpoint_dir / PRETRAINED_MODEL_DIR
     policy.save_pretrained(pretrained_dir)

src/lerobot/configs/__init__.py

@@ -24,6 +24,7 @@ Import them directly: ``from lerobot.configs.train import TrainPipelineConfig``
 from .dataset import DatasetRecordConfig
 from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
+from .recipe import MessageTurn, TrainingRecipe, load_recipe
 from .types import (
     FeatureType,
     NormalizationMode,
@@ -31,6 +32,12 @@ from .types import (
     PolicyFeature,
     RTCAttentionSchedule,
 )
+from .video import (
+    VALID_VIDEO_CODECS,
+    VIDEO_ENCODER_INFO_KEYS,
+    VideoEncoderConfig,
+    camera_encoder_defaults,
+)
 
 __all__ = [
     # Types
@@ -43,7 +50,16 @@ __all__ = [
     "DatasetRecordConfig",
     "DatasetConfig",
     "EvalConfig",
+    "MessageTurn",
     "PeftConfig",
     "PreTrainedConfig",
+    "TrainingRecipe",
     "WandBConfig",
+    "load_recipe",
+    "VideoEncoderConfig",
+    # Defaults
+    "camera_encoder_defaults",
+    # Constants
+    "VALID_VIDEO_CODECS",
+    "VIDEO_ENCODER_INFO_KEYS",
 ]

src/lerobot/configs/dataset.py

@@ -14,10 +14,12 @@
 
 """Shared dataset recording configuration used by both ``lerobot-record`` and ``lerobot-rollout``."""
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from datetime import datetime
 from pathlib import Path
 
+from .video import VideoEncoderConfig, camera_encoder_defaults
+
 
 @dataclass
 class DatasetRecordConfig:
@@ -55,10 +57,9 @@ class DatasetRecordConfig:
     # Number of episodes to record before batch encoding videos
     # Set to 1 for immediate encoding (default behavior), or higher for batched encoding
     video_encoding_batch_size: int = 1
-    # Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1', 'auto',
-    # or hardware-specific: 'h264_videotoolbox', 'h264_nvenc', 'h264_vaapi', 'h264_qsv'.
-    # Use 'auto' to auto-detect the best available hardware encoder.
-    vcodec: str = "libsvtav1"
+    # Video encoder settings for camera MP4s (codec, quality, GOP, etc.). Tuned via CLI nested keys,
+    # e.g. ``--dataset.camera_encoder.vcodec=h264`` (see ``VideoEncoderConfig``).
+    camera_encoder: VideoEncoderConfig = field(default_factory=camera_encoder_defaults)
     # Enable streaming video encoding: encode frames in real-time during capture instead
     # of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding
     streaming_encoding: bool = False

src/lerobot/configs/default.py

@@ -17,7 +17,7 @@
 from dataclasses import dataclass, field
 
 from lerobot.transforms import ImageTransformsConfig
-from lerobot.utils.import_utils import get_safe_default_codec
+from lerobot.utils.import_utils import get_safe_default_video_backend
 
 
 @dataclass
@@ -34,7 +34,7 @@ class DatasetConfig:
     image_transforms: ImageTransformsConfig = field(default_factory=ImageTransformsConfig)
     revision: str | None = None
     use_imagenet_stats: bool = True
-    video_backend: str = field(default_factory=get_safe_default_codec)
+    video_backend: str = field(default_factory=get_safe_default_video_backend)
     # When True, video frames are returned as uint8 tensors (0-255) instead of float32 (0.0-1.0).
     # This reduces memory and speeds up DataLoader IPC. The training pipeline handles the conversion.
     return_uint8: bool = False
@@ -117,3 +117,9 @@ class PeftConfig:
     # the rank used for the adapter. In general a higher rank means more trainable parameters and closer to full
     # fine-tuning.
     r: int = 16
+
+    # Alpha parameter for LoRA scaling (scaling = lora_alpha / r).
+    # In general, a higher alpha means stronger adaptation signal.
+    # If None, the PEFT library defaults to alpha=8, which may dampen high-rank adapters.
+    # Common values are r (alpha == rank) or 2*r.
+    lora_alpha: int | None = None

src/lerobot/configs/eval.py

@@ -18,8 +18,8 @@ from logging import getLogger
 from pathlib import Path
 
 from lerobot import envs, policies  # noqa: F401
-from lerobot.configs import parser
 
+from . import parser
 from .default import EvalConfig
 from .policies import PreTrainedConfig
 
@@ -46,8 +46,11 @@ class EvalPipelineConfig:
         # HACK: We parse again the cli args here to get the pretrained path if there was one.
         policy_path = parser.get_path_arg("policy")
         if policy_path:
-            cli_overrides = parser.get_cli_overrides("policy")
-            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            yaml_overrides = parser.get_yaml_overrides("policy")
+            cli_overrides = parser.get_cli_overrides("policy") or []
+            self.policy = PreTrainedConfig.from_pretrained(
+                policy_path, cli_overrides=yaml_overrides + cli_overrides
+            )
             self.policy.pretrained_path = Path(policy_path)
 
         else:

src/lerobot/configs/parser.py

@@ -13,8 +13,10 @@
 # limitations under the License.
 import importlib
 import inspect
+import json
 import pkgutil
 import sys
+import tempfile
 from argparse import ArgumentError
 from collections.abc import Callable, Iterable, Sequence
 from functools import wraps
@@ -24,6 +26,7 @@ from types import ModuleType
 from typing import Any, TypeVar, cast
 
 import draccus
+import yaml  # type: ignore[import-untyped]
 
 from lerobot.utils.utils import has_method
 
@@ -32,6 +35,29 @@ F = TypeVar("F", bound=Callable[..., object])
 PATH_KEY = "path"
 PLUGIN_DISCOVERY_SUFFIX = "discover_packages_path"
 
+# Storage for path args extracted from YAML/JSON config files, so that
+# get_path_arg() can find them even when they weren't passed via CLI.
+_config_path_args: dict[str, str] = {}
+
+# Storage for non-path YAML overrides so validate() can pass them to from_pretrained.
+_config_yaml_overrides: dict[str, list[str]] = {}
+
+
+def _flatten_to_cli_args(d: dict, prefix: str = "") -> list[str]:
+    """Recursively flatten a nested dict to CLI-style args (e.g. {"lr": 1e-4} -> ["--lr=0.0001"])."""
+    args = []
+    for key, value in d.items():
+        if key in (PATH_KEY, draccus.CHOICE_TYPE_KEY):
+            continue
+        full_key = f"{prefix}.{key}" if prefix else key
+        if isinstance(value, bool):
+            value = str(value).lower()
+        if isinstance(value, dict):
+            args.extend(_flatten_to_cli_args(value, full_key))
+        elif value is not None and not isinstance(value, list):
+            args.append(f"--{full_key}={value}")
+    return args
+
 
 def get_cli_overrides(field_name: str, args: Sequence[str] | None = None) -> list[str] | None:
     """Parses arguments from cli at a given nested attribute level.
@@ -145,7 +171,14 @@ def load_plugin(plugin_path: str) -> None:
 
 
 def get_path_arg(field_name: str, args: Sequence[str] | None = None) -> str | None:
-    return parse_arg(f"{field_name}.{PATH_KEY}", args)
+    result = parse_arg(f"{field_name}.{PATH_KEY}", args)
+    if result is None:
+        result = _config_path_args.get(field_name)
+    return result
+
+
+def get_yaml_overrides(field_name: str) -> list[str]:
+    return _config_yaml_overrides.get(field_name, [])
 
 
 def get_type_arg(field_name: str, args: Sequence[str] | None = None) -> str | None:
@@ -192,6 +225,52 @@ def filter_path_args(fields_to_filter: str | list[str], args: Sequence[str] | No
     return filtered_args
 
 
+def extract_path_fields_from_config(config_path: str, path_fields: list[str]) -> str:
+    """Extract `path` fields from a YAML/JSON config before draccus processes it.
+
+    When a user specifies e.g. ``policy.path: lerobot/smolvla_base`` in a YAML config,
+    draccus will fail because ``path`` is not a valid field on policy config classes.
+    This function extracts those path values, stores them in ``_config_path_args`` for
+    later retrieval by ``get_path_arg()``, and returns a cleaned temp config file path.
+    """
+    config_file = Path(config_path)
+    suffix = config_file.suffix.lower()
+
+    if suffix in (".yaml", ".yml"):
+        with open(config_file) as f:
+            config_data = yaml.safe_load(f)
+    elif suffix == ".json":
+        with open(config_file) as f:
+            config_data = json.load(f)
+    else:
+        return config_path
+
+    if not isinstance(config_data, dict):
+        return config_path
+
+    modified = False
+    for field in path_fields:
+        if field in config_data and isinstance(config_data[field], dict) and PATH_KEY in config_data[field]:
+            _config_path_args[field] = str(config_data[field].pop(PATH_KEY))
+            remaining = config_data[field]
+            if remaining:
+                _config_yaml_overrides[field] = _flatten_to_cli_args(remaining)
+            else:
+                del config_data[field]
+            modified = True
+
+    if not modified:
+        return config_path
+
+    # Write cleaned config to a temp file
+    with tempfile.NamedTemporaryFile(mode="w", suffix=suffix, delete=False) as tmp:
+        if suffix in (".yaml", ".yml"):
+            yaml.dump(config_data, tmp, default_flow_style=False)
+        else:
+            json.dump(config_data, tmp, indent=2)
+    return tmp.name
+
+
 def wrap(config_path: Path | None = None) -> Callable[[F], F]:
     """
     HACK: Similar to draccus.wrap but does three additional things:
@@ -225,6 +304,9 @@ def wrap(config_path: Path | None = None) -> Callable[[F], F]:
                 if has_method(argtype, "__get_path_fields__"):
                     path_fields = argtype.__get_path_fields__()
                     cli_args = filter_path_args(path_fields, cli_args)
+                    # Also extract path fields from the YAML/JSON config file
+                    if config_path_cli:
+                        config_path_cli = extract_path_fields_from_config(config_path_cli, path_fields)
                 if has_method(argtype, "from_pretrained") and config_path_cli:
                     cli_args = filter_arg("config_path", cli_args)
                     cfg = argtype.from_pretrained(config_path_cli, cli_args=cli_args)

src/lerobot/configs/recipe.py

@@ -0,0 +1,206 @@
+#!/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 __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Literal, get_args
+
+MessageRole = Literal["user", "assistant", "system", "tool"]
+MessageStream = Literal["high_level", "low_level"]
+
+DEFAULT_BINDINGS = {
+    "subtask": "active_at(t, style=subtask)",
+    "memory": "active_at(t, style=memory)",
+    "plan": "active_at(t, style=plan)",
+    "speech": "emitted_at(t, role=assistant, tool_name=say)",
+    "interjection": "emitted_at(t, style=interjection)",
+    "vqa": "emitted_at(t, style=vqa, role=assistant)",
+    "vqa_query": "emitted_at(t, style=vqa, role=user)",
+}
+
+PLACEHOLDER_RE = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)\}")
+"""``${name}`` placeholder pattern used by both recipe binding-reference
+discovery (here) and rendered-message substitution (in ``language_render``)."""
+
+_VALID_ROLES = frozenset(get_args(MessageRole))
+_VALID_STREAMS = frozenset(get_args(MessageStream))
+
+
+@dataclass
+class MessageTurn:
+    """A single chat-style turn in a recipe template.
+
+    ``content`` may be a plain string, a list of HF-style multimodal blocks, or
+    ``None`` when ``tool_calls_from`` supplies tool-call payloads instead.
+    ``stream`` tags the turn for downstream filtering, ``target`` flags it as a
+    training target, and ``if_present`` skips the turn when the named binding
+    resolves to ``None``.
+    """
+
+    role: MessageRole
+    content: str | list[dict[str, Any]] | None = None
+    stream: MessageStream | None = None
+    target: bool = False
+    if_present: str | None = None
+    tool_calls_from: str | None = None
+
+    def __post_init__(self) -> None:
+        """Validate role, stream, and content after dataclass construction."""
+        if self.role not in _VALID_ROLES:
+            raise ValueError(f"Unsupported message role: {self.role!r}")
+        # ``stream`` is typed Optional only so the dataclass can keep its
+        # field ordering, but recipes must always tag every turn with a
+        # stream — the renderer's ``_validate_rendered`` would reject
+        # ``None`` later on. Fail at construction so the bad recipe is
+        # caught at YAML load time rather than at the first sample.
+        if self.stream is None:
+            raise ValueError(
+                f"MessageTurn(role={self.role!r}) is missing a stream — "
+                f"every turn must declare one of {sorted(_VALID_STREAMS)}."
+            )
+        if self.stream not in _VALID_STREAMS:
+            raise ValueError(f"Unsupported message stream: {self.stream!r}")
+        if self.content is None and self.tool_calls_from is None:
+            raise ValueError("MessageTurn.content is required unless tool_calls_from is set.")
+        if self.content is not None and not isinstance(self.content, (str, list)):
+            raise TypeError("MessageTurn.content must be a string, a list of HF-style blocks, or None.")
+        if isinstance(self.content, list):
+            for block in self.content:
+                if not isinstance(block, dict) or "type" not in block:
+                    raise ValueError(
+                        "Multimodal content blocks must be HF-style dictionaries with a type key."
+                    )
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> MessageTurn:
+        """Construct a :class:`MessageTurn` from a plain dictionary."""
+        return cls(**data)
+
+
+@dataclass
+class TrainingRecipe:
+    """A recipe describing how to render training samples from language rows.
+
+    A recipe is either a *message recipe* (``messages`` plus optional
+    ``bindings``) or a *blend recipe* (``blend`` mapping names to weighted
+    sub-recipes). ``weight`` is only meaningful inside a blend.
+    """
+
+    messages: list[MessageTurn] | None = None
+    bindings: dict[str, str] | None = None
+    blend: dict[str, TrainingRecipe] | None = None
+    weight: float | None = None
+
+    def __post_init__(self) -> None:
+        """Validate that exactly one of ``messages`` or ``blend`` is set."""
+        if self.messages is not None and self.blend is not None:
+            raise ValueError("TrainingRecipe must set only one of messages or blend.")
+        if self.messages is None and self.blend is None:
+            raise ValueError("TrainingRecipe must set one of messages or blend.")
+
+        if self.messages is not None:
+            self._validate_message_recipe()
+        if self.blend is not None:
+            self._validate_blend_recipe()
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> TrainingRecipe:
+        """Construct a :class:`TrainingRecipe` from a nested dictionary."""
+        data = dict(data)
+        if data.get("messages") is not None:
+            data["messages"] = [
+                turn if isinstance(turn, MessageTurn) else MessageTurn.from_dict(turn)
+                for turn in data["messages"]
+            ]
+        if data.get("blend") is not None:
+            data["blend"] = {
+                name: recipe if isinstance(recipe, TrainingRecipe) else cls.from_dict(recipe)
+                for name, recipe in data["blend"].items()
+            }
+        return cls(**data)
+
+    @classmethod
+    def from_yaml(cls, path: str | Path) -> TrainingRecipe:
+        """Load a :class:`TrainingRecipe` from a YAML file at ``path``."""
+        import yaml  # type: ignore[import-untyped]
+
+        with open(path) as f:
+            data = yaml.safe_load(f)
+        if not isinstance(data, dict):
+            raise ValueError(f"Recipe YAML must contain a mapping at the top level: {path}")
+        return cls.from_dict(data)
+
+    def _validate_message_recipe(self) -> None:
+        """Ensure every templated binding is known and at least one turn is a target."""
+        assert self.messages is not None
+        known_bindings = set(DEFAULT_BINDINGS) | set(self.bindings or {}) | {"task"}
+
+        for turn in self.messages:
+            missing = self._referenced_bindings(turn) - known_bindings
+            if missing:
+                raise ValueError(f"MessageTurn references unknown binding(s): {sorted(missing)}")
+
+        if not any(turn.target for turn in self.messages):
+            raise ValueError("Message recipes must contain at least one target turn.")
+
+    def _validate_blend_recipe(self) -> None:
+        """Ensure each blend component is a non-empty, weighted message recipe."""
+        assert self.blend is not None
+        if not self.blend:
+            raise ValueError("Blend recipes must contain at least one component.")
+
+        for name, recipe in self.blend.items():
+            if recipe.blend is not None:
+                raise ValueError(f"Blend component {name!r} cannot itself define a blend.")
+            if recipe.messages is None:
+                raise ValueError(f"Blend component {name!r} must define messages.")
+            if recipe.weight is None:
+                raise ValueError(f"Blend component {name!r} must define weight.")
+            if recipe.weight <= 0:
+                raise ValueError(f"Blend component {name!r} must have a positive weight.")
+
+    def _referenced_bindings(self, turn: MessageTurn) -> set[str]:
+        """Return the binding names that ``turn`` references via placeholders or attributes."""
+        names: set[str] = set()
+        if turn.if_present is not None:
+            names.add(turn.if_present)
+        if turn.tool_calls_from is not None:
+            names.add(turn.tool_calls_from)
+        names.update(_placeholders_in_content(turn.content))
+        return names
+
+
+def _placeholders_in_content(content: str | list[dict[str, Any]] | None) -> set[str]:
+    """Return the set of ``${name}`` placeholders found anywhere in ``content``."""
+    if content is None:
+        return set()
+    if isinstance(content, str):
+        return set(PLACEHOLDER_RE.findall(content))
+
+    names: set[str] = set()
+    for block in content:
+        for value in block.values():
+            if isinstance(value, str):
+                names.update(PLACEHOLDER_RE.findall(value))
+    return names
+
+
+def load_recipe(path: str | Path) -> TrainingRecipe:
+    """Load a :class:`TrainingRecipe` from a YAML file at ``path``."""
+    return TrainingRecipe.from_yaml(path)

src/lerobot/configs/rewards.py

@@ -27,12 +27,13 @@ from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError
 
-from lerobot.configs.types import PolicyFeature
 from lerobot.optim.optimizers import OptimizerConfig
 from lerobot.optim.schedulers import LRSchedulerConfig
 from lerobot.utils.device_utils import auto_select_torch_device, is_torch_device_available
 from lerobot.utils.hub import HubMixin
 
+from .types import PolicyFeature
+
 T = TypeVar("T", bound="RewardModelConfig")
 logger = logging.getLogger(__name__)
 
@@ -89,9 +90,9 @@ class RewardModelConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
     def reward_delta_indices(self) -> list | None:  # type: ignore[type-arg]
         return None
 
-    @abc.abstractmethod
-    def get_optimizer_preset(self) -> OptimizerConfig:
-        raise NotImplementedError
+    def get_optimizer_preset(self) -> OptimizerConfig | None:
+        """Default optimizer for this reward model, or ``None`` for zero-shot models."""
+        return None
 
     def get_scheduler_preset(self) -> LRSchedulerConfig | None:
         return None

src/lerobot/configs/train.py

@@ -25,11 +25,11 @@ from huggingface_hub import hf_hub_download
 from huggingface_hub.errors import HfHubHTTPError
 
 from lerobot import envs
-from lerobot.configs import parser
 from lerobot.optim import LRSchedulerConfig, OptimizerConfig
 from lerobot.utils.hub import HubMixin
 from lerobot.utils.sample_weighting import SampleWeightingConfig
 
+from . import parser
 from .default import DatasetConfig, EvalConfig, PeftConfig, WandBConfig
 from .policies import PreTrainedConfig
 from .rewards import RewardModelConfig
@@ -144,8 +144,11 @@ class TrainPipelineConfig(HubMixin):
             )
             self.reward_model.pretrained_path = str(Path(reward_model_path))
         elif policy_path:
-            cli_overrides = parser.get_cli_overrides("policy")
-            self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
+            yaml_overrides = parser.get_yaml_overrides("policy")
+            cli_overrides = parser.get_cli_overrides("policy") or []
+            self.policy = PreTrainedConfig.from_pretrained(
+                policy_path, cli_overrides=yaml_overrides + cli_overrides
+            )
             self.policy.pretrained_path = Path(policy_path)
         elif self.resume:
             config_path = parser.parse_arg("config_path")
@@ -269,10 +272,3 @@ class TrainPipelineConfig(HubMixin):
 
         with draccus.config_type("json"):
             return draccus.parse(cls, config_file, args=cli_args)
-
-
-@dataclass(kw_only=True)
-class TrainRLServerPipelineConfig(TrainPipelineConfig):
-    # NOTE: In RL, we don't need an offline dataset
-    # TODO: Make `TrainPipelineConfig.dataset` optional
-    dataset: DatasetConfig | None = None  # type: ignore[assignment] # because the parent class has made it's type non-optional

src/lerobot/configs/video.py

@@ -0,0 +1,235 @@
+# 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.
+# Note: We subclass str so that serialization is straightforward
+# https://stackoverflow.com/questions/24481852/serialising-an-enum-member-to-json
+
+"""Video encoder configurations."""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from typing import Any
+
+from lerobot.utils.import_utils import require_package
+
+logger = logging.getLogger(__name__)
+
+# List of hardware encoders to probe for auto-selection. Availability depends on the platform and the chosen video backend.
+# Determines the order of preference for auto-selection when vcodec="auto" is used.
+HW_VIDEO_CODECS = [
+    "h264_videotoolbox",  # macOS
+    "hevc_videotoolbox",  # macOS
+    "h264_nvenc",  # NVIDIA GPU
+    "hevc_nvenc",  # NVIDIA GPU
+    "h264_vaapi",  # Linux Intel/AMD
+    "h264_qsv",  # Intel Quick Sync
+]
+VALID_VIDEO_CODECS: frozenset[str] = frozenset({"h264", "hevc", "libsvtav1", "auto", *HW_VIDEO_CODECS})
+# Aliases for legacy video codec names.
+VIDEO_CODECS_ALIASES: dict[str, str] = {"av1": "libsvtav1"}
+
+
+LIBSVTAV1_DEFAULT_PRESET: int = 12
+
+# Keys persisted under ``features[*]["info"]`` as ``video.<name>`` (from :class:`VideoEncoderConfig`).
+# ``vcodec``` and ``pix_fmt`` are derived from the video stream directly.
+VIDEO_ENCODER_INFO_FIELD_NAMES: frozenset[str] = frozenset(
+    {"g", "crf", "preset", "fast_decode", "extra_options", "video_backend"}
+)
+VIDEO_ENCODER_INFO_KEYS: frozenset[str] = frozenset(
+    f"video.{name}" for name in VIDEO_ENCODER_INFO_FIELD_NAMES
+)
+
+
+@dataclass
+class VideoEncoderConfig:
+    """Video encoder configuration.
+
+    Attributes:
+        vcodec: Video encoder name. ``"auto"`` is resolved during
+            construction (HW encoder if available, else ``libsvtav1``).
+        pix_fmt: Pixel format (e.g. ``"yuv420p"``).
+        g: GOP size (keyframe interval).
+        crf: Quality level — mapped to the native quality parameter of the
+            codec (``crf`` for software, ``qp`` for NVENC/VAAPI,
+            ``q:v`` for VideoToolbox, ``global_quality`` for QSV).
+        preset: Speed/quality preset. Accepted type is per-codec.
+        fast_decode: Fast-decode tuning. For ``libsvtav1`` this is a level (0-2)
+            embedded in ``svtav1-params``. For ``h264`` and ``hevc`` non-zero values
+            set ``tune=fastdecode``. Ignored for other codecs.
+        video_backend: Python to be used for encoding. Only ``"pyav"``
+            is currently supported.
+        extra_options: Free-form dictionary of additional video encoder options
+            (e.g. ``{"tune": "film", "profile:v": "high", "bf": 2}``).
+    """
+
+    vcodec: str = "libsvtav1"  # TODO(CarolinePascal): rename to codec ?
+    pix_fmt: str = "yuv420p"
+    g: int | None = 2
+    crf: int | float | None = 30
+    preset: int | str | None = None
+    fast_decode: int = 0
+    # TODO(CarolinePascal): add torchcodec support + find a way to unify the
+    # two backends (encoding and decoding).
+    video_backend: str = "pyav"
+    extra_options: dict[str, Any] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        self.resolve_vcodec()
+        # Empty-constructor ergonomics: ``VideoEncoderConfig()`` must "just work".
+        if self.preset is None and self.vcodec == "libsvtav1":
+            self.preset = LIBSVTAV1_DEFAULT_PRESET
+        self.validate()
+
+    @classmethod
+    def from_video_info(cls, video_info: dict | None) -> VideoEncoderConfig:
+        """Reconstruct a :class:`VideoEncoderConfig` from a video feature's ``info`` block.
+        Missing or ``None`` values fall back to the class defaults.
+        """
+        video_info = video_info or {}
+        kwargs: dict[str, Any] = {}
+
+        for src_key, dst_field in (("video.codec", "vcodec"), ("video.pix_fmt", "pix_fmt")):
+            value = video_info.get(src_key)
+            if value is not None:
+                kwargs[dst_field] = value
+
+        for field_name in VIDEO_ENCODER_INFO_FIELD_NAMES:
+            value = video_info.get(f"video.{field_name}")
+            if value is None:
+                continue
+            # Persisted as ``{}`` after merges with disagreeing sources — treat as default.
+            if field_name == "extra_options" and not value:
+                continue
+            kwargs[field_name] = value
+
+        return cls(**kwargs)
+
+    def detect_available_encoders(self, encoders: list[str] | str) -> list[str]:
+        """Return the subset of available encoders based on the specified video backend.
+
+        Args:
+            encoders: List of encoder names to detect. If a string, it is converted to a list.
+        Returns:
+            List of available encoder names. If the video backend is not "pyav", returns an empty list.
+        """
+        if self.video_backend == "pyav":
+            require_package("av", extra="dataset")
+            from lerobot.datasets import detect_available_encoders_pyav
+
+            return detect_available_encoders_pyav(encoders)
+        return []
+
+    def validate(self) -> None:
+        """Validate the video encoder configuration."""
+        if self.video_backend == "pyav":
+            require_package("av", extra="dataset")
+            from lerobot.datasets import check_video_encoder_parameters_pyav
+
+            check_video_encoder_parameters_pyav(self.vcodec, self.pix_fmt, self.get_codec_options())
+
+    def resolve_vcodec(self) -> None:
+        """Check ``vcodec`` and, when it is ``"auto"``, pick a concrete encoder.
+
+        For ``"auto"``, the first hardware encoder in the preference list that is available is chosen; if none are available, ``libsvtav1`` is used. If the
+        resolved codec (explicit or after auto-selection) is not available, raises ``ValueError``.
+
+        Stream-derived canonical codec names listed in :data:`VIDEO_CODECS_ALIASES` are
+        rewritten to their corresponding encoder name (e.g. ``"av1"`` → ``"libsvtav1"``).
+        """
+        self.vcodec = VIDEO_CODECS_ALIASES.get(self.vcodec, self.vcodec)
+        if self.vcodec not in VALID_VIDEO_CODECS:
+            raise ValueError(f"Invalid vcodec '{self.vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
+        if self.vcodec == "auto":
+            available = self.detect_available_encoders(HW_VIDEO_CODECS)
+            for encoder in HW_VIDEO_CODECS:
+                if encoder in available:
+                    logger.info(f"Auto-selected video codec: {encoder}")
+                    self.vcodec = encoder
+                    return
+            logger.warning("No hardware encoder available, falling back to software encoder 'libsvtav1'")
+            self.vcodec = "libsvtav1"
+
+        if self.detect_available_encoders(self.vcodec):
+            logger.info(f"Using video codec: {self.vcodec}")
+            return
+        raise ValueError(f"Unsupported video codec: {self.vcodec} with video backend {self.video_backend}")
+
+    def get_codec_options(
+        self, encoder_threads: int | None = None, as_strings: bool = False
+    ) -> dict[str, Any]:
+        """Translate the tuning fields to codec-specific options.
+
+        ``VideoEncoderConfig.extra_options`` are merged last but never override a structured field.
+
+        Args:
+            encoder_threads: Number of encoder threads set globally for all VideoEncoderConfigs.
+                For libsvtav1, this is mapped to ``lp`` via ``svtav1-params``.
+                For h264/hevc, this is mapped to ``threads``.
+                Hardware encoders ignore this parameter.
+            as_strings: If ``True``, casts values to strings.
+        """
+        opts: dict[str, Any] = {}
+
+        def set_if(key: str, value: Any) -> None:
+            if value is not None:
+                opts[key] = value if not as_strings else str(value)
+
+        # GOP size is not a codec-specific option, so it is always set.
+        set_if("g", self.g)
+
+        if self.vcodec == "libsvtav1":
+            set_if("crf", self.crf)
+            set_if("preset", self.preset)
+            svtav1_parts: list[str] = []
+            if self.fast_decode is not None:
+                svtav1_parts.append(f"fast-decode={max(0, min(2, self.fast_decode))}")
+            if encoder_threads is not None:
+                svtav1_parts.append(f"lp={encoder_threads}")
+            if svtav1_parts:
+                opts["svtav1-params"] = ":".join(svtav1_parts)
+        elif self.vcodec in ("h264", "hevc"):
+            set_if("crf", self.crf)
+            set_if("preset", self.preset)
+            if self.fast_decode:
+                opts["tune"] = "fastdecode"
+            set_if("threads", encoder_threads)
+        elif self.vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
+            if self.crf is not None:
+                opts["q:v"] = max(1, min(100, 100 - self.crf * 2))
+        elif self.vcodec in ("h264_nvenc", "hevc_nvenc"):
+            opts["rc"] = 0
+            set_if("qp", self.crf)
+            set_if("preset", self.preset)
+        elif self.vcodec == "h264_vaapi":
+            set_if("qp", self.crf)
+        elif self.vcodec == "h264_qsv":
+            set_if("global_quality", self.crf)
+            set_if("preset", self.preset)
+        else:
+            set_if("crf", self.crf)
+            set_if("preset", self.preset)
+
+        # Extra options are merged last but never override structured fields (values are kept as given).
+        for k, v in self.extra_options.items():
+            if k not in opts:
+                set_if(k, v)
+
+        return opts
+
+
+def camera_encoder_defaults() -> VideoEncoderConfig:
+    """Return a :class:`VideoEncoderConfig` with RGB-camera defaults."""
+    return VideoEncoderConfig()

src/lerobot/datasets/__init__.py

@@ -31,15 +31,25 @@ from .dataset_tools import (
     modify_features,
     modify_tasks,
     recompute_stats,
+    reencode_dataset,
     remove_feature,
     split_dataset,
 )
 from .factory import make_dataset, resolve_delta_timestamps
 from .image_writer import safe_stop_image_writer
 from .io_utils import load_episodes, write_stats
+from .language import (
+    EVENT_ONLY_STYLES,
+    LANGUAGE_EVENTS,
+    LANGUAGE_PERSISTENT,
+    PERSISTENT_STYLES,
+    STYLE_REGISTRY,
+    column_for_style,
+)
 from .lerobot_dataset import LeRobotDataset
 from .multi_dataset import MultiLeRobotDataset
 from .pipeline_features import aggregate_pipeline_dataset_features, create_initial_features
+from .pyav_utils import check_video_encoder_parameters_pyav, detect_available_encoders_pyav
 from .sampler import EpisodeAwareSampler
 from .streaming_dataset import StreamingLeRobotDataset
 from .utils import DEFAULT_EPISODES_PATH, create_lerobot_dataset_card
@@ -53,12 +63,19 @@ __all__ = [
     "CODEBASE_VERSION",
     "DEFAULT_EPISODES_PATH",
     "DEFAULT_QUANTILES",
+    "EVENT_ONLY_STYLES",
     "EpisodeAwareSampler",
+    "LANGUAGE_EVENTS",
+    "LANGUAGE_PERSISTENT",
     "LeRobotDataset",
     "LeRobotDatasetMetadata",
     "MultiLeRobotDataset",
+    "PERSISTENT_STYLES",
+    "STYLE_REGISTRY",
     "StreamingLeRobotDataset",
     "VideoEncodingManager",
+    "check_video_encoder_parameters_pyav",
+    "detect_available_encoders_pyav",
     "add_features",
     "aggregate_datasets",
     "aggregate_pipeline_dataset_features",
@@ -66,6 +83,7 @@ __all__ = [
     "convert_image_to_video_dataset",
     "create_initial_features",
     "create_lerobot_dataset_card",
+    "column_for_style",
     "delete_episodes",
     "get_feature_stats",
     "load_episodes",
@@ -74,6 +92,7 @@ __all__ = [
     "modify_features",
     "modify_tasks",
     "recompute_stats",
+    "reencode_dataset",
     "remove_feature",
     "resolve_delta_timestamps",
     "safe_stop_image_writer",

src/lerobot/datasets/aggregate.py

@@ -15,6 +15,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import copy
 import logging
 import shutil
 from pathlib import Path
@@ -23,9 +24,11 @@ import datasets
 import pandas as pd
 import tqdm
 
+from lerobot.configs import VIDEO_ENCODER_INFO_KEYS
+
 from .compute_stats import aggregate_stats
 from .dataset_metadata import LeRobotDatasetMetadata
-from .feature_utils import get_hf_features_from_features
+from .feature_utils import features_equal_for_merge, get_hf_features_from_features
 from .io_utils import (
     get_file_size_in_mb,
     get_parquet_file_size_in_mb,
@@ -46,11 +49,54 @@ from .utils import (
 from .video_utils import concatenate_video_files, get_video_duration_in_s
 
 
+def merge_video_feature_info_for_aggregate(all_metadata: list[LeRobotDatasetMetadata]) -> dict[str, dict]:
+    """Create a merged video feature info dictionary for aggregation. The video encoder info is merged field-by-field: each key is kept only when every source agrees; otherwise that key is set to ``null`` (or ``{}`` for ``video.extra_options``) and a warning is logged.
+
+    Args:
+        all_metadata: List of LeRobotDatasetMetadata objects to merge.
+
+    Returns:
+        dict: A dictionary of merged video feature info.
+    """
+    merged_info = copy.deepcopy(all_metadata[0].features)
+    video_keys = [k for k in merged_info if merged_info[k].get("dtype") == "video"]
+
+    for vk in video_keys:
+        video_infos = [m.features.get(vk, {}).get("info") or {} for m in all_metadata]
+        base_video_info = video_infos[0]
+
+        merged_encoder_info: dict = {}
+        fallback_keys: list[str] = []
+        for info_key in VIDEO_ENCODER_INFO_KEYS:
+            values = [info.get(info_key, None) for info in video_infos]
+            first_value = values[0]
+            all_match = all(v == first_value for v in values[1:])
+
+            if all_match:
+                merged_encoder_info[info_key] = first_value
+            else:
+                fallback_keys.append(info_key)
+                merged_encoder_info[info_key] = {} if info_key == "video.extra_options" else None
+
+        if fallback_keys:
+            logging.warning(
+                f"Merging heterogeneous or incomplete video encoder metadata for feature {vk}. "
+                f"Setting these keys to null: {fallback_keys}.",
+            )
+
+        merged_info[vk]["info"] = {**base_video_info, **merged_encoder_info}
+        # TODO(CarolinePascal): make this variable once we have support for other video backends.
+        merged_info[vk]["info"]["video.video_backend"] = "pyav"
+
+    return merged_info
+
+
 def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
     """Validates that all dataset metadata have consistent properties.
 
     Ensures all datasets have the same fps, robot_type, and features to guarantee
     compatibility when aggregating them into a single dataset.
+    Video encoder info is not considered for validation but is merged during aggregation in ``merge_video_feature_info_for_aggregate``.
 
     Args:
         all_metadata: List of LeRobotDatasetMetadata objects to validate.
@@ -74,7 +120,7 @@ def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
             raise ValueError(
                 f"Same robot_type is expected, but got robot_type={meta.robot_type} instead of {robot_type}."
             )
-        if features != meta.features:
+        if not features_equal_for_merge(features, meta.features):
             raise ValueError(
                 f"Same features is expected, but got features={meta.features} instead of {features}."
             )
@@ -274,7 +320,8 @@ def aggregate_datasets(
             LeRobotDatasetMetadata(repo_id, root=root) for repo_id, root in zip(repo_ids, roots, strict=False)
         ]
     )
-    fps, robot_type, features = validate_all_metadata(all_metadata)
+    fps, robot_type, _ = validate_all_metadata(all_metadata)
+    features = merge_video_feature_info_for_aggregate(all_metadata)
     video_keys = [key for key in features if features[key]["dtype"] == "video"]
 
     dst_meta = LeRobotDatasetMetadata.create(
@@ -332,7 +379,6 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
         videos_idx: Dictionary tracking video chunk and file indices.
         video_files_size_in_mb: Maximum size for video files in MB (defaults to DEFAULT_VIDEO_FILE_SIZE_IN_MB)
         chunk_size: Maximum number of files per chunk (defaults to DEFAULT_CHUNK_SIZE)
-
     Returns:
         dict: Updated videos_idx with current chunk and file indices.
     """
@@ -414,9 +460,11 @@ def aggregate_videos(src_meta, dst_meta, videos_idx, video_files_size_in_mb, chu
                 current_dst_duration = dst_file_durations.get(dst_key, 0)
                 videos_idx[key]["src_to_offset"][(src_chunk_idx, src_file_idx)] = current_dst_duration
                 videos_idx[key]["src_to_dst"][(src_chunk_idx, src_file_idx)] = dst_key
+                # TODO(CarolinePascal): Move the check before the loop to avoid failing in the middle + add possibility to re-encode the video if the check fails
                 concatenate_video_files(
                     [dst_path, src_path],
                     dst_path,
+                    compatibility_check=True,
                 )
                 # Update duration of this destination file
                 dst_file_durations[dst_key] = current_dst_duration + src_duration

src/lerobot/datasets/compute_stats.py

@@ -512,7 +512,7 @@ def compute_episode_stats(
 
     ep_stats = {}
     for key, data in episode_data.items():
-        if features[key]["dtype"] == "string":
+        if features[key]["dtype"] in {"string", "language"}:
             continue
 
         if features[key]["dtype"] in ["image", "video"]:

src/lerobot/datasets/dataset_metadata.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import contextlib
+from collections.abc import Callable
 from pathlib import Path
 
 import numpy as np
@@ -23,6 +24,7 @@ import pyarrow as pa
 import pyarrow.parquet as pq
 from huggingface_hub import snapshot_download
 
+from lerobot.configs import VideoEncoderConfig
 from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
 from lerobot.utils.feature_utils import _validate_feature_names
 from lerobot.utils.utils import flatten_dict
@@ -34,12 +36,12 @@ from .io_utils import (
     load_episodes,
     load_info,
     load_stats,
-    load_subtasks,
     load_tasks,
     write_info,
     write_stats,
     write_tasks,
 )
+from .language import DEFAULT_TOOLS, LANGUAGE_COLUMNS
 from .utils import (
     DEFAULT_EPISODES_PATH,
     check_version_compatibility,
@@ -175,7 +177,6 @@ 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)
 
@@ -189,6 +190,29 @@ class LeRobotDatasetMetadata:
         if self.episodes is None:
             self._load_metadata()
 
+    def filter_episodes(
+        self,
+        predicate: Callable[[dict], bool],
+        candidates: list[int] | None = None,
+    ) -> list[int]:
+        """Filter episodes whose metadata satisfies a given predicate.
+
+        Args:
+            predicate: Predicate over per-episode metadata rows used to select episodes.
+            candidates: Optional list of episode indices to restrict evaluation to.
+
+        Returns:
+            List of sorted episode indices that satisfy the predicate.
+        """
+        self.ensure_readable()
+        if candidates is not None:
+            candidate_set = set(candidates)
+            combined = lambda ep: ep["episode_index"] in candidate_set and predicate(ep)  # noqa: E731
+        else:
+            combined = predicate
+        filtered = self.episodes.filter(combined, keep_in_memory=True, load_from_cache_file=False)
+        return sorted(int(idx) for idx in filtered["episode_index"])
+
     def _pull_from_repo(
         self,
         allow_patterns: list[str] | str | None = None,
@@ -318,6 +342,49 @@ class LeRobotDatasetMetadata:
         """Keys to access visual modalities (regardless of their storage method)."""
         return [key for key, ft in self.features.items() if ft["dtype"] in ["video", "image"]]
 
+    @property
+    def has_language_columns(self) -> bool:
+        """Return ``True`` if the dataset declares any language column.
+
+        Used to gate language-aware code paths (collate, render step) so
+        unannotated datasets keep PyTorch's default collate behavior.
+        """
+        return any(col in self.features for col in LANGUAGE_COLUMNS)
+
+    @property
+    def tools(self) -> list[dict]:
+        """OpenAI-style tool schemas declared by this dataset.
+
+        Read from ``meta/info.json["tools"]``. Returns a copy, so callers
+        can mutate the result safely. Falls back to
+        :data:`lerobot.datasets.language.DEFAULT_TOOLS` (the canonical
+        ``say`` schema) when the dataset doesn't declare any — that way
+        unannotated datasets and chat-template consumers
+        (``apply_chat_template(messages, tools=meta.tools)``) keep
+        working out of the box.
+
+        Implementations live under :mod:`lerobot.tools` (one file per
+        tool); see ``docs/source/tools.mdx`` for the authoring guide.
+        """
+        declared = self.info.tools
+        if declared:
+            return [dict(t) for t in declared]
+        return [dict(t) for t in DEFAULT_TOOLS]
+
+    @tools.setter
+    def tools(self, value: list[dict] | None) -> None:
+        """Persist a tool catalog to ``meta/info.json`` and reload metadata.
+
+        Writes ``value`` into the on-disk ``info.json`` (or clears the
+        ``tools`` key when ``value`` is ``None`` or empty), then reloads
+        ``self.info`` so the in-memory metadata matches what's on disk.
+        Saves callers from hand-editing ``info.json`` and re-instantiating
+        the metadata object.
+        """
+        self.info.tools = [dict(t) for t in value] if value else None
+        write_info(self.info, self.root)
+        self.info = load_info(self.root)
+
     @property
     def names(self) -> dict[str, list | dict]:
         """Names of the various dimensions of vector modalities."""
@@ -510,10 +577,23 @@ class LeRobotDatasetMetadata:
         self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
         write_stats(self.stats, self.root)
 
-    def update_video_info(self, video_key: str | None = None) -> None:
-        """
+    def update_video_info(
+        self,
+        video_key: str | None = None,
+        camera_encoder: VideoEncoderConfig | None = None,
+    ) -> None:
+        """Populate per-feature video info in ``info.json``.
+
         Warning: this function writes info from first episode videos, implicitly assuming that all videos have
         been encoded the same way. Also, this means it assumes the first episode exists.
+
+        Args:
+            video_key: If provided, only update this video key. Otherwise update
+                all video keys in the dataset.
+            camera_encoder: Encoder configuration used to produce the
+                videos. When provided, its fields are recorded as
+                ``video.<field>`` entries alongside the stream-derived
+                ``video.*`` entries (see :func:`get_video_info`).
         """
         if video_key is not None and video_key not in self.video_keys:
             raise ValueError(f"Video key {video_key} not found in dataset")
@@ -522,7 +602,7 @@ class LeRobotDatasetMetadata:
         for key in video_keys:
             if not self.features[key].get("info", None):
                 video_path = self.root / self.video_path.format(video_key=key, chunk_index=0, file_index=0)
-                self.info.features[key]["info"] = get_video_info(video_path)
+                self.info.features[key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder)
 
     def update_chunk_settings(
         self,
@@ -633,7 +713,6 @@ class LeRobotDatasetMetadata:
         _validate_feature_names(features)
 
         obj.tasks = None
-        obj.subtasks = None
         obj.episodes = None
         obj.stats = None
         obj.info = create_empty_dataset_info(

src/lerobot/datasets/dataset_reader.py

@@ -295,9 +295,4 @@ class DatasetReader:
         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._meta.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

src/lerobot/datasets/dataset_tools.py

@@ -26,7 +26,7 @@ This module provides utilities for:
 import logging
 import shutil
 from collections.abc import Callable
-from concurrent.futures import ThreadPoolExecutor, as_completed
+from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed
 from pathlib import Path
 
 import datasets
@@ -36,6 +36,7 @@ import pyarrow.parquet as pq
 import torch
 from tqdm import tqdm
 
+from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults
 from lerobot.utils.constants import ACTION, HF_LEROBOT_HOME, OBS_IMAGE, OBS_STATE
 from lerobot.utils.utils import flatten_dict
 
@@ -60,9 +61,14 @@ from .utils import (
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_DATA_PATH,
     DEFAULT_EPISODES_PATH,
+    VIDEO_DIR,
     update_chunk_file_indices,
 )
-from .video_utils import encode_video_frames, get_video_info
+from .video_utils import (
+    encode_video_frames,
+    get_video_info,
+    reencode_video,
+)
 
 
 def _load_episode_with_stats(src_dataset: LeRobotDataset, episode_idx: int) -> dict:
@@ -95,6 +101,11 @@ def delete_episodes(
 ) -> LeRobotDataset:
     """Delete episodes from a LeRobotDataset and create a new dataset.
 
+    Video segments that need re-encoding (because the source file mixes kept and
+    deleted episodes) are re-encoded with the source dataset's existing encoder
+    settings — read back from ``meta/info.json`` — so the output dataset stays
+    consistent with its own metadata.
+
     Args:
         dataset: The source LeRobotDataset.
         episode_indices: List of episode indices to delete.
@@ -157,6 +168,11 @@ def split_dataset(
 ) -> dict[str, LeRobotDataset]:
     """Split a LeRobotDataset into multiple smaller datasets.
 
+    Video segments that need re-encoding (because the source file mixes episodes
+    that fall into different splits) are re-encoded with the source dataset's
+    existing encoder settings — read back from ``meta/info.json`` — so each
+    output split stays consistent with its own metadata.
+
     Args:
         dataset: The source LeRobotDataset to split.
         splits: Either a dict mapping split names to episode indices, or a dict mapping
@@ -578,8 +594,7 @@ def _keep_episodes_from_video_with_av(
     output_path: Path,
     episodes_to_keep: list[tuple[int, int]],
     fps: float,
-    vcodec: str = "libsvtav1",
-    pix_fmt: str = "yuv420p",
+    camera_encoder: VideoEncoderConfig,
 ) -> None:
     """Keep only specified episodes from a video file using PyAV.
 
@@ -593,8 +608,7 @@ def _keep_episodes_from_video_with_av(
             Ranges are half-open intervals: [start_frame, end_frame), where start_frame
             is inclusive and end_frame is exclusive.
         fps: Frame rate of the video.
-        vcodec: Video codec to use for encoding.
-        pix_fmt: Pixel format for output video.
+        camera_encoder: Video encoder settings used to re-encode the kept frames.
     """
     from fractions import Fraction
 
@@ -619,12 +633,13 @@ def _keep_episodes_from_video_with_av(
 
     # Convert fps to Fraction for PyAV compatibility.
     fps_fraction = Fraction(fps).limit_denominator(1000)
-    v_out = out.add_stream(vcodec, rate=fps_fraction)
+    codec_options = camera_encoder.get_codec_options(as_strings=True)
+    v_out = out.add_stream(camera_encoder.vcodec, rate=fps_fraction, options=codec_options)
 
     # PyAV type stubs don't distinguish video streams from audio/subtitle streams.
     v_out.width = v_in.codec_context.width
     v_out.height = v_in.codec_context.height
-    v_out.pix_fmt = pix_fmt
+    v_out.pix_fmt = camera_encoder.pix_fmt
 
     # Set time_base to match the frame rate for proper timestamp handling.
     v_out.time_base = Fraction(1, int(fps))
@@ -687,14 +702,14 @@ def _copy_and_reindex_videos(
     src_dataset: LeRobotDataset,
     dst_meta: LeRobotDatasetMetadata,
     episode_mapping: dict[int, int],
-    vcodec: str = "libsvtav1",
-    pix_fmt: str = "yuv420p",
 ) -> dict[int, dict]:
     """Copy and filter video files, only re-encoding files with deleted episodes.
 
     For video files that only contain kept episodes, we copy them directly.
     For files with mixed kept/deleted episodes, we use PyAV filters to efficiently
-    re-encode only the desired segments.
+    re-encode only the desired segments. The encoder used for re-encoding is
+    derived per video key from the source dataset's ``meta/info.json`` so the
+    destination metadata keeps describing the videos accurately.
 
     Args:
         src_dataset: Source dataset to copy from
@@ -711,6 +726,9 @@ def _copy_and_reindex_videos(
 
     for video_key in src_dataset.meta.video_keys:
         logging.info(f"Processing videos for {video_key}")
+        camera_encoder = VideoEncoderConfig.from_video_info(
+            src_dataset.meta.info.features.get(video_key, {}).get("info")
+        )
 
         if dst_meta.video_path is None:
             raise ValueError("Destination metadata has no video_path defined")
@@ -792,8 +810,7 @@ def _copy_and_reindex_videos(
                     dst_video_path,
                     episodes_to_keep_ranges,
                     src_dataset.meta.fps,
-                    vcodec,
-                    pix_fmt,
+                    camera_encoder,
                 )
 
                 cumulative_ts = 0.0
@@ -1264,11 +1281,7 @@ def _estimate_frame_size_via_calibration(
     episode_indices: list[int],
     temp_dir: Path,
     fps: int,
-    vcodec: str,
-    pix_fmt: str,
-    g: int,
-    crf: int,
-    fast_decode: int,
+    camera_encoder: VideoEncoderConfig,
     num_calibration_frames: int = 30,
 ) -> float:
     """Estimate MB per frame by encoding a small calibration sample.
@@ -1282,11 +1295,7 @@ def _estimate_frame_size_via_calibration(
         episode_indices: List of episode indices being processed.
         temp_dir: Temporary directory for calibration files.
         fps: Frames per second for video encoding.
-        vcodec: Video codec (libsvtav1, h264, hevc).
-        pix_fmt: Pixel format (yuv420p, etc.).
-        g: GOP size (group of pictures).
-        crf: Constant Rate Factor (quality).
-        fast_decode: Fast decode tuning parameter.
+        camera_encoder: Video encoder settings used for calibration encoding.
         num_calibration_frames: Number of frames to use for calibration (default: 30).
 
     Returns:
@@ -1322,11 +1331,7 @@ def _estimate_frame_size_via_calibration(
             imgs_dir=calibration_dir,
             video_path=calibration_video_path,
             fps=fps,
-            vcodec=vcodec,
-            pix_fmt=pix_fmt,
-            g=g,
-            crf=crf,
-            fast_decode=fast_decode,
+            camera_encoder=camera_encoder,
             overwrite=True,
         )
 
@@ -1644,11 +1649,7 @@ def convert_image_to_video_dataset(
     dataset: LeRobotDataset,
     output_dir: Path | None = None,
     repo_id: str | None = None,
-    vcodec: str = "libsvtav1",
-    pix_fmt: str = "yuv420p",
-    g: int = 2,
-    crf: int = 30,
-    fast_decode: int = 0,
+    camera_encoder: VideoEncoderConfig | None = None,
     episode_indices: list[int] | None = None,
     num_workers: int = 4,
     max_episodes_per_batch: int | None = None,
@@ -1663,11 +1664,8 @@ def convert_image_to_video_dataset(
         dataset: The source LeRobot dataset with images
         output_dir: Root directory where the edited dataset will be stored. If not specified, defaults to $HF_LEROBOT_HOME/repo_id. Equivalent to new_root in EditDatasetConfig.
         repo_id: Edited dataset identifier. Equivalent to new_repo_id in EditDatasetConfig.
-        vcodec: Video codec (default: libsvtav1)
-        pix_fmt: Pixel format (default: yuv420p)
-        g: Group of pictures size (default: 2)
-        crf: Constant rate factor (default: 30)
-        fast_decode: Fast decode tuning (default: 0)
+        camera_encoder: Video encoder settings
+            (``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`).
         episode_indices: List of episode indices to convert (None = all episodes)
         num_workers: Number of threads for parallel processing (default: 4)
         max_episodes_per_batch: Maximum episodes per video batch to avoid memory issues (None = no limit)
@@ -1676,6 +1674,9 @@ def convert_image_to_video_dataset(
     Returns:
         New LeRobotDataset with images encoded as videos
     """
+    if camera_encoder is None:
+        camera_encoder = camera_encoder_defaults()
+
     # Check that it's an image dataset
     if len(dataset.meta.video_keys) > 0:
         raise ValueError(
@@ -1699,7 +1700,10 @@ def convert_image_to_video_dataset(
     logging.info(
         f"Converting {len(episode_indices)} episodes with {len(img_keys)} cameras from {dataset.repo_id}"
     )
-    logging.info(f"Video codec: {vcodec}, pixel format: {pix_fmt}, GOP: {g}, CRF: {crf}")
+    logging.info(
+        f"Video codec: {camera_encoder.vcodec}, pixel format: {camera_encoder.pix_fmt}, "
+        f"GOP: {camera_encoder.g}, CRF: {camera_encoder.crf}"
+    )
 
     # Create new features dict, converting image features to video features
     new_features = {}
@@ -1769,11 +1773,7 @@ def convert_image_to_video_dataset(
                 episode_indices=episode_indices,
                 temp_dir=temp_dir,
                 fps=fps,
-                vcodec=vcodec,
-                pix_fmt=pix_fmt,
-                g=g,
-                crf=crf,
-                fast_decode=fast_decode,
+                camera_encoder=camera_encoder,
             )
 
             logging.info(f"Processing camera: {img_key}")
@@ -1815,11 +1815,7 @@ def convert_image_to_video_dataset(
                     imgs_dir=imgs_dir,
                     video_path=video_path,
                     fps=fps,
-                    vcodec=vcodec,
-                    pix_fmt=pix_fmt,
-                    g=g,
-                    crf=crf,
-                    fast_decode=fast_decode,
+                    camera_encoder=camera_encoder,
                     overwrite=True,
                 )
 
@@ -1865,7 +1861,9 @@ def convert_image_to_video_dataset(
                 video_path = new_meta.root / new_meta.video_path.format(
                     video_key=img_key, chunk_index=0, file_index=0
                 )
-                new_meta.info.features[img_key]["info"] = get_video_info(video_path)
+                new_meta.info.features[img_key]["info"] = get_video_info(
+                    video_path, camera_encoder=camera_encoder
+                )
 
         write_info(new_meta.info, new_meta.root)
 
@@ -1888,3 +1886,83 @@ def convert_image_to_video_dataset(
 
     # Return new dataset
     return LeRobotDataset(repo_id=repo_id, root=output_dir)
+
+
+def _reencode_video_worker(args: tuple) -> Path:
+    """Picklable worker for :func:`reencode_dataset`'s process pool."""
+    video_path, camera_encoder, encoder_threads = args
+    reencode_video(
+        input_video_path=video_path,
+        output_video_path=video_path,
+        camera_encoder=camera_encoder,
+        encoder_threads=encoder_threads,
+        overwrite=True,
+    )
+    return video_path
+
+
+def reencode_dataset(
+    dataset: LeRobotDataset,
+    camera_encoder: VideoEncoderConfig,
+    encoder_threads: int | None = None,
+    num_workers: int | None = None,
+) -> LeRobotDataset:
+    """Re-encode every video in a dataset with a new set of encoding parameters.
+
+    Videos are re-encoded in-place and the video information in ``info.json`` is refreshed.
+
+    Args:
+        dataset: An existing :class:`LeRobotDataset` whose videos will be
+            re-encoded.
+        camera_encoder: Target encoder configuration applied to every video
+            file.
+        encoder_threads: Per-encoder thread count forwarded to
+            :func:`reencode_video`. ``None`` lets the codec decide.
+        num_workers: Number of parallel processes. ``None`` or ``0`` means
+            sequential (no multiprocessing); ``1+`` spawns a
+            :class:`~concurrent.futures.ProcessPoolExecutor`.
+
+    Returns:
+        The same :class:`LeRobotDataset` instance with its metadata updated
+        on disk.
+    """
+    meta = dataset.meta
+    video_paths_list = []
+
+    # Only re-encode if the videos are not already encoded with the given video encoding parameters
+    for video_key in meta.video_keys:
+        current_info = meta.info.features[video_key].get("info", {})
+        current_encoder = VideoEncoderConfig.from_video_info(current_info)
+        if current_encoder != camera_encoder:
+            video_paths_list.extend((meta.root / VIDEO_DIR / video_key).rglob("*.mp4"))
+        else:
+            logging.info(f"{video_key} videos are already encoded with {camera_encoder}. Nothing to do.")
+
+    if len(video_paths_list) == 0:
+        logging.warning("Dataset has no videos to re-encode.")
+        return dataset
+    logging.info(f"Re-encoding {len(video_paths_list)} video file(s) with {camera_encoder}")
+
+    worker_args = [(vp, camera_encoder, encoder_threads) for vp in video_paths_list]
+    if num_workers and num_workers > 1:
+        with ProcessPoolExecutor(max_workers=num_workers) as pool:
+            futures = [pool.submit(_reencode_video_worker, args) for args in worker_args]
+            for future in tqdm(
+                as_completed(futures),
+                total=len(futures),
+                desc="Re-encoding videos",
+            ):
+                future.result()
+    else:
+        for args in tqdm(worker_args, desc="Re-encoding videos"):
+            _reencode_video_worker(args)
+
+    # Refresh video info in metadata for every video key.
+    for vid_key in meta.video_keys:
+        video_path = meta.root / meta.get_video_file_path(0, vid_key)
+        meta.info.features[vid_key]["info"] = get_video_info(video_path, camera_encoder=camera_encoder)
+
+    write_info(meta.info, meta.root)
+    logging.info("Dataset metadata updated.")
+
+    return dataset

src/lerobot/datasets/dataset_writer.py

@@ -31,6 +31,8 @@ import PIL.Image
 import pyarrow.parquet as pq
 import torch
 
+from lerobot.configs import VideoEncoderConfig, camera_encoder_defaults
+
 from .compute_stats import compute_episode_stats
 from .dataset_metadata import LeRobotDatasetMetadata
 from .feature_utils import (
@@ -65,14 +67,19 @@ def _encode_video_worker(
     episode_index: int,
     root: Path,
     fps: int,
-    vcodec: str = "libsvtav1",
+    camera_encoder: VideoEncoderConfig | None = None,
     encoder_threads: int | None = None,
 ) -> Path:
     temp_path = Path(tempfile.mkdtemp(dir=root)) / f"{video_key}_{episode_index:03d}.mp4"
     fpath = DEFAULT_IMAGE_PATH.format(image_key=video_key, episode_index=episode_index, frame_index=0)
     img_dir = (root / fpath).parent
     encode_video_frames(
-        img_dir, temp_path, fps, vcodec=vcodec, overwrite=True, encoder_threads=encoder_threads
+        img_dir,
+        temp_path,
+        fps,
+        camera_encoder=camera_encoder,
+        encoder_threads=encoder_threads,
+        overwrite=True,
     )
     shutil.rmtree(img_dir)
     return temp_path
@@ -89,20 +96,22 @@ class DatasetWriter:
         self,
         meta: LeRobotDatasetMetadata,
         root: Path,
-        vcodec: str,
+        camera_encoder: VideoEncoderConfig | None,
         encoder_threads: int | None,
         batch_encoding_size: int,
         streaming_encoder: StreamingVideoEncoder | None = None,
         initial_frames: int = 0,
     ):
-        """Initialize the writer with metadata, codec, and encoding config.
+        """Initialize the writer with metadata, codec, and encoder config.
 
         Args:
             meta: Dataset metadata instance (used for feature schema, chunk
                 settings, and episode persistence).
             root: Local dataset root directory.
-            vcodec: Video codec for encoding (e.g. ``'libsvtav1'``, ``'h264'``).
-            encoder_threads: Threads per encoder instance. ``None`` for auto.
+            camera_encoder: Video encoder settings applied to all cameras.
+                ``None`` uses :func:`~lerobot.configs.camera_encoder_defaults`.
+            encoder_threads: Number of encoder threads (global). ``None``
+                lets the codec decide.
             batch_encoding_size: Number of episodes to accumulate before
                 batch-encoding videos.
             streaming_encoder: Optional pre-built :class:`StreamingVideoEncoder`
@@ -111,7 +120,7 @@ class DatasetWriter:
         """
         self._meta = meta
         self._root = root
-        self._vcodec = vcodec
+        self._camera_encoder = camera_encoder or camera_encoder_defaults()
         self._encoder_threads = encoder_threads
         self._batch_encoding_size = batch_encoding_size
         self._streaming_encoder = streaming_encoder
@@ -284,7 +293,7 @@ class DatasetWriter:
                             episode_index,
                             self._root,
                             self._meta.fps,
-                            self._vcodec,
+                            self._camera_encoder,
                             self._encoder_threads,
                         ): video_key
                         for video_key in self._meta.video_keys
@@ -495,7 +504,7 @@ class DatasetWriter:
 
         # Update video info (only needed when first episode is encoded)
         if episode_index == 0:
-            self._meta.update_video_info(video_key)
+            self._meta.update_video_info(video_key, camera_encoder=self._camera_encoder)
             write_info(self._meta.info, self._meta.root)
 
         metadata = {
@@ -564,7 +573,12 @@ class DatasetWriter:
     def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> Path:
         """Use ffmpeg to convert frames stored as png into mp4 videos."""
         return _encode_video_worker(
-            video_key, episode_index, self._root, self._meta.fps, self._vcodec, self._encoder_threads
+            video_key,
+            episode_index,
+            self._root,
+            self._meta.fps,
+            self._camera_encoder,
+            self._encoder_threads,
         )
 
     def close_writer(self) -> None:

src/lerobot/datasets/feature_utils.py

@@ -13,15 +13,23 @@
 # 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 pprint import pformat
 
 import datasets
 import numpy as np
 from PIL import Image as PILImage
 
+from lerobot.configs import VIDEO_ENCODER_INFO_KEYS
 from lerobot.utils.constants import DEFAULT_FEATURES
 from lerobot.utils.utils import is_valid_numpy_dtype_string
 
+from .language import (
+    LANGUAGE_PERSISTENT,
+    is_language_column,
+    language_events_column_feature,
+    language_persistent_column_feature,
+)
 from .utils import (
     DEFAULT_CHUNK_SIZE,
     DEFAULT_DATA_FILE_SIZE_IN_MB,
@@ -46,7 +54,13 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
     """
     hf_features = {}
     for key, ft in features.items():
-        if ft["dtype"] == "video":
+        if is_language_column(key):
+            hf_features[key] = (
+                language_persistent_column_feature()
+                if key == LANGUAGE_PERSISTENT
+                else language_events_column_feature()
+            )
+        elif ft["dtype"] == "video":
             continue
         elif ft["dtype"] == "image":
             hf_features[key] = datasets.Image()
@@ -108,6 +122,41 @@ def create_empty_dataset_info(
     )
 
 
+def features_equal_for_merge(features_a: dict[str, dict], features_b: dict[str, dict]) -> bool:
+    """Return whether two LeRobotDatasetMetadata ``features`` dicts are compatible for aggregation.
+
+    For video features, keys under ``info`` related to video encoding parameters are ignored during
+    comparison as they do not prevent aggregation.
+    """
+
+    def _without_encoder_info_keys(feature: dict) -> dict:
+        filtered = dict(feature)
+        filtered_info = filtered.get("info")
+        if isinstance(filtered_info, dict):
+            filtered["info"] = {
+                info_key: info_value
+                for info_key, info_value in filtered_info.items()
+                if info_key not in VIDEO_ENCODER_INFO_KEYS
+            }
+        return filtered
+
+    if set(features_a) != set(features_b):
+        return False
+    for key in features_a:
+        fa_key = features_a[key]
+        fb_key = features_b[key]
+        if fa_key.get("dtype") != fb_key.get("dtype"):
+            return False
+        if fa_key.get("dtype") != "video":
+            if fa_key != fb_key:
+                return False
+            continue
+
+        if _without_encoder_info_keys(fa_key) != _without_encoder_info_keys(fb_key):
+            return False
+    return True
+
+
 def check_delta_timestamps(
     delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
 ) -> bool:
@@ -242,6 +291,8 @@ def validate_feature_dtype_and_shape(
         return validate_feature_image_or_video(name, expected_shape, value)
     elif expected_dtype == "string":
         return validate_feature_string(name, value)
+    elif expected_dtype == "language":
+        return validate_feature_language(name, value)
     else:
         raise NotImplementedError(f"The feature dtype '{expected_dtype}' is not implemented yet.")
 
@@ -321,6 +372,30 @@ def validate_feature_string(name: str, value: str) -> str:
     return ""
 
 
+def validate_feature_language(name: str, value) -> str:
+    """Validate a feature that is expected to hold language annotations.
+
+    Language columns (``language_persistent`` / ``language_events``) are
+    populated after recording by the annotation pipeline, not at record time.
+    Any value supplied here is dropped before the frame is written, so a
+    non-empty value almost certainly signals a mistake. We warn rather than
+    fail to keep recording resilient.
+
+    Args:
+        name (str): The name of the feature.
+        value: The value to validate.
+
+    Returns:
+        str: Always an empty string — language values are non-fatal.
+    """
+    if value is not None:
+        logging.warning(
+            f"The feature '{name}' is a 'language' column populated by the annotation pipeline, "
+            f"not at record time. The provided value will be dropped."
+        )
+    return ""
+
+
 def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features: dict) -> None:
     """Validate the episode buffer before it's written to disk.
 

src/lerobot/datasets/io_utils.py

@@ -31,10 +31,10 @@ from torchvision import transforms
 from lerobot.utils.io_utils import load_json, write_json
 from lerobot.utils.utils import SuppressProgressBars, flatten_dict, unflatten_dict
 
+from .language import LANGUAGE_COLUMNS
 from .utils import (
     DEFAULT_DATA_FILE_SIZE_IN_MB,
     DEFAULT_EPISODES_PATH,
-    DEFAULT_SUBTASKS_PATH,
     DEFAULT_TASKS_PATH,
     EPISODES_DIR,
     INFO_PATH,
@@ -186,14 +186,6 @@ 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.
@@ -265,11 +257,13 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
         dict: The batch with items converted to torch tensors.
     """
     for key in items_dict:
+        if key in LANGUAGE_COLUMNS:
+            continue
         first_item = items_dict[key][0]
         if isinstance(first_item, PILImage.Image):
             to_tensor = transforms.ToTensor()
             items_dict[key] = [to_tensor(img) for img in items_dict[key]]
-        elif first_item is None:
+        elif first_item is None or isinstance(first_item, dict):
             pass
         else:
             items_dict[key] = [x if isinstance(x, str) else torch.tensor(x) for x in items_dict[key]]
@@ -304,8 +298,9 @@ def item_to_torch(item: dict) -> dict:
     Returns:
         dict: Dictionary with all tensor-like items converted to torch.Tensor.
     """
+    skip_keys = {"task", *LANGUAGE_COLUMNS}
     for key, val in item.items():
-        if isinstance(val, (np.ndarray | list)) and key not in ["task"]:
+        if isinstance(val, (np.ndarray | list)) and key not in skip_keys:
             # Convert numpy arrays and lists to torch tensors
             item[key] = torch.tensor(val)
     return item

src/lerobot/datasets/language.py

@@ -0,0 +1,242 @@
+#!/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 __future__ import annotations
+
+from typing import Literal
+
+import datasets
+import pyarrow as pa
+
+LANGUAGE_PERSISTENT = "language_persistent"
+LANGUAGE_EVENTS = "language_events"
+LANGUAGE_COLUMNS = (LANGUAGE_PERSISTENT, LANGUAGE_EVENTS)
+PERSISTENT_ROW_FIELDS = ("role", "content", "style", "timestamp", "camera", "tool_calls")
+EVENT_ROW_FIELDS = ("role", "content", "style", "camera", "tool_calls")
+
+CORE_STYLES = {
+    "subtask",
+    "plan",
+    "memory",
+    "motion",
+    "interjection",
+    "vqa",
+    "trace",
+    "task_aug",
+}
+# Project-local styles can be registered at import time by appending to
+# ``EXTENDED_STYLES`` before ``column_for_style`` is called. Anything added
+# here is treated as a known style alongside ``CORE_STYLES`` for resolver
+# validation. Empty by default — populate from a downstream module that
+# also extends ``PERSISTENT_STYLES`` or ``EVENT_ONLY_STYLES`` to declare
+# the new style's column.
+EXTENDED_STYLES: set[str] = set()
+STYLE_REGISTRY = CORE_STYLES | EXTENDED_STYLES
+
+PERSISTENT_STYLES = {"subtask", "plan", "memory", "motion", "task_aug"}
+EVENT_ONLY_STYLES = {"interjection", "vqa", "trace"}
+
+# Styles whose ``content`` is grounded in a specific camera view. Rows of these
+# styles MUST carry a non-null ``camera`` referencing an ``observation.images.*``
+# feature key. Rows of every other style MUST have ``camera=None``. ``motion``
+# is intentionally NOT in this set: motion primitives are described in
+# robot-frame (joint / Cartesian) terms, not pixel space, so they are
+# camera-agnostic. ``trace`` is the pixel-trajectory event style and IS
+# view-dependent. The ``camera`` field nevertheless lives on
+# ``PERSISTENT_ROW_FIELDS`` too so the schema, validator, and resolver
+# behave symmetrically across the two columns; persistent rows simply
+# always have ``camera=None`` in practice today.
+VIEW_DEPENDENT_STYLES = {"vqa", "trace"}
+
+LanguageColumn = Literal["language_persistent", "language_events"]
+
+
+def _json_arrow_type() -> pa.DataType:
+    """Return the Arrow JSON type, falling back to ``string`` on older pyarrow."""
+    return pa.json_() if hasattr(pa, "json_") else pa.string()
+
+
+def _json_feature() -> object:
+    """Return the HF ``datasets`` JSON feature, falling back to a string value."""
+    return datasets.Json() if hasattr(datasets, "Json") else datasets.Value("string")
+
+
+def language_persistent_row_arrow_type() -> pa.StructType:
+    """Return the Arrow struct type for a single persistent language row.
+
+    Persistent rows carry their own ``timestamp`` because they represent a state
+    that became active at a specific moment and remains active until superseded.
+    ``timestamp`` is ``float32`` to match the timestamp dtype LeRobotDataset
+    uses for frame data.
+    """
+    return pa.struct(
+        [
+            pa.field("role", pa.string(), nullable=False),
+            pa.field("content", pa.string(), nullable=True),
+            pa.field("style", pa.string(), nullable=True),
+            pa.field("timestamp", pa.float32(), nullable=False),
+            pa.field("camera", pa.string(), nullable=True),
+            pa.field("tool_calls", pa.list_(_json_arrow_type()), nullable=True),
+        ]
+    )
+
+
+def language_event_row_arrow_type() -> pa.StructType:
+    """Return the Arrow struct type for a single event language row.
+
+    Event rows have no ``timestamp`` field: each event is stored on the dataset
+    row whose frame timestamp is the event's firing time.
+    """
+    return pa.struct(
+        [
+            pa.field("role", pa.string(), nullable=False),
+            pa.field("content", pa.string(), nullable=True),
+            pa.field("style", pa.string(), nullable=True),
+            pa.field("camera", pa.string(), nullable=True),
+            pa.field("tool_calls", pa.list_(_json_arrow_type()), nullable=True),
+        ]
+    )
+
+
+def language_persistent_arrow_type() -> pa.ListType:
+    """Return the Arrow list type for the ``language_persistent`` column."""
+    return pa.list_(language_persistent_row_arrow_type())
+
+
+def language_events_arrow_type() -> pa.ListType:
+    """Return the Arrow list type for the ``language_events`` column."""
+    return pa.list_(language_event_row_arrow_type())
+
+
+def language_persistent_row_feature() -> dict[str, object]:
+    """Return the HF ``datasets`` feature mapping for a persistent language row."""
+    return {
+        "role": datasets.Value("string"),
+        "content": datasets.Value("string"),
+        "style": datasets.Value("string"),
+        "timestamp": datasets.Value("float32"),
+        "camera": datasets.Value("string"),
+        "tool_calls": datasets.List(_json_feature()),
+    }
+
+
+def language_event_row_feature() -> dict[str, object]:
+    """Return the HF ``datasets`` feature mapping for an event language row."""
+    return {
+        "role": datasets.Value("string"),
+        "content": datasets.Value("string"),
+        "style": datasets.Value("string"),
+        "camera": datasets.Value("string"),
+        "tool_calls": datasets.List(_json_feature()),
+    }
+
+
+def language_persistent_column_feature() -> datasets.List:
+    """Return the HF ``datasets`` feature for the ``language_persistent`` column."""
+    return datasets.List(language_persistent_row_feature())
+
+
+def language_events_column_feature() -> datasets.List:
+    """Return the HF ``datasets`` feature for the ``language_events`` column."""
+    return datasets.List(language_event_row_feature())
+
+
+def language_feature_info() -> dict[str, dict]:
+    """Return the ``info["features"]`` entries for both language columns."""
+    return {
+        LANGUAGE_PERSISTENT: {"dtype": "language", "shape": (1,), "names": None},
+        LANGUAGE_EVENTS: {"dtype": "language", "shape": (1,), "names": None},
+    }
+
+
+def is_language_column(key: str) -> bool:
+    """Return ``True`` if ``key`` is one of the dataset's language column names."""
+    return key in LANGUAGE_COLUMNS
+
+
+def is_view_dependent_style(style: str | None) -> bool:
+    """Return ``True`` if rows of ``style`` must be tagged with a ``camera`` key."""
+    return style in VIEW_DEPENDENT_STYLES
+
+
+def validate_camera_field(style: str | None, camera: str | None) -> None:
+    """Enforce the ``camera`` invariant: required iff ``style`` is view-dependent.
+
+    Raises ``ValueError`` if a view-dependent style is missing ``camera`` or if
+    a non-view-dependent style carries one. Pipeline writers and the validator
+    should call this on every emitted row.
+    """
+    if is_view_dependent_style(style):
+        if not camera:
+            raise ValueError(
+                f"Rows of view-dependent style {style!r} require a non-empty 'camera' "
+                f"field referencing an 'observation.images.*' feature key."
+            )
+    elif camera is not None:
+        raise ValueError(f"Rows of style {style!r} must have camera=None; got camera={camera!r}.")
+
+
+# --- Tool registry --------------------------------------------------------
+# Tools declared on a dataset live in ``meta/info.json["tools"]`` as a list
+# of OpenAI-style function schemas. The runtime / training stack reads them
+# through :class:`LeRobotDatasetMetadata.tools` (with these constants as
+# fallback when the dataset doesn't declare any). Implementations live
+# under :mod:`lerobot.tools` (one file per tool); see
+# ``docs/source/tools.mdx`` for the authoring guide.
+
+SAY_TOOL_SCHEMA: dict = {
+    "type": "function",
+    "function": {
+        "name": "say",
+        "description": "Speak a short utterance to the user via the TTS executor.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "text": {
+                    "type": "string",
+                    "description": "The verbatim text to speak.",
+                }
+            },
+            "required": ["text"],
+        },
+    },
+}
+"""Canonical schema for the ``say`` tool emitted by the steerable
+annotation pipeline (PR 2 Module 2). Single source of truth — PR 2's
+writer, PR 3's runtime tool registry, and the dataset visualizer all
+import this constant rather than duplicating the dict."""
+
+DEFAULT_TOOLS: list[dict] = [SAY_TOOL_SCHEMA]
+"""Fallback tools list. Returned by ``LeRobotDatasetMetadata.tools``
+when ``meta/info.json["tools"]`` is unset, so unannotated datasets and
+chat-template consumers (``apply_chat_template(messages, tools=...)``)
+keep working out of the box."""
+
+
+def column_for_style(style: str | None) -> LanguageColumn:
+    """Map a language style to the column where rows of that style are stored.
+
+    Styles in :data:`PERSISTENT_STYLES` route to :data:`LANGUAGE_PERSISTENT`.
+    Styles in :data:`EVENT_ONLY_STYLES` and the implicit ``None`` style route
+    to :data:`LANGUAGE_EVENTS`.
+    """
+    if style is None:
+        return LANGUAGE_EVENTS
+    if style in PERSISTENT_STYLES:
+        return LANGUAGE_PERSISTENT
+    if style in EVENT_ONLY_STYLES:
+        return LANGUAGE_EVENTS
+    raise ValueError(f"Unknown language style: {style!r}")

src/lerobot/datasets/language_render.py

@@ -0,0 +1,545 @@
+#!/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 __future__ import annotations
+
+import copy
+import hashlib
+import re
+from collections.abc import Sequence
+from typing import Any
+
+from lerobot.configs.recipe import DEFAULT_BINDINGS, PLACEHOLDER_RE, TrainingRecipe
+from lerobot.utils.utils import unwrap_scalar
+
+from .language import LANGUAGE_PERSISTENT, column_for_style
+
+LanguageRow = dict[str, Any]
+RenderedMessages = dict[str, list[Any]]
+
+_RESOLVER_RE = re.compile(r"^(?P<name>[A-Za-z_][A-Za-z0-9_]*)\((?P<args>.*)\)$")
+
+
+def active_at(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    style: str | None = None,
+    role: str | None = None,
+    tool_name: str | None = None,
+    camera: str | None = None,
+) -> LanguageRow | None:
+    """Return the persistent row of ``style`` that is active at time ``t``.
+
+    A persistent row is "active" at ``t`` when its own ``timestamp`` is the
+    most recent one ``<= t`` for the given ``style``/``role``/``tool_name``/
+    ``camera`` selector. Only valid for persistent styles.
+    """
+    _validate_persistent_resolver("active_at", style)
+    matches = [
+        row
+        for row in _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera)
+        if _timestamp(row) <= t
+    ]
+    if not matches:
+        return None
+    latest_ts = max(_timestamp(row) for row in matches)
+    return _select_one(
+        [row for row in matches if _timestamp(row) == latest_ts],
+        style=style,
+        role=role,
+        tool_name=tool_name,
+        camera=camera,
+    )
+
+
+EMITTED_AT_TOLERANCE_S = 0.1
+"""Half-window for matching persistent rows to a frame timestamp in
+``emitted_at``. Persistent timestamps come from parquet (float32) and ``t``
+is also a float32 from parquet, so in the ideal hot path an exact match
+would suffice — but any caller that derives ``t`` arithmetically (e.g.
+``frame_idx / fps``) breaks bit-equality. A 0.1 s tolerance covers
+common arithmetic drift without admitting frames that are visibly far
+apart at typical control rates (30–100 Hz). This does mean two persistent
+rows of the same selector emitted within 0.1 s of each other cannot be
+told apart by ``emitted_at`` — acceptable because persistent annotations
+(subtask / plan / memory transitions) change on a human-action timescale,
+not at the camera frame rate."""
+
+
+def emitted_at(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    style: str | None = None,
+    role: str | None = None,
+    tool_name: str | None = None,
+    camera: str | None = None,
+) -> LanguageRow | None:
+    """Return the row of ``style`` emitted at exactly time ``t``.
+
+    For persistent styles, this matches persistent rows whose own ``timestamp``
+    is within ``EMITTED_AT_TOLERANCE_S`` of ``t`` (see that constant for why
+    we use a tolerance instead of bit-equality). For event styles, the
+    ``events`` list is assumed to come from the dataset row at frame ``t``
+    (event rows carry no timestamp of their own), so all matching event rows
+    are considered emitted at ``t``. ``camera`` filters by the row's
+    ``camera`` field — required to disambiguate when multiple view-dependent
+    rows share ``(t, role)`` across cameras.
+    """
+    if column_for_style(style) == LANGUAGE_PERSISTENT:
+        matches = [
+            row
+            for row in _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera)
+            if abs(_timestamp(row) - t) <= EMITTED_AT_TOLERANCE_S
+        ]
+    else:
+        matches = _matching_rows(events, style=style, role=role, tool_name=tool_name, camera=camera)
+    return _select_one(matches, style=style, role=role, tool_name=tool_name, camera=camera)
+
+
+def nth_prev(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    style: str | None = None,
+    offset: int = 1,
+    role: str | None = None,
+    tool_name: str | None = None,
+    camera: str | None = None,
+) -> LanguageRow | None:
+    """Return the persistent row that was active ``offset`` steps before ``t``.
+
+    Walks back through chronologically sorted persistent rows of ``style``
+    (filtered by optional ``role``/``tool_name``/``camera``) and returns the
+    one ``offset`` positions before the row active at ``t``. Only valid for
+    persistent styles.
+    """
+    return _nth_relative("nth_prev", t, persistent, style, -offset, role, tool_name, camera)
+
+
+def nth_next(
+    t: float,
+    *,
+    persistent: Sequence[LanguageRow],
+    style: str | None = None,
+    offset: int = 1,
+    role: str | None = None,
+    tool_name: str | None = None,
+    camera: str | None = None,
+) -> LanguageRow | None:
+    """Return the persistent row that becomes active ``offset`` steps after ``t``.
+
+    Walks forward through chronologically sorted persistent rows of ``style``
+    (filtered by optional ``role``/``tool_name``/``camera``) and returns the
+    one ``offset`` positions after the row active at ``t``. Only valid for
+    persistent styles.
+    """
+    return _nth_relative("nth_next", t, persistent, style, offset, role, tool_name, camera)
+
+
+def render_sample(
+    *,
+    recipe: TrainingRecipe,
+    persistent: Sequence[LanguageRow] | None,
+    events: Sequence[LanguageRow] | None,
+    t: float,
+    sample_idx: int,
+    task: str | None = None,
+    dataset_ctx: Any | None = None,
+) -> RenderedMessages | None:
+    """Render the chat-style messages for a single dataset sample.
+
+    Resolves the recipe's bindings against ``persistent`` and ``events`` rows
+    at frame timestamp ``t``, then expands the recipe's message templates.
+    Returns ``None`` if the resolved sample contains no target message.
+    """
+    persistent_rows = _normalize_rows(persistent or [])
+    event_rows = _normalize_rows(events or [])
+    selected_recipe = _select_recipe(recipe, sample_idx)
+    bindings = _resolve_bindings(
+        selected_recipe,
+        persistent=persistent_rows,
+        events=event_rows,
+        t=t,
+        sample_idx=sample_idx,
+        task=task,
+        dataset_ctx=dataset_ctx,
+    )
+    return _render_message_recipe(selected_recipe, bindings)
+
+
+def _select_recipe(recipe: TrainingRecipe, sample_idx: int) -> TrainingRecipe:
+    """Pick a deterministic blend component for ``sample_idx`` (or return ``recipe``)."""
+    if recipe.blend is None:
+        return recipe
+
+    total_weight = sum(component.weight or 0.0 for component in recipe.blend.values())
+    if total_weight <= 0:
+        raise ValueError("Blend weights must sum to a positive value.")
+
+    digest = hashlib.blake2b(str(sample_idx).encode(), digest_size=8).digest()
+    draw = int.from_bytes(digest, "big") / 2**64 * total_weight
+    cumulative = 0.0
+    last_component: TrainingRecipe | None = None
+    for component in recipe.blend.values():
+        last_component = component
+        cumulative += component.weight or 0.0
+        if draw < cumulative:
+            return component
+    assert last_component is not None
+    return last_component
+
+
+def _resolve_bindings(
+    recipe: TrainingRecipe,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    t: float,
+    sample_idx: int,
+    task: str | None,
+    dataset_ctx: Any | None,
+) -> dict[str, LanguageRow | str | None]:
+    """Resolve every binding in ``recipe`` (plus ``task``) at time ``t``."""
+    bindings: dict[str, LanguageRow | str | None] = {
+        "task": _resolve_task(task, dataset_ctx, persistent=persistent, sample_idx=sample_idx),
+    }
+    specs = {**DEFAULT_BINDINGS, **(recipe.bindings or {})}
+    for name, spec in specs.items():
+        bindings[name] = _resolve_spec(spec, persistent=persistent, events=events, t=t)
+    return bindings
+
+
+def _resolve_task(
+    task: str | None,
+    dataset_ctx: Any | None,
+    *,
+    persistent: Sequence[LanguageRow] = (),
+    sample_idx: int = 0,
+) -> str | None:
+    """Return the task string for ``sample_idx``.
+
+    Resolution order:
+
+    1. Explicit ``task`` override (caller-supplied) wins.
+    2. If ``persistent`` contains rows of style ``task_aug`` (role=user),
+       deterministically pick one by ``sample_idx`` so each frame of an
+       episode rotates through the available rephrasings across an epoch.
+       This realizes Xiao 2022 / CAST-style task-prompt diversity without
+       changing ``meta/tasks.parquet`` and without forcing recipes to opt
+       in: ``${task}`` automatically picks a rephrasing when one exists,
+       and falls back to the canonical task otherwise. Recipes that want
+       the literal canonical task can override the binding.
+    3. Otherwise read the canonical task from ``dataset_ctx`` (which is
+       backed by ``meta/tasks.parquet``).
+    """
+    if task is not None:
+        return task
+
+    aug_rows = [r for r in persistent if r.get("style") == "task_aug" and r.get("role") == "user"]
+    if aug_rows:
+        # Deterministic, blake2b-based pick keyed on sample_idx so the
+        # rotation is reproducible across runs (Python's built-in ``hash``
+        # is process-randomized).
+        digest = hashlib.blake2b(f"task_aug:{sample_idx}".encode(), digest_size=8).digest()
+        idx = int.from_bytes(digest, "big") % len(aug_rows)
+        chosen = aug_rows[idx].get("content")
+        if chosen:
+            return str(chosen)
+
+    if dataset_ctx is None:
+        return None
+    if isinstance(dataset_ctx, dict):
+        return dataset_ctx.get("task")
+    return getattr(dataset_ctx, "task", None)
+
+
+def _resolve_spec(
+    spec: str,
+    *,
+    persistent: Sequence[LanguageRow],
+    events: Sequence[LanguageRow],
+    t: float,
+) -> LanguageRow | None:
+    """Parse a single binding's resolver expression and dispatch to its function."""
+    match = _RESOLVER_RE.match(spec.strip())
+    if match is None:
+        raise ValueError(f"Invalid resolver expression: {spec!r}")
+    name = match.group("name")
+    kwargs = _parse_resolver_args(match.group("args"))
+    kwargs.pop("t_arg", None)
+
+    if name == "emitted_at":
+        return emitted_at(t, persistent=persistent, events=events, **kwargs)
+    if name == "active_at":
+        return active_at(t, persistent=persistent, **kwargs)
+    if name == "nth_prev":
+        return nth_prev(t, persistent=persistent, **kwargs)
+    if name == "nth_next":
+        return nth_next(t, persistent=persistent, **kwargs)
+    raise ValueError(f"Unknown language resolver: {name!r}")
+
+
+def _parse_resolver_args(args: str) -> dict[str, Any]:
+    """Parse a comma-separated resolver argument list into a kwargs dict."""
+    kwargs: dict[str, Any] = {}
+    if not args.strip():
+        return kwargs
+
+    parts = [part.strip() for part in args.split(",") if part.strip()]
+    for part in parts:
+        if part == "t":
+            kwargs["t_arg"] = True
+            continue
+        if "=" not in part:
+            raise ValueError(f"Invalid resolver argument: {part!r}")
+        key, value = (item.strip() for item in part.split("=", 1))
+        if key == "offset":
+            kwargs[key] = int(value)
+        else:
+            kwargs[key] = value.strip("\"'")
+    return kwargs
+
+
+def _render_message_recipe(
+    recipe: TrainingRecipe,
+    bindings: dict[str, LanguageRow | str | None],
+) -> RenderedMessages | None:
+    """Expand ``recipe.messages`` into rendered chat messages using ``bindings``."""
+    assert recipe.messages is not None
+    messages: list[dict[str, Any]] = []
+    streams: list[str | None] = []
+    target_indices: list[int] = []
+
+    for turn in recipe.messages:
+        if turn.if_present is not None and bindings.get(turn.if_present) is None:
+            continue
+
+        message = {"role": turn.role}
+        if turn.content is not None:
+            message["content"] = _render_content(turn.content, bindings)
+
+        if turn.tool_calls_from is not None:
+            row = bindings.get(turn.tool_calls_from)
+            tool_calls = row.get("tool_calls") if isinstance(row, dict) else None
+            if tool_calls:
+                message["tool_calls"] = copy.deepcopy(tool_calls)
+
+        message_idx = len(messages)
+        messages.append(message)
+        streams.append(turn.stream)
+        if turn.target:
+            target_indices.append(message_idx)
+
+    if not target_indices:
+        return None
+
+    rendered = {
+        "messages": messages,
+        "message_streams": streams,
+        "target_message_indices": target_indices,
+    }
+    _validate_rendered(rendered)
+    return rendered
+
+
+def _render_content(
+    content: str | list[dict[str, Any]],
+    bindings: dict[str, LanguageRow | str | None],
+) -> str | list[dict[str, Any]]:
+    """Substitute bindings into a string or each string field of multimodal blocks."""
+    if isinstance(content, str):
+        return _substitute(content, bindings)
+
+    rendered_blocks = []
+    for block in content:
+        rendered_block = copy.deepcopy(block)
+        for key, value in rendered_block.items():
+            if isinstance(value, str):
+                rendered_block[key] = _substitute(value, bindings)
+        rendered_blocks.append(rendered_block)
+    return rendered_blocks
+
+
+def _substitute(template: str, bindings: dict[str, LanguageRow | str | None]) -> str:
+    """Replace ``${name}`` placeholders in ``template`` with their bound values."""
+
+    def replace(match: re.Match[str]) -> str:
+        """Resolve a single ``${name}`` match to its bound string value."""
+        name = match.group(1)
+        if name not in bindings:
+            raise ValueError(f"Unknown template binding: {name!r}")
+        value = bindings[name]
+        if value is None:
+            return ""
+        if isinstance(value, dict):
+            content = value.get("content")
+            return "" if content is None else str(content)
+        return str(value)
+
+    return PLACEHOLDER_RE.sub(replace, template)
+
+
+def _validate_rendered(rendered: RenderedMessages) -> None:
+    """Sanity-check the rendered output for stream/target alignment."""
+    messages = rendered["messages"]
+    streams = rendered["message_streams"]
+    target_indices = rendered["target_message_indices"]
+
+    if len(streams) != len(messages):
+        raise ValueError("message_streams must be aligned with messages.")
+    if not target_indices:
+        raise ValueError("Rendered samples must contain at least one target message.")
+    for idx in target_indices:
+        if idx < 0 or idx >= len(messages):
+            raise ValueError(f"Target message index {idx} is out of bounds.")
+    # ``stream`` is enforced non-None at MessageTurn construction time
+    # (see ``MessageTurn.__post_init__``), so a missing stream here would
+    # mean the dataclass invariant was bypassed; no need to re-check.
+
+
+def _nth_relative(
+    name: str,
+    t: float,
+    persistent: Sequence[LanguageRow],
+    style: str | None,
+    offset: int,
+    role: str | None,
+    tool_name: str | None,
+    camera: str | None,
+) -> LanguageRow | None:
+    """Shared body for ``nth_prev`` / ``nth_next`` with signed ``offset``."""
+    _validate_persistent_resolver(name, style)
+    if abs(offset) < 1:
+        raise ValueError(f"{name} offset must be non-zero.")
+
+    rows = sorted(
+        _matching_rows(persistent, style=style, role=role, tool_name=tool_name, camera=camera),
+        key=_row_sort_key,
+    )
+    if not rows:
+        return None
+
+    anchor_idx = None
+    for idx, row in enumerate(rows):
+        if _timestamp(row) <= t:
+            anchor_idx = idx
+        else:
+            break
+
+    target_idx = (offset - 1 if offset > 0 else None) if anchor_idx is None else anchor_idx + offset
+
+    if target_idx is None or target_idx < 0 or target_idx >= len(rows):
+        return None
+    return rows[target_idx]
+
+
+def _validate_persistent_resolver(name: str, style: str | None) -> None:
+    """Reject calls with missing or event-only ``style`` for persistent resolvers."""
+    if style is None:
+        raise ValueError(f"{name} requires a persistent style.")
+    if column_for_style(style) != LANGUAGE_PERSISTENT:
+        raise ValueError(f"{name} cannot be used with event-only style {style!r}.")
+
+
+def _matching_rows(
+    rows: Sequence[LanguageRow],
+    *,
+    style: str | None,
+    role: str | None,
+    tool_name: str | None,
+    camera: str | None,
+) -> list[LanguageRow]:
+    """Return ``rows`` filtered by optional ``style``/``role``/``tool_name``/``camera`` selectors."""
+    return [
+        row
+        for row in rows
+        if (style is None or row.get("style") == style)
+        and (role is None or row.get("role") == role)
+        and (tool_name is None or _row_has_tool_name(row, tool_name))
+        and (camera is None or row.get("camera") == camera)
+    ]
+
+
+def _select_one(
+    rows: Sequence[LanguageRow],
+    *,
+    style: str | None,
+    role: str | None,
+    tool_name: str | None,
+    camera: str | None,
+) -> LanguageRow | None:
+    """Return the single matching row, or raise if the resolver is ambiguous.
+
+    Multiple matches always raise — even when the caller already passed
+    some selectors — because remaining ambiguity means the data has
+    several rows that look identical to the resolver and the caller
+    needs to pin down a specific one (e.g. add ``camera=...`` for VQA
+    rows shared across cameras).
+    """
+    if not rows:
+        return None
+    if len(rows) > 1:
+        raise ValueError(
+            f"Ambiguous resolver for style={style!r} role={role!r} "
+            f"tool_name={tool_name!r} camera={camera!r}: {len(rows)} matching rows. "
+            f"Add a selector that distinguishes them."
+        )
+    return rows[0]
+
+
+def _row_sort_key(row: LanguageRow) -> tuple[float, str, str]:
+    """Stable sort key for both persistent and event rows.
+
+    Event rows lack ``timestamp`` (it is implicit in the frame), so default
+    to ``0.0`` — within a single frame all event rows share the same sort
+    bucket and are tiebroken by ``(style, role)``.
+    """
+    timestamp = row.get("timestamp")
+    ts = float(unwrap_scalar(timestamp)) if timestamp is not None else 0.0
+    return (ts, row.get("style") or "", row.get("role") or "")
+
+
+def _timestamp(row: LanguageRow) -> float:
+    """Extract a row's ``timestamp`` as a Python float (unwrapping numpy scalars)."""
+    return float(unwrap_scalar(row["timestamp"]))
+
+
+def _row_has_tool_name(row: LanguageRow, tool_name: str) -> bool:
+    """Return ``True`` if any of the row's tool calls invokes ``tool_name``."""
+    for tool_call in row.get("tool_calls") or []:
+        if isinstance(tool_call, str):
+            continue
+        function = tool_call.get("function") if isinstance(tool_call, dict) else None
+        if isinstance(function, dict) and function.get("name") == tool_name:
+            return True
+    return False
+
+
+def _normalize_rows(rows: Sequence[Any]) -> list[LanguageRow]:
+    """Convert pyarrow scalars / mappings into a fresh list of plain dict rows."""
+    normalized = []
+    for row in rows:
+        if row is None:
+            continue
+        if hasattr(row, "as_py"):
+            row = row.as_py()
+        if not isinstance(row, dict):
+            raise TypeError(f"Language rows must be dictionaries, got {type(row).__name__}.")
+        normalized.append(dict(row))
+    return normalized

src/lerobot/datasets/lerobot_dataset.py

@@ -24,6 +24,7 @@ import torch.utils
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.errors import RevisionNotFoundError
 
+from lerobot.configs import VideoEncoderConfig
 from lerobot.utils.constants import HF_LEROBOT_HUB_CACHE
 
 from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
@@ -36,8 +37,7 @@ from .utils import (
 )
 from .video_utils import (
     StreamingVideoEncoder,
-    get_safe_default_codec,
-    resolve_vcodec,
+    get_safe_default_video_backend,
 )
 
 logger = logging.getLogger(__name__)
@@ -49,6 +49,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         repo_id: str,
         root: str | Path | None = None,
         episodes: list[int] | None = None,
+        episode_filter: Callable[[dict], bool] | None = None,
         image_transforms: Callable | None = None,
         delta_timestamps: dict[str, list[float]] | None = None,
         tolerance_s: float = 1e-4,
@@ -58,10 +59,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         video_backend: str | None = None,
         return_uint8: bool = False,
         batch_encoding_size: int = 1,
-        vcodec: str = "libsvtav1",
+        camera_encoder: VideoEncoderConfig | None = None,
+        encoder_threads: int | None = None,
         streaming_encoding: bool = False,
         encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
     ):
         """
         2 modes are available for instantiating this class, depending on 2 different use cases:
@@ -153,6 +154,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 ``$HF_LEROBOT_HOME/hub``.
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list. Defaults to None.
+            episode_filter (Callable[[dict], bool] | None, optional): Predicate over per-episode
+                metadata rows used to select episodes. Evaluated against ``meta/`` without ``stats`` keys
+                (e.g.``task_index``, ``episode_index``, ``length``, ``from_timestamp``, ``to_timestamp``).
+                Intersected with ``episodes`` when both are set. Example: ``lambda ep: ep["length"] >= 100``.
+                Defaults to None.
             image_transforms (Callable | None, optional):
                 Transform applied to visual modalities inside `__getitem__` after image decoding / tensor
                 conversion. This works for both image-backed and video-backed observations and can later be
@@ -177,16 +183,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 You can also use the 'pyav' decoder used by Torchvision, which used to be the default option, or 'video_reader' which is another decoder of Torchvision.
             batch_encoding_size (int, optional): Number of episodes to accumulate before batch encoding videos.
                 Set to 1 for immediate encoding (default), or higher for batched encoding. Defaults to 1.
-            vcodec (str, optional): Video codec for encoding videos during recording. Options: 'h264', 'hevc',
-                'libsvtav1', 'auto', or hardware-specific codecs like 'h264_videotoolbox', 'h264_nvenc'.
-                Defaults to 'libsvtav1'. Use 'auto' to auto-detect the best available hardware encoder.
+            camera_encoder (VideoEncoderConfig | None, optional): Video encoder settings for cameras
+                (codec, quality, etc.). When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults`
+                is used by the writer.
+            encoder_threads (int | None, optional): Number of encoder threads (global). ``None`` lets the
+                codec decide.
             streaming_encoding (bool, optional): If True, encode video frames in real-time during capture
                 instead of writing PNG images first. This makes save_episode() near-instant. Defaults to False.
             encoder_queue_maxsize (int, optional): Maximum number of frames to buffer per camera when using
                 streaming encoding. Defaults to 30 (~1s at 30fps).
-            encoder_threads (int | None, optional): Number of threads per encoder instance. None lets the
-                codec auto-detect (default). Lower values reduce CPU usage per encoder. Maps to 'lp' (via svtav1-params) for
-                libsvtav1 and 'threads' for h264/hevc.
 
         Note:
             Write-mode parameters (``streaming_encoding``, ``batch_encoding_size``) passed to
@@ -199,13 +204,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.reader = None
         self.set_image_transforms(image_transforms)
         self.delta_timestamps = delta_timestamps
-        self.episodes = episodes
         self.tolerance_s = tolerance_s
         self.revision = revision if revision else CODEBASE_VERSION
-        self._video_backend = video_backend if video_backend else get_safe_default_codec()
+        self._video_backend = video_backend if video_backend else get_safe_default_video_backend()
         self._return_uint8 = return_uint8
         self._batch_encoding_size = batch_encoding_size
-        self._vcodec = resolve_vcodec(vcodec)
         self._encoder_threads = encoder_threads
 
         if self._requested_root is not None:
@@ -218,6 +221,23 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.root = self.meta.root
         self.revision = self.meta.revision
 
+        if episodes is not None and any(
+            episode >= self.meta.total_episodes or episode < 0 for episode in episodes
+        ):
+            logger.warning(
+                f"Some episodes in the provided episodes list are out of range for this dataset ({self.meta.total_episodes})."
+            )
+
+        if episode_filter is not None:
+            resolved = self.meta.filter_episodes(episode_filter, candidates=episodes)
+            if not resolved:
+                raise ValueError(
+                    "The episode filter did not match any episode. Make sure the filter and episodes list are valid and compatible."
+                )
+            logger.info(f"The episode filter matched {len(resolved)} episode(s).")
+            episodes = resolved
+        self.episodes = episodes
+
         # Create reader (hf_dataset loaded below)
         self.reader = DatasetReader(
             meta=self.meta,
@@ -251,12 +271,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
             streaming_enc = None
             if streaming_encoding and len(self.meta.video_keys) > 0:
                 streaming_enc = self._build_streaming_encoder(
-                    self.meta.fps, self._vcodec, encoder_queue_maxsize, encoder_threads
+                    self.meta.fps,
+                    camera_encoder,
+                    encoder_queue_maxsize,
+                    encoder_threads,
                 )
             self.writer = DatasetWriter(
                 meta=self.meta,
                 root=self.root,
-                vcodec=self._vcodec,
+                camera_encoder=camera_encoder,
                 encoder_threads=encoder_threads,
                 batch_encoding_size=batch_encoding_size,
                 streaming_encoder=streaming_enc,
@@ -298,17 +321,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
     @staticmethod
     def _build_streaming_encoder(
         fps: int,
-        vcodec: str,
+        camera_encoder: VideoEncoderConfig | None,
         encoder_queue_maxsize: int,
         encoder_threads: int | None,
     ) -> StreamingVideoEncoder:
         return StreamingVideoEncoder(
             fps=fps,
-            vcodec=vcodec,
-            pix_fmt="yuv420p",
-            g=2,
-            crf=30,
-            preset=None,
+            camera_encoder=camera_encoder,
             queue_maxsize=encoder_queue_maxsize,
             encoder_threads=encoder_threads,
         )
@@ -625,7 +644,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         image_writer_threads: int = 0,
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
-        vcodec: str = "libsvtav1",
+        camera_encoder: VideoEncoderConfig | None = None,
         metadata_buffer_size: int = 10,
         streaming_encoding: bool = False,
         encoder_queue_maxsize: int = 30,
@@ -656,20 +675,20 @@ class LeRobotDataset(torch.utils.data.Dataset):
             video_backend: Video decoding backend (used when reading back).
             batch_encoding_size: Number of episodes to accumulate before
                 batch-encoding videos. ``1`` means encode immediately.
-            vcodec: Video codec for encoding. Options include ``'libsvtav1'``,
-                ``'h264'``, ``'hevc'``, ``'auto'``.
+            camera_encoder: Video encoder settings for cameras (codec, quality, etc.).
+                When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used.
+            encoder_threads: Number of encoder threads (global). ``None``
+                lets the codec decide.
             metadata_buffer_size: Number of episode metadata records to buffer
                 before flushing to parquet.
             streaming_encoding: If ``True``, encode video frames in real-time
                 during capture instead of writing images first.
             encoder_queue_maxsize: Max buffered frames per camera when using
                 streaming encoding.
-            encoder_threads: Threads per encoder instance. ``None`` for auto.
 
         Returns:
             A new :class:`LeRobotDataset` in write mode.
         """
-        vcodec = resolve_vcodec(vcodec)
         obj = cls.__new__(cls)
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
@@ -690,23 +709,23 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_transforms = None
         obj.delta_timestamps = None
         obj.episodes = None
-        obj._video_backend = video_backend if video_backend is not None else get_safe_default_codec()
+        obj._video_backend = video_backend if video_backend is not None else get_safe_default_video_backend()
         obj._return_uint8 = False
         obj._batch_encoding_size = batch_encoding_size
-        obj._vcodec = vcodec
         obj._encoder_threads = encoder_threads
 
         # Reader is lazily created on first access (write-only mode)
         obj.reader = None
 
-        # Create writer
         streaming_enc = None
         if streaming_encoding and len(obj.meta.video_keys) > 0:
-            streaming_enc = cls._build_streaming_encoder(fps, vcodec, encoder_queue_maxsize, encoder_threads)
+            streaming_enc = cls._build_streaming_encoder(
+                fps, camera_encoder, encoder_queue_maxsize, encoder_threads
+            )
         obj.writer = DatasetWriter(
             meta=obj.meta,
             root=obj.root,
-            vcodec=vcodec,
+            camera_encoder=camera_encoder,
             encoder_threads=encoder_threads,
             batch_encoding_size=batch_encoding_size,
             streaming_encoder=streaming_enc,
@@ -729,12 +748,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
         force_cache_sync: bool = False,
         video_backend: str | None = None,
         batch_encoding_size: int = 1,
-        vcodec: str = "libsvtav1",
+        camera_encoder: VideoEncoderConfig | None = None,
+        encoder_threads: int | None = None,
         image_writer_processes: int = 0,
         image_writer_threads: int = 0,
         streaming_encoding: bool = False,
         encoder_queue_maxsize: int = 30,
-        encoder_threads: int | None = None,
     ) -> "LeRobotDataset":
         """Resume recording on an existing dataset.
 
@@ -757,13 +776,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
             video_backend: Video decoding backend for reading back data.
             batch_encoding_size: Number of episodes to accumulate before
                 batch-encoding videos.
-            vcodec: Video codec for encoding.
+            camera_encoder: Video encoder settings for cameras (codec, quality, etc.).
+                When ``None``, :func:`~lerobot.configs.video.camera_encoder_defaults` is used.
+            encoder_threads: Number of encoder threads (global). ``None``
+                lets the codec decide.
             image_writer_processes: Subprocesses for async image writing.
             image_writer_threads: Threads for async image writing.
             streaming_encoding: If ``True``, encode video in real-time during
                 capture.
             encoder_queue_maxsize: Max buffered frames per camera for streaming.
-            encoder_threads: Threads per encoder instance. ``None`` for auto.
 
         Returns:
             A :class:`LeRobotDataset` in write mode, ready to append episodes.
@@ -774,7 +795,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 "Writing into the revision-safe Hub snapshot cache (used when root=None) would corrupt "
                 "the shared cache. Please provide a local directory path."
             )
-        vcodec = resolve_vcodec(vcodec)
         obj = cls.__new__(cls)
         obj.repo_id = repo_id
         obj._requested_root = Path(root)
@@ -783,11 +803,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_transforms = None
         obj.delta_timestamps = None
         obj.episodes = None
-        obj._video_backend = video_backend if video_backend else get_safe_default_codec()
+        obj._video_backend = video_backend if video_backend else get_safe_default_video_backend()
         obj._return_uint8 = False
         obj._batch_encoding_size = batch_encoding_size
-        obj._vcodec = vcodec
-        obj._encoder_threads = encoder_threads
 
         if obj._requested_root is not None:
             obj._requested_root.mkdir(exist_ok=True, parents=True)
@@ -796,21 +814,22 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.meta = LeRobotDatasetMetadata(
             obj.repo_id, obj._requested_root, obj.revision, force_cache_sync=force_cache_sync
         )
+
+        obj._encoder_threads = encoder_threads
         obj.root = obj.meta.root
 
         # Reader is lazily created on first access (write-only mode)
         obj.reader = None
 
-        # Create writer for appending
         streaming_enc = None
         if streaming_encoding and len(obj.meta.video_keys) > 0:
             streaming_enc = cls._build_streaming_encoder(
-                obj.meta.fps, vcodec, encoder_queue_maxsize, encoder_threads
+                obj.meta.fps, camera_encoder, encoder_queue_maxsize, encoder_threads
             )
         obj.writer = DatasetWriter(
             meta=obj.meta,
             root=obj.root,
-            vcodec=vcodec,
+            camera_encoder=camera_encoder,
             encoder_threads=encoder_threads,
             batch_encoding_size=batch_encoding_size,
             streaming_encoder=streaming_enc,

src/lerobot/datasets/pyav_utils.py

@@ -0,0 +1,174 @@
+#!/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.
+"""PyAV-based compatibility checks for :class:`VideoEncoderConfig`.
+
+Centralises all :mod:`av` introspection of the bundled FFmpeg build.
+Checks degrade to a no-op when the target codec isn't available locally.
+"""
+
+import functools
+import logging
+from typing import Any
+
+import av
+
+logger = logging.getLogger(__name__)
+
+FFMPEG_NUMERIC_OPTION_TYPES = ("INT", "INT64", "UINT64", "FLOAT", "DOUBLE")
+FFMPEG_INTEGER_OPTION_TYPES = ("INT", "INT64", "UINT64")
+
+
+@functools.cache
+def get_codec(vcodec: str) -> av.codec.Codec | None:
+    """PyAV write-mode ``Codec`` for *vcodec*, or ``None`` if unavailable."""
+    try:
+        return av.codec.Codec(vcodec, "w")
+    except Exception:
+        return None
+
+
+@functools.cache
+def _get_codec_options_by_name(vcodec: str) -> dict[str, av.option.Option]:
+    """Private-option name → PyAV ``Option`` for *vcodec* (empty if unavailable)."""
+    codec = get_codec(vcodec)
+    if codec is None:
+        return {}
+    return {opt.name: opt for opt in codec.descriptor.options}
+
+
+@functools.cache
+def _get_codec_video_formats(vcodec: str) -> tuple[str, ...]:
+    """Pixel formats accepted by *vcodec* in PyAV's preferred order (empty if unknown)."""
+    codec = get_codec(vcodec)
+    if codec is None:
+        return ()
+    return tuple(fmt.name for fmt in (codec.video_formats or []))
+
+
+def detect_available_encoders_pyav(encoders: list[str] | str) -> list[str]:
+    """Return the subset of *encoders* available as video encoders in the local FFmpeg build.
+
+    Each name is probed directly via :func:`get_codec`; input order is preserved.
+    """
+    if isinstance(encoders, str):
+        encoders = [encoders]
+
+    available: list[str] = []
+    for name in encoders:
+        codec = get_codec(name)
+        if codec is not None and codec.type == "video":
+            available.append(name)
+        else:
+            logger.debug("encoder '%s' not available as video encoder", name)
+    return available
+
+
+def _check_option_value(vcodec: str, label: str, value: Any, opt: av.option.Option) -> None:
+    """Range-check numeric *value* and choice-check string *value* against *opt*."""
+    type_name = opt.type.name
+    if type_name in FFMPEG_NUMERIC_OPTION_TYPES:
+        if isinstance(value, bool):
+            raise ValueError(
+                f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
+            )
+        elif isinstance(value, str):
+            try:
+                num_val = float(value)
+            except ValueError as e:
+                raise ValueError(
+                    f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
+                ) from e
+        elif isinstance(value, (float, int)):
+            num_val = value
+        else:
+            raise ValueError(
+                f"{label}={value!r} is not numeric; codec {vcodec!r} expects a number for this option."
+            )
+
+        # Check integer type compatibility
+        if type_name in FFMPEG_INTEGER_OPTION_TYPES and not num_val.is_integer():
+            raise ValueError(
+                f"{label}={num_val!r} must be an integer for codec {vcodec!r} "
+                f"(FFmpeg option {opt.name!r} is {type_name}); float values are not allowed."
+            )
+
+        # Check numeric range compatibility
+        lo, hi = float(opt.min), float(opt.max)
+        if lo < hi and not (lo <= num_val <= hi):
+            raise ValueError(
+                f"{label}={num_val} is out of range for codec {vcodec!r}; must be in [{lo}, {hi}]"
+            )
+
+    elif type_name == "STRING":
+        if isinstance(value, bool):
+            raise ValueError(f"{label}={value!r} is not a valid string value for codec {vcodec!r}.")
+        if isinstance(value, str):
+            str_val = value
+        elif isinstance(value, (int, float)):
+            str_val = str(value)
+        else:
+            raise ValueError(f"{label}={value!r} has unsupported type for STRING option on codec {vcodec!r}")
+
+        # Check string choice compatibility
+        choices = [c.name for c in (opt.choices or [])]
+        if choices and str_val not in choices:
+            raise ValueError(
+                f"{label}={str_val!r} is not a supported choice for codec "
+                f"{vcodec!r}; valid choices: {choices}"
+            )
+    else:
+        return
+
+
+def _check_pixel_format(vcodec: str, pix_fmt: str) -> None:
+    formats = _get_codec_video_formats(vcodec)
+    if formats and pix_fmt not in formats:
+        raise ValueError(
+            f"pix_fmt={pix_fmt!r} is not supported by codec {vcodec!r}; "
+            f"supported pixel formats: {list(formats)}"
+        )
+
+
+def _check_codec_options(vcodec: str, codec_options: dict[str, Any]) -> None:
+    """Validate merged encoder options (typed) against the codec's published AVOptions."""
+    supported_options = _get_codec_options_by_name(vcodec)
+    for key, value in codec_options.items():
+        # GOP size is not a codec-specific option, it has to be validated separately.
+        if key == "g":
+            if isinstance(value, bool) or not isinstance(value, int) or value < 1:
+                raise ValueError(f"g={value!r} must be a positive integer for codec {vcodec!r}")
+            continue
+        if key not in supported_options:
+            continue
+        _check_option_value(vcodec, key, value, supported_options[key])
+
+
+def check_video_encoder_parameters_pyav(vcodec: str, pix_fmt: str, codec_options: dict[str, Any]) -> None:
+    """Verify *config* is compatible with the bundled FFmpeg build.
+
+    Checks pixel format, abstract tuning-field compatibility, and each merged
+    encoder option from :meth:`~lerobot.configs.video.VideoEncoderConfig.get_codec_options`
+    against PyAV (including numeric ``extra_options`` present in that dict).
+    No-op when ``config.vcodec`` isn't in the local FFmpeg build.
+
+    Raises:
+        ValueError: on the first incompatibility encountered.
+    """
+    options = _get_codec_options_by_name(vcodec)
+    if not options:
+        raise ValueError(f"Codec {vcodec!r} is not available in the bundled FFmpeg build")
+    _check_pixel_format(vcodec, pix_fmt)
+    _check_codec_options(vcodec, codec_options)

src/lerobot/datasets/utils.py

@@ -88,7 +88,6 @@ 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"
@@ -130,6 +129,9 @@ class DatasetInfo:
     # Optional metadata
     robot_type: str | None = None
     splits: dict[str, str] = field(default_factory=dict)
+    # OpenAI-style tool schemas declared by the dataset. ``None`` means the
+    # dataset doesn't declare any — readers fall back to ``DEFAULT_TOOLS``.
+    tools: list[dict] | None = None
 
     def __post_init__(self) -> None:
         # Coerce feature shapes from list to tuple — JSON deserialisation
@@ -151,11 +153,15 @@ class DatasetInfo:
         """Return a JSON-serialisable dict.
 
         Converts tuple shapes back to lists so ``json.dump`` can handle them.
+        Drops ``tools`` when unset so existing datasets keep a clean
+        ``info.json``.
         """
         d = dataclasses.asdict(self)
         for ft in d["features"].values():
             if isinstance(ft.get("shape"), tuple):
                 ft["shape"] = list(ft["shape"])
+        if d.get("tools") is None:
+            d.pop("tools", None)
         return d
 
     @classmethod

src/lerobot/datasets/video_utils.py

@@ -22,7 +22,7 @@ import shutil
 import tempfile
 import threading
 import warnings
-from dataclasses import dataclass, field
+from dataclasses import asdict, dataclass, field
 from fractions import Fraction
 from pathlib import Path
 from threading import Lock
@@ -33,90 +33,17 @@ import fsspec
 import numpy as np
 import pyarrow as pa
 import torch
-import torchvision
 from datasets.features.features import register_feature
 from PIL import Image
 
-from lerobot.utils.import_utils import get_safe_default_codec
+from lerobot.configs import (
+    VideoEncoderConfig,
+    camera_encoder_defaults,
+)
+from lerobot.utils.import_utils import get_safe_default_video_backend
 
 logger = logging.getLogger(__name__)
 
-# List of hardware encoders to probe for auto-selection. Availability depends on the platform and FFmpeg build.
-# Determines the order of preference for auto-selection when vcodec="auto" is used.
-HW_ENCODERS = [
-    "h264_videotoolbox",  # macOS
-    "hevc_videotoolbox",  # macOS
-    "h264_nvenc",  # NVIDIA GPU
-    "hevc_nvenc",  # NVIDIA GPU
-    "h264_vaapi",  # Linux Intel/AMD
-    "h264_qsv",  # Intel Quick Sync
-]
-
-VALID_VIDEO_CODECS = {"h264", "hevc", "libsvtav1", "auto"} | set(HW_ENCODERS)
-
-
-def _get_codec_options(
-    vcodec: str,
-    g: int | None = 2,
-    crf: int | None = 30,
-    preset: int | None = None,
-) -> dict:
-    """Build codec-specific options dict for video encoding."""
-    options = {}
-
-    # GOP size (keyframe interval) - supported by VideoToolbox and software encoders
-    if g is not None and (vcodec in ("h264_videotoolbox", "hevc_videotoolbox") or vcodec not in HW_ENCODERS):
-        options["g"] = str(g)
-
-    # Quality control (codec-specific parameter names)
-    if crf is not None:
-        if vcodec in ("h264", "hevc", "libsvtav1"):
-            options["crf"] = str(crf)
-        elif vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
-            quality = max(1, min(100, int(100 - crf * 2)))
-            options["q:v"] = str(quality)
-        elif vcodec in ("h264_nvenc", "hevc_nvenc"):
-            options["rc"] = "constqp"
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_vaapi",):
-            options["qp"] = str(crf)
-        elif vcodec in ("h264_qsv",):
-            options["global_quality"] = str(crf)
-
-    # Preset (only for libsvtav1)
-    if vcodec == "libsvtav1":
-        options["preset"] = str(preset) if preset is not None else "12"
-
-    return options
-
-
-def detect_available_hw_encoders() -> list[str]:
-    """Probe PyAV/FFmpeg for available hardware video encoders."""
-    available = []
-    for codec_name in HW_ENCODERS:
-        try:
-            av.codec.Codec(codec_name, "w")
-            available.append(codec_name)
-        except Exception:  # nosec B110
-            logger.debug("HW encoder '%s' not available", codec_name)  # nosec B110
-    return available
-
-
-def resolve_vcodec(vcodec: str) -> str:
-    """Validate vcodec and resolve 'auto' to best available HW encoder, fallback to libsvtav1."""
-    if vcodec not in VALID_VIDEO_CODECS:
-        raise ValueError(f"Invalid vcodec '{vcodec}'. Must be one of: {sorted(VALID_VIDEO_CODECS)}")
-    if vcodec != "auto":
-        logger.info(f"Using video codec: {vcodec}")
-        return vcodec
-    available = detect_available_hw_encoders()
-    for encoder in HW_ENCODERS:
-        if encoder in available:
-            logger.info(f"Auto-selected video codec: {encoder}")
-            return encoder
-    logger.info("No hardware encoder available, falling back to software encoder 'libsvtav1'")
-    return "libsvtav1"
-
 
 def decode_video_frames(
     video_path: Path | str,
@@ -132,7 +59,9 @@ def decode_video_frames(
         video_path (Path): Path to the video file.
         timestamps (list[float]): List of timestamps to extract frames.
         tolerance_s (float): Allowed deviation in seconds for frame retrieval.
-        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available in the platform; otherwise, defaults to "pyav".
+        backend (str, optional): Backend to use for decoding. Defaults to "torchcodec" when available
+            in the platform; otherwise, defaults to "pyav". The legacy value "video_reader" is
+            accepted for one release as an alias for "pyav" and will be removed in a future version.
         return_uint8 (bool): If True, return raw uint8 frames without float32 normalization.
             This reduces memory for DataLoader IPC; normalization can be done on GPU afterward.
 
@@ -142,88 +71,90 @@ def decode_video_frames(
     Currently supports torchcodec on cpu and pyav.
     """
     if backend is None:
-        backend = get_safe_default_codec()
+        backend = get_safe_default_video_backend()
     if backend == "torchcodec":
         return decode_video_frames_torchcodec(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
-    elif backend in ["pyav", "video_reader"]:
-        return decode_video_frames_torchvision(
-            video_path, timestamps, tolerance_s, backend, return_uint8=return_uint8
-        )
+    elif backend == "pyav":
+        return decode_video_frames_pyav(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
+    elif backend == "video_reader":
+        logger.warning("backend='video_reader' is deprecated and now aliases to 'pyav'.")
+        return decode_video_frames_pyav(video_path, timestamps, tolerance_s, return_uint8=return_uint8)
     else:
         raise ValueError(f"Unsupported video backend: {backend}")
 
 
-def decode_video_frames_torchvision(
+def decode_video_frames_pyav(
     video_path: Path | str,
     timestamps: list[float],
     tolerance_s: float,
-    backend: str = "pyav",
     log_loaded_timestamps: bool = False,
     return_uint8: bool = False,
 ) -> torch.Tensor:
-    """Loads frames associated to the requested timestamps of a video
+    """Loads frames associated to the requested timestamps of a video using PyAV.
 
-    The backend can be either "pyav" (default) or "video_reader".
-    "video_reader" requires installing torchvision from source, see:
-    https://github.com/pytorch/vision/blob/main/torchvision/csrc/io/decoder/gpu/README.rst
-    (note that you need to compile against ffmpeg<4.3)
+    This is the fallback decoder for platforms where torchcodec has no wheel (currently macOS
+    x86_64 and linux armv7l — see the torchcodec block in pyproject.toml for the full matrix).
+    On supported platforms, prefer `decode_video_frames_torchcodec`, which is faster and supports
+    accurate seek.
 
-    While both use cpu, "video_reader" is supposedly faster than "pyav" but requires additional setup.
-    For more info on video decoding, see `benchmark/video/README.md`
+    PyAV doesn't support accurate seek: we seek to the nearest preceding keyframe and decode
+    forward until we have covered the requested timestamp range. The number of key frames in a
+    video can be adjusted at encoding time to trade off decoding speed against file size.
 
-    See torchvision doc for more info on these two backends:
-    https://pytorch.org/vision/0.18/index.html?highlight=backend#torchvision.set_video_backend
+    Args:
+        video_path: Path to the video file.
+        timestamps: List of timestamps (in seconds) to extract frames for.
+        tolerance_s: Allowed deviation in seconds between a queried timestamp and the closest
+            decoded frame.
+        log_loaded_timestamps: When True, log every decoded frame's timestamp at INFO level.
+        return_uint8: When True, return raw uint8 frames (C, H, W). Otherwise, return float32 in
+            [0, 1] range.
 
-    Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
-    the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
-    that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
-    and all subsequent frames until reaching the requested frame. The number of key frames in a video
-    can be adjusted during encoding to take into account decoding time and video size in bytes.
+    Returns:
+        torch.Tensor of shape (len(timestamps), C, H, W).
     """
-    video_path = str(video_path)
-
-    # set backend
-    keyframes_only = False
-    torchvision.set_video_backend(backend)
-    if backend == "pyav":
-        keyframes_only = True  # pyav doesn't support accurate seek
-
-    # set a video stream reader
     # TODO(rcadene): also load audio stream at the same time
-    reader = torchvision.io.VideoReader(video_path, "video")
+    video_path = str(video_path)
 
     # set the first and last requested timestamps
     # Note: previous timestamps are usually loaded, since we need to access the previous key frame
     first_ts = min(timestamps)
     last_ts = max(timestamps)
 
-    # access closest key frame of the first requested frame
-    # Note: closest key frame timestamp is usually smaller than `first_ts` (e.g. key frame can be the first frame of the video)
-    # for details on what `seek` is doing see: https://pyav.basswood-io.com/docs/stable/api/container.html?highlight=inputcontainer#av.container.InputContainer.seek
-    reader.seek(first_ts, keyframes_only=keyframes_only)
+    loaded_frames: list[torch.Tensor] = []
+    loaded_ts: list[float] = []
 
-    # load all frames until last requested frame
-    loaded_frames = []
-    loaded_ts = []
-    for frame in reader:
-        current_ts = frame["pts"]
-        if log_loaded_timestamps:
-            logger.info(f"frame loaded at timestamp={current_ts:.4f}")
-        loaded_frames.append(frame["data"])
-        loaded_ts.append(current_ts)
-        if current_ts >= last_ts:
-            break
+    # Seek + decode. `container.seek(offset)` with no `stream` argument expects the offset in
+    # av.time_base units (microseconds). `backward=True` lands us on the nearest keyframe at or
+    # before `first_ts`, so we can then decode forward until we cover `last_ts`. See:
+    # https://pyav.basswood-io.com/docs/stable/api/container.html#av.container.InputContainer.seek
+    with av.open(video_path) as container:
+        stream = container.streams.video[0]
+        container.seek(int(first_ts * av.time_base), backward=True)
 
-    if backend == "pyav":
-        reader.container.close()
+        for frame in container.decode(stream):
+            if frame.pts is None:
+                continue
+            current_ts = float(frame.pts * stream.time_base)
+            if log_loaded_timestamps:
+                logger.info(f"frame loaded at timestamp={current_ts:.4f}")
+            # Convert to CHW uint8 to match torchcodec's output layout.
+            arr = frame.to_ndarray(format="rgb24")  # H, W, 3
+            loaded_frames.append(torch.from_numpy(arr).permute(2, 0, 1).contiguous())
+            loaded_ts.append(current_ts)
+            if current_ts >= last_ts:
+                break
 
-    reader = None
+    if not loaded_frames:
+        raise FrameTimestampError(
+            f"No frames could be decoded from {video_path} in the timestamp range [{first_ts}, {last_ts}]."
+        )
 
     query_ts = torch.tensor(timestamps)
-    loaded_ts = torch.tensor(loaded_ts)
+    loaded_ts_t = torch.tensor(loaded_ts)
 
     # compute distances between each query timestamp and timestamps of all loaded frames
-    dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
+    dist = torch.cdist(query_ts[:, None], loaded_ts_t[:, None], p=1)
     min_, argmin_ = dist.min(1)
 
     is_within_tol = min_ < tolerance_s
@@ -234,14 +165,14 @@ def decode_video_frames_torchvision(
             " This might be due to synchronization issues with timestamps during data collection."
             " To be safe, we advise to ignore this item during training."
             f"\nqueried timestamps: {query_ts}"
-            f"\nloaded timestamps: {loaded_ts}"
+            f"\nloaded timestamps: {loaded_ts_t}"
             f"\nvideo: {video_path}"
-            f"\nbackend: {backend}"
+            f"\nbackend: pyav"
         )
 
     # get closest frames to the query timestamps
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
-    closest_ts = loaded_ts[argmin_]
+    closest_ts = loaded_ts_t[argmin_]
 
     if log_loaded_timestamps:
         logger.info(f"{closest_ts=}")
@@ -282,7 +213,11 @@ class VideoDecoderCache:
         with self._lock:
             if video_path not in self._cache:
                 file_handle = fsspec.open(video_path).__enter__()
-                decoder = VideoDecoder(file_handle, seek_mode="approximate")
+                try:
+                    decoder = VideoDecoder(file_handle, seek_mode="approximate")
+                except Exception:
+                    file_handle.close()
+                    raise
                 self._cache[video_path] = (decoder, file_handle)
 
             return self._cache[video_path][0]
@@ -400,18 +335,17 @@ def encode_video_frames(
     imgs_dir: Path | str,
     video_path: Path | str,
     fps: int,
-    vcodec: str = "libsvtav1",
-    pix_fmt: str = "yuv420p",
-    g: int | None = 2,
-    crf: int | None = 30,
-    fast_decode: int = 0,
+    camera_encoder: VideoEncoderConfig | None = None,
+    encoder_threads: int | None = None,
+    *,
     log_level: int | None = av.logging.WARNING,
     overwrite: bool = False,
-    preset: int | None = None,
-    encoder_threads: int | None = None,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    vcodec = resolve_vcodec(vcodec)
+    if camera_encoder is None:
+        camera_encoder = camera_encoder_defaults()
+    vcodec = camera_encoder.vcodec
+    pix_fmt = camera_encoder.pix_fmt
 
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
@@ -422,42 +356,18 @@ def encode_video_frames(
 
     video_path.parent.mkdir(parents=True, exist_ok=True)
 
-    # Encoders/pixel formats incompatibility check
-    if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
-        logger.warning(
-            f"Incompatible pixel format 'yuv444p' for codec {vcodec}, auto-selecting format 'yuv420p'"
-        )
-        pix_fmt = "yuv420p"
-
     # Get input frames
     template = "frame-" + ("[0-9]" * 6) + ".png"
     input_list = sorted(
         glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("-")[-1].split(".")[0])
     )
 
-    # Define video output frame size (assuming all input frames are the same size)
     if len(input_list) == 0:
         raise FileNotFoundError(f"No images found in {imgs_dir}.")
     with Image.open(input_list[0]) as dummy_image:
         width, height = dummy_image.size
 
-    # Define video codec options
-    video_options = _get_codec_options(vcodec, g, crf, preset)
-
-    if fast_decode:
-        key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
-        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
-        video_options[key] = value
-
-    if encoder_threads is not None:
-        if vcodec == "libsvtav1":
-            lp_param = f"lp={encoder_threads}"
-            if "svtav1-params" in video_options:
-                video_options["svtav1-params"] += f":{lp_param}"
-            else:
-                video_options["svtav1-params"] = lp_param
-        else:
-            video_options["threads"] = str(encoder_threads)
+    video_options = camera_encoder.get_codec_options(encoder_threads, as_strings=True)
 
     # Set logging level
     if log_level is not None:
@@ -493,8 +403,97 @@ def encode_video_frames(
         raise OSError(f"Video encoding did not work. File not found: {video_path}.")
 
 
+def reencode_video(
+    input_video_path: Path | str,
+    output_video_path: Path | str,
+    camera_encoder: VideoEncoderConfig | None = None,
+    encoder_threads: int | None = None,
+    log_level: int | None = av.logging.WARNING,
+    overwrite: bool = False,
+) -> None:
+    """Re-encode a video file using the given encoder configuration.
+
+    Args:
+        input_video_path: Existing video file to read.
+        output_video_path: Path for the re-encoded file.
+        camera_encoder: Encoder configuration. Defaults to :func:`camera_encoder_defaults`.
+        encoder_threads: Optional thread count forwarded to :meth:`VideoEncoderConfig.get_codec_options`.
+        log_level: libav log level while encoding, or ``None`` to leave logging unchanged. Defaults to WARNING.
+        overwrite: When ``False`` and ``output_video_path`` already exists, skip and log a warning.
+    """
+
+    camera_encoder = camera_encoder or camera_encoder_defaults()
+
+    output_video_path = Path(output_video_path)
+
+    if output_video_path.exists() and not overwrite:
+        logger.warning(f"Video file already exists: {output_video_path}. Skipping re-encode.")
+        return
+
+    output_video_path.parent.mkdir(parents=True, exist_ok=True)
+
+    video_options = camera_encoder.get_codec_options(encoder_threads, as_strings=True)
+    vcodec = camera_encoder.vcodec
+    pix_fmt = camera_encoder.pix_fmt
+
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmp_named_file:
+        tmp_output_video_path = tmp_named_file.name
+
+    if log_level is not None:
+        logging.getLogger("libav").setLevel(log_level)
+
+    try:
+        with av.open(input_video_path, mode="r") as src:
+            try:
+                in_stream = src.streams.video[0]
+            except IndexError as e:
+                raise ValueError(f"No video stream in {input_video_path}") from e
+
+            fps = (
+                in_stream.base_rate
+            )  # We allow fractional fps though LeRobotDataset only supports integer fps
+            width = int(in_stream.width)
+            height = int(in_stream.height)
+
+            with av.open(
+                tmp_output_video_path,
+                mode="w",
+                options={
+                    "movflags": "faststart"
+                },  # faststart is to move the metadata to the beginning of the file to speed up loading
+            ) as dst:
+                out_stream = dst.add_stream(vcodec, fps, options=video_options)
+                out_stream.pix_fmt = pix_fmt
+                out_stream.width = width
+                out_stream.height = height
+
+                for frame in src.decode(in_stream):
+                    frame = frame.reformat(width=width, height=height, format=pix_fmt)
+                    packet = out_stream.encode(frame)
+                    if packet:
+                        dst.mux(packet)
+
+                packet = out_stream.encode()
+                if packet:
+                    dst.mux(packet)
+
+        shutil.move(tmp_output_video_path, output_video_path)
+    except Exception:
+        Path(tmp_output_video_path).unlink(missing_ok=True)
+        raise
+    finally:
+        if log_level is not None:
+            av.logging.restore_default_callback()
+
+    if not output_video_path.exists():
+        raise OSError(f"Video re-encoding did not work. File not found: {output_video_path}.")
+
+
 def concatenate_video_files(
-    input_video_paths: list[Path | str], output_video_path: Path, overwrite: bool = True
+    input_video_paths: list[Path | str],
+    output_video_path: Path,
+    overwrite: bool = True,
+    compatibility_check: bool = False,
 ):
     """
     Concatenate multiple video files into a single video file using pyav.
@@ -507,6 +506,7 @@ def concatenate_video_files(
         input_video_paths: Ordered list of input video file paths to concatenate.
         output_video_path: Path to the output video file.
         overwrite: Whether to overwrite the output video file if it already exists. Default is True.
+        compatibility_check: Whether to check if the input videos are compatible. Default is False.
 
     Note:
         - Creates a temporary directory for intermediate files that is cleaned up after use.
@@ -525,6 +525,22 @@ def concatenate_video_files(
     if len(input_video_paths) == 0:
         raise FileNotFoundError("No input video paths provided.")
 
+    # This check may be skipped at recording time as videos are encoded with the same encoder config.
+    if compatibility_check:
+        reference_video_info = get_video_info(input_video_paths[0])
+        for input_path in input_video_paths[1:]:
+            video_info = get_video_info(input_path)
+            if (
+                video_info["video.height"] != reference_video_info["video.height"]
+                or video_info["video.width"] != reference_video_info["video.width"]
+                or video_info["video.fps"] != reference_video_info["video.fps"]
+                or video_info["video.codec"] != reference_video_info["video.codec"]
+                or video_info["video.pix_fmt"] != reference_video_info["video.pix_fmt"]
+            ):
+                raise ValueError(
+                    f"Input video {input_path} is not compatible with the reference video {input_video_paths[0]}."
+                )
+
     # Create a temporary .ffconcat file to list the input video paths
     with tempfile.NamedTemporaryFile(mode="w", suffix=".ffconcat", delete=False) as tmp_concatenate_file:
         tmp_concatenate_file.write("ffconcat version 1.0\n")
@@ -591,26 +607,20 @@ class _CameraEncoderThread(threading.Thread):
         fps: int,
         vcodec: str,
         pix_fmt: str,
-        g: int | None,
-        crf: int | None,
-        preset: int | None,
+        codec_options: dict[str, str],
         frame_queue: queue.Queue,
         result_queue: queue.Queue,
         stop_event: threading.Event,
-        encoder_threads: int | None = None,
     ):
         super().__init__(daemon=True)
         self.video_path = video_path
         self.fps = fps
         self.vcodec = vcodec
         self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
+        self.codec_options = codec_options
         self.frame_queue = frame_queue
         self.result_queue = result_queue
         self.stop_event = stop_event
-        self.encoder_threads = encoder_threads
 
     def run(self) -> None:
         from .compute_stats import RunningQuantileStats, auto_downsample_height_width
@@ -646,19 +656,9 @@ class _CameraEncoderThread(threading.Thread):
                 # Open container on first frame (to get width/height)
                 if container is None:
                     height, width = frame_data.shape[:2]
-                    video_options = _get_codec_options(self.vcodec, self.g, self.crf, self.preset)
-                    if self.encoder_threads is not None:
-                        if self.vcodec == "libsvtav1":
-                            lp_param = f"lp={self.encoder_threads}"
-                            if "svtav1-params" in video_options:
-                                video_options["svtav1-params"] += f":{lp_param}"
-                            else:
-                                video_options["svtav1-params"] = lp_param
-                        else:
-                            video_options["threads"] = str(self.encoder_threads)
                     Path(self.video_path).parent.mkdir(parents=True, exist_ok=True)
                     container = av.open(str(self.video_path), "w")
-                    output_stream = container.add_stream(self.vcodec, self.fps, options=video_options)
+                    output_stream = container.add_stream(self.vcodec, self.fps, options=self.codec_options)
                     output_stream.pix_fmt = self.pix_fmt
                     output_stream.width = width
                     output_stream.height = height
@@ -724,22 +724,24 @@ class StreamingVideoEncoder:
     def __init__(
         self,
         fps: int,
-        vcodec: str = "libsvtav1",
-        pix_fmt: str = "yuv420p",
-        g: int | None = 2,
-        crf: int | None = 30,
-        preset: int | None = None,
+        camera_encoder: VideoEncoderConfig | None = None,
         queue_maxsize: int = 30,
         encoder_threads: int | None = None,
     ):
+        """
+        Args:
+            fps: Frames per second for the output videos.
+            camera_encoder: Video encoder settings applied to all cameras.
+                When ``None``, :func:`camera_encoder_defaults` is used.
+            encoder_threads: Number of encoder threads (global setting).
+                ``None`` lets the codec decide.
+            queue_maxsize: Max frames to buffer per camera before
+                back-pressure drops frames.
+        """
         self.fps = fps
-        self.vcodec = resolve_vcodec(vcodec)
-        self.pix_fmt = pix_fmt
-        self.g = g
-        self.crf = crf
-        self.preset = preset
+        self._camera_encoder = camera_encoder or camera_encoder_defaults()
+        self._encoder_threads = encoder_threads
         self.queue_maxsize = queue_maxsize
-        self.encoder_threads = encoder_threads
 
         self._frame_queues: dict[str, queue.Queue] = {}
         self._result_queues: dict[str, queue.Queue] = {}
@@ -770,18 +772,17 @@ class StreamingVideoEncoder:
             temp_video_dir = Path(tempfile.mkdtemp(dir=temp_dir))
             video_path = temp_video_dir / f"{video_key.replace('/', '_')}_streaming.mp4"
 
+            vcodec = self._camera_encoder.vcodec
+            codec_options = self._camera_encoder.get_codec_options(self._encoder_threads, as_strings=True)
             encoder_thread = _CameraEncoderThread(
                 video_path=video_path,
                 fps=self.fps,
-                vcodec=self.vcodec,
-                pix_fmt=self.pix_fmt,
-                g=self.g,
-                crf=self.crf,
-                preset=self.preset,
+                vcodec=vcodec,
+                pix_fmt=self._camera_encoder.pix_fmt,
+                codec_options=codec_options,
                 frame_queue=frame_queue,
                 result_queue=result_queue,
                 stop_event=stop_event,
-                encoder_threads=self.encoder_threads,
             )
             encoder_thread.start()
 
@@ -986,8 +987,18 @@ def get_audio_info(video_path: Path | str) -> dict:
     return audio_info
 
 
-def get_video_info(video_path: Path | str) -> dict:
-    # Set logging level
+def get_video_info(
+    video_path: Path | str,
+    camera_encoder: VideoEncoderConfig | None = None,
+) -> dict:
+    """Build the ``video.*`` / ``audio.*`` info dict persisted in ``info.json``.
+
+    Args:
+        video_path: Path to the encoded video file to probe.
+        camera_encoder: If provided, record the exact encoder settings used to encode this
+            video. Stream-derived values take precedence — encoder fields are only written for keys
+            not already populated from the video file itself.
+    """
     logging.getLogger("libav").setLevel(av.logging.WARNING)
 
     # Getting video stream information
@@ -1018,6 +1029,14 @@ def get_video_info(video_path: Path | str) -> dict:
     # Adding audio stream information
     video_info.update(**get_audio_info(video_path))
 
+    # Add additional encoder configuration if provided
+    if camera_encoder is not None:
+        for field_name, field_value in asdict(camera_encoder).items():
+            # vcodec is already populated from the video stream
+            if field_name == "vcodec":
+                continue
+            video_info.setdefault(f"video.{field_name}", field_value)
+
     return video_info
 
 

src/lerobot/policies/__init__.py

@@ -18,13 +18,13 @@ from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .eo1.configuration_eo1 import EO1Config as EO1Config
 from .factory import get_policy_class, make_policy, make_policy_config, make_pre_post_processors
+from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig as GaussianActorConfig
 from .groot.configuration_groot import GrootConfig as GrootConfig
 from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig as MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
 from .pi05.configuration_pi05 import PI05Config as PI05Config
 from .pretrained import PreTrainedPolicy as PreTrainedPolicy
-from .sac.configuration_sac import SACConfig as SACConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .utils import make_robot_action, prepare_observation_for_inference
@@ -32,21 +32,21 @@ from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
 from .wall_x.configuration_wall_x import WallXConfig as WallXConfig
 from .xvla.configuration_xvla import XVLAConfig as XVLAConfig
 
-# NOTE: Policy modeling classes (e.g., SACPolicy) are intentionally NOT re-exported here.
+# NOTE: Policy modeling classes (e.g., GaussianActorPolicy) are intentionally NOT re-exported here.
 # They have heavy optional dependencies and are loaded lazily via get_policy_class().
-# Import directly: ``from lerobot.policies.sac.modeling_sac import SACPolicy``
+# Import directly: ``from lerobot.policies.gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy``
 
 __all__ = [
     # Configuration classes
     "ACTConfig",
     "DiffusionConfig",
+    "EO1Config",
+    "GaussianActorConfig",
     "GrootConfig",
     "MultiTaskDiTConfig",
-    "EO1Config",
     "PI0Config",
     "PI0FastConfig",
     "PI05Config",
-    "SACConfig",
     "SmolVLAConfig",
     "TDMPCConfig",
     "VQBeTConfig",

src/lerobot/policies/eo1/modeling_eo1.py

@@ -28,11 +28,12 @@ import torch.nn.functional as F  # noqa: N812
 import torch.utils.checkpoint
 from torch import Tensor
 
-from lerobot.policies.eo1.configuration_eo1 import EO1Config
-from lerobot.policies.pretrained import PreTrainedPolicy
 from lerobot.utils.constants import ACTION, OBS_STATE
 from lerobot.utils.import_utils import _transformers_available, require_package
 
+from ..pretrained import PreTrainedPolicy
+from .configuration_eo1 import EO1Config
+
 if TYPE_CHECKING or _transformers_available:
     from transformers.activations import ACT2FN
     from transformers.models.qwen2_5_vl import Qwen2_5_VLForConditionalGeneration

src/lerobot/policies/eo1/processor_eo1.py

@@ -22,7 +22,6 @@ from typing import TYPE_CHECKING, Any
 import torch
 
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
-from lerobot.policies.eo1.configuration_eo1 import EO1Config
 from lerobot.processor import (
     AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
@@ -44,6 +43,8 @@ from lerobot.utils.constants import (
 )
 from lerobot.utils.import_utils import _transformers_available, require_package
 
+from .configuration_eo1 import EO1Config
+
 if TYPE_CHECKING or _transformers_available:
     from transformers.models.qwen2_5_vl import Qwen2_5_VLProcessor
 else:

src/lerobot/policies/factory.py

@@ -47,12 +47,12 @@ from lerobot.utils.feature_utils import dataset_to_policy_features
 from .act.configuration_act import ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig
 from .eo1.configuration_eo1 import EO1Config
+from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig
 from .groot.configuration_groot import GrootConfig
 from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
 from .pi0.configuration_pi0 import PI0Config
 from .pi05.configuration_pi05 import PI05Config
 from .pretrained import PreTrainedPolicy
-from .sac.configuration_sac import SACConfig
 from .smolvla.configuration_smolvla import SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig
 from .utils import validate_visual_features_consistency
@@ -88,7 +88,7 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
 
     Args:
         name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-            "multi_task_dit", "vqbet", "pi0", "pi05", "sac", "smolvla", "wall_x".
+            "multi_task_dit", "vqbet", "pi0", "pi05", "gaussian_actor", "smolvla", "wall_x".
     Returns:
         The policy class corresponding to the given name.
 
@@ -127,10 +127,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from .pi05.modeling_pi05 import PI05Policy
 
         return PI05Policy
-    elif name == "sac":
-        from .sac.modeling_sac import SACPolicy
+    elif name == "gaussian_actor":
+        from .gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
 
-        return SACPolicy
+        return GaussianActorPolicy
     elif name == "smolvla":
         from .smolvla.modeling_smolvla import SmolVLAPolicy
 
@@ -167,7 +167,7 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
 
     Args:
         policy_type: The type of the policy. Supported types include "tdmpc",
-                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "sac",
+                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "gaussian_actor",
                      "smolvla", "wall_x".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
@@ -191,8 +191,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return PI0Config(**kwargs)
     elif policy_type == "pi05":
         return PI05Config(**kwargs)
-    elif policy_type == "sac":
-        return SACConfig(**kwargs)
+    elif policy_type == "gaussian_actor":
+        return GaussianActorConfig(**kwargs)
     elif policy_type == "smolvla":
         return SmolVLAConfig(**kwargs)
     elif policy_type == "groot":
@@ -365,10 +365,10 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
-    elif isinstance(policy_cfg, SACConfig):
-        from .sac.processor_sac import make_sac_pre_post_processors
+    elif isinstance(policy_cfg, GaussianActorConfig):
+        from .gaussian_actor.processor_gaussian_actor import make_gaussian_actor_pre_post_processors
 
-        processors = make_sac_pre_post_processors(
+        processors = make_gaussian_actor_pre_post_processors(
             config=policy_cfg,
             dataset_stats=kwargs.get("dataset_stats"),
         )

src/lerobot/policies/gaussian_actor/__init__.py

@@ -12,8 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .configuration_sac import SACConfig
-from .modeling_sac import SACPolicy
-from .processor_sac import make_sac_pre_post_processors
+from .configuration_gaussian_actor import GaussianActorConfig
+from .modeling_gaussian_actor import GaussianActorPolicy
+from .processor_gaussian_actor import make_gaussian_actor_pre_post_processors
 
-__all__ = ["SACConfig", "SACPolicy", "make_sac_pre_post_processors"]
+__all__ = ["GaussianActorConfig", "GaussianActorPolicy", "make_gaussian_actor_pre_post_processors"]

src/lerobot/policies/gaussian_actor/configuration_gaussian_actor.py

@@ -1,4 +1,4 @@
-# !/usr/bin/env python
+#!/usr/bin/env python
 
 # Copyright 2025 The HuggingFace Inc. team.
 # All rights reserved.
@@ -75,18 +75,19 @@ class PolicyConfig:
     init_final: float = 0.05
 
 
-@PreTrainedConfig.register_subclass("sac")
+@PreTrainedConfig.register_subclass("gaussian_actor")
 @dataclass
-class SACConfig(PreTrainedConfig):
-    """Soft Actor-Critic (SAC) configuration.
+class GaussianActorConfig(PreTrainedConfig):
+    """Gaussian actor configuration.
 
-    SAC is an off-policy actor-critic deep RL algorithm based on the maximum entropy
-    reinforcement learning framework. It learns a policy and a Q-function simultaneously
-    using experience collected from the environment.
+    This configures the policy-side (actor + observation encoder) of a Gaussian
+    policy, as used by SAC and related maximum-entropy continuous-control algorithms.
+    By default the actor output is a tanh-squashed diagonal Gaussian
+    (``TanhMultivariateNormalDiag``); the tanh squashing can be disabled via
+    ``policy_kwargs.use_tanh_squash``. The critics, temperature, and Bellman-update
+    logic live on the algorithm side (see ``lerobot.rl.algorithms.sac``).
 
-    This configuration class contains all the parameters needed to define a SAC agent,
-    including network architectures, optimization settings, and algorithm-specific
-    hyperparameters.
+    CLI: ``--policy.type=gaussian_actor``.
     """
 
     # Mapping of feature types to normalization modes
@@ -122,7 +123,7 @@ class SACConfig(PreTrainedConfig):
     device: str = "cpu"
     # Device to store the model on
     storage_device: str = "cpu"
-    # Name of the vision encoder model (Set to "helper2424/resnet10" for hil serl resnet10)
+    # Name of the vision encoder model (Set to "lerobot/resnet10" for hil serl resnet10)
     vision_encoder_name: str | None = None
     # Whether to freeze the vision encoder during training
     freeze_vision_encoder: bool = True
@@ -135,7 +136,13 @@ class SACConfig(PreTrainedConfig):
     # Dimension of the image embedding pooling
     image_embedding_pooling_dim: int = 8
 
-    # Training parameter
+    # Encoder architecture
+    # Hidden dimension size for the state encoder
+    state_encoder_hidden_dim: int = 256
+    # Dimension of the latent space
+    latent_dim: int = 256
+
+    # Online training (TODO(Khalil): relocate to TrainRLServerPipelineConfig)
     # Number of steps for online training
     online_steps: int = 1000000
     # Capacity of the online replay buffer
@@ -146,67 +153,38 @@ class SACConfig(PreTrainedConfig):
     async_prefetch: bool = False
     # Number of steps before learning starts
     online_step_before_learning: int = 100
-    # Frequency of policy updates
-    policy_update_freq: int = 1
 
-    # SAC algorithm parameters
-    # Discount factor for the SAC algorithm
-    discount: float = 0.99
-    # Initial temperature value
-    temperature_init: float = 1.0
-    # Number of critics in the ensemble
-    num_critics: int = 2
-    # Number of subsampled critics for training
-    num_subsample_critics: int | None = None
-    # Learning rate for the critic network
-    critic_lr: float = 3e-4
-    # Learning rate for the actor network
-    actor_lr: float = 3e-4
-    # Learning rate for the temperature parameter
-    temperature_lr: float = 3e-4
-    # Weight for the critic target update
-    critic_target_update_weight: float = 0.005
-    # Update-to-data ratio for the UTD algorithm (If you want enable utd_ratio, you need to set it to >1)
-    utd_ratio: int = 1
-    # Hidden dimension size for the state encoder
-    state_encoder_hidden_dim: int = 256
-    # Dimension of the latent space
-    latent_dim: int = 256
-    # Target entropy for the SAC algorithm
-    target_entropy: float | None = None
-    # Whether to use backup entropy for the SAC algorithm
-    use_backup_entropy: bool = True
-    # Gradient clipping norm for the SAC algorithm
-    grad_clip_norm: float = 40.0
-
-    # Network configuration
-    # Configuration for the critic network architecture
-    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    # Configuration for the actor network architecture
-    actor_network_kwargs: ActorNetworkConfig = field(default_factory=ActorNetworkConfig)
-    # Configuration for the policy parameters
-    policy_kwargs: PolicyConfig = field(default_factory=PolicyConfig)
-    # Configuration for the discrete critic network
-    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    # Actor-learner transport (TODO(Khalil): relocate to TrainRLServerPipelineConfig).
     # Configuration for actor-learner architecture
     actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
     # Configuration for concurrency settings (you can use threads or processes for the actor and learner)
     concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
 
-    # Optimizations
-    use_torch_compile: bool = True
+    # Network architecture
+    # Configuration for the actor network architecture
+    actor_network_kwargs: ActorNetworkConfig = field(default_factory=ActorNetworkConfig)
+    # Configuration for the policy parameters (Gaussian head)
+    policy_kwargs: PolicyConfig = field(default_factory=PolicyConfig)
+    # Configuration for the discrete critic network
+    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
 
     def __post_init__(self):
         super().__post_init__()
-        # Any validation specific to SAC configuration
+        # Any validation specific to GaussianActor configuration
 
     def get_optimizer_preset(self) -> MultiAdamConfig:
+        # Default learning rate used to satisfy the abstract ``get_optimizer_preset()``
+        # contract from ``PreTrainedConfig``. The actual optimizers used during RL
+        # training are built by ``SACAlgorithm.make_optimizers_and_scheduler()`` from
+        # ``SACAlgorithmConfig.{actor_lr,critic_lr,temperature_lr}`` and fully bypass
+        # this preset.
+        default_lr = 3e-4
         return MultiAdamConfig(
             weight_decay=0.0,
             optimizer_groups={
-                "actor": {"lr": self.actor_lr},
-                "critic": {"lr": self.critic_lr},
-                "temperature": {"lr": self.temperature_lr},
+                "actor": {"lr": default_lr},
+                "critic": {"lr": default_lr},
+                "temperature": {"lr": default_lr},
             },
         )
 

src/lerobot/policies/gaussian_actor/modeling_gaussian_actor.py

@@ -15,16 +15,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import math
 from collections.abc import Callable
 from dataclasses import asdict
-from typing import Literal
 
-import einops
-import numpy as np
 import torch
 import torch.nn as nn
-import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 from torch.distributions import MultivariateNormal, TanhTransform, Transform, TransformedDistribution
 
@@ -32,20 +27,20 @@ from lerobot.utils.constants import ACTION, OBS_ENV_STATE, OBS_STATE
 
 from ..pretrained import PreTrainedPolicy
 from ..utils import get_device_from_parameters
-from .configuration_sac import SACConfig, is_image_feature
+from .configuration_gaussian_actor import GaussianActorConfig, is_image_feature
 
 DISCRETE_DIMENSION_INDEX = -1  # Gripper is always the last dimension
 
 
-class SACPolicy(
+class GaussianActorPolicy(
     PreTrainedPolicy,
 ):
-    config_class = SACConfig
-    name = "sac"
+    config_class = GaussianActorConfig
+    name = "gaussian_actor"
 
     def __init__(
         self,
-        config: SACConfig | None = None,
+        config: GaussianActorConfig | None = None,
     ):
         super().__init__(config)
         config.validate_features()
@@ -54,9 +49,8 @@ class SACPolicy(
         # Determine action dimension and initialize all components
         continuous_action_dim = config.output_features[ACTION].shape[0]
         self._init_encoders()
-        self._init_critics(continuous_action_dim)
         self._init_actor(continuous_action_dim)
-        self._init_temperature()
+        self._init_discrete_critic()
 
     def get_optim_params(self) -> dict:
         optim_params = {
@@ -65,11 +59,7 @@ class SACPolicy(
                 for n, p in self.actor.named_parameters()
                 if not n.startswith("encoder") or not self.shared_encoder
             ],
-            "critic": self.critic_ensemble.parameters(),
-            "temperature": self.log_alpha,
         }
-        if self.config.num_discrete_actions is not None:
-            optim_params["discrete_critic"] = self.discrete_critic.parameters()
         return optim_params
 
     def reset(self):
@@ -79,7 +69,9 @@ class SACPolicy(
     @torch.no_grad()
     def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
         """Predict a chunk of actions given environment observations."""
-        raise NotImplementedError("SACPolicy does not support action chunking. It returns single actions!")
+        raise NotImplementedError(
+            "GaussianActorPolicy does not support action chunking. It returns single actions!"
+        )
 
     @torch.no_grad()
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
@@ -92,360 +84,43 @@ class SACPolicy(
         actions, _, _ = self.actor(batch, observations_features)
 
         if self.config.num_discrete_actions is not None:
-            discrete_action_value = self.discrete_critic(batch, observations_features)
-            discrete_action = torch.argmax(discrete_action_value, dim=-1, keepdim=True)
+            if self.discrete_critic is not None:
+                discrete_action_value = self.discrete_critic(batch, observations_features)
+                discrete_action = torch.argmax(discrete_action_value, dim=-1, keepdim=True)
+            else:
+                discrete_action = torch.ones(
+                    (*actions.shape[:-1], 1), device=actions.device, dtype=actions.dtype
+                )
             actions = torch.cat([actions, discrete_action], dim=-1)
 
         return actions
 
-    def critic_forward(
-        self,
-        observations: dict[str, Tensor],
-        actions: Tensor,
-        use_target: bool = False,
-        observation_features: Tensor | None = None,
-    ) -> Tensor:
-        """Forward pass through a critic network ensemble
+    def forward(self, batch: dict[str, Tensor | dict[str, Tensor]]) -> dict[str, Tensor]:
+        """Actor forward pass: sample actions and return log-probabilities.
 
         Args:
-            observations: Dictionary of observations
-            actions: Action tensor
-            use_target: If True, use target critics, otherwise use ensemble critics
+            batch: A flat observation dict, or a training dict containing
+                ``"state"`` (observations) and optionally ``"observation_feature"``
+                (pre-computed encoder features).
 
         Returns:
-            Tensor of Q-values from all critics
+            Dict with ``"action"``, ``"log_prob"``, and ``"action_mean"`` tensors.
         """
-
-        critics = self.critic_target if use_target else self.critic_ensemble
-        q_values = critics(observations, actions, observation_features)
-        return q_values
-
-    def discrete_critic_forward(
-        self, observations, use_target=False, observation_features=None
-    ) -> torch.Tensor:
-        """Forward pass through a discrete critic network
-
-        Args:
-            observations: Dictionary of observations
-            use_target: If True, use target critics, otherwise use ensemble critics
-            observation_features: Optional pre-computed observation features to avoid recomputing encoder output
-
-        Returns:
-            Tensor of Q-values from the discrete critic network
-        """
-        discrete_critic = self.discrete_critic_target if use_target else self.discrete_critic
-        q_values = discrete_critic(observations, observation_features)
-        return q_values
-
-    def forward(
-        self,
-        batch: dict[str, Tensor | dict[str, Tensor]],
-        model: Literal["actor", "critic", "temperature", "discrete_critic"] = "critic",
-    ) -> dict[str, Tensor]:
-        """Compute the loss for the given model
-
-        Args:
-            batch: Dictionary containing:
-                - action: Action tensor
-                - reward: Reward tensor
-                - state: Observations tensor dict
-                - next_state: Next observations tensor dict
-                - done: Done mask tensor
-                - observation_feature: Optional pre-computed observation features
-                - next_observation_feature: Optional pre-computed next observation features
-            model: Which model to compute the loss for ("actor", "critic", "discrete_critic", or "temperature")
-
-        Returns:
-            The computed loss tensor
-        """
-        # Extract common components from batch
-        actions: Tensor = batch[ACTION]
-        observations: dict[str, Tensor] = batch["state"]
-        observation_features: Tensor = batch.get("observation_feature")
-
-        if model == "critic":
-            # Extract critic-specific components
-            rewards: Tensor = batch["reward"]
-            next_observations: dict[str, Tensor] = batch["next_state"]
-            done: Tensor = batch["done"]
-            next_observation_features: Tensor = batch.get("next_observation_feature")
-
-            loss_critic = self.compute_loss_critic(
-                observations=observations,
-                actions=actions,
-                rewards=rewards,
-                next_observations=next_observations,
-                done=done,
-                observation_features=observation_features,
-                next_observation_features=next_observation_features,
-            )
-
-            return {"loss_critic": loss_critic}
-
-        if model == "discrete_critic" and self.config.num_discrete_actions is not None:
-            # Extract critic-specific components
-            rewards: Tensor = batch["reward"]
-            next_observations: dict[str, Tensor] = batch["next_state"]
-            done: Tensor = batch["done"]
-            next_observation_features: Tensor = batch.get("next_observation_feature")
-            complementary_info = batch.get("complementary_info")
-            loss_discrete_critic = self.compute_loss_discrete_critic(
-                observations=observations,
-                actions=actions,
-                rewards=rewards,
-                next_observations=next_observations,
-                done=done,
-                observation_features=observation_features,
-                next_observation_features=next_observation_features,
-                complementary_info=complementary_info,
-            )
-            return {"loss_discrete_critic": loss_discrete_critic}
-        if model == "actor":
-            return {
-                "loss_actor": self.compute_loss_actor(
-                    observations=observations,
-                    observation_features=observation_features,
-                )
-            }
-
-        if model == "temperature":
-            return {
-                "loss_temperature": self.compute_loss_temperature(
-                    observations=observations,
-                    observation_features=observation_features,
-                )
-            }
-
-        raise ValueError(f"Unknown model type: {model}")
-
-    def update_target_networks(self):
-        """Update target networks with exponential moving average"""
-        for target_param, param in zip(
-            self.critic_target.parameters(),
-            self.critic_ensemble.parameters(),
-            strict=True,
-        ):
-            target_param.data.copy_(
-                param.data * self.config.critic_target_update_weight
-                + target_param.data * (1.0 - self.config.critic_target_update_weight)
-            )
-        if self.config.num_discrete_actions is not None:
-            for target_param, param in zip(
-                self.discrete_critic_target.parameters(),
-                self.discrete_critic.parameters(),
-                strict=True,
-            ):
-                target_param.data.copy_(
-                    param.data * self.config.critic_target_update_weight
-                    + 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 compute_loss_critic(
-        self,
-        observations,
-        actions,
-        rewards,
-        next_observations,
-        done,
-        observation_features: Tensor | None = None,
-        next_observation_features: Tensor | None = None,
-    ) -> Tensor:
-        with torch.no_grad():
-            next_action_preds, next_log_probs, _ = self.actor(next_observations, next_observation_features)
-
-            # 2- compute q targets
-            q_targets = self.critic_forward(
-                observations=next_observations,
-                actions=next_action_preds,
-                use_target=True,
-                observation_features=next_observation_features,
-            )
-
-            # subsample critics to prevent overfitting if use high UTD (update to date)
-            # TODO: Get indices before forward pass to avoid unnecessary computation
-            if self.config.num_subsample_critics is not None:
-                indices = torch.randperm(self.config.num_critics)
-                indices = indices[: self.config.num_subsample_critics]
-                q_targets = q_targets[indices]
-
-            # critics subsample size
-            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
-            if self.config.use_backup_entropy:
-                min_q = min_q - (self.temperature * next_log_probs)
-
-            td_target = rewards + (1 - done) * self.config.discount * min_q
-
-        # 3- compute predicted qs
-        if self.config.num_discrete_actions is not None:
-            # NOTE: We only want to keep the continuous action part
-            # In the buffer we have the full action space (continuous + discrete)
-            # We need to split them before concatenating them in the critic forward
-            actions: Tensor = actions[:, :DISCRETE_DIMENSION_INDEX]
-        q_preds = self.critic_forward(
-            observations=observations,
-            actions=actions,
-            use_target=False,
-            observation_features=observation_features,
-        )
-
-        # 4- Calculate loss
-        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
-        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
-        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
-        critics_loss = (
-            F.mse_loss(
-                input=q_preds,
-                target=td_target_duplicate,
-                reduction="none",
-            ).mean(dim=1)
-        ).sum()
-        return critics_loss
-
-    def compute_loss_discrete_critic(
-        self,
-        observations,
-        actions,
-        rewards,
-        next_observations,
-        done,
-        observation_features=None,
-        next_observation_features=None,
-        complementary_info=None,
-    ):
-        # NOTE: We only want to keep the discrete action part
-        # In the buffer we have the full action space (continuous + discrete)
-        # We need to split them before concatenating them in the critic forward
-        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
-        actions_discrete = torch.round(actions_discrete)
-        actions_discrete = actions_discrete.long()
-
-        discrete_penalties: Tensor | None = None
-        if complementary_info is not None:
-            discrete_penalties: Tensor | None = complementary_info.get("discrete_penalty")
-
-        with torch.no_grad():
-            # For DQN, select actions using online network, evaluate with target network
-            next_discrete_qs = self.discrete_critic_forward(
-                next_observations, use_target=False, observation_features=next_observation_features
-            )
-            best_next_discrete_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
-
-            # Get target Q-values from target network
-            target_next_discrete_qs = self.discrete_critic_forward(
-                observations=next_observations,
-                use_target=True,
-                observation_features=next_observation_features,
-            )
-
-            # Use gather to select Q-values for best actions
-            target_next_discrete_q = torch.gather(
-                target_next_discrete_qs, dim=1, index=best_next_discrete_action
-            ).squeeze(-1)
-
-            # Compute target Q-value with Bellman equation
-            rewards_discrete = rewards
-            if discrete_penalties is not None:
-                rewards_discrete = rewards + discrete_penalties
-            target_discrete_q = rewards_discrete + (1 - done) * self.config.discount * target_next_discrete_q
-
-        # Get predicted Q-values for current observations
-        predicted_discrete_qs = self.discrete_critic_forward(
-            observations=observations, use_target=False, observation_features=observation_features
-        )
-
-        # Use gather to select Q-values for taken actions
-        predicted_discrete_q = torch.gather(predicted_discrete_qs, dim=1, index=actions_discrete).squeeze(-1)
-
-        # Compute MSE loss between predicted and target Q-values
-        discrete_critic_loss = F.mse_loss(input=predicted_discrete_q, target=target_discrete_q)
-        return discrete_critic_loss
-
-    def compute_loss_temperature(self, observations, observation_features: Tensor | None = None) -> Tensor:
-        """Compute the temperature loss"""
-        # calculate temperature loss
-        with torch.no_grad():
-            _, log_probs, _ = self.actor(observations, observation_features)
-        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
-        return temperature_loss
-
-    def compute_loss_actor(
-        self,
-        observations,
-        observation_features: Tensor | None = None,
-    ) -> Tensor:
-        actions_pi, log_probs, _ = self.actor(observations, observation_features)
-
-        q_preds = self.critic_forward(
-            observations=observations,
-            actions=actions_pi,
-            use_target=False,
-            observation_features=observation_features,
-        )
-        min_q_preds = q_preds.min(dim=0)[0]
-
-        actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
-        return actor_loss
+        observations = batch.get("state", batch)
+        observation_features = batch.get("observation_feature") if isinstance(batch, dict) else None
+        actions, log_probs, means = self.actor(observations, observation_features)
+        return {"action": actions, "log_prob": log_probs, "action_mean": means}
 
     def _init_encoders(self):
         """Initialize shared or separate encoders for actor and critic."""
         self.shared_encoder = self.config.shared_encoder
-        self.encoder_critic = SACObservationEncoder(self.config)
+        self.encoder_critic = GaussianActorObservationEncoder(self.config)
         self.encoder_actor = (
-            self.encoder_critic if self.shared_encoder else SACObservationEncoder(self.config)
+            self.encoder_critic if self.shared_encoder else GaussianActorObservationEncoder(self.config)
         )
 
-    def _init_critics(self, continuous_action_dim):
-        """Build critic ensemble, targets, and optional discrete critic."""
-        heads = [
-            CriticHead(
-                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
-                **asdict(self.config.critic_network_kwargs),
-            )
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_ensemble = CriticEnsemble(encoder=self.encoder_critic, ensemble=heads)
-        target_heads = [
-            CriticHead(
-                input_dim=self.encoder_critic.output_dim + continuous_action_dim,
-                **asdict(self.config.critic_network_kwargs),
-            )
-            for _ in range(self.config.num_critics)
-        ]
-        self.critic_target = CriticEnsemble(encoder=self.encoder_critic, ensemble=target_heads)
-        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
-
-        if self.config.use_torch_compile:
-            self.critic_ensemble = torch.compile(self.critic_ensemble)
-            self.critic_target = torch.compile(self.critic_target)
-
-        if self.config.num_discrete_actions is not None:
-            self._init_discrete_critics()
-
-    def _init_discrete_critics(self):
-        """Build discrete discrete critic ensemble and target networks."""
-        self.discrete_critic = DiscreteCritic(
-            encoder=self.encoder_critic,
-            input_dim=self.encoder_critic.output_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-        self.discrete_critic_target = DiscreteCritic(
-            encoder=self.encoder_critic,
-            input_dim=self.encoder_critic.output_dim,
-            output_dim=self.config.num_discrete_actions,
-            **asdict(self.config.discrete_critic_network_kwargs),
-        )
-
-        # TODO: (maractingi, azouitine) Compile the discrete critic
-        self.discrete_critic_target.load_state_dict(self.discrete_critic.state_dict())
-
     def _init_actor(self, continuous_action_dim):
-        """Initialize policy actor network and default target entropy."""
+        """Initialize policy actor network."""
         # NOTE: The actor select only the continuous action part
         self.actor = Policy(
             encoder=self.encoder_actor,
@@ -455,21 +130,25 @@ class SACPolicy(
             **asdict(self.config.policy_kwargs),
         )
 
-        self.target_entropy = self.config.target_entropy
-        if self.target_entropy is None:
-            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_discrete_critic(self) -> None:
+        """Initialize discrete critic network."""
+        if self.config.num_discrete_actions is None:
+            self.discrete_critic = None
+            return
 
-    def _init_temperature(self) -> None:
-        """Set up temperature parameter (log_alpha)."""
-        temp_init = self.config.temperature_init
-        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
+        # TODO(Khalil): Compile the discrete critic
+        self.discrete_critic = DiscreteCritic(
+            encoder=self.encoder_critic,
+            input_dim=self.encoder_critic.output_dim,
+            output_dim=self.config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )
 
 
-class SACObservationEncoder(nn.Module):
+class GaussianActorObservationEncoder(nn.Module):
     """Encode image and/or state vector observations."""
 
-    def __init__(self, config: SACConfig) -> None:
+    def __init__(self, config: GaussianActorConfig) -> None:
         super().__init__()
         self.config = config
         self._init_image_layers()
@@ -677,84 +356,6 @@ class MLP(nn.Module):
         return self.net(x)
 
 
-class CriticHead(nn.Module):
-    def __init__(
-        self,
-        input_dim: int,
-        hidden_dims: list[int],
-        activations: Callable[[torch.Tensor], torch.Tensor] | str = nn.SiLU(),
-        activate_final: bool = False,
-        dropout_rate: float | None = None,
-        init_final: float | None = None,
-        final_activation: Callable[[torch.Tensor], torch.Tensor] | str | None = None,
-    ):
-        super().__init__()
-        self.net = MLP(
-            input_dim=input_dim,
-            hidden_dims=hidden_dims,
-            activations=activations,
-            activate_final=activate_final,
-            dropout_rate=dropout_rate,
-            final_activation=final_activation,
-        )
-        self.output_layer = nn.Linear(in_features=hidden_dims[-1], out_features=1)
-        if init_final is not None:
-            nn.init.uniform_(self.output_layer.weight, -init_final, init_final)
-            nn.init.uniform_(self.output_layer.bias, -init_final, init_final)
-        else:
-            orthogonal_init()(self.output_layer.weight)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        return self.output_layer(self.net(x))
-
-
-class CriticEnsemble(nn.Module):
-    """
-    CriticEnsemble wraps multiple CriticHead modules into an ensemble.
-
-    Args:
-        encoder (SACObservationEncoder): encoder for observations.
-        ensemble (List[CriticHead]): list of critic heads.
-        init_final (float | None): optional initializer scale for final layers.
-
-    Forward returns a tensor of shape (num_critics, batch_size) containing Q-values.
-    """
-
-    def __init__(
-        self,
-        encoder: SACObservationEncoder,
-        ensemble: list[CriticHead],
-        init_final: float | None = None,
-    ):
-        super().__init__()
-        self.encoder = encoder
-        self.init_final = init_final
-        self.critics = nn.ModuleList(ensemble)
-
-    def forward(
-        self,
-        observations: dict[str, torch.Tensor],
-        actions: torch.Tensor,
-        observation_features: torch.Tensor | None = None,
-    ) -> torch.Tensor:
-        device = get_device_from_parameters(self)
-        # Move each tensor in observations to device
-        observations = {k: v.to(device) for k, v in observations.items()}
-
-        obs_enc = self.encoder(observations, cache=observation_features)
-
-        inputs = torch.cat([obs_enc, actions], dim=-1)
-
-        # Loop through critics and collect outputs
-        q_values = []
-        for critic in self.critics:
-            q_values.append(critic(inputs))
-
-        # Stack outputs to match expected shape [num_critics, batch_size]
-        q_values = torch.stack([q.squeeze(-1) for q in q_values], dim=0)
-        return q_values
-
-
 class DiscreteCritic(nn.Module):
     def __init__(
         self,
@@ -800,7 +401,7 @@ class DiscreteCritic(nn.Module):
 class Policy(nn.Module):
     def __init__(
         self,
-        encoder: SACObservationEncoder,
+        encoder: GaussianActorObservationEncoder,
         network: nn.Module,
         action_dim: int,
         std_min: float = -5,
@@ -811,7 +412,7 @@ class Policy(nn.Module):
         encoder_is_shared: bool = False,
     ):
         super().__init__()
-        self.encoder: SACObservationEncoder = encoder
+        self.encoder: GaussianActorObservationEncoder = encoder
         self.network = network
         self.action_dim = action_dim
         self.std_min = std_min
@@ -885,7 +486,7 @@ class Policy(nn.Module):
 
 
 class DefaultImageEncoder(nn.Module):
-    def __init__(self, config: SACConfig):
+    def __init__(self, config: GaussianActorConfig):
         super().__init__()
         image_key = next(key for key in config.input_features if is_image_feature(key))
         self.image_enc_layers = nn.Sequential(
@@ -931,12 +532,12 @@ def freeze_image_encoder(image_encoder: nn.Module):
 
 
 class PretrainedImageEncoder(nn.Module):
-    def __init__(self, config: SACConfig):
+    def __init__(self, config: GaussianActorConfig):
         super().__init__()
 
         self.image_enc_layers, self.image_enc_out_shape = self._load_pretrained_vision_encoder(config)
 
-    def _load_pretrained_vision_encoder(self, config: SACConfig):
+    def _load_pretrained_vision_encoder(self, config: GaussianActorConfig):
         """Set up CNN encoder"""
         from transformers import AutoModel
 

src/lerobot/policies/gaussian_actor/processor_gaussian_actor.py

@@ -32,18 +32,18 @@ from lerobot.processor import (
 )
 from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
-from .configuration_sac import SACConfig
+from .configuration_gaussian_actor import GaussianActorConfig
 
 
-def make_sac_pre_post_processors(
-    config: SACConfig,
+def make_gaussian_actor_pre_post_processors(
+    config: GaussianActorConfig,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
 ) -> tuple[
     PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
     PolicyProcessorPipeline[PolicyAction, PolicyAction],
 ]:
     """
-    Constructs pre-processor and post-processor pipelines for the SAC policy.
+    Constructs pre-processor and post-processor pipelines for the Gaussian actor policy.
 
     The pre-processing pipeline prepares input data for the model by:
     1. Renaming features to match pretrained configurations.
@@ -56,7 +56,7 @@ def make_sac_pre_post_processors(
     2. Unnormalizing the output features to their original scale.
 
     Args:
-        config: The configuration object for the SAC policy.
+        config: The configuration object for the tanh-Gaussian policy.
         dataset_stats: A dictionary of statistics for normalization.
 
     Returns:

src/lerobot/policies/pi05/modeling_pi05.py

@@ -441,13 +441,13 @@ class PaliGemmaWithExpertModel(
         if image.dtype != torch.float32:
             image = image.to(torch.float32)
         image_outputs = self.paligemma.model.get_image_features(image)
-        features = image_outputs.pooler_output * self.paligemma.config.text_config.hidden_size**0.5
+        features = image_outputs.pooler_output
         if features.dtype != out_dtype:
             features = features.to(out_dtype)
         return features
 
     def embed_language_tokens(self, tokens: torch.Tensor):
-        return self.paligemma.model.language_model.embed_tokens(tokens)
+        return self.paligemma.model.language_model.get_input_embeddings()(tokens)
 
     def forward(
         self,
@@ -662,8 +662,7 @@ class PI05Pytorch(nn.Module):  # see openpi `PI0Pytorch`
         # Process language tokens
         def lang_embed_func(tokens):
             lang_emb = self.paligemma_with_expert.embed_language_tokens(tokens)
-            lang_emb_dim = lang_emb.shape[-1]
-            return lang_emb * math.sqrt(lang_emb_dim)
+            return lang_emb
 
         lang_emb = self._apply_checkpoint(lang_embed_func, tokens)
         embs.append(lang_emb)

src/lerobot/processor/__init__.py

@@ -95,6 +95,13 @@ from .relative_action_processor import (
 from .rename_processor import RenameObservationsProcessorStep, rename_stats
 from .tokenizer_processor import ActionTokenizerProcessorStep, TokenizerProcessorStep
 
+# RenderMessagesStep is intentionally NOT re-exported here: it pulls in
+# `lerobot.datasets.language`, which requires the `[dataset]` extra
+# (`datasets`, `pyarrow`). Importing it from the processor package would
+# break every base-install consumer of `lerobot.processor`. Users that
+# need it import directly:
+#   from lerobot.processor.render_messages_processor import RenderMessagesStep
+
 __all__ = [
     "ActionProcessorStep",
     "AddTeleopActionAsComplimentaryDataStep",

src/lerobot/processor/batch_processor.py

@@ -174,6 +174,24 @@ class AddBatchDimensionComplementaryDataStep(ComplementaryDataProcessorStep):
             task_index_value = complementary_data["task_index"]
             if isinstance(task_index_value, Tensor) and task_index_value.dim() == 0:
                 complementary_data["task_index"] = task_index_value.unsqueeze(0)
+
+        complementary_data.pop("language_persistent", None)
+        complementary_data.pop("language_events", None)
+
+        if "messages" in complementary_data:
+            messages = complementary_data["messages"]
+            if isinstance(messages, list) and (not messages or isinstance(messages[0], dict)):
+                complementary_data["messages"] = [messages]
+
+        if "message_streams" in complementary_data:
+            streams = complementary_data["message_streams"]
+            if isinstance(streams, list) and (not streams or isinstance(streams[0], str)):
+                complementary_data["message_streams"] = [streams]
+
+        if "target_message_indices" in complementary_data:
+            indices = complementary_data["target_message_indices"]
+            if isinstance(indices, list) and (not indices or isinstance(indices[0], int)):
+                complementary_data["target_message_indices"] = [indices]
         return complementary_data
 
     def transform_features(

src/lerobot/processor/converters.py

@@ -153,26 +153,30 @@ def from_tensor_to_numpy(x: torch.Tensor | Any) -> np.ndarray | float | int | An
     return x
 
 
+_COMPLEMENTARY_KEYS = (
+    "task",
+    "index",
+    "task_index",
+    "episode_index",
+    "timestamp",
+    "language_persistent",
+    "language_events",
+    "messages",
+    "message_streams",
+    "target_message_indices",
+)
+
+
 def _extract_complementary_data(batch: dict[str, Any]) -> dict[str, Any]:
-    """
-    Extract complementary data from a batch dictionary.
+    """Extract complementary data from a batch dictionary.
 
-    This includes padding flags, task description, and indices.
-
-    Args:
-        batch: The batch dictionary.
-
-    Returns:
-        A dictionary with the extracted complementary data.
+    Includes padding flags (any key containing ``_is_pad``) plus the fixed
+    set of metadata / language keys defined in ``_COMPLEMENTARY_KEYS`` —
+    each only when present in ``batch``.
     """
     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}
+    extras = {k: batch[k] for k in _COMPLEMENTARY_KEYS if k in batch}
+    return {**pad_keys, **extras}
 
 
 def create_transition(

src/lerobot/processor/hil_processor.py

@@ -4,7 +4,6 @@
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with 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
@@ -321,6 +320,7 @@ class GymHILAdapterProcessorStep(ProcessorStep):
     This step normalizes the `transition` object by:
     1. Copying `teleop_action` from `info` to `complementary_data`.
     2. Copying `is_intervention` from `info` (using the string key) to `info` (using the enum key).
+    3. Copying `discrete_penalty` from `info` to `complementary_data`.
     """
 
     def __call__(self, transition: EnvTransition) -> EnvTransition:
@@ -330,6 +330,9 @@ class GymHILAdapterProcessorStep(ProcessorStep):
         if TELEOP_ACTION_KEY in info:
             complementary_data[TELEOP_ACTION_KEY] = info[TELEOP_ACTION_KEY]
 
+        if DISCRETE_PENALTY_KEY in info:
+            complementary_data[DISCRETE_PENALTY_KEY] = info[DISCRETE_PENALTY_KEY]
+
         if "is_intervention" in info:
             info[TeleopEvents.IS_INTERVENTION] = info["is_intervention"]
 
@@ -348,18 +351,24 @@ class GymHILAdapterProcessorStep(ProcessorStep):
 @ProcessorStepRegistry.register("gripper_penalty_processor")
 class GripperPenaltyProcessorStep(ProcessorStep):
     """
-    Applies a penalty for inefficient gripper usage.
+    Applies a small per-transition cost on the discrete gripper action.
 
-    This step penalizes actions that attempt to close an already closed gripper or
-    open an already open one, based on position thresholds.
+    Fires only when the commanded action would actually transition the gripper
+    from one extreme to the other (close-while-open or open-while-closed).
+    This discourages gripper oscillation while leaving "stay" and saturating-further
+    commands unpenalized.
 
     Attributes:
         penalty: The negative reward value to apply.
         max_gripper_pos: The maximum position value for the gripper, used for normalization.
+        open_threshold: Normalized state below which the gripper is considered "open".
+        closed_threshold: Normalized state above which the gripper is considered "closed".
     """
 
-    penalty: float = -0.01
+    penalty: float = -0.02
     max_gripper_pos: float = 30.0
+    open_threshold: float = 0.1
+    closed_threshold: float = 0.9
 
     def __call__(self, transition: EnvTransition) -> EnvTransition:
         """
@@ -379,11 +388,15 @@ class GripperPenaltyProcessorStep(ProcessorStep):
         if raw_joint_positions is None:
             return new_transition
 
-        current_gripper_pos = raw_joint_positions.get(GRIPPER_KEY, None)
+        current_gripper_pos = raw_joint_positions.get(f"{GRIPPER_KEY}.pos", None)
         if current_gripper_pos is None:
             return new_transition
 
-        # Gripper action is a PolicyAction at this stage
+        # During reset, the transition may not carry any action yet.
+        if action is None:
+            return new_transition
+
+        # Gripper action is expected as the last action dimension.
         gripper_action = action[-1].item()
         gripper_action_normalized = gripper_action / self.max_gripper_pos
 
@@ -391,9 +404,13 @@ class GripperPenaltyProcessorStep(ProcessorStep):
         gripper_state_normalized = current_gripper_pos / self.max_gripper_pos
 
         # Calculate penalty boolean as in original
-        gripper_penalty_bool = (gripper_state_normalized < 0.5 and gripper_action_normalized > 0.5) or (
-            gripper_state_normalized > 0.75 and gripper_action_normalized < 0.5
-        )
+        #   - currently open  AND target is closed  -> close transition
+        #   - currently closed AND target is open   -> open transition
+        is_open = gripper_state_normalized < self.open_threshold
+        is_closed = gripper_state_normalized > self.closed_threshold
+        cmd_close = gripper_action_normalized > self.closed_threshold
+        cmd_open = gripper_action_normalized < self.open_threshold
+        gripper_penalty_bool = (is_open and cmd_close) or (is_closed and cmd_open)
 
         gripper_penalty = self.penalty * int(gripper_penalty_bool)
 
@@ -409,11 +426,14 @@ class GripperPenaltyProcessorStep(ProcessorStep):
         Returns the configuration of the step for serialization.
 
         Returns:
-            A dictionary containing the penalty value and max gripper position.
+            A dictionary containing the penalty value, max gripper position,
+            and the open/closed thresholds.
         """
         return {
             "penalty": self.penalty,
             "max_gripper_pos": self.max_gripper_pos,
+            "open_threshold": self.open_threshold,
+            "closed_threshold": self.closed_threshold,
         }
 
     def reset(self) -> None:

src/lerobot/processor/normalize_processor.py

@@ -134,6 +134,24 @@ class _NormalizationMixin:
         if self.dtype is None:
             self.dtype = torch.float32
         self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
+        self._reshape_visual_stats()
+
+    def _reshape_visual_stats(self) -> None:
+        """Reshape flat ``(C,)`` visual stats to ``(C, 1, 1)`` for image broadcasting.
+
+        No-op for stats from :func:`~lerobot.datasets.compute_stats.compute_stats`
+        (already ``(C, 1, 1)``). Needed by RL training, which can start without
+        a dataset and supplies stats manually via JSON config.
+        """
+        for key, feature in self.features.items():
+            if feature.type != FeatureType.VISUAL:
+                continue
+            if key not in self._tensor_stats:
+                continue
+            for stat_name, stat_tensor in self._tensor_stats[key].items():
+                if not isinstance(stat_tensor, Tensor) or stat_tensor.ndim != 1:
+                    continue
+                self._tensor_stats[key][stat_name] = stat_tensor.reshape(-1, 1, 1)
 
     def to(
         self, device: torch.device | str | None = None, dtype: torch.dtype | None = None
@@ -152,6 +170,7 @@ class _NormalizationMixin:
         if dtype is not None:
             self.dtype = dtype
         self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)
+        self._reshape_visual_stats()
         return self
 
     def state_dict(self) -> dict[str, Tensor]:
@@ -201,6 +220,7 @@ class _NormalizationMixin:
             # Don't load from state_dict, keep the explicitly provided stats
             # But ensure _tensor_stats is properly initialized
             self._tensor_stats = to_tensor(self.stats, device=self.device, dtype=self.dtype)  # type: ignore[assignment]
+            self._reshape_visual_stats()
             return
 
         # Normal behavior: load stats from state_dict
@@ -211,6 +231,7 @@ class _NormalizationMixin:
             self._tensor_stats.setdefault(key, {})[stat_name] = tensor.to(
                 dtype=torch.float32, device=self.device
             )
+        self._reshape_visual_stats()
 
         # Reconstruct the original stats dict from tensor stats for compatibility with to() method
         # and other functions that rely on self.stats

src/lerobot/processor/render_messages_processor.py

@@ -0,0 +1,84 @@
+#!/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 __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any
+
+from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.configs.recipe import TrainingRecipe
+from lerobot.datasets.language import LANGUAGE_EVENTS, LANGUAGE_PERSISTENT
+from lerobot.datasets.language_render import render_sample
+from lerobot.types import EnvTransition, TransitionKey
+from lerobot.utils.utils import unwrap_scalar
+
+from .pipeline import ProcessorStep, ProcessorStepRegistry
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="render_messages_processor")
+class RenderMessagesStep(ProcessorStep):
+    """Processor step that turns raw language columns into rendered chat messages.
+
+    Reads ``language_persistent`` and ``language_events`` from the transition's
+    complementary data, renders them through ``recipe`` at the sample timestamp,
+    and replaces the raw columns with the resulting ``messages`` /
+    ``message_streams`` / ``target_message_indices`` keys.
+    """
+
+    recipe: TrainingRecipe
+    dataset_ctx: Any | None = None
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition | None:
+        """Render messages for a single transition; return ``None`` to drop it."""
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+        persistent = complementary_data.get(LANGUAGE_PERSISTENT) or []
+        events = complementary_data.get(LANGUAGE_EVENTS) or []
+
+        if not persistent and not events:
+            return transition
+
+        timestamp = complementary_data.get("timestamp")
+        if timestamp is None:
+            raise KeyError("RenderMessagesStep requires sample timestamp in complementary data.")
+
+        sample_idx = complementary_data.get("index", 0)
+        rendered = render_sample(
+            recipe=self.recipe,
+            persistent=persistent,
+            events=events,
+            t=unwrap_scalar(timestamp),
+            sample_idx=int(unwrap_scalar(sample_idx)),
+            task=complementary_data.get("task"),
+            dataset_ctx=self.dataset_ctx,
+        )
+        if rendered is None:
+            return None
+
+        new_transition = transition.copy()
+        new_complementary_data = dict(complementary_data)
+        new_complementary_data.pop(LANGUAGE_PERSISTENT, None)
+        new_complementary_data.pop(LANGUAGE_EVENTS, None)
+        new_complementary_data.update(rendered)
+        new_transition[TransitionKey.COMPLEMENTARY_DATA] = new_complementary_data
+        return new_transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        """Pass features through unchanged; rendering only touches complementary data."""
+        return features

src/lerobot/rewards/__init__.py

@@ -21,11 +21,13 @@ from .factory import (
 )
 from .pretrained import PreTrainedRewardModel as PreTrainedRewardModel
 from .sarm.configuration_sarm import SARMConfig as SARMConfig
+from .topreward.configuration_topreward import TOPRewardConfig as TOPRewardConfig
 
 __all__ = [
     # Configuration classes
     "RewardClassifierConfig",
     "SARMConfig",
+    "TOPRewardConfig",
     # Base class
     "PreTrainedRewardModel",
     # Factory functions

src/lerobot/rewards/classifier/configuration_classifier.py

@@ -30,7 +30,7 @@ class RewardClassifierConfig(RewardModelConfig):
     latent_dim: int = 256
     image_embedding_pooling_dim: int = 8
     dropout_rate: float = 0.1
-    model_name: str = "helper2424/resnet10"  # TODO: This needs to be updated. The model on the Hub doesn't call self.post_init() in its __init__, which is required by transformers v5 to set all_tied_weights_keys. The from_pretrained call fails when it tries to access this attribute during _finalize_model_loading.
+    model_name: str = "lerobot/resnet10"
     device: str = "cpu"
     model_type: str = "cnn"  # "transformer" or "cnn"
     num_cameras: int = 2

src/lerobot/rewards/classifier/modeling_classifier.py

@@ -17,10 +17,11 @@ import logging
 import torch
 from torch import Tensor, nn
 
-from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
-from lerobot.rewards.pretrained import PreTrainedRewardModel
 from lerobot.utils.constants import OBS_IMAGE, REWARD
 
+from ..pretrained import PreTrainedRewardModel
+from .configuration_classifier import RewardClassifierConfig
+
 
 class ClassifierOutput:
     """Wrapper for classifier outputs with additional metadata."""
@@ -105,6 +106,7 @@ class Classifier(PreTrainedRewardModel):
     def __init__(
         self,
         config: RewardClassifierConfig,
+        **kwargs,
     ):
         from transformers import AutoModel
 

src/lerobot/rewards/classifier/processor_classifier.py

@@ -25,7 +25,8 @@ from lerobot.processor import (
     policy_action_to_transition,
     transition_to_policy_action,
 )
-from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
+
+from .configuration_classifier import RewardClassifierConfig
 
 
 def make_classifier_processor(

src/lerobot/rewards/factory.py

@@ -22,9 +22,11 @@ import torch
 
 from lerobot.configs.rewards import RewardModelConfig
 from lerobot.processor import PolicyAction, PolicyProcessorPipeline
-from lerobot.rewards.classifier.configuration_classifier import RewardClassifierConfig
-from lerobot.rewards.pretrained import PreTrainedRewardModel
-from lerobot.rewards.sarm.configuration_sarm import SARMConfig
+
+from .classifier.configuration_classifier import RewardClassifierConfig
+from .pretrained import PreTrainedRewardModel
+from .sarm.configuration_sarm import SARMConfig
+from .topreward.configuration_topreward import TOPRewardConfig
 
 
 def get_reward_model_class(name: str) -> type[PreTrainedRewardModel]:
@@ -36,7 +38,7 @@ def get_reward_model_class(name: str) -> type[PreTrainedRewardModel]:
 
     Args:
         name: The name of the reward model. Supported names are "reward_classifier",
-              "sarm".
+              "sarm", "topreward".
 
     Returns:
         The reward model class corresponding to the given name.
@@ -52,6 +54,10 @@ def get_reward_model_class(name: str) -> type[PreTrainedRewardModel]:
         from lerobot.rewards.sarm.modeling_sarm import SARMRewardModel
 
         return SARMRewardModel
+    elif name == "topreward":
+        from lerobot.rewards.topreward.modeling_topreward import TOPRewardModel
+
+        return TOPRewardModel
     else:
         try:
             return _get_reward_model_cls_from_name(name=name)
@@ -68,7 +74,7 @@ def make_reward_model_config(reward_type: str, **kwargs) -> RewardModelConfig:
 
     Args:
         reward_type: The type of the reward model. Supported types include
-                     "reward_classifier", "sarm".
+                     "reward_classifier", "sarm", "topreward".
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -81,6 +87,8 @@ def make_reward_model_config(reward_type: str, **kwargs) -> RewardModelConfig:
         return RewardClassifierConfig(**kwargs)
     elif reward_type == "sarm":
         return SARMConfig(**kwargs)
+    elif reward_type == "topreward":
+        return TOPRewardConfig(**kwargs)
     else:
         try:
             config_cls = RewardModelConfig.get_choice_class(reward_type)
@@ -161,6 +169,14 @@ def make_reward_pre_post_processors(
             dataset_meta=kwargs.get("dataset_meta"),
         )
 
+    elif isinstance(reward_cfg, TOPRewardConfig):
+        from lerobot.rewards.topreward.processor_topreward import make_topreward_pre_post_processors
+
+        return make_topreward_pre_post_processors(
+            config=reward_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
     else:
         try:
             processors = _make_processors_from_reward_model_config(

src/lerobot/rewards/sarm/compute_rabc_weights.py

@@ -58,9 +58,10 @@ import torch
 from tqdm import tqdm
 
 from lerobot.datasets import LeRobotDataset
-from lerobot.rewards.sarm.modeling_sarm import SARMRewardModel
-from lerobot.rewards.sarm.processor_sarm import make_sarm_pre_post_processors
-from lerobot.rewards.sarm.sarm_utils import normalize_stage_tau
+
+from .modeling_sarm import SARMRewardModel
+from .processor_sarm import make_sarm_pre_post_processors
+from .sarm_utils import normalize_stage_tau
 
 
 def get_reward_model_path_from_parquet(parquet_path: Path) -> str | None:

src/lerobot/rewards/sarm/modeling_sarm.py

@@ -32,13 +32,14 @@ import torch.nn as nn
 import torch.nn.functional as F  # noqa: N812
 from torch import Tensor
 
-from lerobot.rewards.pretrained import PreTrainedRewardModel
-from lerobot.rewards.sarm.configuration_sarm import SARMConfig
-from lerobot.rewards.sarm.sarm_utils import (
+from lerobot.utils.constants import OBS_STR
+
+from ..pretrained import PreTrainedRewardModel
+from .configuration_sarm import SARMConfig
+from .sarm_utils import (
     normalize_stage_tau,
     pad_state_to_max_dim,
 )
-from lerobot.utils.constants import OBS_STR
 
 
 class StageTransformer(nn.Module):

src/lerobot/rewards/sarm/processor_sarm.py

@@ -58,15 +58,16 @@ from lerobot.processor import (
     policy_action_to_transition,
     transition_to_policy_action,
 )
-from lerobot.rewards.sarm.configuration_sarm import SARMConfig
-from lerobot.rewards.sarm.sarm_utils import (
+from lerobot.types import EnvTransition, PolicyAction, TransitionKey
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+from .configuration_sarm import SARMConfig
+from .sarm_utils import (
     apply_rewind_augmentation,
     compute_absolute_indices,
     find_stage_and_tau,
     pad_state_to_max_dim,
 )
-from lerobot.types import EnvTransition, PolicyAction, TransitionKey
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 
 class SARMEncodingProcessorStep(ProcessorStep):

src/lerobot/rewards/topreward/__init__.py

@@ -0,0 +1,19 @@
+# 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 .configuration_topreward import TOPRewardConfig
+from .modeling_topreward import TOPRewardModel
+from .processor_topreward import make_topreward_pre_post_processors
+
+__all__ = ["TOPRewardConfig", "TOPRewardModel", "make_topreward_pre_post_processors"]

src/lerobot/rewards/topreward/compute_rabc_weights.py

@@ -0,0 +1,353 @@
+#!/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.
+
+"""Compute per-frame TOPReward progress curves for a LeRobot dataset.
+
+For each episode, scores trajectory prefixes of increasing length using
+the TOPReward reward model, min-max normalises the raw log-prob rewards per episode,
+and writes a parquet file with one row per frame.
+
+The parquet uses the same schema as SARM's :mod:`lerobot.rewards.sarm.compute_rabc_weights`.
+
+Usage:
+    # Sparse-dense mode (15 anchors per episode, matches upstream)
+    python -m lerobot.rewards.topreward.compute_rabc_weights \\
+        --dataset-repo-id lerobot/libero_10_image \\
+        --num-samples 15
+
+    # Use a different VLM backbone
+    python -m lerobot.rewards.topreward.compute_rabc_weights \\
+        --dataset-repo-id lerobot/libero_10_image \\
+        --vlm-name Qwen/Qwen3-VL-4B-Instruct
+"""
+
+from __future__ import annotations
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import pyarrow as pa
+import pyarrow.parquet as pq
+import torch
+from tqdm import tqdm
+
+from lerobot.datasets import LeRobotDataset
+from lerobot.rewards.topreward.configuration_topreward import TOPRewardConfig
+from lerobot.rewards.topreward.modeling_topreward import TOPRewardModel
+from lerobot.rewards.topreward.processor_topreward import TOPRewardEncoderProcessorStep
+from lerobot.types import TransitionKey
+
+DEFAULT_OUTPUT_FILENAME = "topreward_progress.parquet"
+
+
+def get_reward_model_path_from_parquet(parquet_path: Path) -> str | None:
+    """Read ``reward_model_path`` from parquet metadata if available."""
+    if not parquet_path.exists():
+        return None
+    try:
+        metadata = pq.read_metadata(parquet_path).schema.to_arrow_schema().metadata
+        if metadata and b"reward_model_path" in metadata:
+            return metadata[b"reward_model_path"].decode()
+    except Exception:  # nosec B110
+        return None
+    return None
+
+
+def _resolve_task(sample: dict[str, Any], default: str) -> str:
+    """Best-effort task extraction from a dataset sample."""
+    task = sample.get("task")
+    if isinstance(task, str) and task:
+        return task
+    return default
+
+
+def normalize_rewards(rewards: list[float] | np.ndarray) -> np.ndarray:
+    """Min-max normalise raw log-prob rewards into ``[0, 1]``."""
+    rewards_arr = np.asarray(rewards, dtype=np.float64)
+    if rewards_arr.size == 0:
+        return rewards_arr.astype(np.float32)
+    if rewards_arr.size == 1:
+        return np.array([1.0], dtype=np.float32)
+    r_min, r_max = rewards_arr.min(), rewards_arr.max()
+    if r_max == r_min:
+        return np.ones_like(rewards_arr, dtype=np.float32)
+    return ((rewards_arr - r_min) / (r_max - r_min)).astype(np.float32)
+
+
+def compute_instruction_rewards_for_prefixes(
+    model: TOPRewardModel,
+    encoder: TOPRewardEncoderProcessorStep,
+    dataset: LeRobotDataset,
+    ep_start: int,
+    num_frames: int,
+    task: str,
+    image_key: str,
+    num_samples: int | None,
+    device: str,
+) -> np.ndarray:
+    """Score an episode via prefix sweep and return a per-frame normalised curve."""
+    if num_samples is None or num_samples >= num_frames:
+        prefix_lengths = np.arange(1, num_frames + 1, dtype=np.int64)
+    else:
+        prefix_lengths = np.unique(np.linspace(1, num_frames, num_samples).round().astype(np.int64))
+
+    episode_frames = torch.stack([dataset[ep_start + i][image_key] for i in range(num_frames)])
+    rewards: list[float] = []
+    for length in prefix_lengths:
+        frames = episode_frames[: int(length)].unsqueeze(0)  # (1, T, C, H, W)
+
+        transition = {
+            TransitionKey.OBSERVATION: {image_key: frames},
+            TransitionKey.COMPLEMENTARY_DATA: {"task": task},
+        }
+        encoded = encoder(transition)
+        obs = encoded[TransitionKey.OBSERVATION]
+        batch = {
+            key: value.to(device) if isinstance(value, torch.Tensor) else value for key, value in obs.items()
+        }
+
+        with torch.no_grad():
+            reward = model.compute_reward(batch)
+        rewards.append(float(reward.item()))
+
+    normalized_rewards = normalize_rewards(rewards)
+
+    if prefix_lengths.shape[0] == num_frames:
+        return normalized_rewards
+
+    return np.interp(
+        np.arange(1, num_frames + 1, dtype=np.float64),
+        prefix_lengths.astype(np.float64),
+        normalized_rewards.astype(np.float64),
+    ).astype(np.float32)
+
+
+def compute_topreward_progress(
+    dataset_repo_id: str,
+    reward_model_path: str | None = None,
+    vlm_name: str | None = None,
+    output_path: str | None = None,
+    device: str = "cuda",
+    num_samples: int | None = None,
+    fps: float | None = None,
+    episodes: list[int] | None = None,
+) -> Path:
+    """Run TOPReward over a dataset and write per-frame progress."""
+    if reward_model_path is not None:
+        logging.info(f"Loading TOPReward config from: {reward_model_path}")
+        model = TOPRewardModel.from_pretrained(reward_model_path)
+        config = model.config
+        config.device = device
+        if vlm_name is not None and vlm_name != config.vlm_name:
+            logging.info(f"Overriding vlm_name from config: {config.vlm_name} -> {vlm_name}")
+            config.vlm_name = vlm_name
+            model = TOPRewardModel(config)
+    else:
+        config_kwargs: dict[str, Any] = {"device": device}
+        if vlm_name is not None:
+            config_kwargs["vlm_name"] = vlm_name
+        if fps is not None:
+            config_kwargs["fps"] = fps
+        config = TOPRewardConfig(**config_kwargs)
+        logging.info(f"Constructing TOPReward with VLM: {config.vlm_name}")
+        model = TOPRewardModel(config)
+
+    model.to(device).eval()
+
+    encoder = TOPRewardEncoderProcessorStep(
+        vlm_name=config.vlm_name,
+        image_key=config.image_key,
+        task_key=config.task_key,
+        default_task=config.default_task,
+        max_frames=None,  # no tail-crop: we control prefix length explicitly
+        fps=config.fps,
+        prompt_prefix=config.prompt_prefix,
+        prompt_suffix_template=config.prompt_suffix_template,
+        add_chat_template=config.add_chat_template,
+        max_length=config.max_input_length,
+    )
+
+    image_key = config.image_key
+
+    logging.info(f"Loading dataset: {dataset_repo_id}")
+    dataset = LeRobotDataset(dataset_repo_id, download_videos=True)
+    logging.info(f"Dataset: {dataset.num_episodes} episodes, {dataset.num_frames} frames")
+
+    episode_indices = list(range(dataset.num_episodes)) if episodes is None else episodes
+    logging.info(f"Processing {len(episode_indices)} episode(s)")
+
+    all_index: list[int] = []
+    all_episode: list[int] = []
+    all_frame: list[int] = []
+    all_progress: list[float] = []
+
+    for episode_idx in tqdm(episode_indices, desc="Episodes"):
+        ep = dataset.meta.episodes[episode_idx]
+        ep_start = int(ep["dataset_from_index"])
+        ep_end = int(ep["dataset_to_index"])
+        num_frames = ep_end - ep_start
+        if num_frames <= 0:
+            continue
+
+        first_sample = dataset[ep_start]
+        task = _resolve_task(first_sample, default=config.default_task or "perform the task")
+
+        per_frame = compute_instruction_rewards_for_prefixes(
+            model=model,
+            encoder=encoder,
+            dataset=dataset,
+            ep_start=ep_start,
+            num_frames=num_frames,
+            task=task,
+            image_key=image_key,
+            num_samples=num_samples,
+            device=device,
+        )
+
+        for local in range(num_frames):
+            all_index.append(ep_start + local)
+            all_episode.append(episode_idx)
+            all_frame.append(local)
+            all_progress.append(float(per_frame[local]))
+
+        if device.startswith("cuda"):
+            torch.cuda.empty_cache()
+
+    table = pa.table(
+        {
+            "index": np.asarray(all_index, dtype=np.int64),
+            "episode_index": np.asarray(all_episode, dtype=np.int64),
+            "frame_index": np.asarray(all_frame, dtype=np.int64),
+            "progress_sparse": np.asarray(all_progress, dtype=np.float32),
+        }
+    )
+
+    schema_metadata: dict[bytes, bytes] = {b"vlm_name": config.vlm_name.encode()}
+    if reward_model_path is not None:
+        schema_metadata[b"reward_model_path"] = reward_model_path.encode()
+    table = table.replace_schema_metadata(schema_metadata)
+
+    out = Path(dataset.root) / DEFAULT_OUTPUT_FILENAME if output_path is None else Path(output_path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    pq.write_table(table, out)
+    logging.info(f"Saved {len(table)} frame values to {out}")
+
+    progress_arr = np.asarray(all_progress, dtype=np.float32)
+    if progress_arr.size:
+        logging.info(
+            f"Progress: mean={float(progress_arr.mean()):.4f}, "
+            f"std={float(progress_arr.std()):.4f}, "
+            f"min={float(progress_arr.min()):.4f}, "
+            f"max={float(progress_arr.max()):.4f}"
+        )
+    return out
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Compute per-frame TOPReward progress curves for RA-BC weighting.",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Examples:
+    # Sparse-dense mode (matches upstream TOPReward num_samples=15)
+    python -m lerobot.rewards.topreward.compute_rabc_weights \\
+        --dataset-repo-id lerobot/libero_10_image \\
+        --num-samples 15
+
+    # Use a smaller VLM
+    python -m lerobot.rewards.topreward.compute_rabc_weights \\
+        --dataset-repo-id lerobot/libero_10_image \\
+        --vlm-name Qwen/Qwen3-VL-4B-Instruct
+        """,
+    )
+    parser.add_argument(
+        "--dataset-repo-id", type=str, required=True, help="HuggingFace dataset repo id or local path."
+    )
+    parser.add_argument(
+        "--reward-model-path", type=str, default=None, help="Optional TOPReward LeRobot config."
+    )
+    parser.add_argument("--vlm-name", type=str, default=None, help="Override the VLM backbone (HF Hub id).")
+    parser.add_argument("--output-path", type=str, default=None, help="Output parquet path.")
+    parser.add_argument("--device", type=str, default="cuda", help="Device to use (default: cuda).")
+    parser.add_argument(
+        "--num-samples",
+        type=int,
+        default=None,
+        help="Anchor prefix samples per episode. None = dense. 15 matches upstream.",
+    )
+    parser.add_argument(
+        "--episodes",
+        type=int,
+        nargs="+",
+        default=None,
+        help="Process only these episode indices (e.g. --episodes 0 or --episodes 0 5 10).",
+    )
+    parser.add_argument("--fps", type=float, default=None, help="Override TOPRewardConfig.fps.")
+    parser.add_argument(
+        "--push-to-hub", action="store_true", help="Upload to the dataset repo on HuggingFace Hub."
+    )
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+
+    output_path = compute_topreward_progress(
+        dataset_repo_id=args.dataset_repo_id,
+        reward_model_path=args.reward_model_path,
+        vlm_name=args.vlm_name,
+        output_path=args.output_path,
+        device=args.device,
+        num_samples=args.num_samples,
+        fps=args.fps,
+        episodes=args.episodes,
+    )
+
+    print(f"\nTOPReward progress saved to: {output_path}")
+
+    if args.push_to_hub:
+        from huggingface_hub import HfApi
+
+        api = HfApi()
+        hub_path = DEFAULT_OUTPUT_FILENAME
+
+        print(f"\nUploading to Hub: {args.dataset_repo_id}/{hub_path}")
+        api.upload_file(
+            path_or_fileobj=str(output_path),
+            path_in_repo=hub_path,
+            repo_id=args.dataset_repo_id,
+            repo_type="dataset",
+        )
+        print(
+            "Successfully uploaded to: "
+            f"https://huggingface.co/datasets/{args.dataset_repo_id}/blob/main/{hub_path}"
+        )
+
+        print("\nTo use in training, add to your config:")
+        print("  use_rabc: true")
+        print(f"  rabc_progress_path: hf://datasets/{args.dataset_repo_id}/{hub_path}")
+        print("  rabc_head_mode: sparse")
+    else:
+        print("\nTo use in training, add to your config:")
+        print("  use_rabc: true")
+        print(f"  rabc_progress_path: {output_path}")
+        print("  rabc_head_mode: sparse")
+
+
+if __name__ == "__main__":
+    main()

src/lerobot/rewards/topreward/configuration_topreward.py

@@ -0,0 +1,146 @@
+# 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 __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from lerobot.configs import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.configs.rewards import RewardModelConfig
+from lerobot.utils.constants import OBS_IMAGES
+
+# Default prompt scaffolding from the upstream TOPReward paper / reference
+# implementation (``QwenClient.compute_instruction_reward``). The prompt
+# scores the terminal ``True`` token in ``f"{instruction} ... True"``
+# given the video.
+DEFAULT_PROMPT_PREFIX = (
+    "The above video shows a robot manipulation trajectory that completes the following task: "
+)
+DEFAULT_PROMPT_SUFFIX_TEMPLATE = (
+    "{instruction} Decide whether the above statement is True or not. The answer is: True"
+)
+
+
+@RewardModelConfig.register_subclass("topreward")
+@dataclass
+class TOPRewardConfig(RewardModelConfig):
+    """Configuration for the TOPReward zero-shot reward model.
+
+    TOPReward is **zero-shot**: it has no learnable parameters of its own.
+    The "model" is a generic vision-language model (default
+    ``Qwen/Qwen3-VL-8B-Instruct``) used with a fixed prompt to extract
+    token log-probabilities as a reward signal. There is therefore no
+    fine-tuned checkpoint to host: ``pretrained_path`` is unused at
+    runtime — the model identity is :attr:`vlm_name` (an HF Hub id).
+
+    Args:
+        vlm_name: Hugging Face Hub id of the underlying VLM. Must be a
+            Qwen3-VL family model (the only client implemented in this
+            LeRobot port).
+        torch_dtype: Torch dtype name passed to the VLM loader
+            (``"auto"``, ``"bfloat16"``, ``"float16"``, ...).
+        attn_implementation: ``transformers`` attention implementation
+            (e.g. ``"flash_attention_2"``, ``"sdpa"``). Defaults to
+            ``None`` so the upstream picks the best available.
+        image_key: Observation key that holds the trajectory frames.
+        task_key: Complementary-data key that holds the task instruction.
+        default_task: Fallback instruction when ``task_key`` is absent.
+        max_frames: Cap on the number of frames fed to the VLM per
+            sample. ``None`` = use all frames.
+        fps: Frames-per-second metadata for the Qwen video processor.
+        prompt_prefix: Text shown to the VLM right after the video and
+            before the suffix template.
+        prompt_suffix_template: Suffix appended after ``prompt_prefix``.
+            Must contain ``{instruction}``; the VLM scores the
+            log-likelihood of the tokens that follow the prefix.
+        add_chat_template: If ``True``, wrap the full prompt with the
+            tokenizer's chat template before tokenisation (matches
+            upstream ``add_chat_template=True``).
+        success_threshold: Optional log-prob threshold. If finite,
+            :meth:`TOPRewardModel.compute_reward` returns
+            ``(reward > success_threshold).float()`` instead of the raw
+            log-prob.
+        max_input_length: Hard limit on the total tokenized input length;
+            samples that exceed it raise a ``ValueError``.
+    """
+
+    # Path to a local LeRobot dir or HF repo that holds a ``config.json``
+    # snapshot of this TOPRewardConfig. The VLM weights themselves are
+    # always identified by ``vlm_name``.
+    pretrained_path: str | None = None
+
+    vlm_name: str = "Qwen/Qwen3-VL-8B-Instruct"
+    torch_dtype: str = "auto"
+    attn_implementation: str | None = None
+
+    image_key: str = OBS_IMAGES + ".top"
+    task_key: str = "task"
+    default_task: str | None = None
+    max_frames: int | None = 16
+    fps: float = 2.0
+
+    prompt_prefix: str = DEFAULT_PROMPT_PREFIX
+    prompt_suffix_template: str = DEFAULT_PROMPT_SUFFIX_TEMPLATE
+    add_chat_template: bool = False
+
+    success_threshold: float = float("-inf")
+    max_input_length: int = 32768
+
+    license: str | None = "mit"  # matches upstream TOPReward
+    tags: list[str] | None = field(
+        default_factory=lambda: ["reward-model", "vision-language", "qwen3-vl", "zero-shot"]
+    )
+
+    input_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    output_features: dict[str, PolicyFeature] = field(default_factory=dict)
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "REWARD": NormalizationMode.IDENTITY,
+        }
+    )
+
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        if self.max_frames is not None and self.max_frames < 1:
+            raise ValueError(f"max_frames must be >= 1, got {self.max_frames}")
+        if self.fps <= 0:
+            raise ValueError(f"fps must be > 0, got {self.fps}")
+        if "{instruction}" not in self.prompt_suffix_template:
+            raise ValueError(
+                "prompt_suffix_template must contain `{instruction}` so the model "
+                "scores the log-likelihood of the task suffix."
+            )
+        if self.max_input_length <= 0:
+            raise ValueError(f"max_input_length must be > 0, got {self.max_input_length}")
+
+        if self.image_key not in self.input_features:
+            self.input_features[self.image_key] = PolicyFeature(shape=(3, 224, 224), type=FeatureType.VISUAL)
+        self.output_features.setdefault("reward", PolicyFeature(shape=(1,), type=FeatureType.REWARD))
+
+    @property
+    def observation_delta_indices(self) -> list[int] | None:
+        return None
+
+    @property
+    def action_delta_indices(self) -> None:
+        return None
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None
+
+    def validate_features(self) -> None:
+        if self.image_key not in self.input_features:
+            raise ValueError(f"TOPReward requires image input feature {self.image_key!r}")

src/lerobot/rewards/topreward/modeling_topreward.py

@@ -0,0 +1,238 @@
+# Copyright 2026 Shirui Chen, Cole Harrison, Ying-Chun Lee, Angela Jin Yang,
+# Zhongzheng Ren, Lillian J. Ratliff, Jiafei Duan, Dieter Fox, Ranjay Krishna
+# and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics.
+
+Paper:         https://arxiv.org/abs/2602.19313
+Project:       https://topreward.github.io/webpage/
+Original code: https://github.com/TOPReward/TOPReward
+Backbone:      https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct  (default)
+
+TOPReward is a **zero-shot** reward model: it has no fine-tuned weights of
+its own. Given a video trajectory and a task instruction, it asks an
+off-the-shelf VLM how likely the instruction is, conditioned on the video,
+and returns that log-likelihood as the reward signal.
+
+Inference recipe:
+
+1. The processor builds a chat-style prompt, tokenises it, and emits
+   ``input_ids``, ``attention_mask``, vision tensors, and ``labels``.
+   The processor label-masks everything except the terminal answer token with
+   ``-100``.
+2. Forward the full token sequence through the VLM.
+3. Read the terminal answer token log-probability from the logits as the
+   scalar reward.
+
+With the default ``prompt_suffix_template``, the only unmasked token is the
+literal ``"True"`` at the end — the reward is
+``log P("True" | video + prompt + instruction)``.
+
+This LeRobot port is **inference-only and not trainable** — :meth:`forward`
+is intentionally inherited from :class:`PreTrainedRewardModel` and raises
+``NotImplementedError``, making :attr:`PreTrainedRewardModel.is_trainable`
+return ``False``.
+
+Because the VLM weights live on the Hugging Face Hub under their canonical
+id (``Qwen/Qwen3-VL-8B-Instruct`` etc.) and TOPReward never modifies them,
+:meth:`_save_pretrained` and :meth:`from_pretrained` are overridden so a
+TOPReward LeRobot "checkpoint" is a single ``config.json`` (the VLM is
+re-fetched from the Hub at load time).
+"""
+
+from __future__ import annotations
+
+import builtins
+import logging
+import os
+from pathlib import Path
+from tempfile import TemporaryDirectory
+from typing import TYPE_CHECKING, Any, TypeVar
+
+import numpy as np
+import torch
+from huggingface_hub import HfApi, hf_hub_download
+from huggingface_hub.constants import CONFIG_NAME
+from huggingface_hub.errors import HfHubHTTPError
+from torch import Tensor
+from torch.nn.functional import cross_entropy
+
+from lerobot.configs.rewards import RewardModelConfig
+from lerobot.rewards.pretrained import PreTrainedRewardModel
+from lerobot.rewards.topreward.configuration_topreward import TOPRewardConfig
+from lerobot.rewards.topreward.processor_topreward import TOPREWARD_FEATURE_PREFIX, TOPREWARD_INPUT_KEYS
+from lerobot.utils.import_utils import _transformers_available, require_package
+
+if TYPE_CHECKING:
+    from lerobot.configs.train import TrainPipelineConfig
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import Qwen3VLForConditionalGeneration
+else:
+    Qwen3VLForConditionalGeneration = None  # type: ignore[assignment]
+
+logger = logging.getLogger(__name__)
+
+T = TypeVar("T", bound="TOPRewardModel")
+
+
+def _torch_dtype(name: str) -> torch.dtype | str:
+    """Resolve a torch dtype name; ``"auto"`` is passed through verbatim."""
+    if name == "auto":
+        return "auto"
+    dtype = getattr(torch, name, None)
+    if isinstance(dtype, torch.dtype):
+        return dtype
+    raise ValueError(f"Unknown torch dtype: {name!r}")
+
+
+class TOPRewardModel(PreTrainedRewardModel):
+    """TOPReward zero-shot reward model."""
+
+    name = "topreward"
+    config_class = TOPRewardConfig
+
+    def __init__(self, config: TOPRewardConfig) -> None:
+        require_package("transformers", extra="topreward")
+        super().__init__(config)
+        self.config = config
+
+        torch_dtype = _torch_dtype(config.torch_dtype)
+        model_kwargs: dict[str, Any] = {"dtype": torch_dtype, "trust_remote_code": True}
+        if config.attn_implementation is not None:
+            model_kwargs["attn_implementation"] = config.attn_implementation
+
+        self.model = Qwen3VLForConditionalGeneration.from_pretrained(config.vlm_name, **model_kwargs)
+
+    def compute_reward(self, batch: dict[str, Any]) -> Tensor:
+        """Return one log-prob reward per sample in the batch."""
+        inputs: dict[str, Any] = {}
+        for key in TOPREWARD_INPUT_KEYS:
+            batch_key = f"{TOPREWARD_FEATURE_PREFIX}{key}"
+            if batch_key not in batch:
+                raise KeyError(
+                    f"TOPReward batch missing `{batch_key}`. Make sure the "
+                    "TOPRewardEncoderProcessorStep ran before `compute_reward`."
+                )
+            inputs[key] = batch[batch_key]
+
+        device = next(self.model.parameters()).device
+        inputs = {key: value.to(device) if hasattr(value, "to") else value for key, value in inputs.items()}
+        labels = inputs.pop("labels")
+        inputs["logits_to_keep"] = 2
+
+        self.eval()
+        with torch.no_grad():
+            outputs = self.model(**inputs)
+        logits = outputs.logits
+        rewards = -cross_entropy(logits[:, -2, :].float(), labels[:, -1], reduction="none")
+        if np.isfinite(self.config.success_threshold):
+            rewards = (rewards > self.config.success_threshold).float()
+        return rewards.to(self.config.device or "cpu")
+
+    def _save_pretrained(self, save_directory: Path) -> None:
+        """Save ``config.json`` only."""
+        self.config._save_pretrained(save_directory)
+
+    @classmethod
+    def from_pretrained(
+        cls: builtins.type[T],
+        pretrained_name_or_path: str | Path,
+        *,
+        config: RewardModelConfig | None = None,
+        force_download: bool = False,
+        resume_download: bool | None = None,
+        proxies: dict | None = None,
+        token: str | bool | None = None,
+        cache_dir: str | Path | None = None,
+        local_files_only: bool = False,
+        revision: str | None = None,
+        strict: bool = False,  # noqa: ARG003 — accepted for API parity; unused (no safetensors to load)
+        **kwargs: Any,
+    ) -> T:
+        """Load a TOPReward configuration and instantiate the wrapped VLM."""
+        if config is None:
+            config = RewardModelConfig.from_pretrained(
+                pretrained_name_or_path=pretrained_name_or_path,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                token=token,
+                cache_dir=cache_dir,
+                local_files_only=local_files_only,
+                revision=revision,
+                **kwargs,
+            )
+        if not isinstance(config, TOPRewardConfig):
+            raise TypeError(
+                f"Expected a TOPRewardConfig, got {type(config).__name__}. Make sure "
+                f"`pretrained_name_or_path={pretrained_name_or_path!r}` points at a "
+                "TOPReward checkpoint."
+            )
+
+        model_id = str(pretrained_name_or_path)
+        if not os.path.isdir(model_id):
+            try:
+                hf_hub_download(
+                    repo_id=model_id,
+                    filename=CONFIG_NAME,
+                    revision=revision,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    token=token,
+                    local_files_only=local_files_only,
+                )
+            except HfHubHTTPError as e:
+                raise FileNotFoundError(
+                    f"{CONFIG_NAME} not found on the HuggingFace Hub in {model_id}"
+                ) from e
+
+        instance = cls(config, **kwargs)
+        instance.to(config.device)
+        instance.eval()
+        return instance
+
+    def push_model_to_hub(self, cfg: TrainPipelineConfig):
+        """Push the TOPReward ``config.json`` + model card to the Hub."""
+        api = HfApi()
+        repo_id = api.create_repo(
+            repo_id=self.config.repo_id, private=self.config.private, exist_ok=True
+        ).repo_id
+
+        with TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
+            saved_path = Path(tmp) / repo_id
+            saved_path.mkdir(parents=True, exist_ok=True)
+
+            self.config._save_pretrained(saved_path)
+
+            card = self.generate_model_card(
+                cfg.dataset.repo_id, self.config.type, self.config.license, self.config.tags
+            )
+            card.save(str(saved_path / "README.md"))
+
+            cfg.save_pretrained(saved_path)
+
+            commit_info = api.upload_folder(
+                repo_id=repo_id,
+                repo_type="model",
+                folder_path=saved_path,
+                commit_message="Upload TOPReward config and readme",
+                allow_patterns=["*.json", "*.yaml", "*.md"],
+                ignore_patterns=["*.tmp", "*.log", "*.safetensors"],
+            )
+
+            logger.info(f"Model pushed to {commit_info.repo_url.url}")

src/lerobot/rewards/topreward/processor_topreward.py

@@ -0,0 +1,305 @@
+# 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.
+
+"""TOPReward pre/post processing pipeline."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import torch
+from torch import Tensor
+
+from lerobot.configs import PipelineFeatureType, PolicyFeature
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    policy_action_to_transition,
+)
+from lerobot.rewards.topreward.configuration_topreward import (
+    DEFAULT_PROMPT_PREFIX,
+    DEFAULT_PROMPT_SUFFIX_TEMPLATE,
+    TOPRewardConfig,
+)
+from lerobot.types import EnvTransition, TransitionKey
+from lerobot.utils.constants import (
+    OBS_IMAGES,
+    OBS_PREFIX,
+    POLICY_POSTPROCESSOR_DEFAULT_NAME,
+    POLICY_PREPROCESSOR_DEFAULT_NAME,
+)
+from lerobot.utils.import_utils import _transformers_available, require_package
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoProcessor
+else:
+    AutoProcessor = None
+
+TOPREWARD_FEATURE_PREFIX = f"{OBS_PREFIX}topreward."
+
+_TRUE_ANSWER = "True"
+
+TOPREWARD_VLM_INPUT_KEYS = (
+    "input_ids",
+    "attention_mask",
+    "pixel_values_videos",
+    "video_grid_thw",
+    "mm_token_type_ids",
+)
+TOPREWARD_INPUT_KEYS = TOPREWARD_VLM_INPUT_KEYS + ("labels",)
+
+
+def _prepare_video_batch(video: Tensor, *, max_frames: int | None) -> Tensor:
+    """Return videos as ``(B, T, C, H, W)`` uint8 tensors for Qwen3-VL."""
+    if video.ndim == 4:
+        video = video.unsqueeze(1)
+    elif video.ndim != 5:
+        raise ValueError(
+            f"Expected TOPReward frames with shape (B,C,H,W) or (B,T,C,H,W); got {tuple(video.shape)}"
+        )
+
+    if max_frames is not None:
+        video = video[:, -max_frames:]
+    if video.shape[-1] in (1, 3):
+        video = video.permute(0, 1, 4, 2, 3)
+    elif video.shape[2] not in (1, 3):
+        raise ValueError(f"Expected channel dim of size 1 or 3, got shape {tuple(video.shape)}")
+
+    if video.is_floating_point():
+        video = video * 255.0
+
+    return video.clamp(0, 255).to(torch.uint8).contiguous()
+
+
+def _expand_tasks(task: Any, *, batch_size: int, default: str | None) -> list[str]:
+    if task is None:
+        task = default
+    if task is None:
+        raise KeyError("TOPReward expected a task description in complementary data")
+    if isinstance(task, str):
+        return [task] * batch_size
+    if isinstance(task, tuple):
+        task = list(task)
+    if not (isinstance(task, list) and all(isinstance(item, str) for item in task)):
+        raise TypeError(f"TOPReward task must be a string or list of strings, got {type(task)}")
+    if len(task) == 1 and batch_size > 1:
+        return task * batch_size
+    if len(task) != batch_size:
+        raise ValueError(f"Expected {batch_size} tasks, got {len(task)}")
+    return task
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="topreward_encoder")
+class TOPRewardEncoderProcessorStep(ProcessorStep):
+    """Encode raw frames + task into Qwen-VL tensors for the TOPReward model.
+
+    Loads a :class:`~transformers.AutoProcessor` matching ``vlm_name`` and
+    builds the full chat prompt including the instruction suffix. The
+    resulting ``input_ids``, ``attention_mask``, vision tensors, and
+    ``labels`` are written under the ``observation.topreward.*`` namespace
+    so the model can score without re-tokenising.
+
+    At call time the step reads:
+
+    - ``observation[image_key]``: ``(B, T, C, H, W)`` or ``(B, C, H, W)`` frames.
+    - ``complementary_data[task_key]``: a string or list of strings.
+
+    and writes ``observation[f"{TOPREWARD_FEATURE_PREFIX}<name>"]`` for the
+    Qwen-VL tensors plus ``labels``.
+    """
+
+    vlm_name: str = "Qwen/Qwen3-VL-8B-Instruct"
+    image_key: str = OBS_IMAGES + ".top"
+    task_key: str = "task"
+    default_task: str | None = None
+    max_frames: int | None = 16
+    fps: float = 2.0
+    prompt_prefix: str = DEFAULT_PROMPT_PREFIX
+    prompt_suffix_template: str = DEFAULT_PROMPT_SUFFIX_TEMPLATE
+    add_chat_template: bool = False
+    max_length: int = 32768
+
+    _processor: Any = field(default=None, init=False, repr=False)
+
+    def __post_init__(self) -> None:
+        require_package("transformers", extra="topreward")
+        self._processor = AutoProcessor.from_pretrained(self.vlm_name, trust_remote_code=True)
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        observation = transition.get(TransitionKey.OBSERVATION)
+        complementary = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
+        if self.image_key not in observation:
+            raise KeyError(f"TOPReward expected image key {self.image_key!r} in observation")
+
+        frames = observation[self.image_key]
+        videos = frames.detach().cpu() if isinstance(frames, Tensor) else torch.as_tensor(frames)
+        videos = _prepare_video_batch(videos, max_frames=self.max_frames)
+
+        batch_size = videos.shape[0]
+        tasks = _expand_tasks(
+            complementary.get(self.task_key, self.default_task),
+            batch_size=batch_size,
+            default=self.default_task,
+        )
+
+        encoded = self._encode_batch(videos, tasks, batch_size)
+
+        new_observation = dict(observation)
+        for key, value in encoded.items():
+            new_observation[f"{TOPREWARD_FEATURE_PREFIX}{key}"] = value
+
+        new_transition = transition.copy()
+        new_transition[TransitionKey.OBSERVATION] = new_observation
+        return new_transition
+
+    def _encode_batch(self, videos: Tensor, tasks: list[str], batch_size) -> dict[str, Any]:
+        """Tokenise a batch of (frames, task) pairs into Qwen-VL tensors.
+
+        The loop only builds per-sample chat strings. Tokenisation, padding,
+        video preprocessing, and label construction are batched.
+        """
+
+        texts: list[str] = []
+        video_metadata = [
+            {
+                "total_num_frames": int(videos.shape[1]),
+                "fps": float(self.fps),
+                "frames_indices": list(range(int(videos.shape[1]))),
+            }
+            for _ in range(batch_size)
+        ]
+        eos_token = self._processor.tokenizer.eos_token
+
+        for i in range(batch_size):
+            instruction_suffix = self.prompt_suffix_template.format(instruction=tasks[i])
+            if self.add_chat_template:
+                suffix_for_template = instruction_suffix.removesuffix(_TRUE_ANSWER).rstrip()
+                templated_messages = [
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "video", "video": videos[i], "fps": self.fps},
+                            {"type": "text", "text": f"{self.prompt_prefix}{suffix_for_template}"},
+                        ],
+                    }
+                ]
+                prompt_chat = self._processor.apply_chat_template(
+                    templated_messages, tokenize=False, add_generation_prompt=True
+                )
+                full_text = f"{prompt_chat}{_TRUE_ANSWER}"
+            else:
+                user_messages = [
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "video", "video": videos[i], "fps": self.fps},
+                            {"type": "text", "text": self.prompt_prefix},
+                        ],
+                    }
+                ]
+                prompt_chat = self._processor.apply_chat_template(
+                    user_messages, tokenize=False, add_generation_prompt=False
+                )
+                if eos_token is not None:
+                    prompt_chat = prompt_chat.split(eos_token)[0]
+                full_text = f"{prompt_chat}{instruction_suffix}"
+
+            texts.append(full_text)
+
+        result = self._processor(
+            text=texts,
+            videos=videos,
+            video_metadata=video_metadata,
+            do_sample_frames=False,
+            padding=True,
+            padding_side="left",
+            return_tensors="pt",
+        )
+        input_ids = result["input_ids"]
+
+        if input_ids.shape[-1] > self.max_length:
+            raise ValueError(
+                f"TOPReward input length {input_ids.shape[-1]} exceeds max_length "
+                f"{self.max_length}; lower `max_frames` or raise `max_length`."
+            )
+
+        labels = torch.full_like(input_ids, -100)
+        labels[:, -1] = input_ids[:, -1]
+        result["labels"] = labels
+        return result
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        return {
+            "vlm_name": self.vlm_name,
+            "image_key": self.image_key,
+            "task_key": self.task_key,
+            "default_task": self.default_task,
+            "max_frames": self.max_frames,
+            "fps": self.fps,
+            "prompt_prefix": self.prompt_prefix,
+            "prompt_suffix_template": self.prompt_suffix_template,
+            "add_chat_template": self.add_chat_template,
+            "max_length": self.max_length,
+        }
+
+
+def make_topreward_pre_post_processors(
+    config: TOPRewardConfig,
+    dataset_stats: dict[str, dict[str, Any]] | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Pipeline that pre-encodes frames + task into Qwen-VL tensors.
+
+    The preprocessor adds a batch dimension if needed, runs TOPReward's
+    encoder (which tokenises the full prompt and emits ``labels``), and
+    moves everything to the configured device. The postprocessor is
+    the identity since TOPReward outputs a single reward tensor.
+    """
+    preprocessor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+        steps=[
+            AddBatchDimensionProcessorStep(),
+            TOPRewardEncoderProcessorStep(
+                vlm_name=config.vlm_name,
+                image_key=config.image_key,
+                task_key=config.task_key,
+                default_task=config.default_task,
+                max_frames=config.max_frames,
+                fps=config.fps,
+                prompt_prefix=config.prompt_prefix,
+                prompt_suffix_template=config.prompt_suffix_template,
+                add_chat_template=config.add_chat_template,
+                max_length=config.max_input_length,
+            ),
+            DeviceProcessorStep(device=config.device or "cpu"),
+        ],
+        name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+    )
+    postprocessor = PolicyProcessorPipeline(
+        name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+        to_transition=policy_action_to_transition,
+    )
+    return preprocessor, postprocessor

src/lerobot/rl/__init__.py

@@ -12,23 +12,33 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""
-Reinforcement learning modules.
+"""Reinforcement learning modules.
 
-Requires: ``pip install 'lerobot[hilserl]'``
-
-Available modules (import directly)::
-
-    from lerobot.rl.actor import ...
-    from lerobot.rl.learner import ...
-    from lerobot.rl.learner_service import ...
-    from lerobot.rl.buffer import ...
-    from lerobot.rl.eval_policy import ...
-    from lerobot.rl.gym_manipulator import ...
+Distributed actor / learner entry points (``actor``, ``learner``,
+``learner_service``) require ``pip install 'lerobot[hilserl]'``. Algorithms,
+buffer, data sources and trainer are gRPC-free and usable standalone.
 """
 
-from lerobot.utils.import_utils import require_package
+from .algorithms.base import RLAlgorithm as RLAlgorithm
+from .algorithms.configs import RLAlgorithmConfig as RLAlgorithmConfig, TrainingStats as TrainingStats
+from .algorithms.factory import (
+    make_algorithm as make_algorithm,
+    make_algorithm_config as make_algorithm_config,
+)
+from .algorithms.sac.configuration_sac import SACAlgorithmConfig as SACAlgorithmConfig
+from .buffer import ReplayBuffer as ReplayBuffer
+from .data_sources import DataMixer as DataMixer, OnlineOfflineMixer as OnlineOfflineMixer
+from .trainer import RLTrainer as RLTrainer
 
-require_package("grpcio", extra="hilserl", import_name="grpc")
-
-__all__: list[str] = []
+__all__ = [
+    "RLAlgorithm",
+    "RLAlgorithmConfig",
+    "TrainingStats",
+    "make_algorithm",
+    "make_algorithm_config",
+    "SACAlgorithmConfig",
+    "RLTrainer",
+    "ReplayBuffer",
+    "DataMixer",
+    "OnlineOfflineMixer",
+]

src/lerobot/rl/actor.py

@@ -49,39 +49,53 @@ https://github.com/michel-aractingi/lerobot-hilserl-guide
 import logging
 import os
 import time
+from collections.abc import Generator
 from functools import lru_cache
 from queue import Empty
+from typing import TYPE_CHECKING, Any
+
+from lerobot.utils.import_utils import _grpc_available, require_package
+
+if TYPE_CHECKING or _grpc_available:
+    import grpc
+
+    from lerobot.transport import services_pb2, services_pb2_grpc
+    from lerobot.transport.utils import (
+        bytes_to_state_dict,
+        grpc_channel_options,
+        python_object_to_bytes,
+        receive_bytes_in_chunks,
+        send_bytes_in_chunks,
+        transitions_to_bytes,
+    )
+else:
+    grpc = None
+    services_pb2 = None
+    services_pb2_grpc = None
+    bytes_to_state_dict = None
+    grpc_channel_options = None
+    python_object_to_bytes = None
+    receive_bytes_in_chunks = None
+    send_bytes_in_chunks = None
+    transitions_to_bytes = None
 
-import grpc
 import torch
 from torch import nn
-from torch.multiprocessing import Event, Queue
+from torch.multiprocessing import Queue
 
 from lerobot.cameras import opencv  # noqa: F401
 from lerobot.configs import parser
-from lerobot.configs.train import TrainRLServerPipelineConfig
-from lerobot.policies import make_policy
-from lerobot.policies.sac.modeling_sac import SACPolicy
+from lerobot.policies import make_policy, make_pre_post_processors
+from lerobot.processor import TransitionKey
 from lerobot.robots import so_follower  # noqa: F401
 from lerobot.teleoperators import gamepad, so_leader  # noqa: F401
 from lerobot.teleoperators.utils import TeleopEvents
-from lerobot.transport import services_pb2, services_pb2_grpc
-from lerobot.transport.utils import (
-    bytes_to_state_dict,
-    grpc_channel_options,
-    python_object_to_bytes,
-    receive_bytes_in_chunks,
-    send_bytes_in_chunks,
-    transitions_to_bytes,
-)
-from lerobot.types import TransitionKey
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.process import ProcessSignalHandler
 from lerobot.utils.random_utils import set_seed
 from lerobot.utils.robot_utils import precise_sleep
 from lerobot.utils.transition import (
     Transition,
-    move_state_dict_to_device,
     move_transition_to_device,
 )
 from lerobot.utils.utils import (
@@ -89,19 +103,24 @@ from lerobot.utils.utils import (
     init_logging,
 )
 
+from .algorithms.base import RLAlgorithm
+from .algorithms.factory import make_algorithm
 from .gym_manipulator import (
-    create_transition,
     make_processors,
     make_robot_env,
+    reset_and_build_transition,
     step_env_and_process_transition,
 )
 from .queue import get_last_item_from_queue
+from .train_rl import TrainRLServerPipelineConfig
 
 # Main entry point
 
 
 @parser.wrap()
 def actor_cli(cfg: TrainRLServerPipelineConfig):
+    # Fail fast with a friendly error if the optional ``hilserl`` extra is missing.
+    require_package("grpcio", extra="hilserl", import_name="grpc")
     cfg.validate()
     display_pid = False
     if not use_threads(cfg):
@@ -212,7 +231,7 @@ def actor_cli(cfg: TrainRLServerPipelineConfig):
 
 def act_with_policy(
     cfg: TrainRLServerPipelineConfig,
-    shutdown_event: any,  # Event,
+    shutdown_event: Any,  # Event
     parameters_queue: Queue,
     transitions_queue: Queue,
     interactions_queue: Queue,
@@ -252,22 +271,24 @@ def act_with_policy(
     logging.info("make_policy")
 
     ### Instantiate the policy in both the actor and learner processes
-    ### To avoid sending a SACPolicy object through the port, we create a policy instance
+    ### To avoid sending a policy object through the port, we create a policy instance
     ### on both sides, the learner sends the updated parameters every n steps to update the actor's parameters
-    policy: SACPolicy = make_policy(
+    policy = make_policy(
         cfg=cfg.policy,
         env_cfg=cfg.env,
     )
-    policy = policy.eval()
+    policy = policy.to(device).eval()
     assert isinstance(policy, nn.Module)
 
-    obs, info = online_env.reset()
-    env_processor.reset()
-    action_processor.reset()
+    # Build the algorithm
+    algorithm = make_algorithm(cfg=cfg.algorithm, policy=policy)
 
-    # Process initial observation
-    transition = create_transition(observation=obs, info=info)
-    transition = env_processor(transition)
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=cfg.policy,
+        dataset_stats=cfg.policy.dataset_stats,
+    )
+
+    transition = reset_and_build_transition(online_env, env_processor, action_processor)
 
     # NOTE: For the moment we will solely handle the case of a single environment
     sum_reward_episode = 0
@@ -291,8 +312,17 @@ def act_with_policy(
 
         # Time policy inference and check if it meets FPS requirement
         with policy_timer:
-            # Extract observation from transition for policy
-            action = policy.select_action(batch=observation)
+            normalized_observation = preprocessor.process_observation(observation)
+            action = policy.select_action(batch=normalized_observation)
+            # Unnormalize only the continuous part.
+            if cfg.policy.num_discrete_actions is not None:
+                continuous_action = postprocessor.process_action(action[..., :-1])
+                discrete_action = action[..., -1:].to(
+                    device=continuous_action.device, dtype=continuous_action.dtype
+                )
+                action = torch.cat([continuous_action, discrete_action], dim=-1)
+            else:
+                action = postprocessor.process_action(action)
         policy_fps = policy_timer.fps_last
 
         log_policy_frequency_issue(policy_fps=policy_fps, cfg=cfg, interaction_step=interaction_step)
@@ -326,7 +356,8 @@ def act_with_policy(
 
         # Check for intervention from transition info
         intervention_info = new_transition[TransitionKey.INFO]
-        if intervention_info.get(TeleopEvents.IS_INTERVENTION, False):
+        is_intervention = bool(intervention_info.get(TeleopEvents.IS_INTERVENTION, False))
+        if is_intervention:
             episode_intervention = True
             episode_intervention_steps += 1
 
@@ -334,6 +365,7 @@ def act_with_policy(
             "discrete_penalty": torch.tensor(
                 [new_transition[TransitionKey.COMPLEMENTARY_DATA].get("discrete_penalty", 0.0)]
             ),
+            TeleopEvents.IS_INTERVENTION.value: is_intervention,
         }
         # Create transition for learner (convert to old format)
         list_transition_to_send_to_learner.append(
@@ -354,7 +386,7 @@ def act_with_policy(
         if done or truncated:
             logging.info(f"[ACTOR] Global step {interaction_step}: Episode reward: {sum_reward_episode}")
 
-            update_policy_parameters(policy=policy, parameters_queue=parameters_queue, device=device)
+            update_policy_parameters(algorithm=algorithm, parameters_queue=parameters_queue, device=device)
 
             if len(list_transition_to_send_to_learner) > 0:
                 push_transitions_to_transport_queue(
@@ -390,14 +422,7 @@ def act_with_policy(
             episode_intervention_steps = 0
             episode_total_steps = 0
 
-            # Reset environment and processors
-            obs, info = online_env.reset()
-            env_processor.reset()
-            action_processor.reset()
-
-            # Process initial observation
-            transition = create_transition(observation=obs, info=info)
-            transition = env_processor(transition)
+            transition = reset_and_build_transition(online_env, env_processor, action_processor)
 
         if cfg.env.fps is not None:
             dt_time = time.perf_counter() - start_time
@@ -408,10 +433,10 @@ def act_with_policy(
 
 
 def establish_learner_connection(
-    stub: services_pb2_grpc.LearnerServiceStub,
-    shutdown_event: Event,  # type: ignore
+    stub: "services_pb2_grpc.LearnerServiceStub",
+    shutdown_event: Any,  # Event
     attempts: int = 30,
-):
+) -> bool:
     """Establish a connection with the learner.
 
     Args:
@@ -441,12 +466,14 @@ def establish_learner_connection(
 def learner_service_client(
     host: str = "127.0.0.1",
     port: int = 50051,
-) -> tuple[services_pb2_grpc.LearnerServiceStub, grpc.Channel]:
-    """
-    Returns a client for the learner service.
+) -> "tuple[services_pb2_grpc.LearnerServiceStub, grpc.Channel]":
+    """Return a client for the learner service.
 
     GRPC uses HTTP/2, which is a binary protocol and multiplexes requests over a single connection.
     So we need to create only one client and reuse it.
+
+    Returns:
+        tuple[services_pb2_grpc.LearnerServiceStub, grpc.Channel]: The stub and the channel.
     """
 
     channel = grpc.insecure_channel(
@@ -461,16 +488,18 @@ def learner_service_client(
 def receive_policy(
     cfg: TrainRLServerPipelineConfig,
     parameters_queue: Queue,
-    shutdown_event: Event,  # type: ignore
-    learner_client: services_pb2_grpc.LearnerServiceStub | None = None,
-    grpc_channel: grpc.Channel | None = None,
-):
+    shutdown_event: Any,  # Event
+    learner_client: "services_pb2_grpc.LearnerServiceStub | None" = None,
+    grpc_channel: "grpc.Channel | None" = None,
+) -> None:
     """Receive parameters from the learner.
 
     Args:
         cfg (TrainRLServerPipelineConfig): The configuration for the actor.
         parameters_queue (Queue): The queue to receive the parameters.
         shutdown_event (Event): The event to check if the process should shutdown.
+        learner_client (services_pb2_grpc.LearnerServiceStub | None): Optional pre-created stub.
+        grpc_channel (grpc.Channel | None): Optional pre-created channel.
     """
     logging.info("[ACTOR] Start receiving parameters from the Learner")
     if not use_threads(cfg):
@@ -513,12 +542,11 @@ def receive_policy(
 def send_transitions(
     cfg: TrainRLServerPipelineConfig,
     transitions_queue: Queue,
-    shutdown_event: any,  # Event,
-    learner_client: services_pb2_grpc.LearnerServiceStub | None = None,
-    grpc_channel: grpc.Channel | None = None,
-) -> services_pb2.Empty:
-    """
-    Sends transitions to the learner.
+    shutdown_event: Any,  # Event
+    learner_client: "services_pb2_grpc.LearnerServiceStub | None" = None,
+    grpc_channel: "grpc.Channel | None" = None,
+) -> None:
+    """Send transitions to the learner.
 
     This function continuously retrieves messages from the queue and processes:
 
@@ -526,6 +554,13 @@ def send_transitions(
         - A batch of transitions (observation, action, reward, next observation) is collected.
         - Transitions are moved to the CPU and serialized using PyTorch.
         - The serialized data is wrapped in a `services_pb2.Transition` message and sent to the learner.
+
+    Args:
+        cfg (TrainRLServerPipelineConfig): The configuration for the actor.
+        transitions_queue (Queue): The queue to receive the transitions.
+        shutdown_event (Event): The event to check if the process should shutdown.
+        learner_client (services_pb2_grpc.LearnerServiceStub | None): Optional pre-created stub.
+        grpc_channel (grpc.Channel | None): Optional pre-created channel.
     """
 
     if not use_threads(cfg):
@@ -563,18 +598,24 @@ def send_transitions(
 def send_interactions(
     cfg: TrainRLServerPipelineConfig,
     interactions_queue: Queue,
-    shutdown_event: Event,  # type: ignore
-    learner_client: services_pb2_grpc.LearnerServiceStub | None = None,
-    grpc_channel: grpc.Channel | None = None,
-) -> services_pb2.Empty:
-    """
-    Sends interactions to the learner.
+    shutdown_event: Any,  # Event
+    learner_client: "services_pb2_grpc.LearnerServiceStub | None" = None,
+    grpc_channel: "grpc.Channel | None" = None,
+) -> None:
+    """Send interactions to the learner.
 
     This function continuously retrieves messages from the queue and processes:
 
     - Interaction Messages:
         - Contains useful statistics about episodic rewards and policy timings.
         - The message is serialized using `pickle` and sent to the learner.
+
+    Args:
+        cfg (TrainRLServerPipelineConfig): The configuration for the actor.
+        interactions_queue (Queue): The queue to receive the interactions.
+        shutdown_event (Event): The event to check if the process should shutdown.
+        learner_client (services_pb2_grpc.LearnerServiceStub | None): Optional pre-created stub.
+        grpc_channel (grpc.Channel | None): Optional pre-created channel.
     """
 
     if not use_threads(cfg):
@@ -613,7 +654,11 @@ def send_interactions(
     logging.info("[ACTOR] Interactions process stopped")
 
 
-def transitions_stream(shutdown_event: Event, transitions_queue: Queue, timeout: float) -> services_pb2.Empty:  # type: ignore
+def transitions_stream(
+    shutdown_event: Any,  # Event
+    transitions_queue: Queue,
+    timeout: float,
+) -> "Generator[Any, None, services_pb2.Empty]":
     while not shutdown_event.is_set():
         try:
             message = transitions_queue.get(block=True, timeout=timeout)
@@ -629,10 +674,10 @@ def transitions_stream(shutdown_event: Event, transitions_queue: Queue, timeout:
 
 
 def interactions_stream(
-    shutdown_event: Event,
+    shutdown_event: Any,  # Event
     interactions_queue: Queue,
-    timeout: float,  # type: ignore
-) -> services_pb2.Empty:
+    timeout: float,
+) -> "Generator[Any, None, services_pb2.Empty]":
     while not shutdown_event.is_set():
         try:
             message = interactions_queue.get(block=True, timeout=timeout)
@@ -652,7 +697,8 @@ def interactions_stream(
 #  Policy functions
 
 
-def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device):
+def update_policy_parameters(algorithm: RLAlgorithm, parameters_queue: Queue, device):
+    """Drain the latest learner-pushed weights into ``algorithm.policy``."""
     bytes_state_dict = get_last_item_from_queue(parameters_queue, block=False)
     if bytes_state_dict is not None:
         logging.info("[ACTOR] Load new parameters from Learner.")
@@ -667,18 +713,7 @@ def update_policy_parameters(policy: SACPolicy, parameters_queue: Queue, device)
         # - Send critic's encoder state when shared_encoder=True
         # - Skip encoder params entirely when freeze_vision_encoder=True
         # - Ensure discrete_critic gets correct encoder state (currently uses encoder_critic)
-
-        # Load actor state dict
-        actor_state_dict = move_state_dict_to_device(state_dicts["policy"], device=device)
-        policy.actor.load_state_dict(actor_state_dict)
-
-        # Load discrete critic if present
-        if hasattr(policy, "discrete_critic") and "discrete_critic" in state_dicts:
-            discrete_critic_state_dict = move_state_dict_to_device(
-                state_dicts["discrete_critic"], device=device
-            )
-            policy.discrete_critic.load_state_dict(discrete_critic_state_dict)
-            logging.info("[ACTOR] Loaded discrete critic parameters from Learner.")
+        algorithm.load_weights(state_dicts, device=device)
 
 
 #  Utilities functions

src/lerobot/rl/algorithms/__init__.py

@@ -0,0 +1,20 @@
+# 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 .sac import SACAlgorithm, SACAlgorithmConfig
+
+__all__ = [
+    "SACAlgorithm",
+    "SACAlgorithmConfig",
+]

src/lerobot/rl/algorithms/base.py

@@ -0,0 +1,207 @@
+# 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 __future__ import annotations
+
+import abc
+import builtins
+import os
+from collections.abc import Iterator
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, TypeVar
+
+import torch
+from huggingface_hub import hf_hub_download
+from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
+from huggingface_hub.errors import HfHubHTTPError
+from safetensors.torch import load_file as load_safetensors, save_file as save_safetensors
+from torch.optim import Optimizer
+
+from lerobot.types import BatchType
+from lerobot.utils.hub import HubMixin
+
+from .configs import RLAlgorithmConfig, TrainingStats
+
+if TYPE_CHECKING:
+    from torch import nn
+
+    from ..data_sources.data_mixer import DataMixer
+
+T = TypeVar("T", bound="RLAlgorithm")
+
+
+class RLAlgorithm(HubMixin, abc.ABC):
+    """Base for all RL algorithms."""
+
+    config_class: type[RLAlgorithmConfig]
+    name: str
+    config: RLAlgorithmConfig
+
+    @abc.abstractmethod
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """One complete training step.
+
+        The algorithm calls ``next(batch_iterator)`` as many times as it
+        needs (e.g. ``utd_ratio`` times for SAC) to obtain fresh batches.
+        The iterator is owned by the trainer; the algorithm just consumes
+        from it.
+        """
+        raise NotImplementedError
+
+    def configure_data_iterator(
+        self,
+        data_mixer: DataMixer,
+        batch_size: int,
+        *,
+        async_prefetch: bool = True,
+        queue_size: int = 2,
+    ) -> Iterator[BatchType]:
+        """Create the data iterator this algorithm needs.
+
+        The default implementation uses the standard ``data_mixer.get_iterator()``.
+        Algorithms that need specialised sampling should override this method.
+        """
+        return data_mixer.get_iterator(
+            batch_size=batch_size,
+            async_prefetch=async_prefetch,
+            queue_size=queue_size,
+        )
+
+    @abc.abstractmethod
+    def make_optimizers_and_scheduler(self) -> dict[str, Optimizer]:
+        """Build and return the optimizers used during training.
+
+        Called once on the learner side after construction.
+        """
+        raise NotImplementedError
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        """Return optimizers for checkpointing / external scheduling."""
+        return {}
+
+    @property
+    def optimization_step(self) -> int:
+        """Current learner optimization step.
+
+        Part of the stable contract for checkpoint/resume. Algorithms can
+        either use this default storage or override for custom behavior.
+        """
+        return getattr(self, "_optimization_step", 0)
+
+    @optimization_step.setter
+    def optimization_step(self, value: int) -> None:
+        self._optimization_step = int(value)
+
+    def get_weights(self) -> dict[str, Any]:
+        """Policy state-dict to push to actors."""
+        return {}
+
+    @abc.abstractmethod
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        """Load policy state-dict received from the learner."""
+        raise NotImplementedError
+
+    @abc.abstractmethod
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        """Algorithm-owned trainable tensors.
+
+        Must return a flat tensor mapping for everything the algorithm owns
+        that is not part of the policy (e.g. critic ensembles, target networks,
+        temperature parameters). Algorithms with no training-only tensors
+        should explicitly return an empty dict.
+        """
+        raise NotImplementedError
+
+    @abc.abstractmethod
+    def load_state_dict(
+        self,
+        state_dict: dict[str, torch.Tensor],
+        device: str | torch.device = "cpu",
+    ) -> None:
+        """In-place load of algorithm-owned tensors.
+
+        Implementations MUST keep the identity of any ``nn.Parameter`` that an
+        optimizer references (e.g. SAC's ``log_alpha``) by using ``.copy_()``
+        rather than rebinding the attribute.
+        """
+        raise NotImplementedError
+
+    def _save_pretrained(self, save_directory: Path) -> None:
+        """Persist the algorithm's tensors and config to ``save_directory``.
+
+        Writes ``model.safetensors`` (algorithm tensors via :meth:`state_dict`)
+        and ``config.json`` (via :meth:`RLAlgorithmConfig.save_pretrained`).
+        """
+        tensors = {k: v.detach().cpu().contiguous() for k, v in self.state_dict().items()}
+        save_safetensors(tensors, str(save_directory / SAFETENSORS_SINGLE_FILE))
+        self.config._save_pretrained(save_directory)
+
+    @classmethod
+    def from_pretrained(
+        cls: builtins.type[T],
+        pretrained_name_or_path: str | Path,
+        *,
+        policy: nn.Module,
+        config: RLAlgorithmConfig | None = None,
+        force_download: bool = False,
+        resume_download: bool | None = None,
+        proxies: dict | None = None,
+        token: str | bool | None = None,
+        cache_dir: str | Path | None = None,
+        local_files_only: bool = False,
+        revision: str | None = None,
+        device: str | torch.device = "cpu",
+        **algo_kwargs: Any,
+    ) -> T:
+        """Build an algorithm and load its weights from ``pretrained_name_or_path``."""
+        if config is None:
+            config = cls.config_class.from_pretrained(
+                pretrained_name_or_path,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                token=token,
+                cache_dir=cache_dir,
+                local_files_only=local_files_only,
+                revision=revision,
+            )
+        if hasattr(config, "policy_config"):
+            config.policy_config = policy.config
+
+        instance = cls(policy=policy, config=config, **algo_kwargs)
+
+        model_id = str(pretrained_name_or_path)
+        if os.path.isdir(model_id):
+            model_file = os.path.join(model_id, SAFETENSORS_SINGLE_FILE)
+        else:
+            try:
+                model_file = hf_hub_download(
+                    repo_id=model_id,
+                    filename=SAFETENSORS_SINGLE_FILE,
+                    revision=revision,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    token=token,
+                    local_files_only=local_files_only,
+                )
+            except HfHubHTTPError as e:
+                raise FileNotFoundError(
+                    f"{SAFETENSORS_SINGLE_FILE} not found on the HuggingFace Hub in {model_id}"
+                ) from e
+
+        tensors = load_safetensors(model_file)
+        instance.load_state_dict(tensors, device=device)
+        return instance

src/lerobot/rl/algorithms/configs.py

@@ -0,0 +1,138 @@
+# 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 __future__ import annotations
+
+import abc
+import builtins
+import logging
+import os
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, TypeVar
+
+import draccus
+from huggingface_hub import hf_hub_download
+from huggingface_hub.constants import CONFIG_NAME
+from huggingface_hub.errors import HfHubHTTPError
+
+from lerobot.utils.hub import HubMixin
+
+T = TypeVar("T", bound="RLAlgorithmConfig")
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class TrainingStats:
+    """Returned by ``algorithm.update()`` for logging and checkpointing."""
+
+    losses: dict[str, float] = field(default_factory=dict)
+    grad_norms: dict[str, float] = field(default_factory=dict)
+    extra: dict[str, float] = field(default_factory=dict)
+
+    def to_log_dict(self) -> dict[str, float]:
+        """Flatten all stats into a single dict for logging."""
+
+        d: dict[str, float] = {}
+        for name, val in self.losses.items():
+            d[name] = val
+        for name, val in self.grad_norms.items():
+            d[f"{name}_grad_norm"] = val
+        for name, val in self.extra.items():
+            d[name] = val
+        return d
+
+
+@dataclass
+class RLAlgorithmConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
+    """Registry for algorithm configs."""
+
+    @property
+    def type(self) -> str:
+        """Registered name of this algorithm config (e.g. ``"sac"``)."""
+        choice_name = self.get_choice_name(self.__class__)
+        if not isinstance(choice_name, str):
+            raise TypeError(f"Expected string from get_choice_name, got {type(choice_name)}")
+        return choice_name
+
+    @classmethod
+    @abc.abstractmethod
+    def from_policy_config(cls, policy_cfg: Any) -> RLAlgorithmConfig:
+        """Build an algorithm config from a policy config.
+
+        Must be overridden by every registered config subclass.
+        """
+        raise NotImplementedError(f"{cls.__name__} must implement from_policy_config()")
+
+    def _save_pretrained(self, save_directory: Path) -> None:
+        """Serialize this config as ``config.json`` inside ``save_directory``."""
+        with open(save_directory / CONFIG_NAME, "w") as f, draccus.config_type("json"):
+            draccus.dump(self, f, indent=4)
+
+    @classmethod
+    def from_pretrained(
+        cls: builtins.type[T],
+        pretrained_name_or_path: str | Path,
+        *,
+        force_download: bool = False,
+        resume_download: bool | None = None,
+        proxies: dict[Any, Any] | None = None,
+        token: str | bool | None = None,
+        cache_dir: str | Path | None = None,
+        local_files_only: bool = False,
+        revision: str | None = None,
+        **algo_kwargs: Any,
+    ) -> T:
+        model_id = str(pretrained_name_or_path)
+        config_file: str | None = None
+        if Path(model_id).is_dir():
+            if CONFIG_NAME in os.listdir(model_id):
+                config_file = os.path.join(model_id, CONFIG_NAME)
+            else:
+                logger.error(f"{CONFIG_NAME} not found in {Path(model_id).resolve()}")
+        else:
+            try:
+                config_file = hf_hub_download(
+                    repo_id=model_id,
+                    filename=CONFIG_NAME,
+                    revision=revision,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    token=token,
+                    local_files_only=local_files_only,
+                )
+            except HfHubHTTPError as e:
+                raise FileNotFoundError(
+                    f"{CONFIG_NAME} not found on the HuggingFace Hub in {model_id}"
+                ) from e
+
+        if config_file is None:
+            raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
+
+        with draccus.config_type("json"):
+            instance = draccus.parse(RLAlgorithmConfig, config_file, args=[])
+
+        if cls is not RLAlgorithmConfig and not isinstance(instance, cls):
+            raise TypeError(
+                f"Config at {model_id} has type '{instance.type}' but was loaded via "
+                f"{cls.__name__}; use the matching subclass or RLAlgorithmConfig.from_pretrained()."
+            )
+
+        for key, value in algo_kwargs.items():
+            if hasattr(instance, key):
+                setattr(instance, key, value)
+        return instance

src/lerobot/rl/algorithms/factory.py

@@ -0,0 +1,99 @@
+# 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 __future__ import annotations
+
+import torch
+
+from .base import RLAlgorithm
+from .configs import RLAlgorithmConfig
+
+
+def make_algorithm_config(algorithm_type: str, **kwargs) -> RLAlgorithmConfig:
+    """Instantiate an `RLAlgorithmConfig` from its registered type name.
+
+    Args:
+        algorithm_type: Registry key of the algorithm (e.g. ``"sac"``).
+        **kwargs: Keyword arguments forwarded to the config class constructor.
+
+    Returns:
+        An instance of the matching ``RLAlgorithmConfig`` subclass.
+
+    Raises:
+        ValueError: If ``algorithm_type`` is not registered.
+    """
+    try:
+        cls = RLAlgorithmConfig.get_choice_class(algorithm_type)
+    except KeyError as err:
+        raise ValueError(
+            f"Algorithm type '{algorithm_type}' is not registered. "
+            f"Available: {list(RLAlgorithmConfig.get_known_choices().keys())}"
+        ) from err
+    return cls(**kwargs)
+
+
+def get_algorithm_class(name: str) -> type[RLAlgorithm]:
+    """
+    Retrieves an RL algorithm class by its registered name.
+
+    This function uses dynamic imports to avoid loading all algorithm classes into
+    memory at once, improving startup time and reducing dependencies.
+
+    Args:
+        name: The name of the algorithm. Supported names are "sac".
+
+    Returns:
+        The algorithm class corresponding to the given name.
+
+    Raises:
+        ValueError: If the algorithm name is not recognized.
+    """
+    if name == "sac":
+        from .sac.sac_algorithm import SACAlgorithm
+
+        return SACAlgorithm
+    raise ValueError(
+        f"Algorithm type '{name}' is not available. "
+        f"Known: {list(RLAlgorithmConfig.get_known_choices().keys())}"
+    )
+
+
+def make_algorithm(cfg: RLAlgorithmConfig, policy: torch.nn.Module) -> RLAlgorithm:
+    """
+    Instantiate an RL algorithm.
+
+    This factory function looks up the :class:`RLAlgorithm` subclass that matches
+    ``cfg.type`` and instantiates it with the provided policy. It also enforces
+    that ``cfg.policy_config`` has been populated before construction (this is
+    normally handled by :meth:`TrainRLServerPipelineConfig.validate`).
+
+    Args:
+        cfg: The algorithm configuration. Must have ``policy_config`` set.
+        policy: The policy module the algorithm will train.
+
+    Returns:
+        An instantiated :class:`RLAlgorithm`.
+
+    Raises:
+        ValueError: If ``cfg.policy_config`` is ``None`` or ``cfg.type`` is not
+            registered.
+    """
+    if getattr(cfg, "policy_config", None) is None:
+        raise ValueError(
+            f"{type(cfg).__name__}.policy_config is None. "
+            "It must be populated (typically by TrainRLServerPipelineConfig.validate) "
+            "before calling make_algorithm()."
+        )
+    cls = get_algorithm_class(cfg.type)
+    return cls(policy=policy, config=cfg)

src/lerobot/rl/algorithms/sac/__init__.py

@@ -0,0 +1,18 @@
+# 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 .configuration_sac import SACAlgorithmConfig
+from .sac_algorithm import SACAlgorithm
+
+__all__ = ["SACAlgorithm", "SACAlgorithmConfig"]

src/lerobot/rl/algorithms/sac/configuration_sac.py

@@ -0,0 +1,99 @@
+# 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 __future__ import annotations
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.policies.gaussian_actor.configuration_gaussian_actor import (
+    CriticNetworkConfig,
+    GaussianActorConfig,
+)
+
+from ..configs import RLAlgorithmConfig
+
+
+@RLAlgorithmConfig.register_subclass("sac")
+@dataclass
+class SACAlgorithmConfig(RLAlgorithmConfig):
+    """Soft Actor-Critic (SAC) algorithm configuration.
+
+    SAC is an off-policy actor-critic deep RL algorithm based on the maximum
+    entropy reinforcement learning framework. It learns a policy and a Q-function
+    simultaneously using experience collected from the environment.
+
+    This configuration class contains the algorithm-side hyperparameters: critic
+    ensemble, target networks, temperature / entropy tuning, and the Bellman
+    update loop. The policy-side (actor + observation encoder) lives in
+    :class:`~lerobot.policies.gaussian_actor.GaussianActorConfig` and is
+    referenced via :attr:`policy_config`.
+    """
+
+    # Optimizer learning rates
+    # Learning rate for the actor network
+    actor_lr: float = 3e-4
+    # Learning rate for the critic network
+    critic_lr: float = 3e-4
+    # Learning rate for the temperature parameter
+    temperature_lr: float = 3e-4
+
+    # Bellman update
+    # Discount factor for the SAC algorithm
+    discount: float = 0.99
+    # Whether to use backup entropy for the SAC algorithm
+    use_backup_entropy: bool = True
+    # Weight for the critic target update
+    critic_target_update_weight: float = 0.005
+
+    # Critic ensemble
+    # Number of critics in the ensemble
+    num_critics: int = 2
+    # Number of subsampled critics for training
+    num_subsample_critics: int | None = None
+    # Configuration for the critic network architecture
+    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+    # Configuration for the discrete critic network
+    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
+
+    # Temperature / entropy
+    # Initial temperature value
+    temperature_init: float = 1.0
+    # Target entropy for automatic temperature tuning. If ``None``, defaults to
+    # ``-|A|/2`` where ``|A|`` is the total action dimension (continuous + 1 if
+    # there is a discrete action head).
+    target_entropy: float | None = None
+
+    # Update loop
+    # Update-to-data ratio. Set to >1 to enable extra critic updates per env step.
+    utd_ratio: int = 1
+    # Frequency of policy updates
+    policy_update_freq: int = 1
+    # Gradient clipping norm for the SAC algorithm
+    grad_clip_norm: float = 40.0
+
+    # Optimizations
+    # torch.compile is currently disabled by default
+    use_torch_compile: bool = False
+
+    # Policy config
+    policy_config: PreTrainedConfig | None = None
+
+    @classmethod
+    def from_policy_config(cls, policy_cfg: GaussianActorConfig) -> SACAlgorithmConfig:
+        """Build an algorithm config with default hyperparameters for a given policy."""
+        return cls(
+            policy_config=policy_cfg,
+            discrete_critic_network_kwargs=policy_cfg.discrete_critic_network_kwargs,
+        )

src/lerobot/rl/algorithms/sac/sac_algorithm.py

@@ -0,0 +1,672 @@
+# 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 __future__ import annotations
+
+import math
+from collections.abc import Callable, Iterator
+from dataclasses import asdict
+from typing import Any
+
+import einops
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor
+from torch.optim import Optimizer
+
+from lerobot.policies.gaussian_actor.modeling_gaussian_actor import (
+    DISCRETE_DIMENSION_INDEX,
+    MLP,
+    DiscreteCritic,
+    GaussianActorObservationEncoder,
+    GaussianActorPolicy,
+    orthogonal_init,
+)
+from lerobot.policies.utils import get_device_from_parameters
+from lerobot.types import BatchType
+from lerobot.utils.constants import ACTION
+from lerobot.utils.transition import move_state_dict_to_device
+
+from ..base import RLAlgorithm
+from ..configs import TrainingStats
+from .configuration_sac import SACAlgorithmConfig
+
+
+class SACAlgorithm(RLAlgorithm):
+    """Soft Actor-Critic. Owns critics, targets, temperature, and loss computation."""
+
+    config_class = SACAlgorithmConfig
+    name = "sac"
+
+    def __init__(
+        self,
+        policy: GaussianActorPolicy,
+        config: SACAlgorithmConfig,
+    ):
+        self.config = config
+        self.policy_config = config.policy_config
+        self.policy = policy
+        self.optimizers: dict[str, Optimizer] = {}
+        self._optimization_step: int = 0
+
+        action_dim = self.policy.config.output_features[ACTION].shape[0]
+        self._init_critics(action_dim)
+        self._init_temperature(action_dim)
+
+        self._device = torch.device(self.policy.config.device)
+        self._move_to_device()
+
+    def _init_critics(self, action_dim) -> None:
+        """Build critic ensemble, targets."""
+        encoder = self.policy.encoder_critic
+
+        heads = [
+            CriticHead(
+                input_dim=encoder.output_dim + action_dim,
+                **asdict(self.config.critic_network_kwargs),
+            )
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_ensemble = CriticEnsemble(encoder=encoder, ensemble=heads)
+        target_heads = [
+            CriticHead(
+                input_dim=encoder.output_dim + action_dim,
+                **asdict(self.config.critic_network_kwargs),
+            )
+            for _ in range(self.config.num_critics)
+        ]
+        self.critic_target = CriticEnsemble(encoder=encoder, ensemble=target_heads)
+        self.critic_target.load_state_dict(self.critic_ensemble.state_dict())
+
+        # TODO(Khalil): Investigate and fix torch.compile
+        # NOTE: torch.compile is disabled, policy does not converge when enabled.
+        if self.config.use_torch_compile:
+            self.critic_ensemble = torch.compile(self.critic_ensemble)
+            self.critic_target = torch.compile(self.critic_target)
+
+        self.discrete_critic_target = None
+        if self.policy_config.num_discrete_actions is not None:
+            self.discrete_critic_target = self._init_discrete_critic_target(encoder)
+
+    def _init_discrete_critic_target(self, encoder: GaussianActorObservationEncoder) -> DiscreteCritic:
+        """Build target discrete critic (main network is owned by the policy)."""
+        discrete_critic_target = DiscreteCritic(
+            encoder=encoder,
+            input_dim=encoder.output_dim,
+            output_dim=self.policy_config.num_discrete_actions,
+            **asdict(self.config.discrete_critic_network_kwargs),
+        )
+        # TODO(Khalil): Compile the discrete critic
+        discrete_critic_target.load_state_dict(self.policy.discrete_critic.state_dict())
+        return discrete_critic_target
+
+    def _init_temperature(self, continuous_action_dim: int) -> None:
+        """Set up temperature parameter (log_alpha) and target entropy."""
+        temp_init = self.config.temperature_init
+        self.log_alpha = nn.Parameter(torch.tensor([math.log(temp_init)]))
+
+        self.target_entropy = self.config.target_entropy
+        if self.target_entropy is None:
+            total_action_dim = continuous_action_dim + (
+                1 if self.policy_config.num_discrete_actions is not None else 0
+            )
+            self.target_entropy = -total_action_dim / 2
+
+    def _move_to_device(self) -> None:
+        self.policy.to(self._device)
+        self.critic_ensemble.to(self._device)
+        self.critic_target.to(self._device)
+        self.log_alpha = nn.Parameter(self.log_alpha.data.to(self._device))
+        if self.discrete_critic_target is not None:
+            self.discrete_critic_target.to(self._device)
+
+    @property
+    def temperature(self) -> float:
+        """Return the current temperature value, always in sync with log_alpha."""
+        return self.log_alpha.exp().item()
+
+    def _critic_forward(
+        self,
+        observations: dict[str, Tensor],
+        actions: Tensor,
+        use_target: bool = False,
+        observation_features: Tensor | None = None,
+    ) -> Tensor:
+        """Forward pass through a critic network ensemble
+
+        Args:
+            observations: Dictionary of observations
+            actions: Action tensor
+            use_target: If True, use target critics, otherwise use ensemble critics
+
+        Returns:
+            Tensor of Q-values from all critics
+        """
+
+        critics = self.critic_target if use_target else self.critic_ensemble
+        q_values = critics(observations, actions, observation_features)
+        return q_values
+
+    def _discrete_critic_forward(
+        self, observations, use_target=False, observation_features=None
+    ) -> torch.Tensor:
+        """Forward pass through a discrete critic network
+
+        Args:
+            observations: Dictionary of observations
+            use_target: If True, use target critics, otherwise use ensemble critics
+            observation_features: Optional pre-computed observation features to avoid recomputing encoder output
+
+        Returns:
+            Tensor of Q-values from the discrete critic network
+        """
+        discrete_critic = self.discrete_critic_target if use_target else self.policy.discrete_critic
+        q_values = discrete_critic(observations, observation_features)
+        return q_values
+
+    def update(self, batch_iterator: Iterator[BatchType]) -> TrainingStats:
+        """Run one SAC training step (critic / discrete-critic / actor / temperature).
+
+        Pulls ``utd_ratio`` batches from ``batch_iterator``, computes the relevant
+        losses, backpropagates each, and updates target networks.
+
+        Args:
+            batch_iterator: yields batches each containing
+                - ``action``: Action tensor
+                - ``reward``: Reward tensor
+                - ``state``: Observations tensor dict
+                - ``next_state``: Next observations tensor dict
+                - ``done``: Done mask tensor
+                - ``observation_feature``: Optional pre-computed observation features
+                - ``next_observation_feature``: Optional pre-computed next observation features
+                - ``complementary_info`` (optional): per-step extras like discrete penalties
+
+        Returns:
+            TrainingStats with per-component losses and grad norms.
+        """
+        clip = self.config.grad_clip_norm
+
+        for _ in range(self.config.utd_ratio - 1):
+            batch = next(batch_iterator)
+            fb = self._prepare_forward_batch(batch, include_complementary_info=True)
+
+            loss_critic = self._compute_loss_critic(fb)
+            self.optimizers["critic"].zero_grad()
+            loss_critic.backward()
+            torch.nn.utils.clip_grad_norm_(self.critic_ensemble.parameters(), max_norm=clip)
+            self.optimizers["critic"].step()
+
+            if self.policy_config.num_discrete_actions is not None:
+                loss_dc = self._compute_loss_discrete_critic(fb)
+                self.optimizers["discrete_critic"].zero_grad()
+                loss_dc.backward()
+                torch.nn.utils.clip_grad_norm_(self.policy.discrete_critic.parameters(), max_norm=clip)
+                self.optimizers["discrete_critic"].step()
+
+            self._update_target_networks()
+
+        batch = next(batch_iterator)
+        fb = self._prepare_forward_batch(batch, include_complementary_info=False)
+
+        loss_critic = self._compute_loss_critic(fb)
+        self.optimizers["critic"].zero_grad()
+        loss_critic.backward()
+        critic_grad = torch.nn.utils.clip_grad_norm_(self.critic_ensemble.parameters(), max_norm=clip).item()
+        self.optimizers["critic"].step()
+
+        stats = TrainingStats(
+            losses={"loss_critic": loss_critic.item()},
+            grad_norms={"critic": critic_grad},
+        )
+
+        if self.policy_config.num_discrete_actions is not None:
+            loss_dc = self._compute_loss_discrete_critic(fb)
+            self.optimizers["discrete_critic"].zero_grad()
+            loss_dc.backward()
+            dc_grad = torch.nn.utils.clip_grad_norm_(
+                self.policy.discrete_critic.parameters(), max_norm=clip
+            ).item()
+            self.optimizers["discrete_critic"].step()
+            stats.losses["loss_discrete_critic"] = loss_dc.item()
+            stats.grad_norms["discrete_critic"] = dc_grad
+
+        if self._optimization_step % self.config.policy_update_freq == 0:
+            for _ in range(self.config.policy_update_freq):
+                loss_actor = self._compute_loss_actor(fb)
+                self.optimizers["actor"].zero_grad()
+                loss_actor.backward()
+                actor_grad = torch.nn.utils.clip_grad_norm_(
+                    self.policy.actor.parameters(), max_norm=clip
+                ).item()
+                self.optimizers["actor"].step()
+
+                loss_temp = self._compute_loss_temperature(fb)
+                self.optimizers["temperature"].zero_grad()
+                loss_temp.backward()
+                temp_grad = torch.nn.utils.clip_grad_norm_([self.log_alpha], max_norm=clip).item()
+                self.optimizers["temperature"].step()
+
+            stats.losses["loss_actor"] = loss_actor.item()
+            stats.losses["loss_temperature"] = loss_temp.item()
+            stats.grad_norms["actor"] = actor_grad
+            stats.grad_norms["temperature"] = temp_grad
+            stats.extra["temperature"] = self.temperature
+
+        self._update_target_networks()
+        self._optimization_step += 1
+        return stats
+
+    def _compute_loss_critic(self, batch: dict[str, Any]) -> Tensor:
+        # Extract common components from batch
+        observations = batch["state"]
+        actions = batch[ACTION]
+        observation_features = batch.get("observation_feature")
+        # Extract critic-specific components
+        rewards = batch["reward"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+        next_observation_features = batch.get("next_observation_feature")
+
+        with torch.no_grad():
+            next_action_preds, next_log_probs, _ = self.policy.actor(
+                next_observations, next_observation_features
+            )
+
+            # 2- compute q targets
+            q_targets = self._critic_forward(
+                observations=next_observations,
+                actions=next_action_preds,
+                use_target=True,
+                observation_features=next_observation_features,
+            )
+
+            # subsample critics to prevent overfitting if use high UTD (update to date)
+            # TODO: Get indices before forward pass to avoid unnecessary computation
+            if self.config.num_subsample_critics is not None:
+                indices = torch.randperm(self.config.num_critics)
+                indices = indices[: self.config.num_subsample_critics]
+                q_targets = q_targets[indices]
+
+            # critics subsample size
+            min_q, _ = q_targets.min(dim=0)  # Get values from min operation
+            if self.config.use_backup_entropy:
+                min_q = min_q - (self.temperature * next_log_probs)
+
+            td_target = rewards + (1 - done) * self.config.discount * min_q
+
+        # 3- compute predicted qs
+        if self.policy_config.num_discrete_actions is not None:
+            # NOTE: We only want to keep the continuous action part
+            # In the buffer we have the full action space (continuous + discrete)
+            # We need to split them before concatenating them in the critic forward
+            actions: Tensor = actions[:, :DISCRETE_DIMENSION_INDEX]
+        q_preds = self._critic_forward(
+            observations=observations,
+            actions=actions,
+            use_target=False,
+            observation_features=observation_features,
+        )
+
+        # 4- Calculate loss
+        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
+        td_target_duplicate = einops.repeat(td_target, "b -> e b", e=q_preds.shape[0])
+        # You compute the mean loss of the batch for each critic and then to compute the final loss you sum them up
+        critics_loss = (
+            F.mse_loss(
+                input=q_preds,
+                target=td_target_duplicate,
+                reduction="none",
+            ).mean(dim=1)
+        ).sum()
+        return critics_loss
+
+    def _compute_loss_discrete_critic(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        actions = batch[ACTION]
+        rewards = batch["reward"]
+        next_observations = batch["next_state"]
+        done = batch["done"]
+        observation_features = batch.get("observation_feature")
+        next_observation_features = batch.get("next_observation_feature")
+        complementary_info = batch.get("complementary_info")
+
+        # NOTE: We only want to keep the discrete action part
+        # In the buffer we have the full action space (continuous + discrete)
+        # We need to split them before concatenating them in the critic forward
+        actions_discrete: Tensor = actions[:, DISCRETE_DIMENSION_INDEX:].clone()
+        actions_discrete = torch.round(actions_discrete)
+        actions_discrete = actions_discrete.long()
+
+        discrete_penalties: Tensor | None = None
+        if complementary_info is not None:
+            discrete_penalties = complementary_info.get("discrete_penalty")
+
+        with torch.no_grad():
+            # For DQN, select actions using online network, evaluate with target network
+            next_discrete_qs = self._discrete_critic_forward(
+                next_observations, use_target=False, observation_features=next_observation_features
+            )
+            best_next_discrete_action = torch.argmax(next_discrete_qs, dim=-1, keepdim=True)
+
+            # Get target Q-values from target network
+            target_next_discrete_qs = self._discrete_critic_forward(
+                observations=next_observations,
+                use_target=True,
+                observation_features=next_observation_features,
+            )
+
+            # Use gather to select Q-values for best actions
+            target_next_discrete_q = torch.gather(
+                target_next_discrete_qs, dim=1, index=best_next_discrete_action
+            ).squeeze(-1)
+
+            # Compute target Q-value with Bellman equation
+            rewards_discrete = rewards
+            if discrete_penalties is not None:
+                rewards_discrete = rewards + discrete_penalties
+            target_discrete_q = rewards_discrete + (1 - done) * self.config.discount * target_next_discrete_q
+
+        # Get predicted Q-values for current observations
+        predicted_discrete_qs = self._discrete_critic_forward(
+            observations=observations, use_target=False, observation_features=observation_features
+        )
+
+        # Use gather to select Q-values for taken actions
+        predicted_discrete_q = torch.gather(predicted_discrete_qs, dim=1, index=actions_discrete).squeeze(-1)
+
+        # Compute MSE loss between predicted and target Q-values
+        discrete_critic_loss = F.mse_loss(input=predicted_discrete_q, target=target_discrete_q)
+        return discrete_critic_loss
+
+    def _compute_loss_actor(self, batch: dict[str, Any]) -> Tensor:
+        observations = batch["state"]
+        observation_features = batch.get("observation_feature")
+
+        actions_pi, log_probs, _ = self.policy.actor(observations, observation_features)
+
+        q_preds = self._critic_forward(
+            observations=observations,
+            actions=actions_pi,
+            use_target=False,
+            observation_features=observation_features,
+        )
+        min_q_preds = q_preds.min(dim=0)[0]
+
+        actor_loss = ((self.temperature * log_probs) - min_q_preds).mean()
+        return actor_loss
+
+    def _compute_loss_temperature(self, batch: dict[str, Any]) -> Tensor:
+        """Compute the temperature loss"""
+        observations = batch["state"]
+        observation_features = batch.get("observation_feature")
+
+        # calculate temperature loss
+        with torch.no_grad():
+            _, log_probs, _ = self.policy.actor(observations, observation_features)
+
+        temperature_loss = (-self.log_alpha.exp() * (log_probs + self.target_entropy)).mean()
+        return temperature_loss
+
+    def _update_target_networks(self) -> None:
+        """Update target networks with exponential moving average"""
+        for target_p, p in zip(
+            self.critic_target.parameters(), self.critic_ensemble.parameters(), strict=True
+        ):
+            target_p.data.copy_(
+                p.data * self.config.critic_target_update_weight
+                + target_p.data * (1.0 - self.config.critic_target_update_weight)
+            )
+        if self.policy_config.num_discrete_actions is not None:
+            for target_p, p in zip(
+                self.discrete_critic_target.parameters(),
+                self.policy.discrete_critic.parameters(),
+                strict=True,
+            ):
+                target_p.data.copy_(
+                    p.data * self.config.critic_target_update_weight
+                    + target_p.data * (1.0 - self.config.critic_target_update_weight)
+                )
+
+    def _prepare_forward_batch(
+        self, batch: BatchType, *, include_complementary_info: bool = True
+    ) -> dict[str, Any]:
+        observations = batch["state"]
+        next_observations = batch["next_state"]
+        observation_features, next_observation_features = self.get_observation_features(
+            observations, next_observations
+        )
+        forward_batch: dict[str, Any] = {
+            ACTION: batch[ACTION],
+            "reward": batch["reward"],
+            "state": observations,
+            "next_state": next_observations,
+            "done": batch["done"],
+            "observation_feature": observation_features,
+            "next_observation_feature": next_observation_features,
+        }
+        if include_complementary_info and "complementary_info" in batch:
+            forward_batch["complementary_info"] = batch["complementary_info"]
+        return forward_batch
+
+    def make_optimizers_and_scheduler(self) -> dict[str, Optimizer]:
+        """
+        Creates and returns optimizers for the actor, critic, and temperature components of a reinforcement learning policy.
+
+        This function sets up Adam optimizers for:
+        - The **actor network**, ensuring that only relevant parameters are optimized.
+        - The **critic ensemble**, which evaluates the value function.
+        - The **temperature parameter**, which controls the entropy in soft actor-critic (SAC)-like methods.
+
+        It also initializes a learning rate scheduler, though currently, it is set to `None`.
+
+        NOTE:
+        - If the encoder is shared, its parameters are excluded from the actor's optimization process.
+        - The policy's log temperature (`log_alpha`) is wrapped in a list to ensure proper optimization as a standalone tensor.
+
+        Args:
+            cfg: Configuration object containing hyperparameters.
+            policy (nn.Module): The policy model containing the actor, critic, and temperature components.
+
+        Returns:
+            A dictionary mapping component names ("actor", "critic", "temperature")
+            to their respective Adam optimizers.
+        """
+        actor_params = self.policy.get_optim_params()["actor"]
+        self.optimizers = {
+            "actor": torch.optim.Adam(actor_params, lr=self.config.actor_lr),
+            "critic": torch.optim.Adam(self.critic_ensemble.parameters(), lr=self.config.critic_lr),
+            "temperature": torch.optim.Adam([self.log_alpha], lr=self.config.temperature_lr),
+        }
+        if self.policy_config.num_discrete_actions is not None:
+            self.optimizers["discrete_critic"] = torch.optim.Adam(
+                self.policy.discrete_critic.parameters(), lr=self.config.critic_lr
+            )
+        return self.optimizers
+
+    def get_optimizers(self) -> dict[str, Optimizer]:
+        return self.optimizers
+
+    def get_weights(self) -> dict[str, Any]:
+        """Send actor + discrete-critic state dicts."""
+        state_dicts: dict[str, Any] = {
+            "policy": move_state_dict_to_device(self.policy.actor.state_dict(), device="cpu"),
+        }
+        if self.policy_config.num_discrete_actions is not None:
+            state_dicts["discrete_critic"] = move_state_dict_to_device(
+                self.policy.discrete_critic.state_dict(), device="cpu"
+            )
+        return state_dicts
+
+    def load_weights(self, weights: dict[str, Any], device: str | torch.device = "cpu") -> None:
+        """Load actor + discrete-critic weights into the policy."""
+        actor_sd = move_state_dict_to_device(weights["policy"], device=device)
+        self.policy.actor.load_state_dict(actor_sd)
+        if "discrete_critic" in weights and self.policy.discrete_critic is not None:
+            discrete_sd = move_state_dict_to_device(weights["discrete_critic"], device=device)
+            self.policy.discrete_critic.load_state_dict(discrete_sd)
+
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        """Algorithm-owned trainable tensors.
+
+        Encoder weights are stripped because they are owned by the policy
+        (``policy.encoder_critic``) and already saved via ``policy.save_pretrained``.
+        """
+        bundle: dict[str, torch.Tensor] = {}
+        for k, v in _strip_encoder_keys(self.critic_ensemble.state_dict()).items():
+            bundle[f"critic_ensemble.{k}"] = v
+        for k, v in _strip_encoder_keys(self.critic_target.state_dict()).items():
+            bundle[f"critic_target.{k}"] = v
+        if self.discrete_critic_target is not None:
+            for k, v in _strip_encoder_keys(self.discrete_critic_target.state_dict()).items():
+                bundle[f"discrete_critic_target.{k}"] = v
+        bundle["log_alpha"] = self.log_alpha.detach()
+        return bundle
+
+    def load_state_dict(
+        self,
+        state_dict: dict[str, torch.Tensor],
+        device: str | torch.device = "cpu",
+    ) -> None:
+        """In-place load of algorithm-owned tensors.
+
+        ``log_alpha`` is restored via ``Parameter.data.copy_`` so the
+        ``temperature`` optimizer's reference to the parameter object stays
+        valid after resume.
+        """
+        critic_ensemble_state = _split_prefix(state_dict, "critic_ensemble.")
+        critic_target_state = _split_prefix(state_dict, "critic_target.")
+        self.critic_ensemble.load_state_dict(critic_ensemble_state, strict=False)
+        self.critic_target.load_state_dict(critic_target_state, strict=False)
+
+        if self.discrete_critic_target is not None:
+            discrete_target_state = _split_prefix(state_dict, "discrete_critic_target.")
+            self.discrete_critic_target.load_state_dict(discrete_target_state, strict=False)
+
+        if "log_alpha" in state_dict:
+            self.log_alpha.data.copy_(state_dict["log_alpha"].to(self.log_alpha.device))
+
+    def get_observation_features(
+        self, observations: Tensor, next_observations: Tensor
+    ) -> tuple[Tensor | None, Tensor | None]:
+        """
+        Get observation features from the policy encoder. It act as cache for the observation features.
+        when the encoder is frozen, the observation features are not updated.
+        We can save compute by caching the observation features.
+
+        Args:
+            policy: The policy model
+            observations: The current observations
+            next_observations: The next observations
+
+        Returns:
+            tuple: observation_features, next_observation_features
+        """
+
+        if self.policy.config.vision_encoder_name is None or not self.policy.config.freeze_vision_encoder:
+            return None, None
+
+        with torch.no_grad():
+            observation_features = self.policy.actor.encoder.get_cached_image_features(observations)
+            next_observation_features = self.policy.actor.encoder.get_cached_image_features(next_observations)
+
+        return observation_features, next_observation_features
+
+
+def _strip_encoder_keys(state: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
+    """Drop ``encoder.*`` keys from a critic-module state dict."""
+    return {k: v for k, v in state.items() if not k.startswith("encoder.")}
+
+
+def _split_prefix(state: dict[str, torch.Tensor], prefix: str) -> dict[str, torch.Tensor]:
+    """Return the subset of ``state`` whose keys start with ``prefix``, prefix-stripped."""
+    return {k.removeprefix(prefix): v for k, v in state.items() if k.startswith(prefix)}
+
+
+class CriticHead(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dims: list[int],
+        activations: Callable[[torch.Tensor], torch.Tensor] | str = nn.SiLU(),
+        activate_final: bool = False,
+        dropout_rate: float | None = None,
+        init_final: float | None = None,
+        final_activation: Callable[[torch.Tensor], torch.Tensor] | str | None = None,
+    ):
+        super().__init__()
+        self.net = MLP(
+            input_dim=input_dim,
+            hidden_dims=hidden_dims,
+            activations=activations,
+            activate_final=activate_final,
+            dropout_rate=dropout_rate,
+            final_activation=final_activation,
+        )
+        self.output_layer = nn.Linear(in_features=hidden_dims[-1], out_features=1)
+        if init_final is not None:
+            nn.init.uniform_(self.output_layer.weight, -init_final, init_final)
+            nn.init.uniform_(self.output_layer.bias, -init_final, init_final)
+        else:
+            orthogonal_init()(self.output_layer.weight)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.output_layer(self.net(x))
+
+
+class CriticEnsemble(nn.Module):
+    """
+    CriticEnsemble wraps multiple CriticHead modules into an ensemble.
+
+    Args:
+        encoder (GaussianActorObservationEncoder): encoder for observations.
+        ensemble (List[CriticHead]): list of critic heads.
+        init_final (float | None): optional initializer scale for final layers.
+
+    Forward returns a tensor of shape (num_critics, batch_size) containing Q-values.
+    """
+
+    def __init__(
+        self,
+        encoder: GaussianActorObservationEncoder,
+        ensemble: list[CriticHead],
+        init_final: float | None = None,
+    ):
+        super().__init__()
+        self.encoder = encoder
+        self.init_final = init_final
+        self.critics = nn.ModuleList(ensemble)
+
+    def forward(
+        self,
+        observations: dict[str, torch.Tensor],
+        actions: torch.Tensor,
+        observation_features: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        device = get_device_from_parameters(self)
+        # Move each tensor in observations to device
+        observations = {k: v.to(device) for k, v in observations.items()}
+
+        obs_enc = self.encoder(observations, cache=observation_features)
+
+        inputs = torch.cat([obs_enc, actions], dim=-1)
+
+        # Loop through critics and collect outputs
+        q_values = []
+        for critic in self.critics:
+            q_values.append(critic(inputs))
+
+        # Stack outputs to match expected shape [num_critics, batch_size]
+        q_values = torch.stack([q.squeeze(-1) for q in q_values], dim=0)
+        return q_values

src/lerobot/rl/buffer.py

@@ -97,8 +97,8 @@ class ReplayBuffer:
         Args:
             capacity (int): Maximum number of transitions to store in the buffer.
             device (str): The device where the tensors will be moved when sampling ("cuda:0" or "cpu").
-            state_keys (List[str]): The list of keys that appear in `state` and `next_state`.
-            image_augmentation_function (Optional[Callable]): A function that takes a batch of images
+            state_keys (list[str]): The list of keys that appear in `state` and `next_state`.
+            image_augmentation_function (Callable | None): A function that takes a batch of images
                 and returns a batch of augmented images. If None, a default augmentation function is used.
             use_drq (bool): Whether to use the default DRQ image augmentation style, when sampling in the buffer.
             storage_device: The device (e.g. "cpu" or "cuda:0") where the data will be stored.
@@ -634,7 +634,7 @@ class ReplayBuffer:
                 If None, you must handle or define default keys.
 
         Returns:
-            transitions (List[Transition]):
+            transitions (list[Transition]):
                 A list of Transition dictionaries with the same length as `dataset`.
         """
         if state_keys is None: