MINT-SJTU/RW-RL-Dataset

RW-RL Dataset: Real-World Reinforcement Learning for Robots

<p> <a href="https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset"> <img alt="Hugging Face Dataset" src="https://img.shields.io/badge/Hugging%20Face-RW--RL--Dataset-yellow"> </a> <a href="https://github.com/MINT-SJTU/Evo-RL"> <img alt="GitHub" src="https://img.shields.io/badge/GitHub-Evo--RL-black"> </a> <img alt="Data scale" src="https://img.shields.io/badge/Planned%20scale-1000%2B%20real--robot%20hours-blue"> <img alt="Modality" src="https://img.shields.io/badge/Modality-video%20%2B%20state%20%2B%20action-green"> </p>

Human-intervention companion dataset: RW-RL-HIL-Dataset (BodenAI) is a separately hosted release of real-world policy rollouts and human corrections. Download it from its own dataset page.

RW-RL Dataset is a real-world robot interaction dataset released by Boden Intelligence, Junpu Innovation Center, and the MINT Lab at Shanghai Jiao Tong University. It is designed for a bottleneck that imitation-learning-only datasets do not directly solve: how can robot policies keep improving in the physical world after they leave the ideal demonstration trajectory?

The broader RW-RL data program targets 1000+ hours of real robot interaction data across multiple robot embodiments, scene domains, task templates, and data regimes. In addition to teleoperated demonstrations, the dataset design emphasizes human intervention, autonomous rollouts, reward signals, done labels, and structured task semantics so that researchers can study policy correction, recovery, and closed-loop improvement in real environments.

The current Hugging Face repository contains R1Lite, G2A, PiperX, and SO101 releases. The full RW-RL program is designed as an expanding real-robot RL data base, and future subsets can follow the same file layout.

Dataset Coverage and Demo

<p align="center"> <img src="https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset/resolve/main/assets/sceneembodimentcoverage_en.png" alt="Scenario and embodiment coverage" width="100%"> <br> _{Scenario and embodiment coverage across robot series, scene domains, task templates, and data types.} </p>

<p align="center"> <video src="https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset/resolve/main/assets/0609RW-RL.mp4" controls muted width="100%"></video> <br> _{RW-RL real-robot task demonstration.} </p>

Visual Examples

The examples below are loaded directly from files already present in this repository.

<table> <tr> <td align="center" width="50%"> <video src="https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset/resolve/main/r1liteplugchargertopowerstrip/videos/observation.images.cameratop/chunk-000/file-000.mp4" controls muted width="100%"></video> <br> _{Plug charger to power strip - top view} </td> <td align="center" width="50%"> <video src="https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset/resolve/main/r1litescanitemtobag/videos/observation.images.cameratop/chunk-000/file-000.mp4" controls muted width="100%"></video> <br> _{Scan item to bag - top view} </td> </tr> <tr> <td align="center" width="50%"> <video src="https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset/resolve/main/r1liteshelfrestockfromcart/videos/observation.images.cameratop/chunk-000/file-000.mp4" controls muted width="100%"></video> <br> _{Shelf restock from cart - top view} </td> <td align="center" width="50%"> <video src="https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset/resolve/main/r1litetissueboxreplace/videos/observation.images.cameratop/chunk-000/file-000.mp4" controls muted width="100%"></video> <br> _{Tissue box replacement - top view} </td> </tr> </table>

Why Real-World RL Data?

Most embodied-AI data pipelines still start from teleoperated imitation learning. A human operator provides a clean demonstration, and the policy learns to imitate that standard path. This is effective for bootstrapping, but it leaves a core limitation: when the robot drifts away from the demonstrated state distribution, the policy often has no data that explains how to recover.

Real-world reinforcement learning needs a different kind of data. A robot must execute, explore, make mistakes, be corrected, recover, and try again under real physical constraints. RW-RL Dataset is built around this loop. The goal is not only to record what success looks like, but also to record when help is needed, how the system is rescued, and whether the next policy iteration can require less intervention.

Dataset Program at a Glance

Three Data Regimes

Reward and Quality Signals

RW-RL Dataset is designed around a multi-signal reward and quality framework instead of a single success/failure label.

Each episode is expected to pass both automated checks and human review before final publication. The automated quality checks include motion validity, derivative spike detection, physical joint and gripper limits, outlier screening, multi-camera timing consistency, and minimum episode completeness.

Current Hub Release

The current repository snapshot contains the R1Lite task folders below. It follows a LeRobot-style layout with one parquet file per episode and one video file per camera stream per episode.

The current R1Lite release uses 15 FPS H.264 videos. The visible camera streams are:

• —observation.images.camera_top
• —observation.images.camera_top_right
• —observation.images.camera_wrist_left
• —observation.images.camera_wrist_right

G2A Release

The G2A release uses the same one-episode-per-parquet and one-episode-per-video layout at 30 FPS with eight camera streams. Project03 adds episodes to two existing tasks and introduces the key insertion and unlocking task.

The eight G2A camera streams are:

• —observation.images.camera_fisheye_left
• —observation.images.camera_fisheye_right
• —observation.images.camera_stereo_left
• —observation.images.camera_stereo_right
• —observation.images.camera_top
• —observation.images.camera_top_back
• —observation.images.camera_wrist_left
• —observation.images.camera_wrist_right

SO101 Dual-Arm Release

The SO101 release contains six dual-arm manipulation tasks in LeRobot v3 format. It uses the bi_so_follower robot type, 30 FPS H.264 video at 640 x 480, and 12-dimensional action and observation.state vectors. Each task folder contains one parquet file per episode and one video file per camera stream per episode.

All SO101 episodes expose the same three camera streams:

• —observation.images.left_wrist
• —observation.images.right_front
• —observation.images.right_wrist

One historical power-cord episode without a right_wrist stream was excluded from this release. No synthetic, duplicated, or black replacement frames were introduced.

File Layout

Each task folder is organized as follows:

<task_name>/
  data/
    chunk-000/
      file-000.parquet
      file-001.parquet
      ...
  videos/
    observation.images.camera_top/
      chunk-000/
        file-000.mp4
        file-001.mp4
        ...
    observation.images.camera_top_right/
      chunk-000/
        file-000.mp4
        ...
    observation.images.camera_wrist_left/
      chunk-000/
        file-000.mp4
        ...
    observation.images.camera_wrist_right/
      chunk-000/
        file-000.mp4
        ...
  meta/
    info.json
    tasks.parquet
    episodes/
      chunk-000/
        file-000.parquet

The task metadata defines the path templates:

data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4

The visible R1Lite release uses chunks_size = 1000, so:

chunk_index = episode_index // 1000
file_index = episode_index % 1000

Modalities and Fields

The R1Lite subset exposes synchronized video, state, action, and indexing fields. A representative task metadata file declares the following feature groups:

Downloading a Subset

Use huggingface_hub to download a single task folder without pulling the entire repository:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="MINT-SJTU/RW-RL-Dataset",
    repo_type="dataset",
    allow_patterns=[
        "r1lite_plug_charger_to_power_strip/**",
    ],
    local_dir="RW-RL-Dataset",
)

To inspect only metadata first:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="MINT-SJTU/RW-RL-Dataset",
    repo_type="dataset",
    allow_patterns=[
        "*/meta/**",
    ],
    local_dir="RW-RL-Dataset-meta",
)

How RW-RL Differs from Imitation-Only Corpora

RW-RL Dataset is positioned as a large-scale real-world RL dataset under a specific definition: the key requirement is not only real robot trajectories, but trajectories that support reinforcement learning through intervention, autonomous rollout, reward, done, and recovery-oriented labels.

This distinction is important. RW-RL Dataset is not meant to replace imitation-learning data. It complements it by adding the missing pieces needed by real-world RL researchers: intervention traces, autonomous attempts, structured recovery information, and reward-oriented signals.

Physical AI Data Infrastructure

The data program is supported by a real-robot data infrastructure stack:

Project Links

• —Human-intervention companion dataset: <https://huggingface.co/datasets/BodenAI/RW-RL-HIL-Dataset>

• —GitHub: <https://github.com/MINT-SJTU/Evo-RL>
• —Hugging Face Dataset: <https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset>

Citation

If you use this dataset, please cite the project repository and this Hugging Face dataset page. A formal citation entry will be added when the authors provide one.

@misc{rw_rl_dataset_2026,
  title        = {RW-RL Dataset: Real-World Reinforcement Learning Dataset},
  author       = {Boden Intelligence and Junpu Innovation Center and MINT Lab, Shanghai Jiao Tong University},
  year         = {2026},
  howpublished = {https://huggingface.co/datasets/MINT-SJTU/RW-RL-Dataset}
}