MINT-SJTU/RW-RL-Dataset

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

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 the R1Lite release files listed below. 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

RW-RL real-robot task demonstration

Scenario and embodiment coverage across robot series, scene domains, task templates, and data types

Visual Examples

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

Plug charger to power strip - top view	Scan item to bag - top view
Shelf restock from cart - top view	Tissue box replacement - top view

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

Dimension	Description
Planned open-source scale	1000+ hours of real-robot interaction data
Robot families	4 robot series, covering major forms such as single-arm, dual-arm, and humanoid-style systems
Scene domains	9 domains, including industrial production, warehouse logistics, retail/service settings, and research environments
Task templates	30+ templates, including grasping, placing, insertion, tightening, arrangement, stacking, and related manipulation skills
Data regimes	Human teleoperation, human intervention, and autonomous rollout/inference
Modalities	Multi-view video, robot state, action, reward, done signal, and task semantic labels

Three Data Regimes

Regime	What is recorded	Why it matters
Human teleoperation	A human controls the robot through the full task while multi-view video, joint states, and action sequences are recorded.	Provides high-quality priors for imitation-learning pretraining and policy initialization.
Human intervention	A policy runs on the real robot, enters an off-distribution or failing state, and a human takes over, corrects the behavior, then hands control back.	Captures the recovery process that RL policies need in order to learn from errors.
Autonomous rollout	The model runs by itself in the real environment, including both successful and incomplete attempts.	Covers the policy's own state distribution, which is essential for iterative RL improvement.

Reward and Quality Signals

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

Signal layer	Examples
Rule and statistics signals	Task completion, execution speed percentiles, action smoothness, jitter, pauses, backward motions, gripper validity, and joint-range checks.
Event signals	Control-mode switches, intervention start/end frames, task phases, and structured intervention reasons when available.
Model-assisted scores	Trajectory ranking, stage completion estimates, and signals that can support reward-model pretraining.

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.

Quality check	Purpose
Abnormal stillness check	Ensures that the robot is not static for an excessive fraction of the episode.
Spike detection	Detects abnormal first- or second-order joint derivative jumps.
Physical range check	Confirms that gripper and joint values stay within valid ranges.
Outlier detection	Filters angles, velocities, and accelerations outside statistical thresholds.
Frame-rate consistency	Checks multi-camera temporal alignment, targeting <= 10 ms timing error.
Episode completeness	Filters episodes shorter than the required minimum duration.

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.

Task folder	Episodes	Video files	Camera streams	Chunks
r1lite_plug_charger_to_power_strip	1,792	7,168	4	chunk-000, chunk-001
r1lite_scan_item_to_bag	1,989	7,956	4	chunk-000, chunk-001
r1lite_shelf_restock_from_cart	2,280	9,120	4	chunk-000 to chunk-002
r1lite_spray_bottle_head_install	3,053	12,212	4	chunk-000 to chunk-003
r1lite_tissue_box_replace	3,222	12,888	4	chunk-000 to chunk-003
Total currently visible on Hub	12,336	49,344	4 per episode	1000 episodes per chunk

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

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:

Group	Fields
Multi-view video	observation.images.camera_top, observation.images.camera_top_right, observation.images.camera_wrist_left, observation.images.camera_wrist_right
Arm state	observation.state.left_arm, observation.state.left_arm.velocities, observation.state.right_arm, observation.state.right_arm.velocities
Gripper state	observation.state.left_gripper, observation.state.right_gripper
Torso and chassis state	observation.state.torso, observation.state.torso.velocities, observation.state.chassis, observation.state.chassis.velocities, observation.state.chassis.imu
End-effector pose	observation.state.left_ee_pose, observation.state.right_ee_pose
Actions	action.left_arm, action.right_arm, action.left_gripper, action.right_gripper, action.torso, action.chassis.velocities
Episode indexing	timestamp, frame_index, episode_index, index, task_index

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.

Dataset direction	Representative examples	Primary focus	RL-specific intervention / rollout data	Reward-oriented labels
Large imitation-learning corpora	Open X-Embodiment and similar cross-embodiment demonstration datasets	Broad robot demonstrations for behavior cloning and generalist policy pretraining	Not the central design target	Usually not the central design target
Large real-robot pretraining corpora	Large-scale robot datasets such as AgiBot World	High-volume real-robot data for general policy pretraining	Not the main organizing principle	Not the main organizing principle
RW-RL Dataset	This dataset program	Real-world policy improvement, correction, recovery, and closed-loop iteration	Core design target	Core design target

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:

Component	Role
BRIC Robo	A unified robot registration, streaming, joint mapping, and sensor-alignment platform for heterogeneous robot bodies.
BASE Omni	An automated data annotation pipeline with many small annotation models and domain-specific large annotation models.
Blink	A data management platform for lifecycle tracking, versioning, quality traceability, and multi-format output.
Real-robot validation loop	Rollouts are evaluated in real environments, successful and failed trajectories flow back into the data platform, and the next policy iteration uses the returned data.

Project Links

• 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}
}