BodenAI, PIA, and the Shanghai Jiao Tong University MINT Lab jointly announce the open‑source release of the first batch of the Real‑World Reinforcement Learning Dataset (RW‑RL-Dataset) for embodied policy iteration.

The dataset targets a core bottleneck in the physical AI industry: how to make robots continuously improve in real environments, rather than merely learning a specific task.

This first open‑source release contains 1,000+ hours of real‑robot interaction data, covering 4 robot families, 9 scenario domains, 30+ task templates, and 3 core data types. To the best of our knowledge, this is currently the largest publicly available open‑source dataset oriented specifically toward real‑robot reinforcement learning (RL).

1. Why Imitation Learning Alone Is No Longer Enough

Today, most embodied AI systems are built around teleoperation-based imitation learning. Human demonstrations provide a reference trajectory for the policy to imitate. This approach provides a strong initialization signal and relatively clean training data. However, it has a fundamental limitation: the policy primarily learns to reproduce behaviors observed in the demonstrations. Once the execution deviates from the demonstrated state distribution, recovery often becomes difficult or impossible.

Reinforcement learning seeks to overcome this limitation by allowing robots to learn directly from interaction with the real world through iterative cycles of:

• executing autonomously,
• attempting and deviating,
• being taken over by a human when necessary,
• recovering,
• and attempting again.

Only through these cycles can a policy learn when to self‑recover and when to request human assistance.

However, this research direction has long been constrained by a persistent bottleneck: where does the data come from? Teleoperation trajectories typically lack failure and recovery; simulation data cannot fully capture the physical details of the real world.

This is precisely the design starting point of this RW-RL-Dataset.

Core design principle: the dataset records not only what went right, but also when human intervention is needed, how recovery is performed, and whether the system can require less help over time.

2. Dataset Specifications: A Large‑scale Open Foundation For Robot RL

2.1 Scale and Coverage

‍

The complete technical documentation and download method will be published on the open‑source platform, with support for direct loading into mainstream RL training frameworks.

2.2 Three Data Types (beyond “success‑only” trajectories)

Unlike many existing public datasets that contain only successful trajectories, RW‑RL-Dataset explicitly distinguishes three regimes from the start:

(1) Teleoperation data

• A human controls the robot throughout the task.
• Records multi‑view video, joint states, and action sequences.
• Functionality: provides high‑quality priors for imitation‑learning pretraining or policy initialization.
• Positioning: the policy’s starting point — solves “how to begin.”

(2) Human‑intervention data

• A policy executes on a real robot.
• When it deviates from the intended trajectory or encounters an out-of-distribution state, a human takes over to correct, then returns control to the policy.
• Records: intervention onset, intervention termination, intervention rationale, and state changes before/after intervention.
• Functionality: the most critical training material for RL — it directly captures how humans recover robots from failure states.
• Positioning: error‑correction samples — key data for learning recovery mechanisms.

(3) Autonomous inference data

• Trajectories generated by autonomous policy execution in real-world environments.
• Includes complete processes for both completed and incomplete attempts.
• Functionality: these trajectories reflect the policy’s actual state-action distribution and are indispensable for RL policy iteration.
• Positioning: a mirror of the policy’s real behavior — core source for training‑distribution coverage.

2.3 Reward Modeling and Quality Calibration: A Multi-Signal Framework

RW-RL-Dataset does not rely on a single reward annotation mode. Instead, it adopts a three‑layer signal fusion framework:

• Layer 1 — Rule‑based and statistical signals: beyond task completion, includes execution speed percentiles, action smoothness (jitter/pause/rollback detection), and sanity checks on gripper and joint behavior.
• Layer 2 — Event signals: the system automatically records structured information such as control‑handoff events, intervention frames, and task phases; some intervention reasons are supplemented with manual annotations.
• Layer 3 — Model‑assisted scoring: used for trajectory quality ranking, phase‑completion estimation, and reward‑model pretraining.

2.4 Data Quality Assurance: Automated Validation and Human Review

BodenAI established a strict quality control pipeline. Before ingestion, each episode must pass the following checks:

‍

After automated validation, data enters a manual review stage focused on edge cases that are difficult to judge with rules alone (e.g., semantic reasonableness of actions, trajectory smoothness). Only data that passes both automated validation and manual review is included in the final dataset release.

3. Technical Capabilities: BodenAI’s Physical AI Data Infrastructure

Producing large‑scale, high‑quality robot data cannot be achieved by “manual scaling.” For this open‑source release, BodenAI provided a complete infrastructure stack spanning capture, production, and verification.

3.1 BRIC Robo: unified capture and runtime management

BodenAI’s in‑house BRIC Robo platform addresses a core challenge in heterogeneous robot data capture.

It standardizes integration through a four‑step workflow:

1. robot registration,
2. streaming protocol integration,
3. joint‑space mapping,
4. sensor alignment.

BRIC Robo has been deeply adapted to multiple leading robot platforms and intelligent production lines (including ZERITH, Astribot, Galaxea, AgileX Robotics, PIA, among others). New robot models can be integrated via one‑click configuration without repeated code adaptation.

Across BodenAI’s three physical AI innovation centers, more than 500 physical robots have been deployed; among them, 20+ distinct models already support unified data capture through BRIC Robo.

This capability enables broader robot family coverage and reduces the risk of overfitting to a single embodiment.

3.2 BASE Omni and Blink: automated data production pipeline

BodenAI upgraded data production from “manual and coarse” to engineering‑grade automation:

• BASE Omni: includes 200+ automated annotation small models and 10+ domain annotation large models, covering the full pipeline from cleaning to labeling. The built-in intelligent labeling engine delivers up to a 7-fold increase in processing efficiency.
• Blink (data management platform): provides full lifecycle management from raw captured data to finalized dataset artifacts, including version control, validation traceability, and multi‑format export.

Together, this workflow frees operators from repetitive cleaning and labeling, allowing experts to focus on high‑value edge‑case judgment and quality acceptance.

3.3 A “real‑robot validation → data feedback” closed loop

Across the three innovation centers, BodenAI has deployed hundreds of real‑world scenario setups (production workstations, retail shelves, warehouse turnover areas, hotel service settings, home environments, etc.). In these environments:

1. The dataset supports rollout validation for RL policies.
2. Successful and failed rollout data is fed back into the capture platform.
3. Feedback data is used for the next round of policy iteration and dataset expansion.

This dataset is not static; it is produced by a continuously evolving real‑robot interaction system.

4. Why We Describe It as “Currently the Largest Open Real-world RL Dataset for Robotics”: A Fair, Technical Definition

While preparing this release, we observed community discussions comparing open datasets. To keep this claim fair and technically acceptable, we state the following facts:

• Open X‑Embodiment: released by Google DeepMind and 21 labs; covers 1 million real‑robot trajectories across 22 robots. However, it is fundamentally an imitation-learning dataset and doesn't include key components required for reinforcement learning, such as human‑intervention data, autonomous rollout trajectories, and reward annotations.
• AgiBot (AgiBot World): an open, million‑scale real‑robot dataset with a total duration of approximately 2,976.4 hours, covering five major scenario categories. It primarily targets large‑scale imitation learning and general policy pretraining.
• RW‑RL-Dataset: the first large‑scale open dataset defined explicitly around real‑robot reinforcement learning. In addition to teleoperation data, it explicitly includes human‑intervention data and autonomous inference data and provides reward signals and a complete RL training label system. Under the real‑robot RL definition, this dataset, to the best of our knowledge, is the largest publicly announced open dataset specifically designed for real-world robot reinforcement learning at the time of writing.

In short, RW-RL-Dataset is not merely a larger imitation-learning dataset. It is designed around a fundamentally different objective: enabling reinforcement learning on real robots at scale. Our goal is to provide the missing infrastructure layer that has historically limited large-scale real-world robot RL research.

We welcome researchers, developers, and industry partners to join us in building an open and scalable data infrastructure for the next generation of Physical AI systems.

‍

Download the RW‑RL Dataset and access the full technical documentation from:

https://github.com/MINT-SJTU/Evo-RL

https://huggingface.co/datasets/MINT-SJTU/RW-RL-Datasets