NINGBO, China — Junpu Intelligence announced today that its Ningbo Embodied AI Robot Innovation Center, in collaboration with Bodeng Intelligence and the SJTU MINT Lab, has officially open-sourced the first batch of the world's first large-scale Real-World Reinforcement Learning (RW-RL) dataset, RW-RL-Dataset.
Unlike traditional robotic datasets that only record "successful trajectories," this dataset captures the complete execution process including successes, failures, and recovery, providing a reproducible real-world data foundation for the global embodied AI community.
Key specifications
The first version of RW-RL-Dataset contains 1,000+ hours of real robot data, covering 4+ robot series, 9+ scenario domains, 30+ task templates, and 3 data modalities. It supports human-in-the-loop, autonomous robot exploration, and offline/online reinforcement learning training.
The data is derived from real manufacturing tasks including unpacking and boxing, threading and insertion, parts sorting, and other typical industrial operations, transformed into high-quality real robot data covering four core skills: grasping, insertion, placement, and tightening.
Bridging the industrial gap
"The last-mile gap between general-purpose robot bodies and industrial scenarios lies in the lack of high-quality industrial data," said He Chuan, head of the Junpu Robot Innovation Center. "This open-source release is just the starting point."
The center plans to expand the dataset to 3,000+ hours by the end of 2026, with a focus on high-value scenarios such as industrial precision assembly, to drive embodied AI toward true large-scale industrial deployment.
