RoboMIND (Multi-embodiment Intelligence Normative Data for Robot Manipulation)

A multi-embodiment real-robot manipulation dataset from the Beijing Humanoid Robot Innovation Center and Peking University, collected under a unified protocol.

RoboMIND was released in December 2024 by the Beijing Humanoid Robot Innovation Center, Peking University's School of Computer Science, and the Beijing Academy of Artificial Intelligence, with the paper accepted at RSS 2025. Under a unified collection protocol, it gathered 107,000 demonstrations on four embodiments — Franka, UR5e, an AgileX dual-arm platform, and a humanoid with two dexterous hands — covering 479 tasks and 96 object categories, plus 5,000 failed demonstrations with the reason labeled, and it built a digital-twin environment in Isaac Sim; every trajectory is stored as one HDF5 file. RoboMIND 2.0, from December 2025, expanded to six embodiments, 739 tasks, and more than 310,000 dual-arm trajectories, adding 12,000 tactile segments, 20,000 mobile-manipulation trajectories, and 20,000 simulated trajectories, and proposed a hierarchical dual-system framework called MIND-2. The data is downloadable on ModelScope and Hugging Face.

ExampleA researcher could take only RoboMIND's UR5e single-arm trajectories to train an imitation-learning policy, then compare against the failed demonstrations for the same task to see where the policy tends to go wrong.

Also called

RoboMIND 2.0

Related

Cross-Embodiment Data · Real-Robot Data · Failure Data · AgiBot World · Open X-Embodiment · Beijing Humanoid Robot Innovation Center

Sources

RoboMIND: Benchmark on Multi-embodiment Intelligence Normative Data for Robot Manipulation
RoboMIND 2.0: A Multimodal, Bimanual Mobile Manipulation Dataset for Generalizable Embodied Intelligence
RoboMIND 项目主页 (Chinese)

As of

2025-12