RoboMIND — Multi-Embodiment Intelligence Normative Data

RoboMIND is a comprehensive multi-embodiment robot manipulation dataset featuring 107,000 real-world demonstration trajectories spanning 479 distinct tasks and 96 unique object classes. Released under Apache 2.0 for commercial use, it integrates teleoperation data from four distinct robot platforms: Franka Emika Panda, Tien Kung humanoid, AgileX Cobot Magic V2.0, and UR-5e. The inclusion of the Tien Kung humanoid robot makes RoboMIND one of the few large-scale datasets with real humanoid manipulation demonstrations rather than arm-only data. At 479 distinct tasks it provides the broadest task coverage of any multi-embodiment manipulation dataset currently available, enabling systematic study of cross-embodiment and cross-task generalisation.

Dataset specifications
Year2024
Episodes107,000
EmbodimentsFranka Emika Panda, Tien Kung humanoid, AgileX Cobot Magic V2.0, UR-5e
Modalitiesrgb, depth, proprioception
Task categoriesmanipulation, pick-and-place, dexterous-manipulation, bimanual, long-horizon
Data formathdf5, mp4, json
LicenseApache 2.0
Accessopen — commercial use permitted
MaintainerX-Humanoid, Beijing Innovation Center for Humanoid Robotics
Origin countryCN

What is it?

RoboMIND is a comprehensive multi-embodiment robot manipulation dataset featuring 107,000 real-world demonstration trajectories spanning 479 distinct tasks and 96 unique object classes. Released under Apache 2.0 for commercial use, it integrates teleoperation data from four distinct robot platforms: Franka Emika Panda, Tien Kung humanoid, AgileX Cobot Magic V2.0, and UR-5e. The inclusion of the Tien Kung humanoid makes RoboMIND one of the few large-scale datasets with real humanoid manipulation demonstrations.

Who is it for?

Researchers studying cross-embodiment generalisation and multi-task learning at scale. The 479-task coverage enables systematic evaluation of whether policies trained on diverse tasks transfer better than single-task specialists. The humanoid embodiment inclusion is specifically valuable for teams developing humanoid manipulation policies who need real-hardware data beyond the limited humanoid datasets currently available.

Key specifications

How it compares

RoboMIND has the broadest task coverage of any multi-embodiment manipulation dataset (479 tasks vs Open X-Embodiment's diverse but less structured task set). The humanoid embodiment inclusion is shared only with AgiBot-Alpha (1,979 hours) among large open datasets. RoboMIND's Apache 2.0 license gives it a commercial advantage over AgiBot-Alpha's more restrictive academic license.

Limitations and access notes

Chinese-origin dataset — geopolitical flag applies for IP awareness. Apache 2.0 permits unrestricted commercial use with no restrictions.

Linked professions

Frequently asked questions

What makes RoboMIND different from Open X-Embodiment?

RoboMIND provides 479 clearly distinct manipulation tasks with 107K trajectories across 4 embodiments including a humanoid. Open X-Embodiment covers 22 embodiments with 1M+ episodes but with less structured task organisation. RoboMIND is stronger for systematic multi-task generalisation research; Open X-Embodiment is stronger for cross-embodiment scale.

Which humanoid robot is in RoboMIND?

RoboMIND includes data from the Tien Kung humanoid robot alongside Franka Panda, AgileX Cobot Magic V2.0, and UR-5e. Tien Kung is a full-size bipedal humanoid, making RoboMIND one of the few large-scale datasets with real humanoid manipulation demonstrations at this scale.

Can RoboMIND be used commercially?

Yes. Apache 2.0 licensed, permitting unrestricted commercial use, modification, and redistribution.

How do I access RoboMIND?

RoboMIND is available on Hugging Face at huggingface.co/datasets/OpenDriveLab/RoboMIND. No registration required.

What is the AgileX Cobot Magic V2.0?

AgileX Cobot Magic V2.0 is a dual-arm collaborative robot platform from AgileX Robotics, designed for bimanual manipulation research. Its inclusion in RoboMIND alongside single-arm platforms (Franka, UR-5e) and a humanoid enables study of how bimanual coordination capabilities transfer across embodiment types.