Gen-HumanWholeBody

Gen-HumanWholeBody Dataset Card

Dataset Summary

Gen-HumanWholeBody is a high-precision whole-body manipulation dataset captured from DAS-Ego, released by GenRobot.AI. It contains 600 hours of continuous whole-body human manipulation data reconstructed via GenRobot's proprietary Data Foundation Model (DFM), with an average joint position error of approximately 3 cm.

The dataset covers 16,976 real-world scenarios and 1,281 distinct human skills, including tasks that require not only two-handed manipulation but also locomotion, body posture adjustment, and whole-body coordination.

By sharing this dataset, GenRobot aims to provide Physical AI researchers with training-ready whole-body manipulation data and to catalyze a scalable data production pipeline based on lightweight egocentric devices.

Dataset at a Glance

Supported Tasks and Leaderboards

Gen-HumanWholeBody is designed to support research on:

• Whole-Body Motion Learning: Learning continuous and coordinated whole-body motion.
• Human-to-Robot Transfer: Exploring skill transfer from human demonstration to robot control.
• Embodied Foundation Models: Providing rich real-world interaction data for embodied foundation model pre-training.

The multi-modal signals in the dataset—whole-body motion, two-handed operations, object interactions, contact states, visual observations, and task semantics—are unified in a shared spatiotemporal coordinate system, making the data directly usable for robot policy training.

Precision Metrics

Quantitative evaluation against ground truth:

The DFM has been validated on challenging conditions including bending, fast motion, long-horizon locomotion, and partial occlusion.

Contact

Questions, suggestions, and requests for additional scenarios or skills are welcome.

• X: @GenrobotAI
• LinkedIn: GenRobot
• Email: [email protected]
• Discord: https://discord.gg/rSSb5thgu