Shanghai AI Lab's InfiniHand reconstructs 3D hands from 5,000 hours of egocentric video at 11.19 FPS

量子位 · wechat · 2026-10-14

Shanghai AI Laboratory with Shanghai Jiao Tong University and others propose InfiniHand, a single streaming model that jointly estimates hand position, hand pose and camera trajectory, lifting egocentric hand motions into a shared world coordinate frame to scale up 3D demonstrations for embodied AI.

Method

Data and training

Results

Paper and code are public (arXiv:2609.35743); after quality screening and embodiment adaptation, the pipeline can feed dexterous-hand data augmentation and imitation learning.

Original post →

More from Embodied

Embodied channel →