Extra Data
Humanoid robots should study to steadiness, transfer, and work together with objects throughout an unlimited vary of conditions, however the knowledge out there for this coaching has clear limits. Web video exhibits various habits however can not seize exact bodily states. Laboratory movement seize programs file correct motion however normally cowl solely a slender set of actions. This white paper examines HiPHI, a 617.5-hour whole-body human movement dataset captured with optical movement seize at sub-millimeter accuracy. It contains 245.7 hours of human-object interplay with synchronized object trajectories and meshes, and organizes protection utilizing FrameNet, a linguistic framework for human motion. The paper additionally introduces a benchmark suite for measuring movement variety and interplay grounding, and experiences outcomes from insurance policies skilled on the dataset and deployed on a bodily Unitree G1 humanoid robotic.
