Ask or search anything...

RoboWheel, developed by researchers from Tsinghua University and Synapath, is a data engine that transforms real-world human hand-object interaction (HOI) videos into high-fidelity, robot-executable supervision for cross-embodiment robotic learning. It achieves significantly lower hand jitter (0.92 cm/s²) and higher real-robot retargeting success (macro average 91.7%) compared to baselines, enabling robot policies trained on HOI data to perform comparably to those using expensive teleoperation.