SimHum Co-training for Bimanual Robot Manipulation
September 18, 2026
SimHum extracts kinematic priors from simulation and visual priors from human observations to enable data-efficient bimanual manipulation. The method achieves 62.5% success on out-of-distribution tabletop tasks using only 80 real-robot episodes.
HOW THIS AFFECTS YOU
●
researcherYou can use this co-training recipe to bridge the sim-to-real and human-to-robot embodiment gaps.