humanoid dexterous manipulationExpert
EmbodyHOI: Embodiment-Aware Generation of Hand-Object Interactions for Humanoid Dexterous Manipulation
Carnegie Mellon researchers report in a preprint that steering a human hand-motion generator toward what a specific humanoid arm can actually reach raises the success of the robot's learned manipulation policies from 68% to 79% on 50 random tasks with a Unitree G1, and from 62% to 76% on a Unitree H1-2, all in simulation. The fix works at generation time, so one generator trained on human hands serves different robots.