Harvard Computational Robotics Group introduces GRASP for physics-based motion retargeting

GRASP aims to give dexterous hands and humanoids more physically workable motion references before policy learning. The Harvard Computational Robotics Group reports pen-spinning and humanoid–object demonstrations. Its hardware tests differ: Sharpa hands replayed trajectories open-loop, while a Unitree G1 ran learned tracking policies. The group’s initial post gives no numerical performance results for those tests.