Event

EmbodyHOI introduces embodiment reachability constraints during generation

1 reports · 1 sources

Chaitanya Chawla et al. propose EmbodyHOI, which runs a differentiable inverse kinematics solver for a specific humanoid embodiment inside the sampling loop of a frozen flow-matching hand-object interaction trajectory generator, placing embodiment reachability constraints into the generation process rather than performing post-hoc retargeting. Because the guidance acts only at inference time, a generator trained on human hand poses can be reused across humanoid embodiments without retraining. On ARCTIC trajectories with two humanoid embodiments, Unitree G1 and H1-2, the method reduces retargeting wrist error from 8.9 cm to 0.6 cm without harming text-to-hand-object interaction generation quality; task success rate rises from 68% to 79% on 50 random G1 tasks and from 62% to 76% on a four-task suite for H1-2. The authors state that post-hoc wrist refinement and best-of-N selection with comparable compute can achieve lower IK energy, but the trained policies perform worse.