A humanoid motion-retargeting method directs extra optimization effort to keyframes and selected body parts, first reported by Arxiv. That targets a limitation of uniform per-frame budgets for object-interaction tasks; experiments found better geometric contact coverage and continuity at lower computational cost, with resulting references improving downstream task completion across diverse object-interaction motions.
Motion retargeting converts human demonstrations into reference trajectories for whole-body control and loco-manipulation. In object-interaction tasks, execution depends on preserving body–object relations at critical stages, while uniform per-frame iteration budgets do not account for those localized accuracy requirements.
Called Spatiotemporal Retargeting (STR), the method uses a vision-language model to select relevant motion signals and body parts. Changes in the selected signals identify keyframes; STR assigns additional solver iterations to those keyframes and higher interaction weights to the selected parts.
Mechanics-Aware Contact Anchoring (MCA) refines object-surface contact positions using an approximate object-wrench model while retaining the selected surface directions. For downstream control, Spatiotemporal Reinforcement Learning (STRL) combines adaptive key-stage sampling with a contact reward based on target proximity and object contact forces.
The components address different parts of the method: STR allocates optimization effort across keyframes and body parts, while MCA refines object-surface contact positions.
Simulation experiments on a Unitree G1 and real-world experiments on a Unitree R1 found that STR improved geometric contact coverage and continuity at lower computational cost. The resulting references improved downstream task completion across diverse object-interaction motions.