UMR unifies humanoid motion retargeting with point clouds

Humanoid imitation learning has been severely bottlenecked by brittle, manual skeleton retargeting, and treating robots as geometric surface envelopes rather than rigid kinematic skeletons addresses a major obstacle in embodied AI. By decoupling motion translation from specific skeletal conventions, the framework allows arbitrary human motion capture to map onto diverse humanoid morphologies without manual keypoint tuning. Furthermore, preserving surface-level geometric alignment maintains fine-grained contact dynamics across limbs and objects, enabling humanoid robotics to tap into massive human motion capture datasets to scale foundation policy pretraining.