UMR: open-source unified motion retargeting for humanoids via learned point-cloud correspondence

ChongZzZhang · x · 2026-09-11

Researchers from HKUST(GZ), Noitom Robotics, Hanyang University and HKU released UMR (Unified Motion Retargeting), an open-source framework for transferring human motion to humanoid robots. Instead of hand-crafted sparse keypoints or body-part mappings, UMR treats exterior point clouds as a unified interface and learns dense point-cloud correspondence, decoupling retargeting from source-specific skeletal semantics and robot topology. The dense correspondence provides fine-grained geometric anchors for constrained matching optimization, enabling surface-level pose alignment and direct transfer of interaction contacts. Experiments show it unifies retargeting across heterogeneous motion sources, robot embodiments, and downstream tasks from locomotion to contact-rich interaction, with higher motion fidelity. Paper, UMR Studio, and code are public.

Original post →

More from Embodied

Embodied channel →