UMR learns dense point-cloud correspondences in canonical poses and binds matched points to the human and robot meshes. Constrained geometric optimization then aligns motion while transferring contact information. This separates the retargeting interface from the original motion format and the target robot topology.
The framework supports motion-data preparation for locomotion, object interaction, and contact with the surrounding scene. Its Robot Retargeting Studio exposes an interactive workflow for configuring assets and applying reference motion. Released code also supports batch processing and visualization for research pipelines.
