An arXiv paper proposes spending more computation on the moments when a robot touches or moves an object.
Human-motion retargeting converts a person’s movement into reference motion that a humanoid robot can follow. For object-interaction tasks, small errors in body position or contact can disrupt the whole action.
A paper submitted to arXiv on October 21 proposes Spatiotemporal Retargeting, or STR. Instead of giving every motion frame the same optimization effort, STR uses a vision-language model to select relevant motion signals and body parts. It then identifies keyframes—important moments in the motion—and assigns them extra solver iterations and greater interaction weight, according to the researchers.
The paper also proposes Mechanics-Aware Contact Anchoring. This step refines where an object should contact a surface using an approximate model of the forces acting on the object, while preserving selected surface directions.
A related method, Spatiotemporal Reinforcement Learning, samples important stages adaptively. Its contact reward considers both how close the robot is to its target and the forces produced during contact.
The researchers report simulation experiments on a Unitree G1 and real-world experiments on a Unitree R1. They say STR improved geometric contact coverage and continuity at lower computational cost, while the resulting motion references improved downstream task completion across object-interaction motions.
The abstract does not give the size of those improvements, and this is a research paper—not a product or deployment. The practical question is whether the approach continues to work across more robots, objects, and contact-heavy tasks.