SemAnCorr: Semantic Anchored Correspondence for Zero-Shot Manipulation Skill Transfer

Researchers introduced SemAnCorr, a training-free method that enables zero-shot robot skill transfer by finding semantic and geometric correspondences across object instances with the same function but different shapes.

The method combines semantic anchor regions, pose-correspondence optimization, and functional maps to build dense correspondences between objects.

On the PartNet-Mobility benchmark, SemAnCorr achieved 90.8% semantic accuracy and stronger geometric coherence than prior methods.

It also improved real-world manipulation transfer from a single demonstration, reducing the need for retraining or large demonstration datasets.