Brief

Researchers have introduced CoMPAS3D, a new dataset and benchmark designed to evaluate interactive motion generation for humanoid robots. This dataset captures 3 hours of improvised partner salsa dancing from 18 dancers across different proficiency levels, featuring over 2,800 expert-annotated segments. The benchmark includes metrics for kinematic quality, move legibility, and proficiency appropriateness, aiming to assess generated motion's social context understanding beyond simple kinematic accuracy. Initial results show that fine-tuned vision-language models perform well on objective metrics, but human evaluations highlight a gap between generated and ground-truth motion. AI