Title: CoMPAS3D: Establishing a New Standard for Interactive Motion Benchmarking

Abstract: For socially interactive humanoid robots to effectively engage with humans, they must adapt their physical movements in real-time to their partner’s capabilities, intentions, and actions. This necessitates AI models that comprehend not only the mechanics of bodily movement but also the semantic meaning of those movements within a shared social framework. However, current evaluation frameworks for interactive motion generation fall short; they fail to assess whether generated follower motion is legible within a common movement vocabulary or if it is suited to the partner’s skill level. This deficiency stems from two primary issues: existing metrics, such as FID and beat alignment, focus solely on kinematics and cannot evaluate these critical social properties, while available datasets lack the necessary move annotations and diversity in proficiency levels.

We propose salsa as an ideal domain for evaluation due to its improvised, dyadic nature and its reliance on a structured vocabulary and judging criteria that encompass timing, musicality, technique, difficulty, partnering, and originality. To address these challenges, we introduce CoMPAS3D, a comprehensive motion capture dataset paired with a robust evaluation framework. This framework assesses kinematic quality through two objective metrics—move legibility and proficiency appropriateness—as well as six subjective dimensions derived from competition standards.

The CoMPAS3D dataset comprises three hours of improvisational partner salsa, performed by 18 dancers across beginner, intermediate, and professional tiers. It features over 2,800 segments annotated by experts, detailing move types, errors, and stylistic nuances. We establish three distinct benchmarks: move classification (functioning similarly to transcription), proficiency estimation (serving as a fluency assessment), and follower generation (akin to dialogue response).

Our results indicate that fine-tuned vision-language models excel in objective metrics when applied to ground-truth motion sequences. However, when these metrics are applied to models like Duolando and InterGen, they expose failures that traditional kinematic metrics overlook. Furthermore, human evaluations validate the significant disparity between generated motion and ground-truth data. CoMPAS3D, along with its annotations, benchmark code, and baseline results, is publicly accessible.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC