FreeMo: Motion Generation with Structured Joint-Collision Energy

Yifei Yang, Zikai Huang, Wenxi Liu, Shengfeng He

Keywords: motion generation

TL;DR: The first self-collision-aware motion generator.

Abstract: In this paper, we present FreeMo, a motion generation framework that produces physically plausible human motion by explicitly addressing self-collisions, where body parts intersect in unrealistic ways. Existing physics-aware generation models primarily handle external interactions, such as foot-ground contact, but are not capable of managing internal body dynamics. Although self-collisions can be corrected using post-hoc methods, these approaches are computationally expensive, difficult to scale, and compromise the differentiability and editability of the generative process. FreeMo integrates structured spatiotemporal constraints into the diffusion sampling process through a differentiable trajectory-level energy function that detects and penalizes persistent joint-level collisions. By directly optimizing joint positions in the latent space, FreeMo guides the generation away from physically implausible motions without compromising semantic alignment or motion naturalness. Experimental results show that FreeMo consistently reduces self-collisions while maintaining high-quality, controllable, and efficient motion synthesis.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7065