S4R: Scaling for Rigid-Body Interpenetration Resolution

arXiv cs.GR details S4R, a scale-continuation method for static rigid-body interpenetration repair in procedurally assembled or generated scenes. The core idea is simple but practical: shrink each body to a collision-free starting scale, then grow back to full size through a sequence of minimum-norm convex contact QPs that solve a linearized separation problem along the way.

For developers building simulation-heavy tools, scene assembly pipelines, or generative content workflows, the appeal is in how S4R turns one hard deep-correction problem into many smaller shallow-contact steps. The method also uses a conservative scale-event bound and frozen-witness gap predictions to reduce exact mesh queries, then finishes with a full-scale evaluator check and bounded tail refinement.

The paper evaluates S4R on Kubric, HY3D-Bench, and Thingi10K with a shared mesh-level evaluator and unified timing protocol. In the main comparisons, it reports zero penetration up to 5,000 bodies, with displacement that stays small and nearly independent of scene size, while also landing at the lowest wall time within each hardware tier among the compared methods.

A GPU implementation extends the approach to larger scenes, which makes the work especially relevant for teams trying to keep authored or generated layouts physically plausible before handing them off to simulation, baking, or downstream runtime systems.