This paper introduces a robust collision detection framework for SDF-encoded geometry, which has seen widespread adoption as a geometric representation in computer graphics. Our approach is designed to satisfy the competing requirements of numerical stability and computational efficiency of interactive and real-time applications. Our collision management strategy pairs temporal coherence caching with curvature-guided importance sampling. This scheme reliably forms sparse, stable contact manifolds, even for complex non-convex geometries. Quantitative evaluations against existing baselines demonstrate significant improvements in both runtime performance and simulation stability. While compatible with a range of numerical backends, our framework demonstrates how specific solver choices, such as derivative-free optimizers, yield additional gains in speed and convergence stability. Furthermore, the algorithm is highly parallelizable and well-suited for GPU architectures, scaling effectively to complex environments containing thousands of rigid bodies while maintaining real-time frame rates. Our proposed framework is validated through comparisons on complex simulation scenes and ablation studies examining the influence of various optimizers and sampling strategies.
BibTeX
@inproceedings{Giles2026,
author = {Giles, Chris and Andrews, Sheldon},
title = {Real-Time Collision Handling for Signed Distance Fields via Caching and Importance Sampling},
year = {2026},
booktitle = {Proc. of the 34th Pacific Conference on Computer Graphics and Applications},
doi = {}
}