Collision detection for physics / ray tracing for arbitrary meshes is slow, unless they are pre-processed as SDF or decomposed into convex parts. However, there was no suitable solution for animated meshes until now
We are developing a Houdini plugin for approximate decomposition of the animated mesh into non-rigid convex parts. Animations were pre-processed (trained) in 30s - 4min (it could be more for more detailed animations). The approximation error can be chosen afterwards interactively (0-2s). The method is robust to topologically incoherent meshes (degenerate triangles, open meshes, self-intersections, etc), which commonly happened in all these animations. The model can be trained for one video (or a set of videos) and evaluated on another, if vertices correspond to each other (for example a rigged character for a game or a movie). Evaluation runs in real-time.
I don’t know how desired this is (simpler solutions are typically used, such as predefined set of capsules), but results can be used in real-time applications. We originally developed the decomposition for our fluid solver ( Physics Forests - A New Real-time Fluid Solver - Download demo NOW! - Game Development - Epic Developer Community Forums ), which is still under progress:)