The reason why is that mesh vs mesh essentially should never be used in video games. The accuracy in hitbox bounds and ease of use (making a model mesh your collision mesh) you get from mesh collisions are completely and utterly overshadowed by the performance and lack of accuracy in determining the time of impact. At 60 fps it's really hard to tell the difference in collision between a typical dodecahedron and a sphere. Just use a sphere! (Or, more likely, some aggregate of AABBs, spheres and capsules)

Oh yea, for a game I would almost certainly decompose into some sort of union of easily handled convex shapes. There are definitely applications -- robotics, engineering, hell, even film -- where more accurate mesh-mesh is needed, however, and most solutions are super domain specific and ad hoc.

Haha, I keep forgetting that there are other applications for collision detection than just video games :)

> The reason why is that mesh vs mesh essentially should never be used in video games

Or, alternatively, the reason why mesh vs mesh is never used in video games is that it hasn't been solved well yet and we have to stick to simplified collision geometry and the resulting glitches we get in games?

Mesh vs mesh collisions will never be faster than primitive collisions. Also, simplified collision geometries do not produce glitches in the same way mesh vs mesh collisions do. At most you may have a collision between two primitives where the models don't actually seem to be touching. But this can easily be circumvented. In fact, I would argue that a lot of the physics glitches we've seen are probably the result of using mesh colliders when a simplified geometry would have sufficed.

Check the DOOM 3 source code, they have a pretty good mesh-vs-mesh solver.