For 3D collisions, most physics engines like to work with convex shapes, so this presents a problem when you want to use arbitrary 3D meshes (
which may be concave).
The solution is a process called "convex decomposition", which takes a concave shape and splits it up into a collection of convex ones.
Many physics engines have tools where you can provide a 3D mesh (e.g. a visual model) and have it be decomposed into a collection of convex polyhedra, which can then be used by the physics engine.
However, on most games that I've worked on, as an optimisation, the artists have always created a 2nd low-poly mesh specifically for physics, because it's wasteful to have perfect collisions
The Sui Generis guy mentioned that he wants to save as much development time as possible, so he's using the actual visual meshes instead of hand-crafting optimised collision meshes. As long as his convex-polyhedra collision code is fast, then it won't be too much of a problem for him.
