http://www.cs.unc.edu/~geom/PDG/pdg.pdf
Unfortunately this paper, and other UNC papers make some dubious claim about EPA:
Bergen's expanding polytope algorithms is exact, not just an approximation. Bullet also implements and uses this EPA algorithm by default, and it seems to be a very fast and a reliable penetration depth method for general convex objects (not just polyhedra). The penetration depth calculated by EPA is exact, within a user-defined epsilon.Bergen proposes a quick lower bound estimation to PDt between two convex polytopes by iteratively expanding a polyhedral approximation of the Minkowski sum [van den Bergen 2001].
I'm anxious to learn about the details about this claim that EPA is just a 'quick approximation', rather then a 'very good penetration depth calculation'. When I asked Gino van den Bergen recently, he confirmed that EPA is an exact method.