What I've gleaned over the years - familiarity with exposed elements in script hooking - descriptions of function and a convo or two with a friend in the game industry that was reviewing such solutions - exact specifics elude me but I do know that their approach involves machine learning in order to give high quality results in complex 3D situations - be that in research, genetic algorithms or a functioning neural network. There's probably much in common between this solution and theirs - but their approach produces direct quality results - it doesn't need 'fix ups' - I think at least the one that shipped with GTA4 - there was some 'fixing' of feet but that was more slope and minor height adjustment.

I believe there's more to the transition between animation and physic aware rigs than just lerping in an animation manager - and different paths for special cases