As I understand it, the program tells the robot where to move his foot, and inverse kinematics compute how the robot should bend its knees, etc accordingly.
As I understand it (again- not sure), L2 can do this--I think there was a blog post with Turner sitting at a table grabbing objects with inverse cinematics. But it also does something else: where a foot should go is not entirely predetermined - it depends on physics (for instance, compensating gravity center to maintain balance while running). Everything is based on physics, so perhaps things like speed, style, strength, terrain inclination, opponent weight, might influence the animation - and I guess the moves it remains realistic in all circumstances (like if it's interrupted or hindered by some object). (I kinda extrapolate from what David and Jeff said, I don't really know).
How it all works together exactly, and how ragdoll, inverse kinematics, stored animations mix together, I don't think David has said.
But it's much more than just inverse kinematics.
It's sicker.
? HUH? (That did not sound right if you're the right type of person 
