Surely, the model could go from a motion-captured stance / action (e.g. attempting to land a blow), to a ragdoll model, inheriting the dynamics of the former for a period of time?
Mixing the ragdoll and animations is a much harder task then it might initially seem. Each animation relies on a set of keyframes, and if you are not in one of the ‘start’ keyframe states, any move an AI will make will either have to be computed dynamically (basically eliminating the animation to begin with), or will have to take shortcuts causing a very poor looking string of frames. For example, how would you generate a ‘get up’ animation if you dont know the position of the (ragdoll) model after it has fallen down? (games to date dont generally have people getting back up after they go rag-dolley).