Old graphics paper, but relevant to my interests. Main contribution at a high level is that there are some motions that the current controller does poorly in, so it proposes a set of tasks to the user, which must identify one to improve. The user does motion capture by literally moving around in physical space.
This is faster than trying to learn how to do this without selective sampling.