In this paper, the authors describe how they built a system for online, real-time puppetry. In building such a system they overcame several challenges:

One is the adaptation, or retargeting, of the motion of the actor to fit a skeleton that may have slightly different sized bones. The major difficulty is the decision between preserving joint angles (and thus posture) or end-effector positions. The latter is considered important when an actor is interacting with the environment, so they present a solution that variably preserves position depending on how close to an object a particular end effector is.

Though Gleicher gave a solution to this problem in an earlier paper, the requirement that this be performed on-line without the benefit of future knowledge made for a harder problem.

Another challenge is that the data itself could be noisy. They propose using a Kalman filter to smooth out noisy data, and show how it's ability to predict future input data is useful for online filtering.

Finally, since the character must be updated quickly, they give a IK solver that employs closed-form solutions whenever possible -- and so is reasonably fast.