The title of this paper says it all: they introduce a system for interactive control of 3D motion data. Also studied are three user-control methods and their requirements: 1) by continually choosing the next motion, 2) by sketching a path and 3) by acting a desired motion.

Chief among their requirements is that the database contain extended, unlabeled sequences of motion capture data. They show how by organizing this data for efficient search in a preprocess, they can acheive low-latency control. In this preprocess, they first cluster the motion -- this cluster is then used as a generalization of the original data for efficient search. These clusters are also linked to the original motion, so that they don't lose any fidelity when playing that motion back.