Our research is motivated by the problem of teaching a graphical model to perform a realistic motion. This problem has its roots in the art of cartoon animation and has modern applications to computer animation, virtual reality, teleconferencing, and robotics. The task is to endow a graphical model with the knowledge of how to perform a repertoire of interesting motions. Once learned, these motions can be invoked directly via high-level cues (e.g. "smile", "walk"), inferred via an abstract goal (e.g. "go to the store"), or cued off of low-level events from a virtual input device or an image sequence.
Other projects include the analysis of periodic motion and tracking of rigid and nonrigid objects.