ICML-98 Submission #179

Learning the Grammar of Dance

Joshua Stuart and Elizabeth Bradley

Department of Computer Science
University of Colorado
Boulder, Colorado, USA 80309-0430

Abstract

Human motion sequences that are generated by computer algorithms may
contain abrupt transitions: places where consecutive body positions
would require physically impossible or stylistically illegal moves.
We use graph-theoretic methods to learn the ``grammar'' of joint
movements in a given corpus and then apply memory-bounded A* search to
the resulting transition graphs --- using an influence diagram that
captures the topology of the human body in order to reduce the search
space --- to find appropriate interpolation sequences.  The
application that motivated the development of these methods is an
algorithm that uses the mathematical properties of chaos to generate
variations on dance and martial arts sequences.  Chaos's {\sl
sensitive dependence on initial conditions} introduces abrupt
transitions in these variations, and the goal of the intelligent
interpolation schemes described here is to smooth those transitions in
a kinesiologically and stylistically consistent manner.


Keywords: animation, interpolation, corpus-based learning techniques

Contact author: lizb@cs.colorado.edu; (303)492-5355