Lara S. Crawford, Graduate Student
(Professor S. Shankar Sastry)
One of the amazing successes of biological systems is animals' ability to learn to control the complicated dynamics of their muscles and joints smoothly and efficiently. Traditional control techniques often do not perform well when confronted with intrinsically complex dynamical systems such as are present in human (or robot) behaviors like running or diving. In these systems, the control goal is not to follow some specific trajectory, but rather to satisfy some criterion such as "stay balanced while moving forward," or "execute a one-and-a-half somersault dive with a full twist." The goal of this project is to generate new approaches to the control of such systems using hierarchical and synergistic organization of biologically-inspired learning controllers.
In our control design, subtasks at the level of individual joints are dealt with by low-level controllers, which are coordinated by a higher-level controller. The system learns the best control signal for a particular task by searching among parameterizations of biologically plausible control families. The control families we have chosen are based on the controls used by biological pattern generators for fast, goal-directed movements. The learning controllers themselves are networks of radial basis functions. We are using a dynamic simulation of a 25 DOF human diver as a platform on which to train the networks and test our control and learning strategies. Our work on this controller will be presented in a workshop at the upcoming Neural and Information Processing Systems 1996.