Interactive Simulation of Goal-Driven Behavior

Abstract

Physical simulation is a powerful tool for evaluating the design of robots and vehicles, training people in medicine and emergency scenarios, and synthesizing motions in virtual worlds. However, simulating goal-driven behavior such as the activities of virtual humans, robots, or other creatures is challenging. In addition to accurately modeling system dynamics, interactive simulation of a goal-driven behavior requires a motion controller that efficiently decides how to activate muscles to produce coordinated movement.

In this talk, I will describe an approach to creating a controller from a single reference motion such as a motion capture recording. The controller adapts that motion to new dynamic environments with varied terrain or random disturbances. I will also describe a method to generalize controllers to new tasks. The method linearly combines multiple existing controllers to create a new controller that is optimal for a new task. I will demonstrate the application of these methods to human motions such as stepping in various styles and jumping. Finally, I will present three research objectives that will improve simulation as a tool for studying and animating goal-driven behaviors.