Abstract
Physical simulation of character behavior is a powerful tool for animating virtual worlds, evaluating robotic designs, and testing bio-mechanical hypotheses. However, simulating these behaviors is difficult for three reasons. First, given a description of a task, we must derive a controller that manipulates forces to achieve that task for a non-linear, non-smooth dynamical system. Second, even when we succeed in designing one controller for a task, the controller may only succeed within a small subset of the character's high-dimensional state space. Finally, many methods do not scale well to the task of controlling high-dimensional characters that are typical in character simulation.

In this talk, I will describe methods to derive controllers for new tasks from motion capture data and existing controllers. I will also describe how to create a controller robust for a larger subset of state space by combining multiple controllers. Finally, I will describe how to apply these techniques to high-dimensional characters needed for many practical scenarios. This work makes it possible to simulate human locomotion activities in various styles in new and challenging dynamic environments. There are still many behaviors that cannot be simulated well. I will outline some next steps for simulating more behaviors more robustly. First, by analyzing more diverse data such as swimming and other animal locomotion, we could begin to simulate more natural behaviors. Second, using model reduction approaches would allow us to apply robust control methods that do not scale well to high-dimensional problems. I believe this research agenda will make simulation of character behavior as powerful as simulation of passive dynamical systems is today.