Trajectory Optimization and Physical Realism

An animated human figure seeking the optimal path from point A to point B typically relies on computationally expensive hard constraints that force the trajectories to be physically realistic. But contact-invariant optimization (CIO), as applied by Igor Mordatch, a graduate student in computer science at the University of Washington, can achieve physical realism more efficiently by changing the contact forces from binary (touching/not touching, which numerical optimizers can’t handle in a smooth way) to a softer constraint that is more like a guideline.  “It’s like you have a jet-pack on your hands or feet,” Mordatch says. As the optimization proceeds, it discovers for itself that the contact/no contact solution is optimal, while still preserving the physical realism of a smooth transition. “The gradual transition between contact and non-contact makes sure the numerical behavior is nice,” he says. “That’s kind of the primary trick.”

Mordatch has used the approach to create animated figures that can stand from a prone position, do handstands, climb over walls, and pass objects. More recently, he has been adding physics-based muscle models in an effort to make the work useful for biomechanics researchers. He envisions a two-step process in which the simple models achieve the general motion that is then refined with a full physics-based model. “We haven’t really tried that yet,” he says. “It’s exciting stuff for the future.”

Full movies of Mordatch's work are viewable at http://homes.cs.washington.edu/~mordatch/.