Learning Control Policies for Grasping Animation
Sheldon Andrews , PhD student, SOCS McGill

Area: computer animation

Abstract:

Synthesizing motions for human grasping tasks remains an open problem in computer animation. In this presentation, I begin with a short overview on character animation using a multibody dynamical simulation. I emphasize aspects which make grasping particularly interesting (and difficult). I then present our method for performing one-handed manipulation of objects. Our approach uses offline simulations to learn control policies for a multiphase controller framework. Parameters for individual controllers are found by minimizing phase appropriate objective functions. For this purpose, a derivative-free numerical optimization method is used. The generated motions are physically plausible and the resulting policy is general enough to be used in real time. I will show some videos of our results and discuss some of the techniques used to improve performance.