Optimal robot motions for physical criteria

This paper presents an optimization-based framework for emulating the low-level capabilities of human motor coordination and learning. Our approach rests on the observation that in most biological motor learning scenarios some form of optimization with respect to a physical criterion is taking place. By appealing to techniques from the theory of Lie groups, we are able to formulate the equations of motion of complex multibody systems in such a way that the resulting optimization problems can be solved reliably and efficiently-the key lies in the ability to compute exact analytic gradients of the objective function without resorting to numerical approximations. The methodology is illustrated via a wide range of optimized, "natural" motions for robots performing various humanlike tasks-for example, power lifting, diving, and gymnastics. (C) 2001 John Wiley & Sons, Inc.