An Inverse Optimal Control Approach to Human Motion Modeling

Abstract

In this paper, we present inverse optimal control as a promising approach to transfer biological motions to humanoid robots. Inverse optimal control serves to identify the underlying optimality criteria of human motions from measurements. Based on these results optimal control models are established that can be used to control robot motion. Inverse optimal control problems are hard to solve since they require the simultaneous treatment of a parameter identification problem and an optimal control problem. We propose a bilevel approach to solve inverse optimal control problems which efficiently combines a direct multiple shooting technique for the optimal control problem solution with a derivative free trust region optimization technique to guarantee the match between optimal control problem solution and measurements. We apply inverse optimal control to determine optimality principles of human locomotion path generation to given target positions and orientations, using new motion capture data of human subjects. We show how the established optimal control model can be used to enable the humanoid robot HRP-2 to autonomously generate natural locomotion paths.