Exploiting Vestibular Output during Learning Results in Naturally Curved Reaching Trajectories

Abstract

Teaching a humanoid robot to reach for a visual target is a complex problem in part because of the high dimensionality of the control space. In this paper, we demonstrate a biologically plausible simplification of the reaching process that replaces the degrees of freedom in the neck of the robot with sensory readings from a vestibular system. We show that this simplification introduces errors that are easily overcome by a standard learning algorithm. Furthermore, the errors that are necessarily introduced by this simplification result in reaching trajectories that are curved in the same way as human reaching trajectories.