A FAST AND EFFICIENT MODEL FOR LEARNING TO REACH

摘要

This paper proposes a self-supervised model which enables a humanoid robot to learn to reach to visual targets. Only 400 training samples are used to learn a forward kinematic model of the six degree-of-freedom (DOF) arm. The forward model is represented compactly with just 150 hidden neurons and enables high accuracy reaching in real time. We provide an optimization process for the learning parameters and a careful analysis of reaching errors. An extension of the model is presented to address additional DOFs in the neck. The consistency of the model with physiological and psychological observations is elaborated.