Neural Network Architecture for Robot Hand Control

A robot hand control system, called GeSAM, is under development at the University of Southern California. The goal of the GeSAM architecture is to provide a generic robot hand controller that is based on a model of human prehensile function. It focuses on the relationship between geometric object primitives and the ways a hand can perform prehensile behaviors. This paper shows how the relationship between object primitives and a useful set of grasp modes can be learned by an adaptive neural network. By adding training points as necessary, system performance can be improved, avoiding the tedious job of computing every relationship individually.