Connectionist networks for learning coordinated motion in autonomous systems

Abstract

A central problem inherent to autonomous systems is the absence of an external reference frame in which sensory inputs can be interpreted. It is hypothesized that, in natural systems, sensory information is transformed into a consistent internal representation that serves as an internal invariant reference frame. This paper presents a hierarchical connectionist network for learning coordinated motion in an autonomous robot. The robot model used in the adaptation studies consists of three subsystems: an eye-like visual receptor, a head, and an arm. The network contains a hierarchy of adaptive subnetworks for processing sensory information. The performance of the hierarchical system was observed to improve towards an asymptotic value. The performance was found to be one order of magnitude better than that of non-hierarchical systems. This suggests that the intermediate layers may be serving as an internal invariant reference frame for the robot.