Matthias Scheutz and Thomas Naselaris
In this paper we describe the vision system of a robot which has to accomplish a path following task. The vision system combines three different learning methods: reinforced, competitive, and supervised. A genetic algorithm is used to determine the regions of the visual field essential for discriminating the path from its surroundings. These regions, in turn, serve as input for a neural network that categorizes the information present in the visual stimulus. The output of the network is used by a motor system to navigate the robot on the path.