Path planning with neural subgoal search

Abstract

A system for robot path planning is presented, that is able to find useful and efficient subgoals in an arbitrary environment. The system consists of two pairs of separately trained networks and an underlying layer of learning units. The network's training is completely based on the most elementary sensoric informations. The created solutions in two and three dimensional simulation environments prove the networks capability to build up a meaningful world model that is effectively applied to the tasks.<>