Robot Motion Planning in Unknown Environments Using Neural Networks

Abstract

. We present two approaches to the motion planning problem for car-like robots using an extended Kohonen Self-Organizing Map (SOM). No prior knowledge about the positions of obstacles is assumed. We incrementally build a path from the starting point of the robot towards the goal, using the SOM as a situation-action map. The first approach uses a trial and error strategy to train the SOM. This method is simple but is not always able to escape from dead-end situations. As an improvement a new training-algorithm is proposed that uses edge detection on the visible objects to generate possible motions. Backtracking is used to choose from different possibilities. Experiments show that this new method realizes a considerable increase in performance and speed. 1 Introduction When building autonomous robot-systems an important problem to solve is the motion planning problem: given a robot R and an environment with a set of obstacles, find a path from a given source to a given goal such that R ...