Robot Learning with Neural Self-Organization

In this paper we describe a learning technique for reactive path-planning and guidance of autonomous robots using a Kohonen self-organizing neural network. The intended application is mineral exploration or agricultural monitoring using simple lowcost robots which do not contain a complete stored map of the environment, and may have unreliable sensor information. The network learns based on a simple sense of value, and our experiments with a simulated world show that this performs signi cantly better than rule-based learning.