Abstract
This paper is to propose a new architecture of a self-creating and organizing neural network for workspace recognition and navigation of an autonomous mobile robot. The robot assumed in this study learns its workspace with sonar sensors but without dead reckoning. The network proposed is organized into a graph structure rather than a tree with forgetting old links so that few dead nodes and dead links are created; the space of input signal can be covered efficiently by unsupervised learning. Methods of path planning and navigation based on a topological map created by learning has been proposed as well. The proposed architecture was tested by simulation of an autonomous mobile robot with eight sonar sensors, and it was demonstrated that the architecture is useful for the purpose.
