View-based route-learning with self-organizing neural networks

This paper describes a view-based mobile robot navigation system relying on self-organizing neural networks. Route navigation was presumed to consist of a chain of view-action associations: A sequence of view images from a test route was obtained, pre-processed, and used to train a system of self-organizing maps. The converged networks consisted of a set of weights representing the learned views and a set encoding the actions to be carried out at those views. A view presented to the trained networks can thus associatively elicit the action coupled to it, allowing autonomous execution of the route. The data were presented to the system of networks using a simple place-dependent scheme, and a context-sensitive decision-maker was used to minimize potential recognition ambiguity during the execution stage. (C) 2001 Elsevier Science B.V. All rights reserved.