Some Artificial Neural Networks for Visual Navigation

Abstract

In this paper we illustrate how we can use artificial neural networks (ANN's) to con struct topologically accurate physical maps of environments from visual observations, and how we can use these maps for navigation. The method we use to build our ANN's is inspired by biological processes, making it more robust than comparable machine methods. Since the nodes, weights, and threshold functions of our ANN's all have physical meanings, we can easily predict network topolo gies and build networks, avoiding traditional trial-and-error training. This makes our method easy to use in constructing solutions to robotics problems.