Neural networks for sonar and infrared sensors fusion

Abstract

The main goal of our work is to have a robot navigating in unknown and not-specially-structured environments, and performing delivery-like tasks. This robot has both unreliable sonar and infrared sensors. To cope with the unreliability, a sensor fusion method is needed. The main problem when applying classical fusion methods is that there is no a-priori model of the environment, because the robot first carries out a map-building process. Some simple methods for sensor fusion exist but, as we show, they do not address all the specific issues of our desired robot task. This is why we use neural networks for such fusion, and so we obtain more reliable data. We discuss some important points related to the neural network training procedure and the results we obtained.