Development of a Visually-Guided Autonomous Underwater Vehicle

Abstract

We are developing autonomous underwater vehicles for exploration and inspection tasks. Our objectives are to enable these submersible robots to autonomously search in regular patterns, follow along fixed natural and artificial features, and swim after dynamic targets. We have developed a method of visually-guiding autonomous land vehicles using correlation-based feature tracking to follow targets of interest. We have used this method to fixate on visual features and generate motion commands for the robot. We propose to use feedback from the motion of the visual feature as reinforcement in a network that learns stable control. We are now applying these techniques to the control of our underwater vehicle, Kambara.