Unmanned surface vehicle navigation through generative adversarial imitation learning

Abstract

In the artificial intelligent and big data technology era, the marine industry among others is inevitably developing in this direction, aiming at becoming autonomous and completing tasks without relying on human involvement while providing safety. The technology of small unmanned surface vehicles (USVs) is relatively mature but with a large development potential and wide research interest expecting significant benefits such as safety and high efficiency in shipping and transportation systems. This article addresses these issues and utilizes an imitation learning algorithm to resolve autonomous navigation for USVs even in complex environmental conditions. We formulate the trajectory modeling as a data-driven imitation learning problem where we employ a state of the art imitation learning algorithm. Experiments are performed in a particular simulated environment tailored to match the specific weather conditions of the local area. The simulation results show the potential of the proposed imitation learning scheme to create advanced intelligent agents for USVs under real-world environmental settings, and USV actuation constraints that allow to predict trajectories with high accuracy and safety.In addition, we evaluated the method’s robustness in generating successful trajectories under environmental conditions that differed from those encountered during training, thereby promoting knowledge reusing without the need for retraining.