Adaptive and extendable control of unmanned surface vehicle formations using distributed deep reinforcement learning

Abstract

Future ocean exploration will be dominated by a large-scale deployment of marine robots such as unmanned surface vehicles (USVs). Without the involvement of human operators, USVs exploit oceans, especially the complex marine environments, in an unprecedented way with an increased mission efficiency. However, current autonomy level of USVs is still limited, and the majority of vessels are being remotely controlled. To address such an issue, artificial intelligence (AI) such as reinforcement learning can effectively equip USVs with high-level intelligence and consequently achieve full autonomous operation. Also, by adopting the concept of multi-agent intelligence, future trend of USV operations is to use them as a formation fleet. Current researches in USV formation control are largely based upon classical control theories such as PID, backstepping and model predictive control methods with the impact by using advanced AI technologies unclear. This paper, therefore, paves the way in this area by proposing a distributed deep reinforcement learning algorithm for USV formations. More importantly, using the proposed algorithm USV formations can learn two critical abilities, i.e. adaptability and extendibility that enable formations to arbitrarily increase the number of USVs or change formation shapes. The effectiveness of algorithms has been verified and validated through a number of computer-based simulations.