A Method for Multi-AUV Cooperative Area Search in Unknown Environment Based on Reinforcement Learning

Abstract

As an emerging direction of multi-agent collaborative control technology, multiple autonomous underwater vehicle (multi-AUV) cooperative area search technology has played an important role in civilian fields such as marine resource exploration and development, marine rescue, and marine scientific expeditions, as well as in military fields such as mine countermeasures and military underwater reconnaissance. At present, as we continue to explore the ocean, the environment in which AUVs perform search tasks is mostly unknown, with many uncertainties such as obstacles, which places high demands on the autonomous decision-making capabilities of AUVs. Moreover, considering the limited detection capability of a single AUV in underwater environments, while the area searched by the AUV is constantly expanding, a single AUV cannot obtain global state information in real time and can only make behavioral decisions based on local observation information, which adversely affects the coordination between AUVs and the search efficiency of multi-AUV systems. Therefore, in order to face increasingly challenging search tasks, we adopt multi-agent reinforcement learning (MARL) to study the problem of multi-AUV cooperative area search from the perspective of improving autonomous decision-making capabilities and collaboration between AUVs. First, we modeled the search task as a decentralized partial observation Markov decision process (Dec-POMDP) and established a search information map. Each AUV updates the information map based on sonar detection information and information fusion between AUVs, and makes real-time decisions based on this to better address the problem of insufficient observation information caused by the weak perception ability of AUVs in underwater environments. Secondly, we established a multi-AUV cooperative area search system (MACASS), which employs a search strategy based on multi-agent reinforcement learning. The system combines various AUVs into a unified entity using a distributed control approach. During the execution of search tasks, each AUV can make action decisions based on sonar detection information and information exchange among AUVs in the system, utilizing the MARL-based search strategy. As a result, AUVs possess enhanced autonomy in decision-making, enabling them to better handle challenges such as limited detection capabilities and insufficient observational information.