关键词:
cooperative pursuit
multi-agent soft actor-critic
multi-stage reward
unmanned surface vehicle (USV)
摘要:
This paper is concerned with the cooperative pursuit of unmanned surface vehicles (USVs) against the dynamic escaping target using multi-agent reinforcement learning. The Markov game process is established for pursuit-evasion, and the success criteria for cooperative capture of USVs are given by using distance and angle constraints. By virtue of the centralized training and decentralized execution framework as well as the long short-term memory network, cooperative pursuit training is conducted using the multi-agent soft actor-critic reinforcement learning, which can optimize capture performance of USVs against the escaping target. Besides, to avoid the occurrence of lazy capturer and increase the capture success rate, a multi-stage reward guidance method is developed, where the training process can be optimized according to the current states of both sides, effectively guiding vehicle to achieve the capture task from easy to difficult. Simulations are provided to illustrate the effectiveness of the proposed reinforcement learning method for cooperative pursuit of USVs. © Shanghai Jiao Tong University 2025.