A Bio-Inspired Reinforcement Learning Method for Pursuit-Evasion in Unmanned Surface Vehicles

The pursuit-evasion problem for underactuated unmanned surface vehicles (USVs) in complex three-dimensional (3D) environments with obstacles is particularly challenging due to the limitations in control authority and the dynamics of the environment. This paper proposes a reinforcement learning approach to solve the pursuit-evasion (PE) problem for USVs, where the pursuer must efficiently approach and capture the evader while avoiding obstacles. First, the method designs an observation space based on sensor and controller inputs, enhancing the adaptability of the USV in such environments. Then a biologically-inspired reward function, modeled after shark predation behavior, is introduced to guide the learning process and improve pursuit performance. Finally, simulation results demonstrate that the proposed approach enables the pursuer USV to effectively track and capture the evader in a 3D environment with obstacles, showing its robustness and effectiveness.