Pursuit-Evasion Game for Underactuated ASVs: An Online Deep Learning Strategy

An online learning-based pursuit-evasion game (PEG) strategy is proposed for underactuated autonomous surface vehicles (ASVs). Unlike conventional PEG strategies, the PEG problem is redefined in the differential game framework, considering ASV underactuation characteristics, physical constraints, and unknown system dynamics. Since underactuated systems have fewer control inputs than degrees of freedom, traditional PEG strategies cannot be directly applied to underactuated ASVs. Within the framework of differential games, a gradient-driven velocity control law is proposed for the underactuated pursuers, utilising optimisation techniques. Consequently, the complex PEG problem is transformed into a solvable optimal control problem. Dynamic task allocation for multi-ASV systems is achieved by integrating the Hungarian algorithm. A cooperative pursuit-evasion control strategy with a priority switching mechanism is designed, which can effectively balance collision avoidance and target capture requirements. An online learning module is constructed based on the gated recurrent units (GRUs), which enables real-time learning of unknown ASV dynamics and effectively captures temporal dependencies. The effectiveness of the proposed strategy is verified through the simulation cases and hardware-in-the-loop (HIL) experiment.