Reinforcement Learning-Based Target Interception for USVs with Experimental Validation

Intercepting hostile targets in complex scenarios is a crucial task for unmanned surface vehicles (USVs). This paper proposes an interception algorithm based on deep reinforcement learning (DRL) and tests it in a real-world environment. This work establishes a DRL framework in a virtual environment based on the Deep Deterministic Policy Gradient (DDPG). A simplified state and a composite reward function are designed for the tasks involving multiple objectives, while the attacker is navigated using an artificial potential field (APF). The proposed policy shows an interception success rate of over 99% against evading attackers in the simulation environment. By constructing an interception hardware architecture based on Matlab and high-precision motion capture system, we transferred the policy from virtual to real environments. With the incorporation of randomly varied hydrodynamic parameters, the defender achieved rapid interception under complex disturbances. The interception and obstacle avoidance algorithm in this paper performs well in both simulation and real-world environments.