Motion Control of Loitering Munition Based on Deep Reinforcement Learning in Autonomous Evasive Attack Task

Loitering munitions are important weapons in modern warfare, but their motion control still relies on manual command, lacking intelligent autonomous decision-making capabilities. In this paper, we propose a motion control scheme that utilizes Deep Reinforcement Learning (DRL), aiming at autonomous evasive attack on a maritime target using a loitering munition. This task requires accurate strike on target ship while evading interceptors and operates in a three-dimensional combat scenario, increasing the search space. To optimize the strategy learning efficiency and achieve superior strategies, this paper introduces a Beta strategy-based Proximal Policy Optimization (BetaPPO) algorithm that exploits the unbiased nature of the beta distribution over the strategy gradient to improve the learning efficiency. The combat entities and task environment are first modeled and then we design state information to represent the battlefield posture and the control actions of the loitering munition. The real-time detection of loitering munition is combined with domain knowledge rules to construct a multi-level reward. Simulation results show that the proposed BetaPPO algorithm outperforms DQN, DDPG, and PPO algorithms with a higher task success rate.