Path Planning for Underactuated Unmanned Surface Vehicle Swarm Based on Deep Reinforcement Learning

This research addresses the challenge of inadequately reflecting the kinematic constraints inherent in underactuated unmanned surface vehicles (USVs) within existing deep reinforcement learning (DRL) simulation platforms. A virtual ocean simulation platform is constructed using the Unity3D engine, specifically designed to incorporate kinematic constraints in underactuated USV swarm points-to-points path planning scenarios. Subsequently, an enhanced multi-agent deep deterministic policy gradient (MADDPG) algorithm is introduced for effective path decision-making. Finally, the Peaceful-pie toolkit facilitates seamless data interaction between Python and Unity, enabling cost-effective model training. Simulation experiments demonstrate that, in unknown environments, underactuated USV swarms can utilize sensor information to navigate collision-free paths towards designated target points.