基于规则操作性PPO的无人船避碰方法研究OA

To address the curse of dimensionality and the challenges of dynamic adaptability in unmanned surface vehicle collision avoidance path planning under complex port environments,a deep reinforcement learning-based collision-avoidance strategy that integrated the international regulations for preventing collisions at sea(COLREGs)and ship maneuverability was proposed,namely RM-PPO.By constructing a reward function framework incorporating rule-based constraints,a dynamically optimized objective system was established for head-on,crossing,and overtaking encounter scenarios,ensuring that the collision-avoidance decisions complied with the maneuvering requirements of COLREGs under different encounter situations.Based on a three-degree-of-freedom ship dynamics model,a continuous action space was designed to achieve the coordinated optimization of propulsion power and rudder angle commands,thereby enhancing the physical realizability of the actions.The network architecture of the proximal policy optimization(PPO)algorithm was improved by introducing a gated recurrent unit(GRU)to enhance temporal decision-making capability,and a rule-flag-based policy generation mechanism was proposed to balance the exploration and stability of the policy.Simulation and experimental results show that the proposed RM-PPO algorithm is capable of achieving efficient and safe collision avoidance in both basic encounter scenarios and complex environments,with significantly superior performance in terms of path length,collision-avoidance distance,and navigation stability compared to conventional reinforcement learning algorithms.