Enhancing autonomous pursuit-evasion with targeted navigation in UAVs using asynchronous multi-stage deep reinforcement learning

Abstract

This dissertation explores the application of reinforcement learning (RL) to the multi-agent pursuit-evasion with targeted navigation (MPETN) problem, focusing on Unmanned Aerial Vehicles (UAVs) operating in environments with obstacles. The study improves upon an existing asynchronous multi-stage deep reinforcement learning (AMS-DRL) algorithm by enhancing the observation space and optimizing the reward mechanism to better accommodate dynamic and static obstacles. The proposed algorithm trains pursuer and evader UAVs asynchronously, allowing the system to adaptively learn optimal strategies for both pursuit and evasion. A high-fidelity simulation environment built in Unity is used for training, where drones face various obstacles while pursuing a target. The results demonstrate that the improved AMS-DRL algorithm maintains excellent performance in environments containing obstacles. The dissertation also examines the algorithm’s limitations and outlines directions for future research, particularly in handling dynamic obstacles and improving real-world deployment capabilities.