Abstract
Autonomous navigation and target search for unmanned aerial vehicles (UAVs) have extensive application potential in search and rescue, surveillance, and environmental monitoring. Reinforcement learning (RL) has demonstrated excellent performance in real-time UAV navigation through dynamic optimization of decision-making strategies, but its application in large-scale environments for target search and obstacle avoidance is still limited by slow convergence and low computational efficiency. To address this issue, a hybrid framework combining RL and artificial potential field (APF) is proposed to improve the target search algorithm. Firstly, a task scenario and training environment for UAV target search are constructed. Secondly, RL is integrated with APF to form a framework that combines global and local strategies. Thirdly, the hybrid framework is compared with standalone RL algorithms through training and analysis of their performance differences. The experimental results demonstrate that the proposed method significantly outperforms standalone RL algorithms in terms of target search efficiency and obstacle avoidance performance. Specifically, the SAC-APF hybrid framework achieves a 161% improvement in success rate compared to the baseline SAC model, increasing from 0.282 to 0.736 in obstacle scenarios.