This paper explores and investigates the lack of exploration performance and generalization performance common to multi-intelligence reinforcement learning algorithms during multi-UAV cooperative reconnaissance to investigate the problem. The principle of evolutionary learning is proposed to improve the performance of the algorithms. Unlike traditional deep reinforcement learning, which typically struggles with tasks that have few rewards, evolutionary approaches excel in this context by reducing the risk of premature convergence. The key advantage is the inherent ability of evolutionary methods to incorporate prior knowledge, which significantly improves the algorithm’s search and generalization capabilities. By integrating these evolutionary mechanisms, this research aims to improve the robustness and adaptability of IDQN algorithms. In this study, Airsim is used as a simulation experiment environment to meet the requirements of complex dynamic environments, and the experimental results show that evolutionary reinforcement learning effectively improves the performance of UAV model reconnaissance and achieves more effective decision-making in complex dynamic environments.