Multi-UAV Cooperative Pursuit Based on Auxiliary Target Prediction and Curiosity-Inspired Hybrid Intrinsic Rewards

To address the passive trailing behavior induced by target maneuverability and the difficulty of multi-agent credit assignment in multi-UAV cooperative pursuit, this paper proposes a multi-agent reinforcement learning algorithm that integrates auxiliary target prediction and curiosity-inspired hybrid intrinsic rewards, namely AP-HIR-MADDPG. The proposed algorithm embeds a target position prediction auxiliary task into the centralized critic network, enabling the shared feature layer to capture the target's motion tendency and thereby alleviating passive trailing behavior. Meanwhile, an intrinsic reward mechanism that combines target position prediction error with counterfactual evaluation is designed to promote effective exploration and improve credit assignment by assessing the contribution of individual actions. To further examine the feasibility of the proposed method with real-UAV participation, a cloud IoT based hybrid validation platform with a one physical and multiple virtual architecture is constructed. Experimental results show that the proposed algorithm improves the pursuit success rate by 20.7% and reduces the average number of pursuit steps by 8.05 compared with the baseline method, thereby enhancing cooperative pursuit performance.