Cooperative Defense Against Suicide Drones for a Moving Asset

초록

As the use of unmanned combat systems increases, protecting moving assets from adversarial suicide drones has become an important challenge. This paper proposes a multi-agent reinforcement learning-based cooperative defense method under partial observability. The key novelty lies in the cooperative defense formulation and a Proportional Navigation (PN)-inspired reward design. We model the escort defense scenario as a decentralized MARL problem and design a reward that encourages minimizing the line-of-sight (LOS) rate while maintaining positive closing velocity, leading to effective and coordinated interception behaviors. To handle partial observability, we adopt a recurrent multi-agent proximal policy optimization (rMAPPO) algorithm. Experimental results show that the proposed method achieves an 84% defense success rate, outperforming baseline methods and demonstrating improved coordination and safety in protecting the moving asset. © 2026 IEEE.

키워드