A Collaborative Multi-Agent Framework with Retrieval-Augmented Generation for Urban Counter-UAV Situation Assessment and Interception Decision-Making

With the rapid expansion of the urban low-altitude economy, the threat posed by rogue Unmanned Aerial Vehicles (UAVs) has intensified, necessitating robust Counter-UAV (C- UAS) systems for civil and security infrastructure. Traditional rule-based decision-making often lacks the flexibility to handle high-dynamic and unpredictable urban scenarios, while standalone Large Language Models (LLMs) are prone to safety- critical hallucinations and frequent violations of Rules of Engagement (ROE). This paper proposes a collaborative MultiAgent System (MAS) integrated with Retrieval-Augmented Generation (RAG), termed RAG-MAS, to automate situation assessment and Weapon-Target Assignment (WTA). The framework partitions the decision-making process into three specialized agents—Intelligence, Strategy, and Command—all grounded by a vector-indexed ROE knowledge base. Experimental results across 200 difficulty-progressive scenarios demonstrate that RAG-MAS reduces the overall Rule Violation Rate (RVR) to 7.0% (compared to 21.2% for traditional baselines) and achieves a 13.3% violation rate in complex swarm-attack scenarios, providing a reliable foundation for autonomous urban defense with acceptable latency trade-offs.