Abstract
Low-altitude wireless networks (LAWNs) enable three-dimensional unmanned aerial vehicle (UAV) swarm operations but face challenges from communication degradation and electronic warfare (EW) in contested airspace. This paper proposes a dual multi-agent swarm system (MASS) architecture that couples vehicle-level decentralized control with a supervisory large language model (LLM) layer under human-in-the-loop oversight. At the vehicle level, we develop a 3D communication-aware, gradient-based formation controller and integrate it with behavior-based navigation and global 3D path planning with dynamic replanning around discovered jamming zones. The controller uses a dominant-LoS air-to-air link-quality model whose effective quality smoothly degrades with inter-agent distance and, inside jamming zones, is further reduced by a multiplicative degradation factor. At the supervisory level, the LLM consumes swarm telemetry and is triggered upon abnormal or jamming conditions. It outputs structured escape guidance, which is blended with the default controller via severity-dependent override logic. Operator commands are supported through graduated autonomy. We develop an open-source Swarm Squad platform and evaluate six 3D planners (A*, Theta*, Dijkstra, Breadth-First Search (BFS), Greedy Best-First Search (GBFS), Bidirectional A*) across 30 randomized EW scenarios per algorithm. Results show that A* with LLM assistance reduces mission completion time by 10.5% and path length by 11.6% while maintaining high communication quality, demonstrating a practical path toward resilient, operator-supervised autonomy in contested LAWNs.