LLM-supported UAV Swarm Mission Planner

Abstract

Disaster response increasingly relies on UAV swarms for surveillance, victim location,and damage assessment, but coordinating such swarms in dynamic, partially observable environments remains a planning challenge that fixed strategies handle poorly. This thesis investigates the use of Large Language Models (LLMs) as high-level mission planners that translate natural-language mission goals into adaptive UAV command sequences in disaster scenarios.A complete planning pipeline is designed and implemented, comprising simulated perception, a hierarchical environment representation that exposes the world to the LLMas a typed semantic graph, an LLM-based mission planner, and a closed-loop mission controller. Five system components are evaluated through a progressive ablation acrosss configurations and four mission goals expected to differ in cognitive demand: mission context injection, structured prompt design, execution feedback, hybrid graph parsing with pre-computed semantic abstractions, and event-driven replanning. The evaluation is grounded by a privileged-information Best-Effort baseline and a non-adaptive static-LLM baseline, and extended through a cross-hazard generalization study to pluvial flood scenarios and a comparison between a cloud-hosted and a locally deployed LLM.The results show that no component is universally dominant: each addresses a distinctbottleneck, and the relative importance of components depends on the cognitive demand of the mission goal. The full adaptive system matches or exceeds the privileged Best Effort baseline on the area assessment task (IoU = 0.948 vs. 0.870, where the baseline’s fixed-sweep coverage primitive leaves systematic gaps) while remaining below it on tasks focused on discrete target localization. The full adaptive system raises infrastructure decision accuracy from 24% to 92%, with pre-computed semantic abstractions providing the decisive component-level uplift (+24 percentage points). Smaller locally hosted models retain detection capability but collapse on tasks requiring a held structured goal and a grounded decision: on infrastructure assessment the local model never commits a decision, drifting off-task despite the scene graph containing the needed information.The findings suggest that the next bottleneck for LLM-supported UAV planning lies insemantic pre-computation and schema-aware adaptation rather than in raw model scale.