autonomous control of drone swarms using ai agents: design and experimentation

Large Language Models (LLMs) are being increasingly investigated as high-level reasoning components for cyber-physical systems, but applying them to real-time UAV swarm management is still difficult because of heterogeneous interfaces, weak grounding, and the need for persistent, closed-loop operation. This thesis introduces a mission-independent, agent-augmented LLM framework for UAV swarm control in which users specify objectives in natural language and the system autonomously carries them out through grounded, real-time interactions. The architecture integrates an LLM-driven Agent Core with a Model Context Protocol (MCP) gateway and a Web-of-Drones layer built on W3C Web of Things (WoT) standards. By representing drones, sensors, and services as standardized WoT Things, the framework supports structured tool-based interaction, continuous state monitoring, and safe actuation without depending on code generation. We assess the framework in an ArduPilot-based simulation across four swarm missions and six state-of-the-art LLMs. The results indicate that, although modern general-purpose LLMs exhibit strong reasoning capabilities, they still have difficulty delivering dependable execution, even on relatively simple swarm tasks, when explicit grounding and execution support are absent. Incorporating task-specific planning tools and runtime safeguards markedly increases robustness, and token usage by itself does not correlate with execution quality or reliability. Overall, the findings underscore both the promise and the present constraints of LLM-driven swarm control, showing that agent-enhanced execution and standardized device abstractions are key to turning natural language intent into reliable swarm behavior.