Automated UAV Controller Synthesis via LLM-Generated Control Logic and Particle Swarm Optimization

Large language models guided evolutionary frameworks for program synthesis have demonstrated strong results across combinatorial benchmarks; however, such frameworks have not been widely explored for robotic control systems. This article presents an automated framework for synthesizing control logic and optimizing numerical coefficients, enabling the deployment of generated controllers onto robotic platforms. To produce deployable controllers, the framework separates control logic synthesis from numerical parameter optimization, with controller coefficients optimized via particle swarm optimization using a Markov decision process reward signal. The resulting controllers are evaluated in a custom uncrewed aerial vehicles (UAV) simulation environment, validated using PX4 software-in-the-loop, and subsequently deployed on a physical UAV. Controller performance is evaluated on lemniscate and lissajous trajectory-tracking tasks and compared against proportional–integral–derivative with disturbance observer and linear quadratic regulator baselines. Across both trajectories, the framework-generated control law achieves improved tracking accuracy relative to the baseline controllers, with a minimum reduction in mean squared error of approximately 38% in real-world experiments. These results demonstrate the feasibility of deploying automatically synthesized control logic on a physical UAV, bridging automated program synthesis and real-world control deployment.