Graduation Semester and Year
Summer 2026
Language
English
Document Type
Thesis
Degree Name
Master of Science in Aerospace Engineering
Department
Mechanical and Aerospace Engineering
First Advisor
Dr Kamesh Subbarao
Abstract
This thesis presents a closed-loop control architecture for uncrewed aerial vehicles (UAVs) in which a large language model (LLM) serves as a high-level decision module operating over a persistent, metric 3D world model.
Rather than generating low-level commands or open-loop plans, the LLM selects one parameterized maneuver per decision step from a small, verified library of flight primitives conditioned on a structured representation of the drone state, tracked object positions, and mission specification.
Translational motion is executed by a planar model predictive controller (MPC) with soft obstacle avoidance, using obstacle hypotheses provided by the LLM, so that safety-critical constraint handling is delegated to the optimization-based controller rather than the language model.
The architecture is evaluated in simulation using a PX4 X500 quadrotor with RGB-D sensing in ROS~2 and Gazebo across multiple LLM backends including Gemini-3.5, GPT-4o, GPT-5.2, and a locally deployed Qwen2.5-3B model.
Experiments spanning target approach, constrained navigation, yaw-aligned approach, and object search tasks indicate that constraining the LLM to symbolic maneuver selection over an explicit 3D world model yields more interpretable and more constraint-respecting behavior in the tested scenarios, while preserving flexible natural-language task specification.
An MPC ablation study demonstrates that obstacle avoidance must be enforced at the planner level: with the MPC disabled, a correct obstacle declaration by the LLM is not turned into a safe trajectory and the direct-setpoint fallback contacts the obstacle, showing that a competent execution layer, and not obstacle declaration alone, is what yields safe motion.
A key emergent behaviour is also observed in which the closed loop compensates for the small terminal offset introduced by the soft-constrained MPC formulation through a subsequent corrective maneuver selected by the LLM, validating the combined design without any explicit residual-compensation logic.
The findings establish that constraining the LLM to a verified action space and delegating safety enforcement to a deterministic planner are both necessary conditions for reliable LLM-driven UAV autonomy.
Keywords
Large language models, Uncrewed aerial vehicles, Model predictive control, Closed-loop autonomy, Robot task planning
Disciplines
Navigation, Guidance, Control and Dynamics
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Yasar, Halimcan, "LLM-Driven Closed-Loop UAV Control with Obstacle-Aware Model Predictive Control" (2026). Mechanical and Aerospace Engineering Theses. 9.
https://mavmatrix.uta.edu/mechaerospace_theses2/9