Multi-Agent Deep Reinforcement Learning Control of the TAT Nasnas Mk.1 UAV: 6-DOF Modeling and Autonomous Trajectory Tracking

This paper proposes a multi-agent reinforcement learning (MARL) framework for autonomous trajectory tracking of the TAT Nasnas Mk.1 fixed-wing unmanned aerial vehicle (UAV), using a nonlinear six-degree-of-freedom (6-DOF) model implemented in MATLAB/Simulink. The control architecture employs two cooperating agents for longitudinal and lateral-directional flight control, trained using the Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3) algorithm under a centralized training and decentralized execution (CTDE) paradigm. A shared multi-objective reward function balances trajectory-tracking accuracy, attitude regulation, control effort, and flight-envelope safety. The framework is evaluated against proportional-integral-derivative (PID), nonlinear backstepping, and single-agent Twin Delayed Deep Deterministic Policy Gradient (TD3) controllers across five representative trajectories: straight-line tracking with heading change, lawnmower surveying, spiral search, figure-eight tracking, and helical climbing. Robustness is assessed under ±10% parametric uncertainty and external wind disturbances. Across the nominal scenarios reported in the results tables, the proposed MARL controller achieves three-dimensional position-tracking root-mean-square error (RMSE) values of 0.52–1.10 m, corresponding to approximately 35–39% lower RMSE than single-agent TD3 and 57–78% lower RMSE than conventional controllers, depending on the scenario and baseline. Coupling-rejection values of 64–70% are reported for the four scenarios in which this metric is evaluated. Under parametric uncertainty, the reported increase in position RMSE is 12%, compared with 20–45% for the benchmark controllers. Under wind disturbances, the increase is 15%, and the reported recovery time is 1.5 s. These are simulation-based findings from a model whose aerodynamic and inertial parameters are partly estimated; hardware-in-the-loop and flight validation remain necessary before operational conclusions can be drawn.