Learning-Enhanced Model Predictive Control for Fixed-Wing Small Uncrewed Aircraft Systems Path-Following

This paper presents a learning-enhanced nonlinear model predictive control approach for 3-D path-following guidance of a fixed-wing small uncrewed aircraft system along with flight test evaluation. A generic framework is introduced that leverages learned meta-level policies for dynamically adjusting inputs to an underlying model predictive controller’s cost function during online operation. The meta-level policy is trained via imitation learning using data generated offline by an expert meta-level policy based on approximate dynamic programming. The framework is applied to path-following guidance by dynamically scheduling reference path parameters over the controller horizon in a cost function that combines competing objectives of minimizing path error and maximizing airspeed. Simulation results assess the performance of the expert policy and show that the learning-enhanced model predictive controller simultaneously reduces the computational cost of the optimal control problem and improves path-following performance over statically parameterized state-of-the-art controllers. Flight experiments further demonstrate the effectiveness of the approach for following challenging 3-D paths, demonstrating real-world feasibility and robustness under significant wind disturbances and model mismatch while achieving strong performance.