Robust Attitude Control for PAVs using DNN with Exponentially Stabilizing Control Lyapunov Functions

Abstract

This research proposes an exponentially stabilizing deep neural Lyapunov control framework (es-DNLC) for robust attitude control of the KP-1 eVTOL personal aerial vehicle (PAV). To begin, es-DNLC is a deep neural network (DNN) learning-based control that was created by employing exponentially stabilizing control Lyapunov functions(es-CLF) which can significantly ensure stability and robustness as compared to using locally asymptotic Lyapunov functions as a learning policy. In addition, the nonlinear flight dynamic model of KP-1 is constructed using mathematical methods to build a controller, and data from CFD results are used to improve the dynamic model's fidelity. Based on the established model, the multi-copter mode attitude control system and the fixed wing mode control system in the longitudinal and lateral directions are constructed. The efficacy of robust control of es-DNLC is then compared to common controls such as linear quadratic regulator (LQR) in MATLAB/Simulink to highlight the improvement of the proposed control framework. The controller that suggested the framework es-DNLC increased the area of the region of attraction (ROA) compared to LQR. Real-world flight tests are performed on a scaled model of the KP-1 eVTOL UAM vehicle using the PX4 autopilot implemented on the flight control computer Pixhawk board for the comparison of LQR and es-DNLC regarding roll motions as a case-study. The suggested es-DNLC control framework can guarantee a substantially greater level of resilience in KP-1 motions against external disturbance and model uncertainty. The proposed robustness control framework can be beneficial for future development of large-scale UAM vehicles for safety and resiliency of air transportation in urban areas.