Design and Flight Test Validation of an AI-Based Longitudinal Flight Controller for Fixed-wing UASs

Abstract

Applications of deep reinforcement learning (RL) based control have mainly focused on multi-rotor unmanned aerial systems (UASs). The limited number of research on AI-based flight controllers for fixed-wing UASs platforms lacks rigorous validation and verification (V&V) flight tests. Simulation V&V are conducted for AI-based controllers and only AI-based guidance has been implemented for the flight test verification. The complexity and scarcity of research on AI-based flight controllers for fixed-wing UASs are multilayers, however high speed and inertia of this class of aircraft, structural load limitations, stall speed and angle of attack, and large impact of aerodynamic forces and moments are among the main reasons. Tight tracking requirements and nonlinear and unsteady aerodynamics in presence of external disturbances only exacerbate the complexity. The traditional controller design methods (e.g., Modern and Classical Control) heavily rely on the quality of aircraft physics-based dynamic models. Robust controllers address the uncertainties in the dynamic model by trading off the performance of flight controllers. This work presents the development of a policy-based flight controller (longitudinal-directional) for a fixed-wing UAS using a model-free deep RL algorithm called PPO (Proximal Policy Optimization). The UAS has a 2.66 m wingspan, weighs 6.4 Kg, and cruises at 15.3 m/s. The controller is trained in simulation using a low-fidelity dynamic model with appropriate dynamic randomization to tackle the issue of parametric uncertainties. To further improve the transferability of the controller, we also incorporate memory functions into the policy using RNN (recurrent neural network) architecture. We opt for the PPO Deep RL algorithm for its monotonic and stable policy improvement characteristics and less sensitivity to hyper-parameters. The developed controller performance is first validated in HiTL (hardware-in-the-loop) simulations. Several actual flight tests are conducted for the verification of the developed controller. The AI-based controller is flight tested in different weather conditions to quantify its robustness toward external disturbances. Both simulation and flight test results are presented in this work.