Learning-based fixed-wing UAV reactive maneuver control for obstacle avoidance

Abstract

Complex 3D obstacle environments raise high requirements of the rapid reaction capability and safety aiming at the maneuver control for obstacle avoidance (MCOA) of fixed-wing unmanned aerial vehicles (FUAVs). Considering the above demands, a learning-based reactive MCOA framework is proposed in this paper. First, the fundamental controller in this framework composed of the interfered fluid dynamical system (IFDS) guidance law and back-stepping control loops is established. Then, aiming at the rapid reaction requirement, a deep-reinforcement-learning-based (DRL-based) reactive online decision-making mechanism for obstacle avoidance matched with the IFDS guidance law is proposed. And aiming at the high safety requirement, the closed-loop system composed of the above fundamental controller and the FUAV 6-DOF nonlinear dynamic model is introduced into the DRL training environments so that the FUAV state transitions considering the characteristics of controllers and dynamics can be realized, and the corresponding reward functions can be calculated accordingly. On this basis, the optimal guidance instructions with high trackability can be resolved in real time using the actor networks trained by the proposed DRL-based mechanism, and the safe maneuver control in complex obstacle environments can be achieved. Finally, a normative modeling method of DRL training environments matched with the reactive MCOA framework is proposed to promote training efficiency. The effectiveness of the proposed framework is demonstrated by simulations.