Autonomous Control of Fixed-wing Unmanned Aerial System by Reinforcement Learning

This paper presents a near real-time aerial maneuver trajectory planning framework for fixed-wing UAS, which combines the Multi-Objective Monte Carlo Tree Search (MOMCTS) algorithm with parallel computing and receding horizon control techniques. Experimental validation over typical maneuvers including loop and break turn on a fixed-wing aircraft model are carried out with varying initial flight states. A calculation time of 2 seconds is taken to achieve a 2-seconds-long action plan for the maneuver mission in the successive time step, which enables the uninterrupted real-time flight maneuver control.