Reinforcement learning reward function in unmanned aerial vehicle control tasks

Abstract This paper presents a new reward function that can be used for deep reinforcement learning in unmanned aerial vehicle (UAV) control and navigation problems. The is based on the construction estimation of time simplified trajectories to target, which are third-order Bezier curves. applied unchanged solve problems both two-dimensional three-dimensional virtual environments. effectiveness was tested newly developed environment, namely, environment describing dynamics UAV flight, taking into account forces thrust, inertia, gravity, aerodynamic drag. In this formulation, three tasks were successfully solved: flight given point space, avoidance interception by another UAV, organization one another. most relevant modern algorithms, Soft actor-critic, Deep Deterministic Policy Gradient, Twin Delayed Gradient used. All algorithms performed well, indicating selected function.