Learning Spacecraft Control with Reinforcement Learning

Since 2004 the amount of space launches has been gradually increasing. There are more than 100 satellites launched into space every year. The number of objects in the low Earth orbit will grow several times due to the planned constellations of thousands of satellites aimed at providing global Internet (Starlink, OneWeb). The increased amount of objects in space will lead to a higher probability of collisions. This trend means that automated collision avoidance algorithms will soon be essential for safe use of space. This paper is devoted to analyzing the reinforcement learning techniques and to designing a system for automatic optimal collision avoidance maneuver estimation. The main contributions of this paper are the following. First, a software simulation of space objects motion is described. Second, evolution strategies method is studied in application to the problem of satellite collision avoidance.