Reinforcement learning for accurate two-stage pointing control of CubeSats in photometric research

Abstract

Star photometry performed on satellites requires high pointing accuracy and stability. However, pointing control on CubeSats is especially affected by the spacecraft’s own movements and vibrations. Therefore, the goal of this work is to explore the application of Reinforcement Learning on the pointing control of CubeSats that perform photometry, considering a two-stage approach. A simulated environment was developed to model a staged approach to pointing control, and this work explored the appropriate parameters to be accounted for. The model achieved subpixel-to-subpixel accuracy and stability, and the results demonstrated the high impact of centroiding observations in achieving a higher learning rate and precision. In addition, this work provided evidence of the need for considering parameters from multiple subsystems and different stages of information processing in a spacecraft.