Learning-Based Adaptive MPC for Magnetorquer-Only Satellite Attitude Control

Abstract

Nanosatellites (e.g., CubeSats) have the potential to carry out useful science and exploration activities for outer solar system exploration, but their small size leads to tradeoffs between attitude control and science capabilities. Magnetorquers offer a unique compromise as compact actuators that enable three-axis attitude control, but suffer from a loss of control authority and require full knowledge of the planetary magnetic field. In this work, we present a learning-based adaptive model predictive control (MPC) algorithm that plans rotational trajectories for the satellite while accounting for this loss of control authority. Key to our approach is the inclusion of a learned magnetic field model that is trained in simulation using low-fidelity magnetic field data. Using measurements collected online, our proposed algorithm adapts this learned magnetic field model using a simple last-layer update scheme, allowing the MPC controller to quickly adapt to the true planetary magnetic field model and more effectively plan attitude trajectories. We demonstrate the results of our proposed approach for two use cases, (1) Earth-orbit and (2) Uranus-orbit, and show that our learning-based adaptive MPC algorithm improves controller success for these system simulations by 5% and 61%, respectively.