A Novel Approach to an Autonomous and Dynamic Satellite Control System Using On-Orbit Machine Learning

Classical control methods require deep analytical understanding of the system to be successfully controlled. This can be particularly difficult to accomplish in space systems where it is difficult, if not impossible, to truly replicate the operational environment in a laboratory. As a result, many missions, especially in the CubeSat form factor, fly with control systems that regularly fail to meet their operational requirements. Failure of a control system might result in diminished science collection or even a total loss of mission in severe circumstances. Additionally, future SmallSat use cases (such as for orbital debris collection, repair missions, or deep space prospecting) shall place autonomous spacecraft in situations where mission operations cannot be fully simulated prior to deployment and a more dynamic control scheme is required. This paper explores the use of a student/teacher machine learning model for the purpose of training an Artificial Intelligence to fly a spacecraft in much the same way a human pilot may be taught to fly a spacecraft. With dedicated Artificial Intelligence & Machine Learning hardware onboard the satellite, it is also hypothesized that deploying an active learning algorithm in space may allow it to rapidly adapt to unforeseen circumstances without direct human intervention. Full development of a magnetorquer only control scheme was conducted with testing ranging from a software-in-the-loop 3D physics engine to a hemispherical air bearing, and finally a planned on-orbit demonstration. Further work is planned to expand this research to translational operations in future missions