Deep Reinforcement Learning Applied to Spacecraft Attitude Control and Moment of Inertia Estimation via Recurrent Neural Networks

This study investigated two distinct problems related to unknown spacecraft inertia. The first problem explored the use of a recurrent neural network to estimate spacecraft moments of inertia using angular velocity measurements. Initial results showed that, for the configuration examined, the neural network can estimate the moments of inertia when there is a known external torque. The second problem trained a reinforcement learning agent, via proximal policy optimization, to control the attitude of a spacecraft. The results demonstrated that reinforcement learning may be a viable option for guidance and control solutions where the spacecraft model may be unknown. The trained agents displayed a degree of autonomy with their ability to recover from events never experienced in training