Neural networks for minimum time satellite reorientation

This work investigates the use of Guidance and Control Networks (G&CNETs) for time-optimal reorientation of an asymmetric rigid spacecraft. Classical optimal-control approaches generally require the online solution of computationally demanding optimization problems, which may limit their applicability on resource-constrained platforms. Motivated by the promising performance of G&CNETs in related guidance and control applications, this study assesses their ability to approximate optimal slewing policies. To this end, several lightweight G&CNET architectures are designed and trained offline using expert trajectories generated by the GPOPS-II optimal control solver. Both regression- and classification-based formulations are considered, together with different activation functions. The resulting controllers are evaluated through an extensive simulation campaign in terms of maneuver time degradation relative to the optimal reference solution, robustness to spacecraft inertia uncertainties, and computational efficiency on commercial off-the-shelf embedded CPU platforms. Results show that the proposed networks achieve near-optimal maneuver times, with increases of only a few percent relative to the optimal reference solutions, representing a competitive alternative to state-of-the-art methods. Furthermore, the measured inference times of only a few milliseconds make them compatible with onboard execution on computationally limited hardware. These findings demonstrate the practical viability of G&CNET-based controllers for time-optimal spacecraft reorientation and highlight their potential for future autonomous spacecraft missions.