Guidance and control networks with periodic activation functions

Building on the remarkable representational power of sinusoidal representation networks (SIRENs), we introduce a neural architecture that efficiently encodes the optimal control policy for general dynamical systems and propose its use, specifically for space flight guidance and control tasks. Our approach, a class of Guidance and Control Networks (G&CNETs), significantly accelerates training under the behavioral cloning paradigm while achieving substantially lower training error across three distinct space flight scenarios. Previous studies have demonstrated the superior approximation capabilities of SIRENs, particularly in image and video reconstruction. We contextualize our findings by drawing a parallel between optimal control learning and high-fidelity signal reconstruction, highlighting the ability of SIREN-based architectures to capture the intricate, discontinuous nature of optimal control landscapes.