Action-Regularized Reinforcement Learning for Adaptive Optics in Optical Satellite Communication

Optical satellite-to-ground communication enables terabit-scale data transmission through the atmosphere. However, atmospheric turbulence distorts the optical wavefront, significantly reducing the coupling efficiency into standard long-haul telecommunications fibers. While adaptive optics systems can correct these distortions, they are costly and complex. In this work, we introduce the State-Adaptive Policy Smoothness (APS) method, a reinforcement learning-based approach for wavefront sensorless adaptive optics tailored for low-cost optical satellite communication. Our results demonstrate that APS consistently maintains high coupling efficiency with low action fluctuation compared to baseline methods, enabling more stable and reliable operation under challenging atmospheric conditions.