Reinforcement learning-based control law on PAPYRUS: simulations using different atmospheric conditions

Predictive control laws for Adaptive Optics (AO) using Artificial Intelligence has been recently explored as an alternative to the classic methods, such as the integrator law. Reinforcement Learning excels in predictive control tasks by enabling systems to learn optimal control strategies through continuous interaction with their environment, adapting to dynamic conditions and achieving effective decision-making in real-time. In our previous work, a Model-based Reinforcement Learning (MBRL) method called Policy Optimization for Adaptive Optics (PO4AO) was used in conjunction with the Object-Oriented Python Adaptive Optics (OOPAO) to simulate the Provence Adaptive Optics Pyramid Run System (PAPYRUS) optical bench. PO4AO demonstrated high adaptability to turbulence and rapid convergence, achieving optimal corrections after just 500 frames of interaction, outperforming a simulated integrator in different atmospheric conditions. Building upon this, our current study explored PO4AO's capability to adapt to sudden atmospheric changes by worsening turbulence conditions during evaluation, notably the wind speed and the seeing. In the result's section, we compare PO4AO's performance in terms of Strehl Ratio (SR) to the integrator. Further description of the experiments are present in the paper.