Abstract: Adaptive optics (AO) improves the performance of optical systems in astronomical observation, laser communication, biomedical imaging, and other applications by sensing and correcting wavefront aberrations in real time. In recent years, advances in reinforcement learning (RL) have provided new approaches to AO control, particularly for systems operating under dynamic disturbances and model uncertainties. This review summarizes recent progress in RL-based AO control. The closed-loop correction process is first formulated as a sequential decision-making problem, and the key modeling elements, including states and observations, action spaces, reward functions, and policy representations, are summarized. According to the availability of explicit wavefront measurements, existing RL-based AO methods are then classified into wavefront-sensor-based (WFS-based) AO and wavefront sensorless (WFSless) AO. Representative studies in both categories are reviewed, with emphasis on their control strategies, methodological characteristics, and reported performance. The major challenges of RL-based AO, including training efficiency, real-time implementation, and system reliability, are further discussed. Finally, future research directions are outlined, with particular attention to the integration of RL with AO physics and conventional control methods, high-dimensional real-time control, safe learning, and experimental validation.