This study presents a robust and intelligent control method for flexible satellites operating under underactuated conditions. When the number of actuators is fewer than the system's degrees of freedom, issues like instability and vibrations arise. To address these problems, a combination of super-twisting sliding mode control and high-order adaptive sliding mode control, along with reinforcement learning, is used. Reinforcement learning helps to adaptively adjust the control gains, improving the system’s performance in the presence of disturbances and actuator failures. Quaternion parameters are utilized to avoid singularity issues when modeling the satellite's angular orientation. In this approach, the control inputs for the first and second axes are adjusted to reduce the error in the third axis without requiring direct control. Various simulations have shown that the proposed method outperforms classical approaches in reducing errors, minimizing chattering, and enhancing system stability. Furthermore, the high-order adaptive sliding mode control demonstrates greater stability against model uncertainties, although with longer settling times. These results indicate the high potential of the proposed methods for use in sensitive space missions.