<p>An improved deep deterministic policy gradient (I_DDPG) algorithm is developed to enhance the accuracy and robustness of toolface angle control in the stabilized platform of a rotary steerable system. The intrinsic frictional torques of the platform are analyzed, and the stability of the DDPG-controlled system is verified through a Lyapunov function. To address value overestimation and minimize error accumulation, a clipped double Q-learning strategy with a delayed update mechanism is integrated into the DDPG framework. Additionally, the maximum entropy principle is employed to enhance exploration capabilities, leading to the I_DDPG algorithm. Comparative simulation results show that I_DDPG outperforms traditional methods, reducing tracking error by 40.64%, increasing response speed by 78.31%, and significantly reducing overshoot by 97.33% compared to the PID algorithm. Furthermore, the algorithm demonstrates strong robustness and adaptability, effectively mitigating the effects of variations in armature resistance and viscous friction coefficient. This approach provides a reliable solution for precise toolface angle control in complex and dynamic drilling environments.</p>