Design of Reinforcement Learning-Based Dual-Servo Controller for Semiconductor Inspection Equipment

Ultra-precision stages for semiconductor inspection must simultaneously achieve long-stroke motion, nanometer-level accuracy, and fast settling. Dual-servo stages comprising coarse and fine stages are widely used, but coordination performance degrades in practice due to dynamic mismatch, alignment errors, and nonlinear coupling. Conventional approaches such as feedforward compensation and model-based control rely on accurate models and show limited robustness to disturbances. This study proposes a reinforcement learning-based PID (RL-PID) controller for the fine stage of a one-degree-of-freedom dual-servo system, while a conventional PID is applied to the coarse stage. The RL-PID optimizes PID scaling factors using DDPG , with stability ensured through a Lyapunov-based reward and FRF -derived safe exploration bounds. Simulation and dSPACE-based experiments demonstrate significant improvements in settling time, overshoot, and tracking accuracy, achieving up to a 33% reduction in RMSE under high-speed scanning. The results validate the effectiveness of safe RL-PID in mitigating nonlinear coordination issues and enhancing industrial applicability.