Research on dual-loop fast steering mirror control technology based on reinforcement learning parameter adaptation

The fast steering mirror (FSM) is indispensable in applications requiring ultra-precise pointing accuracy; however, its intrinsic third-order underdamped dynamics with nonlinear hysteresis pose significant challenges in achieving high-performance control. To overcome these limitations, we propose a reinforcement learning (RL)-driven dual-loop adaptive control framework. First, to resolve the inherent trade-off problem between response speed and stability, a tracking differentiator is introduced to estimate velocity signal for feedback augmentation; then, the inner velocity loop utilizes proportional-derivative control to counteract first-order oscillatory pole, and the outer position loop employs proportional–integral control to form a type I system for tracking control; this approach elevates system bandwidth from 270 to 315 Hz, reduces settling time to 2.31 ms, and eliminates step-response overshoot. Further addressing nonlinear hysteresis effects, the adaptive control algorithms are developed based on particle swarm optimization and RL, and the system bandwidths are increased to 363 and 391 Hz. Comparative analyses under step and sinusoidal commands confirm the superior adaptability of RL in mitigating hysteresis-induced uncertainties. Finally, a high-dynamic FSM prototype is implemented, with experimental results validating the control framework's efficacy. The methodology offers a generalized solution and with potential applications extending to advanced optical instrumentation and robotics.