Reinforcement Learning for Aligning Laser Optics with Kinematic Mounts

Aligning laser beams in optical systems is typically challenging, demanding both substantial time and expertise. This paper introduces a novel approach utilizing reinforcement learning (RL) architectures, specifically Deep Q-Learning (DQL), to align laser beams through controlling laser spot positions in optical setups. Our method employs RL for mirror position control, enabling the proposed solution to tune the mirror angles seamlessly. Real-time camera feedback informs the machine learning model's decisions. We propose a twostep laser alignment procedure: initially aligning the laser spot on the front aperture using the first RL model, and then once the laser is within the first aperture area, aligning the laser through both apertures using the second RL model. We verify the success of the RL model for the first step through experiments with a physical laser alignment setup of various configurations. We verify the RL model for the second step through simulations.