Reinforcement Learning Approach to Optimizing Profilometric Sensor Trajectories for Surface Inspection

High-precision surface defect detection in manufacturing often relies on laser triangulation profilometric sensors for detailed surface measurements, providing detailed and accurate surface measurements over a line. Accurate motion between the sensor and workpiece, usually managed by robotic systems, is critical for maintaining optimal distance and orientation. This paper introduces a novel Reinforcement Learning (RL) approach to optimize inspection trajectories for profilometric sensors based on the boustrophedon scanning method. The RL model dynamically adjusts sensor position and tilt to ensure consistent profile distribution and high-quality scanning. We use a simulated environment replicating real-world conditions, including sensor noise and surface irregularities, to plan trajectories offline using CAD models. Key contributions include designing a state space, action space, and reward function tailored for profilometric sensor inspection. The Proximal Policy Optimization (PPO) algorithm trains the RL agent to optimize these trajectories effectively. Validation involves testing the model on various parts in simulation and performing real-world inspection with a UR3e robotic arm, demonstrating the approach's practicality and effectiveness.