Reinforcement Learning Approach to Optimizing Profilometric Sensor Trajectories for Surface Inspection.

Area of Science:

• Manufacturing Technology
• Robotics
• Artificial Intelligence

Background:

• High-precision surface defect detection is crucial in manufacturing.
• Laser triangulation profilometric sensors provide detailed surface measurements.
• Accurate robotic motion is essential for optimal sensor performance.

Purpose of the Study:

• To develop a novel Reinforcement Learning (RL) approach for optimizing inspection trajectories.
• To enhance surface defect detection using profilometric sensors.
• To ensure consistent profile distribution and high-quality scanning through dynamic trajectory adjustment.

Main Methods:

• Utilized a Reinforcement Learning (RL) model, specifically the Proximal Policy Optimization (PPO) algorithm.
• Designed a tailored state space, action space, and reward function for profilometric sensor inspection.
• Employed a simulated environment with realistic conditions (sensor noise, surface irregularities) for offline trajectory planning using CAD models.

Main Results:

• The RL agent successfully optimized inspection trajectories for profilometric sensors.
• The approach demonstrated effectiveness in ensuring consistent profile distribution and high-quality scanning.
• Validation in simulation and real-world testing with a UR3e robotic arm confirmed the model's practicality.

Conclusions:

• The proposed RL-based method offers an effective solution for optimizing surface inspection trajectories.
• This approach significantly improves the precision and efficiency of defect detection in manufacturing.
• The dynamic trajectory optimization enhances the performance of laser triangulation profilometric sensors in real-world applications.