Manipulating XXY Planar Platform Positioning Accuracy by Computer Vision Based on Reinforcement Learning

Abstract

With the rise of Industry 4.0 and artificial intelligence, the demand for industrial automation and precise control has increased. Machine learning can reduce the cost of machine parameter tuning and improve high-precision positioning motion. In this study, a visual image recognition system was used to observe the displacement of an XXY planar platform. Ball-screw clearance, backlash, nonlinear frictional force, and other factors affect the accuracy and reproducibility of positioning. Therefore, the actual positioning error was determined by inputting images captured by a charge-coupled device camera into a reinforcement Q-learning algorithm. Time-differential learning and accumulated rewards were used to perform Q-value iteration to enable optimal platform positioning. A deep Q-network model was constructed and trained through reinforcement learning for effectively estimating the XXY platform’s positioning error and predicting the command compensation according to the error history. The constructed model was validated through simulations. The adopted methodology can be extended to other control applications based on the interaction between feedback measurement and artificial intelligence.