NBP-CalibDRL: an automated hand-eye calibration method using deep reinforcement learning for industrial automation

Purpose As the demand for high-precision calibration increases in industrial automation and smart manufacturing, particularly for precision tasks, the accuracy and robustness of hand-eye calibration methods become crucial for ensuring reliable performance. Hence, this paper aims to present next best pose calibration with deep reinforcement learning (NBP-CalibDRL), an automated hand-eye calibration method based on deep reinforcement learning (DRL), designed to enhance both calibration accuracy and stability. Design/methodology/approach NBP-CalibDRL integrates three key components: an initial pose generation algorithm using K-means clustering, a viewpoint-distance adjustment strategy driven by DRL and a learning process for the viewpoint-distance value function. These elements form a calibration framework that guides the robot to select optimal viewpoints and distances using a reward mechanism and experience replay. Findings Experimental results show that NBP-CalibDRL outperforms traditional methods, such as random pose generation and K-means clustering, in reducing reprojection errors and improving stability. It maintains robust calibration even with varying calibration board positions and poses. The proposed method has been validated in both simulation and physical environments, demonstrating its potential for real-world industrial applications such as precision assembly and welding. Originality/value NBP-CalibDRL innovatively applies DRL to hand-eye calibration, proposing a method to improve calibration accuracy by adjusting the viewpoint and distance and introducing a new viewpoint-distance value function fitting approach. By leveraging autonomous learning and decision-making, it improves both calibration precision and robustness. In addition, this paper enhances existing hand-eye calibration methods based on K-means clustering, significantly enhancing calibration accuracy.