Residual Position Error Compensation Method Based on JEP-DNN for Robot Vision Measurement System

Abstract

Accurate positioning of surface geometric features is essential for quality inspection of automotive components. While vision-based robotic measurement systems are extensively employed, residual errors arising from nonlinear factors, kinematic constraints, and environmental influences pose significant challenges. This paper introduces a novel compensation method leveraging a joint error propagation deep neural network (JEP-DNN) to overcome these limitations. First, we develop a comprehensive error model integrating both kinematic and non-kinematic factors, including geometrical parameters and gravity-induced deformations. Then, to address the residual model error, we establish a joint error propagation model to describe error propagation across different parts of the robotic system. The JEP-DNN utilizes the error propagation model to perform residual error compensation, combining corrections in joint space and Cartesian space to achieve high interpretability and precision. Experimental results demonstrate the effectiveness of the proposed method. In a laser tracker calibration experiment, the JEP-DNN reduced average position errors by 30.4%, while in a vision-based inspection task for an automotive cornering light, it achieved a 12.4% improvement in positional accuracy compared to uncompensated results. Finally, accuracy and repeatability evaluation of automotive sheet metal parts illustrate that the robotic vision system achieved a maximum position error of 0.143 mm, with high repeatability (less than 0.066 mm).