Vision-based and real-time calibration of industrial robot by using deep learning and dimension-reduced models

A real-time calibration method for six-DoF industrial robot is presented. It aims to address the issue that calibration always suspends the work of robot and is quite costly due to repeated utilization of expensive measurement instrument. For this, a pose measurement strategy with cost-effective cameras is proposed where the work trajectory of robot is considered as a strict constraint on configurations to meet the requirement of real time. However, such efficiency improvement is accompanied by the accuracy decrease because the camera brings large measurement error and the trajectory constraint results in the reduced parameter identifiability. Firstly, to pull up the camera performance to the same level of precise instrument like the laser tracker, a set of neural networks are designed to map the inaccurately measured poses by camera to the accurate poses of concerned joints. Secondly, a dimension-reducing method is proposed to truncate the constrained kinematic-error model into two, which commonly achieve high parameter identifiability. Simulation shows that the prediction accuracy of joint pose with the well-trained networks is much higher than the measurement accuracy of laser tracker. Besides, the calibration accuracy with the dimension-reduced models is validated to be higher than that with the single high-dimension model of typical calibration. Finally, a generation strategy of training data with deeper regularity is proposed to meet the accuracy requirement in real scene, and the comparative results in the experiment validate a good potential of the real-time calibration method in manufacturing that requires high accuracy of robot and no suspension of work.