High-accuracy kinematic calibration of robot manipulator by compensating geometric and non-geometric errors

Purpose This study aims to solve the problem of high-precision kinematic calibration of manipulators. Kinematic calibration is an effective means to improve the absolute positioning accuracy of manipulators. However, the calibration accuracy of traditional methods still has limitations under several working conditions. To overcome this problem, a hybrid approach of calibration combining kinematic model and convolutional neural network is proposed in this paper to improve the calibration accuracy of a manipulator. Design/methodology/approach A hybrid approach of calibration combining a kinematic model and a convolutional neural network is proposed in this paper to improve the calibration accuracy of a manipulator. Specifically, as the first step, a sequential quadratic programming-based kinematic calibration process is carried out to primarily identify the geometric parameter errors. On the basis of this identification, a hybrid approach of calibration based on a convolutional neural network (CNN) is proposed. Afterward, the kinematic calibration integrated CNN approach is adopted for comprehensive compensation of both geometric and non-geometric parameter errors. Findings The performance of the proposed method is experimentally verified and compared with nine benchmarked methods, demonstrating a relatively high calibration accuracy. Meanwhile, several key issues are discussed, including the generalization capabilities of our proposed method, the probability density of the position error as well as the influence of the input format of the CNN model. Originality/value A hybrid calibration method combining kinematic modeling and neural networks is proposed, which is capable of fully compensating geometric and non-geometric parameter errors.