Data-efficient, Fast and Autonomous Hybrid Calibration for Robot Positioning Accuracy Enhancement using Active Learning-based Method

Industrial robots generally have good repeatability and poor accuracy. Calibration is a known technique to improve robot accuracy. This technique is traditionally model-based and aims to identify the parameters of an analytical model of the robot using real measurements. More recent calibration methods, called hybrid calibration, combine model-based and data-driven approaches. These methods are more efficient than the traditional ones in integrating phenomena that are difficult to model. The most widely used data-driven approach for hybrid calibration is Artificial Neural Networks (ANN). The main drawback is that ANN is data-consuming, leading to long and complex measurement processes. In this paper, a new active learning hybrid calibration method is introduced. The hybrid calibration method relies on kinematics calibration to identify the geometric parameters of the robot, and on Gaussian Process Regression (GPR) for residual position error approximation. A fine-tuning of the method’s design parameters made on a measured dataset has led to a relevant and practical stopping condition. Experimental validations on two robotic manipulators demonstrate that the positioning error could be reduced by 97%, while using less than 30 measurements to train the GPR. Thus, the proposed method is data-efficient and provides better and faster results than current ANN-based hybrid calibration methods.