Positioning Error Calibration of Industrial Robots Based on Random Forest

抄録

Because most industrial robots are taught using the direct teaching and playback method, they are unsuitable for variable production systems. Alternatively, the offline teaching method has limited applications because of the low accuracy of the position and posture of the end-effector. Therefore, many studies have been conducted to calibrate the position and posture. Positioning errors of robots can be divided into kinematic and non-kinematic errors. In some studies, kinematic errors are calibrated by kinematic models, and non-kinematic errors are calibrated by neural networks. However, the factor of the positioning errors has not been identified because the neural network is a black box. In another machine learning method, a random forest is constructed from decision trees, and its structure can be visualized. Therefore, we used a random forest method to construct a calibration model for the positioning errors and to identify the positioning error factors. The proposed calibration method is based on a simulation of many candidate points centered on the target point. A large industrial robot was used, and the 3D coordinates of the end-effector were obtained using a laser tracker. The model predicted the positioning error from end-effector coordinates, joint angles, and joint torques using the random forest method. As a result, the positioning error was predicted with a high accuracy. The random forest analysis showed that joint 2 was the primary factor of the X- and Z-axis errors. This suggests that the air cylinder used as an auxiliary to the servo motor of joint 2, which is unique to large industrial robots, is the error factor. With the proposed calibration, the positioning error norm was reduced at all points.