A deep learning based robot positioning error compensation

Robot position accuracy plays a very important role in advanced industrial applications, nowadays, most of the industrial robots have excellent repeatability, however, it still always remain some absolute position error that are due to non geometric calibration parameters that are hard to model and identify. The present work studied a method to reduce the absolute position error of robots using conventional identification procedures as well as neural networks.In order to increase the robot accuracy, we propose to first identify determinable error sources (geometric errors and joint deflection errors), then, use deep learning based methods to identify the non-geometric error sources such as link compliance, gear backlash, and others, which are difficult to model correctly and completely. The algorithm is tested on simulation with the UR-10 robot and is able to identify some predefined parameters with a high level of accuracy using only measurements data and deep learning methods.