A machine learning-based visual servoing approach for fast robot control in industrial setting

Abstract

Industry 4.0 aims to make collaborative robotics accessible and effective inside factories. Human–robot interaction is enhanced by means of advanced perception systems which allow a flexible and reliable production. We are one of the contenders of a challenge with the intent of improve cooperation in industry. Within this competition, we developed a novel visual servoing system, based on a machine learning technique, for the automation of the winding of copper wire during the production of electric motors. Image-based visual servoing systems are often limited by the speed of the image processing module that runs at a frequency on the order of magnitude lower with respect to the robot control speed. In this article, a solution to this problem is proposed: the visual servoing function is synthesized using the Gaussian mixture model (GMM) machine learning system, which guarantees an extremely fast response. Issues related to data size reduction and collection of the data set needed to properly train the learner are discussed, and the performance of the proposed method is compared against the standard visual servoing algorithm used for training the GMM. The system has been developed and tested for a path following application on an aluminium bar to simulate the real stator teeth of a generic electric motor. Experimental results demonstrate that the proposed method is able to reproduce the visual servoing function with a minimal error while guaranteeing extremely high working frequency.