A hybrid approach towards automated textile sorting

摘要

Abstract Each citizen of the European Union discards approximately 12 kg of garments annually. This prompted the European Commission to start an initiative on how to handle and recycle old clothing. To this purpose, machine handling of the sorting and processing of such textiles is preferred due to the high amount of garments. As these clothing articles are soft, of different sizes, shapes, colors and articles like socks preferably need to be paired, a major challenge in the field of robotics and artificial intelligence arises. Current approaches specialize in one part of this challenging task. State of the art for the sorting of textiles part is to use cameras and pre-trained networks. This paper presents a hybrid approach on using Convolutional Neural Networks in conjunction with Visual Language Models to find and pick single pieces of garments from a convoluted pile, inspect the individual clothing article and sort them into dedicated containers. The selection is done based on type and color of garment as well as including a pairing algorithm for socks. Successful grasping and dropping as well as potential rapid object loss detection during manipulation is achieved by incorporating Wireless Capacitive Tactile Sensor Arrays to the robots’ end-effector. The approach is demonstrated in a robotic gripper use-case, showcasing grasping and sorting of garments with accuracies of 96.72% and 93.44% for type and color respectively, and pairing of textiles on the example of socks with an accuracy better than 81.90%.