Automatic sorting of randomly arranged garments

The Institute for Information and Communication Technologies DIGITAL of JOANNEUM RESEARCH Forschungsgesellschaft mbH from Graz (Austria) presents a demonstrator that shows the functional principle of the new »CoboSort« system for the automated sorting of randomly arranged fully packaged, partially packaged and unpackaged garments. The camera system, the recognition of the garments and the calculation of the optimum gripping position for separating the textiles will be on display. The proposed solution has a positive impact on the reuse of returned clothes on the fashion market and enables new business models with a smaller environmental footprint.

The core components of the complete system are a robot, a gripper, a sensor system and a special image processing system for analyzing the sensor data and the control logic, which is based on AI tools. A »Fanuc CRX-Cobot« with a customized gripper is used to ensure safety and acceptance in the work area. A »Zivid One Plus depth camera« is used for sensory input. Processing is based on a customized machine learning model and gripper control logic. Additional sensors provide continuous feedback on the »success« of the gripping. This forms the basis for continuous monitoring of the gripping quality and a constantly increasing gripping success rate by retraining and fine-tuning the involved machine learning models when required.

Common systems for automatic handling of garments are often based on single plastic packaging, so that suction grippers can efficiently pick and handel individual items. Such an approach no longer works for items returned from end customers, where the packaging may be open, broken or missing completely. The proposed system is able to pick randomly arranged, stacked, fully and partially packed as well as unpacked clothing from a pile, trolley or box for subsequent sorting activities.

While robotic pick-and-place operations are standard in applications with rigid objects, dealing with textiles adds a number of challenges to the process with image processing. Garments are deformable and come in a wide variety of fabrics, colours and shapes with very different characteristics. In addition, fastenings and self-fasteners complicate the identification and the complex interaction properties make it hard to predict how garments will behave when gripped in certain positions. For the presented specific application, it is also necessary that only individual garments are picked up, as duplicate grasps would later lead to an overload in the individual evaluation of each returned garment.

By implementing intelligent systems for the management of new, returned or used garments, the fashion market can operate in a more sustainable, recyclable and environmentally friendly way. Under current industry standards, the sorting operations required to manage returned garments involve teams of operators for the repeated, intensive and heavily wearing picking and sorting tasks. The project shows that the use of a collaborative robotic assistant equipped with image sensors, grippers and artificial intelligence is a viable alternative.

Due to its small dimensions, modular architecture, self-safety and configurability, this system represents a moderate investment compared to current sorting solutions. It therefore paves the way for decentralized and flexible redistribution systems to support the emergence of new forms of e-commerce for unused, unwanted and new garments.

By using the system, the consumption of raw materials for the production of textiles can be reduced and the amount of waste generated can be cut. At the same time, the workers involved in the current sorting solutions are spared repetitive, wear-and-tear activities and given a proactive role as they improve the machine learning models themselves. This is done via an intuitive interface that can also be operated by non-robotic experts.