Robotics has advanced significantly, but small and medium-sized enterprises face challenges in adopting it in high-mix, low-volume environments. Primarily due to the manual programming required for robot integration and the time it takes for a process to run stably and produce high-quality output without rejections. The high scrap rate in unstable processes leads to low productivity, which is detrimental to competitiveness. While computer vision technologies offer some flexibility, they often require large, manually labelled datasets and only provide binary quality feedback or error classes. This paper proposes an approach to automate robotic processes and quality control using synthetic data. Computer-aided design data and assembly sequences initiate object recognition training. A quality model is automatically derived from the data, which is used to evaluate the manufactured quality during the process and to continuously improve the process via iterative learning control. The goal of this procedure is to allow for a change of variant without the need for time-consuming re-teaching by a robotics expert.
Geiser et al. (Thu,) studied this question.