This paper focuses on how to generate digital twins (DTs) starting from image or video inputs through the use of modern AI methods. To address the problem of lack of public datasets, this paper reports on the creation of 2 new image datasets featuring 244 and 471 images, respectively, of collaborative robots (Cobots), annotated based on the 3 categories of Cobots, joints and links. Based on metrics obtained through the logging of training runs and the real-time visualization of network parameters, we demonstrate the impact of the number of images had on the learning capacity of the network. Pretrained results have been saved and made available for future inference as well as transfer learning.