I want to create a image dataset for training YOLOv8 model for an autonomous robot working in steel structure assembly site. As it is difficult to get specific dataset related to steel structure assembly robot, I am planning to use synthetic images for training and then deploy it in the real world. I am currently using free assets such as trucks, forklift, crane, cement mixer, human etc. from unity asset store and Unity perception package for automatic labelling. I tried different randomizers. But, I am not satisfied with the trained model as it fails to detect almost 50% time while testing with real images.

Any suggestions /expert opinion for creating an effective synthetic dataset will be highly appreciated.