Smart robots play a crucial role in assisting human workers within manufacturing units (like Industry 4.0) by perceiving and analyzing their surroundings using Deep Learning (DL) models for Computer Vision (CV) applications. On the one hand, training DL models requires extensive annotated data. On the other hand, the scarcity and specificity of publicly available industrial datasets as well as the ethical, privacy, technical, and security challenges for capturing and annotating real images in industrial setups raise the problem of finding an alternative to train DL models for CV applications. In previous work, we proposed a simulation-based synthetic data generation (SDG) pipeline to render 200,000 images of eight industrial assets using NVIDIA Omniverse. In this study, we leverage the SDG pipeline to build and maintain dynamic and modular scenes, resulting in large-scale complex industrial simulation scenes. Furthermore, they feature Domain Randomization (DR) to increase content variability, and hence to bridge the reality gap. Inspired by real assembly lines, production areas, storage rooms, warehouses, offices set up, etc., we extensively render photorealistic images, rich in variations, capable of generalizing DL models to new unseen environments. Consequently, we introduce SORDI.ai, a comprehensive synthetic industrial image dataset for object detection applications. It comprises over a million images covering more than one hundred object classes belonging to logistics, transportation, signage, tools, and office assets. For evaluation purposes, we trained object detection DL models with our synthetic dataset, and inferred over a target dataset containing real/synthetic images. We gradually tested different levels of DR to demonstrated how does the reality gap bridge. Afterward, we showed the importance of mixing multi-domain training dataset to achieve better generalization, and the efficiency of our SDG pipeline to increase prediction accuracies in low real data regimes.