Investigating the Impact of Domain Randomization on Synthetic-to-Real Deep Learning Performance

Deep-learning vision models need large labelled datasets, and collecting and annotating real images is slow and expensive. Synthetic data from simulators and rendering pipelines offers a scalable alternative, but models trained only on synthetic images often lose accuracy on real ones, a problem known as the synthetic-to-real domain gap. Domain randomization (DR) attacks the gap by varying rendering parameters such as lighting, textures, backgrounds, camera viewpoint, and object pose so widely that the real world looks like one more variation of the training data. This paper is a critical narrative review of how DR affects synthetic-to-real performance, mainly in object detection. It compares ten studies on their synthetic-data approach, randomization factors, architectures, metrics, findings, and limitations. Across these studies, models trained only on randomized synthetic data reached useful real-world accuracy in some constrained settings, and synthetic pretraining followed by a small amount of real-data fine-tuning often beat real-only training, although the margin can be small when real data is plentiful. Textures, materials, lighting, post-processing, and scene context are each reported as important in individual studies, but the rankings disagree, and differences in datasets, detectors, and metrics prevent a universal ordering of randomization factors. One study also shows that near-perfect synthetic validation scores can sit alongside noticeably lower real-world scores, so synthetic validation alone should not be trusted as a proxy for deployment performance. The review points to gaps in standardized evaluation, controlled factor-level ablation, and reproducibility.