Synthetic-to-Real Domain Adaptation

The field studies how to make models trained on simulated or generated data work reliably on real images, sensor readings, and physical systems, where labels may be scarce or costly. Common approaches include physics-based simulation and domain randomization, style translation, feature alignment with adversarial learning or pseudo-labels, and self-supervised or test-time adaptation. Researchers evaluate transfer in settings such as GTA-to-Cityscapes driving, LiDAR perception, robotic manipulation, and medical imaging, often comparing synthetic-only training with mixtures that include real data. A central tension is that visual realism alone may not capture the real domain’s diversity: synthetic data can provide useful supervision, but small amounts of well-matched real data and task-specific adaptation often determine whether gains transfer.