Industrial Visual Sim-to-Real Review Categorizes by CAD Availability
A new article on arXiv, titled arXiv:2605.30581, takes a fresh look at the sim-to-real challenge in industrial settings, framing it as a domain-gap issue based on how accessible CAD models are. It argues that when implementing systems in industry, the gap between available data and required decisions is much larger than just converting synthetic images to real ones. The systems can start from various sources like CAD models or simulated data, but they often face different conditions like lighting, materials, and types of defects. In cases where CAD is available, it helps with tasks like rendering and calibration. However, without CAD, other methods like normal-reference appearances and synthetic anomaly assumptions take over. The article provides a thorough classification for those working in this field.
Key facts
- Paper reframes industrial visual sim-to-real as domain-gap problem organized by prior availability.
- Distinguishes CAD-available and CAD-unavailable settings.
- Industrial deployment involves broader mismatch beyond synthetic-to-real image transfer.
- Systems built from CAD renderings, simulated RGB-D, normal references, synthetic defects, pretrained features, or language prompts.
- Deployment conditions vary: sensors, lighting, materials, fixtures, calibration, production variation, rare defect modes.
- CAD-available: explicit geometry supports rendering, calibration, pose estimation, segmentation, geometric verification.
- CAD-unavailable: geometry replaced by normal-reference appearance, feature distributions, teacher-student residuals, synthetic anomaly assumptions.
- Review provides structured taxonomy for researchers and practitioners.
Entities
Institutions
- arXiv