Synthetic vision system recognizes assembly steps with 92.4% accuracy
No real-world data needed—just CAD models and step descriptions.
A team from ETH Zurich (Zhang et al.) has developed a synthetic-driven vision system that can recognize assembly steps in real time without any real-world training data. The system, described in a paper accepted at CASE 2026, automatically generates realistic assembly sequences using only CAD models and simple step descriptions. It combines a physics-based motion generation module (to simulate human assembly variations), domain-randomized rendering (to handle diverse lighting and backgrounds), and an object-detection-based step recognition module for robust sim-to-real transfer.
On a real-world industrial assembly case, the system achieved 92.4% accuracy. Compared to existing methods, it delivered a 46.7% improvement in handling motion variance, a 15.8% improvement in environmental robustness, and a 61.2% improvement in sim-to-real transfer. Setup can be completed in under an hour, making it a practical solution for factories needing quality control without the expense and labor of collecting and annotating real assembly data.
- System uses only CAD models and step descriptions to generate synthetic training data — no real-world images needed.
- Achieves 92.4% accuracy on a real industrial assembly task with three key improvements: 46.7% (motion variance), 15.8% (environment robustness), 61.2% (sim-to-real transfer).
- Setup takes under an hour, enabling rapid deployment for assembly quality control.
Why It Matters
Eliminates costly real-world data collection, making AI-powered assembly inspection accessible to any factory with CAD models.