As NVIDIA paints its Physical AI inspection vision, DaoAI has already put it on the line. In December 2025, NVIDIA published a technical article noting that the CNNs (convolutional neural networks) that dominated AOI for a decade have hit a development ceiling.
Three structural limits of traditional CNNs
- High data threshold: each defect type needs thousands of labeled images, and rare defects lack enough samples.
- Limited semantic understanding: the model recognizes images but can't understand context or reason about root cause.
- Constant retraining: switching product lines means re-labeling and re-training, with maintenance cost piling up.
The direction NVIDIA validated
NVIDIA uses a pre-trained visual foundation model, first adapting it to the domain with a million unlabeled factory images, then fine-tuning with a small amount of labeled data. The result: PCB-defect detection accuracy rose from 93.84% to 98.51%.
DaoAI's implementation: built into hardware
DaoAI builds this approach into hardware as a plug-and-play solution:
- Based on a visual foundation model (VGG) architecture
- Trained on 1M+ real SMT factory images
- Specifically domain-adapted for PCBA manufacturing
- Feature extraction in feature space, not pixel space
On the line, the key numbers: programming time −97%, false-call rate −80%, operating cost −60%.
Manufacturers face a choice: adopt a tech stack you have to assemble yourself, or deploy a ready-to-run, continuously self-optimizing inspection system.
FAQ
What are the structural limitations of traditional CNN in the AOI field?
Traditional CNN has three major structural limitations in the AOI field. It has a high data threshold, requiring thousands of labeled images for each defect type. It has limited semantic understanding and can't understand context or infer root causes. Also, it needs continuous retraining, increasing maintenance costs when switching product lines.
What is the AOI technology direction verified by NVIDIA?
NVIDIA uses a pre-trained visual foundation model. It first performs domain adaptation with millions of unlabeled factory images and then fine-tunes with a small amount of labeled data. The experiment has significantly improved the PCB defect detection accuracy.
How does DaoAI implement the technology on the production line and what are the effects?
DaoAI integrates the technology into hardware. Based on the VGG architecture, after training on over 1 million real SMT factory images and domain adaptation, it reduces programming time by 97%, false alarm rate by 80%, and operating costs by 60% on the production line.