AiMatics Uses World Models to Fix Factory Defect Data Gaps

Building an AI that spots faulty products normally requires examples of faults. On factory floors those are scarce: defects are rare, and flawed items are repaired and shipped quickly. South Korean company AiMatics says it has worked around the problem by teaching its system what normal products look like and flagging anything that deviates.

Chae Jung-hoon, head of AiMatics' AI future technology development group, said some customers refused to hand over defect data at all. "Sites can give us normal data but not defect data," he said, so the company built its solution around that constraint.

From dashcams to factories

AiMatics built its video recognition expertise on Roadscope, an AI dashcam for vehicles, and is now expanding into factory inspection and healthcare. The common thread is a "world model": an AI that first learns basic patterns and relationships from unlabeled video and signals, then picks up specific tasks from a small amount of labeled data.

Its vision inspection system AIM-T1 photographs electronic parts such as smartphone PCBs, camera modules and fingerprint sensors. According to the company, it judges defects as small as 3 micrometers within 0.1 seconds and can inspect up to 75,000 items a day. On one customer line, work previously done by about 40 people is now run by two.

Changing products usually means swapping jigs, adjusting cameras and lighting, and retraining the AI, which used to take three days to two weeks. AIM-T1 is designed around replaceable jigs, and the world model cuts the data and time needed. New products can now be put into inspection within hours, and the company targets under one hour from training to deployment.

Vehicles, healthcare and logistics

The in-vehicle product warns of drowsy driving, phone use, inattention and signal violations. In a 2024 trial on 500 intercity buses with the Korea Transportation Safety Authority, the company says drowsy driving fell 99.7% and accidents 29.2%. Its aimNet model is about 1MB so it can run on the device.

In healthcare, AiMatics is working with an affiliate on technology to find suspected arrhythmia segments in long ECG recordings.

Next, the company wants to automate logistics between processes, linking a process-operation platform with a robot-arm mobile robot under development to build autonomous factories step by step. Its goal is a general-purpose world model usable across industries by 2028.