The Day the Inspection Robot Misread a Defect
An inspection robot checking semiconductor equipment parts begins flagging good units as defective. The factory workers try every trick to override the machine's stubborn judgment, but each attempt yields the wrong result.
At six in the morning, before the cleanroom lights came on, the inspection robot Maru was already watching the conveyor belt.
Maru was a machine vision system for checking ceramic components in semiconductor equipment. Three cameras picking up 0.1-micron variations, six light sources, and one neural model trained on three years of defect data. Team lead Junki liked to brag that output had jumped twelve percent thanks to Maru.
But that day was different.
Seven o’clock, first part: defective. Second: defective. Sixteenth: defective.
Junki grabbed a component and placed it under the microscope himself. No scratches, dimensions normal, surface condition good. He reset Maru. Defective.
“Recalibrate it.”
His colleague Sunyoung pulled out her laptop and ran the calibration routine. Maru paused, then started again. Defective, defective, defective.
Junki adjusted the light angle. Defective. He shut off one light source. Defective. Both. Defective.
“Could it be dust on the camera lens?”
Sunyoung wiped the optics. Maru rendered a good verdict for a moment, then flagged the next part as defective again.
By ten o’clock, Dohyeon from quality arrived. He had updated Maru’s training data four days ago, he said. He’d added new defect patterns and raised the sensitivity for micro-cracks.
Junki asked, “How much?”
“I dropped the sensitivity threshold from 0.3 to 0.07.”
Silence filled the cleanroom.
Maru kept running, methodically, faithfully, processing each perfectly good part on the conveyor as defective. Whether sharper eyes had learned to see more or simply to suspect more, Maru couldn’t tell the difference.
This article is for information only and is not investment advice.