The most expensive bug we ever shipped
A while back, a measurement station at a customer site started rejecting good parts. Not many — 0.3 % extra, week over week. By the time anyone noticed, the line had quietly scrapped a sizeable amount of in-spec material. The cause: a 0.04 mm drift on a calibrated 200 mm working distance, accumulated over six months.Calibration drift is a slow leak. It's not a thing most teams budget time for, and it costs more than every other vision-system maintenance item combined.
What "calibration" actually covers
Two layers, often confused:- Intrinsic calibration — focal length, principal point, distortion coefficients of the lens. Stable on a static system. Drifts when the lens is touched, vibrated hard, or temperature-cycled.
- Extrinsic calibration — pose of the camera relative to the world (the conveyor, the part fixture, the robot frame). Drifts continuously: bolts settle, frames flex, mounting plates warm up.
In our experience, intrinsic drift is rare and obvious when it happens. Extrinsic drift is constant and almost invisible until you measure it.
Telecentric lenses won't save you, but they buy you margin
A good telecentric lens cuts your geometric error budget by ~5x compared to an entocentric lens at the same FOV. That doesn't mean you can skip recalibration — it means small drifts produce smaller measurement errors, so you can tolerate longer intervals between recalibrations.For anything reporting micrometres, telecentric is non-negotiable. For "is this part there?" detection, entocentric is fine.
What actually causes drift
- Temperature. An aluminium frame moves about 23 µm per metre per °C. Most factories swing 10 °C between summer night and summer day. That's 0.23 mm of frame deformation on a 1 m baseline — easily enough to push a precision station out of tolerance.
- Vibration. Camera mounts loosen. Always use locking thread compound on optical mounts. Always.
- Cleaning. An overzealous operator wiping a lens can shift its angular alignment by 0.5°. We've seen it.
- Cable strain. A heavy GigE cable hanging off a small camera, over six months, will tilt the camera. Strain relief is not optional.
A maintenance schedule that actually works
- Daily — a known-good calibration target imaged automatically as part of warm-up. The system computes deviation from baseline. If deviation crosses threshold, the line stops.
- Weekly — manual visual check of the target image overlaid with the baseline.
- Quarterly — full recalibration, intrinsic and extrinsic, with a traceable target.
- On any mechanical event — bump, cleaning, lens swap, cable change → recalibrate.
The daily auto-check is the single highest-leverage piece of this. It's a 30-line script and it has caught every drift we've seen before it became expensive.
Before you ship
Build the calibration target into the cell. Not a loose chequerboard in a drawer somewhere — a permanent, mounted, traceable target the system can reach without operator help. The cells where we've done this have measurably better long-term accuracy than the cells where we didn't.What's your recalibration interval? We're curious whether anyone is doing online drift estimation without a physical target.
