Why Robot Vision Fails In Ordinary Lighting

Vision is the most common addition to an industrial robot and the most common source of intermittent faults. The reason is that a camera measures light, and light in a real building is not constant.

The camera sees illumination, not objects

An image is a record of how much light reached each pixel. Change the light and the same object produces different numbers, even though nothing about the part has changed.

Algorithms that threshold or match on brightness therefore inherit every fluctuation in the room. A cloud passing a skylight is, to the system, a change in the part.

This is why integrators fight so hard to control lighting rather than to improve algorithms. Controlling the input is cheaper than compensating for it.

Shiny parts are the recurring problem

Machined metal, coated plastic and anything wet reflect a highlight that moves as the part or the camera moves. The highlight is bright, sharp and looks like an edge.

Edge-finding routines lock onto it, and the reported position of the part shifts. The failure is intermittent because it depends on the exact angle the part happened to land at.

The usual remedies are diffuse illumination, polarizing filters, or lighting from an angle that keeps the reflection out of the lens entirely.

The plant changes around the camera

A cell commissioned in winter afternoon light behaves differently in summer. A neighboring line adds a work lamp, a bay door opens, a bulb ages and shifts color.

None of these are recorded as changes to the robot, so when the reject rate rises, the investigation starts at the wrong end. Vision faults are frequently building faults.

Enclosing the inspection station, even loosely, removes most of this class of problem at once.

Dust, film and the slow drift

Lenses and lights accumulate a film of whatever the process produces. The image dims and softens gradually, so the system does not fail; it simply gets worse.

Because the change is slow, operators adapt by loosening tolerances until something breaks. Cleaning schedules for optics are as important as lubrication schedules for bearings.

Some systems monitor image statistics over time specifically to catch this drift before it reaches the reject rate.

Learned models move the failure rather than removing it

Neural approaches handle variation that hand-written rules cannot, which is why they now do surface inspection and bin picking that were previously impractical.

They still fail when conditions leave the range they were trained on, and they fail less predictably, because there is no threshold to inspect. Diagnosis becomes collecting images of the failures.

Either way the discipline is the same: fix the lighting first, and treat the camera as an instrument that needs calibration and cleaning.