Run the Defect-Definition Test
Automated inspection is sold as a camera problem. It is almost always a definitions problem: two experienced inspectors on the same line will disagree about borderline parts, and a vision system has to be told which of them is right.
The Bot Scout defect-definition test comes before any vendor visit. Collect fifty borderline parts, have your inspectors classify them independently, and measure how often they agree. That agreement rate is the ceiling on what any automated system can achieve, because it is the quality of the ground truth you will train and judge against.
Where agreement is high, automation is straightforward and worth doing. Where it is low, the project is a standards exercise first, and buying a camera will simply relocate the argument.
| Inspection type | What decides feasibility | Common blocker | What to specify |
|---|---|---|---|
| Dimensional check | Fixturing and repeatability | Part moves between shots | Tolerance and datum scheme |
| Surface defect | Lighting and contrast | Defect invisible under flat light | Lighting geometry and sample set |
| Presence or absence | Simple and reliable | Rarely the real requirement | Complete part list |
| Assembly verification | Angle and occlusion | Feature hidden at the inspection pose | Views required per part |
| Cosmetic grading | Human agreement rate | Inspectors disagree with each other | Written grading standard first |
Lighting Is the Real Purchase
Most surface-defect projects are decided by illumination rather than by camera resolution. A defect that is invisible under diffuse light can be obvious under a grazing angle, and no amount of processing recovers information the lighting never captured.
Ask vendors to demonstrate on your worst parts, in your ambient conditions, with the lighting they intend to install. A demonstration in a controlled booth proves the camera works, not that the cell will.
Cycle time is the second constraint. Inspection that adds seconds to a line running at pace can cost more than the scrap it catches, which is the baseline arithmetic in the robotics in manufacturing guide.
- Measure inspector agreement on borderline parts before buying.
- Write the grading standard before requesting a demonstration.
- Require testing on your worst parts in your ambient light.
- Specify views per part, including occluded features.
- Name who reviews rejects and retrains the system.
Who Owns the Data Afterwards
An inspection cell produces a stream of images, classifications, and reject records. That data is the asset, and it is worthless if nobody has been assigned to read it.
Ask where images are stored, how long they are retained, who can retrain the classifier, and what happens to performance when a supplier or process changes. A vendor-only retraining model turns every material change into a support ticket with a lead time.
Compare the inspection cell against a mobile or route-based approach where the requirement is coverage rather than per-part checking, using the industrial robots hub and the inspection robots guide.
Bottom Line
Robotic quality inspection succeeds where the defect is defined, the lighting reveals it, and a named person owns the data. Establish inspector agreement before comparing vision vendors.
Classify fifty borderline parts with two inspectors and measure their agreement before any vendor visit.
FAQs
Why do automated inspection projects fail?
Usually because the defect was never defined precisely enough. If experienced inspectors disagree on borderline parts, the system has no reliable ground truth to learn or be judged against.
Is camera resolution the most important factor?
Rarely. Lighting geometry decides most surface-defect applications, because processing cannot recover information the illumination never captured.
How should a vision system be demonstrated?
On your worst parts, in your ambient conditions, with the lighting the vendor intends to install. A controlled-booth demonstration proves the camera works, not the cell.
What happens when the process or supplier changes?
Performance can degrade without anything mechanical changing. Confirm who retrains the classifier, how quickly, and at what cost before signing.