Vision Error Proofing versus Mechanical Poka-yoke: How to Choose

Mechanical error proofing and vision error proofing are frequently framed as a binary choice, and that framing is where most of the argument comes from. The two do genuinely different things. A fixture makes an incorrect action physically impossible — the part does not fit unless it is oriented correctly, the door does not close unless the tool is in its holder. A camera makes nothing impossible; it judges whether something happened and reports a verdict. One prevents, one detects. Once that distinction is clear, the selection question becomes tractable: does this requirement mean "it must not happen", or does it mean "we must know whether it happened"? The first is a fixture problem, the second is a camera problem.

Where the two genuinely differ

The fourth row is the one that decides most real cases. A fixture is very reliable within its own scope and tells you nothing outside it — when a defect escapes, you have no idea whether the fixture worked and something got past it, or the fixture was never involved at all. A camera is less reliable per condition but leaves a record of every judgement, which is what makes root cause analysis and process improvement possible at all. So the choice is not only about error prevention; it is also about whether you want process data afterwards. Where a quality system or a customer audit requires traceable evidence, that consideration alone often settles it.

When a fixture is clearly the right answer

• The error has a definite physical form, such as a part that can only go in one way. A locating feature solves it completely and permanently, with no threshold and no false alarms.
• The consequence is a safety risk. Where someone could be injured, physical prevention is the right layer; no vision system should be the last line of defence for personal safety.
• The cost of an escape is extremely high and the error form is stable, as in high-value assembly where a single orientation mistake is catastrophic.
• The takt is extremely tight and the judgement is simple, where the few milliseconds of image processing are themselves a problem.

The limitation on all four is the same: a fixture can only prevent an error whose shape it was built for. Add a new variant, change a material, introduce a part that looks similar but differs in a way the fixture cannot feel, and it stops working — quietly, since a fixture that no longer applies still feels like it is doing its job. That is the failure mode nobody notices until it is expensive.

When vision is clearly the right answer

Vision wins wherever the requirement is about knowing rather than preventing. Sequence and order — a fixture can enforce order within its own cycle, but cannot report that step four was skipped when it has no idea what happened at step four. Action correctness in manual assembly, where the whole point is that a person performs the work and you want to know whether they did it as the standard specifies. High-mix production, where a fixture per variant would mean a rack of fixtures and a changeover for each, while a camera can handle a new variant with a criterion update. And multi-condition stations, where one camera covers the presence of a part, the attitude of a tool and the completion of a step in the same frame. The specific patterns are covered in preventing missing operations and preventing skipped steps.

Cost accounting has been known to mislead

A fixture often looks cheaper at quotation and more expensive over its life. Design and machining are a one-off, and running costs are almost nil, which suits a single fixed product. But a product change means a new fixture, and high-mix production means a rack of them plus changeover labour. Vision inverts this: moderate hardware and setup cost, plus ongoing threshold tuning, but a product change may only need a criterion update. The comparison should also include the cost of a fixture's silent failure — a fixture that has stopped matching the product does not break or alarm, it simply fails to stop the error it used to stop. Placing that on the same sheet as the quotation is what makes the comparison honest.