Poka-Yoke: Mistake-Proofing in Manufacturing

Inspection asks whether a defect happened. Poka-yoke makes the question irrelevant, because the mistake that would have caused the defect can't be made in the first place, or can't go unnoticed for more than a second if it is.

The three types

Contact. A fixture, sensor, or physical shape that detects contact, non-contact, or dimension. A connector housing molded asymmetrically so a wire harness only plugs in one way is contact poka-yoke: the mistake isn't corrected, it's physically unavailable. A locating pin sized to one specific part revision works the same way, refusing to seat if the wrong revision shows up on the line.

Fixed-value. Verifies that the correct count of something happened. A kitting tray with a cutout for every fastener a job requires is fixed-value poka-yoke: if a cavity isn't empty when the operator finishes, a fastener is missing from the assembly, not lost somewhere on the floor. A torque driver that counts fastening cycles and won't release the tool until it registers the required number works on the same principle.

Motion-step. Verifies that steps happened, and happened in the right order. A two-hand anti-tie-down interlock on a press won't cycle unless both palm buttons are pressed simultaneously, which prevents a hand from being in the die when it closes. A sequenced torque wrench that won't unlock the next fastener position until the prior one registered in spec is the same idea applied to an assembly sequence instead of a safety interlock.

Prevention beats detection, structurally

The argument isn't about a multiplier, it's about what each stage has already spent by the time a defect is caught. A contact fixture that stops a wrong part at the station costs the interruption of that one cycle. The same defect caught at final inspection means every downstream operation between that station and the inspector has already added labor and machine time to a unit that's now being scrapped or reworked, value that poka-yoke never put at risk because the part never left the station in the wrong state. Caught later still, after shipment, the cost is logistics, a warranty claim, and a customer relationship, on top of everything upstream. Poka-yoke isn't a better inspection method. It's the argument that inspection is the expensive fallback for the mistakes prevention didn't catch.

The modern version: vision as software poka-yoke

Physical poka-yoke depends on a fixture that matches one specific geometry. That works when the part is simple and stable. It works less well when the thing being verified is a completed assembly with dozens of components in variable positions, which is exactly the case vision-based assembly verification exists for.

DePuy Synthes automates orthopedic surgical tray inspection with computer vision, cutting inspection time 47.3%. To be precise about what this is: it's a quality-inspection deployment in our corpus, not a documented poka-yoke case study, but it's doing the same job a shadow-foam tray outline does physically, verifying every required instrument is present and correctly placed, implemented in software instead of a cutout. The functional logic is identical even though the category tag isn't.

The same substitution shows up in Federal Package's edge-learning inspection, which reports over 99% defect detection accuracy at 100% product inspection coverage. That last figure matters more than it looks: traditional sampling inspection accepts a certain escape rate by design, because checking every unit isn't practical by hand. Poka-yoke never accepted that trade-off, and 100% machine-vision coverage is the software version of the same refusal.

The honest limitation is worth stating plainly. A physical fixture prevents the next step from starting at all; a camera mostly detects after the fact, once a frame has been captured. A vision system that only flags a bad assembly on a dashboard, without stopping the part, is inspection wearing poka-yoke's reputation. The deployments that actually earn the comparison pair the computer vision verification with a hard interlock, a station that won't advance or a conveyor that won't release, so the software stops the part the way a fixture would, instead of just noting that it should have.

Where to go next

• The quality control use cases and benchmarks pages cover outcomes across the full 125-deployment corpus.
• Vendors building vision-based verification are ranked on the quality inspection software hub.
• Schaeffler automates EV component inspection with machine vision across 4,800 annual projects, another example of verification moving from a physical fixture to a camera.