Combining Digital Work Instructions and Poka Yoke: The AI Setup That Catches Errors Before Assembly Completes

Digital work instructions and poka yoke devices have been deployed independently in manufacturing for decades. Work instructions deliver process knowledge to operators; poka yoke devices prevent or detect errors during execution. When they are implemented separately, there is a gap between them: the operator receives the instruction, begins execution, and the poka yoke device catches errors only at specific, pre-defined check points.

AI digital poka yoke manufacturing closes this gap by combining instruction delivery with continuous camera-based process observation, so that every step is not just displayed but verified in real time as it is performed.

Why separate implementations leave an error gap

A standard digital work instruction system displays the required steps sequentially and advances when the operator confirms each step. The weakness is that the confirmation is operator-reported: the system accepts that the step was completed because the operator pressed a button, not because it verified the physical outcome.

A standard poka yoke device checks one physical condition at one point in the process. A photosensor confirms that a component is present before the machine cycles. A torque tool interlock confirms that a fastener reached specification before releasing. These checks are highly reliable within their defined scope and blind to everything outside it.

An operator who skips a step not covered by a physical poka yoke device, confirms it in the work instruction system, and moves the assembly to the next station has created a defect that neither system caught.

How the combined system works

The combined digital work instruction and AI poka yoke system uses a camera monitoring the assembly station to observe the physical execution of each step as the instruction displays it. The workflow:

1. The work instruction displays step N with a visual and text description.
2. The operator performs the step.
3. The camera observes the physical result: was the component placed, was the fastener driven, was the sub-assembly correctly oriented?
4. The AI model compares the observed state against the reference for step N.
5. If the physical state matches the reference, the instruction advances to step N+1 automatically.
6. If it does not match, the instruction holds and an alert indicates which element of the step requires correction.

The instruction does not advance based on operator confirmation alone. It advances based on observed physical completion. This eliminates the category of errors created by operators who press confirmation buttons without completing steps.

What types of steps can be visually verified

Not every assembly step has a camera-observable physical outcome. Steps that are well-suited to vision verification include:

• Component placement (was the part placed in the correct location?)
• Component presence and count (are all required components present?)
• Label application (was the label applied in the correct position and orientation?)
• Fastener presence (are fasteners visible in the correct locations?)
• Cable and harness routing (does the routing match the reference path?)
• Sub-assembly orientation before the next operation

Steps that are not well-suited to vision verification include internal torque values, electrical continuity, and pressure tests. These require sensor-based verification, which can be integrated into the same workflow even when the vision system handles other steps.

Deployment experience with the Nagare platform

Nagare’s digital work instruction and poka yoke use case was deployed at an automotive sub-assembly plant producing 14 variants of a steering column sub-assembly. The plant had previously used paper work instructions and two photosensor-based poka yoke devices covering four of the 22 assembly steps.

After deploying the combined digital instruction and camera verification system, the plant’s in-line rejection rate dropped from 3.8% to 0.9% in the first 90 days. The primary error categories eliminated were sequence errors (steps performed out of order) and missing component errors that the photosensor coverage did not reach.