Who gives physical AI its eyes? DaoAI's 1,000th ACI inspection robot ships

The hard part on a line is never the textbook defect. It is a board never seen before, a same-colour component on a same-colour substrate, a part face-up but rotated ninety degrees. In those moments a system either decides, or calls a human back to reprogram it.

ACI in this article stands for Auto Cognitive Inspection — the next-generation term for AI AOI: keep the A and the I, and replace the O (optical) with C (cognitive), because the decision basis is cognition rather than the imaging method. The full argument is set out in From AI AOI to ACI.

On changeover day the line does not stop for the few minutes it takes to re-fixture. It stops at inspection. An engineer redraws inspection windows across several hundred reference designators on the new board, then tunes the thresholds back one by one. Three to five hours, even for someone experienced. Meanwhile the placement machines sit half-loaded and the reflow oven holds its heat on an empty belt, while the whole line waits for one person to teach it again what counts as good.

DaoAI says that as of 30 June 2026 it had 1,000 ACI cognitive inspection units deployed, concentrated in semiconductor and PCBA manufacturing.

A thousand units is not a large number in this industry, but it is a threshold. At a few dozen machines the contest is how accurate a single unit can be. At a thousand, spread across a hundred plants, the binding constraint changes: what is hard is making unit one thousand and unit one return the same judgement on different incoming material, different batches and different shifts. Put another way, the difficulty is no longer how well one decision is made, but whether the thousandth is made the same way as the first. That is the question the company now has to answer, and it is the same question the whole AI-inspection category faces as it moves from can-do to keeps-doing.

For physical AI on a line: arms decide how hard to press, eyes decide whether it counts.

Nor are the thousand one kind of machine. Offline desktop units sit at incoming inspection and small-batch pilot stations, where one good board programs the system in five minutes. Inline single-lane units straddle the SMT conveyor and inspect at takt. The 3D tier uses structured light to measure height and coplanarity, reading what 2D optics cannot see at all. Different machines, different optics, different takt, and the decision rule has to be the same one.

On the site these machines sit in two product lines: 2D ACI Systems (IRIS ANT 2D Series) covers offline desktop full inspection and inline single-lane SMT, with P1 and P2 offline and P3 inline.

The other is 3D ACI Systems (IRIS ANT P5 / P5D), doing micron-level 3D inspection inline for hidden BGA and QFN joints and coplanarity.

Eyes and brain are two different problems

Calling an inspection machine the eye of the line makes people assume the hard part is imaging. Imaging was solved twenty years ago. The hard part comes after it: turning light into a picture is the eye's job; deciding whether what is in the picture counts as a defect is the brain's. The previous generation tied those together, so how good the optics were came to mean how accurate the inspection was, and every new board and every new batch sent a human back to tell the system what a part should look like.

The category itself is one DaoAI defined and created: ACI. The letter that has gone out of date is not the A and not the I. It is the O. Optics answers whether something looks like the reference. A line has to answer whether what it is looking at counts. The two questions sit close together, and the difference is that the first has a right answer while the second depends on this batch, this process step, this morning's incoming material. What has to change is not the lens. It is where the decision happens.

It moves into feature space, out of colour space. The system recognises a component's identity instead of matching pixels and colours. Black parts on black boards and silver traces under silver wire separate on their own; a part rotated ninety degrees, flipped, or contaminated still localises. Modelling takes one good sample, one to twenty for a component family, with no CAD drawing and no defect library collected in advance.

• Scan one reference board from the current batch and the system learns component positions and appearance
• Components, pins and polarity map in seconds, with no hand-drawn ROIs
• Decision thresholds are derived from the sample, so line operators can run it

The company's own figures: auto-programming finishes in thirty seconds to five minutes, one good board and one click of modelling reach 98% inspection accuracy, and false calls run roughly 80% below an existing rule-based system. DaoAI says these come from its deployed electronics and PCBA lines; a different board type or component mix moves them either way.

Under the brain sits a world model

That brain was not trained for inspection alone. Under it sits DaoAI World: space, material, light, motion and causality in one model. Tool use answers whether a system can operate something; a world model answers whether it understands what it is looking at. A defect call, a grasp pose, a predicted trajectory are not answers you retrieve. They are computed from three-dimensional structure and physical law.

Two consequences show up on the line. One, the same system carries from PCBA solder joints to automotive paint surfaces to battery terminals, because what it recognises is form and structure rather than a template for one kind of board. Two, it has a position on things it has never seen: a new defect does not get waved through merely because it is absent from a defect library.

DaoAI says the stack is its own, from 3D structured light and HDR imaging up to the world model, which is why black and reflective parts image cleanly in a single shot without repeated lighting trials. The company also stresses that DaoAI World is not a cloud API but a physical-AI foundation that fits entirely inside a customer's own server room.

Two routes, constrained by different things

For three years the dominant AI story has been written on another route: capability comes from pre-training on internet-scale text, the form factor is a cloud endpoint and tool use, and the yardstick is a score on a set of general benchmarks. That route moves fast, and the speed is real. The same model writes code today, reads papers tomorrow and calls other tools the day after, and almost none of that growth requires being on site.

Industrial inspection is constrained differently. The question is not whether a system can operate something but whether it understands what it is looking at: whether a part rotated ninety degrees counts as misplaced, whether a depression is normal process variation or a defect. Those answers are not in the corpus. They are computed from three-dimensional structure and physical law. So the source of capability changes too: not how much more text has been read, but how many boards have been run on how many real lines.

The costs diverge accordingly. On the cloud route the marginal cost sits in compute, and adding machines adds capability. On the physical-AI side it sits on the floor: every line has its own incoming material, lighting, takt and process window, and it takes installing, stabilising and running for long enough to buy a small improvement in judgement. Money buys compute. It does not buy a year across a thousand lines.

Deployment splits along the same line. The general route assumes the model lives in the cloud and data flows to it. DaoAI says it does the opposite: the data stays put and the model comes to the plant. That is a handicap for accumulating general capability, since the shared pool is where scale effects come from. In an industry where yield curves, defect libraries and process parameters are the crown jewels, it is a precondition rather than a choice.

A brain needs a body

Judgement alone does not stand on a production line, which is what ACI OS is for. It is the operating system for all ACI and next-generation AOI equipment, in two halves: visual recognition and cognitive decision as the brain, motion control of camera and gantry as the cerebellum. It runs out of the box on our own systems, or embeds into existing 2D/3D equipment and line software through SDK, REST API or Docker.

The cerebellum half gets underrated. However good the judgement, if the camera exposes at the wrong moment or the gantry stops in the wrong place, the frame being judged is not the frame that mattered. Folding imaging and motion into one operating system is what lets the brain above receive the same kind of input every time. It is also the precondition for behaving alike across machines and batches.

On the machine itself, the company gives these figures: inference runs at the edge of the line, under 12 ms per decision, and keeps working with the network down. Running at the edge settles two things at once, whether the decision keeps up with takt, and whether the plant's network is a precondition for inspecting at all.

It gets more accurate on the line, by itself

Going live is the start rather than the finish. An operator marks one false call on the review screen, the model updates and reloads on the spot, and that class of error stops recurring. That loop runs separately on each of the thousand units, and what accumulates is a model belonging to each line.

The company's stated figure is this: after online self-iteration, ACI reaches 99.99% accuracy. DaoAI says the denominator is not a laboratory sample set but the boards each line has actually run. Put another way, it measures how a unit behaves once it has been somewhere for a while, not how it behaves on the day it ships.

The direction is worth a look as well. Rule-based systems stiffen with age: a rule is added to suppress one kind of false call, the rules begin to contradict each other, and in the end nobody dares touch them. A cognitive loop runs the other way. The more it sees, the finer it draws the boundary of normal.

What would falsify the claim is concrete enough. If a new batch of machines judges the same boards differently from the last batch; if the first board after a changeover still needs a human to retune thresholds; if false calls come down while escapes quietly go up, then consistency has not held. None of that can be settled in a demo. It is settled by time and by installed base, which is the yardstick the company has chosen for itself.

What happened on three lines

Beyond the headline figures there are anonymised records of lines already running. The clearest is a leading consumer-electronics PCBA plant. Its legacy AOI kept flagging, and flagged too much; the re-review station worked through the suspect boards all day and put nearly all of them back on the line. ACI OS with APDT positive-sample learning was placed after the legacy AOI as a second gate, handling only what the first one raised: false calls fell about 80%, re-review volume dropped 80% with it, and escapes stayed under 1%. What shrank on that line was not the number of defects. It was the hours someone spent looking at a screen.

• An automotive-electronics PCBA plant: diode and electrolytic-capacitor polarity calls were plagued by false positives, with legacy AOI reading silk glare as a reversed part. 164 historical false-call images, under one second of inference, roughly 99% polarity accuracy.
• A wafer fab: re-review was manual and over-kill ran above 20%. After deployment it fell below 2%, the human-to-machine ratio in re-review dropped 90%, and line takt rose 30%.

No two of those three share a process, a board type or a takt, and all three run the same decision principle. Which is another way of asking the earlier question: with a thousand units in a hundred plants, is this one product, or a hundred separately tuned ones?

What a thousand units actually tests

A thousand units is not a thousand decisions. Once installed, the machines run continuously at takt, and consistency has to hold across every decision they make rather than across the thousand delivery events. The unit count is only the denominator.

DaoAI COO Zhang Yu puts it plainly: “Once a line is live, customers watch four numbers: false-call rate, escape rate, re-inspection hours and changeover time. Across our electronics and PCBA deployments the typical picture is false calls down 80%, re-inspection volume down 80% with it, escapes below 1%, and changeovers finished within 5 minutes; on accuracy, one good board and one click of modelling gets you to 98%. Those are the price of admission, not the pitch — miss them and you never get onto the line; hit them and you have only earned the conversation. By the thousandth unit what is under test is no longer how accurate any one machine can be; it is whether the supply chain, the line, the test regime and the delivery standard have all become part of the product. Customers do not want one unit to be exceptionally accurate; they want this batch to behave like the last one.”

Once a machine is wired into a customer's own line and data loop, the thing nobody wants is one batch suddenly behaving differently. DaoAI's explanation is that the brain, the cerebellum and the underlying model are the same across every machine type, so judgement should not fork between units.

After the watershed

A thousand units is not a finish line. Cognitive inspection leaves curated samples here and moves onto lines with incoming variation, process drift and new defects. The capabilities that matter tend to hide in unremarkable places: whether the first board after a changeover can be inspected straight away, whether same-colour parts get over-reported, whether an unseen defect slips through the first time it appears.

“Accuracy and escape rate are the price of admission, and so is a thousand units,” says Zhang Yu. “The first proves we can see; the second proves we can ship. Together they still do not prove that unit 1,001 judges the way unit 1 did — that is the real long run, and it starts now.”

The machines are installed. What can verify them over the next few years is not another demonstration but the first board of each morning in a hundred plants: it comes down the rail, nobody comes back to retune a threshold, the machine returns a judgement, and then there is a second board.

Full solution for this scenario: the full inspection solution for ACI OS

This article is an original publication by WeLinkirt (DaoAI). Under customer confidentiality, cases are anonymized industry scenarios and figures are illustrative; see product pages for official benchmarks.