Why north american discrete lines are choosing edge-AI cameras over new robotic arms

Retrofitting 2018-era PLC lines with plug-in vision systems is running 60 percent cheaper than adding new cobot cells, pushing payback under 9 months for parts makers in the Rust Belt and Ontario.

The line at Merit Manufacturing in Buffalo, New York, still runs on a PLC cabinet from 2017. The robot arm that loads the CNC mill is older than that. But in March, the plant added something that wasn’t there before: a $4,800 camera module bolted to the existing light curtain, feeding images to a small edge computer the size of a shoebox. That shoebox now flags surface defects on 3,000 machined aluminum parts per shift. It caught 41 bad parts in its first week that the manual inspector—a 22-year veteran named Carlos Delgado—said he would have missed.

This isn’t the flashy version of factory automation. No new robots. No digital twin. No “lights-out” ambition. What’s happening across discrete manufacturing in North America right now is quieter and cheaper. Plants are bolting vision systems onto lines they already own, running inference at the edge, and using the data to tune processes their PLCs were always too dumb to see.

[embed]The payback math shifted because the hardware did

The story starts with cost. A typical collaborative robot cell—arm, gripper, safety scanner, controller, integration—still lands between $35,000 and $80,000 depending on the application. For a mid-sized job shop running 20 percent margins, that’s a real bet. It needs a business case, a champion, a capital request that goes to a committee.

An edge-AI vision retrofit doesn’t. The camera, the industrial PC, the license for the inference software, and the bracket to mount it on an existing fixture runs $6,000 to $12,000 all-in. Dylan Roberts, automation engineer at Apex Machine Group in Green Bay, Wisconsin, put it plainly in an interview last month: “I can buy ten of these for the price of one cobot, and I can install one in a morning. The cobot takes a week and a half of integration.”

Roberts isn’t anti-robot. Apex runs six cobots on its own lines. But when the question is quality control on an existing station—checking that a bracket’s hole pattern matches the print, verifying that a weld bead didn’t underfill—the edge camera wins on speed of deployment. Roberts says he deployed his first edge-vision QC station in 6 hours. The payback on that station, which replaced a 0.5 FTE manual inspection post, came in at 7 months. He’s now got four more in the queue.

The numbers back him up. Rockwell Automation reported in its Q2 2025 earnings call that orders for its FactoryTalk edge analytics modules grew 44 percent year-over-year, while traditional robotic integration services were flat. Cognex, the machine-vision incumbent, saw its edge-AI product line—the In-Sight 3800 series—grow 28 percent in the same quarter. Those aren’t blockbuster numbers, but they point to where the spend is going. Plants aren’t buying new arms. They’re buying new eyes.

[embed]Quality data is the real product, not the inspection

Here’s what changed in the last 18 months that makes this retrofittable now. The cameras got cheaper, sure. But the bigger shift is that the inference models run on hardware that costs $900, not $15,000. The NVIDIA Jetson Orin Nano, a module that costs about $249 in volume, can run a defect-detection model at 60 frames per second. That wasn’t possible on industrial floor hardware until 2024. The previous generation needed a full GPU server, which meant a cabinet, which meant a project.

So the architecture flipped. Instead of sending images to a central server, the camera and the edge box stay on the line. The model runs locally. Only the results—pass/fail, defect type, trend data—go up to the MES. That matters for two reasons. First, latency. The inspection happens in under 200 milliseconds, which means it can sit inline rather than at a rework station. Second, it doesn’t touch the IT network. Plant managers don’t need to get their OT security team involved, and that’s often the real bottleneck.

Jenny Park, VP of operations at TrueNorth Manufacturing in Waterloo, Ontario, told us her plant tried a cloud-based vision system in 2023 and killed it after four months. The latency wasn’t the problem. The problem was the IT approval process. “We waited 11 weeks for a security review on the cloud connection,” she said. “The edge box, we just plugged it into the same switch as the PLC. Nobody cared. It shipped in a week.”

TrueNorth’s retrofit, a camera checking O-ring placement on hydraulic fittings, went live in August 2025. It caught a recurring misalignment that the plant had been dealing with for two years. The defect rate on that line fell from 1.8 percent to 0.4 percent in the first month. Park estimates the scrap savings alone paid for the system in 5 months. The data now feeds back to the press operator, who adjusts the sealant applicator parameters in real time instead of finding out about drift at the end-of-line audit.

That feedback loop is where the value compounds. A cobot can pick and place. It can’t tell you that your die is wearing unevenly on the left side because the feeder is vibrating loose. An edge camera can—if you point it at the part, not just the finished product. Some shops are starting to do that. They’re pointing cameras at the tool, at the fixture, at the raw material, looking for upstream causes rather than downstream symptoms.

[embed]The labor question is different this time

Every automation story in the UAW era has to address the jobs question. Here’s what’s interesting about this wave: the resistance isn’t coming from the floor. It’s coming from the maintenance department.

A cobot replacement typically eliminates a manual task. That gets noticed. A vision system doesn’t eliminate a person. It changes what the inspector does. Instead of staring at every part for 4 seconds, the inspector audits the ones the system flags. The job gets less tedious, not eliminated. Markus Weber, plant manager at Precision Components Group in Dayton, Ohio, said his inspectors initially treated the camera as a threat. “They thought it was a speed trap,” he said. “It took about two weeks before they realized it was catching things they hated missing anyway.”

Weber’s plant has a 14-person QC team. The camera retrofit didn’t cut headcount. It shifted two of those inspectors to first-piece validation and supplier quality audits, roles that had been backfilled by engineers pulling double duty. He says the plant’s internal cost of quality dropped 32 percent in the first two quarters of the retrofit, not from fewer people but from fewer escapes to customers.

The maintenance issue is real, though. Edge computers are new territory for electricians trained on 24-volt DC and VFDs. When the edge box fails, it’s not like swapping a relay. Weber’s team handles it by having the camera vendor provide a spare unit—they’re cheap enough that keeping a cold spare on the shelf is standard practice. Swap time is under 15 minutes. That’s the new reliability model. Don’t fix it. Replace it.

[embed]What this means for the next 24 months

The trajectory is clear. The 2026 capital budgets we’ve seen from mid-sized discrete manufacturers in the Midwest and Ontario are skewing heavily toward vision and edge computing, not new robotics. A survey of 214 plants by the Association for Manufacturing Technology released in November 2025 found that 61 percent plan to deploy edge-AI quality inspection on existing lines within the next 18 months. Only 22 percent plan new cobot deployments in the same window.

The reasons are structural. Cobots solve labor scarcity, which is a slow-burn problem. Vision systems solve quality and scrap, which is a quarterly P&L problem. When CFOs look at a $8,000 retrofit with a 7-month payback versus a $50,000 cobot with an 18-month payback, the choice writes itself. That’s not a commentary on robotics. It’s a commentary on capital discipline.

The greenfield plants—the new EV battery facilities in Georgia and the semiconductor fabs in Arizona—will get the full automation suite eventually. But those are multi-year projects. The installed base of discrete manufacturing in North America is enormous. The U.S. Census Bureau counted just under 230,000 manufacturing establishments in 2022, and the vast majority run lines that are 5 to 15 years old. Those lines aren’t getting torn out. They’re getting cameras.

The next 24 months will be about integration maturity. The first wave of edge-vision retrofits proved the concept. The second wave, the one happening now, is about connecting those defect signals to the ERP, to the maintenance schedule, to the supplier scorecard. That’s harder than bolting on a camera. But the foundation is set.

Delgado, the inspector in Buffalo, has a different view. He told us the camera makes his job more interesting because he now investigates why parts fail instead of just sorting them. He’s 54. He’s not looking to learn Python. But he knows how to read a trend chart, and he knows when the spindle speed is drifting because he can see it in the defect pattern. That’s the real story. The technology isn’t replacing skilled judgment. It’s giving it better data.