Robots already do the easy welds
What is left is free-form

What forty-year-old technology solved

People ask whether welding automation is not already done. Half of that is true. Vertical fillet welding, which follows the corner where two faces meet, is solvable by rules. The intersection of two planes is defined geometrically, so a touch or arc sensing pass finds the line. It takes time but it is not hard, and shipyards already have robots doing it.

The trouble starts after that. The moment it becomes a butt weld, two plates brought edge to edge, a rule-based system has no idea where to aim. There is no intersection. Every part has a different shape, gaps and misalignments differ each time. A person looks and judges from here to there; a rule cannot make that judgement. We bought every welding camera on the market, Chinese products included, and tested them all. Not one of them automated a free-form joint.

What we proved in a shipyard

So when verification with HD Hyundai Samho began, there was scepticism inside the consortium that went in with us. This will not work, was the view. It was also a task another company had approached and withdrawn from.

The result: it worked. From heavily noisy sensor data the model infers the start and end of the weld as keypoints, and the robot takes those coordinates and welds without teaching. We completed on-site automation of free-form flat butt welding. As far as we know it is the first case in Korea.

Why others could not: the sin of splitting perception from execution

This market has an old trap. The vision company says its recognition was accurate. The robot company says its repeatability is excellent. And on the floor nothing runs. The coordinate frames of sensor, CAD, robot and tool drift slightly against one another, the error accumulates, and a person ends up teaching the path again. Then each side says the other gave bad coordinates, or the other's accuracy was off.

That is why LiOps built the full stack from perception to control. Across the whole sequence, from the coordinates the vision sensor produces to the robot moving there and welding, cumulative error is held within 1 mm.

One more thing. Conventional wisdom in this field held that it could not be done without an expensive sensor, and competing solutions do use units costing tens of thousands of dollars. We went the other way. If the spatial model overcomes the noise, the sensor can be cheap. Our welding perception now runs on a sensor costing a few hundred dollars. Precision comes from software rather than hardware specification, and that changes the cost structure completely.

What is left is the work only people did

On a shipyard subassembly line, fit-up and welding take 75% of the labour deployed and roughly 77% of the working hours. That holds even though the structured sections are already automated, because everything left is free-form: overhead welds, end treatments, curved members. These are the jobs where existing welding robots and automation equipment stop and only people continue, and the share grows as work moves into mid and large assembly.

So we decided not to sell robots

If the technology has crossed this stretch, the way it is sold can change too. Instead of selling equipment and withdrawing, we sell the welded result itself, on the same completed-volume terms shipyards have used to buy labour for decades. We call the model the robot welding crew. Because no result means no payment, it is a bet only a company whose technology actually works on free-form joints can place.

Sources: Ministry of Employment and Labour occupational labour force survey (first half of 2024), Ministry of Employment and Labour shipbuilding subcontracting survey (2024), and man-hour data from Professor Jonghun Woo, Department of Naval Architecture and Ocean Engineering, Seoul National University. Site images are from our own verification work. This article sets out our technical and market view and does not describe contract terms with any specific customer.