2–5 mm scatter within a series: the parts are never identical, our program already is.

The problem, as it looks on the floor

You run short series of bigger parts. The parts come off cutting and bending, and no two are the same: a flange sits a few millimetres off, a tube is a touch longer, a bracket leans. Within one series the scatter between parts is 2 to 5 mm.

A welding robot does not care about the series. It cares about the path. If the seam is not where the taught path says it is, the torch welds air, or metal that should not be welded.

So the shop does one of three things, and all three cost money:

• builds a precision fixture for each part, so every part sits exactly where the first one sat;
• buys a 3D laser scanner and has an engineer write the measuring program for each part;
• keeps welding that part by hand.

For short series, the fixture and the scanner rarely pay off. The part is still welded by the welder.

What path adaptation does

You teach the cobot by hand, like any cobot. You jog the torch along the seam of a reference sample and save the points. That is the taught path.

While you teach, the camera on the torch looks at every taught point and memorises what it sees there: the edges, the gap, the surroundings of the seam.

On the next part, the cobot goes to the places where it photographed the reference sample. It takes the pictures again, and the AI model finds the characteristic points so the system can compute their correction.

With the found points, the program is recomputed for this part. Then the cobot welds it.

The next part gets the same treatment. No fixture holds it to a tenth of a millimetre. No scanner is programmed. The part is clamped so it does not move, and that is enough.

What it is, and what it is not

• It compensates the scatter between parts of one series, on parts where a few millimetres of difference would otherwise put the torch off the seam. That is where the 2 to 5 mm come from.
• It works before the arc: the path is recomputed, then the part is welded. It does not steer the torch while welding. For a part that distorts under heat or turns while it is welded, there is a separate function, real-time correction, described on its own card.
• The taught path is the program. The person who knows the weld teaches it. The system finds it again on every next part.

Who does this

The person who teaches the reference sample is your welder or your welding engineer, after one day of training at your station. Teaching a new part is a hand-teach, not a programming project you order from an integrator.

On the station

• Rocket Retina welding camera, mounted at the torch.
• Rocket Neuron AI computer, which runs the vision model and recomputes the program.
• A Fairino cobot, or your own robot where we support it.
• A Megmeet MIG/MAG or TIG welding source with the robot communication module.