Why robots need to watch a plumber

Ask a robotics team where their data comes from and you get the same answer over and over. Lab rigs. Teleoperation. A pile of clips scraped off the internet that show the finished result and almost none of the actual work. That is fine for a demo. It falls apart the moment a robot has to do something useful in a real building.

A plumber under a sink is a good example. The job looks simple from the outside. In practice it is a fast sequence of decisions made by feel. Which way the fitting turns. How much force before something strips. When to stop tightening because the next quarter turn cracks the housing. None of that is written down anywhere. It lives in the hands of people who have done it ten thousand times, and it has never been recorded from their point of view.

This is the part people miss. The footage a robot actually needs is not rare because it is secret. It is rare because nobody ever had a reason to film it. A working plumber is not standing there with a camera on their chest documenting how they clear a trap. They are getting the job done and moving to the next call.

So the most valuable training data on the planet is being thrown away every single day, in millions of small jobs, because there was never a way to capture it and never a reason to. That is the gap. Real work, seen from the worker's own eyes, with the small corrections and dead ends left in.

We want the messy middle. The part where the first approach does not work and you try another. The reach into the bag for the right tool. The pause to check before you commit. A clean highlight reel teaches a model that work is smooth and obvious. It is not. The recoveries are where the real skill is, and they are the hardest thing to fake or stage.

That is why we ask for first-person video, chest or head mounted, of actual jobs. Plumbing, electrical, HVAC, carpentry, appliance repair, all of it. Not a tutorial. The work itself, as it happens.

If you do this work, you already own the thing every robotics company is short on. You do not apply to record. You film a real job, submit the footage to an open bounty, and up to five people per campaign get paid in USDC. Inception keeps 10 percent. The other 90 percent is the prize pool that goes to the crowd. The robotics company funds it and never touches the people doing the filming.

Robots will eventually do a lot of this work. Before they can, they have to watch someone who already knows how. That someone is you.

Back to Notes from the field