Train robots.
Own the upside.

Flat-pack from box to finished, including the bit where a screw goes in crooked and comes back out. Recovery is rare in clean datasets and worth a lot.

Wet, reflective, cluttered, outdoors. Every one of those breaks a model trained on a lab bench, which is exactly why the footage is valuable.

An animal that moves on its own while you work. Nothing about it is repeatable, and that is the point.

Tight spaces, awkward angles, two hands doing different jobs. Long enough to show the whole sequence, start to finish.

Repeated grasp and place with slightly different objects each time. Plain, and among the most useful things you can film.

A tool loaded, applied, and reloaded. The full loop matters more than a clean single stroke.