Skip to content
LearnSign up

Learn / The Academy · 1 min read

World models

AI that guesses what happens next

The short answer

A world model is AI that has learned how things move and change, so it can guess what happens next. Tip a full mop bucket and it predicts the spill. Robot teams use these models to let robots practice in a made-up copy of the world before they touch a real one.

First-person view of a cleaner mopping a tiled hallway, with the wet shine spreading behind the mop.

What you already know without thinking

You know a wet floor stays slick for a few minutes. You know a trash bag full of paper is light and one full of coffee cups will drip. Nobody taught you that from a book. You learned it from years on the job.

A world model tries to learn that same kind of sense from video. It watches what happens, over and over, until it can predict the next few seconds.

Practice without breaking anything

A robot that knocks over a real bucket makes a real mess. In a world model, it can knock over 10,000 buckets and nobody has to mop.

NVIDIA offers a set of these called Cosmos. It calls them "world foundation models" and says teams use them for robots and self-driving cars. They can also create made-up video to add to the real video used for training.

Why made-up video still needs real video

A world model can only guess well about things it has seen. If it has never watched a mop on tile, its guess about a mop on tile will be off.

Ropedia describes its 10,000-hour first-person set as built for world models, among other uses. Real footage is the starting point. The made-up video comes after.

Why this matters to your crew

A world model learns how things behave from watching them. Your crew works on carpet, tile, sealed concrete and glass, in buildings with very different layouts. Real footage from all of those places teaches it how a mop, a rag or a trash liner behaves in each one.

What other owners ask

Is a world model the same as a video game?

It is a bit like one. A game follows rules someone wrote. A world model learns its rules from watching real video.

If a world model can make video, why pay for real video?

It can only copy what it has learned. Real video from real buildings is what it learns from in the first place.

Where to read more

These are public examples from other companies. They are not Colby partners or customers.

Next lessonWho uses this data

Still have questions?

Talk to our team.

Call us and ask anything about pay, your clients or your crew. Or sign up and we will call you.