World models train robots and self-driving cars in simulations, while humans collect their data

If you have trained a model on data that lacked coverage of rare cases, you know the problem: the real world is too expensive for experimentation. You cannot drop a robot a million times or deliberately send a self-driving car onto icy roads. World models fill precisely this gap, and forecasts suggest they will move significantly closer to everyday life in 2027.

How this differs from LLMs

Oliver Cameron, CEO of the Odyssey research lab, explains the difference simply. A language model takes a sentence and predicts the next word. A world model takes a visual state of the world, such as an image, and predicts the next state. For this to work, the model needs to understand physics, cause and effect, and how people and objects respond to our actions. That is how Jamie Shotton, Wayve's chief scientist, defines the task.

In practice, the result is a generated 3D environment that looks and behaves like the real thing. It is a training ground where mistakes cost nothing.

What this looks like in practice

Brian McClendon, CTO of Niantic Spatial, describes a scenario involving a household robot. You photograph the room where it will work, a 3D model is built from the images, and the software then runs millions of simulations so the robot can learn the layout: where the table is, where the threshold is, and what can be moved.

Niantic is well positioned here. Pokémon Go players spent years scanning streets, and the company, which sold the game last year, is now betting on reconstructing the real world. Wayve is taking a different approach: its generative model draws partly on dashcam recordings to recreate dangerous roads and conditions that cannot safely be tested in real life.

Gaming is a third application. Physics and 3D rendering are fundamental to games, which is why Tencent is investing in its own world models.

Money and the hidden cost

The deals show investor interest. Yann LeCun, Meta's former chief AI scientist, raised more than a billion dollars for his startup Advanced Machines Intelligence. AMD acquired Fei-Fei Li's World Labs, founded in 2024, for more than $8 billion this year. Google continues to regularly tout improvements in this field.

There is also a less visible side. To simulate the physical world, labs need recordings of that world, and real people make them. The author of a Wired article strapped an iPhone to his forehead and spent several hours filming himself folding laundry, washing dishes and pouring cola, then sold the footage through an app for a modest sum. If companies keep pursuing higher quality, this vast data-collection gig economy will only grow.

Limitations and what to do

The article is primarily an overview: it provides no specific metrics for simulation quality, comparisons with real-world testing, or timelines for the widespread arrival of household robots. The forecast for 2027 remains a forecast.

For practitioners, the takeaway is this. If you work in robotics, autonomous vehicles or games, world-model simulation is worth considering as an inexpensive way to cover rare and dangerous cases. If you are a product manager or founder, pay attention to the data shortage: whoever has unique recordings of the physical world will have an advantage. And those collecting this data should think ahead about consent, privacy and fair compensation, because this is where regulators and users are most likely to start asking questions.