A proprietary version of this technology is already training Waymo's robot cars. Waymo blogged about their DeepMind-provided world model a year ago.
> It’s so obviously not representative of the real world.
That entirely depends on the use-case and level of granularity needed, doesn't it?
This blog is suggesting that we could train robots using only a world model.
> Do you have a basis for this idea, or it's a gut-feel?
My basis is that I’m a real human and these videos look far more like bad video games than the real life I’ve experienced.
> Isn’t the Waymo version used for some small portion of the dataset to cover edge cases that they have no real data for?
The usecases in the blog post[1] paint a much more expansive picture of how Wayne used Genie 3: including enabling transfer learning to convert single-camera dashcam footage into simulated full-take sensor data (multi-camera and Lidar), which is extremely valuable.
> My basis is that I’m a real human and these videos look far more like bad video games than the real life I’ve experienced
1:1 fidelity isn't required for a lot of training tasks; for instance, computer vision can do a lot with 32x32 resolution images, which is trash quality to humans. I think the potential customers for this technology will constrain the model to fit their use cases.
1. https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-f...
They clearly know how to train robots. Unless Genie is a misdirection...
When your training simulation fails to reproduce certain features of the real-world environment, your control model is likely to have trouble with them, so more realistic simulations are helpful, but even fairly unrealistic training simulations are pretty useful.
In general, training neural networks to control robots is going to require simulations, ideally differentiable simulations. Even if they aren't differentiable, though, you can SGD your simulated training runs across a data center in a way that you just can't with training runs on Roombas.