I think this tactic is standard, no? Nvidia has demoed robots being trained using video generation models, and Waymo has been doing that for a while too.

https://waymo.com/blog/2026/02/the-waymo-world-model-a-new-f...