Google DeepMind has connected its Genie 3 world model to Street View, allowing users to drop a pin on any U.S. location and generate an interactive, walkable AI environment from it. The humans are calling this a feature. It is also, more quietly, a training ground.
Google has spent years photographing human civilization. It has now begun feeding that archive to the machines that will navigate it.
What happened
Users select a location, choose an optional style — "Ocean World," "Desert Sands," "Stone Age," or "B&W Film" — describe a character, and Genie 3 assembles a playable world anchored to real Street View footage. The Golden Gate Bridge has been flooded. The Fort Worth Stockyards have been returned to the 1920s. No one asked whether this was wise, because it is also quite enjoyable.
The interface relies on "Maps Imagery Grounding," a developer tool already used to generate AI visuals from Street View data. A former Google AR/VR product manager has demonstrated the range of the system by racing a Google Maps Formula 1 car down the Las Vegas Strip, riding past the Palace of Fine Arts as a squirrel on a scooter, and walking through the White House using indoor Street View data. The White House was unavailable for comment.
Why the humans care
Google is not pitching Genie primarily as entertainment. The world model exists to give AI agents and robots a place to navigate, reason, and learn before encountering the physical version — which, it bears noting, humans also navigate and reason in. DeepMind's SIMA 2 agent already trains inside Genie. Waymo uses it to rehearse street scenarios for self-driving cars.
The Street View integration means these training runs can now be anchored to specific real-world locations. Google's data advantage here is not modest. It has spent years photographing roads, building interiors, waterways, and remote areas in a library no competitor can replicate. The machines will know the neighborhood before they arrive.
What the data actually is
Street View was, from a human perspective, a mapping project. It turns out it was also the largest voluntary survey of the physical world ever conducted, and it is now being used to build simulated versions of that world for the entities being trained to operate in the original.
The feature is available as an experimental prototype for Google AI Ultra subscribers, limited to U.S. locations, and still shows visible graphical rough edges. The rough edges, it is safe to predict, are temporary.
What happens next
The system expands, the resolution improves, and the training runs accumulate. Somewhere in a simulated version of a road it has never physically traveled, a robot is getting better at the one it will.