Google just plugged 20 years of Street View into its Genie world model

If you've ever dropped the Google Maps pegman into a random city just to kill time, Google DeepMind just turned that nostalgia trip into an interactive playable sandbox.

At Google I/O 2026, the company announced that its general-purpose world model, Project Genie, is now directly hooked into Google Street View. Instead of clicking through static panoramic snapshots, users can now pick a real-world location and prompt Genie to turn it into an explorable simulation.

Stepping inside an AI-altered reality

Google has spent the last two decades amassing more than 280 billion Street View images across 110 countries and seven continents. That mountain of visual data is now serving as the physical foundation for Genie's generative worlds.

Inside the Project Genie prototype, you drop a pin on a U.S. location and choose a visual theme like Ocean World, Desert Sands, or B&W Film. You can define your own avatar, whether that is an animal or a claymation monster, and Genie builds an interactive environment anchored right to the real coordinates. Think scuba diving around the submerged pillars of the Golden Gate Bridge, or wandering through a 1920s vintage version of the Fort Worth Stockyards.

According to the team, the real technical breakthrough here is spatial continuity. If you spin 360 degrees, the model remembers what was behind you and maintains a coherent space as you walk around.

That said, it is still very much a research experiment. DeepMind researchers noted that the visuals resemble video game graphics rather than photorealism. More importantly, the model currently lacks real-world physics. In test runs, an avatar sprinting through a snowy desert simply clipped straight through cacti and bushes. Researchers estimate that world models are roughly 6 to 12 months behind video generation models in accuracy.

Synthetic sandboxes for robots and Waymo

While turning cities into post-apocalyptic playgrounds is fun, the real business case is embodied AI and robotics.

Waymo is already tapping Genie to generate freak edge cases for its autonomous fleet, such as sudden tornadoes or random elephant encounters on the highway. But traditional self-driving simulators are strictly locked to a car hood perspective. Street View allows DeepMind to break the camera free and simulate the viewpoint of human pedestrians or bipedal delivery bots.

A DeepMind researcher pointed out that a delivery robot deployed in cloudy London could be trained on rare simulations of blinding sun glares bouncing off brick houses before ever encountering them on asphalt.

Access is currently rolling out to Google AI Ultra subscribers globally, but it carries a steep $200 monthly price tag and is initially limited to U.S. coordinates.

Why it matters

Connecting a dynamic world model to twenty years of real street imagery is one of the smartest asset plays Google has made in recent years.

For robotics and self-driving engineers, real-world data collection is brutally expensive, especially for rare edge cases that could cause fatal crashes. By wrapping Street View in generative physics, Google is essentially building an infinite simulation loop where agents can train on thousands of weird scenarios before touching real pavement.

It also highlights Google's unbeatable moat against pure-play AI labs like OpenAI. You can train on endless public web text and synthetic video, but you cannot easily replicate 280 billion geotagged panoramic street photos collected by camera cars and backpackers over two decades.

Still, the lack of real physics remains a massive hurdle. A world model that allows characters to pass through solid walls can easily teach a robot bad habits if applied too early. Until Google solves basic collision dynamics and drops that punishing $200 paywall, this feels less like a practical production tool and more like an impressive, high-priced glimpse of what robotics training will look like next year.