Google DeepMind just plugged Street View into Project Genie, its general-purpose world model, turning two decades of mapping imagery into interactive, simulatable environments. The integration launched at Google I/O 2026, according to TechCrunch AI, which got an early look at the system. What stands out here is the scale of the data feeding it: more than 280 billion Street View images across 110 countries and seven continents, now hooked into a model that can generate explorable worlds on demand.
What launched
Starting today, select Google AI Ultra subscribers in the U.S. can pull a real-world location into Genie and walk through it as a simulated environment. Global Ultra access rolls out over the next few weeks. Users can prompt the model to alter conditions, dropping snow on a Manhattan block they’re about to visit or putting the sun over a London street the local weather rarely delivers.
Key capabilities
- Real-place anchoring. Unlike pure text-to-world generation, Genie now grounds simulations in actual Street View imagery, so the block you load looks like the block you know.
- Condition swapping. Change weather, time of day, or environment on the fly. The team demoed an underwater version of a real neighborhood and a snowy run through Joshua Tree.
- Spatial continuity. Google Maps director Jonathan Herbert told TechCrunch AI the real breakthrough is memory. Turn 360 degrees and the AI keeps the environment behind you consistent, then builds from there.
- Multi-agent perspectives. Waymo’s own simulator runs from the car’s point of view. Genie can shift the camera to a pedestrian or a robot, which matters for training non-vehicle agents.
- Rare-event training. Genie 3 already feeds Waymo’s simulator scenarios like tornadoes and stray elephants. Street View data extends that to specific cities Waymo hasn’t launched in yet.
How it compares
Genie 3 hit research preview last August and opened to Ultra subscribers in January for text-to-world game generation. Adding Street View is the jump from “imagine a world” to “simulate this exact corner of the real one.” Compared to Google’s own Veo video model, which already grasps that smoke disperses and fabric drapes, Genie is still catching up on physics. Research scientist Jack Parker-Holder told TechCrunch AI the gap is roughly six to 12 months.
Who can use it
- Today: A subset of Google AI Ultra subscribers in the U.S.
- Coming weeks: Global Ultra users.
- Pricing: Bundled into the existing Ultra subscription, no separate fee announced.
- Target uses: Robotics training, autonomous vehicle simulation, gaming, and education.
The caveats
Product manager Diego Rivas called the feature an experiment, and the TechCrunch AI demos backed that up. Output is recognizable but video-game quality, not photorealistic. The model isn’t physics-aware yet, so a simulated runner can plow straight through a cactus. Herbert also noted Genie can’t yet produce a faithful reconstruction of a street, only a believable one.
Why it matters
This is Google leaning on a moat almost nobody else has. Twenty years of camera cars and tracker backpacks become training fuel for a world model, and the payoff shows up in robotics and self-driving long before it shows up in consumer gaming. If Waymo can rehearse a launch city in Genie before sending real cars, the cost curve of geographic expansion bends. Same logic applies to any robot that needs to handle a place before it gets there.
The physics gap is the thing to watch. Once Genie learns cause and effect the way Veo has started to, the line between “interactive map” and “functional simulator” gets thin fast. Full details at the original TechCrunch AI report.