Google DeepMind connected Street View to Genie 3, its general-purpose world model, letting users simulate real-world streets with interactive, adjustable conditions. The integration draws on 20 years of Street View data — 280 billion images collected across 110 countries and 7 continents — and launched today during Google I/O for AI Ultra subscribers in the United States. Global rollout follows over the next few weeks.
Users can drop into a real street, adjust the weather, shift the time of day, or simulate disaster scenarios on top of recognizable geography. The robotics use case is immediate: a robot being deployed in London can train on rare sunny conditions so sunlight reflections off Victorian housing don't confuse its vision system when they occur. Genie 3 already powers one of Waymo's simulators, training self-driving cars on edge cases like tornadoes and elephant encounters across 11 US cities.
The Street View integration adds spatial continuity to Genie's world-building. When a user turns 360 degrees inside a simulation, the model correctly remembers and reconstructs the environment behind them, a step beyond static panoramas. From that anchor, the model can layer new conditions — snow in New York City, flooding in Joshua Tree — on top of a spatially coherent base.
Key facts
- 01Google DeepMind connected Street View to Genie 3, drawing on 280 billion images collected over 20 years across 110 countries.
- 02The integration launched today for Google AI Ultra subscribers in the US, with global rollout over the next few weeks.
- 03Genie 3 already powers Waymo simulators across 11 US cities, training autonomous vehicles on rare events.
- 04Spatial continuity is the breakthrough — the model remembers the environment behind you when you turn 360 degrees.
- 05Accuracy lags video generators by 6 to 12 months; the model is not yet physics-aware.
Google released Genie 3 for research preview in August 2025 and opened access to AI Ultra subscribers in January 2026. The model generates interactive game worlds from text prompts or images, targeting educational experiences, gaming, and robotics training. Street View extends that foundation by grounding simulations in real-world geometry rather than purely synthetic environments.
Accuracy remains a bottleneck. The samples Google demonstrated — including an underwater simulation of a recognizable neighborhood — are video-game quality, not photorealistic. The model is not yet physics-aware: in a Joshua Tree simulation, a woman ran through cacti and bushes without collision detection. Compare that to Google's video generator Veo, which intuitively models how paper boats drift on water currents and smoke disperses into air.
Google says the quality gap between Genie and its video models is 6 to 12 months, a timeline consistent with the pace at which world models are closing on video generators in other domains. Physics understanding is learned through passive observation rather than hard-coded, so the lag is expected to narrow as training data and compute scale.
“I think for this kind of model, it's maybe six to 12 months behind video in terms of the accuracy and quality, so I think it's something we will solve.”— Jack Parker-Holder, DeepMind research scientist
Waymo's existing simulator operates from the car's point of view. Genie with Street View allows shifting perspective to other agents — a pedestrian, a delivery robot, a wheelchair user — on the same street. That flexibility matters for training systems that navigate shared spaces rather than highways.
Jonathan Herbert, director of Google Maps and a 12-year Street View veteran, said the team has long considered how to build a richer model of the world on top of Street View data. The Genie integration is the first public demonstration of that ambition at scale.
The model can't yet create a faithful reconstruction of a street — it's building a plausible environment anchored to a real place, not a photogrammetry-level replica. The distinction matters for use cases like urban planning or infrastructure inspection, where millimeter accuracy is required. For robotics training and consumer exploration, plausible coherence is the bar.
Google positions Genie as an experiment, not a finished product. Diego Rivas, a product manager at DeepMind, emphasized the accuracy gap and the need to put the tool in front of users to surface edge cases. The Street View rollout is a test of whether grounding world models in real geography accelerates their usefulness or simply compounds their hallucination risks.
The robotics and autonomous-vehicle markets are the near-term commercial anchor. Waymo's simulator workload demonstrates demand for location-specific training data that doesn't require physically instrumenting every edge case. If Genie's accuracy curve holds, the model becomes a cheaper substitute for real-world data collection in controlled environments.
Google's move also signals where world models compete next: not against video generators on photorealism, but against game engines and simulators on interactivity and spatial coherence. The Street View integration is the first time a frontier world model has married real-world imagery at this scale with continuous, controllable simulation. If the 6-to-12-month accuracy timeline holds, the gap between synthetic training environments and real-world deployments narrows sharply by year-end.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
