DeepMind links Genie 3 world model to Street View's 280B images

280 billion photographs, taken across 110 countries over two decades, now power a generative world model that simulates real locations in real time. At Google I/O this week, Google DeepMind announced that Project Genie -- its system for creating interactive artificial intelligence environments -- has been wired directly into the full Google Street View archive.

The result: users can drop into a simulated New York block blanketed in snow, or a London street on a rare sunny afternoon, and navigate it interactively. Genie 3, the latest iteration of the model, first appeared as a research preview in August 2025 and opened access to Google AI Ultra subscribers in the US in January 2026. The Street View integration is rolling out to some Ultra users now, with a global expansion expected over the coming weeks, per The Next Web.

Street View's archive spans all seven continents and represents one of the most geographically diverse visual datasets ever assembled. Prior world models have mostly trained on curated synthetic data or narrow game environments. Connecting Genie 3 to this corpus is a direct argument that grounding in real-world imagery changes what these systems can produce.

Two audiences, one model

Jack Parker-Holder, a research scientist on DeepMind's open-endedness team, told The Next Web that the feature is designed to serve two distinct groups. Robotics and autonomous vehicle engineers can use it to generate simulated environments that mirror actual locations, training agents on scenarios that rarely appear in physical testing. Regular users, by contrast, can simply explore -- wandering through places they have never visited.

Waymo is already in the first group, using Genie 3 to train its self-driving systems on rare and dangerous edge cases. When a specific road configuration or weather condition appears only once in millions of miles of real driving, a world model seeded with Street View imagery from comparable locations can generate thousands of training variations. It is a synthetic data pipeline, but one grounded in the visual texture of the actual world.

That framing connects this announcement to a broader trend in physical artificial intelligence research. NVIDIA's Cosmos platform is built around a nearly identical premise: physical AI systems need simulated environments to fill gaps that real-world data cannot cover cost-effectively, as NVIDIA's blog has outlined. DeepMind's approach differs in that it anchors simulations to photorealistic human-scale imagery rather than physics-engine primitives.

What practitioners need to watch

Generative world models compress and reconstruct the world -- they do not store it. Simulations derived from Street View will inherit whatever gaps exist in the underlying dataset: dense urban coverage versus sparse rural areas, seasonal and temporal inconsistencies, and the fixed vantage point of a car-mounted camera. How well the model generalizes beyond its training distribution is a question that engineers relying on these environments for safety-critical work will need to stress-test carefully.

DeepMind's announcement also arrives amid growing government scrutiny of advanced artificial intelligence. Google, Microsoft, and xAI have all agreed to voluntary pre-release evaluations through the US Department of Commerce's Center for AI Standards and Innovation, as BBC News reported earlier this month. Whether generative world models fall within the scope of those assessments remains unclear.

The central question for machine learning engineers is when a simulated environment becomes a reliable enough proxy for reality that policies trained inside it can be trusted. Street View offers unprecedented geographic breadth. Whether that breadth translates into the right kind of coverage for safety-critical robotics is still unresolved, and the next wave of developer evaluations will be the first real test.

Frequently asked questions

Q: What is Project Genie?
A: Project Genie is Google DeepMind's system for generating interactive AI environments. Genie 3, the current version, can now draw on Google Street View data to create navigable simulations of real-world locations.

Q: Who can access the Street View integration right now?
A: The feature is rolling out to Google AI Ultra subscribers in the United States, with a broader global expansion expected over the coming weeks.

Q: How is Waymo using Genie 3?
A: Waymo uses Genie 3 to generate simulated training scenarios for its self-driving systems, focusing on rare or dangerous edge cases that are impractical to capture through physical testing alone.

Q: What are the main limitations of world-model simulations derived from Street View?
A: Generative models reconstruct rather than store the world, so their outputs reflect the biases of the source dataset -- including uneven geographic coverage, limited viewpoints, and temporal gaps in the imagery.