Google Lets Users Walk Real Streets in a Virtual World Using Street View and Genie

Google DeepMind has begun stitching Street View imagery into its Genie virtual-environment engine, allowing people to explore recognizable city blocks in an interactive simulation that can change weather, time and other conditions. Announced at the Google I/O developer conference, the feature is available to some Google AI Ultra subscribers in the U.S. starting today, with a broader rollout planned in the coming weeks.

For everyday users the change means more than dropping the little Street View figure onto a map. People will be able to create scenarios — picture a summer block in winter or a rainstorm on a familiar avenue — and move through that scene from viewpoints other than a car’s windshield. DeepMind and Google describe use cases that range from trip planning and education to imaginative play. The company also sees practical value for robotics and autonomous vehicles: training systems can replay rare or unusual conditions in a controlled, place-specific setting.

The update builds on two decades of Street View collection. Google has gathered upward of 280 billion photos from roughly 110 countries and all seven continents using vehicles equipped with cameras and individuals wearing portable camera rigs. DeepMind researchers say that wealth of imagery gives Genie a rich factual base to anchor simulated worlds, letting models reference real locations instead of inventing them from scratch. Jack Parker-Holder, a DeepMind research scientist, told reporters that blending this dataset with Genie helps machines and people test how the same street looks under different circumstances — for instance, to show a robot the rare moments when sunlight reflects off a façade so the machine won’t be surprised by the glare.

Genie itself has been in development and public preview for months. DeepMind released a research preview of Genie 3 last August and opened early access to certain subscribers in the U.S. this January. The tool already powers experimental simulators for training self-driving systems on highly unusual events, such as severe weather or other low-probability hazards, and the Street View layer could expand that work to more cities. Google staff emphasized that a key distinction between Genie and many vehicle simulators is viewpoint flexibility: traditional driving simulators recreate scenes from a car’s perspective, whereas Genie can represent how a human, a delivery robot or a drone would perceive the same block.

That said, the current results still look more like a polished game than a photo. In demonstrations shown to journalists, environments were clearly recognizable but lacked full photorealism and a true grasp of physical interactions. Agents in the simulations sometimes moved through solid objects or ignored terrain in ways that real-world actors would not. DeepMind engineers compared that limitation to earlier gaps between images and video models: the system must learn physical cause and effect through observation rather than be told explicit rules, and they estimate it may take several more months to a year to close that gap with the video-generation side of the company’s work.

DeepMind managers stressed the experimental nature of the integration. Diego Rivas, a product manager on the team, said the company plans to expand access while improving fidelity and accuracy. Jonathan Herbert, director of Google Maps who began his career on the Street View team, highlighted another technical advance: the simulation maintains spatial memory as users rotate or move, so the model preserves what was behind you and can extend the scene consistently. Outside experts see both promise and hazards. A robotics researcher who reviewed the system noted its potential to accelerate training for urban robots but warned that simulated perfection can breed overconfidence unless models learn robust physics and edge cases.

As the rollout continues, consumers, city planners and companies running autonomous fleets will be watching closely to see whether the combined dataset and simulation tools deliver useful, reliable views of the places people actually live and work.