Genie 3's connection to Street View is not just a better demo. The multi agent perspective shift it enables is the capability that turns world models into robotics grade training infrastructure — and Google has twenty years of the physical world to back it up.
Waymo's existing simulators can show a robotaxi what the world looks like from the driver's seat. They cannot show what the world looks like from the pedestrian stepping off the curb, the cyclist filtering through traffic, or the delivery robot hugging the sidewalk. Google DeepMind solved that problem, and the way it solved it is the actual story.
Jack Parker-Holder, a research scientist on DeepMind's open-endedness team, described the specific value of the new integration during the Google I/O 2026 demonstration. Waymo's production simulators are locked to the vehicle's point of view. The connection to Street View — the photographic record Google has been building since 2007 — allows Genie 3 to render the same scene from any agent's perspective in the environment, producing multi-agent training scenarios that car-centric simulators structurally cannot generate. This is not a cosmetic improvement to an existing capability. It is a different kind of simulation.
Waymo is already using the system. Diego Rivas, group product manager at Google DeepMind, confirmed that Waymo has been using Genie 3 to simulate scenarios too dangerous, illegal, or logistically impossible to stage on real roads: tornado conditions, animals mid-crossing, simultaneous equipment failures. The value is not in showing the car what the road looks like. It is in showing the car what the world around it looks like — and training the full stack on how other agents perceive and react to it.
Google has been building that world for nearly two decades. Jonathan Herbert, principal product manager at Google Maps, said during the I/O demo that the Street View archive represents the largest street-level image collection in existence, spanning 280 billion photographs across 110 countries and all seven continents — a figure that traces to Google's own presentation and has not been independently verified. According to Herbert, no other AI laboratory has announced a comparable dataset.
The Street View grounding capability went live on May 19, 2026, announced at Google I/O and available first to Google AI Ultra subscribers at $200 per month — down from $250 after a coinciding price reduction on the top-tier plan. Genie 3 access remains exclusive to that tier.
Genie 3 is the third generation of a research prototype that first shipped as a research preview in August 2025. The jump from Genie 2 to Genie 3 is not incremental. Genie 2, released in late 2024, held scene coherence for roughly ten seconds before losing spatial coherence. Genie 3 sustains scene memory across several minutes of real-time navigation — the difference between a demo and a training-grade simulation environment. The model renders the path ahead as the user moves, rather than pre-computing a static environment. At 720p and twenty to twenty-four frames per second, the output is visually fluent enough to be useful for training robotics systems on real-world spatial complexity.
Google's own model page acknowledges the limit: Genie 3 cannot yet create a faithful reconstruction of any given street. It maintains spatial continuity — spinning 360 degrees inside a generated environment, the model correctly remembers what was behind the viewer rather than regenerating it from scratch — but it does not achieve perfect geographic fidelity to a specific location. Herbert described this plainly: the model knows what a street should look like, trained as it is on billions of images, but it generates a plausible world rather than a faithful replica of the one photographed.
This matters for how Waymo is actually using it. Genie 3 with Street View grounding is a scenario generator, not a digital twin. It produces training conditions — unusual weather, unpredictable agents, edge-case configurations — that a fleet can encounter in the real world, without requiring the real world to provide them on demand. That is genuinely useful for autonomous vehicle training, and it is distinct from what satellite imagery, HD maps, or synthetic data can produce alone.
Who else could match this? The honest answer is not yet clear. Meta has published research on visual world models and has enormous visual datasets. Apple has street-level data from its mapping operations. Tesla's fleet generates continuous real-world driving data at a scale no competitor can match, though under a fundamentally different architectural paradigm. Whether the twenty-year Street View head start constitutes a durable competitive moat, or whether the capability is portable to different data architectures, is a question nobody in public has answered.
For robotics developers, autonomous vehicle companies, and simulation infrastructure teams, the relevant question is not whether Genie 3 is impressive. By Google's own specs, it is. The question is whether Google's world model — grounded in a photographic record no competitor has assembled — is the right foundation to build on, and whether Google's tiered access model leaves enough room to find out. The data is the moat. The question is who gets to use it.