A self-driving car has never seen an elephant wander into its lane. It has probably never rolled across snow piled high on the Golden Gate Bridge, or met a tornado on a suburban cul-de-sac. But according to Waymo, its cars are now learning to handle all three — and thousands of other vanishingly rare scenarios — inside a hyper-realistic virtual world built on Google DeepMind's Genie 3.

The Alphabet-owned robotaxi company unveiled the Waymo World Model in a blog post on 6 February, describing it as "a frontier generative model that sets a new bar for large-scale, hyper-realistic autonomous driving simulation." The technology was spotlighted again this week by Ars Technica, as Waymo pushes into tougher new markets including Boston and Washington, D.C.

At its heart, the announcement rests on a deceptively simple idea: if you want to teach a car to cope with the unexpected, you need a way to show it the unexpected.

What is a "world model"?

In plain English, a world model is an AI system that has learned how the physical world tends to look and behave — gravity, shadows, how cars move, how pedestrians cross — and can generate fresh, playable video of scenes it has never been given directly. You type a prompt, feed it a driving input, and it produces footage that feels like looking out of a car's windscreen.

Genie 3, which Google DeepMind revealed last year, is the company's most advanced general-purpose world model. Unlike older systems that forgot details the moment you looked away, Genie 3 can "remember" objects in a scene for several minutes — a crucial trick for realistic driving simulation.

Billions of virtual miles

Waymo says its vehicles have now driven nearly 200 million fully autonomous miles on real roads, but billions more in simulation. The Waymo World Model is the new engine for that simulated mileage.

Crucially, Waymo and DeepMind retrained the base model to output not just 2D camera video but also 3D lidar data — the laser-based depth-sensing that Waymo's robotaxis rely on, and which rivals such as Tesla have long shunned. Engineers can change the time of day, the weather, the traffic, or drop in objects ranging from a Texas longhorn to a pedestrian in a T-rex costume.

They can also rerun real drives with different decisions — a "what if" rewrite that, Waymo claims, holds up visually where older reconstruction techniques break down.

Why it matters for safety

Autonomous driving has always had a "long tail" problem. Most miles are mundane; the dangerous moments are rare, scattered and hard to collect. A system that has never seen flooding, dense fog, or a malfunctioning truck facing the wrong way may handle them badly when it finally does.

Simulating those edge cases in volume, Waymo argues, creates a more rigorous safety benchmark before the car ever meets them on a public road.

The catch, as Ars Technica notes, is that the benefit "will depend on how accurately Genie 3 can simulate the real world." Published Genie clips range from convincing to uncanny, and Waymo has not yet released independent benchmark results showing the simulated training translates into measurably safer driving.

Rough road ahead

Waymo's record is not spotless. WIRED reports the company has hit "a rough patch" in Washington, D.C., where its service has drawn complaints and regulatory scrutiny. The district is preparing a major robotaxi study this summer, according to Axios. Federal safety probes have continued into 2026 even as Waymo expanded to ten US cities on the back of a $16bn funding round.

A better simulator is not a substitute for a clean real-world record. But if the Waymo World Model lives up to its billing, it could quietly become one of the most important pieces of infrastructure in the race to make driverless cars something cities can live with.