Waymo Trains Self-Driving Cars in AI-Generated Virtual Worlds

Before navigating the streets of major American cities, Waymo’s autonomous vehicles have already accumulated extraordinary experience, not only on the road but—more importantly—in entirely AI-generated virtual worlds.

On February 6, 2026, Waymo announced the intensive training regimen its autonomous vehicles undergo, revealing that its robotaxis spend more time training in virtual worlds generated by Google’s Genie 3 than driving in the real world.

Waymo Uses Virtual Worlds to Hone its Autonomous Vehicles

Recent days have seen striking images circulating online of digital worlds brought to life by user imagination—versions of Grand Theft Auto VI, for example, or a pack of cigarettes strolling through the New York City subway. Behind these creations lies Google DeepMind’s Project Genie, a research prototype capable of creating, modifying, and exploring interactive virtual universes from simple prompts.

While access to Genie 3 remains limited, some companies are already leveraging it to train their artificial intelligence systems. Waymo is among them, utilizing these virtual worlds to prepare its autonomous vehicles before real-world deployment.

Waymo reports that its vehicles have already traveled billions of miles in these simulations—far exceeding the approximately 200 million miles accumulated on open roads.

How Waymo Trains its Vehicles with Genie 3

The goal isn’t to create fantastical, unrealistic worlds. Waymo aims to closely mirror reality, faithfully recreating real-world driving conditions and routes. To achieve this, the company relies on real-world images and data collected from cameras, lidar, and mapping systems to reconstruct similar environments within its simulators. These scenes then serve as the basis for replaying the same route hundreds, or even thousands, of times, with slight variations to test different scenarios directly based on the real world.

Engineers can modify parameters at will: changing the route, disrupting traffic, varying the behavior of other drivers, or even adding an obstacle like an elephant. The objective is to observe how the vehicle reacts to unique situations, some impossible to test in real-world conditions. This allows Waymo to replay a scenario thousands of times with minor adjustments, identifying weaknesses in its system and refining its algorithms.

Genie 3 also enables vehicles to encounter varied weather conditions, such as rain, fog, or night. In scenarios shared by Waymo, one simulation is simply titled “Fleeing a Devastating Fire,” depicting a Waymo vehicle evacuating from a blaze.

As Waymo approaches its European rollout, these tools likely played a significant role in preparing for the fleet’s arrival in London later this year. Genie 3 has likely allowed Waymo to recreate European environments, with potentially narrower roads, different signage, and, crucially, driving on the right side of the road.

Waymo’s approach to autonomous driving prioritizes demonstrably safe AI, building safety into the core of its AI ecosystem. This holistic approach, centered around the Waymo Foundation Model, powers a unified AI system that drives continuous learning and improvement. The Driver, Simulator, and Critic components are all fueled by the same underlying AI, creating a virtuous cycle of development. The Simulator, powered by tools like Genie 3, is critical to this process, allowing for rigorous testing and refinement in a controlled environment.

The use of virtual worlds isn’t simply about accumulating mileage. It’s about creating a robust and reliable system capable of handling the unpredictable nature of real-world driving. By exposing its vehicles to a vast range of scenarios—both common and edge cases—Waymo aims to ensure its autonomous technology is prepared for anything the road might throw its way. The company achieved a more than ten-fold reduction in crashes with serious injuries compared to human drivers, and continues to improve safety through this method.