Waymo World Model

Waymo's Genie-3-based generative driving simulator, able to generate camera and lidar data together.

The Waymo World Model is a generative world model that Waymo announced on February 6, 2026, for large-scale autonomous-driving simulation. It is built on top of Genie 3, Google DeepMind's interactive world model, leveraging Genie 3's pretraining on huge amounts of video to generate photorealistic, interactive 3D driving environments, and outputting both camera and LiDAR sensor data at once. It mainly targets autonomous driving's long-tail problem: dangerous situations like extreme weather, natural disasters, or rare objects are seldom encountered during real-world test driving, yet must still be tested for in advance. It supports three kinds of control: changing the self-driving car's own actions to run 'what if we had driven differently' counterfactuals; controlling road layout and traffic conditions; and using language to change the time of day or weather, or even generate an entirely synthetic scene. It illustrates how general-purpose world models are starting to be used as both a test environment and a data source for embodied systems.

ExampleA single sentence can turn a real, sunny daytime test-drive scene into a rainy night, and the self-driving car can be made to change lanes differently, to observe how the autonomous-driving system reacts on the same stretch of road.

Also called

The Waymo World Model (built on Genie 3)

Related

Genie 3 · World Model · Autonomous Driving · Interactive World Model · Long-tail Problem · LiDAR

Sources

The Waymo World Model: A New Frontier for Autonomous Driving Simulation (Waymo Blog)

As of

2026-02