Waymo, Google's self-driving unit, on Feb. 6 (local time) unveiled the Waymo World Model, a frontier generative model that creates large-scale, highly realistic autonomous driving simulations.

Waymo said on its official blog that the world model is a key component in its AI ecosystem for validating safety. Based on Google DeepMind's universal world model Genie 3, which generates realistic interactive 3D environments, it was tuned to match the complexity and demanding conditions of real-world driving environments.

A key feature is that it can use Genie 3's extensive world knowledge to simulate rare events that are difficult to collect at scale in the real world, such as extreme weather like tornadoes or encounters with wild animals.

Engineers can control simulations using simple language prompts, driving inputs and scene layouts. It also generates high-fidelity multi-sensor outputs, including camera and lidar data.

Waymo explained that while existing simulations were limited because they relied only on real-road driving data, the Genie 3-based approach can explore situations the vehicle has not directly observed. It also adds post-training that converts world knowledge learned from 2D video into 3D lidar outputs suited to Waymo hardware, reproducing a broad range from everyday driving to long-tail scenarios.

Waymo said it expects this will allow precise validation of robotaxis' response capabilities in extreme weather and natural disaster situations.