Waymo Expands Autonomous Testing With AI
Waymo has introduced a generative world model that simulates rare and high-risk driving scenarios beyond real-world fleet experience.
Topics
Waymo has unveiled the Waymo World Model, a generative simulation system built on Google DeepMind’s Genie 3.
The model is designed to simulate complex and unusual traffic situations that are difficult to capture at scale through physical road testing. Examples include extreme weather conditions, unexpected obstacles and atypical traffic patterns.
Waymo said the system differs from traditional simulation models that rely primarily on proprietary driving data. Instead, it leverages Genie 3’s broader world knowledge, developed through pre-training on large and diverse video datasets.
Also Read: Salesforce Announces Genie
Through post-training, the model converts two-dimensional video understanding into three-dimensional lidar outputs aligned with Waymo’s sensor hardware. It generates both camera and lidar data, allowing multimodal simulations that reflect how the Waymo Driver interprets its environment.
The company stated that its autonomous system has logged nearly 200 million fully autonomous miles on public roads. However, billions of additional miles are accumulated in simulated environments before vehicles encounter real-world deployment scenarios.
Waymo said simulation forms one of three core pillars of its safety strategy. The company argues that modelling rare events strengthens system preparedness for complex driving conditions. The announcement did not provide benchmark comparisons or independent validation data.
The Waymo World Model allows engineers to control simulations in multiple ways. Driving action control enables testing of alternative outcomes in specific situations. Scene layout control adjusts road design, traffic signals and surrounding vehicle behaviour. Text prompts can generate variations in lighting, weather and entirely synthetic scenes.
Also Read: Disney’s Genie Park Planner Will Let Guests Personalise Experience
The system can also convert standard dashcam or mobile phone footage into multimodal simulations showing how the Waymo Driver would perceive a scene through its sensors.
For longer-duration scenarios, such as navigating narrow lanes, Waymo has developed a lighter version of the model designed to reduce computational requirements while supporting large-scale simulation.
The development reflects broader industry use of generative AI models to expand training environments for safety-critical autonomous systems.
