Waymo is training its robotaxis on streets that only exist inside a computer, and it is doing it with help from DeepMind's Genie 3. The company says it has built a new "Waymo World Model" on top of Genie 3 that can spin up photorealistic, virtual versions of city streets so its vehicles can practice rare, dangerous and unusual situations long before they show up in real traffic. The idea is to speed up expansion while toughening its software against long‑tail edge cases that are hard to capture in normal driving. For San Francisco riders, Waymo says that means robotaxis that are better prepared for fog, steep grades and chaotic intersections.
In a blog post today, the company described the World Model as "built upon Genie 3" and said it can generate temporally consistent camera images and lidar point clouds that feed into its test systems, according to Waymo. Engineers can then run counterfactual "what‑if" experiments, tweaking weather, time of day or even road layout with simple language prompts. Waymo says these multi‑sensor simulations let the Waymo Driver learn from a blend of real drives and synthetic events without putting anyone at risk.
How Genie 3 Powers Hyper‑Realistic Driving Simulations
Genie 3 is DeepMind's world model that can spin up minutes‑long, interactive 3D environments from text prompts and keep objects consistent as a user moves through a scene, which makes it attractive for training embodied systems, according to Google’s Project Genie. Waymo says it has post‑trained Genie 3 on driving so the outputs line up with its own sensor suite and the kinds of rare situations its fleet might actually encounter on the road.
What It Means For San Francisco
Waymo runs a sizable depot in Bayview and keeps its corporate headquarters in Mountain View, a setup that provides local driving logs the company can convert into targeted simulations for stress testing, city reporting shows. The Toland Street facility serves as a hub for the Bay Area fleet, and Waymo's materials highlight city‑specific scenes such as the Golden Gate Bridge dusted with light snow to showcase how the system can remix familiar locations, per the SF Chronicle. Waymo argues that combining those local logs with synthetic variations will shorten the learning curve when its Driver software rolls into unfamiliar neighborhoods.
Limits And Safety Questions
Reporters and researchers following Genie 3 point out that generative world models are not perfect. They can hallucinate small details, still struggle with clearly rendered text inside scenes and do not yet create geographically exact twins of real‑world streets. That means any lessons drawn from simulation need to be checked against on‑road data before they are used for safety‑critical decisions, a reminder that simulation is a powerful aid but not a full stand‑in for real‑world testing, as Ars Technica notes.
What's Next
Waymo says the World Model's scale could help it move faster into new robotaxi markets, a push that is part of a broader growth strategy that includes major fundraising and a target of roughly 20 expansion cities, according to Bloomberg. Riders in San Francisco may not see obvious changes right away, but the company is betting that more focused stress testing in synthetic worlds will translate into fewer surprises as its fleet encounters new roads and unfamiliar weather, Bloomberg reports.
