Genie 3 Google Deepmind

New Feature: Waymo unveiled a generative simulation model built on Google DeepMind’s Genie 3 to create hyper-realistic driving scenarios. - Capabilities: The system generates scenarios from elephant encounters to tornadoes that Waymo’s autonomous fleet has never encountered on real roads. - Technical Innovation: Unlike traditional methods limited to company data, the model leverages Genie 3’s pre-trained world knowledge for broader scenario coverage. - Testing Scale: Waymo has logged nearly 200 million fully autonomous miles while accumulating billions of miles in virtual simulation.

Waymo unveiled a generative simulation model Thursday that uses Google DeepMind’s Genie 3 to create hyper-realistic driving scenarios. The system generates scenarios from elephant encounters to tornadoes that its autonomous fleet has not yet encountered on real roads. Called the Waymo World Model, this system represents a major expansion of how the company tests its autonomous vehicles before they operate on public streets.

“Genie 3’s strong world knowledge, acquired through its pre-training on an extremely large and diverse set of videos, allows us to explore situations never directly observed by our fleet” Waymo (via Waymo Blog) Breaking from Industry Standards This approach marks a substantial departure from conventional autonomous vehicle testing methodologies. Most simulation models in the industry train only on a company’s own driving data. This limitation restricts their ability to prepare for rare or unusual scenarios. Waymo sees simulation as one of the three core pillars of its safety approach.

The other two pillars are real-world testing and rigorous validation. The Waymo Driver has logged nearly 200 million fully autonomous miles so far. However, the company racks up billions of miles in virtual worlds before facing scenarios on public roads. We’re excited to introduce the Waymo World Model—a frontier generative mode for large-scale, hyper-realistic autonomous driving simulation built on @GoogleDeepMind’s Genie 3.

— Waymo (@Waymo) February 6, 2026 By simulating the “impossible”, we proactively prepare the Waymo Driver for some of the most rare and… pic.twitter.com/Pl80OMDqLC By tapping into Genie 3’s pre-trained world knowledge rather than relying solely on proprietary driving data, Waymo creates a competitive advantage. Traditional simulation approaches cannot easily replicate this capability. The billions of virtual miles accumulated before real-world deployment illustrate how the company treats simulation. It is not merely a supplementary testing tool but a fundamental pillar that enables safe scaling.

Technical Innovation The competitive edge stems from Genie 3’s unique architecture. Genie 3 is Google DeepMind’s latest general world model. Waymo World Model is built on this frontier generative system for large-scale, hyper-realistic autonomous driving simulation. Genie 3 was pre-trained on an extensive and diverse video dataset. This training gave it broad world knowledge that extends far beyond typical autonomous driving data. Through specialized post-training, 2D video knowledge gets translated into 3D lidar outputs tailored to Waymo’s proprietary hardware. The system generates both camera and lidar data.

Cameras provide visual details while lidar delivers precise depth information. Traditional simulation methods have clear limitations. Unlike 3D Gaussian splats which fall apart visually when routes deviate, the generative model keeps things realistic and consistent. This represents a major advancement over its predecessor. The new system offers improved consistency and visual fidelity.

“Unlike reconstructive methods like 3D Gaussian splats, which fall apart visually when routes deviate, the generative model keeps things realistic and consistent” Waymo (Waymo Blog) While competitors like NVIDIA offer world foundation models through Cosmos, Waymo’s specialized post-training creates hardware-specific fidelity. This tailored approach gives Waymo advantages that generic platforms struggle to match. Control and Capabilities Building on this technical foundation, the Waymo World Model offers three control methods. These methods give engineers substantial flexibility in designing test scenarios. Driving action control lets engineers test counterfactual scenarios.

This capability allows them to explore how the Waymo Driver might respond to different situations. Scene layout control adjusts road layouts, traffic light conditions, and how other road users behave. Meanwhile, text prompts can generate different times of day, weather conditions, or entirely synthetic scenes. The model can also convert ordinary dashcam or cell phone videos into multimodal simulations. This conversion allows Waymo to learn from real-world incidents captured by others. The system transforms these incidents into training scenarios.

As a result, every dashcam incident on the internet becomes a potential training scenario. This capability multiplies the diversity of edge cases the system can encounter. Waymo vehicles do not need to experience these situations firsthand. Expanding the Testing Frontier These capabilities open new possibilities for safety validation. They extend testing far beyond traditional limits. Waymo is leveraging DeepMind’s general world model and adapting it for autonomous driving simulation. This integration enables the exploration of rare and even impossible driving conditions.

Such conditions would be unsafe or impractical to test in the real world. Waymo built a leaner version of the model. This version achieves a major reduction in compute for large-scale simulations. This means running millions of scenario variations becomes practical and cost-effective. By tapping into Genie 3’s pre-trained world knowledge, Waymo can now prepare its autonomous fleet for situations it has not yet directly observed. This advancement closes a key gap in simulation-based safety validation. The integration of this advanced technology positions Waymo to accelerate its testing capabilities.

At the same time, the company maintains all the rigorous safety standards required for full autonomous vehicle deployment on public roads worldwide.

People Also Asked

• Genie3—GoogleDeepMind
• GoogleDeepMind’sGenie3Builds Interactive Worlds from a Text...
• Waymo TapsGoogleDeepMindGenie3for Self-Driving Simulation
• Genie3AI - AI Video Generator | Create Stunning Videos from Text
• Google News -GoogleDeepMindunveils new AI world model,Genie...
• Genie3: Create Interactive 3D Worlds from Text
• Genie3:GoogleDeepMindreinvents interactive AI
• Genie3creates real-time AI worlds from simple text prompts