Waymo, Alphabet's self-driving car company, recently unveiled its latest secret weapon.Waymo World Model (Waymo World Model). This system is based on the latest Google DeepMind...Genie 3 World Model TechnologyIt can generate highly realistic and interactive 3D virtual worlds, allowing Waymo's self-driving cars to undergo billions of miles of extreme testing in this AI-generated "Matrix" before actually hitting the road, without having to take risks in the real world.
Based on Genie 3: Not only can it generate videos, but it can also be "interactive".
The core of Waymo's world model originates from Genie 3, developed by Google DeepMind. Unlike OpenAI's Sora, which primarily generates videos for "viewing," Genie 3 is a "General World Model" whose generated environments are interactive.
Waymo has made targeted adjustments to enable it not only to generate visual images but also to simultaneously output LiDAR and radar sensing data, allowing self-driving cars to learn not only through image training systems but also through data recorded by multiple sensors.
Specializing in the "long tail effect": tornadoes and oncoming vehicles can be generated at will.
Waymo has accumulated nearly 2 million miles of autonomous driving mileage on real roads, but that's not enough. This is because in the real world, "long-tail scenarios," such as sudden tornadoes, pedestrians dressed as dinosaurs, or vehicles driving in the wrong direction on highways, have extremely low probability of occurring, making it difficult to collect large amounts of data.
However, with the help of the Genie 3 world model, Waymo will be able to build more realistic digital environments and tailor suitable scenarios for specific learning and training needs.
Through generative world models, Waymo engineers can act like "creators":
• Extreme climate events are generated:Blizzard, flood, and blindingly bright light.
• Create unexpected situations:Set up scenarios where large animals rush out or cargo falls from vehicles ahead.
• Counterfactual Simulation:This is the most powerful feature. The system can recreate a real-life scenario, but change the variables (e.g., "What if that car hadn't yielded?"), allowing the AI to try different decisions in a parallel universe and assess which reaction is safer.
Google's AI strategy: Project Genie leads the way.
It's worth mentioning that Google also launched a consumer-facing product at the end of January this year.Project Genie Prototype (Currently available to Google AI Ultra subscribers). This is a tool that allows ordinary users to "build a world with words." You can think of it as a "game version" of the Waymo world model, showing that Google is splitting its Genie technology into two paths: one for entertainment and creation, and the other for high-precision industrial applications.
Analysis of viewpoints
The introduction of this technology signifies that autonomous driving training has entered a new era of "generative simulation".
Past simulators were mostly manually created based on game engines (such as Unreal or Unity). Although they were beautiful, the physics rules were hardcoded, and the scenes were relatively limited. The concept of a "world model" is that AI learns the "rules of how the world works" by observing a large number of videos. For example, it knows that a car will stop when it hits a wall, glass will break, and water will splash.
This is similar to the end-to-end training that Tesla is currently doing (Tesla is also training its own world model to predict future shadows). However, Waymo's advantage lies in having more accurate sensor data (LiDAR) as ground truth.
For Waymo, this solves the bottleneck of expansion speed. Previously, entering a new city (such as from Phoenix to San Francisco) required a fleet to travel there for several months to collect data; now, theoretically, with basic map data, the world model can generate various traffic conditions for the city, allowing the cars to "get familiar" with the city in the virtual world before being deployed. This may explain why Waymo's expansion speed has increased significantly recently.
