Waymo published the ten architecture rules behind 200 million driverless miles

Waymo published a post on August 26 setting out ten positions on how its driving system is built, drawing on the more than 200 million fully autonomous miles the company has driven. It was written by Srikanth Thirumalai.

Several of the positions were architectural commitments rather than lessons. Waymo said multimodal sensing — cameras, lidar and radar together — was indispensable. It described high-definition maps as a prior that an automated mapping system keeps current, and said it had consolidated a modular perception stack into fewer, larger foundation models, then distilled them through a teacher-student process to fit the compute available in the vehicle.

On validation, Waymo said an independent onboard layer checks every trajectory the driving system proposes against physical constraints and traffic law, and acts as a backstop when one fails. The company called that layer non-negotiable for safely scaling at Level 4, and set it against pure end-to-end neural architectures, which it described as a black box.

Waymo also detailed an automated reviewer it calls the Waymo Critic, which it said grades millions of on-road miles each week and tens of billions of simulated miles for safety, traffic-law compliance and ride quality. Simulation, it said, is run closed-loop rather than as replay of recorded data. Under what Waymo calls its Foundation Model, deliberative reasoning is handled by vision-language models trained on Gemini, while sensor fusion handles real-time control.

The tenth position addressed the path from driver assistance to autonomy. Waymo said improving a Level 2 system until it becomes Level 4 was “a false summit”.