How Waymo Made Self-Driving Cars a Reality | Dmitri Dolgov, Co-CEO

Waymo's bet on fundamentals made self-driving real

Waymo turned self-driving from a research question into a commercial service by building the machine that makes the product, not just a snapshot of it. The driver runs on the car while off-board teachers and a high-fidelity simulator train and validate it, and an eval-first culture spreads across the company. Reliability gains cost roughly ten times more per nine, so Waymo invested in fundamentals that let it adopt new architectures without breaking safety. It now runs in 15 cities and about half a million trips a week.

Takeaways

• Eval-first culture is Waymo's real advantage. Waymo built the infrastructure to evaluate and simulate its driver so it can adopt new architectures without breaking product momentum. Dolgov says strong fundamentals let the company take big technical swings and land disruptive change without regressions.
• Only the driver runs on the car. Off-board high-capacity teacher models are distilled into on-board students. Driving needs closed-loop training and evaluation, so Waymo relies on a high-fidelity simulator and an actor-critic split. Training never happens in the real world.
• Each nine of reliability costs ten times more. Dolgov says reliability gains follow a rough tenfold effort rule. Each additional nine of reliability takes about ten times the effort, which is why Waymo invests heavily in evaluation and validation before each release.
• Fleet behaves as one coordinated system. Waymos do not talk to each other directly. Coordination happens at fleet level over passenger pickup and prepositioning before rush hour. Cars also return to depots to charge and clean, including charging bays.
• San Francisco came first for complexity and market. Dolgov cites its massive commercial potential as one of the top cities in the US and worldwide, plus its driving complexity. Proximity to the team also mattered, and San Francisco became Waymo's first city.
• Waymo scales across cities, vehicles and sensors. The Waymo Driver generalizes across 15 US cities, the fifth and sixth generations, the JLR-based vehicle and the custom Ojai. Dolgov says Waymo does about half a million trips a week and five million fully autonomous miles.
• Pick problems that can be simplified. Dolgov tells engineers with Waymo-sized ambition but a startup budget to choose problems that can be simplified, such as warehouse robotics, and to invest in fundamentals if they aim at full autonomy.

Milestones on the way to autonomy

1. In 2009, the project starts with the question of whether to try at all, not what product to build.
2. An early challenge was to drive 100,000 autonomous miles and 10 hundred-mile Bay Area routes with no intervention.