How Waymo Solved Self-Driving | Dmitri Dolgov, Co-CEO

Summary · 3 min read

Dmitri Dolgov is co-CEO of Waymo. He describes Waymo’s 17-year journey from early experiments to fully autonomous vehicles on public roads.

Dolgov argues that Waymo’s advantage is not a model snapshot but the data, infrastructure, simulation and evaluation machinery that lets it iterate and deploy with confidence. He explains how the on-car driver works with off-board simulation and evaluation, and why safety-critical systems need closed-loop training and testing. He says teams can build impressive driving demos quickly, but reaching the reliability required for deployment at scale takes far more work. For physical-AI founders, he recommends starting with simpler applications unless they are prepared to invest in the fundamentals.

Minute by minute

1. 0:23

   First rides made autonomy tangible. Aditya Agarwal recalls his first Waymo ride as a striking experience. Dmitri Dolgov remembers a driverless car on public roads in 2015, fully autonomous operations in Chandler in 2017, using Waymo around San Francisco in 2022, and opening freeways to autonomous rides.

2. 3:29

   Waymo coordinates at fleet level. Dolgov says each vehicle drives autonomously, while the fleet coordinates pickups, prepositions cars for demand and sends others to depots to charge and clean. Vehicles share information about events such as accidents, and they coordinate around charging stations in depots.

3. 5:07

   Safety shapes the rider experience. Dolgov says the stakes include pedestrians, cyclists and other vehicles, not only passengers. He says people talk to, wave to and thank Waymos, while some test or bully them; riders also use the private cabin for work calls, family trips and music.

4. 9:23

   Vehicles and cities shape the product. Dolgov says Waymo’s new custom vehicle is designed around riders, with wide sliding doors, more space and a sixth-generation driver. He says San Francisco offered commercial potential and challenging driving conditions, while being nearby let employees use the product themselves.

5. 12:49

   Early milestones tested feasibility. When Waymo began in 2009, Dolgov says its goal was to understand the problem, not build a product. The team set milestones of 100,000 autonomous miles and 10 Bay Area routes of 100 miles each, completed with a human behind the wheel but without intervention.

6. 15:00

   The driver needs a training ecosystem. Dolgov describes an on-car driver that turns sensor input into driving decisions, supported by off-board models for the driver, simulator and critic. Only the driver runs in the car; the simulator and critic support training and evaluation. He says closed-loop evaluation is necessary when open-loop testing cannot capture the task’s complexity.

7. 20:11

   Fundamentals let Waymo adopt new models. Dolgov says Waymo moved toward high-capacity off-board models distilled into an on-car student as machine-learning methods advanced. He argues that strong data, infrastructure and evaluation let Waymo make major technical changes without disrupting a safety-critical product. A model alone, he says, cannot be validated or iterated on without that machinery.

8. 25:18

   Deployment demands far more than a demo. Dolgov says small teams can now build an impressive driving system in days or weeks, but that does not make it ready for a product; each additional nine of reliability takes roughly 10 times the effort. He identifies 2015, when Waymo began public-road operations without a human driver, as a major step that required redundancy and a data-driven readiness framework.

9. 30:08

   Evaluation must shape engineering. Dolgov says safety and evaluation must permeate engineering rather than be added afterward, much like test-driven development. He describes model-level and system-level metrics, billions of simulated miles, and multi-week release evaluations that include physical-world tests and expert review.

10. 35:06

    Start physical AI with a simpler task. Dolgov advises resource-constrained physical-AI teams to start with applications that simplify the task, such as warehouse systems with controlled environments or guardrails. Teams pursuing fully autonomous systems, he says, need to invest in world models, realistic simulators and evaluation infrastructure.

11. 37:27

    Stable metrics signal deployment readiness. Dolgov says teams can gauge progress by whether the problems stabilize and metrics improve consistently, rather than shifting every six months. A team aiming to deploy within six months or a year should already have a scorecard that would support a confident launch. He clarifies that Waymo’s early milestones were for learning, not a product schedule.

12. 41:04

    Easier physical AI work will spread. Dolgov expects today’s AI tools to help teams build simpler products much faster, including work that once took a large company a year and could take five people a month. He cautions that safety-critical domains with many interacting components and very slim margins for error are less likely to become fully commoditized.

13. 44:24

    The driver generalizes within driving. Dolgov says the Waymo Driver generalizes across 15 U.S. cities and across vehicle generations with different sensors. He sees potential in personal vehicles, trucking and deliveries, but says simulators and planning need adaptation for different physical agents. He reports about half a million trips each week and about 5 million fully autonomous miles.

Published Oct 8, 2026.