#147 – Dmitri Dolgov: Waymo and the Future of Self-Driving Cars

Episode Summary

Executive Summary: Dmitri Dolgov traces Waymo’s path from the DARPA challenges to a fully driverless commercial service in Phoenix, explaining how the company evolved from proving autonomy was possible to building a scalable, safe product. The conversation covers robotics origins, hardware/software stack, machine learning, fleet data, passenger experience, trucking, regulation, and why Waymo believes fully driverless transport can be both safe and delightful.

Main Topics: Origins in computer science and robotics (Priority: 5/5): Dolgov describes his early fascination with computers, programming, and later robotics, including graduate work and Stanford’s DARPA Urban Challenge team, which first exposed him to self-driving cars. DARPA Urban Challenge and early autonomy lessons (Priority: 5/5): He explains the structure of the competition, the technical difficulties, and memorable bugs from Stanford’s vehicle, emphasizing that these early experiments proved autonomy was feasible while revealing its complexity. Waymo’s founding and evolution into a product company (Priority: 5/5): The discussion covers Google’s 2009 self-driving project, the shift from learning milestones to a driverless commercial vision, and the creation of Waymo as an independent entity within Alphabet. Waymo’s driver stack: sensors, compute, and ML (Priority: 5/5): Dolgov details Waymo’s sensor suite, custom hardware, onboard compute, and the hybrid ML approach used for perception, prediction, planning, and safety-critical decision-making. Commercial deployment, rider experience, and feedback loops (Priority: 4/5): He explains fully driverless service in Phoenix, how users summon rides, how Waymo collects feedback, and how the company refines pickup/drop-off and UX based on real-world usage. Scaling beyond Phoenix and future commercialization (Priority: 4/5): The conversation focuses on how Waymo plans to copy and paste the Phoenix model to new cities through improvements in core technology, evaluation/deployment systems, and operational excellence. Broader philosophy: safety, ethics, books, and meaning (Priority: 3/5): The interview ends with discussion of the trolley problem, safety around pedestrians, favorite Russian literature, and Dolgov’s view that meaning comes from fun, learning, impact, and family.

Key Arguments: Autonomy became real when the team demonstrated zero-intervention driving on hard routes, which gave confidence that a full driver could be built. Waymo’s success depends on combining custom sensors, onboard compute, ML, mapping, simulation, and evaluation rather than relying on a single end-to-end black box. LiDAR is not a crutch; using cameras, radar, and LiDAR together improves safety and capability, especially for fully driverless operation. Machine learning is central across the stack, but not everything should be learned end-to-end; Waymo uses hybrid systems with engineered semantics where appropriate. A fully driverless service changes the user experience fundamentally: no safety driver, no friction, more trust, and a more predictable product. Phoenix was chosen as a proving ground to learn every hard part of deployment—technical, operational, and commercial—before scaling to more cities. Passenger feedback is essential and comes through in-app ratings, live support, UX studies, and real-world usage patterns like groceries, bikes, schools, and errands. The trolley problem is philosophically interesting but not how self-driving systems are engineered; the real goal is to build a driver that avoids such dilemmas through better defensive driving. Waymo sees trucking as sharing the same core autonomy problems as ride-hailing, with different vehicle geometry and product specialization. Driving should be safe, efficient, smooth, assertive, and predictable—more like a professional limo driver than a reckless human driver.

Data Points: Google self-driving project start: 2009 - Dolgov says the Google Self-Driving Car Project began in 2009. Waymo founding: 2016 - Waymo became the independent company in 2016. DARPA Urban Challenge year: 2007 - The third DARPA competition in the series was the Urban Challenge in 2007. Team size at project start: about a dozen people - Dolgov says the early Google project started with roughly 12 people. Learning milestone: autonomous miles: 100,000 miles - One early goal was to drive 100,000 miles in autonomous mode. Learning milestone: routes: 10 routes of 100 miles each - The second milestone was 10 difficult 100-mile routes with zero intervention. Time to complete early milestones: just under 2 years - He says the milestones took about a year and a half to a little more than that. First fully driverless ride: 2015 - The first fully driverless public-road ride in a custom vehicle happened in Austin in 2015. Waymo commercial launch: 2018 - Waymo One launched in Phoenix in 2018. Early rider driverless access: 2019 - In 2019, Waymo offered fully driverless rider-only rides to early riders. External financing round: $3.2 billion - Dolgov says Waymo raised $3.2 billion in 2020. Vehicle sensor suite: 29 cameras, 5 LiDARs, 6 radars - He cites these figures for Waymo’s vehicles. Phoenix service geography: larger than San Francisco - He says the Phoenix service area is bigger than San Francisco. Initial public driverless fleet mileage: 100 miles in one day - On the first day of public driverless operations, Waymo drove 100 miles. Testing footprint: more than 25 cities - Dolgov says Waymo has tested in over 25 cities.

Pivotal Quotes: "We believe that it's doable." — Dmitri Dolgov: He summarizes the biggest lesson from the early autonomous-driving milestones. "I wouldn't characterize it exactly that way. No, I think LiDAR is very important." — Dmitri Dolgov: His response to Elon Musk’s claim that LiDAR is a crutch. "You want to build a driver that is safe, comfortable, smooth, and predictable." — Dmitri Dolgov: He describes the ideal Waymo driving style and operating philosophy.

Implications: Waymo’s story suggests fully autonomous transport is no longer a lab demo but a scalable service. The next battle is replication: better hardware, better evaluation, better operations, and trusted public adoption across cities and use cases.