Dolgov, Dmitri

CTO of Waymo; formerly Google's self-driving car project

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

Lex Fridman Podcast · Dec 20, 2020

• Dmitri Dolgov
• Waymo history
• DARPA Urban Challenge
• driverless versus driver-assist pivot
• lidar versus camera-only debate
• sensor fusion and end-to-end learning
• Phoenix robotaxi deployment
• pedestrian safety edge cases
• trolley problem critique
• precision-recall driving behavior analogy

Core ideas

Dolgov traces his path into autonomous vehicles from a childhood dream of becoming a traffic-control police officer in 1980s Russia, through a physics/math education at Moscow's MIPT (following both his parents), to joining Stanford's team for the 2007 DARPA Urban Challenge as a postdoc — the third and most complex of DARPA's driving competitions, requiring vehicles to navigate a simulated city shared with human-driven cars rather than an empty desert. He recounts the moment he considers formative to his career: running his free-space parking-lot motion-planning code on a real car for the first time at 3 a.m. on Stanford's campus, and being unable to sleep afterward from excitement. He shares self-deprecating competition anecdotes — a rules dispute with the winning CMU team over whether cars had to reverse out of parking spots, and his own team's "Stanford victory lap" bug, where a planning flaw caused the car to loop an extra lap around a checkpoint oval because it wouldn't plan the next waypoint until physically reaching the prior one.

He narrates Waymo's institutional history: founded in 2009 as a ~12-person Google project after Larry Page and Sergey Brin were impressed by the DARPA Urban Challenge, it set two deliberately ambitious early milestones (100,000 autonomous miles, and ten distinct 100-mile routes driven start-to-finish with zero interventions) achieved in about 18 months — a "prototype and learn" phase Dolgov calls the most fun of his career. A pivotal strategic decision came around 2013, when the team abandoned pursuing driver-assist ("Level 3") systems in favor of a fully driverless vision, having concluded the two approaches require fundamentally different engineering paths. Key milestones followed: the 2015 ride for blind passenger Steve Mann in a custom driverless "Firefly" vehicle (no steering wheel or pedals) in Austin; Waymo's 2016–2017 spinout as an independent Alphabet company; October 2017's first regular driverless (empty driver's seat) public-road operations; the 2018 launch of the Waymo One commercial service in Phoenix; and October 2020's opening of fully driverless, rider-only trips to the general public in a Phoenix service area larger than San Francisco. He distinguishes this from competitors' typical "safety driver" model, emphasizing Waymo's use of genuinely empty vehicles as a qualitatively different and harder achievement.

On hardware, Dolgov describes Waymo's fifth-generation sensor suite (built largely in-house, especially lidar) as a "qualitative jump" in capability, manufacturability, and cost designed to enable scaling beyond Phoenix, while declining to discuss specific compute/GPU details for IP reasons. He pushes back directly on Elon Musk's characterization of lidar as a "crutch," arguing that cameras, lidar, and radar are complementary due to their different physical properties (passive vs. active sensing, different wavelengths), that cost and aesthetics are not fundamentally prohibitive objections, and that deliberately excluding a capable sensing modality is only a reasonable trade-off for lower-stakes driver-assist products, not for a fully driverless system. On the technical architecture question of end-to-end learning, he rejects the "purest" interpretation (one model mapping raw sensor input directly to steering/braking) in favor of a hybrid approach: heavy machine learning throughout the stack (perception, prediction, decision-making, and increasingly sensor fusion across modalities and time), balanced with explicit rule-based/physics-based components for cases with unambiguous semantics (e.g., a red traffic light means stop for the ego vehicle, even though predicting whether other drivers will actually stop remains a genuinely hard ML problem). He notes growing interest in applying transformer/language-model-style architectures to behavior prediction and planning, given structural similarities (sequential order, local versus global context) between language and driving scenes.

Dolgov describes real-time inter-vehicle data sharing (e.g., one car reporting a new construction zone or accident, propagated fleet-wide as an update to shared map "priors") as helpful but not required, since each Waymo vehicle must be capable of fully autonomous operation independent of connectivity — distinguishing this from "teleoperation" (which Waymo does not use) and clarifying that human "fleet response" assistance is limited to non-safety/non-latency-critical confirmation in rare, off-nominal scenarios, alongside separate live customer-support ("live help") functionality for riders. He discusses product design lessons drawn directly from rider feedback — in-app ratings, live support conversations, and longitudinal/UX research studies — citing a memorable failure case where a naive "nearest pickup point" heuristic sent a passenger walking around a parking lot full of cacti in 110-degree heat, illustrating why predictability and reliability (the car arriving exactly where indicated) matter more to user trust than marginal optimization. He frames Waymo's goal as building a driving style that is simultaneously safe, smooth, assertive, and predictable — rejecting the binary of "cautious versus aggressive" in favor of an analogy to precision-recall trade-offs, where true capability improvements move the entire performance curve rather than just shifting an operating point, comparing ideal driving behavior to a skilled professional (e.g., a limo driver) whose smoothness and efficiency coexist rather than trade off.

On scaling beyond Phoenix, Dolgov frames the challenge along three axes: core technology (the fifth-generation hardware/software platform designed for manufacturability and generalization), evaluation and deployment infrastructure (release processes, simulation, and validation pipelines mature enough that improvements can ship routinely without heavy manual overhead), and product/commercial/operational excellence — with Phoenix serving as an intentional end-to-end proving ground for de-risking each dimension before broader "copy-paste" expansion, which he says has not yet begun in earnest. He confirms Waymo Via (goods movement, including trucking since 2017) shares almost all of the same technical core as Waymo One (passenger transport), with hardware differences mainly involving sensor placement (e.g., dual side-mounted lidar on trucks to eliminate blind spots) rather than fundamentally different software challenges, and notes zero-occupancy vehicle movements are treated nearly identically to occupied ones at the first-order driving-behavior level. On pedestrian and cyclist safety, he describes it as one of Waymo's highest priorities, citing lidar's equal night/day performance as a key safety advantage and recounting a real Phoenix incident where a cyclist fell directly into the vehicle's path at 45 mph — both the human test driver and, in later simulation, the autonomous system independently executed the same correct combined steering-and-braking response. He dismisses the classic "trolley problem" as largely a philosophical distraction from actual engineering practice, arguing real systems focus on probabilistic collision-risk reasoning and, more importantly, on defensive driving that avoids ever entering such forced-choice scenarios in the first place, rather than on solving discrete ethical dilemmas at decision time.

He closes with book recommendations — Bulgakov's The Master and Margarita (which he has only read in Russian and is uncertain translates well), the Strugatsky brothers' science fiction (Roadside Picnic, Hard to Be a God, Monday Begins on Saturday), and Orwell's 1984 (paired with Animal Farm, which he connects to 2020-era concerns about "doublethink" and the fragility of open societies) — and a reflection on meaning, referencing Asimov's short story The Last Question before offering a personal, evolving account: meaning shifts across life stages, from early sensory novelty and fun, to the joy of learning, to contributing back to society, to the perspective shift of having children (he has three) — framed as ideally additive rather than a replacement of what came before.

Excerpts