Lex Fridman PodcastDmitri Dolgov: Waymo and the Future of Self-Driving Cars | Lex Fridman Podcast #147
CHAPTERS
- 0:00 – 4:55
Waymo’s mission and Dmitri’s early fascination with computers
Lex sets the stage for Waymo’s impact and Dmitri’s long involvement since the Google Self-Driving Car Project. Dmitri recounts getting an early PC, moving from playing games to programming, and even building a small game that was nearly sold commercially.
- •Waymo’s place in fully driverless deployment and why it matters
- •Early home computing in the late 1980s and learning by tinkering
- •From playing games to writing software and a Pascal game
- •A young lesson in business models (freeware vs selling outright)
- 4:55 – 9:55
Games, creativity, and childhood dreams (traffic control cop)
The conversation drifts into what makes games magical—simplicity, imagination, and world-building. Dmitri shares an unexpectedly grounded childhood dream: becoming a traffic control officer, which amusingly foreshadows his future in transportation automation.
- •Simple graphics vs imagination: why certain games feel timeless
- •Minecraft and creativity as a recurring theme
- •Dmitri’s childhood dream: directing traffic in Moscow
- •How early interests can later map onto real careers
- 9:55 – 13:00
From abstract CS to robotics: education path and MIPT (Phystech)
Dmitri explains that robotics came late—self-driving cars were his first real robotics work. He and Lex discuss Moscow Institute of Physics and Technology (MIPT/Phystech), its intensity, and Dmitri’s unusual decision to return to Russia for university after finishing high school in the US.
- •Robotics arrived post-grad; early focus was math/physics/CS theory
- •Grad school at Michigan, postdoc at Stanford as the inflection point
- •MIPT culture: fundamentals-first, rigorous training
- •Choosing to return to Russia for six years at Phystech
- 13:00 – 22:58
DARPA Urban Challenge: rules, team dynamics, and ‘victory lap’ bug
Dmitri recounts joining Stanford’s DARPA Urban Challenge effort, working on motion planning, and the competition’s structure (dynamic vehicles, no pedestrians). He shares vivid moments from nighttime testing, the anxiety of race day, and memorable bugs—including the famous ‘victory lap’ behavior.
- •Urban Challenge setup: abandoned airbase ‘city,’ human-driven and robotic cars
- •Stanford’s role and key contributors (e.g., Mike Montemerlo)
- •Free-space planning in parking lots and interpretation of rules
- •‘Victory lap’ bug and what it taught about planning/waypoints
- 22:58 – 30:05
Google Self-Driving Car Project beginnings: early milestones and what they proved
The discussion turns to how DARPA momentum and Larry Page/Sergey Brin’s interest led to the Google self-driving project in 2009. Dmitri describes the small founding team and the learning-driven milestones—100,000 autonomous miles and ten 100-mile ‘spicy’ routes with zero interventions.
- •2009: a small team formed to understand the real problem space
- •Milestones designed to maximize learning, not product quality
- •Ten 100-mile routes across diverse conditions (freeways, mountains, dense cities)
- •Prototype success vs the harder path to a reliable consumer product
- 30:05 – 38:50
Pivot to full autonomy and the road to Waymo (2013–2020 milestones)
Dmitri outlines the key strategic shift: moving from driver-assist ideas to fully driverless vehicles around 2013. He traces major milestones, including the first driverless public-road ride in 2015 (Firefly), Waymo’s founding in 2016/17, and ongoing driverless operations leading to Waymo One.
- •2010–2013: exploring driver-assist before choosing full autonomy
- •2015: Firefly ride with a blind passenger—no steering wheel/pedals
- •2016/17: Waymo created as a dedicated Alphabet company
- •2017 onward: regular driverless operations and scaling confidence
- 38:50 – 45:22
Waymo’s self-driving hardware stack: sensors, redundancy, and on-car compute
Lex presses into the engineering details of Waymo’s fifth-generation platform. Dmitri describes their approach to hardware-software co-design, in-house lidar development, sensor diversity (cameras, lidars, radars), and the need for powerful, redundant onboard computing—while staying cautious on IP details.
- •Separating vehicle platform from the self-driving ‘driver’ hardware suite
- •Fifth-generation hardware as a qualitative jump over prior versions
- •Sensor mix and scale: 29 cameras, 5 lidars, 6 radars
- •Onboard compute requirements: real-time processing and redundancy
- 45:22 – 50:25
Fleet data, edge cases, simulation, and ‘connected cars’ without dependency
They discuss the less-visible backbone: massive data pipelines, mining edge cases, and simulation/evaluation infrastructure at scale. Dmitri explains how fleet connectivity helps propagate learned information (construction, accidents) while emphasizing that each car must still operate safely without connectivity.
- •Offboard infrastructure: data mining, training, evaluation, and simulation
- •Finding long-tail edge cases and building custom internal frameworks
- •Real-time fleet updates as map priors (construction zones, incidents)
- •Connectivity as an advantage, not a requirement for safe driving
- 50:25 – 53:23
Human support without teleoperation: Live Help and fleet assistance
Lex asks about the human role behind a driverless service. Dmitri draws a clear line: Waymo does not teleoperate cars, but provides rider support (Live Help) and a form of fleet assistance where humans can confirm context in rare off-nominal scenarios—while safety-critical driving stays onboard.
- •No teleoperation: safety/latency-critical decisions remain onboard
- •Live Help improves rider confidence and first-trip usability
- •Fleet assistance can provide confirmation/context in unusual scenes
- •Designing interfaces and guidance for rider-only experiences
- 53:23 – 57:44
Waymo One in Phoenix: public rider-only service, demand, and product learnings
Dmitri describes opening fully driverless rider-only trips to the public in Phoenix and the strong early demand. They discuss service area flexibility, common rider use cases, and why real paid rides are irreplaceable for learning and improving the product.
- •Public access to fully driverless rides and managing capacity constraints
- •Service area model: flexible A-to-B within a defined geography
- •Real-world rider behavior: errands, schools, bars, families with car seats
- •Why paid operations accelerate learning compared to controlled pilots
- 57:44 – 1:11:49
User feedback loops and UX details: pickup/dropoff, predictability, and delight
The conversation focuses on how Waymo gathers rich feedback—during rides, after rides, through support, and via UX research studies. Dmitri shares how seemingly small issues (e.g., pickup points across parking lots) require careful mapping and product design, and how predictable computer-to-computer coordination can outperform human rideshare.
- •In-ride feedback UI, post-ride ratings, and qualitative comments
- •UX research: longitudinal studies to track changes over time
- •Pickup/dropoff is deceptively hard; mapping and walking directions matter
- •Predictability as a product advantage vs human-driven rideshare variability
- 1:11:49 – 1:18:33
Driving style: assertiveness without rule-breaking, and what ‘good driving’ means
Lex asks whether autonomous cars need to ‘bend rules’ like humans. Dmitri argues for a professional-driver ideal: safe, smooth, predictable, and still efficient—shifting the tradeoff curve rather than choosing between timid and aggressive behaviors.
- •Following rules while remaining efficient and assertive
- •Professional limo driver analogy: smooth precision over aggression
- •Avoiding false tradeoffs by improving core capability (‘move the curve up’)
- •Comfort and predictability as part of safety and user trust
- 1:18:33 – 1:24:11
Waymo Via and trucking: what transfers and what changes
They turn to autonomous trucking as a second major product line (moving goods). Dmitri explains that most of the hardest problems—perception, prediction, planning, ML infrastructure, simulation—carry over, while hardware placement and domain specifics differ (e.g., sensor geometry, freeway emphasis).
- •Waymo One vs Waymo Via: moving people and moving goods
- •Core autonomy stack transfers across vehicle types and domains
- •Truck sensor configuration differs to manage geometry and blind spots
- •Freeway domain shifts priorities (e.g., range) but not fundamentals
- 1:24:11 – 1:36:05
Scaling beyond Phoenix: three axes—core tech, evaluation/deployment, operations
Lex asks how Waymo expands to new cities; Dmitri frames scaling as a deliberate, staged process. He describes three dimensions: improving core hardware/software (including Gen 5), building robust evaluation and release processes, and achieving operational/product excellence learned from end-to-end service in Phoenix.
- •Phoenix as the intentional ‘full-stack’ learning platform
- •Three scaling axes: core driver tech, evaluation/deployment, operations/product
- •Gen 5 designed for manufacturability, reliability, and unit economics
- •Release confidence: simulation, validation, and efficient deployment pipelines
- 1:36:05 – 1:41:34
Policy partnerships and the lidar debate: why multiple modalities matter
Dmitri emphasizes the importance of proactive engagement with governments at local, state, and federal levels. He then responds to the ‘lidar is a crutch’ critique: lidar, radar, and cameras complement one another physically, and Waymo’s in-house lidar advances reduce both cost and integration concerns.
- •Deployment requires sustained relationships with regulators and officials
- •Sensor diversity: passive cameras plus active lidar/radar for robustness
- •Cost and aesthetics objections to lidar are not seen as fundamental blockers
- •Manufacturing scale and technology improvements reduce lidar cost dramatically
- 1:41:34 – 1:54:11
Machine learning across the stack: fusion, modularity, and structured prediction
They dig deep into ML philosophy: end-to-end vs modular designs, early sensor fusion, and injecting inductive bias where appropriate. Dmitri describes ML’s role from detection to scene understanding and behavior prediction, and notes parallels between driving trajectories and sequence modeling breakthroughs like transformers.
- •ML is pervasive: perception, semantic understanding, prediction, planning, simulation
- •Early fusion across modalities and time improves capability
- •Rejecting ‘raw sensors to steering torque’ as too brittle; favor hybrid structure
- •Transformers/sequence models as promising tools for behavior and interaction modeling
- 1:54:11 – 2:01:00
Pedestrians, cyclists, and safety in the long tail (real incident example)
Dmitri discusses protecting vulnerable road users as a top priority, highlighting sensing advantages (e.g., lidar at night) and the need for low-latency reactions. He shares a specific Phoenix incident where a cyclist fell into the lane, illustrating how perception, tracking, and emergency maneuvers must come together in milliseconds.
- •Vulnerable road users drive stringent safety requirements and validation
- •Night and occlusion scenarios where active sensors add crucial margin
- •Long-tail events require fast perception updates and strong braking/steering
- •Real-world example: cyclist falls into roadway; system behavior validated via simulation
- 2:01:00 – 2:23:09
Trolley problem as a distraction, then books and the meaning of life
Lex raises the trolley problem; Dmitri argues real safety comes from defensive driving and avoiding such forced dilemmas through better capability, not encoding stylized moral choices. They close with Dmitri’s book recommendations (Bulgakov, Strugatsky brothers, Orwell) and a reflective discussion on how purpose evolves through life.
- •Ethics in AVs as emergent from risk modeling and defensive design
- •Focus on reducing probability of catastrophic dilemmas rather than ‘choosing’ victims
- •Book picks: The Master and Margarita; Strugatsky sci-fi; Orwell’s 1984
- •Meaning of life as additive phases: experience, fun, learning, impact, family
