Dmitri Dolgov is Waymo’s co-CEO and an early engineer on Google’s self-driving car project. He describes full autonomy as a long-horizon engineering and deployment challenge: advances in AI, complementary sensors, simulation and safety evaluation must work together to support a driverless ride-hailing service. — Cheeky Pint — Dmitri Dolgov Interview.

Visual summary of operating lessons from Dmitri Dolgov.

Part 1: The Foundations of Autonomy

  1. On Early Inspiration: Dolgov traced his entry into autonomous driving to Stanford’s participation in the 2007 DARPA Urban Challenge, which he called a pivotal moment for the field and for his own career. — a16z — Building the World’s Most Trusted Driver.
  2. On Foundational Education: After moving to the United States, Dolgov returned to Russia in 1994 for undergraduate and master’s study in physics and applied mathematics, which he described as a strong foundation before pursuing graduate work in the United States. — Cheeky Pint — Dmitri Dolgov Interview.
  3. On Early Exploration: Dolgov described the early Google self-driving team as a small research effort that used ambitious driving milestones to learn what full autonomy would require. — Y Combinator — Seven Lessons from Building Physical AI.
  4. On Initial Benchmarks: The early Google project set itself two tests: 100,000 autonomous miles and ten 100-mile routes without intervention. Dolgov recalled that meeting both took roughly 18 months. — Y Combinator — Seven Lessons from Building Physical AI.
  5. On Corporate Conviction: Dolgov credited Larry Page, Sergey Brin and Alphabet’s leadership with the vision and stamina to support the long development path of fully autonomous driving. — Cheeky Pint — Dmitri Dolgov Interview.
  6. On Patience in Development: Dolgov argued that successive AI breakthroughs make early demonstrations easier, but the long tail of safety-critical cases still requires iterative work; there is no single technical shortcut to full autonomy. — Cheeky Pint — Dmitri Dolgov Interview.
  7. On Committing to Full Autonomy: Waymo describes a focus on systems that complete a trip without a human driver, rather than handing responsibility back to one; this makes long-tail safety cases central to its roadmap. — Waymo — Experienced Urban Driver.

Part 2: The AI Architecture

  1. On the AI Backbone: Dolgov said the fifth-generation Waymo Driver made a larger bet on AI as the software backbone than the fourth generation, which had more smaller ML subsystems. — Cheeky Pint — Dmitri Dolgov Interview.
  2. On Foundation Models: Dolgov describes an off-board foundation model of the physical driving world that feeds specialized Driver, Simulator and Critic teachers; smaller models are distilled for faster inference. — Cheeky Pint — Dmitri Dolgov Interview.
  3. On Teacher and Student Models: The Waymo stack specializes a shared large foundation model into Driver, Simulator and Critic teachers, then distills task-specific students for deployment or evaluation. — Cheeky Pint — Dmitri Dolgov Interview.
  4. On Real-Time Inference: Dolgov said driving decisions are processed on the vehicle rather than requiring a real-time cloud connection; non-driving tasks, such as checking for forgotten items after a ride, can run off-board. — Cheeky Pint — Dmitri Dolgov Interview.
  5. On the Limits of Simple End-to-End Driving: Dolgov said a basic pixels-in, trajectories-out model can drive well in ordinary cases yet remains far below the safety bar for fully autonomous operation across rare edge cases. — Cheeky Pint — Dmitri Dolgov Interview.
  6. On Intermediate Representations: Waymo augments learned model representations with structured concepts such as objects, roads, signs and speed limits, enabling richer simulation, training and safety validation. — Cheeky Pint — Dmitri Dolgov Interview.
  7. On Runtime Validation: Dolgov says structured intermediate representations give Waymo additional real-time safety-validation layers; they do not amount to a guarantee that every downstream decision is verified or safe. — Cheeky Pint — Dmitri Dolgov Interview.
  8. On Early Deep-Learning Gains: In a 2018 recap, Dolgov said Waymo’s work with Google Brain applied deep networks to pedestrian detection and reduced the system’s error rate by 100-fold within months. — Waymo — Google I/O AI Recap.

Part 3: Sensor Fusion and Hardware

  1. On Complementary Sensors: Dolgov describes cameras, lidar and radar as complementary sensing modalities whose data are jointly encoded to form a stronger view of the driving environment. — Cheeky Pint — Dmitri Dolgov Interview.
  2. On Radar in Low Visibility: Radar can retain useful vehicle returns in dense fog where cameras degrade; Dolgov notes that each sensor has different noise characteristics and Waymo fuses their data rather than switching between them. — Cheeky Pint — Dmitri Dolgov Interview.
  3. On Sensor-Fusion Reasoning: In Dolgov’s Waymo essay, a camera’s stop-sign detection can be checked against lidar data to distinguish a real sign from a reflection in a storefront. — Waymo — Experienced Urban Driver.
  4. On Filtering Weather Noise: Dolgov wrote that rain and snow add noise to autonomous-vehicle sensor data and that Waymo uses machine learning to filter it while identifying road users. — Waymo — Google I/O AI Recap.
  5. On the Firefly Milestone: A Waymo Firefly prototype completed a fully driverless public-road ride in Austin with Steve Mahan on October 20, 2015, an important early milestone rather than proof that general autonomy was solved. — Waymo — Firefly Prototype Journey.
  6. On Passenger-Centered Hardware: Dolgov described Waymo’s sixth-generation custom vehicle as a shift from designing around a human driver toward designing the cabin around passengers, with more interior space. — Cheeky Pint — Dmitri Dolgov Interview.
  7. On Long-Range Perception: Dolgov wrote that dense, long-range sensor data can help the Waymo Driver detect a truck door beginning to open before a person emerges. — Waymo — Experienced Urban Driver.
  8. On Reducing Hardware Cost: Dolgov described sixth-generation hardware as an effort to simplify the sensor stack and lower costs while preserving the capabilities needed for broader deployment. — Cheeky Pint — Dmitri Dolgov Interview.

Part 4: Real-World Testing and Simulation

  1. On the Demo-to-Product Gap: Dolgov cautions that an autonomous-driving demonstration can arrive quickly, while making a driverless service reliable across the long tail of rare, safety-critical events takes much longer. — Y Combinator — Seven Lessons from Building Physical AI.
  2. On Training Infrastructure: In a 2018 recap, Dolgov said Waymo used TensorFlow and Google’s TPU-equipped data centers to train neural networks up to 15 times more efficiently. — Waymo — Google I/O AI Recap.
  3. On Simulation Scale: Dolgov reported in 2018 that Waymo was testing its ML models in simulation at an equivalent of 25,000 cars driving throughout each day. — Waymo — Google I/O AI Recap.
  4. On Learning from Disengagements: Dolgov wrote that Waymo used safety-related disengagements as learning events, generating hundreds or thousands of variations of each event in simulation to test fixes. — Waymo — Accelerating the Pace of Learning.
  5. On Local Driving Behavior: Dolgov described how Waymo’s accumulated city driving experience helps its system adapt to local patterns, including slower driving on steep San Francisco slopes. — Waymo — Experienced Urban Driver.
  6. On Intersection Nuance: Waymo’s planning system adapts its staging position for a turn using prior experience of how drivers navigate that particular intersection. — Waymo — Experienced Urban Driver.
  7. On Cautious Driving Near Occlusions: Dolgov wrote that Waymo may slow near an occluded crosswalk or a low-visibility hillcrest, behaviors intended to make its driving more predictable and comfortable. — Waymo — Experienced Urban Driver.
  8. On Scaling Autonomous Miles: In a 2026 talk, Dolgov said Waymo needed roughly 15 years for its first 100 million fully autonomous miles but about seven months for the next 100 million. — Y Combinator — Seven Lessons from Building Physical AI.
  9. On Entering New Markets: Dolgov says the Waymo Driver now generalizes across operating domains more effectively, but each new market still requires data collection, specialization and rigorous local validation. — Cheeky Pint — Dmitri Dolgov Interview.

Part 5: Strategy, Business, and the Future

  1. On Driver Assistance Versus Autonomy: Dolgov treats driver assistance and fully driverless service as qualitatively different problems: the latter must handle the trip and the safety-critical long tail without a human fallback. — Cheeky Pint — Dmitri Dolgov Interview.
  2. On Sequential De-risking: Dolgov describes Waymo’s progression through successive vehicle and driver generations as deliberately expanding the operating domain only after learning how to build, evaluate and deploy the prior one. — Cheeky Pint — Dmitri Dolgov Interview.
  3. On Focusing on Ride-Hailing: In 2023 Waymo’s co-CEOs said the company would focus near-term effort and investment on ride-hailing, postponing trucking operations and most of that business unit’s technical work. — Waymo — Doubling Down on Waymo One.
  4. On Trucking Spillovers: Waymo said its continued work on the core Driver, especially freeway capabilities, could later benefit autonomous trucking even as the trucking business timeline was pushed back. — Waymo — Doubling Down on Waymo One.
  5. On Cities and Parking: Dolgov suggested that, if fully autonomous vehicles become more common, cities might need less land devoted to parking because shared vehicles spend less time idle. — Cheeky Pint — Dmitri Dolgov Interview.
  6. On the Road-Safety Mission: Dolgov frames the high global toll of road crashes as a reason to pursue safer driving systems, while acknowledging the need to earn trust through measurable, real-world safety evidence. — Y Combinator — Seven Lessons from Building Physical AI.
  7. On Generalizable Deployment: Dolgov wrote that experience from Waymo’s Arizona driverless service fed its safety-readiness process and carried into other cities, although local environments still differ. — Waymo — Experienced Urban Driver.

Part 6: Experience, Trust, and Public Deployment

  1. On Rider Comfort: Waymo’s city-driving system learns small behaviors—such as where to stage a turn and when to slow near an occluded crosswalk—that Dolgov says add up to a more natural and predictable ride. — Waymo — Experienced Urban Driver.
  2. On Driving as Social Interaction: Dolgov compares driving with a multi-agent social interaction in which history, context and other road users’ likely responses matter, not merely the geometry of the next maneuver. — Cheeky Pint — Dmitri Dolgov Interview.
  3. On Accessibility: In a 2023 letter, Dolgov and co-CEO Tekedra Mawakana told the story of rider Sharon Giovinazzo, who said driverless rides gave her more independence after ride-hail denials involving her guide dog. — Waymo — 2023 Community Letter.
  4. On Community Learning: Waymo’s co-CEOs wrote that operating in a city requires learning from riders and local organizations and demonstrating trustworthiness alongside technical deployment. — Waymo — 2023 Community Letter.
  5. On Public Safety Evidence: Dolgov argues that public safety claims should be backed by evaluation, field data and published research rather than a demonstration alone. — Y Combinator — Seven Lessons from Building Physical AI.

Part 7: AI Progress and Long-Horizon Leadership

  1. On Successive AI Waves: Dolgov says convolutional networks improved perception and that transformers later expanded Waymo’s ML work into behavior prediction, planning and simulation. — a16z — Building the World’s Most Trusted Driver.
  2. On Driving as Sequence Modeling: Dolgov likens road-user prediction and vehicle planning to sequence modeling: local continuity matters, but actions also depend on the wider social and spatial scene. — a16z — Building the World’s Most Trusted Driver.
  3. On Milestones and Maturity: Dolgov contrasts early driverless demonstrations with later public rides and faster city launches to show that capability, validation and commercial operation matured over many years. — Y Combinator — Seven Lessons from Building Physical AI.
  4. On Practical Full Autonomy: For Dolgov, the defining product question is whether the vehicle can complete a trip without human driver supervision within a specified operating domain; wider deployment still requires validation. — Waymo — Experienced Urban Driver.