Interview
Dmitri Dolgov: Waymo and the Future of Self-Driving Cars | Lex Fridman Podcast #147
- Current Operational Status: Waymo has achieved the first at-scale deployment of fully autonomous (Level 4/5) vehicles with no safety driver or human in the driver's seat, offering commercial "Waymo One" services in Phoenix, Arizona.
- Public Launch Date: The fully driverless, rider-only public service officially launched on October 8, marking a shift from a controlled early rider program to an open commercial product.
- Market Demand: Post-launch demand exceeds current fleet capacity, indicating high user interest and adoption potential.
- Fleet Size & Testing History: Waymo has conducted testing across more than 25 cities, including San Francisco, Austin, and Michigan (for snow conditions), prior to full-scale commercialization.
- Vehicle Specifics: The service utilizes the Jaguar I-PACE equipped with Waymo's proprietary fifth-generation self-driving hardware stack.
- Hardware Architecture: The sensor suite consists of 29 cameras, 5 LiDAR units, and 6 radars, all custom-designed or heavily integrated to maximize in-house control over sensing and compute.
- Compute Capabilities: The onboard computer features significant redundancy and massive compute power designed to process raw sensor data in real-time for perception, prediction, and decision-making.
- Fifth Generation Upgrade: The new hardware platform (Gen 5) offers qualitative improvements in reliability, simplicity of architecture, and manufacturability compared to previous generations, designed specifically for massive scaling.
- Strategic Pivot: In 2013, the team pivoted from developing L3 driver-assist systems to fully driverless vehicles, recognizing the fundamental engineering differences between the two approaches.
- DARPA Origins: Dmitry Dolgov joined Stanford's team for the 2007 DARPA Urban Challenge; his team placed second, with a specific "Victory Lap" bug caused by the car completing an extra lap around the oval track, and a debate with CMU regarding the interpretation of parking rules in the competition.
- Project Milestones (2009-2010): The initial Google self-driving car project set aggressive milestones to drive 100,000 autonomous miles and complete 10 difficult 100-mile routes with zero interventions, achieving this in under two years.
- Data & Learning Infrastructure: Waymo leverages Alphabet's infrastructure for massive-scale data storage, simulation, and machine learning training, allowing the fleet to share real-world data to update maps and improve behavior prediction.
- Vehicle-to-Infrastructure (V2X) Strategy: Vehicles currently operate independently, but the system uploads real-time data about dynamic obstacles (accidents, construction) to the fleet to update priors; full vehicle-to-vehicle interaction is viewed as a future advantage, not a current requirement.
- Teleoperation Stance: Waymo explicitly rejects remote teleoperation for safety reasons; instead, it uses a "Life Help" feature where riders connect with human support for guidance, while fleet assistance is limited to non-latency-critical situational confirmation.
- Sensing Philosophy: Waymo defends the necessity of LiDAR, arguing that a multi-modal approach (LiDAR, Radar, Cameras) provides a safer, more capable system than vision-only approaches, noting that cost and aesthetics are not prohibitive for scaled deployment.
- Machine Learning Integration: ML is integrated throughout the stack, from object detection to high-level behavior prediction, utilizing a hybrid approach that combines deep learning with model-based and rule-based systems to inject semantic bias (e.g., traffic light rules).
- Pedestrian Safety: The system prioritizes vulnerable road users, using sensor fusion to detect erratic behaviors (e.g., falling cyclists) and executing rapid combined steering and braking maneuvers to avoid collisions.
- Ethical Framework: Waymo rejects the "trolley problem" as a primary engineering driver, focusing instead on defensive driving strategies to avoid such scenarios entirely rather than making binary life-and-death algorithmic choices.
- Expansion Strategy: The company is currently converting its Phoenix operations into a "copy-paste" template, focusing on scaling three axes: core technology (Gen 5 hardware/software), evaluation/deployment processes, and product/commercial excellence.
- Trucking Division: Waymo Via handles autonomous trucking and goods delivery; while the application differs, the core technology, sensors, and ML stacks are shared with the passenger vehicle division.
- User Experience Optimization: The system prioritizes "professional limo driver" behavior—smooth, assertive, efficient, and predictable—avoiding overly cautious or aggressive human driving habits to build user trust and delight.
- Feedback Loops: User feedback is gathered in real-time via in-car screens, post-trip ratings, and "Life Help" interactions, which directly inform improvements in pickup/drop-off location logic and map integration.
- Future Books: Dmitry Dolgov cites The Master and Margarita by Bulgakov, the works of the Strugatsky brothers, and Orwell's 1984 as impactful readings that influenced his perspective on culture, ethics, and societal fragility.
- Meaning of Life: Dolgov views the meaning of life as evolving stages: new experiences in youth, the joy of learning and discovery in adulthood, and the desire to contribute back to society through impact and technology.