Conference Presentation, Interview, Fireside Chat, Keynote
Waymo's Dmitri Dolgov: 20 Million Rides and the Road to Full Autonomy
- Founder & Background: Dmitry Dolgov, Waymo co-founder, has led autonomous vehicle (AV) development for 21 years, tracing roots to the 2005 DARPA Grand Challenge and a Stanford Automotive Lab project starting in 2009.
- Early Motivation: A PhD in AI and experience in Moscow's physics program instilled an ability to learn independently, while the 2004/2005 DARPA challenges provided the "light switch moment" that defined his career focus on a tangible, high-impact product.
- Foundational Goals (2009–2011): The initial Google self-driving car project set two specific milestones: driving 100,000 total autonomous miles and completing 10 distinct 100-mile Bay Area routes without human intervention.
- Team & Execution: A core team of ~12 people achieved these early goals in 18 months, operating 24/7 to build hardware, calibrate sensors, and develop core driving algorithms.
- Hype Cycle Navigation: Dolgov observed that AV cycles are driven by breakthroughs (e.g., convolutional nets, transformers) that accelerate early progress but often obscure the difficulty of achieving the "long tail" of full autonomy.
- Mission-Driven Persistence: The team persisted through industry slumps by anchoring on the critical mission statistic: one road fatality occurs globally every 26 seconds, necessitating a refusal of "quick wins" or silver bullets.
- Waymo Foundation Model: The core AI ecosystem utilizes a multimodal, world-action language model acting as a "Waymo Foundation Model" which powers three distinct pillars: the Driver, the Simulator, and the Critic.
- Model Capabilities: This model integrates physics/dynamics reasoning, 3D spatial understanding of agents (pedestrians, cyclists), and language alignment to leverage general world knowledge for semantic driving tasks.
- End-to-End Architecture Strategy: Waymo employs a structured, end-to-end model augmented with materialized intermediate representations, rejecting "vanilla" end-to-end systems as insufficient for superhuman safety at scale.
- Key Architectural Advantages: The structured approach enables runtime agent validation, closed-loop training/evaluation, and ratio reward functions for reinforcement learning, which are critical for commercial deployment.
- Sixth Generation Hardware: The new "Waymo Driver" (6th gen) focuses on performance, drastic cost reduction, and volume production, powering the new "Waymo Origin" vehicle (based on the Chrysler Pacifica).
- Vehicle Design: The Waymo Origin features a "living room" interior, automatic sliding doors, and rear screens, currently serving employees before a general public rollout later this year.
- Exponential Scaling Metrics:
- It took 16 years to reach 100 million full autonomous miles.
- Only ~6 months were required to reach 200 million miles.
- Over 20 million full autonomous rides have been delivered to date, with 10 million occurring in the last seven months.
- Operational Expansion: The service now operates in 11 cities; four new cities were launched simultaneously earlier this year, with a global strategy to expand internationally to London and Tokyo.
- Safety Performance:
- The Waymo driver is 13 times safer than human drivers regarding serious injury-causing collisions in operational cities.
- The fleet currently drives >4 million full autonomous miles per week.
- At current scale, the system prevents a serious injury every eight days.
- Sensor Capabilities: The LIDAR system demonstrated the ability to detect pedestrians hidden behind a bus by analyzing sparse signal returns from footsteps moving underneath the vehicle.
- Future Business Model (5–10 Year Outlook): Waymo has shifted from sequential de-risking to "rapid parallel global commercialization," prioritizing widespread deployment across the US and international markets over the next decade.
- Personal Integration: Dolgov and his family use Waymo as their primary mode of transportation; his children, raised with the technology, express a preference for autonomous vehicles over human-driven cars.