- Waymo now provides over 500,000 fully autonomous rides weekly across 10 US cities.
- The latest 'Gen 6' system features custom vehicles with simpler, more capable, and significantly lower-cost hardware.
- Waymo's AI leverages foundation models, specialized 'teachers,' and augmented intermediate representations for superhuman safety.
Dmitri Dolgov, co-CEO of Waymo, offers a rare glimpse into the two-decade journey that transformed Google's ambitious self-driving car project into a commercial reality, highlighting critical AI advancements and strategic shifts.
Waymo's path to widespread autonomy has been a testament to sustained vision and iterative technical evolution, not a series of dead ends. Dmitri Dolgov, who joined the project in 2009, emphasizes that the core technology is now mature enough to support full autonomy, shifting the company's focus from scientific research to accelerated global scaling and deployment. This transition was marked by a 'big bet on AI' with the fifth generation of the Waymo driver, which moved from disparate AI subsystems to a unified AI backbone, enabling much greater generalizability across diverse operating domains, not just specific cities.
The Waymo driver operates using a sophisticated fusion of three primary sensor modalities: cameras, lidars, and radars, each offering complementary physical properties and 360-degree coverage. This data feeds into an AI system that processes the world, makes driving decisions, and actuates the vehicle. While real-time inference occurs locally, non-critical tasks like detecting a mess in the car can leverage off-board cloud models. Dolgov also clarified the nuanced debate around 'end-to-end' systems, explaining that while a purely end-to-end approach (pixels to trajectories) can get a system driving nominally, achieving superhuman safety and scale requires augmenting learned representations with structured intermediate representations, crucial for robust simulation, safety validation, and defining reward functions for reinforcement learning.
The company's upcoming sixth generation system represents a significant leap in deployment efficiency. It pairs a custom-designed passenger-centric vehicle, dubbed the 'O Hype,' with a new, simpler, and drastically lower-cost sensor hardware stack. This hardware, which leverages advancements in camera, radar, and lidar technology, is designed to be comparable in cost to advanced driver-assist systems. Dolgov highlights that while the hardware and sensor configurations evolve, the underlying Waymo driver software demonstrates remarkable generalizability, adapting to different vehicle platforms and environmental conditions with specialization and validation work. This generalizability is key to Waymo's imminent international expansion into cities like London and Tokyo.
Looking ahead, Dolgov envisions a future where Waymo's technology fundamentally reshapes urban landscapes and daily life. With over 3,000 cars performing half a million rides weekly and accumulating over 4 million autonomous miles, the company is rapidly expanding, recently launching in four new US cities in a single day. He firmly believes that full autonomy and driver-assist systems are fundamentally different problems, not points on a continuum, and that the long-term impact will include smoother traffic flow, reduced need for parking infrastructure, and eventually, personally owned autonomous vehicles. This ambitious vision, he credits, is sustained by Google's enduring culture of big vision, stamina, and conviction.
“I would say that we've clearly moved past the stage of scientific research and kind of deep core technology development to this new phase of accelerated global scaling and deployment.”
- Dmitri Dolgov, co-CEO of Waymo
