From Models to Mobility: Building Waymo with Dmitri Dolgov

Full Title

From Models to Mobility: Building Waymo with Dmitri Dolgov

Summary

This episode features Waymo co-CEO Dmitri Dolgov discussing the evolution of Waymo's autonomous driving technology, from deep research to global scaling.

Dolgov highlights the core technological advancements, the importance of simulation and AI models, and the challenges and future of scaling autonomous mobility.

Key Points

• Waymo has transitioned from research to a phase of accelerated global scaling and deployment, now completing nearly half a million autonomous rides weekly.
• The Waymo Driver system uses a modular approach with three primary sensors (cameras, LiDAR, radar) and specialized AI models for perception, decision-making, and actuation.
• Real-time inference is performed locally on the vehicle, with cloud processing reserved for non-critical tasks like car cleaning and status checks.
• The development of the Waymo Driver relies on a foundational, off-board AI model that is then specialized into onboard "teachers" (driver, simulator, critic) and distilled into smaller, efficient models.
• Technological breakthroughs, particularly in AI and compute power, have been critical, but the journey involved iterative learning rather than a single "right path."
• The debate around end-to-end vs. modular systems and cameras-only vs. multi-sensor approaches is nuanced, with Waymo employing a sophisticated modular system augmented by AI.
• While an end-to-end approach can achieve basic driving, achieving superhuman safety and handling edge cases requires a more complex system with intermediate representations and specialized models like simulators and critics.
• The core technology of the Waymo Driver is considered generalizable across different environments, but specialization and validation are needed for new markets like London and Tokyo.
• Waymo's sixth-generation vehicle and sensor stack represent a significant leap in simplification, cost reduction, and capability, designed from the ground up for passenger experience.
• The company's long-term vision and investment in technical talent, coupled with a culture of not accepting the status quo, have been crucial to its development.
• The future of autonomous driving is seen as a convergence of ride-hailing services and personal ownership, with significant implications for urban planning, such as the reduced need for parking.

Conclusion

The transition from research to scaled deployment signifies a critical milestone for Waymo, with the core technology now robust enough for global expansion.

The continuous evolution of AI, especially foundational models and simulation, is key to solving complex driving scenarios and achieving higher levels of safety and efficiency.

Future developments will likely involve highly integrated, cost-effective hardware and a further redefinition of urban spaces as autonomous vehicles reduce the need for parking.

Discussion Topics

• How will the widespread adoption of autonomous vehicles reshape urban landscapes and public transportation?
• What are the biggest ethical considerations for companies developing and deploying fully autonomous driving technology?
• Beyond safety and efficiency, what novel passenger experiences can autonomous vehicles enable in the future?

Key Terms

LiDAR

Light Detection and Ranging; a remote sensing method that uses light in the form of a pulsed laser to measure variable distances to the Earth or other objects.

Radar

Radio Detection and Ranging; a system that uses radio waves to determine the range, angle, or velocity of objects.

AI

Artificial Intelligence; the simulation of human intelligence processes by machines, especially computer systems.

Foundation Model

A large-scale machine learning model trained on a broad dataset that can be adapted to a wide range of downstream tasks.

End-to-end

An approach in machine learning where a single model takes raw input and produces the desired output, without intermediate steps.

Reinforcement Learning (RL)

A type of machine learning where an agent learns to make decisions by performing actions in an environment to maximize a reward.

VLM

Vision-Language Model; an AI model that can process and understand both visual and textual information.

Emergent Behavior

Complex behaviors or capabilities that arise from the interaction of simpler components in a system, which were not explicitly programmed.

Timeline