Real Alpha
Watch mode

Minus One · 3 min · Oct 8, 2026

Perception transfers more easily than a physical agent's actions

Dmitri Dolgov

Co-CEO, Waymo

  • Aditya AgarwalHost · Partner, South Park Commons

    One of Facebook's first engineers, he helped build its first search engine; later CTO of Dropbox.

  • RS

    Ruchi SanghviHost · Partner, South Park Commons

About this segment

Dolgov says the Waymo Driver transfers across cities, vehicles and sensor setups. Perception and training infrastructure may extend to other robots, but simulators and action-generating plans need embodiment-specific work.

From “How Waymo Solved Self-Driving | Dmitri Dolgov, Co-CEO”, Oct 8, 2026. Watch the full episode

“The specifics of how do you build a simulator for a humanoid robot versus how you build it for a car, there's going to be a lot of work there.”
Dmitri Dolgovat 46:49 · Minus One46:49

Takeaways

  1. The driver already runs across multiple cities and vehicle types.
  2. Perception and training infrastructure may transfer to other agents.
  3. Simulators and action plans need work for each embodiment.
Watch 3:00 from 44:24

In this episode

Full episode
  1. 15:00Closed-loop validation requires a driver, simulator and critic4:15
  2. 22:07Waymo’s moat is its system for improving the product5:01
  3. 26:59Each additional safety nine demands more rigorous evaluation8:07
  4. 37:27A stable scorecard can show a system is nearing deployment3:37
  5. 35:06Start physical AI in settings where the problem can be simplified2:32
  6. 41:04AI will speed simple builds, while safety-critical work stays hard3:30
  7. 12:37Waymo began by testing feasibility, not building a product2:22

  8. 44:24Perception transfers more easily than a physical agent's actions3:00

Up next · Latent SpaceMessy metagenomic data can improve protein models2:50

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. 44:30

    I think you've already answered this question, but it's crushing my dreams a little bit. Does the Waymo driver ever generalize? Is there some sort of general physical agent hiding in your foundation model that can have different embodiments, not in the form of a car? Yeah, well, okay, does the Waymo driver generalize when we talk about generalizability? Are you going to build robots? or like delivery bots or like, you know, phones, let me answer it maybe at different levels.

  2. 45:04

    There is the most specific version of generalizability where we think about the Waymo driver generalizing to different cities, different countries, different vehicles, different sensor configuration. That is done. It generalizes extremely well. We're in 15 cities across the U.S. right now. We're operating the fleet on the fifth generation and the sixth generation of the driver. It has different sensors, different vehicles, the GLR base and the Ojai, and it's the same driver.

  3. 45:38

    That will continue. We are focusing on ride hailing today. There's north of 13 trillion miles being driven worldwide, It's predicted to be about 20 by the end of the decade. So why wouldn't you want to have the benefits of autonomous vehicles in all of the driving modalities, whether it's ride-hailing, whether it's personally-owned vehicles, whether it's trucking and deliveries. So that will, same product, slight modifications will generalize extremely well, something that we're working towards.

  4. 46:10

    then there's the question of will it generalize to other versions of physical AI agents and there I think my answer is a little bit more nuanced and the perceptual parts I would expect to generalize really well sensing and understanding the environment I think the fundamentals that we talked about, like training recipes, like how do you properly do closed loop eval and RL in those domains, the fundamental infrastructure that supports all of that, that will generalize pretty well.

  5. 46:56

    The specifics of how do you build a simulator for a humanoid robot versus how you build it for a car, there's going to be a lot of work there. And the actual generative part, if your physical embodiment of your agent is very different, that planning generative part, that will require some adaptation as well. I have one on behalf of my son, who's here on the first row. He's nine years old.