Person
1 episodes · 26 quoted claims
Dolgov's father, a physicist, took a one-year visiting position at Kyoto University as the Soviet Union was falling apart, and the family moved there.
“My dad is a physicist, so the Soviet Union started falling apart, and then he had a visiting position in Kyoto University for a year.”
Dolgov returned to Russia after high school and earned his bachelor's and master's degrees there.
“So I went back to Russia, and I got my bachelor's and master's there.”
Waymo's vehicles use three main sensing types, cameras, lidar and radar, all with 360-degree coverage, plus microphones.
“We use three different sensing modalities. There's cameras, there's lighters or lasers, and there are radars.”
Waymo uses an off-board cloud model to detect items left behind or messes in a car after a ride, which is not a real-time driving task.
“So that we do by asking a model that actually lives off-board as opposed to having to put it on the car, right?”
Waymo specializes a large foundation model into three off-board teacher models: the Waymo driver, a simulator and a critic.
“Then we specialize it into what we call three main off-board teachers. There are still large, high-capacity off-board models. There's the Waymo driver, there is the simulator, and then there's the critic.”
Waymo distills each large teacher model into a smaller student model for the driver, the simulator and the critic.
“Then you specialize in fine-tune, still off-board model. Those are the teachers, and then you distill.”
Waymo published the EMMA paper, in which a fine-tuned vision-language model generates driving trajectories and drives well in normal conditions.
“we published a paper called EMMA that did exactly that. And it will actually, in the nominal case, drive pretty darn well”
Dolgov says imitation learning alone is not enough for Waymo and that reinforcement-learning-based fine-tuning in closed-loop simulation is needed.
“Mind that that's not enough. You have to do something in closed loop.”
Waymo augments its learned encoder-decoder representations with structured representations such as objects, roads and speed limits.
“So this is where augmenting that learned representation, those learned embeddings from the encoder decoder with that more structured representation is what we do.”
Dolgov says freeway driving is mostly uneventful but has rare, severe events, because speeds are higher and consequences grow with speed.
“where the consequences of a bad event are much more severe speed is much higher everything is quadratic in speed”
Waymo began offering a fully autonomous commercial service for the first time in Chandler, Arizona, in 2020.
“our initial deployment where we started offering a fully autonomous commercial service for the first time was in 2020 in Chandler, Arizona.”
Waymo's 2020 Chandler service ran on the fourth-generation Waymo Driver, which used Pacifica minivans.
“And that was on what we called the fourth generation of the Waymo Driver. This was, if you remember, the Pacifica minivans with different hardware, different software.”
Dolgov says the fourth-generation Waymo Driver used many small AI and machine-learning models, whereas the fifth generation made AI the backbone.
“There was a lot more, you know, little AI models and ML models in the fourth generation.”
Dolgov says Waymo's sixth-generation sensor hardware is simpler, more capable and much cheaper than the fifth generation.
“It is very different from the fifth generation. It is simpler, it is more capable, it is much lower cost”
Dolgov says Waymo's new custom vehicle has a footprint barely larger than the Jaguar I-PACE, despite a much roomier interior.
“But the actual footprint of that is barely, barely, barely larger than the I-PACE.”
Dolgov says radar has lower resolution than lidar but holds up better in fog, snow and heavy rain.
“Radar has much lower resolution, but because of the physics of it, it degrades much better in adverse weather conditions.”
Dolgov says cameras degrade in darkness, sun glare or oncoming headlights, while lidar is unaffected.
“then camera will degrade. There's still some noisy signal, but it will degrade. And LiDAR is completely unaffected.”
In one San Francisco case, Waymo's peripheral lidars picked up reflections of a pedestrian's feet under a bus, so the car detected the person before they emerged.
“our peripheral lighters bounce under the bus and there was just a little bit of very, very noisy reflection of the movement of the person's feet”
Dolgov said Waymo was doing about half a million rides and over 4 million fully autonomous miles per week, with about 3,000 cars.
“We're doing about half a million rides per week. That translates to about over 4 million fully autonomous miles.”
Waymo operated fully autonomously in 11 US cities, with public riders in 10 of them; Nashville had just started.
“We are operating in a fully autonomous mode in 11 cities in the US. In 10 of those, we have riders, public riders.”
Dolgov said it took about eight years from Waymo's first rider-only operation to having external riders in four cities, and it launched four cities in a single day.
“That's about eight years. And then the other week, we just launched four in one day.”
Dolgov says Waymo's depots are automatically orchestrated, with cars returning when low on energy or needing cleaning.
“nowadays, if you look at one of our depots, it's like a fully automatically orchestrated dance of autonomous vehicles.”
Dolgov said Waymo did not yet have automated charging and that a person plugged in a cable.
“We don't yet have automated charging. In the future, you can imagine that being fully automated. But a person will come in and plug in a cable”
Dolgov said privately owned cars sit unused about 90 percent of the time, and that widespread autonomy could free land used for parking.
“Well, because, again, your car is just sitting there 90 % of the time, right?”
Dolgov cited the arrival of ImageNet in 2013, followed by transformers and vision-language models, as waves of AI progress that reshaped self-driving.
“we've seen many waves of technology breakthroughs in 2013 ImageNet came around”
Dolgov cited the engineering rule of thumb that each additional nine of reliability takes ten times more effort.
“and there's the standard engineering rule of thumb that every next nine takes 10x more.”