Waymo chief product officer on progress, competition vs Cruise

Waymo, letter’s self-driving car division has had a relatively good few months — at least compared to one of its main competitors: General MotorsCruise.

Waymo, formerly Google’s self-driving car project, is now an independent subsidiary of Google parent company Alphabet and has been operating in some form since 2009. Five years ago, the company launched what it calls “the world’s first commercial self-driving car” as a ride-hailing service in the Phoenix metropolitan area, and last year expanded to San Francisco. The company is soon planning a commercial rollout in a fourth city, Austin, and recently began test driving vehicles during the winter in Buffalo, New York.

For much of this period, Cruise seemed to be in tight competition: When Waymo raised money in 2020 at a $30 billion valuation, Cruise did the same in 2021 at the same valuation. When Cruise begins offering fully self-driving service in San Francisco in the winter of 2022, Waymo will follow in the fall. In August, California regulators voted to allow two companies to provide robo-taxi services around the clock in San Francisco, making it the first major U.S. city to allow two robo-taxi companies to compete for “round-the-clock” service.

Now, after a series of safety issues and accidents involving Cruise’s self-driving cars in recent months, the situation looks very different. Cruise has suspended all public road operations — both supervisory and manual operations, laid off contractors and recalled nearly 1,000 robotaxis after pedestrian collisions. In October, the California Department of Motor Vehicles suspended Cruise’s self-driving vehicle deployment and testing licenses, effective immediately; last week, General Motors announced that it would significantly reduce Cruise’s spending in 2024.

After the news broke, Waymo Chief Product Officer Saswat Panigrahi told CNBC that the attitude of regulators has not changed, and neither has the public’s view of the company.

Clearly, Waymo appears to be performing better than some of its competitors. What exactly do you think you do differently?

There are no shortcuts. I mean, it’s not a question you ask an app or a web page and they give you the answer. This is a vehicle that weighs thousands of pounds and is driving in the real world—yes, it’s an application of artificial intelligence, but it’s a very different application of artificial intelligence. Regarding time, experience and rigor, no matter how hard you try, it takes time to do it.

So what I’m saying is, the amount of data you test on your own – you can always test more, but the scale of testing is staggering – I sometimes say building Waymo drivers is a difficult thing, but it’s almost the same with evaluating drivers difficulty. The amount of simulation we had to do…has taken ten years. It requires Google’s level of infrastructure because even to do a simulation of this scale, as you and I are talking about right now, the 25,000 cars in our simulator are learning how to drive better. In order to do that, you need incredible infrastructure capabilities, because even if you have AI capabilities, it’s very difficult to make that skill work without the infrastructure — and before AI became cool, it was A decade of investment in artificial intelligence.

The computing infrastructure that powers these simulations?

Yeah, some of it is just the raw scale of computing, how many computers you can carry, that kind of thing. But some of it is also – think about how realistic old-school video games are compared to today’s video games, it’s a metaphor for the way things are. Suppose we see a person speeding 60 mph on a 45 mph (street) in Phoenix, and then imagine we see a very narrow intersection in San Francisco – can you realistically compare the two? Combined to challenge your driver after millions of miles could something more serious happen in the real world?

(Among other things), being able to increase rainfall, for example – well, when you’re driving on a nice day through this narrow intersection with a speeding officer, you’re safe enough. Can you do this on a rainy day? Can it be done at night? When you want to push a system this way, you can’t wait for real-life rain to happen, but being able to simulate rain requires infrastructure, but also enough algorithms and realism to push that.

Can you specify how much calculation is required?

Before Waymo, I worked on quite large-scale systems at Google and Ericsson, and this is a pretty staggering scale. But the only number I can tell you is over 25,000 virtual vehicles driving continuously, 24/7, learning from each other, and driving (tens of) billions of miles in simulation. Think about how many kilometers you or I drive in a year – in any given year we drive, what, 10,000 miles…? Now consider billions of miles of experience—a difference of nearly seven orders of magnitude.

Let’s talk about how traffic has changed over the past month. Do you see the increase? reduce?

Things are growing – to give you an idea, we’ve had over 10 (public passenger trips) this year…Phoenix and San Francisco are both experiencing increased ridership. We’re well ahead of 10,000 trips per week (in every city)… so things are going great. We’re taking time to respond to feedback and expand thoughtfully.

(Note: Waymo recently revealed that Waymo passengers will take more than 700,000 rides in self-driving cars in 2023.)

How has all the controversy in recent months affected public perception of your plans?

It was a very positive response from the riders. We look at their ratings, we look at their usage patterns, we look at what they tell us qualitatively, we talk to them in focus groups, and they’re all very positive…

For the people we share our city with—communities, groups like first responders, firefighters, etc.—we are constantly in contact with them. We are listening to their feedback. In SF alone, we’ve trained more than 5,000 first responders, held multiple training sessions, and built on that with new capabilities. For example, now we can signal intensity to firefighters: “Hey, we’re going to turn around and leave this scene.”

During the same period, have regulatory agencies’ requirements for the Waymo team changed?

We had a very open dialogue with the regulators and submitted more data than they requested… So we had a very positive engagement with them, but the tone has not changed.

Three years ago, we were the first company to publicly release a security framework, which was our mechanism for testing the effectiveness of our systems and how we determined when they were ready for deployment. We were also the first to publish all crash data from a fully autonomous service… before any regulatory agency asked us to do so. Yes, we submit ongoing reports to them as well.

In terms of your AI pipeline and how it works, are you running deep learning on neural networks? Import training data from simulation? Give me a synopsis.

There is a lot of artificial intelligence that can help us detect pedestrians, children, cyclists, pedestrians on scooters, pedestrians on e-scooters (that’s why it goes faster), elderly people pushing strollers. Being able to predict in which direction a car that creates an unusual curvature will jump… being able to predict where different objects will appear in the next few seconds.

All of this is a lot of artificial intelligence, and a lot of it is dedicated to looking at the differences between the way children behave, the way adults behave, the way cyclists behave… you name it from deep learning, reinforcement learning, all these areas, We use it in several parts of the system.

Most self-driving cars have remote operations teams. How does Waymo work?

I want to clarify that the driving is done by the Waymo Driver in the car – there is no remote person driving the car. In a way, you can think of it like air traffic control. Air traffic control does not fly the aircraft, but pilots may ask air traffic control questions, “Hey, I’m observing a very unusual situation here, what’s the intent?” When the situation isn’t immediately clear on the scene, there are a few things to ask. Very basic binary question so people can respond to provide clarification.

For example, you might have a set of cones blocking a street, but there might be a gap big enough for you to get in, so it’s a bit ambiguous whether you should get in or stop – that’s the kind of question that can be asked and answered… …it does the right thing even without support.

What has been Waymo’s biggest internal obstacle in the past year?

One thing I would say that’s definitely going to be interesting this year is cost reduction.

During past expansions, I got the impression that Waymo was looking for “Goldilocks cities”, by which I mean cities where launching a driverless car service would not be too difficult but would also be challenging in some way, e.g. With growing populations or fun road maneuvers, but no snow or ice. What are you looking for when you’re looking for your next city? What are the possibilities for these cities besides Phoenix?

You touched on a key thing. Phoenix has been great for us…if the roads are really tight, you don’t need to look that far, but when you’re going 45 mph, sometimes people are going 50 to 60 mph, You do need to see further, anticipate objects, make unexpected turns, and so on. One thing we realized as we traveled from Phoenix to San Francisco, the ultra-high density of narrow pedestrian streets, dual-cabs, etc., is that this is also the case in every other city with good weather in the U.S., and at least, some internationally. So, it’s just a linear combination of the two. So if you take Los Angeles as an example, West Hollywood is kind of like a denser area of ​​San Francisco, but its roads to the suburbs are a lot like Phoenix.

On the weather axis, we have rain and fog now…and then the next one will eventually be snow…what we want to make sure is that we go to a city not just to rubber stamp it, but just to be able to say we are there autonomical.