Welcome to the AI podcast. My name is Katie Burke Washabaugh, and I'll be your host of today's self-driving episode.
Now to be clear, there is a human at the wheel of this podcast, but cars are a different story.
Autonomous vehicles have been referred to as one of the most complex AI challenges of our time.
They incorporate high performance compute, surround sensors, and a massive amount of software, including redundant and diverse deep neural networks to deliver incredible benefits, the most important of which is road safety.
Intelligent software-defined vehicles will be rolling out over the next few years.
Companies from Mercedes-Benz and Volvo to newcomers such as Xpeng and NIO are bringing this technology to production.
As we enter this new autonomous era, It's important to learn how we got here and what's on the road ahead.
I'm joined by Sarah Tarek, vice president of autonomous driving software at NVIDIA.
Sarah started at NVIDIA in 2011, working on the developer technology team, focusing on rendering and simulation techniques for gaming GPUs.
She soon began working on one of NVIDIA's first forays into AV development, leading the effort on self-parking on the original Drive platform.
In 2018, Sarah joined Zoox, one of the premier robo taxi startups in the industry, where she specialized in computer vision and perception.
Now, Sarah is back at NVIDIA and at the helm of our self-driving software development, which looks quite a bit different from her work nearly 10 years ago.
I've droned on long enough. Sarah, welcome to the podcast.
How are you this afternoon? I'm great, Katie.
How are you? I'm great. I've really been looking forward to this conversation.
I actually got into this industry as a reporter. covering autonomous vehicles back when Waymo was...
The Google self-driving project and Zoox was still in stealth.
So it's been interesting to see how it's evolved from. kind of this more sci-fi high-tech vision to actual deployment that we're seeing today and kind of the different applications that have that have grown over the past decade.
So I'm really looking forward to diving into this progression and kind of what's on the road ahead with with someone who's been on the ground floor of this development.
So just to kick it off, I was wondering if you could talk a little bit about what drew you to autonomous vehicles in the first place.
Yeah, I mean, autonomous vehicles is such an interesting combination, right?
Because it's... It's this technology that has a huge positive impact globally on so many people.
At the same time, it's so interesting Technically, there's so many problems that are yet unsolved and the technology that powers it and the work that we do is just evolving so rapidly.
This combination makes a big impact and is also technically challenging.
That's something that I've been sort of chasing for most of my professional career.
I've switched between a couple of different industries computer graphics, high-performance computing, and AV.
And I think this is the place where I found the best intersection So yeah, if I'm doing anything with my life, why not do something that can make a huge difference? in not just safety, right?
AV is not just about safety, it's just improving the quality of life of so many people.
Yeah. Along those lines, I mean, safety is of course the primary benefit of autonomous driving technology.
I mean, especially As we look at the most recent numbers of traffic fatalities, they've started to go up during the pandemic.
There really is this imperative to improve road safety and kind of address these kinds of dangerous conditions.
But aside from the clear safety benefits, what else do you see in autonomous driving, what makes it a technology worth investing in?
Right, so autonomous technology is already getting to be ubiquitous and it's going to be even more so in the coming. in the coming decades.
It's not just consumer cars, it's robotaxis, it's delivery trucks, delivery robots, even farming.
It has this impact over so many industries, It not only makes all of these tasks safer, but makes them less mundane, gives back time to people to do other things. provides accessibility where before maybe, you know, there was none from everything from providing mobility to people that would not have had it to, autonomous trucking that you know has the potential for being able to transport much more goods to places that people need them so that's what i said right like it's holistically improving quality of life for people everywhere.
Yeah, no, before COVID, I had to commute an hour and a half both ways from San Francisco to Santa Clara.
My quality of life could definitely be improved with autonomous vehicles.
But yeah, so just. Rewinding a little bit to kind of the early days of this technology, I've heard a few people speak about the development of AV perception and the different unexpected cases that arose.
And I know I've heard stories of kind of chickens running in the road and how the car perceives and detects them.
I remember there's also a story, there's a California DMV report of a pedestrian walking up to a cruise test vehicle and just punching the windshield.
So there's so many unpredictable factors in the world.
So I was just wondering in your experience, especially kind of in the early days of this development, was there anything you experienced or witnessed that would surprise people?
Yeah, I mean, so much has changed over the past 10 years.
I remember probably about 10 years ago, right?
It's fun to try to do this, step in front of an autonomous car and see what happens.
When I did, it actually did. it didn't break on purpose.
It's like, it didn't break by itself. And I wonder like, what happened here?
It's like, am I, do I just look slightly different from every other that we've trained these networks on.
And there was definitely a lot more work that we needed to do to make these algorithms more generalizable you know, more robust.
And we've gotten to the place from there to now where basically, right?
When we look at false positives that these networks are generating, so basically a false positive is any time that the network says there's something here and it wasn't.
We think it might be an error. We look at these false positives and it turns out Oh, there was actually something there.
And I, as a human, I'm wrong in thinking that there is no object.
I mean, maybe it's very dark and there's a person in the shadows behind. behind some occlusion, but they're there and the networks are able to catch what it is that humans are no longer able to label and catch.
So that's such an impressive change over just a decade.
Yeah, no, and we were talking about 10 years and it sounds like a long time, but in terms of the type of technology we're talking about, I mean, that's, That's a relatively fast pace of development.
What breakthroughs needed to happen in these past years in order to get to where we are today? just the amount that deep learning has improved is amazing.
So I think the three things, the three teams that have really pushed innovation here are the drastically increased amounts of data and compute, and then the increase in the sophistication of network architectures and the methods that we're using to train them.
So interestingly, this year is the 10th anniversary of AlexNet, which was a real turning point in the popularity of deep learning methods, right?
So basically this was, AlexNet was a, By today's standards, a very simple network, but it really changed the accuracy that people were able to achieve on simple image classification tasks.
And it just beat hands down every other approach that was there at the time.
And the thing that Alex net, like the combination of factors that helped it really beat everything else was the an increase in the amount of data that they use, because now they started getting access to millions of images. an increase in the amount of computational power they were able to use because they started using GPUs and then just using a CNN.
And the decade that has followed since this paper was published is really just an exploration of the same themes, right?
Like we have gotten exponentially more amounts of data and more amounts of compute that we're using to train on this data and to run inference on this data.
And we have made better and better and more sophisticated algorithms to be able to exploit this compute and this data.
When you look back now, I was talking about AlexNet.
This was a network that was trained on two GPUs, a couple of 100 gigaflops, couple of million images.
And now we have networks that are trained on supercomputers.
They have hundreds of billions of parameters.
They're trained on trillions of things. And just in 10 years, having that sort of massive explosion, imagine what it has done for the power of what we're able to do, right?
So from being able to just classify simple images, we are now able to look at the whole world through multiple sensors and extract just 3D information, right?
Like what is all of the objects in the world?
What are their bounds? attributes, all of this stuff in the same amount of time that we were in the past in 10 years ago, able to say that this image is a cat.
So just, you know, it's just massive. Yeah, and especially talking in terms of billions and trillions.
I mean, it's hard to wrap your head around when you really kind of try and sit and visualize it.
But so what are you working on now? What are you excited about?
Yeah, so I am. part of the larger EV organization at NVIDIA and NVIDIA in general is doing so many exciting things. not just in AV, but in many other areas.
But inside AV, NVIDIA is powering not just so many companies that are developing AV stacks, but we are developing a whole platform that enables AV on consumer cars This is such an exciting project.
We are able to specify the hardware, the sensors, and build all of the software stack to enable any consumer car to do the whole spectrum of levels of autonomy starting from very simple stuff like active safety, basically emergency braking, blind spot monitoring,
The same stack that does this also does much more complex AV functions like L2, L2+, and level 3 driving, depending on the ODDs.
Being able to create this entire stack together and being able to ship it on consumer cars worldwide is such an exciting project.
Again, the impact is so massive. Because you get to make a difference in people's lives all over the world.
And at the same time, you get to learn what are the things that these tax are not able to do well.
And then improve them all the time because the software that lives on a car, it's not a one and done thing, right?
You're learning from the world. and you're constantly improving the algorithms, improving the models and shipping better and better software that is providing better functionality, more expansion of ODD, even a better service, better comfort.
For a consumer, it's like you get a car and it's forever getting better and better rather than depreciating, which is how we tend to think about cars.
Being able to work on a project like that I think is so interesting and of course, At Nvidia, it's this combination of we get to work on the software, but in deep collaboration with with basically the entire stack, right?
Like how the hardware is architected, how it enables software to run well.
And so this is for me, I think for most people, a very interesting challenge.
Definitely. I think the idea of a software-defined vehicle It makes it makes so much sense when you think about it, because I think the way we consume, I guess you would say cars now.
It's such a waste of money. I mean, you buy your car and it slowly depreciates and it's such an investment that you never really. kind of get the full value out of.
But the idea that it just keeps getting better and there's just, there's always something new.
I think is really exciting. And I think it really redefines what it means to own a car or own a car or or travel by car.
I think that's a very exciting concept. Aside from kind of software-defined vehicles, what else do you see on the horizon for intelligent transportation?
Right. So we talked a lot about AV, right?
And that's a big part in my mind, obviously, because I work on it for intelligent transportation, but there's there's so many other things like the whole experience inside the car.
I think there's a lot that will be improved.
NVIDIA is working on Concierge, but basically every occupant in the car should be able to access Very seamlessly, all of the services that they're used to in any other environment, in their homes, with their phones and so forth,
So yeah, again, you want this, like being in a car to be this, this really positive experience and you want to be able to leverage that time for your leisure, for working, for whatever as seamlessly as possible.
Now to kind of switch gears, so to speak, Out of curiosity, are you a car enthusiast or do you like to drive as someone who's developing self-driving technology?
Do you have any interest in human driving?
Yeah, so I mean, I love the look of classic 60s sports cars, right? the Jaguar E-Type or the Shelby Cobra.
I have zero idea of how to maintain such a car, but in some ideal world, Such a car would exist with a completely modern interior.
It would be electric. And perhaps this is heresy to say, but it would even be self-driving.
Well, do you know how to drive stick shift?
I do, I do, but I'm no race car driver. Still, I think that's a rare skill in today's world.
That's what I grew up learning to drive.
So yeah, definitely, definitely tough. All right, and then any predictions for the next 10 years of intelligent transportation?
What do you foresee coming down the road?
Yeah, so I think we talked a little bit about the last 10 years and in some sense, maybe 10 years ago, we were all young and naive and thought by now, We would have had robo taxis out in the roads and maybe part of our community got a little bit disillusioned in the past.
But I do think we are now at the point where In the next 10 years, absolutely, I expect to see more and more consumer cars that have these higher levels of autonomy very specialized use cases that are running autonomous, whether it is like shuttles, which is already happening in certain cases, but it would become much more prevalent In restricted ODDs, having fully driverless taxi, shuttles, robots, and then starting to see more and more expansion of full L4 in bigger ODDs.
I'll restrict my enthusiasm to saying full level four will be everywhere because I don't really know.
It's this last... couple of percentage, right?
Like who knows when it all gets solved everywhere is a, is a big statement, but everything else, right?
Like, absolutely. Yeah. No, I mean, to think that We're even talking about this type of technology as widespread, maybe not everywhere, but to say accessible, I think is a huge, huge progress.
And I think 10 years ago, if you were to talk about this type of technology would seem in a way just totally out there, totally sci-fi.
But I think it's really exciting how far we've come and, in what's up ahead.
I mean, transportation is definitely gonna look a lot different.
For sure, right. Yeah, I mean, and we see even now, right?
The amount of complexity that we can already handle, I think, is tremendous.
And we don't have any I don't really see Moore's law kicking in.
I don't see that we're going to stop getting the increase in computational power that we've been seeing over the past past years.
So yeah, we have, I think, all of the building blocks that we need for starting to put this technology really out there.
Very exciting. Thank you so much, Sarah, for your time.
This was such a pleasure and Hopefully we can have you back in fewer than 10 years and we'll see what you've accomplished since then.
It was such a pleasure, Katie. Thank you.
Thank you. Thank you.
