Hello, and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. Before we get started, a quick reminder, you can subscribe to the AI podcast virtually anywhere you get your podcasts.
Apple Podcasts, Google Play, CastBox, SoundCloud, you name it, we're there.
And if we're not, Let us know and we'll get on there.
You can send us an email at AIpodcast at nvidia.com.
You can also rate and review the show on Apple Podcasts.
And you can find all of our episodes on the show's homepage on the NVIDIA blog. type NVIDIA AI podcast into your search engine of choice.
You'll get there. You can binge on all the episodes and get up to speed with everything going on in the world of AI.
We're all aware of the race to deliver self-driving cars that will allow consumers like you and me to sit back and be chauffeured to work, to school, or on our weekend errands.
But the implications of self-driving trucks may actually be bigger in the short term to the global economy.
My guest today is Chuck Price. He's the chief product officer of Too Simple.
Too Simple is a global self-driving truck company based in San Diego, California and operating self-driving trucks out of Tucson, Arizona They also have another headquarters in Beijing, China.
Founded in 2015, Too Simple is developing commercial We'll be right back.
TuSimple's trucks are the first and only capable of self-driving from depot to depot, and they do so every day for its customers.
At NVIDIA's GTC this past spring, TuSimple announced the development of an automotive-grade camera with night vision capabilities.
They're the largest autonomous truck company in the world.
And yes, they're a certified unicorn with a valuation of $1.1 billion after their latest financing round.
So let's talk trucks. Chuck Bryce, thank you for joining the NVIDIA AI podcast.
Thank you for having me. So let's start with something I dropped in the intro there.
Commercial ready, level four, fully autonomous driving solution.
What does that mean? What's a level four autonomous driving solution?
Level four is a designation by the SAE, which indicates sort of the autonomy. say that there are really only two relevant levels in the scale from one to five, one being like standard cruise control, five being sort of the infinite level of The two relevant levels we think are level two, which is driver assist and level four. which is full driver out autonomy within a defined operating domain.
Okay. Our operating domain is on highways and surface streets that have been mapped for our autonomous solution. so we can operate from a depot's parking lot onto the surface streets, onto the highway. back on surface streets and into the parking lots day and night in good weather and bad weather.
And so where is Too Simple right now? Are you operating level four trucks?
We are operating level four trucks today in Arizona.
These are being operated with safety drivers and safety engineers, two people in place because we're still, of course, under development.
Our goal is to be able to operate in a driver out fashion, out of the vehicle fashion, completely autonomous by the end of 2020.
So by the end of next year? By the end of next year.
And just kind of for reference, you mentioned level two is driver assist that's available now in... commercial passenger vehicles, and then the things that companies like Tesla, everybody knows Tesla, they're working on kind of level three going into level four.
That seems to be where they're going. They never call it by the levels.
They say full self-driving. So that is sort of a less specific. specific name, which gives them, I guess, some flexibility.
Sure. But TuSimple, you have level four.
Autonomous trucks that are out on the roads in Arizona in kind of supervised conditions.
And as you said, aiming for end of next year, 2020 for a driver out situation.
That's correct. That's correct. freight.
We're actually operating for revenue with commercial shippers three to five times a week.
And so before we get into the hows of all this, let's look at the why.
Why does this matter? What would an autonomous truck do for the world?
So there's a tremendous driver shortage and it's only getting worse.
This year, it's projected to be in the U.S. alone a shortage of 50,000 drivers.
By the mid 20s, that number will rocket up to approaching 200,000 drivers. drivers missing from availability for commercial operations.
And so what this means for fleets is they have lost opportunities, you know, opportunities that they can't fill for shipping goods.
And for consumers, it means higher prices.
If you're an operator of a fleet, your capital costs are very high and your margins are low.
Because of the shortage, the driver cost is going up. and the margins are already squeezed.
So this represents an opportunity to take advantage of shipping opportunities that they could not before because they couldn't fill the seats.
It's also a more fuel efficient They can save fuel while they're operating.
And it also allows fleets to increase utilization. of vehicles, the average utilization of a truck today is about 50%.
That means it sits in a truck stop or in a parking lot half the time. we believe we can increase utilization to over 80% with this technology.
So this seems like a pretty clear cut case just to kind of grab one of the headlines going around that this isn't, you know, the robots are coming for our jobs.
This is something where. Hey, look, there's already a shortage and it's impacting everything from fleet operations to the prices that people like me pay for all the stuff I order online that gets delivered. let alone the supply chain of delivering things to the grocery store and everywhere else.
We've already got a shortage of 50,000 drivers and that's just going up.
So this isn't taking away jobs. This is filling a need that's a very kind of vital sounding need.
That's right. And it also affects the supply chain of the trucking manufacturing industry as well. they don't buy a truck, it doesn't get manufactured.
So there are many jobs impacted by the present shortage.
We believe this is a kind of technology that it is quite disruptive.
But we think it's a very positive disruption, and it will lift all boats in the economy.
So let's talk about how it works. And, you know, I'm assuming there are things that you can't and don't want to talk about secret sauce wise, but kind of on a On a sort of high level, how does an autonomous driving system work?
And specifically, what might be unique or different to getting a truck to drive on its own as opposed to a passenger vehicle?
Yeah, well, an autonomous system this vehicle's relationship with the other objects in the environment that are relevant.
From there, there is a predictive stage which around the vehicle.
And then there is a planning process which takes the predictions and plans, sort of the next move for the vehicle. then that next move is executed through a low-level vehicle control system. that controls the longitudinal and lateral motion of the vehicle.
So that is an autonomous system. Our particular system is challenged by the fact that this is a 40-ton vehicle. that is articulated, it's hinged in the middle, and it is very long.
It's also... unstable in different ways than a passenger car it's it's you know it's a taller vehicle tends to be roll unstable.
So it's more likely that you will roll a truck rather than have it slide out.
But you also have to worry about, you know, the things that people understand, you know, jackknifing of the trailer.
Right. So we have particular control challenges.
We have particular... behavioral challenges on the highway because it is a large, heavy, and long vehicle.
So our interactions with traffic are sort of dominated by those characteristics.
As a result, we built... a technology that has a much longer perception range than systems you'd find that are on passenger cars.
So typical passenger car perception system can perceive objects perception distance out of the vehicle.
We think that practically speaking the average passenger passenger car autonomous system can see in the 150 to 200 meter range.
Okay. That range is really too short for a large truck operating at highway speeds.
Because at highway speeds, it takes at least 100 meters for the truck to come to a stop.
Right. Every object is a crisis in the making.
You don't know what it's going to do. You need to study it.
So with our system, our perception range is 1,000 meters.
This allows us to see twice as far as the typical truck driver sees when they're operating.
Of course, you can see out to the nearest galaxy.
But a truck driver typically studies the objects that are about a quarter mile away.
Right. And you're talking about half mile, two thirds of a mile almost.
Yeah, over half a mile. And this allows us to not only operate Which we think is important for trucks, but it allows us to apply strategic decisions for efficiency. and sometimes for safety based on information that we're gathering a half mile away.
So when you're actually developing this for industry, obviously safety is always the primary concern with... any vehicle, any autonomous system, but it's not just about optimizing for not hitting an object in the road or being able to react to an unpredictable vehicle that's around you, but you're also optimizing for efficiencies.
That's correct. Safety is of course the primary concern.
But once we have achieved safe operation with a large vehicle, acceleration.
So you want to try to keep it at a constant speed.
That requires, first, a very strong perception system to allow us to make choice that minimize needs to decelerate and trajectories for minimum energy to avoid cases where we would need to make change in the vehicle's movement.
So, for example, swerving is very dangerous with a truck.
You want to avoid that. So it allows us to avoid those kinds of states.
Earlier this year at GTC, you announced the development of a night vision system.
And I know that your trucks are also capable of operating to some degree in bad weather.
What are the specific challenges that come up?
You think about human truck drivers and...
I would imagine part of the driver shortage is that it's a tough job and it's the actual physical conditions of driving are tough. also psychologically, overnight hauls, driving through bad weather, that kind of thing.
Anybody who's driven- at night or in bad weather can relate to that a little bit.
When you're talking about a computer, doing the driving and the autonomous system, everything goes into it.
What are some of the specific challenges involved in dealing with night and bad weather conditions?
Sure. Well night, obviously the computer doesn't need caffeine.
It doesn't get more tired. Yeah. We never, we never get drowsy.
Right. I should say our system, because of the long range we require, we're using cameras as our primary sensor.
There is no LIDAR or radar that reaches the distances that we achieve with our camera suite.
This means that for night operations... we needed a camera system Mm-hmm.
When it is night and moonless and cloudy, the roads get very dark then.
Now, of course, we have headlights, and we use the headlights, but that introduces another problem, and that is the bloom. or glare that comes from oncoming headlights, other lights in the scene.
Especially when you're trying to process images, With low light conditions and suddenly you're affected by very bright lights, you need to have a very high definition, a very high dynamic range capability. with the camera. is sort of a bright light sort of exploding into the adjacent pixels of the camera.
Are you using computer vision or other machine learning related tech to kind of process and correct the images?
Yes, we do a tremendous amount of processing.
That's why we use NVIDIA. solutions in the perception system where we're using an AI approach. to classify objects, but we also have a tremendous amount of software that is stabilizing the image, correcting for anomalies in the image. before we get to the AI portion of the system.
Is AI, or all the technologies that fall under this umbrella we call AI, Where else is that in your stack?
Is it helping to predict outcomes that then inform decisions to accelerate, decelerate? move that kind of thing.
Yes. So we use various forms of AI. We use CNNs for the object recognition and classification. in the perception system.
In the prediction and planning part of the environment, we're using different form of AI, which I won't get into the details of, but it is not a neural net, but it is an AI system.
This allows for a very rich interpretation or prediction of the environment. and provides a very sophisticated general driving model that gives us a very interesting, a very powerful vehicle behaviors on the highway that makes it much more natural than the typical driving system.
I'm speaking with Chuck Price. Chuck is the Chief Product Officer at TuSimple.
TuSimple is the largest autonomous truck company in the world They're headed towards a milestone of their first driverless operations by the end of 2020, the end of next year.
Right now they've got... Level four autonomous driving solutions on the roads in Arizona, but in supervised conditions with humans on board. to make sure everything's safe, but they're looking towards what they call a driver out situation, fully autonomous driverless by the end of next year.
Chuck, let's talk about your background for a minute, how you got into the world of autonomous trucks.
Have you ever driven a truck? study autonomous systems.
How did you get into this line of work? Great question.
So I actually did a count. This is my 10th startup.
Wow. And I've been in the Valley for a while.
And in 2010, I actually moved to China to start a company there. and lived there for three years.
There's nothing much to speak about that particular company.
But when I returned, A friend of mine who is an investor suggested I should meet the founders of a company That was in 2013.
That was another truck accident. automation company.
That particular company was Peloton Technology that was doing truck platooning.
I got very excited about the concept of doing automation with trucks after I met them. and joined the company within a few months, became their VP of engineering.
And I stuck with that through the first product development. over four years.
And along the way, autonomy and eventually discovered Too Simple. and decided I wanted to be a part of it.
So I, a couple of years ago, joined Too Simple.
So I'd say I've got six years in Before that, I was all over the valley, different companies.
As people do in this neck of the woods. That's right.
You're an engineer by trade or by training.
I'm a software engineer by trade. Yeah. And when did you start working with machine learning, deep learning, things that we think of in the AI stack?
Full-time involved in this kind of technology seems too simple.
Oh, okay. At the previous company, it's a controls problem.
It's not so much an AI. AI problem. And so we didn't focus on the AI, the deep learning, the neural nets there.
We focused on very precise vehicle control.
Maybe focusing then on your time at Too Simple, which is relatively short, what, if anything, has surprised you? whether it's about where AI kind of more broadly has been going or the rate of sort of adoption and acceleration of AI tech, or specific to the domain of autonomous vehicles and trucking?
Is there something along the way that really kind of, you know, You're expecting things to go right and they took a left or even looking back, you think, yeah, wow, I didn't see that coming, but it really, really kind of changed things.
Well, I'm going to say two things. The first thing that surprised me was that Two Simple existed and I didn't know about them.
They had been very quiet. Their headquarters being in San Diego helped them with that.
Right. But I thought I knew every player in the space in Silicon Valley, but that just showed my sort of my attitude that Silicon Valley is where everything happens.
So I was first surprised by that. And then I was surprised at how... fast this company was moving and just how quickly they could pull together such a sophisticated system starting from the absolute latest AI techniques.
So it's not so much... Things that we did wrong, it's things that we did right and did right very quickly.
We went from, you know, in three years from. stunning to see how quickly it has come together.
And that's, I think, a testament not only available to us today, but also to the, you know, the great team that Shaodi Ho and Mo Chen pulled together to make this happen.
And so there's this stated goal of first driverless operations by the end of next year, 2020, What about looking a couple of years beyond that and whether specific to Too Simple or sort of the impact? of autonomous trucks on the industry, on society and commerce and such.
Where do you see this headed in the next 30, say, five years?
We've made a concerted effort to the incorporate this technology into factory-built trucks.
So we think In the mid-20s, we're going to see OEMs rolling level four autonomous trucks online. out of their factories.
And we believe that once we achieve driver out capability, that there will be a fairly rapid uptake by fleets because they have a need right now for more capacity and for the efficiencies that this kind of system will provide.
So I think we'll see it emerge fairly quickly once we have demonstrated that we can achieve this.
So I'm going to start saving my pennies now.
So that way, when the level four trucks all roll out, consumer prices come down because, you know, the. the logistics are better for the trucking companies, then I can really take advantage of the Amazon Prime deals.
So... I'm looking forward to that. Chuck, if people want to find out more about what Too Simple is doing now and kind of the state of the industry.
Where can they go online? I would assume there's information on the company homepage.
Sure. Tootsimple.ai is a great place to start. the industry.
We're hiring like crazy. We're at 400 people and growing.
So look us up. We love to end the podcast on a help-wanted note.
That's always good for everybody involved.
Chuck Price, thank you so much for joining the podcast.
It's great to dig into this stuff and start to understand a little more of. you know, not just what goes into getting a 40 ton truck to operate. safely and driverlessly and efficiently on the road, but also what that means to the larger picture.
Thank you for your time. Best of luck to you and everybody at Too Simple.
Well, thank you for having me on the show.
¶¶
