Welcome to NVIDIA's AI Podcast, and we are recording from the floor of the 2017 GPU Technology Conference. known otherwise as GTC, a gathering of the AI faithful here in Silicon Valley.
We all know about driverless cars. Driverless cars get all the love and all the attention and, you know, for good reason because we all don't want to drive.
But we're going to talk in this segment about autonomous trucks, how and why we need driverless trucks in many ways, just as much as we need driverless cars.
And to help us do that, we have Xiaodi Hao, who is the CTO and co-founder of Too Simple, a company that is bringing driverless trucks to the road.
Xiaodi, welcome. Thank you, Michael. Okay, why trucks?
Trucks are, let's be frank, not as sexy as cars.
We get packages dropped off and things delivered all the time by them.
Okay, first of all, we are a startup company.
So we're not ready to solve a bunch of small, rare corner cases by a small group of people.
So what we really want is we want to go deep.
If we wanted to go deep, we have to narrow down our problems.
And to narrow down our problems, we focus on highway transportation instead of urban transportation.
Because urban transportation sometimes could be a nightmare.
I'm trying to define the routes or the road lanes in Beijing Road.
Yeah, you can think of having a 12 pages for a problem definition.
This is the definition of what it looks like to drive down a bedroom.
How do you define lanes in Beijing? Could be just defining that problem could, you know, spend me about 12 pages.
That's a nightmare. Before we get to that 12 pages, what's your background and then how did you get to autonomous driving?
Oh, my background was computer vision. My undergrad was in Shanghai Jiao Tong University and Yeah, you know that.
I actually taught there for two years. Excited, right?
Yeah. So after that, I come to the US in 2008.
I joined the Computation and Neurosystems Department at Caltech under the supervision of Professor Christoph Koch.
Professor Christoph Koch now is the president of Allen Institute of Brain Sciences.
So you can imagine my background is a mixture of neuroscience, cognitive science, and computer vision.
I see. And so those things combine then to help you navigate trucks down the road in Beijing.
You say the problem is 12 pages long. How do you even start?
As a starting point, I just like make several phone calls to several intimate friends who sometimes we know each other, either like my high school friends who is now the CEO of the company.
It's like classmate. And we also had our principal scientist.
We know each other from the conferences.
I read his paper. I love this paper. So I say, hey. you know i loved your paper and we talked a lot on conferences and that's how we formed the technical team Were you always clear on the problem of highway navigation and driverless trucks?
Or did you start with, okay, here's what we're good at.
Now let's figure out What are the big problems that we can go deep on, like you say, and go and solve?
Ah, okay. That's a good question. So initially we're like sniffing around about our topic. has several rules for where we narrow down our topics.
So rule number one is that we have to focus on a problem where 99% accuracy really means a huge difference, comparing to 90% accuracy.
Because nowadays, deep learning is just out there.
A lot of open source tools. So basically an undergrad pull out a network, train it by himself, wait for several days, could get something, you know, decent results out of it.
And pretty good. And that's fine when you're, you know, looking for cats on the internet or gifts or whatever.
Exactly. That's number one. So we really wanted to go deep to do the best.
Our topic has to be that type of topic so that we can reflect our value.
So number one is 99% accuracy makes a difference.
Number two is that it has to be an unsolved problem.
For example, now in 2017, we're entering the face recognition. chapter.
That's not a super exciting topic, like mostly solved problem.
So we really wanted to enter an unknown territory where nobody knows how to solve a problem and I was talking about early 2016, where we just formed our team and thinking about different directions.
And the third thing is that we're really thinking of something.
Whenever we do, we do it three years. For every year, we have to guarantee there is a significant change growth, maybe 3x of the company.
And we have to find a market that is large enough to support the growth of technology.
And there we go, autonomous driving be a perfect topic for all those three topics.
But we didn't really think about going directly into highway driving.
In the beginning, I just said, okay, we do autonomous driving.
This sounds like a moonshot project. And we talked to all the investors and all the collaborators about our project.
Nobody believes us. Believe you could do it?
Human can do autonomous driving by the end of, say, 2020 or so.
Nobody believes that in early 2016. People just thought we're just crazy about it.
Have people changed their minds about that now?
We're talking about, hey, we're going to roll out in 2018.
And people, some of the investors says, can you make that this year?
Investors are always optimistic. And then you prove them wrong.
Pardon the interruption, but we have a favor to ask. leave us a review on iTunes, Google play music, or whatever your favorite podcast platform of choices.
It helps more people find us as always. Thanks for listening.
Now, back to the good stuff. So highway driving was an unsolved problem.
And that's distinct from, you know, I'm an automotive manufacturer.
Yes, I need to think about highway driving, but you guys are going after...
Not cars, but trucks. Right. Because trucks are like tools.
It's always easier if you're thinking about a production tool rather than something that you're utilizing every day. it's always easier to talk to the direct utilization of the tool rather than just like a consumer level thing.
Because consumer level thing is like more abstract. it's harder to convince people to buy something.
But if you can show people that this tool can replace your existing tool and can make your tool chain cheaper, more productive, it's much easier to negotiate with people like that.
So what was the response and people want that?
And by people, I mean, what kinds of companies are interested in driverless or autonomous trucks?
Oh yeah, we have talked to a lot of truck companies and logistics companies in China.
And currently, some of the logistics companies, their profit percentage will be like 2% profits.
And the cost for labor costs will be like 40%, which is huge.
And usually in China, the situation is that there are two drivers in the road. per truck because they're long distance driving and one truck driver sleeps, the other take turns.
And also there is a shortage in truck drivers.
So there are a lot of routes that the people don't want it to go because there are two deserted places.
There's not a lot of hotels down the road.
You have to plan the road very carefully or the weather is so bad.
That's sometimes big wind and very, very dirty area.
So you're not talking about delivery trucks like to my door in Shanghai.
No. You're talking about bringing goods between places on the highway.
Right. How does that then pose a different problem or, you know, you talked about winds and bad weather and You know, what does your system need to learn how to do and how does it learn to be a truck driver essentially?
Oh, it basically based on computer vision technology because, you know, human can drive.
And humans are using visible range of the spectrum of light rays.
Why can't computers use that? That's the same thing.
The basic assumption is that humans can drive using visible lights and the So should be computers using the same type of input.
Is a truck an easier problem than a car, you know, in city streets?
Is it in some sense... more solvable and therefore we're going to have driverless trucks sooner than we'll have driverless cars for passengers or for consumers?
I think there are overlaps and they also have their distinct problems in both topics.
For example, for car driving, you will encounter a lot of urban scenarios, like different other drivers interacting with you and also unknown objects.
For example, maybe you can have a wheelchair All those weird things that would happen more often in urban scenario.
So that's a challenge in urban scenario.
But for highway driving, if we're talking about a fully loaded truck, the total brake distance will be over 100 meters.
Therefore, we have to see even further than the best LiDAR system that you can buy.
Currently, the best LiDAR system that we can buy is like a Velodyne 64 lines LiDAR, which costs $65,000 and that could allow us to see maybe 60 meters.
I mean, maybe other companies could have a more advanced ladder technology could see even push it further to 80 meters, but that's still... too short in terms of highway freight transportation driving.
What we want is over 200 meters of visible range.
And by doing that, I think the only option currently available for us is cameras.
We're highly customizing all the cameras.
We're customizing the lens set and we're customizing our own stereopsis system in order to see depth and the FOV, and the shutter time, and all those parameters are carefully chosen.
So we spend a whole lot of time in... camera and sensing technology so that we can see reliably 200 meters.
So, and you've done that? Yeah. What's the next obstacle that's different for highway and truck driving as opposed to consumer passenger cars?
The only challenge is reliability. We have all the components ready.
Last year, we participated in public benchmarks and we ranked number one in all of the categories that we participated in.
It's about 10. which is pretty good, which means that we have the ability to do the algorithm correctly.
I never tell people that we are the number ones in Tom's driving industry.
Of course, we're not. But the thing is that we have to put everything together and try to combine them in a reliable way. just by combining everything together is still not enough.
We have more redundancy, more reliability.
For example, we now have five, six, seven components just for localization.
Because most of the time you can use GPS, but GPS is not really stable.
And sometimes say under bridge or in a tunnel, the GPS signal loss.
And traditionally you have the IMU, which is inertia, momentum, unit, and that would correct something that helped a little bit.
You can consider GPS and AMU as a two component combination, and now we have a seven component combination. which makes everything perfectly reliable and increasing the reliability.
And we're so greedy in doing that. So I think it's like a never-ending job.
And how does that change the business of trucking, right?
We talked about the labor cost, you know, if you don't pay two people to be in the truck at all times.
So there's the labor cost, but you know, Reliability, efficiency, fuel savings.
I mean, can you be a better driver in different dimensions?
Yeah, I think so. Because currently in the US regulations, I think the driver has to stop driving after four hours of driving and the maximum driving time would be 11 or I can't remember exactly, but it's either 11 hours or seven hours, but that's a relatively short period of time.
With the advance of sensing technology, camera is actually sensing better in the dark than human light compensation and also more advanced the CMOS sensors.
We can drive on days, on nights, rainy days, snowy days.
On extremely foggy days, we can't drive, but usually the highway will be just closed because of the... bad weather.
And so are you saying also that you can drive just continuously as long as there's school?
Yes. What are the regulatory, whether it's China or here, what are the regulatory hurdles to having trucks drive down the road by themselves?
Okay, I think the U.S. regulation is more conservative, I have to say.
You know, the Chinese Communism Party is more like prone to the productivity side. of the regulation.
So I don't think there is any strong regulation that puts against driving continuously as long as there is a driver driving in China.
And we further got a permission from the government in China that allows us to drive autonomously in public roads.
60 kilometers of public road. And we can either do tests, close the road, do our tests, or we just drive in open public road.
So is it likely that your trucks and the vehicles that have your gear are going to show up in China first, potentially, or...
In this year. Yeah, in this year. In the latter half of the year, we're going to roll out in China.
And also, I think there are several states in the US because every single state has a different regulation.
And there are several states more devolved to autonomous driving.
By more developed, I mean they are more prepared to a downstream because it's a new thing and the legislation takes a long time.
And Arizona is one of our favorite state and we're planning to do a lot of road tests in Arizona and There's an anecdote.
The city of Tucson in Arizona, the Chinese translation is Tucson, which is exactly the same as our company name.
So we love the city. Destiny. Well, I look forward to seeing your trucks on the road.
I guess we won't know the difference, you know, as a person passing one by, but, or really we.
Is there going to be like a crown of censors that we'll be able to recognize?
Very low profile thing. So you see that most of the trucks up the windshield, there is a board that reduces air traction.
Basically, our sensors are installed there.
So there's small pinholes opening and our sensors will sneak in the world silently without letting you know.
And you're retrofitting existing trucks.
You don't have to build a new truck from scratch.
We are. actually not thinking about retrofitting because first of all, the retention rate of trucks are about 80% a year.
So there are 20% of new truck coming in.
And second, our truck has to be controlled by wire.
So there are two types of trucks. One type is hydraulically controlled.
The other type is electronically controlled.
For every electronically controlled truck, once we got the protocol, we can certainly fit.
However, there are a lot of hydraulically controlled trucks in China. not thinking about retrofitting those trucks.
Because if we do so, we have to install a robotic arm.
Yeah, reliability on the mechanical side is not tolerable for our truck.
So not retrofitting, but going with new trucks.
I understand. That makes sense. Xiaodi Hao, thank you so much.
And we'll keep an eye out for your trucks on the road, even though we won't recognize them.
Thank you.
