Self Driving Cars: Under the Hood | udelv CEO Daniel Laury

So I first would like to start with, for those who are not very familiar with it, what is autonomous driving? There are 5 levels of autonomous driving, starting with level 0, which is zero, all the way to level 5, which means that a vehicle can be driving by itself under any circumstances, any condition, anywhere in the world without the help of a human driver. Level 5 of autonomy is a big deal. It's not— I don't think it's going to happen anytime soon. It'll take probably another decade or 2 maybe before we get autonomous driving everywhere in the world under any condition. But what is happening now today at a very, very fast pace is Level 4. Which is full autonomy of a vehicle without a human driver having to take control over the vehicle in specific geographies. What the Department of Transportation calls an ODD, an operating design domain. And this is happening right now. The leading countries in the world that are Leading the world in autonomous vehicles are the United States, that started the trend more than 10 years ago with the DARPA Grand Challenge in 2007, where some teams from universities in the United States were tasked with the idea of taking a vehicle around a a loop and drive it autonomously. And that was— the idea was that the military needed that kind of technology and they wanted to see how far the university had been going in terms of that technology. And 3 teams, Stanford, Carnegie Mellon, and I think MIT, came up first out of this challenge and then eventually became the companies that you might know of today, Waymo, Delphi, and GM and Cruise and all those great companies that have emerged these days. The second largest player in the world for autonomous vehicles is China. China is very big, very isolated in the sense that they are right now not letting anyone else, any other foreign company doing autonomous driving in China. But they have their own companies that are very, very large like Baidu, Tencent, Alibaba that have been working on autonomous driving for a few years now. And because of the size of the country, have been making a lot of progress recently. And then you have a few other areas such as Israel which is very technologically advanced, and the Israelis are very, very good at developing hardware, and they've helped develop a lot of the sensors that are needed for autonomous driving. And then you have Europe that's currently a little bit behind because of the fragmented situation of the European Union and the absence of regulation that is conducive to autonomous driving over there. But with companies like the German automakers, we have no doubt that they're going to be catching up very, very quickly. So what are the tools of autonomous driving? Well, you need essentially 4 types of tools. You need very, very powerful computing equipment. You need a sensor suite. You need algorithms. And data, and you need very, very robust networking to communicate with the vehicles. The computing equipment is what— the improvement in the computing equipment over the last 2 or 3 years is what made autonomous driving happen so quickly because of the amount of data that can be crunched with those new computers. So here you have an example of the newest NVIDIA chip. NVIDIA is one of the leaders in computing equipment for autonomous driving. And this is the picture of the new chip that's going to come out at the end of this year called Pegasus. The Pegasus chip has 2 GPUs and 2 CPUs embedded into one chip that has an enormous amount of computing power. At a pretty low energy level, about 30 watts if they can achieve that, and an enormous amount of crunching, capability for crunching data, as well as redundancies for automotive-grade functional safety. If one GPU fails, the other one can take over. If one CPU fails, the other one can take over. It's the first chip that has that kind of functional safety and redundancy on the market. The sensor suite is a bunch of things that you may or may not have heard about, but you have like, first of all, you have a series of cameras that give you a picture, surround picture, 360 view of what's going on around the vehicle. The second type of sensor which is extremely important for autonomous driving are LIDARs, light-emitting radars. And those devices, LIDARs, which are very expensive these days and hopefully will come down in price as the automobile market matures in this technology and starts ordering millions of them. But the LiDARs are devices that enable you to make a point cloud image of your environment and be very precise, and a 3D image of the environment that enables you to localize the vehicle. It enables 3D mapping, which is necessary in this business, in this industry. And it basically enables the vehicle to see where, like a human eye would see, to see where it's going. So that means literally understanding the size of a curb, you know, where are traffic lights located, buildings, etc. Anything that's And then finally you add other types of sensors to that suite which are more traditional radars to see with a longer range, about 200, 300 meters of range, and then ultrasonic sensors that are a very short range that gives you a view of what's going on around the vehicle in a 2 to 2.5 meter range around the vehicle. So with that suite of sensors we kind of see the vehicle From very close to pretty far out. That's a view of what LiDAR will see, what the computer sees when it decrypts a LiDAR. Algorithm and data is another thing that's improved considerably, and it's essentially what so-called machine learning Or artificial intelligence. And it's using convolutional neural networks to be able to essentially perceive— it helps in perception, to decrypt what the cameras and the LiDARs are seeing and to make sense of it. And that's where artificial intelligence and machine learning takes— a lot of— make— are so critical for autonomous driving. It enables to look at objects and classify them. Say this is a cat, this is a dog, this is a child, this is a car, you know, and then help the computer make predictive decisions based on what the object He is. And then finally, networking is very important. We need to know where the vehicle is. The vehicle needs to communicate back to the base. And in our case, for example, where we have developed a very sophisticated teleoperation system, when we could— where we can actually drive the vehicle remotely from our office, this requires a lot of data. And so you need to have a very robust communication system. We generate about 4 terabytes of data per day of driving, which is considerable.