We've got David today. David and Emma are Cambridge alumni, right? But one majored in sciences and the other in the arts. David, can you introduce yourself?
Hello, my name is David, I am currently working at this automated driving company in an "all in" state. I worked at Baidu for three years, and before that I stayed in the United Kingdom for six or seven years. Materials science had been my major in those years, and so the redirection to automated driving is also a big leap for me.
Anyway, the technology is always changing, so no one said what to study in university is useful in technology.Anyway, to do what you want to do because I think enthusiasm rather than what you have learned is the one that really pushes everyone forward and the starting cost is actually easily compensated in the follow-up.
I think our audience should all be familiar with Emma. Would Emma like to say a few words?
I was actually a liberal arts student but I was most interested in electric cars. And then I find all this new power of the EVs, all coincidentally, have been putting a lot of effort into the automated driving field. And because of that, I'm particularly concerned about the relationship between EVs and automated driving.
And I also want to ask David that, there are many companies have been involved in the development of automated driving, both traditional car companies and new car-making forces, and, for example, tech companies like Google and Apple, and transportation network companies like Uber and Didi. What is the difference between these companies in respect of automated driving, what are their strengths and weaknesses?
The core of the business: what are companies doing
Ok. To know that, we need to go back and see what's the core of the business. Or what are they doing?
Clearly, the traditional car companies including non-tranditional car companies like Tesla, at the end they sell cars. That is to say, their ultimate goal is not developing automated driving but to sell their cars. And automated driving happens to have the greatest added value when selling cars. So they will develop the automated driving well and sell their cars well at the same time.
We come back to see companies like Didi, also known as express traveling companies. Their business model is to provide services and attract more customers. It's still a question that does it make money? Because this kind of business model is a bit like sharing, you have to invite customers on both sides. Your first customers are those who will serve the other half of your customers.
This business model is problematic. Maybe we will talk about it later. But for such a company, their current business model, including Uber and Didi in China, their ultimate goal is to save costs. Then we need to figure out the cost relationship between an automated driving car and a US local driver.
We'll then evaluate whether the technology can be implemented or not. And there's another category of vendors that are the traditional Tier 1 providers, like Bosch, the Delphi, Denso Japan, and Magna. What they're trying to do? They're actually trying to sell the packaged subgradiented technologies to the OEMs so that they keep having gimmicks to sell cars, but will they perfect the technology before selling it?
They won't. Because their goal is to treat this as like the traditional ABS, like the various recommended techs that we've heard when they selling a car. These technologies are sold to you bit by bit as well, and the technology itself is also incremental bit by bit, which fits in with the capitalists' business sense.
And this is like, I have a good thing broken into 8 sections, and they are sold to you separately in 8 years, that's what it is. So actually there will be three different types of business models as the automated driving actually be implemented.
There is other companies like Didi to do vertical service, but it is not a driverless taxi, it may be in port or logistics. Also, it may be a more subdivided project. Recently, there is a company doing autonomous driving of mine cars which actually saves a driver’s salary.
So if he says need to pay wages adding various additional money as standard in remote areas, then it is worthwhile to install autonomous driving into the added value to replace a driver. So, it's no different from Didi's business model in essence. That is, whether the driver's money can be saved, and whether this account can be calculated back.
Right. David just talked about various groups, like Uber and Didi, some traditional car companies, and also tech companies like Google's Waymo. So for these groups. they all have different implementation plan. David, I'd like to know, from your professional perspective, which way may be more promising?
I believe to come back to the light-assets mode. We're looking first at why these companies in the Bay Area have such high profit margins.
This is because of the boundaries of the product are particularly wide not clear all at one glace, and they are constantly widenable, meaning that their margin is very high. For example, engineers at Google or Facebook, they makes money wherever their products go, and this is their business model, not like that of the traditional OEM's hardware model.
Because the OEM has the same batch of people to produce hardware, and when the hardware is outdated, the same batch of people are going to make another batch of hardware. So it It will always be week after week after week, and it won't be particularly profitable. But these internet companies in the Bay Area has very strong business model.
Let’s see if we want to become a very high-margin company. In the current automated driving business, Waymo may be the most right one for it only develops system and the supported hardware.
It will not get too close to car companies. Let's say sell its tech to car companies or signing for a strategic cooperation plan, and you can use my technology. Definitely not. Will it be possible for Google to operate the business itself? No. Since operating is quite laborious, like a servant right?
So we come to Tesla. Tesla needs to sell its cars, and we need to keep in mind that Tesla is a car company and its aim is selling cars. At present, it can develops the automated driving tech by itself, but it may also cooperate with other Tier 1 vendors in the future.
And there might be some chipmakers come back and assist its work. But not Google. Google won't do that. It must develop a very closed system, and wherever it moves, there is its business field and its source of profits.
I want to ask that how much will it cost for Google to build up an automated driving car? We keep talking about whether the cost can be lower than a driver. So I really want to know.
它的激光雷达，加上其他的传感器，因为它所有东西都是自己造的，我们也很难说按照市场价估一个价格。但是像Cruise和 Uber ATG做的产品，我们是可以按照市场价做一个breakdown，然后把它的BOM拿出来看多少钱的。但是Waymo确实很难一下子说有多少钱，但是十几万美金是必须要拿下来的。
This question can be divided into several parts. According to my understanding, the focus is whether the BOM cost is lower than a driver rather than the big team. Its BOM, as I read, is about twice of a driver's annual salary, around $10,000.
Taking the US as an example, $10,000 includes the cost of its lidar and other sensors. Well, since it builds up everything by its own, so it's hard to estimate the cost at market price. But with products like Cruise and Uber ATG, we can do a breakdown at market price and then take out the BOM to see how much it costs.
But Waymo is really hard to say at once how much it cost. Somehow the cost has to be limited within several hundreds of thousands of dollars.
The lidar is still very expensive, like that of Waymo's solution.
But after it's increased, the cost of lidar is also going down. But the reason why it's so expensive is because he may have to make all the parts by himself for a system. If only to make a certain piece like a camera or a chip program, the volume goes up and price will be cheap.
It seems that the price of lidar has been cut by about half in the past two years?
Lidar also has different types, which can be presented as another topic. The current lidars are all mechanical, and the basic principle is TOF (time of flight). It is very expensive to apply a solid-state radar, so most automated driving companies are still using mechanical rotary radars.
This kind of radar was introduced to the world by Velodyne's owner at the time, who was a brilliant inventor also called David. Now, ten years have passed, the radar is still popular among most manufacturers, especially in the field of unmanned taxi segment, or automated driving in traveling services.
At present, no company is able to combine various advanced technologies at a low cost, such as solid-state radar with vision. These more advanced technical methods are constantly being studied by companies.
For example, many algorithms and new lidar functions are developed in the CV (computer vision) field recently. However, is it possible to keep the cost low? It is actually a matter of time, and temporarily the answer is no.
At present, a multi-sensor fusion solution that equips mechanical radar with a visual camera and millimeter-wave is given priority, but it requires HD maps as an assistant, so there is still much to do.
The HD maps also contribute to the high cost. You can’t be like Tesla who just installs the camera and the hardware and leave it aside without follow-up actions, right? It’s not going to work.
In which areas will automated driving first make breakthrough?
Yes, the development stage of autonomous driving is currently between L2 and L3 for these main car manufacturers and technology companies. There are some like Google that develops the L4 directly but it may still be a certain distance away from actually use. The public is still skeptical about this technology, right?
The ordinary consumers still have to overcome a great psychological barrier to really enjoy a self-driving car, right? So my question is, where do you think autonomous driving can first make a breakthrough in terms of application?
Good question. Breakthroughs should first occur where human beings can already be replaced. For example, the closed field of driving, because there is less uncertainty. The vehicles in the scene, the flow of people, and even the size of the map are all controllable. Therefore, this will be the first place where automated driving can be put into operation.
Let’s take a theme park as an example. The shuttle service between the hotel and the theme park, (Emma: shuttle bus), is easy to be replaced by autonomous driving, but if it is on a highway or city main road, there is still a long way to go for automated driving.
One of the scene on the city main road I might think of is to drive at a low-speed and during the period of a relatively small traffic volume, such as the street-sweeping car.
Yes. The sweepers along the street are what we think will be first on the road.
And at a low speed.
Yes, but actually, it is not accurate to view it as an application of automated driving. When SAE (Society of Automotive Engineers) defines automated driving, it requires that the car should be used for human beings instead of a functional one. What does it mean? That is, if the automation is for a truck to ship containers, it is not within the definition range of automated driving from L1 to L5.
This is often overlooked by many people in the industry and they only focus on the levels of automated driving, but not on the service objects. In fact, if its service target is a port, it should not be within SAE's definition.
The definition emphasizes having people sit in the vehicles, regardless of what level it is from L0 to L5. So if in sweeping cars or for ports in the mining areas, this kind of "automated driving" does not apply to our topic today. But people are used to the concept of levels because it is easier to understand.
Exactly. It needs a scene that is related to people.
Right. People must be involved, because the security is for people, right? If the service targets are objects, we don’t need to pay too much attention to security guarantee.
Is the current automated driving reliable?
Speaking of people, I have another question. Now the so-called automated driving, such as that of Tesla, is actually semi-autonomous, or even just driving-assisted. Part of driving is done by people, part is done by the car. So is it an anti-human design if the division of labor between people and vehicles is unclear.
The question you asked is very good. I often feel that if a machine only partially replaces the manpower, people still need to monitor the whole process. It does not actually save labors. It just frees our hands, but not brains.
I don’t know if you understand. Imagine yourselves sitting in a self-driving car, you still have to intervene and look ahead absorbedly all the way, which is also energy-consuming. Let’s make it more precisely.
It feels like your washing machine cannot be fully automatic, and you need to check every 50 seconds if it is still working. Then it flashes the green light, telling you that “Oh, I am still washing”. Do you think that you are freed from it? Not at all!
对，一个是在精力上没有解放，另外在你的精神上也没有完全解放。我以前听过别人说的，坐这种半自动驾驶车是什么样的体验，然后有一个很高赞的回答就是说：有点像别人给你挖耳朵，你感觉是有点爽的，但是其实心里是很担心的，因为你是不确定他 O不 OK。
True dat. It doesn’t save our energy nor mentality. I have heard a saying about how one feels when driving a semi-autonomous car. One of the answers which got the most likes was: It feels like someone is doing ear-cleaning for you, and you feel cool and worried at the same time because you are not sure if your technician can do it well.
This example is very appropriate.
Yes, from another side, like the L2 level now, actually, If you hit a person, the responsibility is still on you rather than car, right?
It depends. My dad crashed my Tesla, and he put the blame on Autopilot.
But in terms of the law, the law needs a subject.
Yes, like L4, the responsibility is the car, but it's still in a vague area in L3.
Yes, you are talking about a question how to transfer the main responsibility in fact? We held an autopilot meeting in Detroit last year. Some people talked about this issue and said that autonomous driving is still safe as it currently is. Then an engineer from a car factory said, how can you judge that your Autopilot technology is safe and this company has tested autonomous driving for million miles in the United States.
It's nonsense when you compared it with the number of cars we sell. We sold 10 generations of a certain model in the past 50 years. We built an absolute value between accumulated mileage and the number of serious injury and death of this car in the 10 generations. How do you compare with my absolute value?
Your millions of miles are just one portion of my car's annual sales so the autonomous driving company is "dumb" this time. If you like way, I can't compare with absolute value but I can compare with relative value that drive millions of miles safely.
So that means the follow-up insurance companies need data to help him build the model and transfer risk. Who will the insurance company believe? The insurance company is so sophisticated and it will certainly not believe these words said by the autonomous driving company
There was a driverless test held in California last year. A lot of car companies and tech companies were in it. Waymo and then Chinese company Pony AI ranked quite high.
Baidu's ranking is actually relatively high and this report is actually on my computer. In fact, you can download it from DMV California’s official website. This report is worth considering because the scene can be defined by myself.
This is not comparable when I turned around our yard 100,000 kilometers to hand off once and I drove uphill and downhill around San Francisco a complicated scene to hand off frequently. The DMV report actually ignores this aspect, including one of its collision reports, which is the crash report.
You can find companies like Waymo often reported dozens of accidents hundreds of times a year, but some domestic companies are not. Does that mean that those company's technology is not good? It is completely difference between those reported the test in the most complicated situation and others ran tens of thousands of miles around backyard road with an accident.
And the report of the crash is written about why and where and the car crashed, whether it was because of the system or the person so the report is not disclosed. Everyone feels like how important the rankings are. In fact, this ranking is not quite comparable in the industry because this scene is not fixed. Unless we only fix it in San Francisco, right?
Everyone is the same at the same time from morning to the evening with the same state, so it is possible to control the variables. If you do not control the variables, the comparability is actually not high.
That's right, when it comes to this judgment, I just saw someone online asking how to judge the level of autonomous driving technology. Do we have one now? For example, what kind of competition can specifically see which company is doing better.
The judgment? In fact, it is impossible to say a standard can be evaluated at the national level, including the American Society of Automotive Engineers.
The definition of L is only a definition. As for the specific technology, they just take this test report in California or Phoenix and pointed at each other's nose to say, "Look, I've been testing for so long time, nothing happened, and you didn't test much, you still have trouble." But as I said earlier, there are a lot of greasy existences in it, so there is no absolute standard an absolute bar like food and medicine at least.
There are so many companies working on automated driving now. Currently, it may have hundreds of or even thousands of companies. From your perspective, will the automated driving tech company like the mobile phone or computer system, maybe just two or three company will survive in the future? What are your opinion?Will that happen?
Let us recall the reasons why only few operating systems left, as you mentioned earlier. Because it requires an ecosystem, and its developers do not allow me to develop software for ten operating systems simultaneously. However, for most of the autonomous driving companies that we know are called the full-stack company.
It does not open the ecology to you. Of course, there are some systems made by companies like Baidu, Apollo, NVIDIA, and including a Japanese company called Otwell. These three significant comparisons, we call the three major open source systems.
It is open-source. But the mainstream automated driving companies in the United States, they do not share with you or their platform. It is isolated.
Therefore, in this isolated environment, it does not involve that I will merge you, or you will consolidate me because no one knows other operation systems. They are made by themselves. But you mentioned that the bottom layer is the fundamental logic that may be a little bit similar. I don’t think those giants, or we called those leading enterprises.
They will eventually become like traditional operating systems. And in the end, it will have mass companies in the vertical field, but only a few leading enterprises make autonomous taxis. Because the problems of autonomous taxi does not like the sweeping solution we mentioned before. It may not be interested in this field, or it does not want to pay more attention to these subdivided segments.
I also see another question, it's how you think the autonomous driving taxi of WeRide( WeRide and AutoNavi launched together) in Guangzhou.
To spell out the name of the company and then evaluate it isn't my favorite style. I prefer to do this kind of thing in private. Let’s talk about how Waymo operates in the San Francisco Bay Area firstly.
As the world's top company, they still don't have driverless taxis. They still need a very closed and relatively controllable environment. Waymo does not have a safety controlled officer, but I believe that this can't happen in China now. Another point is the operation time. Do they set a threshold? Do they meet various conditions? Like Waymo said, this car can be called in most cases. Buts I think it should not be in China definitely.
China vs. America's automated driving technology
Now there is a pal who is a fan of cross-border traveling, and he asked about the gap between automated driving in China and foreign countries?
就硬核上来说，跟Waymo还有 Cruise automation，还有另外几家美国公司比，可能在算法上还有一些差距，但是这个算法上实际上是要按时间和里程数不断再去往上提高。这个还好，我觉得这个是需要时间，因为他毕竟做得早，比你在这方面有经验。
但是你说从硬件上来堆，就是说 BOM，倒不一定说两边有多少差距了。因为其实这里的BOM就是供应商，很多都是重叠的，在硬件上的差异并不大，而是在软件和针对这些corner case的处理上，中文叫什么？
If you talk about the hardcore, there are still some gaps in the algorithm when comparing with Waymo, Cruise automation and some other American companies. But the algorithm could be developed by continuing practicing with time and mileage. It's not that crucial. I think in terms of time, because they did it earlier than us, so they have more experience in this area.
If you talk about the hardware, there are not too many differences of BOM between us. Because the BOM means the supplier, most of them are overlapping. So, the difference in hardware is not that big. The difference is when you are dealing with corner cases in software. How to say corner case in Chinese?
It describes “Extreme.”
Yeah, for an extreme case.
Yes, we were talking about one-car intelligence. However, for example, the complexity of the road. China is more complicated than in other foreign countries. The report said the way in China is about 30 times more complicated than that of other countries. So, does it mean that foreign countries have natural advantages?
Actually, we can't only classify it as domestic and American. We have to turn it into a place of compliance and an unruly area. This place could be any place; it could be Thailand or Vietnam, right?
That's right, not even Paris, France.
If in the case, you should think about it the other way around. If we could survive in a difficult situation, our tech should be more reliable.
Recently, some newly emerging car companies claiming that Tesla and Waymo of Google probably won't be applicable in China. And this is because their HD map is not localized enough. How do you think of this statement? Is it reliable?
80% I guess. The HD map or most visual program are, especially those based on lidar, crucial and necessary. Here, we put Tesla's plan aside for a while. The HD map is like a pre-installed sensor, and this means that when a car operating on the road, it dose not requires a full-set functions since some feature points have been captured with the HD map.
So the car only needs to match the feature points, and in this way, the positioning and perception function will be brought up to a higher level. Then, based on this principle, the large enough measured and up-to-date map is required.
And this is a problem in China, especially the frequency of update. For example, in Shenzhen, I feel like the road changed and upgraded very often. Say if I work hard on a HD map and it’s only 3-4 months later, the roads have been changed. And my work will be in vain. So you can see the cost is pretty high.
On the other hand, when I measured map, I encountered with this frequent changing problem. This will make the measurement difficult to conduct. And if it is in Guangzhou, the one-way roads are surrounded with hawkers. You can imagine how hard the measurement will be. This is too complicated.
Therefore, it should be simplified. In Guangzhou, they are operating on an island, I guess? This will simplify the situation, which is like the Palo Alto in the US. On the island, there will be much less situations out of control.
So the HD map in this situation is applicable. As for the cost of the HD map now, I’d say it’s expensive. This kind of solution based on multi-sensor fusion, I mean the lidar multi-sensor fusion, will costs more than that of Tesla. But its reliability and subsequent upgradability is better than Tesla's mode.
I heard you. For the full automation, most of the people, like my mother often asks that does it really need no human assistance to do anything? If that come true, I just need to tell the car where I want to go and I will get there. There is still a long way to go to get to this level.
Yes, first of all, you should ask your mother about that idea. Where does she want to use it? It will take a long time to make it happen in Guangzhou. However, if we say that we will now build a new city.
For example, in a Middle Eastern country, we are starting to build this road, we will put everything we need for autonomous driving. The traffic lights can communicate with the car. There are also sensors on the crosswalk that can communicate with the vehicle. All of them are connected. It is much more convenient.
Why? Because most of the problems we solve with algorithms and engineering are what we call the starting cost. Those cities are already like that. I'm going to solve some of the original problems for that, and the cost is very high.
It's not like putting the technology in a newly emerging city, and the infrastructure is entirely following my standards. It's okay. In that case, it will make the full automation happening faster. To achieve the progress that you could do nothing but lie there, you could get your company from your home.
That kind of automatic driving is easier to achieve. In China's current cities, it is more difficult because it means we need to spend a lot of starting costs.
Why the US has advantages? The gap between the urban and rural areas in the United States is relatively small, and the pattern is almost the same. That means all cities in the United States, from east to the west coast, and I think they look the same, except for New York City.
That makes a replicable pattern possible. But the cities are very different from east to west in China. From the small Alley in Sichuan to the one-way line along the Haihe River in Tianjin, who can tell me how to do that? It's hard.
So what's your plan as a practitioner?
That's good question.
As the host who I’m glad to see that Emma asked David an interesting and sensitive question, and due to the length of the video, David will reveal his answer in the next episode.
As the public experience the power of technology, often as it evolves from 1 to N. The guests that I invite to DannyPal show are the frontline practitioners, and hopefully our conversations help you better understand the development of technology from 0 to 1. It's like doing too many quiz questions in daily life, occasionally, you need to go back to the textbook knowledge and think about "first principles" over and over again.
Although Mark Twain said, "To be literate without reading a good book is to be literate in vain," the premise is that you must first learn to read and write. We are in the age of fast-food knowledge, where everyone can comment on national events, technology trends, and industry development, and everyone seems to be the "king of knowledge", but if people continue to ask them questions, they will easily show their weaknesses.
I agree with Charlie Munger's pursuit and flexible application of the fundamental principles of various disciplines, commonly known as the " Latticework of Mental Models." Whether it’s history, psychology, physics, chemistry, statistics, etc., we should all go first to grasp the principle of getting from 0 to 1, and then to learn the flexible use from 1 to N.
There is a proverb that says, "To the man with a hammer in his hand, the world looks like a nail."I hope you won't be the one with only a hammer in your hand.
我是小丹尼，谈车说科技。如果你觉得我们聊的对你有帮助，别忘了帮我点赞关注，这是让我做好下一个视频的最大动力。Techs Never Die，回见。
I'm Danny, talking about cars and tech. If you find what we talked about helpful, don't forget to subscribe my channel. That's the biggest motivation for me to make the next video.
Techs Never Die, see you next time.