Leading Li Wenguang: Pure vision is a huge gap from L4 automatic driving

On May 21, the 13th Annual Conference on Intelligent Connected Vehicle Technology, co-sponsored by the China Automotive Engineering Society, the National Intelligent Connected Vehicle Innovation Center, the School of Vehicles and Transportation of Tsinghua University, Shanghai International Automobile City (Group) Co., Ltd., and the National Key Laboratory of Intelligent Green Vehicles and Transportation, was grandly opened at Shanghai International Automobile City. At the plenary session, Li Wenguang, president of the intelligent driving product line, shared the keynote speech "Safety-oriented, towards the development of autonomous driving". Li Wenguang believes that the current intelligent networked cars are difficult to leapfrog from "human-computer sharing" to advanced autonomous driving. Huawei relied on more than 11.1 billion kilometers of smart driving mileage big data to run out of a new height of safety: its latest ads version has reduced the serious accident rate to once every 7.2 million kilometers, which is 4.37 times higher than the safety level of people driving; while the purely unmanned park VPD (valet parking) has been in commercial use for more than a year, and the safety performance has reached 30 times that of people driving parking. However, the brightness of human-computer sharing data is not the same as the true maturity of L4 unmanned. The entry threshold for advanced autonomous driving is extremely high, and the availability of L4 requires the system to have unsupervised stability of millions of kilometers and the deterministic scheduling capability of the operating system of "Never Strike". In response to the two major focal points of the industry's hot discussion, Huawei has given a clear tone: First, it has a lot of sensor fusion. In the realistic extreme superposition scenario of "night + dark object + backlight + heavy rain and dirt", the gap from L4 is huge; the second is to clearly oppose the "cabin driving integration" of high-end vehicles, pointing out that the two are completely different in computing power concurrency, operating system and safety level. Blind sealing not only does not save costs, but will not pay for reducing the chip yield. Based on rational technical research, Huawei predicts that L3 will start pilots in 2026, high-end intelligent driving will be commercially available in 2027, and Robotaxi, unmanned trunk logistics, and full-speed L4 in urban areas will usher in an outbreak by 2028. The ultimate goal of autonomous driving is to reconstruct a convenient ecology for the whole scene, including payment and charging, on the basis of safety, so as to truly save time for users and rewrite their traveling life. The following is a shorthand for the speech: Li Wenguang: First of all, let's talk about our latest progress, because we have a safety report, which should be the data we ran yesterday. There are about 50-60 million intellectual driving mileage updates every day, and now it has reached 11.1 billion. The one thing I want to focus on here is that our intellectual driving safety level. At present, our statistics show that China has an average of 1.8 million kilometers and a serious accident. Some of the airbags have exploded. Now the situation in human driving is about 5.28 million. We have also counted the latest version of ADS4.1, which can now reach 7.2 million. In the case of intelligent driving, our average refers to the past year, because it is rolling every month. This should be the data rolling to April 30. We are seeing that the latest version should have achieved 1.2 million. The progress we have made in each version is still relatively large. We can see that we are now driving about 4.37 times better than China on average. At this point, I would like to talk about parking in particular. Because parking tells you less, we looked at the parking of people, about 6,700 times. According to the normal parking of people, it is about 1,000 times a year. Generally, it will be scratched once in 6-7 years. This is a statistical data. We should have climbed 410,000 times from the last few years to the latest, which is equivalent to nearly 61 times of human driving. Moreover, we had APA and EPA last year, and EPA is actually unmanned, including VPD. We have actually commercially used VPD in the park. I took a look at the data that has been run more than 18 million times, running for more than a year, and various data. Because VPD in the park is purely unmanned, including cruising and parking in the park. In fact, it has relatively high requirements, and it has reached a level of 30 times. In fact, we are still very confident in low-speed and parking, and we can achieve automatic driving as soon as possible. But the above mentioned ones do not mean that we have now reached the level of L4, so I would like to talk about this objectively, because the above is about the safety data of human-computer co-riding, but it is not that there is no one, we have talked about the intellectual driving level of our industry now, I looked at the high-speed, we talk about safety supervision, if you do not supervise, there will be safety problems, it is difficult to say, I feel that some are thousands of kilometers in the industry, some may achieve tens of thousands of kilometers, better high-speed can achieve hundreds of thousands of kilometers, the safety supervision mileage of the city is about hundreds of thousands to tens of thousands, there are others in this industry, I am not particularly clear, we are about this level of hundreds of thousands to tens of thousands. At this level, according to our normal view, it should be 300,000 kilometers. There are also requirements for the industry standard of autonomous driving itself. Its requirements are 10,000 hours of driving, according to the high-speed 66 kilometers, the average should be 660,000 kilometers and about 300,000 kilometers in the city. Of course, this is an entry threshold requirement for autonomous driving. This is a safety requirement. L3 and L4 have the same safety requirements, but there is actually a higher requirement from the usability point of view. L2 According to our standard definition, it is still driving, it is still an auxiliary driver, just to do an auxiliary, people should be online at any time. L3's system, in fact, you can ask people to supervise the system, but you have to tell people 10 seconds in advance, if people do not take over, you have to pull over safely, which is why L3 is a limited autonomous driving, which is one of its requirements. The requirements for L4 are higher, and the usability we see is at least an order of magnitude higher than L3, because for unmanned people, you can't just strike, and the entire transportation system is chaotic, so the usability requirements from L4 are also higher. Let me talk about the challenge. This is that we ran about 10 billion kilometers. We counted a data. We only took the highway and the speed in here. If you want to reduce the number of accidents, then how can you eliminate these problems? We can all have a prediction. In addition, we took a look at it. In fact, the car emergency brake is relatively easy to solve. We have some solutions, including the second most small obstacles. For this, we also specifically counted the size of the obstacles, why did we specifically make a lidar to solve the 14cm problem? Because we see that 15 to 30 centimeters is the most in our small obstacles, accounting for more than 48%. Further down, it is actually only 2.5%, so it can greatly reduce the probability of small obstacles. The previous is just a gesture, the high is to be eliminated, and the low is also to be eliminated. If we want to do this, we require four capabilities for the automatic driving system. One is to see more clearly, the second is to think more clearly, and the action is more in place. The system should be more reliable, equivalent to each requirement It is facing a class of problems. For that system, you need to be able to see more clearly. This piece is mainly for some complex proximity, such as the negative ladder of parking, overhead, including some bad weather, and some small targets. This is our key list. Around this, in fact, we did a lot of sensory software and hardware to do some things, especially including we did some heavy rain and fog distributed radar, including some of our solid state, in chronological order, last year we have launched so far there are four products, of course, there will be some new products behind. Including this piece of lidar we made, we are considering using domestic self-developed chips. It usually takes three years from the establishment of our chip to the launch of our product. For example, the 896 line we made is actually doing this thing in early 2023. That is, three years ago, if it is pre-studied, it will be earlier, so this cycle is still relatively long. The second is in terms of algorithms, we are for autonomous driving, we came out with a 1.0 last year, we did another iteration this year, the core is that we did something stronger with online reinforcement learning this year, because the mileage just caught our long tail problem, but it does not mean that you can solve it, because if you want the model to learn, one or two cases are not enough. In order to solve this problem, we introduced an online reinforcement learning to solve this problem, which will replicate a large number of the same cases, so that we can solve this problem. Second, we have made some predictions. We have to make a substantial improvement to the risk field to solve some problems. For example, if this child passes through a stationary truck, then you should have predicted that the back of him may be worn out from the front of the truck. Therefore, many similar cases have been solved. We are particularly grateful to Tsinghua University. We have also done a lot of in-depth research with Professor Wang Jianqiang. We are also especially grateful to Academician Li Jun and Academician Keqiang for doing a lot of projects. There are movements in place, this is good to understand, because you can see clearly, the brain thinks clearly, it does not mean that you can do well with your movements. When the self-driving car is driving on the road, it will inevitably encounter ice and snow weather and foggy weather. These things will also require our car to have better performance on ice and snow. We have achieved in-depth integration with the chassis, including we have to solve some low-attached pavements, such as open pavements, such as punctured tires. In various scenarios, of course, we have to do in-depth cooperation with the actuator manufacturers here, and there will be a large number of products this year. Finally, let's talk about the ability of our operating system, because assisted driving may not be able to see how high the requirements for the entire system are, because you have a lot of problems in front of you that obscure your system, but we all require stability of nearly one million kilometers after automatic driving. Then you solve your system, first of all, you must solve the reliability of the entire system. This includes how your system isolates such faults, safely isolates, and recovers autonomously. In these aspects, at the same time, you must prevent intrusion, tampering, leakage, and proliferation by others. This is another aspect of our requirements, because there is no one, and the requirements for this aspect are naturally up. Another is that as a real-time operating system, then how do you solve your real-time scheduling problem is to solve the certainty of scheduling, you must complete the tasks you must complete in a short period of time, you can not have obvious jitter, which is the core mission of the three aspects of the operating system. Combined with Che Yun, I would like to say that we are in the road collapse, we have also done some research on this case, because we are in the ABS cloud, we have a large cloud, we now do a bicycle, our ability to see how to improve. However, in some cases, such as the road collapse at night, it is still difficult to find a good example. So far, we have not found a good solution, but we have to solve this problem of group deaths and injuries. Just like Meida Expressway, which is a problem that brings group deaths and injuries, the impact is very bad. We have to solve this problem in combination with the cloud. Our current car discovers an accident or an accident itself, and will inform our surrounding vehicles of this accident within one second, which is equivalent to establishing a rapid problem awareness ability. Of course, we do not only solve the safety and comfort of users from point A to point B, we still have to solve the problem of how to make life better and more convenient after users have a car. At the same time, we have also built a lot of ecology including payment and insurance. With these ecology, combined with the ability of autonomous driving, we will help owners save time in this area, so that you can have more time to experience a good life. What we are currently doing here, of course, we will put the major cities in the country, including the fourth and fifth tier cities, should be gradually improved, including parking lots, charging stations and car wash shops. This is our current progress. This should have been covered well in many large cities. There may be another problem. I saw that Musk sent 12 bits a while ago, saying that multi-sensor fusion can also be solved through images. There is a point of view. We have to talk about multi-sensor fusion. He said no, you see that I can also solve this problem with that pure vision. Of course, the 12 bits he sent must have improved. If you use 8 bits, SP processing must have lost a lot of information. But what I want to say is that 12 bits is not enough. From our large number of statistics, from the visual point of view, the most difficult thing is the identification of night and dark objects. The object that Musk sent is a bright light object during the day. You may not see that information on the display screen, but the information transmitted from itself is actually there during the day. It is only 8 bits of high brightness plus 8 bits of darkness. The two are processed, but for the whole night, because the dark object is hidden in the background noise, it is itself submerged by the noise. If you process it again, you can't see this object. Of course, this was not the most difficult. The most difficult thing was actually the night plus dark objects, plus backlight, and there was a strong light on the other side. The most difficult thing may be that at this time you are just in time for a rainy day, your camera is hanging water, or dirty, or dust. In fact, in the real world you can encounter a variety of cases, it is far from the real thing to achieve L4. I think the gap between pure vision is still quite large, which is the point I just said. Some people often ask me, do you want to do autonomous driving in one cabin? What I want to say is that the two positioning gaps are very large, which is not the same thing at all. Regardless of the algorithm used, the operating system used, and the requirements for functional safety and network security are different. The cockpit needs to develop an ecology. Now it is not running Hongmeng or Android. You can't run to a highly functional and safe operating system with high operating system requirements, so we will not put it together. Of course, the most discussed may be cost savings. We can see that for high-end cars, because this is definitely not cost-saving, because for high-value cars, the cockpit must also have strong capabilities and strong intellectual driving ability, and the key is that these two are concurrent. We have also discussed it. Because we are both chips in the cockpit and chips for intelligent driving, we have been discussing whether we should also study one. The results of the study are that these two are concurrent in computing power and can not be saved. So what can we save? By sealing the two bands together, you can save a package cost by combining one chip. However, after the packaging is large, the inevitable result is the yield of the piece. You may have saved a little packaging cost, but you have reduced the yield, and the overall benefit is negligible. So we think it may be a low-end car, but the tram province is also to use it this way after letting go of the two films. In that case, there will be a saving of about two or three hundred yuan. From our analysis, from all aspects of PCB. But the complexity is still great, because you involve different operating systems, running on a chip, for our software, at least I now see that we Huawei should not be able to run these two operating systems together, we do not have this software capability, this requirement is too high. This is the question I want to answer here. We should belong to the more radical side of autonomous driving, but we are more reliable than Musk. He often says what is coming soon, and the year is coming soon, but our predictions are relatively reliable. We should do a pilot this year. Next year, we LG will be commercial on a large scale. L4 We think that 2B is relatively simple, and the low speed of L4 is relatively simple, because after you put the speed down, this complexity is a square or even three-way decline. After the speed down, because combined with our VPD campus, we all have this experience, we actually have business experience, so we dare to say so. So by 2028, we believe that the commercial use of Robot Taxi, including our D45, L4 commercial use, including the pilot of unmanned trunk lines, and the pilot of these parts of the city at full speed in the urban area, we believe that it is certainly achievable.