Possible path from ADAS to driverless
The 8th China International New Energy Vehicle forum hosted . CST Sanjiao automobile Xiaojun invites you to think first:
first, when will driverless in the automobile industry come?
Second, from now on, what challenges do we still face;
Third, what is the realistic path of driverless realization?
Rational view on the development of automobile science and technology
At present, the whole industry is quite excited, because the four modernizations of our automobile industry “electrification, intelligence, intelligent networking and sharing” are indeed promoting the reconstruction of the whole industry. As an automobile industry, it has a history of more than 120 years since Mercedes Benz invented the first car in 1886.
However, from the 1970s to 2000, the development speed of the whole automobile industry was relatively slow. I don’t know if all of you here have come to Stuttgart to participate in the museum. The cars designed by Mercedes Benz in the 1970s and 1980s are not much different from those in 2000. Even in terms of core performance, power performance and vehicle architecture, there are no essential differences. In the thirty or forty year period, the automobile industry has been gradually improved. After 2000, the automobile industry has carried out a new stage of development, and the whole industry has ushered in a breakthrough innovation in technology and industrial reconstruction.
We should also take a very rational view of development, whether it is the development of electrification, including the development of intelligence and networking we are discussing today. We should grasp the trend of future development and even the rhythm of future development from the perspective of technology and industry, so that all participants can better comply with the basic law of the development of this industry in terms of technology and the preparation of our business model. This is the main content that I share my thoughts and hope to discuss with you today.
When will driverless car industry come?
To get to today’s topic, first of all, when will driverless cars come?
I also heard president Li’s prediction of the driverless schedule. I very much agree with President Li’s judgment. At present, there are actually two camps for the providers of driverless solutions. One is the traditional vehicle enterprises such as Mercedes Benz, including the leading enterprises such as Mercedes Benz, BMW, Audi and even Ford and GM. Another camp is the technology company, Apollo in China.
At present, every enterprise has put its solutions to the market, and the level is also different. The traditional vehicle enterprises are more what we call L2 or L2 Of course, the maturity of technology is different. Most people in technology companies aim at solutions above L3 or L4, but they are more in the development process of some scheme verification.
Each enterprise has basically announced the landing time of their driverless scheme. Is there any difference between the landing of vehicle factories and that of technology companies? You can see that most enterprises launch driverless schemes above L4 before 2020 and 2021. From my own judgment, this is also a very special time point.
Why do you think so? Let’s look at the data. These enterprises that have launched these driverless programs talk more about L4 level programs. This is the case that we selected them to do the road test in California. Because California road test has one advantage, both the test mileage and speed have a leading level. We selected the test mileage in 2017, which is about 350000 miles. The total mileage designed by other enterprises is still relatively small. Waymo’s test mileage is 352545, with 63 interruptions and 18 interruptions per 100000 miles.
About 100000 miles is within this range. If the current driving scheme is used by us, there are 18 interruptions in this life cycle. Of course, the interruption does not necessarily lead to serious traffic accidents. It may be the suspension of the system or other potential risks. Our test drivers actively interrupt the driverless system.
We made a reference: functional safety. Our definition of time is security. If we convert it into driving mileage, we make an assumption that at a speed of 40 miles per hour, which requires us to interrupt at the level of 0.00004 per 100000 miles. I think the result converted to me is also relatively acceptable. It means that there is an interruption accident for every 40000 vehicles. This is when you give your life to a machine in the future. This is acceptable. Our level and real ability to achieve safe driving still have great factors, including hardware failure and the perceived behavior of other vehicles.
In conclusion, we can make a driverless scheme and pilot scheme, but there is still a long way to go to truly meet the requirements of our vehicle regulation level, reliable and safe driving.
What challenges do we have to overcome to achieve driverless driving?
What are the specific challenges of unmanned driving at present?
First of all, in terms of hardware, our maturity is relatively high, but there are still some challenges, including the reliability of a single sensor, including laser, radar and camera head. They all have certain limitations in specific scenes. We are required to make a multi-sensor combination scheme, but in fact, through analysis, there are still different disputes and routes for the driverless schemes of various enterprises, which still needs more verification to achieve the optimal solution.
In addition, the decision-making link, the calculation in the high-performance field, including some development and application of the controller. Of course, although there is no large-scale mass production of actuators, we find it more difficult in this regard. Including the perception link, our algorithm, the accuracy of computer vision, including the method of multi-sensor data fusion, including the application of Surveying and mapping and real-time update, we still have many challenges and much work needs to be done in these fields. In depth learning and role link, we optimize very fast, but in the field of neural network and machine learning, we still need a lot of virtual verification to continuously improve the accuracy of our algorithm.
Our hardware has a good foundation, but there is a lot of work to do. Our software challenge is still very big. Here are several typical driverless system schemes, including waymo, cruise and Audi.
Cruise I define it as a technology company, and more will choose high-precision lidar as one of the core hardware of perception. As a supplement to the high-precision laser radar, our enterprises rely more on the high-precision laser radar. Different enterprises have different starting points and paths, which I will discuss further later.
An important question, lidar and computer vision, which route is more suitable for the final solution of unmanned driving in the future? In fact, looking at lidar, its advantages are very obvious. It performs well in terms of detection distance, detection accuracy and different industries and mines. The first cost requirement is relatively high, but we see the recent development of lidar industry, including the emergence of solid-state lidar, and the cost will be eliminated soon.
The cost of computer vision is relatively low, but its scenes, such as passing through tunnels and darkness, have great defects in this kind of scene. In addition, computer vision depends on the speed of our algorithms and processors, which is also a disadvantage.
In my personal judgment, I will not say that one route or which route must defeat another route in the future. I mean that lidar and computer vision will certainly find the best fusion, because lidar itself has different route choices, and finally find the best route choice. This route choice is the most important consideration.
First, cover all perceived scenes.
Second, it has a good reliability, including the redundancy of hardware and software.
Third, the cost is acceptable. This industry is still a stage of exploration and finding a more balanced stage.
Now many enterprises are doing driverless system research and development, often ignoring the role of the so-called Internet technology in the field of driverless. Why do we pay more attention to the intelligence of bicycles? First of all, it is relatively easy to solve, but from the perspective of the whole society and the whole transportation system, we must first put the car into the whole society or transportation system in the future, because the car itself also needs to be integrated with other means of transportation.
Second, we should take into account that infrastructure can provide some help to driverless and assisted driving. With the development of Internet connection technology, there are many application scenarios in the information interaction with drivers. However, with the development of technology, the actual Internet connection can also play a lot of roles. Through the development of v2x, the accuracy of our perception can be greatly improved and our security can be improved through the mutual perception, At the same time, it also reduces the technical cost of realizing effective perception. In the real driverless stage, IOT can also play a very important role in decision-making and control interaction. The vehicle and infrastructure can improve the accuracy of the algorithm through the interaction of decision-making, and can also reduce our computational requirements for hardware.
Another challenge is also the direction of our discussion. The new generation of automotive electronic and electrical architecture scheme, from distributed ECU control unit to our automotive architecture with bus as the core, is our main architecture scheme at present. However, in the next step, with the intelligent network connection, the current architecture can not meet the needs of future development. These needs first of all, the development of bus technology itself, including Ethernet technology, because we currently see that large-scale data transmission has great limitations. At this time, our new transmission media should be introduced into it.
In addition, we the concept of domain controller, and even the concept of central processing unit and central processing unit in the future. Domain processor has many different domains, such as vehicle body domain, powertrain domain, Infotainment domain, etc. each system has a domain controller. On the one hand, it also has the function of information processing and decision co cooling. The domain controller has computing power. Through this capability, there is a risk of technical impact on the whole electronic and electrical system. Second, it meets the needs of a large number of distributed computing power in the future in the stage of intelligent networking, which should be the development of automotive electronics and electricity as a whole vehicle. How to adapt to the development of intelligent networking technology in the future puts forward new challenges.
Our company has done a lot of work in this field and is also doing research in this field.
Another factor is the cost factor. Let’s take an example of radar from unmanned driving. The cost of radar is not the same product (it may be wrong), but it also has certain explanatory significance. In fact, the optional lidar scheme in the market, from a very high stage, from the lidar used by waymo in 2012 to the solid-state lidar used in production, the lidar price will be greatly reduced and the cost will be greatly reduced. Cost is currently a major factor, but this factor will soon be overcome.
To sum up, in the second stage, the main challenges we need to overcome to realize unmanned driving:
1. Hardware is basically relatively mature, but the hardware combination of perception process and the most reasonable architecture still need to be further explored. For computing power, the development of high computing power and low-power processors is also a further innovation field that needs to be done at present. Networking technology still has a lot of work to do for our infrastructure related hardware solutions.
2. There is a greater gap in the field of software, especially in the link of in-depth learning, which requires a large amount of data for further training, and there is a large gap in the possibility, function and safety redundancy of software.
3. Test verification, we can not only rely on the actual road test, the first time is long, and the second cost is high. The biggest weakness of road test is that it can’t solve the problem of extreme scenarios. It is often possible to predict more controllable scenes, but the most important thing for us to learn deeply is to solve the scenes we can’t expect at ordinary times. On the contrary, these scenarios can not be solved through the actual road test. At this time, we should rely more on other training methods, including virtual testing and simulation warehouse testing, which is an important factor to accelerate the degree of our software algorithm.
4. The cost of this industry is the law of 20 and 80. We do a lot of work to realize the functions we need. From a technical point of view, it only needs 20% R & D investment. Although we have 20% of technical barriers to overcome, we should not be too optimistic. From the perspective of technology industry, it takes 80% or even higher investment to overcome 20% of technology. I’m not so optimistic about when driverless will mature. In terms of scale, it will be at least after 2030.
What is the realistic path of driverless realization?
For the real L4 level experiment, we should consider the concept of geographic fence. When we talk about any automatic driving, we can’t ignore the concept of geographic fence. The real unmanned driving in the whole scene will be finally realized in ten years. We must not wait for this process to mature in the final stage. Reshape a scene, let’s apply a scene, so that our advanced driverless and automatic driving scheme can benefit our consumers earlier and faster, so that we can use it in certain scenes, including whether it can be used in congested conditions and special scenes just introduced by President Li.
Different camps take different ways to achieve our ultimate goal of driverless. Traditional OEMs still consider the problem of vehicle mass production. It is often our L2 as the core of our current work, so that more auxiliary driving functions and conditional driverless functions can be applied in mass production on our vehicles. As a technology company, it focuses more on L4 standard and above.
Here, I would like to give a suggestion to an enterprise that participates in the whole driverless industry, whether it is hardware or software solutions, including system integration. For truly driverless Enterprises above L4 level, as some leading enterprises, whether waymo, baidu or Apollo, in ten or fifteen years, Platform level development can be done at this stage, whether it is software platform, hardware platform, including algorithm platform. Because it has time to make long-term technical investment, it also has the ability to lead the whole industry to slowly gather towards its platform.
Although I am not willing to admit this as a business person, the field of car driverless is the model to realize our mobile phone solution, and I think it is still very possible. Google, as a provider of Android system and our mobile phone enterprises as application software and hardware, is a relatively more feasible solution with lower industrial investment, and this solution is more likely to succeed, so we need enterprise platforms such as Google and Baidu, but this platform is commercial, From now on to the commercialization of technology research and development, it is a long process.
As a small and medium-sized technology company, when you make hardware products and software solutions, you don’t have so much time. Your R & D focuses more on modularization and reduction
If you want to develop your own fatigue monitor system products, brands or software; please don’t hesitate to contact [email protected] , we won’t let you down.
Leave a ReplyWant to join the discussion?
Feel free to contribute!