October 27, 2017, by the Tsinghua University, Suzhou Institute of guidance, fresh cars, research car co-sponsored by the 'automatic driving fast industrialization of the password' theme salon held in Suzhou. Tsinghua University, Suzhou Automobile Research Institute Cheng Bo Around the 'China Intelligent Network of auto industry development opportunities' as the theme of the speech.
At the meeting, the intelligent network car is the choice of gradual development or leap-forward development of the wave said that the traditional car prices and Internet companies have chosen a different path for their development.For the existing business debate, L3 stage in the end do not want? He said: 'The original OEMs are insisting on a gradual, Volvo some time ago said that across the L3, from L2 directly to the L4. But I feel this is difficult to cross, from L2 to L4 this gap is still very large.' In addition, He also revealed that China is forming a national level of intelligent network development platform for innovation.
The following is a speech by the president:
Tsinghua University, Suzhou Automobile Research Institute, Cheng Bo
The intelligent network of automotive industry technology development volume amplification
Today we are the first time in the form of salon to share some topics.As we have been doing in the intelligent network car research for many years, domestic and international also do some forward-looking technical research, including the scene also has some achievements But we feel that these things still need to do together, whether it is to do the technology, including the industry, etc., or need everyone to participate in. We actually have a large annual meeting with the Automotive Engineering Society, but feel the annual meeting of the year Not enough, so we have a vision, hoping to be able to pass this salon relatively loose, more flexible way, including the technical researchers, including the education of talent, including all the investors to explore, A basic mind.
During the period, we will also publish industry research report by Tsinghua University, Suzhou Automobile Research Institute through research and complete with the fresh car team.In addition, through the industry hot event discussion, build Around the automotive, transportation, Internet, communications and other areas of communication platform in this, Tsinghua University, Suzhou Automobile Research Institute also hired fresh car editor Liu Xiong for industrial research director, to speed up with the Institute of internal industry research team depth interaction, through the fresh Automotive media advantages, the intelligent network of automotive industry, the development of the sound volume amplification, and promote the industry is developing.
Introduction to the concept of intelligent network
Today we discuss this topic, belonging to the intelligent network car.
The original we said that smart cars, more or automatic driving, it is through the car sensor to perceive, decision and then control the definition of the network which does not need intelligence, you can only network, and the two together is the intelligent network Car, which is the source of a concept.
The SAE Automotive Engineering Society of the United States made a definition of the degree of intelligence from L1 to L5. The first level is a single function of the prediction, there are obstacles in front to help you emergency braking, or help you to play the steering wheel to circumvent, or give you Warning. The second level is integrated, the third level part of the vehicle to achieve automatic driving, the fourth level is the general structure of the road can not need people to drive.
Last year's technical road map inside, it is divided into three levels of this definition, first-class is through the network to assist the car driving, such as providing information, said the front detours when people can not see, to provide information. Perception, and the third is decision-making and control, which constitutes our basic concept of intelligent networking.
There are some concepts, one is that we know the smart cars, intelligent transportation, as well as car networking, intelligent network car is in the middle, they are related, and have their own more precise definition.
Smart network development needs cross-border integration
The other is the smart car it is cross-border integration, the light industry has done a smart network car, the other car without traffic environment, no matter how good the use of intelligent network vehicles themselves need to perceive, Control, to be able to self-navigation, to be able to interact with people, which is its own function, but its industry itself and the ecological environment requires everyone's integration, there are positioning, maps, V2X, cars and cars to contact , Car and people to contact, car and infrastructure to contact, as well as information and communication, car sharing services, large data, of course, there are other, constitute the entire industry ecology.
In addition, once the trip has an accident, there is a responsibility, unmanned car hit people how to do? Who will take responsibility? Unmanned car can not be on the road, the right way to define? Especially now immature Product, who dares to put it on the road? So there are legal issues, there are standardized problems, the standard problem, the insurance problem, there are moral problems.
What is the moral problem? We may have heard that the unmanned driving vehicle is controlled by the engineer's program, which is the risk of the future is to choose who hit it? Is it against the wall or hit people? For the car and the outside, this design is Altruism or selfishness? This involves a moral issue.
By looking at this more detailed map of the whole ecology, it is found that it involves upstream hardware and software problems, processors, storage, and driving hardware, software, business applications, including technical solutions, as well as traditional vehicle manufacturing, Ecology, our car manufacturers only accounted for a corner.
It is expected that the future of the future profitability may be the largest mobile service providers in the entire value chain which may be the highest occupies. There are platform service providers and terminal software equipment, while for other hardware, This is not dominant.
Progressive development VS leapfrog development? Traditional car prices and Internet companies choose different
For the smart network car in the end should be how to develop? This has been controversial, or everyone in the choice of a number of different directions and path, is also inevitable, because are in. Intelligent network car is a portrait of what the future is now Also did not draw too clear, just said sub-five, divided into six, which is also an ideal state of our.
From now on to the whole industry, one is the traditional car factory, including its parts suppliers.There are some new Internet, ICT and some service enterprises in the development of the time there are two different paths, For the traditional host factory is certainly progressive, because it is too heavy burden of transformation, he can not throw away all the cars now, one step into the unmanned, it is impossible. In addition to some of its suppliers Must go together, one is the commercial benefit sharing mechanism, another host factory's core technology is not in the hands of parts suppliers.
Another Internet company, the choice is directly from the height of the automatic driving here to do, because the manufacturing is not his strengths, so this is two different paths.
L3 level fear is difficult to cross
We say how many levels, in fact, now really get down, do not need so many levels, a high degree of automatic driving and fully automatic driving can be made into a grade, 1 and 2 can actually make a level, 3 may be a transition, In fact, it is enough to divide three levels, and it is 100% when we go to level 3 and level 2, and it is human that we do not use driving support system. The control can be transferred, I am the command or system requirements, the system can take over the car driving. That is, during this time the accident, may be the responsibility to re-division.
In addition to the system technology requirements, more of its reliability problems, Fail safe is the system failure to ensure its safety, first of all to stop the car, to stop or emergency braking to the vehicle responsibility entirely by the system Control, this time, this time and space he is completely responsible for this time after the failure, the car had to drive, have to travel, the speed may be down. The aircraft is now empty to the air, it must require the system Is the number of sets, this for the reliability of the entire system made far beyond what we now imagine the request.
And now there are enterprises in the debate, L3 stage or not? The original OEMs are insisting to step by step, Volvo some time ago released L3 they crossed, from L2 directly to L4, there are these claims, but I feel this is difficult to cross , From 2 to 4 this gap is very large.
The Game of Two Technical Routes of 'Cycling Intelligence' and 'Network Intelligence'
Tesla is now the level of intelligence is actually 2 level, it is now basically do not rely on infrastructure, more is to use it to carry the sensor to the car to the smart network, how to make a way? To achieve automatic driving.Go Google is relying on the beginning of a high-precision map, it will be a city to use three-dimensional lidar to build all the models, the car is actually running on the digital map, it can be compared and accurate Positioning, to control the car.Now it is generally accepted that Google is more, Europe, North America are in a lot of use of road data.As Tesla do this may also have its reasons, really want to wait for infrastructure improvement , The car does not know when to be on the road, the need for limited conditions, the first car can run up, this is two different technical route selection.
The new platform technology is now relatively thin
We now say that intelligent network car is a subversive product, the so-called subversion is not only from the use, in fact, we in the vehicle's architecture, some of the core technology, there have been some new breakthroughs, , We said that the traditional car did not, as the support of the current intelligent network of automotive technology, not a single technology. Perception, decision-making, electronic and electrical architecture is the original, but in the case of intelligent driving, electronic and electrical structure to be re- There are also car terminals, and now we are all one way, just playing the message, after it is two-way interaction, after the voice of the interaction, there are some visual interaction, it will have a lot, including information interaction.
There are car computing platform, and now we are using micro-calculation ECU to the future in the car needs a lot of data processing, to judge, so the calculation platform must have. High-precision maps, large data, cloud control, As well as information security, artificial intelligence, these are indispensable to the future, which is now accumulated from the global terms, or relatively thin.
Artificial intelligence is vital
In which the artificial intelligence, the reason why smart cars so fast development, in particular, recently began to accelerate, artificial intelligence in which played a relatively large role.From artificial intelligence, we can now see some outstanding performance, such as AlphaGo defeated the championship players, it is in the calculation, in the local game in this regard to the extreme. To the next step to perceive intelligence, one is to listen to, one is to see, now from the data speaking, the performance is more than Master, and the future will certainly be stronger than the people. Cognition of this piece it is necessary to understand, now in the car has not, in some medical useful. The second round of artificial intelligence when the outbreak is 60 years, is also medical Expert system, its recognition accuracy is much higher than people.
In addition to the perception of this, in the behavior of decision-making, direct control of this piece it can also play a role, but at present, or in the perception of this now there are some more clear results.For the enhanced learning and migration learning this, now NVIDIA Do the car in the attempt, it ran on the highway 10 km, or in the special primary stage.
We go back to the simple talk about the autopilot, its basic task we think that one is the perception of the environment, one is driving the decision, one is the control of the vehicle, the vehicle control which is actually the traditional car which has basically been solved. Perception, one is the decision-making, which is the need to add. Aware of it inside the environment to identify, there are scenarios to understand, there is behavior decision-making decision-making path planning, which is subdivided.
There are two kinds of methods to do this in this way, one is that we speak the traditional, it is based on the rules, some complex road environment to identify it into a separate problem, and then one by one to solve, this is us One of the basic logic of solving complex problems is to decompose and simplify this complex problem, but I remember that it was suggested a few decades ago that complex problems do not have to be solved by this result. When he said that the theory of practicality is not strong.Now we say that the end of the machine for the depth of learning, it may not have strict logic or mathematical logic, the so-called end-to-end is that I have an input image, I can identify this road On these obstacles, and I can also directly extend to produce some of my instructions, which is end to end.
Here we have the advantages and disadvantages of the two methods, we say that now end to end this piece of recognition, the accuracy is indeed very high, but it has some fatal weaknesses, one is its interpretability is bad, is the black box, the results In the end is reasonable and unreasonable? Not clear, why is this result, it is not clear, because what is generated, the reasons for the basis, also do not understand. The biggest problem is for our car manufacturers, a problem I would like to quickly find this problem and to replace it.If we control the system after the study, say the problem, how do you find this fault? That is not find, so this can be interpreted Has become a major future tracing the quality of the difficulties.
Also brought a problem, the future is online learning, that my car to sell you is a kind of time, you open every day that place, you perennial to open the streets of the streets and the streets of Suzhou to learn the results of the year is not the same, you In Beijing to practice out of the car, to New York does not necessarily run the US car certainly finished training to China, the pace of these things to bring a lot of uncertainty.
Its advantage is for the algorithm, for the complexity is very high, people do not have to write code, and you have a basic structure I have to carry out some adaptive training, this is sought after by everyone.
This is when it comes to the perception of the environment, we will break down again, with everyone professional science point of view, environmental perception can actually be divided into three levels, a level is to identify the car or someone on the road, the tree is the second scene Understand, there are these things, it is a thing, in front of the car is in front of me or to turn right, or to turn around, the need for understanding of this scenario. The third for other objects of the forecast, pedestrian Ye Hao, the vehicle Yeah, only this problem is solved, we say that perception, knowledge is perception, perception, cognition, and then these are resolved after the decision, is the control, only to the vehicle. Now most of us work in this piece.
This is the information obtained, and now Baidu through some competition, said the signal recognition accuracy has reached 3 9, pedestrian detection has more than 95%, there are other indicators, including vehicle identification may also be very high accuracy, this than our traditional Do the algorithm of environmental identification is much higher, the future they are in the challenge of pedestrian recognition rate to reach a few 9.
Semantic understanding, this is part of the situational understanding, because the car is running on the road, which is where I can run, one is on the road, the other is no obstacles, now no, the future I did not, It is predicted that there are some people on the edge of his behavior to be predicted, through different colors to distinguish, and then gathered after it can draw the outline of the region to provide a very convenient conditions for decision-making, which through the depth Learning can be resolved.
For behavioral decisions more difficult, it is traditional, this is the United States SAE get a hierarchical structure, each actually have to subdivide, so this thing to solve or very difficult.
End of the end of this time is not long, or in 1989 on campus, followed by the recent years, with some of the AI after the algorithm to enhance the current is the implementation of the NVIDIA test, the other to the control side of the less.
Has started to establish a national smart car innovation and development platform
Let's take a brief look at some of the opportunities we have in the future because we have just said some of the technical routes to choose, because a product to launch the community, in particular, to interact with people, its entire ecological chain, including the social system must have. Attention, and are it as a national strategy for foreign countries for a product it is difficult to be listed as a national strategy, from the nineties of last century, around the traffic, the intelligent traffic and smart cars, continued this twenty thirty Year in the United States, Germany, Japan are basically in accordance with a rhythm in the Tokyo 2020 Winter Olympics will launch automatic driving service vehicles, our country now set the goal is in the 2022 Winter Olympics will also have automatic driving vehicle.
China now put forward the goal, L3 below the loading rate by 2020 is 50%, this is very high, can not achieve this is really hard to say, now also 3 years .2025 to 80%, 2030 is the new standard , And the car network to 100%, which can be expected, we can not talk about how to achieve the path, as a country, especially the Chinese government is talking, so we have the opportunity.
At present, I feel the opportunity or some, around the definition of the concept now, automatic driving without large-scale application, this piece left us a certain time. The second, the development of the country are leading, industry collaboration such a national strategy , That is, Baotuan fight gangster, China must be strong, because the government's strong appeal, the government has money, so this piece as long as the direction is good, we say that the core technology can wait for another break, just like high-speed rail, I first The road to run you up, the car to run up to you.There are some, one is our car market, any depot can come to us to sell. Second, we are complex scenes, the scene is complicated, what is the benefit? I am here to train the car to other places can not run.You like Europe, the United States trained out to China can not run the third, we now do a unified platform, especially information security This is the next just a straightforward barrier, all Car to the Chinese run must be network, the data must be handed out, we must follow our rules, so I feel that there is a chance. In addition, China is now the wisdom of the city also Is the focus of the next step in economic development.
The latest dynamic, and now the Development and Reform Commission is the force, the Ministry of Industry in the technical innovation to promote this relatively large, Development and Reform Commission is now in the industry this area, is now in the start to establish a national level smart car innovation and development platform.
Another depot is also moving up, this I feel is a chance.
The future will be service to lead innovation
There are several trends, to share with you, to discuss, one is driving gradually into the people, from the last century we said smart cars, intelligent traffic, that is, by intelligent people less involved in car driving, or do not participate in car driving Because the car accident, 10% of the car itself is the fault, Europe and the United States may be less than 5% of these cars, the road environment is about 30%, some places may be more. The reason is 90%, so let people May be less driving or not driving, which is to solve the traffic safety accident focus, and now a year traffic accident casualties 1.2 million to 1.3 million, this is huge.
In addition to improve efficiency, as long as not people driving, there is no temper, the car will be a unified path to go.
The second is certainly gradually to share, China is so big population base, we are in accordance with the foreign per capita car ownership that is absolutely unreasonable, travel needs is just a need, so by sharing, now sharing in the country to promote or So the share, the brand will reduce the value of the customer is to buy the service is not to buy the brand, the other travel because of convenience, we travel without any concerns, it is estimated that short distance, long distance will increase.
The third, moved to the service, we all think there is a barrier to its transformation, in the existing soil to grow a number of new OEMs to replace the future should be by the service to lead innovation, rather than technology Driving innovation. Electric cars may find it better for development opportunities.