news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Wei brand Blue Mountain Intelligent Driving Edition

In the past few years, the label of "intelligent driving" has been more closely tied to emerging car companies. Many new brands such as Xiaopeng, Huawei, and Ideal have all made efforts in the field of intelligent driving, with the aim of being the first to occupy the minds of users in the new energy era.

In comparison, the voices of traditional car companies seem much smaller, but in fact they have not laid low in this era. Car companies represented by Great Wall are also secretly accumulating strength, trying to win back a place in the competition of intelligent driving.

Since the beginning of this year, Wei Jianjun, President of Great Wall Motors, has conducted live tests on intelligent driving in Baoding and Chongqing. As the head of the company, he not only took part in person, but also brought in many colleagues from the technical department to endorse the company's latest intelligent driving technology.

Objectively speaking, if we go back to a year ago, Great Wall Motors' intelligent driving generally did not have much presence in the industry, and the L2+ assisted driving products used in mass-produced vehicles were mostly mature solutions from suppliers.

Until April this year, Great Wall released the SEE integrated intelligent driving technology architecture, the core goal of which is to cope with the L2+ assisted driving from highways to urban scenarios. Under this architecture, it is optimized on the basis of the traditional segmented algorithm module, and adopts an integrated large model to solve decision-making problems in more scenarios, but it also has artificial rules in it.

From the perspective of technology, Great Wall Motors' SEE model is no longer a traditional rule-based decision-making and planning model, but an integrated large model with artificial intelligence decision-making. Therefore, its obstacle avoidance capability demonstrates a certain level of a veteran driver. Compared with the previous clumsy method of relying on a team of thousands of people to patch each scenario, the efficiency has been greatly improved.

So now, after the entire group’s massive R&D investment, to what level has Great Wall’s intelligent driving capabilities evolved?

The following is the transcript of a recent conversation between Great Wall Motors Vice President of Intelligence Wu Huixiao, Senior Director of the Intelligent Platform Development Center Jiang Haipeng, and Intelligent Platform Development Center expert Wu Guosuzhou and others and the media, slightly edited by the Titanium Media App.

Q: How does Great Wall’s current intelligent driving capabilities compare to other brands?

Jiang Haipeng:First of all, let's talk about our own advantages. Starting from the second half of last year, we are very grateful to Huawei and Xiaopeng for their support. We originally predicted that Urban NOA would be implemented in 2025. It was precisely because of the in-depth promotion of our domestic leading enterprises. At first, it was with a map, and then gradually evolved into a large-model map-free architecture. It was precisely because of their relatively radical promotion that educated our users and pushed Urban NOA to the hottest scene of intelligent driving.

We conducted an in-depth test drive of Tesla during CES this year. After returning to China, we conducted in-depth evaluations and test drives of all models with intelligent driving functions. I feel that our current status is not inferior to any other company. If we have to rank us, I think we are in the top three. There is no exaggeration at all.

Q: What do you think of the current mainstream end-to-end large model technology route in the industry?

Jiang Haipeng:Now almost every algorithm company or OEM is talking about end-to-end big models. I can tell you responsibly that there are no more than three companies in the world that are truly end-to-end and have already built in accordance with the end-to-end architecture. Even under the end-to-end architecture, many rules and security issues are embedded inside. Because the end-to-end concept itself is not something that Tesla proposed last year or this year. When we first started working on autonomous driving, we knew that there must be modeling, but at that time we did not have enough capabilities, models, chips, or data.

Therefore, we first model the perception, and then further model the model, and then gradually push it to the decision-making end. When we didn't have a model, we relied on rules. Later, we found that once we entered the city, if we still followed the rules, we couldn't meet the needs of the city. If we pile up codes without limit, it may take 3,000 or 5,000 people to write the city rules for an autonomous driving system. Even if you write them, it will be difficult to solve if there is a change. So today, everyone knows clearly that we must use data-driven operation to reduce the number of human codes and reduce costs, otherwise it will not work.

Q: Automakers are relatively optimistic about the end-to-end situation, but there are still some uncertainties. How does Great Wall view risk avoidance?

Jiang Haipeng:End-to-end is definitely the future, but it is not the end. There are even more advanced things now. The end point of true intelligent driving must be to think like a human, to understand the scenarios, and to perform intelligent driving actions based on the understanding of the scenarios.

Let's take a simple example. When we are driving, there is a black plastic bag in front of us. If there are cars on both sides and we can't go around it, we will definitely run over it, because we know it is a soft object and we can run it over. First, there will be no accident and second, it will not cause damage to our vehicle. Now it is different. We have no way of knowing that there is a soft object in front of us. We only know that it is an obstacle and we either brake or avoid it.

In the future, autonomous driving will definitely be based on scene understanding. We are also making plans and developing this aspect. Now we have some foundations. First of all, from the perspective of chips, such as Nvidia's Thor, and we understand that some major domestic chip companies are also defining their own chips in this direction. You need to support chips running similar to large language models. From the perspective of models, there are what we call language models, similar to open AI, which support us to understand, perceive and make comprehensive judgments. This aspect is the final outcome in the future.

End-to-end refers to a set of technical logic based on autonomous driving itself, which is segmented from perception modeling to fusion modeling and finally to scale modeling. Because people are getting lazier and lazier, doing development, especially software code and algorithm, is a very brain-burning job. What should they do if they want to make their work easier? Students who work on AI are particularly annoyed to touch code because they think writing code is a waste of their time. So they are extremely self-driven. I want to convert the code into a model. Once the model is formed, the cloud will adapt itself. Therefore, we believe that end-to-end is a stage of technical development, but not the end of intelligent driving.

Q: Why does Great Wall SEE architecture 2.0 retain both modular end-to-end and fully end-to-end characteristics?

Suzhou of Wu State:Both of these are concepts. True end-to-end does not mean that there are no rules to cover it, but the advantage of end-to-end is that your data can be learned from the trajectory and characteristics of human driving from beginning to end. With the current state of technology, no one can train a completely end-to-end thing that can be used in all scenarios. There is also a modular end-to-end thing that has a perception interface in the middle.

For example, there are lane lines and similar obstacles. The biggest advantage of modular end-to-end is that the model is easy to train, but it does not have the advantage of complete end-to-end, so the two parts are combined. To put it bluntly, running two models, the two models are independent, the computing power requirements are doubled, and a mechanism for data sharing and exchange needs to be designed. This is also a point that makes the model more difficult to make. Finally, there will be a corresponding arbitration mechanism to determine in what scenario I believe a certain model export requires capabilities, so that these scenarios can be seamlessly combined.

Q: What is the basis for Great Wall Intelligent Driving’s specific city opening strategy?

Jiang Haipeng:The logic of opening a city is actually very simple. We ranked the top 20 cities where our Blue Mountain sales are the best, and Chongqing, Chengdu, etc., have an advantage after visiting these four cities, with different types of cities occupying one category each. Chongqing is a mountain city, and the road conditions are indeed complicated. Chengdu is a typical representative of congestion scenes. The city lanes are very narrow, and the traffic volume is huge. Each lane has a bus lane and a waiting area. Shenzhen is a typical representative of high-tech cities, which is very similar to Shanghai. Baoding is a representative of second- and third-tier cities, so each city represents a type of city. After we generalize a few cities to other cities, we will get twice the result with half the effort.

We are very envious of Wei, Xiaoli and Li Auto’s promotion methods, because they already have cars on the market that provide users with this function through OTA. It is not starting from scratch. Now there is 0.5, and adding 0.5 will make it 1. We currently do not have cars on the market, so we can only bring this function from the beginning of sales. These are two different methods.

Q: What adjustments did Great Wall Motors face in the process of promoting urban NOA?

Wu Huixiao:Large-scale popularization will be in 2025. Last year, we estimated that the node is the first half of 2024. Later, we worked with the whole vehicle, including making some adjustments to the model architecture during the opening process. Then, during the implementation process, we also found that the entire industry will face the stage of users accepting human-machine co-driving. The Blue Mountain Intelligent Driving Edition will have this function when it is launched, and there will be some optimization and adjustment in the interaction after the OTA on the vehicle.

Q: What is the current size of Great Wall’s intelligent driving R&D team?

Wu Huixiao:My boss told me before that he is in charge of more than a thousand people, including drivers and TSTs, and the driving personnel are not the right direction for the future. The development method based on understanding of larger model scenarios will actually rely less and less on engineering investment and human wave tactics. The future must be to use a higher quality, higher density, higher talent echelon, coupled with our infrastructure.

Q: How do you view the role of lidar in intelligent driving?

Jiang Haipeng:LiDAR can only solve 1% of the problem. Should car companies spend thousands of dollars on 1%? We think they should. It’s not that our company doesn’t need to reduce costs. The brand and the boss have been asking if we can reduce some costs, but we still insist that we can’t reduce them. This 1% is precisely related to safety. So we don’t plan to get rid of LiDAR in the next one or two years. We still have to ensure the bottom line of safety.

Q: What is the order in which Great Wall Motors’ intelligent driving models will be launched?

Wu Huixiao:Currently, the Blue Mountain Intelligent Driving Edition is equipped with Coffee Pilot Ultra, in addition there are max and pro. We also need to look at the needs of users. Some users feel that they have concerns about using it in this situation, so we will decide whether to promote highway NOA or urban NOA.

Q: When will consumers be able to pay for smart driving with real money?

Wu Huixiao:My view is that business is the greatest charity. Your car must be able to operate positively in order to maintain the positive and sustainable healthy development of your business and industrial chain. You cannot force sales. First, you must make the product well. After you make it well, you solve many problems for users, and they feel that the money spent is worth it. We also need to do a lot of work within our company. The process, IT system, and payment channel must be built. In fact, the industry has gone through this process.

Whether it is intelligent driving or the cockpit, the experience must be improved. Based on the healthy closed loop of the future business model, the road must be paved first from the organization and technology chain. Once you really improve the experience, users will be willing to pay for it. Let users truly enjoy the convenience of technology. They will feel that the money is worth spending and will naturally be willing to pay.

Car companies are in a very competitive situation this year, and the software industry is making losses. I think our entire country, including me personally, must learn to learn what kind of value we should pay for. We pay for food, clothes, and houses, so are we willing to pay for music and intellectual property? The entire society must make this effort.

Q: Why is next year the first year of the intelligent driving industry?

Suzhou of Wu State:It can be viewed from four aspects.

The first is the evolution of algorithms. Many years ago, we worked on intelligent driving and felt that the first thing to do to do intelligent driving well is to do a good job of perception. Perception cannot reach the things behind what cannot be seen. Later, the perception level was improved and it was said that although there was perception, there was no cognition. Today, we are talking about cognition, which is the understanding of the scene by the intelligent driving system. In the past, this was very difficult, and we have done countless explorations, but today's language models and visual language technologies have given us a grasp of the technology of macro-scenario cognition. If we want to move this model to the car, it is not to copy it directly and use it. This is impossible. At least from a technical point of view, this is the means. The large model of intelligent driving may also be able to solve the ability of zero samples or rare samples. This is the algorithm aspect.

The second is computing power. Today, all high-end autonomous driving companies, except for Huawei, which has its own chips and is far ahead, use Nvidia chips. In fact, its core designers completed the design in 2019, but the deadline was delayed. So they have been adapting to the Nvidia chip architecture from four years ago. In fact, there will be a very important iteration of chips next year, whether it is Nvidia or domestic high-computing chips, and algorithms will be put into the car.

The third is data. Over the years, both traditional OEMs and new forces have more or less accumulated a certain amount of smart driving data. When the amount of data accumulates to a certain level, it in turn creates the possibility of a qualitative change in the training of large models.

Finally, there is the cognitive aspect. In the past, people were divided into two camps: those who believed in autonomous driving and those who did not. Now everyone believes that data-driven is the future. Data-driven has one advantage. Now the system is taken over once every 100 kilometers. Assuming that it is optimized 10 times every year, everyone has confidence and cognition in this matter. This cognition, on the one hand, will in turn promote everyone's investment in technology, and on the other hand, it will indirectly affect the supporting soft things of laws, regulations and ethics. This point is about to arrive.

Q: What do you think about Tesla FSD entering China? What advantages do domestic intelligent driving technologies such as Great Wall have over Tesla?

Wu Huixiao:Tesla has been leading the field of assisted driving. We have also seen its performance in the United States. We can only say that there are challenges. First of all, it is difficult to achieve a perfect result in the short term. The roads in the United States are quite different from those in China. There are very few mixed traffic in the United States. When we experienced it on the streets of San Francisco, if there were people, it would still cause great interference to the driver. Once there is a lot of data, it will definitely improve faster based on the end-to-end large model development model.(This article was first published on Titanium Media App, author: Li Yupeng, editor: Zhang Min)