The CEO of Jiyue took over the baton from Li Bin to personally test pure visual intelligent driving: why did I finally abandon lidar?

Tencent News "High Beam" author Xie Wan

A month ago, Li Bin, the founder, chairman and CEO of NIO, called on the CEO to test the product to become the industry standard on the way to challenge the 1000-kilometer range, and a month later, on January 14, Xia Yiping, CEO of Jidu and CEO of Jiyue, responded to Li Bin's call and drove from Shanghai to Hangzhou with Huang Chenxia, general manager of Tencent News Operations, to experience the latest upgraded high-end intelligent driving solution of Jiyue for the first time.

Xia Yiping revealed in the live broadcast of Tencent News's "CEO Test" that he had just experienced the FSD V12 launched by Tesla recently, and he believed that Jiyue's intelligent driving was not inferior to Tesla in terms of product experience. At the same time, he said that the competition in the automotive industry is shifting from the competition of the three electrics to the competition of core AI capabilities.

Faced with the question that "car robots are marketing gimmicks", Xia Yiping said that car robots are the definition of real products, and Jiyue believes that the future of cars will definitely be robotized.

Here's a highlight of the conversation between the two on the test road.

Automotive robots are not a marketing gimmick

Huang Chenxia: Jiyue said that it was a car robot to the outside world, and frankly speaking, from the time it was first proposed, I think there was also discussion and controversy. Is this a gimmick or what? Do you think the focus of the future of "car robots" will be on cars or robots? Do you want to explain the original intention at that time, and whether there was any change in your thinking after you had the original intention at that time?

Xia Yiping: The earliest original intention and thinking about doing this was to think about what is the direction of the evolution and iteration of automobiles in the future? At that time, I saw intelligence, and I also saw the impact of future autonomous driving, including AI technology on cars. At the same time, when I started planning this car in March 2021, I also saw another change in the entire industry at that time, that is, the large computing power chip at the vehicle specification level, whether it is the cockpit or intelligent driving, it has begun to be mass-produced and put on the car, which is actually very important.

Why did Jiyue firmly choose Qualcomm 8295 at that time, when 8295 was not favored by everyone in the market, we want to realize the localization of AI, and let AI provide a better product experience, we must be able to have enough hardware to support the technical capabilities of AI to provide locally, which is the key. At that time, we resolutely said that we would land high-computing power chips on the first car.

First of all, it's not a marketing gimmick, it's the definition of a real product, because we believe that the car of the future will definitely be robotic.

Huang Chenxia: How do you describe exactly what is the robotization of automobiles?

Xia Yiping: Robotization is very simple, you will find that AI is driving every item around us, including the sweeping robot you go to see at home today, in fact, it can already be called a small robot at home. In essence, you can understand that at that time, we were thinking about one thing very simply, let's not talk about cars, that is, what kind of capabilities will the robots of the future have. What we thought at the time was that no matter what form a robot is, whether it is a bipedal robot or a quadruped robot, it should coexist with humans, and the basic core capabilities are the three categories summarized at that time.

First, how humans give instructions to robots, voice must be the most efficient. We don't believe that you put a touch screen on a robot, and you can't command it with gestures, so when communicating with robots, we think that natural language is the best way to communicate with robots in the future.

Second, the robot needs to be self-moving in any situation and at any time when it coexists with humans, and it is able to walk on its own. For example, when a robot horse runs on the street, it is necessary to recognize traffic lights, pedestrians, and zebra crossings, and to know the basic traffic rules. You will find that a robot can move freely in various scenarios in the future, which is the matter of autonomous driving, which is essentially the foundation.

Third, if it is an agent, it needs to evolve itself, that is, iterate itself. In fact, human beings are the same, from a baby to a child of three or five years old, and then to a teenager of ten years old.

Pure vision itself is not a matter of cost, it is a matter of technology

Huang Chenxia: In 2021, when you started to make Jidu products, you considered automotive robots, and did you choose pure vision for intelligent driving at that time? After all, the whole industry was still discussing that pure vision might be stuck, and lidar might make up for shortcomings.

Xia Yiping: We are gradually moving towards visualization, we have a general direction, we all know that it is impossible to install so many lidars in user-oriented models, because the cost is too high, so we must use vision to solve the cost problem. In fact, you will find that using vision to solve this matter is not a cost problem in itself, and in the end it is essentially a technical problem.

We still had lidar when we started, and the first was 12 cameras plus 2 lidars. At that time, the reason for putting 2 lidars was that I felt that it was the first time to do intelligent driving, and I still wanted to have a little more redundancy to ensure safety. But in fact, the further back you go, you can understand that LiDAR is a bit like a "walker", you can use the walker to walk quickly when you are a toddler, but the walker may also make you dependent, I don't want to go on my own. If you don't have a walker and stick to the wall from the beginning to walk on your own, you're likely to learn to walk quickly.

In January 2023, a decision was made to turn into a vision plan. First, we found that lidar is a very sophisticated device, and in terms of long-term vehicle operation and maintenance costs in terms of long-term vehicle operation and maintenance costs, it is not particularly friendly to the entire intelligent driving or car. At that time, after these thoughts, we decided to simply throw away the crutches and go purely visual and do it in a data-driven way.

Huang Chenxia: At this stage, if the pure visual intelligent driving experience is good enough, the requirements for the computing power behind it, the self-iteration of large models, data quality, and parameters are not low, and the cost is not very cheap from a certain point of view.

Xia Yiping: LiDAR can be understood as the ability to use the end to push the cost to the user, in fact, the cost of pure vision is not low, because it relies on a large amount of data and training, and the training depends on a lot of computing power, and the cost of computing power is not low at all. When we do computing power once, the training cost itself is still very expensive.

In addition, after the emergence of artificial intelligence like ChatGPT in the past two years, including OCC occupancy network and other technologies, the 2D images that we can see now can be well restored to 3D, with the ability to measure and restore 3D reality, and the richness of the image must be greater than the amount of information contained in the image generated by the lidar.

Huang Chenxia: It is more difficult for me to transform into something more informative.

Xia Yiping: This content is far more than the content of lidar, and the cost behind it may not be borne by the consumer, but in fact, the R&D cost behind it is huge. But what are the benefits? We can continue to use AI learning and rapid growth capabilities to make cars continue to iterate rapidly and grow rapidly. You may find that our autonomous driving capabilities have improved every month, and all kinds of intelligent driving will make you feel more and more intelligent and understand the road conditions.

Huang Chenxia: I told a few friends before, for example, people's mobile phones are handed over to people, but people's cars are handed over to products, so this product needs to be strong enough to ensure that people are relatively safe in any scenario, which is very different. Similarly, as an end product, the safety of the car cannot be overemphasized.

Xia Yiping: I think this is also a very important feature of the automotive industry, that is, no matter what kind of car you make, no matter how advanced the safety is, ensuring the safety of users is the first priority. No matter how far you can drive, you must always put safety first.

When building a car now, we must think about the changes brought about by AI and large models

Huang Chenxia: From R&D and design to the launch of a car, in fact, it is just the beginning, and many of the initial ideas need to be verified with product strength, and continuous successful products to go through the cycle.

Xia Yiping: Yes, we had a very deep underlying theory before, and we firmly believed that good technology can bring good experience. First of all, technology must bring excellent experience and help others solve pain points, and at the same time, the moat and product experience barriers brought by this technology must become the core competitiveness of the product.

For example, why does Jiyue use pure vision to solve the problem? Because from the perspective of long-term development, this matter is right. Over time, it will become a moat and a barrier for your product. Then you find that your product will iterate faster and faster on the right path, and we still have to use product power to do marketing, not to use technology as a marketing gimmick to do external marketing in the name and language.

Huang Chenxia: In the long run, to become a friend of time, it must be that the underlying technology and product strength are solid enough. In October last year, I talked to the person in charge of investment in the largest public fund in China about a few tracks, in the smart car track, he is very firm that pure vision represents the future trend, and he has his underlying logic.

Xia Yiping: In fact, the underlying logic is very simple, and you can understand it as the driving force behind it. In the past, we used to build cars to find various solutions from the market, but now we don't think about rebuilding cars like that, when you discuss smart cars, you can't do without AI, large models, and these kinds of things are driving the development of technology.

Huang Chenxia: It is possible to transform everything.

Xia Yiping: It has been transformed, and this year it has been found that AI has penetrated into all walks of life, and many applications have been made to improve efficiency. When making a car now, we must think from the highest level, what kind of changes will the future AI and large models have on the car, and how to fully integrate the car with the entire AI capabilities, which is very important. If you reason based on this first principle, you will come to the conclusion that the person who will do it must be the future.