Recently, Jiyue 07 was officially unveiled at the Beijing Auto Show. Intelligence and AI have always been the labels of Jiyue Automobile. At the press conference, Jiyue announced the progress of the current opening of the intelligent driving system, and also announced that by the end of 2024, Baidu The smart driving system can be used wherever the map can navigate.
We also interviewed Mr. Pan Yunpeng, Vice President of Intelligent Software of Jiyue Auto, to answer questions about intelligent driving and intelligent cockpits.
Pan Yunpeng, Vice President of Jiyue Automotive Intelligent Software
Media question: Is there any gap between the AEB performance of pure visual solutions and LiDAR? Or is there no gap?
Pan Yunpeng: During our testing of AEB, we found that lidar has little impact on braking distance and time. It mostly increases the upper limit of AEB in some turning scenes, including rapid scenes. We are also constantly increasing the upper limit of AEB.
Media question: Two years ago can be called the first year of lidar. In the past two years, high-definition mapping has been implemented and purely visual routes have been published. Do you think that apart from these three trends, the future development of autonomous driving will be What is the method, which route should be taken, or both?
Pan Yunpeng: The views in this industry are becoming more and more similar. There are several key words, one is de-graphics, one is lidar, and one is end-to-end. Everyone talks more about it, and it may be in a long time. There will be some time.
There are more and more pure visual players this year, but lidar will also exist. I believe they will coexist for a long time. Each company’s data accumulation is different, and each company’s ecology is also different. , I believe it must be multiple plans. The same scenario does not necessarily mean who is better than the other. For example, if you talk about pure vision solutions, everyone will talk about cost and reliability when talking about lidar. Our pure vision reduces after-sales costs, but the research and development costs will increase, and the amount of data will increase. The requirements for computing power are getting bigger and bigger, and we try our best to expand the entire computing power to 2.21.
OverallIt must be the most appropriate route chosen by each company based on its own characteristics.
Second, everyone has experienced many versions without pictures or versions with fine pictures, and they can all find a problem, that is, where there is more experience and information, the smart driving experience will be better and more secure, especially at some intersections. Scenarios, some intersections are particularly complex, and problems may arise if there is no experience information at all. We still feel that the availability of experience information has a very direct impact on the user experience at this stage, so many of the picture-less solutions we are talking about are essentially hybrid solutions. Some places have no pictures, and some places have pictures, and pictures refer to them. It's a high-precision map.
Third, each company’s ecology is different. Baidu has lane-level navigation. Everyone should have used Baidu’s lane-level navigation. For us, this is a first-mover advantage. Although we also have no map, generally speaking, we naturally have more information in the map than ordinary maps, so our smart driving experience and sense of security are better. In last week’s smart driving competition, Jiyue 01 won the first place in the Beijing Station. The contestants at that time included Jihu, Wenjie, Xiaopeng and Weilai. We scored 95 and ranked 1st. The second place is 93 points.
As for the issue of having pictures and not having pictures, it must be a step-by-step process. Even if a non-picture solution is developed, it is impossible to remove all the existing solutions with pictures. The entire industry has a unanimous judgment on this point. .
Finally, end-to-end. It is divided into two types, but generally speaking, the basic logic behind it must have a large amount of high-quality data. It is still a bit early to realize end-to-end, and it still requires the accumulation of a large amount of high-quality data. Everyone is still in the process of exploration, and of course this is also the focus of Jieyue’s next step.
Media question: Regarding the voice issue, you said that it can identify objects, but now there are many wake-up-free voice devices. After talking to the car machine, if you talk to the passenger next door, it will also misrecognize it. We can Can you identify the object of the conversation?
Pan Yunpeng: Solving this problem is still quite difficult. It requires a better understanding of semantics and context, and also requires the accumulation of data. For example, two people are chatting and Simo starts talking. The user says Simo to step back or stop. By understanding the semantics and analyzing the scene, these scenes can be gradually reduced. Generally speaking, it is also a data closed-loop and data-driven idea, which is difficult to kill with a stick through rules.
Media question: First, as a user, many companies now have smart cars, and they are all in urban areas. How should we judge whether the smart car is easy to use or not? How does an ordinary person judge the criteria for judging good or bad? Second, in the future, when it can be driven across the country, I will no longer need to drive. If you don’t pay great attention to the driving quality and are not a driving enthusiast, the requirements for the chassis of the car will not be so high. From a development perspective,In other words, does a good chassis help improve smart driving? Third, some companies say they can open it nationwide, so they all open it. Why do we open it like this when we have 50 seats per 100 seats?
Pan Yunpeng: Let me answer the second question first. Everyone's needs are different. Some people care about driving quality, while others don't. A good chassis must be helpful to a smart driving experience, including its braking response and speed-up response, which are 100% helpful. If a chassis responds quickly and includes a softer calibration, it will definitely be different in the scenarios of smart driving and human driving. This is relatively easy to judge.
Not only the chassis of Zhijia, including motors and electric drives, has a direct response to customers, but from a safety perspective, the braking rate, braking response time delay, and the higher the speed, the greater the impact every second. .
The first question and the third question are put together, why do people say that it can be opened without pictures and can be opened nationwide, but we are still taking it step by step. This is related to our safety judgment standards. First, we think it is safe. Opening it for a short period of time does not help the user experience. Second, if you open it and are forced to take over every day, it will not be good for your reputation, especially for newer brands. At this time, Frighten users several times, and if unfortunately several accidents occur, the damage to the brand will be greater. Our standards for driving are still very strict, so you can rest assured.
We mainly serve users. We are not a marketing-oriented company. We still selected the cities with the largest number of users. We will conduct relatively complete tests in these cities. Because of Baidu’s blessing, our The map has a lot of experience and information. We will pick out the places where the map is not very accurate, and gradually open the city based on the standards of ensuring safety and security. The overall logic of opening the city is this.
One hundred cities, two hundred cities, and three hundred cities are essentially a process of data iteration. With more data, on the basis that 95% of scenarios can be covered very well, it is possible to It will also take 500% of energy to cover the remaining 5% scenes, which will take longer. Kaicheng also follows the logic of this standard and satisfies According to the rules of engineering research and development, data is gradually iterated.
Media question: I have experienced some NOA products before, from our friend. They may be OK in scene processing, but their starting and parking are very easy to make people get motion sickness. We have to consider the comfort level and the control logic. How was it handled?
Pan Yunpeng: This is a very good engineering problem. We have also encountered it. Many users have many problems with subjective body feeling. Some users feel that the brakes are too sharp and they nod their heads, which is uncomfortable. Some users feel that the start is too slow. , lost the opportunity, and was very angry. There was also the user. We found that he wanted to follow closer in the morning, but braked too sharply at night and felt unsafe. This is a veryA difficult question to answer.
First of all, the problem of nodding can definitely be solved, that is, whether the brakes should be soft, including the logic and method of automatic parking. For other problems, we still learn our own behavior through data algorithms. This is the only way to overcome many subjective scenarios.
Media question: We are opening 100 cities very quickly. I would like to ask what the standards are for opening 100 cities, such as coverage rate?
Pan Yunpeng: We have previously opened major urban roads in Beijing, Shanghai, Hangzhou, Shenzhen, and Guangzhou. Probably 95% of the covered roads have been covered. This standard has not changed because Baidu Maps will Through high-frequency driving areas, these areas will be opened. The size of each city is different. Thousands of kilometers of pure ground road coverage is necessary. This is a big city.
Media question: At what level is the computing power utilization of dual ORin now? Yesterday, it was officially announced that Simo will be used in the future, including subsequent planning and control. It may be done more with models, and then it will be switched to end-to-end. The entire Occupancy is still relatively high.
Pan Yunpeng: Simo’s entire computing power will be upgraded. Now it can only be said to be sufficient. It is a more scientific way to use these two chips. As mentioned just now, we merge many small models. Let it have better performance and lower computing power requirements.
In addition, we will also have more end-to-end algorithms. From which end to which end, each car company is not exactly the same, and the requirements for computing power may become higher and higher.
Editor in charge: Shi Tianqi