news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Zhu Xichan said that the performance of the large model is amazing, but it is estimated that it is unlikely to be installed on vehicles in the next two years, because the high computing power required by the large model is difficult to deploy on the vehicle side, and the large model also cannot solve the long-tail problem of safety.

Reporter |Zhou Xin

Image source |Tuchong Creative

Tesla launched the V12 version of FSD (Full-Self Driving) in March this year. This version deleted hundreds of thousands of lines of manual rule codes and adopted an "end-to-end" autonomous driving solution.

End-to-end technology has become the hottest topic at the moment, and domestic car companies have also begun to roll out "end-to-end". Xiaopeng Motors announced at the "520 AI DAY" press conference that its first large-scale end-to-end model in China has been mass-produced. Ideal Auto set up a separate team for end-to-end models at the end of last year. Li Xiang, CEO of Ideal Auto, also said in early June that Ideal will push end-to-end + VLM (visual language model) intelligent driving solutions to test users in the third quarter of this year.

SAIC Motor also recently stated that pure vision and end-to-end technology have entered the development stage for mass production. So far, almost all leading domestic autonomous driving companies have released their own "end-to-end" solutions and vehicle deployment plans.

On August 16, at the 4th Shenyang Intelligent Connected Vehicle Challenge, Zhu Xichan, a professor at the School of Automotive Engineering of Tongji University, told Economic Observer: "Tesla's end-to-end and artificial intelligence make autonomous driving more possible. Tesla has made end-to-end popular. Domestic car companies are all talking about end-to-end. But now, if anyone announces end-to-end (mass production), don't buy his car."

A week ago, at the launch of the Enjoy S9, Yu Chengdong, chairman of Huawei's terminal and intelligent vehicle solutions BU, commented on Tesla's FSD, saying that the commercial version has made great progress, with a high upper limit but a low lower limit. "We went to test it, and we ran over a stationary white truck on the road without slowing down. We also ran over a green truck without slowing down."

The so-called end-to-end autonomous driving generally means that after the control system reads the data from the original sensor, it directly calculates the control instructions only through the neural network, which does not contain any artificially designed rule modules.

Mainstream autonomous driving systems will adopt a modular approach, dividing the AD system into perception, planning and control. They will first accurately perceive the surrounding dynamic and static traffic participants and road network structure, then plan the vehicle's driving trajectory, and finally perform closed-loop control of the vehicle through actuators.

It is understood that Tesla FSD's end-to-end large model eliminates the sections between the perception and positioning, decision-making and planning, and control and execution of the autonomous driving system, and combines the three modules together to form a large neural network.

"The end-to-end model removes all the criterion models. After the criterion models are removed, the generalization ability of AI is stronger," said Zhu Xichan. Then the various AI modules are linked together by transformers, and information is transmitted implicitly. This will reduce the loss of information during transmission and optimize the entire network. This is why Yu Chengdong said that the upper limit of Tesla's FSD can be increased.

"But what Yu Chengdong said about the 'lower limit is also very low' means that after the standard model is thrown away, it is difficult to know whether AI has learned the lower limit of safety when training it," Zhu Xican further stated that Tesla FSD V12 version, especially version 12.4, does have a very high upper limit and strong generalization ability. It can handle more things, making driving more like an experienced driver, but there are still many problems.

In April, a Tesla owner in the United States hit and killed a motorcyclist. The 56-year-old driver admitted that he had turned on Autopilot at the time but was distracted while driving.

Zhu Xichan said that Tesla's FSD V12 version was pushed to the United States in March, but the NTSB (National Transportation Safety Board) is still investigating, and it is not clear whether it is V11 or V12. If the user who caused the accident updated, then the version used was V12, which is a big problem in terms of safety.

He believes that China has achieved end-to-end perception or segmented end-to-end, but compared with Tesla, domestic companies still have a big gap in data volume and AI training computing power. "Xpeng Motors' AI computing power is relatively large in China, but it is still dozens of times lower than Tesla."

In terms of data collection, at present, car companies rely on user vehicles to collect data. The greater the vehicle sales, the richer the data. Tesla has sold more than 1.7 million vehicles in the United States, completing the closed loop from data to users. "Whoever breaks 1 million vehicles for smart driving and artificial intelligence data collection first will be the first to break through," said Zhu Xichan. Huawei Hongmeng Zhixing and Ideal Auto may reach 1 million vehicles by next year.

It is worth mentioning that at the recent 4th Yanzhi Automobile Annual Meeting, Zhu Xican said that the performance of the large model is very amazing, but it is estimated that it is unlikely to be installed on the car in the next two years, because the high computing power required by the large model is difficult to deploy on the car side, and the large model also cannot solve the long-tail problem of safety.

Zhu Xican believes that based on the data capabilities and AI training computing power currently developed by domestic automakers in intelligent driving, it is a more reasonable technical route to develop end-to-end perception solutions that integrate lidar perception solutions and retain safety criteria models in regulatory control algorithms. This segmented end-to-end solution model is a relatively reasonable technical route.

The END

WeChat ID |eeojjgcw

Sina Weibo |@Economic Observer