news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

After Baidu's driverless taxi "Robot Run" became popular, autonomous driving seems to have really come to ordinary people. However, in the eyes of most practitioners, the current L4 Robotaxi or smart electric car's intelligent driving assistance system is still a long way from being easy to use and large-scale commercialization.

"At present, we see that most of the current urban NOA are still in the usable stage and have not yet reached the stage of being easy to use. The main problem is still the low traffic efficiency and non-human behavior. Automakers are rapidly increasing the number of cities covered, resulting in a decrease in system availability and a low pass rate. A truly easy-to-use Intelligent Driving 2.0 system must provide an anthropomorphic intelligent driving experience." Horizon President Chen Liming said at the recently held 2024 China Automotive Forum that in 2025, autonomous driving will usher in the ChatGPT moment.

Starting from the beginning of 2023, Tesla claimed to have adopted a fully end-to-end large-model V12 version of FSD, which triggered widespread discussions in the industry; subsequently, car companies such as Xiaopeng and Weilai, as well as supplier companies such as Horizon and Yuanrong Qixing announced one after another that they would deploy end-to-end intelligent driving systems.

Recently, Chentao Capital jointly released the "End-to-End Autonomous Driving Industry Research Report" (hereinafter referred to as the "Report") with three parties. The Report shows that among the more than 30 front-line experts in the autonomous driving industry interviewed, 90% said that the companies they work for have invested in the research and development of end-to-end technology, and most technology companies believe that they cannot afford to miss this technological revolution.

Yu Qian, co-founder and CEO of QINGZHOU Zhihang, believes that recent end-to-end technology has created a clearer trend in the evolution of intelligent driving technology.

Public information shows that end-to-end is a concept in deep learning, referring to an AI model that can output the final result as long as the raw data is input. A typical example is the popular ChatGPT. The application of end-to-end technology in autonomous driving has transformed the original architecture of a combination of multiple models such as perception, prediction, and planning into a single model architecture of "perception and decision-making integration." In layman's terms, the past autonomous driving route is like multiple people driving a car, while end-to-end technology is a single person driving, which is closer to real human driving.

"The end-to-end large model is based on a probabilistic model training. One problem is that for relatively simple and easy-to-describe scenarios, its output is often not so accurate and the bottom line is relatively low. Tesla has done a pretty good job in this regard, but has not completely solved this problem. We believe that in the current lack of sufficient data, it is still necessary to gradually achieve end-to-end, one module at a time, replacing each other, completing the end-to-end while providing a safety net. With this relatively solid engineering infrastructure and rapid iteration method, we can step by step improve the performance upper limit of the system, while also ensuring the lower limit of the system performance." Chen Liming said.

Yu Qiankun, CTO of Sai Ke Intelligent, told reporters that the intelligent driving application of the end-to-end model can be divided into two stages: the first stage is the two-model solution, which is currently the mainstream direction used in the industry. However, the two-model solution will inevitably suffer from some information loss due to the explicit output in the middle, and it is difficult to fully utilize the information of the sensor; the second stage is the one-model solution, which is a one-step solution. Currently, many people are doing preliminary research, which is also closer to the direction of AGI, but this direction is more difficult and is estimated to be 3-5 years later before some large-scale applications can be obtained. It should be pointed out that many car companies have not completely abandoned the traditional "rule control" after announcing the "end-to-end" vehicle.

An autonomous driving algorithm engineer from a certain car company said that because the output of the neural network has a certain probability, it cannot be guaranteed that the output is absolutely safe. Therefore, after the end-to-end large model is installed on the car, the rule method will still be necessary. At present, most end-to-end intelligent driving systems still use some rule methods to perform secondary verification on the output of the neural network.

At present, the industry generally believes that the research and development progress gap between domestic car companies and Tesla is about 1.5 to 2 years. Gu Junli, deputy general manager of Chery Automobile Co., Ltd., believes that in order to catch up with Tesla in terms of business model, it is necessary to form a large-scale product. "When the data reaches more than one million Tesla-level, through intensive training of the model, the intelligent driving can learn the video stream and directly tell the driver the driving direction, just like the popular ChatGPT." Gu Junli said.

Compared with Tesla, which is almost the only company in overseas markets, China's intelligent driving systems are in a stage of flourishing development. It is difficult to connect data between different models and different technical solutions, which brings certain difficulties to the data-driven end-to-end large model.

Chen Liming said: "The difficulty we are facing now is that the architecture of many models and sensors, as well as the layout and adoption of sensors, are constantly changing. Although we have collected a lot of data, this data is not accumulated in a high-quality manner and can be used continuously. This is a problem that we will explore not only for a certain company, but for the entire industry. In other words, how can OEMs and technology companies work together to solve this problem? This is something we all need to discuss together."

Yu Qiankun also believes that the current end-to-end engineering applications face problems such as poor completeness of data collection, poor reusability of data collection, and low training computing power.