Ideal Auto's Intelligent Driving Reshuffles Again: The End-to-End Department Becomes Independent to Strengthen Technological Pre-Research | Exclusive from 36Kr
Text｜Angie Li
Editor｜Qin Li, Xuan Yang
After fully promoting the "end-to-end" solution, Li Auto's intelligent driving has once again ushered in a personnel adjustment.
36Kr exclusively learned that the person in charge of Li Auto's "end-to-end" intelligent driving mass production has officially emerged: Xia Zhongpu, with a job level of 21, reporting directly to the intelligent driving head, Lang Xianpeng.
Previously, the intelligent driving algorithm research and development department of Li Auto was overall led by Jia Peng, reporting to Lang Xianpeng. After the organizational adjustment, the intelligent driving team is mainly divided into three groups: Xia Zhongpu is responsible for the "end-to-end" model algorithm and implementation; Jia Peng is responsible for the world model, with more focus on technical pre-research; Wang Jiajia is responsible for mass production research and development. All three departments report directly to Lang Xianpeng.
This is an organizational adjustment made by Li Auto for long-term technical pre-research after achieving phased results in the "end-to-end" solution. "The division of labor in the intelligent driving business is further refined," an informed source said.
Among them, the "behavioral intelligence" that Xia Zhongpu is responsible for has been upgraded from a second-level department to a first-level department in Li Auto's intelligent driving. According to 36Kr Auto, Xia Zhongpu previously worked at Baidu's autonomous driving department, and the technical direction he was responsible for was the integration of the intelligent driving system planning and control links with the AI neural network, which is one of the key challenges for the realization of the "end-to-end" solution.
After joining Li Auto, Xia Zhongpu is responsible for the implementation of the "end-to-end" model. In just two years, his job level has risen from P9 (corresponding to the new job level system of Li Auto at level 19) to level 21, with a two-level promotion.
Li Auto is also further betting on the next-generation end-to-end solution and L4-level autonomous driving.
Previously, Li Auto disclosed in the second-quarter earnings conference call that it has internally initiated a longer-term end-to-end VLA model (Vision-Language-Action Model, that is, the vision-language-action model). According to 36Kr Auto, this solution has a more in-depth integration with the multimodal large model and is regarded by industry insiders as the next-generation end-to-end solution.
Lang Xianpeng, the head of Li Auto's intelligent driving, also told 36Kr in an interview that the company is internally pre-researching L4-level autonomous driving. CEO Li Xiang publicly stated that with the evolution of technology and the enhancement of computing power, unsupervised L4-level autonomous driving will be realized within three years.
The technical pre-research part is led by Jia Peng. According to 36Kr Auto, Zhan Kun, a senior intelligent driving algorithm expert in Jia Peng's team, previously worked in Baidu's autonomous driving team Apollo on the development of Robotaxi (autonomous driving taxis).
After this adjustment, an organizational path with parallel end-to-end intelligent driving mass production and L4 research and development has been formed within Li Auto. A similar parallel strategy of mass production and pre-research was also reflected in Xiaomi Auto previously.
In response to the above information, 36Kr Auto sought verification from Li Auto, but the official declined to comment.
According to the exclusive understanding of 36Kr Auto, Li Auto's end-to-end intelligent driving solution was approximately initiated and promoted in November 2023, and it was originally planned to be mass-produced and installed in vehicles by the end of 2024.
At the beginning of this year, after the senior intelligent driving executives personally went to the United States to see the end-to-end intelligent driving effect of Tesla, they decided to fully promote the end-to-end solution. The team built a technical framework of end-to-end + VLM (Visual Language Model). In April, Li Auto experienced a company-level contraction and adjustment, and the intelligent driving team also shrank from a thousand-person level to several hundred people.
However, due to the end-to-end effect solution exceeding expectations, on October 23 this year, Li Auto has fully pushed the end-to-end intelligent driving solution to users of the intelligent driving Max version.
The "end-to-end" solution was first used by Tesla, integrating the AI neural network into the perception, decision-making, path planning, control and other links of intelligent driving. Driven by massive data and large computing power, it has a higher technical ceiling than the traditional route based on engineers' handwritten rules, and the driving performance is also closer to that of human drivers.
Currently, the "end-to-end" has become the mainstream direction of the evolution of intelligent driving technology in automotive companies, and even the starting point for L4-level autonomous driving and even the broader robotics industry, with stronger possibilities in the commercial value aspect.
In this competition, automotive companies not only launch technical competitions but also turn it into another kind of user experience competition, that is, intelligent driving "from parking space to parking space". This lowers the understanding threshold of intelligent driving for consumers but also tests the polishing of intelligent driving products of automotive companies.
On November 28, Li Auto fully pushed the latest achievement of the end-to-end + VLM dual-system architecture to users of the intelligent driving Max version - from parking space to parking space, becoming the first automotive company in the industry to fully push this function.
The official said that this function can deal with narrow community roads, roundabouts, U-turns, or complex construction scenarios, etc.; at the same time, the latest intelligent driving version of Li Auto also supports independent passage through ETC at national highway toll stations, allowing users to achieve an intelligent driving experience with no breakpoints and full-scene coverage from parking space to parking space.
In the industry, companies such as Huawei, Xiaopeng, Xiaomi, Zeekr, NIO, IM Motors, Changan, and Great Wall have also long been involved in this end-to-end wave.
In August this year, Huawei launched the ADS 3.0 system based on the end-to-end technology on the Enjoyment S9 for the first time, featuring the ability to be directly enabled in the underground parking space, and it can independently enter and exit the gates, pass through roundabouts, and stop and start on the roadside. Huawei plans to gradually push it to all users starting from September 23.
Xiaopeng's end-to-end solution also focuses on the "from parking space to parking space" function. On November 15 this year, Xiaopeng announced that its intelligent driving solution that realizes "from parking space to parking space" through a set of software has started the first test in the whole network.
On the same day, at the Guangzhou Auto Show, Xiaomi Auto also introduced that its intelligent driving system is based on the end-to-end technology that combines the mapless NOA technology, user driving data training, and the VLM visual language large model, which can realize the intelligent driving experience in the full scenes of driving and parking, with the "from parking space to parking space" function. Xiaomi Auto expects that this solution will be pushed to the Pioneer Edition at the end of December.
Zeekr Auto is closely following the progress of the first echelon. Also at the Guangzhou Auto Show in November, Zeekr said that its mapless NZP (Urban Navigation Pilot Assist) will be pushed to users nationwide by the end of the year, and the "from parking space to parking space" function will be pushed in batches in January next year. Changan Automobile has also established a thousand-person intelligent driving team to create an end-to-end intelligent driving solution.
IM Motors, which adopts the solution of the supplier Momenta, announced on October 17 that the mapless urban NOA function has been fully enabled nationwide; Great Wall Automobile, which also cooperates with the supplier DeepRoute, also announced the national opening of the full-scene NOA on November 15.
NIO's end-to-end technical route was publicly disclosed in August this year. According to 36Kr Auto, at this year's NIO Day in December, NIO may release more mass production information about the end-to-end solution.
Undoubtedly, the "end-to-end" intelligent driving solution has pushed the intelligent driving competition of automotive companies to a climax. However, most automotive companies are facing the balance problem between the maintenance of the mass-produced intelligent driving software and the development of the new end-to-end solution.
But the phased achievements of Li Auto prove that under the premise of reducing the burden of past mass production, a clear technical idea can be quickly verified with the support of massive data and algorithms. This may be a reference idea for traditional automotive companies to continue to be involved in the "end-to-end" wave.