End-to-end intelligent driving ignites Zhongguancun! The first Autonomous Driving Summit in 2025 successfully concluded, with Professor Li Shengbo from Tsinghua University and the person in charge of Ideal Intelligent Driving providing analysis and judgment.
The First Autonomous Driving Summit in 2025 Ignites Beijing!
On January 14, the 4th Global Autonomous Driving Summit, jointly initiated by Zhixingxing under Zhiyi Technology and Che Dongxi, successfully concluded at the Conference Center of the Exhibition and Trading Center of the Zhongguancun National Independent Innovation Demonstration Zone in Beijing.
As a conference IP created by Zhiyi Technology for the autonomous driving field, the 4th Global Autonomous Driving Summit, with the theme of "New Technological Cycle, New Industrial Journey", is set up with three sections: "Main Venue + Sub-venues + Exhibition Area".
Among them, the main venue holds the opening ceremony, the End-to-End Autonomous Driving Innovation Forum, and the Urban NOA Special Forum; the sub-venues hold the Autonomous Driving Visual Language Model Technology Seminar and the Autonomous Driving World Model Technology Seminar, presenting the scientific research achievements, technological exploration, product solution innovation and future trends in the new cycle of end-to-end autonomous driving in the era of全民智驾 in an all-round way.
Nearly 30 academic leaders, industry experts and young scholars in the field of autonomous driving gathered at the event to jointly discuss the hottest topics in the industry, such as end-to-end, world model, visual language model, and urban NOA.
▲Main Venue Site
There were continuous brilliant viewpoints from the guests at the scene. Li Shengbo, a professor and doctoral supervisor at the School of Vehicle and Mobility / School of Artificial Intelligence, Tsinghua University, believes that the vehicle-road-cloud integration is the development framework of the autonomous driving base model, and it requires the government, enterprises, and universities to form a consortium to solve problems such as data and computing power.
Lang Xianpeng, Vice President of Autonomous Driving R & D of Ideal Automobile, said that Ideal Automobile will be committed to "the automobilization of artificial intelligence", improving the autonomous driving ability from the aspects of model, engineering, and product.
Wang Panqu, Intelligent Driving Partner of Zero One Automobile, Liu Minjun, Co-founder and CTO of Shengqi Technology, and Li Zhanbin, Deputy General Manager of Langge Technology, all believe that end-to-end has brought about an innovation in the R & D model.
Nearly 900 audiences came to the scene throughout the day of the summit. Whether it was the morning and afternoon sessions of the main venue or the two closed-door seminars of the sub-venues, all seats were occupied. Especially at the opening ceremony, many audiences stood in the aisles on both sides of the audience seats and in the back of the venue. At the same time, more than 20 media platforms and video accounts live-streamed this summit, with the live-streaming viewership exceeding one million.
▲Sub-venue - Autonomous Driving Visual Language Model Technology Seminar
▲Sub-venue - Autonomous Driving World Model Technology Seminar
As the first autonomous driving summit in China in 2025, this conference not only systematically presented the development trend and mass production progress of intelligent driving in the past year, but also pointed out the forward direction for the new round of changes initiated by end-to-end autonomous driving, and also ignited the research and development enthusiasm of the world model in the domestic autonomous driving field.
1. The Commercialization of Intelligent Driving Enters an Important Node, and End-to-End + Large Model Usher in New Industry Changes
At the summit, Gong Lunchang, Co-founder and CEO of Zhiyi Technology, delivered the opening speech as the representative of the organizer. He said that in the past year, end-to-end has collaborated with large language models and visual language models to become the main technical route of autonomous driving. At the same time, the world model has also received more attention.
▲Gong Lunchang, Co-founder and CEO of Zhiyi Technology
Based on this, the summit hopes to sort out the current development status of the industry and further clarify the future development trend through full discussions.
This summit has set up multiple topics around the directions of end-to-end autonomous driving, urban NOA, autonomous driving visual language model technology, and autonomous driving world model technology.
Digitalization and intelligence are becoming important driving forces for a new round of high-quality development in China. Since its establishment, Zhiyi Technology has been focusing on the core technologies and industry needs behind this driving force, and has built Che Dongxi, Xin Dongxi and Che Dongxi three media brands, continuously providing professional and high-quality graphic, text and video content, and has a wide influence in related fields.
At the same time, Zhiyi Technology has built the Zhixingxing brand around enterprise services, and provides high-quality technical content in the form of online public classes and seminars.
Finally, Gong Lunchang thanked the Zhongguancun Science City Management Committee for its strong support for this summit.
After Gong Lunchang's speech, Li Shengbo, a professor and doctoral supervisor at the School of Vehicle and Mobility / School of Artificial Intelligence, Tsinghua University, took the stage first to share the development history, key technologies and future trends of "Data-Driven End-to-End Autonomous Driving".
▲Li Shengbo, Professor and Doctoral Supervisor at the School of Vehicle and Mobility / School of Artificial Intelligence, Tsinghua University
Then, Li Shengbo pointed out that in a narrow sense, "autonomous driving" mainly refers to a high-level intelligent driving system for complex urban traffic conditions.
This type of system has extremely high safety requirements for perception, decision-making, and control technologies. Because the control right, risk monitoring, and failure response are all handled by the system itself, the system must be able to independently complete all driving tasks.
Li Shengbo said that insufficient safety is the core problem for the practical application of the existing autonomous driving system. A typical indicator is the number of takeovers per 10,000 kilometers during driving, which is far from reaching the average level of human drivers. The core problem lies in how to deal with edge driving scenarios (that is, scenarios with a small number of occurrences but high risk).
In order to solve this problem, the consensus in the industry is to take the technical route of "end-to-end" autonomous driving based on the data closed loop and using the neural network as the strategy carrier.
The purpose is to continuously optimize the driving strategy of the autonomous driving system through the collection, upload, cleaning, training and deployment of edge scene data, so as to achieve the effect of learning while driving and improving driving performance.
Li Shengbo pointed out that the essence of end-to-end is "neural networkization", rather than only a neural network with a black box.
He introduced the technical advantages of end-to-end autonomous driving. Compared with the traditional modular design, the end-to-end design solution can more effectively transfer information, reduce information loss, and fully tap the potential of data resources on the one hand; on the other hand, it has more neural network parameters, greater training freedom, and a higher upper limit of performance.
Li Shengbo mentioned that the technical research of end-to-end autonomous driving by Chinese scientific research institutions is not later than that of foreign countries. The intelligent vehicle team of the Vehicle College of Tsinghua University has been exploring this technical path since 2018.
He pointed out that when this technology was launched six years ago, the conditions were not as good as they are today. Insufficient data and lack of computing power are challenging problems that restrict the training performance of the model.
The team was also the first in the industry to propose the R & D idea of "Compensating for Insufficient Data with Simulation, and Overcoming Insufficient Computing Power with Algorithm", and has made a series of important progress in many aspects such as simulation software and AI trainers, including the development of the large-scale autonomous driving training software LasVSim with independent intellectual property rights and the first optimal strategy reinforcement learning solver GOPS for industrial control.
Li Shengbo shared the research results of Tsinghua University in reinforcement learning and neural network training, especially the improvements in stability and efficiency.
The team proposed the algorithm DSAC (Distributional Soft Actor-Critic) with the top performance in the field of reinforcement learning, which continuously distributes the overfitting state-action value, equivalently learning infinitely many value functions, effectively suppressing the overestimation problem caused by the traditional reinforcement learning that only fits a single value function, and the performance is significantly improved compared with the existing reinforcement learning algorithms; it developed the neural network optimizer RAD (Relativistic Industrialization Adaptive Gradient Descent) with the top performance in the field of industrial control, which models the optimization process of neural network parameters as the evolution process of the state of a multi-particle relativistic system, ensuring the training stability and convergence of reinforcement learning from a dynamic perspective.
Based on this, Tsinghua University successfully developed the first three-stage end-to-end autonomous driving system iDrive in China, and took the lead in completing the open road test in urban conditions. This technical solution has been successfully deployed in real vehicles in enterprises such as GAC, Dongfeng, and Zhixingzhe, which can realize behaviors such as giving way to non-motor vehicles and bypassing roadside parking in complex urban road conditions such as congestion and intersections without signals.
Finally, Li Shengbo looked forward to the future development direction, pointing out that the vehicle-road-cloud integration is the development framework of the autonomous driving base model, and it is urgent to establish a unified data platform to promote data sharing. At the same time, he emphasized the importance of computing power for autonomous driving and called on the government, enterprises, and universities to form a consortium to jointly solve these bottleneck problems that restrict the development of autonomous driving.
Professor Li Shengbo shared his thoughts on autonomous driving from an academic perspective, while Lang Xianpeng, Vice President of Autonomous Driving R & D of Ideal Automobile, from an industrial perspective, delivered a speech with the theme of "Ideal Autonomous Driving Technology Innovation and Application".
▲Lang Xianpeng, Vice President of Autonomous Driving R & D of Ideal Automobile
Lang Xianpeng introduced the R & D progress of Ideal Autonomous Driving in the past year.
Ideal Automobile innovatively proposed the dual-system architecture of end-to-end autonomous driving + VLM, and combined with the training and evaluation system based on the world model, it took the lead in completing the full-volume push from parking space to parking space. And the improvement of the performance of the autonomous driving model also conforms to the Scaling Law. The model based on 10 million clips training is about to be pushed to users, and the full-scene takeover rate (MPI) will exceed 100 kilometers.
He said that the total mileage base of the current autonomous driving model training exceeds 3 billion kilometers, and the cloud computing power exceeds 8E Flops. In the next stage, the focus will be on improving the resource utilization efficiency through innovative technical means.
Lang Xianpeng also pointed out that from end-to-end + VLM to VLA, it is another advancement of AI ability from "behavioral intelligence" to "spatial intelligence". In the follow-up, Ideal Automobile will focus on "the automobilization of artificial intelligence", and improve the comprehensive ability of autonomous driving from the three dimensions of model, engineering, and product.
In order to achieve L3, the full-scene takeover rate (MPI) should exceed 500km, which is equivalent to a takeover once every two weeks, the average accident mileage (MPA) should exceed 3.5 million kilometers, which is about 5 times the safe mileage of humans, and the AD mileage penetration rate should be greater than 25% to achieve a leap.
To achieve L4, the capabilities in these three aspects need to be further improved to a higher dimension.
In order to achieve L4, the full-scene takeover rate (MPI) should exceed 2500km, which is equivalent to a takeover once every quarter, the average accident mileage (MPA) should exceed 6.6 million kilometers, which is about 10 times the safe mileage of humans, and the AD mileage penetration rate should be greater than 60%.
Lang Xianpeng emphasized at the end that the mission and vision of Ideal is to become a global leading artificial intelligence enterprise, and continue to carry out the innovation and application of autonomous driving technology and products in this direction. Please wait and see.
Some representatives of the supply chain industry also shared some thoughts on autonomous driving.