He Xiaopeng abandons L3.
During the Two Sessions, as a deputy to the National People's Congress, He Xiaopeng's suggestion this year is to accelerate the policy and regulatory system to achieve the leap from L2 to L4, simplify the intermediate L3 process, and create more policy space for technological innovation and business exploration.
XPeng Motors has indeed put this into practice.
On March 2nd, He Xiaopeng launched the second - generation VLA, claiming it to be a large physical - world model that "opens the era of L4", aiming directly at the L4 end - game.
This means that He Xiaopeng has officially abandoned L3.
XPeng Motors has clearly skipped L3.
On the afternoon of March 2nd, this new car - making force that has always stood firmly at the forefront of intelligent driving officially released the result of months of hard work in the past year - the second - generation VLA.
He Xiaopeng, the chairman and CEO of XPeng Motors, announced that this solution will start the test - drive experience on March 11th and be fully pushed in April.
The second - generation VLA is a base model developed by XPeng with billions of investment last year. Judging from the information released by the official, this new technological paradigm has brought an unprecedented upgrade of intelligent driving experience to XPeng: it is not only better than XPeng's previous XNGP but also far ahead of the industry.
He Xiaopeng's exact words were: "Through our internal comparison and evaluation, I think it is nearly five times better than the top players in the industry, whether in terms of takeover rate, smoothness, or the supported reachable range."
However, XPeng did not make the second - generation VLA, which has a "qualitative leap" in effect after heavy investment, into an L3 solution.
According to XPeng Motors, the second - generation VLA is a large physical - world model that "opens the era of L4", aiming directly at the L4 end - game.
As a deputy to the National People's Congress, one of He Xiaopeng's suggestions during this year's Two Sessions is also to accelerate the policy and regulatory system to achieve the leap from L2 to L4, simplify the intermediate L3 process, and create more policy space for technological innovation and business exploration.
XPeng is one of the few car companies in the automotive industry that clearly aims to "skip L3 and directly enter L4".
L3 is a level in the classic core standards of autonomous driving and is largely regarded as a necessary stage for the iteration of autonomous driving technology and market promotion.
With the development of the domestic intelligent driving industry, it is generally expected that this year will be a turning point for the transition from L2+ to L3. Leading car companies and industrial chain companies such as Huawei, Geely, and GAC have successively announced their L3 mass - production plans.
Even at the policy level, it has started to give the green light. At the beginning of the year, two models were announced to have obtained the access permission for L3 - level products from the Ministry of Industry and Information Technology.
In the first year of L3 mass - production, why did XPeng Motors make a seemingly "anti - consensus" choice?
Upgrade to L3
First, let's explain the concept of L3.
L0 - L5 is a globally recognized autonomous driving classification system formulated by the Society of Automotive Engineers (SAE). It has been referenced and adopted by regulatory agencies such as the US Department of Transportation and the Chinese Ministry of Industry and Information Technology and is a widely recognized industry standard.
Put simply, L2 is pure assisted driving. Although the system can drive, the driver needs to continuously supervise. L3 is "autonomous driving within a limited range". Within the Operational Design Domain (ODD), the system can fully execute driving tasks.
The most fundamental difference is that the subject of responsibility has shifted - the driver is the subject of responsibility for L2, while within the ODD, the system becomes the subject of responsibility for L3. Therefore, L3 is regarded to some extent as a qualitative change node in autonomous driving and a prelude to full - fledged L4 autonomous driving.
The European car company Mercedes - Benz was the first to launch an L3 system in the industry. As early as 2021, Mercedes - Benz obtained the access permission for the L3 system in Germany, and in May of the following year, some of its models could be optionally equipped with the Drive Pilot system.
At that time, Tesla and many domestic car companies were also actively promoting various intelligent driving functions. However, no matter how cool the function demonstrations were, in essence, they were still L2 systems where the driver was responsible. Mercedes - Benz, which was the first to launch the L3 system, was once regarded as a representative of traditional car companies that were responsible and accountable.
In the past few years, the level of high - end intelligent driving in China has been in the L2+ echelon.
Some aggressive intelligent driving models presented the "parking space to parking space" usage scenario when they were launched, which to some extent could be regarded as the prototype of the ODD. However, no car company has made the "parking space to parking space" into a real L3 product or service.
The reason behind this is that neither the technological breakthrough nor the product experience has reached that "critical point".
In the past two years, intelligent driving technology has been iterating rapidly. Especially with the application of AI, the performance of high - end intelligent driving has been significantly improved and is being installed in more and more models. The industry has a faster expectation for the implementation of autonomous driving, and L3 has been put on the agenda.
In a small - scale exchange last year, Zhu Xichan, a professor at Tongji University and an industry expert, said that the main task of intelligent driving in 2026 is to upgrade from L2+ to L3 and complete the gradual process from "human - machine co - driving" to "machine - centric and human - retreating".
In April last year, Huawei's Inroad launched the HUAWEI ADS 4 high - speed L3 commercial solution, showing its technological reserves and commercial plans for L3. Yu Chengdong, the person in charge of HarmonyOS Auto, also cooperated by saying that "Huawei will enter the L3 and L3+ era in 2026", and drivers will be able to sleep in the car.
At the beginning of this year, the Ministry of Industry and Information Technology announced the first batch of two L3 - level conditional autonomous driving vehicle product access permissions. One of them is the ArcFox Alpha S (L3 version) equipped with Huawei's Qiankun Intelligent Driving ADS.
In addition to Huawei, car companies such as Changan Automobile, Geely Automobile, GAC Group, and Chery Automobile have also released their layouts for L3 to varying degrees.
Take GAC as an example. Its high - end intelligent brand Hyper took the first national test license for L3 highway autonomous driving at a speed of 120 km/h at the end of last year. The first model of its new brand Qijing, which cooperates with Huawei's Inroad, is also clearly equipped with an L3 architecture.
AI Bridges the Gap
The generally recognized standard system in the industry implies a technological logic: L4 will evolve from L3.
Currently, while intelligent driving solutions for passenger cars are blooming everywhere, a typical application scenario of L4 - Robotaxi - largely operates within the ODD and achieves true autonomous driving by gradually expanding the operating range.
This is the ancestral experience accumulated in the rule - based era, using a sufficient number of fleets, a large amount of data, and a large number of image captures to expand the ODD range.
However, with the application of AI and the change of technological paradigms, this logic is being shaken.
Why skip L3? He Xiaopeng's explanation is: "In today's global technological development, basically the next step after L2 is L4. Adding an L3 in the middle poses challenges to hardware, software, and laws and regulations."
That is to say, the change of technological paradigms is starting to bridge the gap between L2 and L4 - while the L2 assisted driving function is iterating and improving, it will infinitely approach and even reach the true L4 function. L3, which is stuck in the middle, is neither sufficient nor necessary.
The second - generation VLA launched by XPeng Motors this time, according to He Xiaopeng's introduction, can achieve a comprehensive upgrade in three experience dimensions: peace of mind, smoothness, full - scenario capabilities, and high efficiency.
Looking at the specific indicators, it can reduce 99% of heavy braking, 98% of sudden acceleration, 60% of safety takeovers, increase the recognition ability of road obstacles by 124%, increase the recognition ability of vehicles on the side and rear by 118%, and improve the accuracy of night - time decision - making by 96%.
Although this system has not fully reached the 100% L4 level, it has shown quite good anthropomorphism and generalization. He Xiaopeng did not disclose the latest MPI (average takeover mileage), but he said it was in the "thousands of kilometers" range.
According to Liu Xianming, the person in charge of XPeng's General Intelligence Center, the second - generation VLA has built a very general and efficient architecture. New versions are iterated every day, and the progress speed is beyond imagination. So, there is confidence to reach the L4 level in the future.
He Xiaopeng believes that this process only takes 1 - 3 years.
Behind this is the reconstruction of a whole set of systems including models, computing power, data, and ontology.
Especially the model. The second - generation VLA, a base model developed by XPeng with billions of investment, is a native multi - modal architecture that can encode all signals in a more efficient and original way and has created a highly real - time thinking chain technology to achieve efficient reasoning.
In addition, XPeng also uses the world model for simulation training and then uses reinforcement learning to further iterate the model through the mechanism of positive and negative feedback.
It is reported that since its debut in November last year, the second - generation VLA has been updated 468 versions in 3 months, with an average of nearly 4 versions per day.
As the proportion of data - driven in algorithm iteration increases, L4 is expected to come faster than expected. Liu Xianming believes that building a good base model is a compulsory course for a company aiming at L4.
"There are many interesting cases of Waymo in San Francisco. If we still follow the past path, everyone may think it is a technological route with extremely high costs and poor generalization. So, Waymo is also doing things like building a base model." Liu Xianming said. However, for an enterprise, it also needs to see whether it has the determination to do so and how quickly it can complete the transformation.
In addition to XPeng and Tesla, which is regarded as an industry model, Horizon, which mass - produced HSD in November last year, also said that the new technological architecture is expected to unify the methodologies of L2 and L4. Su Qing, the vice - president and chief architect of Horizon, said at the Horizon Technology Ecosystem Conference last year:
"After the arrival of the new paradigm, using the current method and working for another 2 - 3 years at most, there is a great chance to increase the MPI to 50,000 or even 100,000 kilometers, while still maintaining human - like performance and being able to generalize automatically in all regions."
Small - scale Pilot in the First Year of L3
As the technology and application of autonomous driving have developed to date, the industry has generally formed two major camps:
Car companies and suppliers represented by Huawei regard L3 as a necessary path, believing that liability guarantee will be the core technical point of intelligent driving competition in the next stage;
Technological radicals represented by Tesla and XPeng tend to believe that "there is no L3 stage" because the AI - driven algorithm has fundamentally revolutionized the evolution path of autonomous driving.
In fact, for both camps, what is provided to users at the current stage is still essentially an assisted driving function that requires the driver to be responsible. Even L3 is a limited - function autonomous driving under specific conditions.
In the early stage of technological development, L3 was destined to be a "semi - finished product" with a greatly discounted experience. The L3 system of Mercedes - Benz has very strict usage restrictions, making it seem rather useless - for example, the road section must be a highway, the natural conditions must be good weather, and at the beginning, the speed could not even exceed 60 km/h...
The mass - production of L3 systems by domestic car companies that started at the end of last year is only at the beginning:
Only a few models have obtained licenses from the Ministry of Industry and Information Technology and test access permissions from local governments. However, there is no clear information about the specific ODD settings and related supporting measures such as insurance and regulations.
A real - world problem is that the commercialization of L3 faces a paradox - in order to reach the L3 level, the hardware and software configurations must reach nearly the L4 standard, but commercially, it can only operate under the limited scenarios and responsibility rules of L3.
It is worth mentioning that Mercedes - Benz, which was the first to conduct L3 pilots, has temporarily abandoned L3. At the beginning of this year, Mercedes - Benz announced that some models will no longer offer the L3 option and will be uniformly replaced with a new - generation L2 - level system.
Many car companies and intelligent driving suppliers have chosen to implement L3 despite the difficulties. The main consideration is market promotion. To quote an industry insider: "The functions of L3, such as regulatory adaptation, responsibility definition, insurance system, and user education, are the gaps that must be filled before the implementation of L4."
For the pioneers, liability guarantee is also a safety net for intelligent driving functions to move from niche to mass adoption. If clear responsibility attribution can be achieved, it will have the opportunity to significantly improve consumer trust and sales, forming a positive cycle.
For XPeng Motors, which has directly skipped L3, how to dispel consumers' doubts to a greater extent? The only way is to widely promote the latest solution and spare no effort in market promotion.
In terms of product implementation rhythm, the second - generation VLA will start to be pushed at the end of March, covering models with the Ultra/Ultra SE versions, and will cover more product lines in April.
In mid - March, XPeng will conduct test - drive experiences of the second - generation VLA intelligent driving solution in first - tier stores. He Xiaopeng said confidently, "Welcome the whole network to conduct comparative evaluations."
In order to have an objective and quantifiable system for intelligent driving experience, XPeng has also specially developed a testing software called "Samo". It can generate an objective comfort score for each trip by monitoring the vehicle's movement status in real - time.
In the past one or two years, the fierce competition in the industry has not substantially improved the experience of intelligent driving systems, and the intelligent driving "silver bullet" has not exerted the influence expected by the industry.
However, with the implementation of solutions under the new technological paradigm, the establishment of the industry system, and the improvement of market recognition, intelligent driving is still expected to be the decisive factor in the fierce competition in the automotive market.
This article is from the WeChat official account "Yiou Auto". The author is Peng Suping, and the editor is Hao Qiuhui. It is published by 36Kr with authorization.