AI News

Executives from 5 major companies including XPeng, NIO and BYD debate: around 2028 is the key window for L3 and above autonomous driving

YicaiUpdated Nov 16 19:03

AI summary

At the World Intelligent Connected Vehicles Conference, executives from XPeng, NIO, BYD, Momenta and Horizon Robotics judged that around 2028 is the key window for the deployment of L3 and above autonomous driving, with 2026 seen as an industry inflection point. The penetration rate of domestic L2-level combined driving assistance in new passenger cars has exceeded 70%, L3 access pilots are being advanced in batches, and the first mandatory national standard is planned to take effect in July 2027. He Xiaopeng advocates skipping L3, with L4 to be deployed in the second half of 2028, while Cao Xudong said the pace of intelligent driving capability improvement has accelerated to 10 times per year.

Body · AI translation

At the World Intelligent Connected Vehicle Conference, multiple industry figures regarded 2026 as the inflection point for the intelligent driving industry, with competition about to enter its second half.

On one hand, the penetration rate of new passenger cars with L2 combined driving assistance functions in China has exceeded 70%, and urban NOA functions have become accessible to mass-market economy models; on the other hand, L3 market access pilots are being advanced in batches, and the first mandatory national standard has been released and is scheduled to take effect in July 2027. Wan Gang, honorary chairman of the China Association for Science and Technology, said at the conference that high-level autonomous driving technology is accelerating its deployment; as of the first half of 2026, 57,000 kilometers of test roads have been opened nationwide, more than 20,000 test demonstration licenses have been issued, and cumulative test mileage has exceeded 220 million kilometers. He expects that by 2030, the penetration rate of L3 autonomous driving will exceed 35%.

What is the second half testing?

He Xiaopeng, CEO of XPeng Motors, believes that in the first half of autonomous driving, the industry relied on software and algorithms to achieve L2 assisted driving in some areas. The second half will see three key changes.

First, technology will shift from traditional software to AI models and data; hardware will no longer be about partial optimization but about full-chain hardware coupling from chips, chassis, motion control, cockpit, and intelligent driving to the whole vehicle, with software and hardware transforming in sync, in a logic similar to robots. Second, regulations will expand from assisted driving rules in China and some countries to global efforts to promote L2-L4 autonomous driving regulations and clarify deployment scenarios for L4 and L5. Urban RoboTaxi is relatively easy to deploy; whether scenarios such as logistics robots and commercial vehicles can be deployed is an important area of scenario innovation in the second half. Third, three core indicators absent in the first half will be added: safety collision mileage, full-scenario coverage capability (autonomous driving achievable in 9,999 out of 10,000 kilometers of driving scenarios), and user payment rate. This set of quantitative indicators will guide the industry to balance experience, safety, and commercial returns.

William Li, founder of NIO, believes that the core of the second half of autonomous driving is whether it can create value for users. "There are two important indicators: first, users' willingness to pay; second, safety."

In other words, competition in the second half is no longer just about whether functions exist or how high the penetration rate is, but about whether safety can be verified and whether users are willing to pay for it. The rise in safety requirements is directly related to the change in the responsible party. Industry observers believe that as the responsible party for L3 and above autonomous driving shifts to the autonomous driving system, safety requirements for autonomous driving will become higher, meaning L3 systems must have the capability threshold to provide a safety backstop.

Yang Dongsheng, senior vice president of BYD Company Limited, said that going from L2 to L3 is a major threshold to cross. China's L3 regulations are highly forward-looking, with standards far higher than L2, and still require extensive demonstration and exploration. L2 is the starting point of the second half, used for technological exploration. L3 is no longer just assisted driving; the system needs to have driving capabilities surpassing humans. L2 mainly addresses corner cases such as "ghost probe" incidents, reducing accident harm through prediction. L3 faces more stringent challenges: in addition to ghost probes, it must work normally in severe weather and also handle various failure and extreme long-tail scenarios such as tire blowouts and slippery roads. Safety is not a matter of doubling but of hundreds of times the requirement, making it very difficult.

Around 2028 will be a key window for high-level autonomous driving deployment

Regarding the deployment pace of L3 and above autonomous driving, 2027 and 2028 were frequently mentioned.

"True L3 is scenario-limited, while L4 and L5 are not scenario-limited. Technological iteration is also very fast. Scenario-limited L3 has hope in 2028, and non-scenario-limited capability will be ready after another year or two of iteration," Yang Dongsheng said.

He Xiaopeng holds a different view on the technology route; he believes autonomous driving should skip L3. In L3 mode, safety responsibility shifts to the system, but it only shifts responsibility; the user experience is not optimized, and the key point is that safety does not improve. Regulatory systems in various countries, if they want to control autonomous driving and support the technology's evolution to higher levels, must ensure safety and controllability. He said L4 technology will be deployed in the second half of 2028, requiring the development of the entire system's software and hardware together; from a policy perspective, this will occur between 2028 and 2032, advancing in batches across different countries, regions, and scenarios. Robotaxi and fully unmanned driverless operations will be faster, but for driverless operations with a human present, "the timing may require several more years in different places."

"2027 to 2028 will be a key inflection point for L3 and L4, and penetration rates will see substantial growth," said Cao Xudong, CEO of Momenta. According to the company's internal intelligent driving Moore's Law, performance previously improved 10-fold every two years; across the two cycles of 2022-2024 and 2024-2026, a hundredfold improvement was achieved in four years, and intelligent driving gradually moved from being nearly unusable like a toy to usable and good to use. After entering the second half, the capability improvement speed of the "intelligent driving Moore's Law" will further accelerate, upgrading from "10-fold every two years" to 10-fold every year.

On the commercialization level, Cao Xudong believes that subscription models for L3 will become more widespread. When products truly create user value, users' willingness to pay will also increase accordingly. Taking Tesla V13 as an example, the early subscription rate was about 20%, and it has now risen to 55%; progress in safety capability has driven the subscription rate up. In China, Huawei and NIO have already carried out forward-looking exploration of subscription models, completing market cultivation; more companies will participate later, promoting the deployment of this model.

Horizon Robotics CEO Yu Kai took a more cautious stance. He believes that 2026 will become an industry inflection point, with passenger cars in the 100,000 yuan class listing urban NOA as standard, this technology beginning to reach ordinary consumers, and the market continuing to expand. But the autonomous driving revolution has only just begun, and every stage of technological leap is full of challenges. Yu Kai said that the industry's leading urban NOA MPCI (critical interventions per thousand miles) has just surpassed 100 kilometers, with one intervention every 100 kilometers. Users who test-drive for half an hour in urban areas at a 4S dealership without intervention consider the function usable. But "usable" does not equal "easy to use"; performance may be stable over the first 100 kilometers of a test drive, but abnormal conditions may still occur when actual driving reaches 1,000 kilometers. Robotaxi and L4, L5 autonomous driving require an MPCI of more than 100,000 kilometers.

"The journey of the revolution is still very long," Yu Kai said. Breaking through from 100 kilometers to 1,000 kilometers MPCI still leaves technical problems not fully solved; even reaching 1,000 kilometers and advancing to 10,000 kilometers may not necessarily proceed linearly along the existing route. He believes one can remain optimistic, but the technology itself is very difficult. "Every tenfold leap in MPCI is like crossing the Pacific Ocean, and we may need to cross several Pacific Oceans before truly achieving what we hope for," Yu Kai said. Currently only "usable" has been achieved; 1,000-10,000 kilometers MPCI corresponds to "easy to use"; only reaching the 100,000-kilometer level can truly achieve users' "love to use," and he hopes to complete this series of leaps in the next 4-5 years.

Source

Yicai