On the evening of March 16, He Xiaopeng made a rare disclosure of the company's financial commitment. The second-generation VLA from XPeng, defined by He as "the first version built for L4 capabilities," operates on a core logic of solving problems entirely with AI rather than predefined rules. This development comes amid calls from delegates during the Two Sessions to accelerate revisions to the Road Traffic Safety Law.

To sufficiently power the second-generation VLA with computing resources and data, XPeng has been investing 300 million yuan per month, a practice sustained for over a year. Following the merger of the autonomous driving and smart cockpit business lines, the General Intelligence Center, led by Liu Xianming, has a clear mandate: to convert this investment into user trust. Indeed, showrooms have seen a noticeable increase in test drive interest specifically focused on the autonomous driving features.

However, just as momentum builds, a viral video showing four children lying on a road has thrust the capabilities and limitations of autonomous driving into the spotlight. In the incident, a XPeng vehicle with its autonomous driving system engaged detected the anomaly, decelerated, and attempted to avoid a collision, but ultimately required human intervention to come to a complete stop.

The focus of the autonomous driving race is shifting. It is no longer about overwhelming competitors with technical specifications, but about clarifying for the public where the safety boundaries of the technology lie and when human drivers need to take control.

During a live stream, the recent viral video was discussed. The data retrieved and interpreted by Liu Xianming confirmed that the vehicle did identify the obstacle and initiated braking. However, he pointed out that the rate of deceleration was insufficient for a complete stop without driver intervention. He Xiaopeng added that the driver was an XPeng employee testing the system, who took over upon noticing the unusual deceleration.

The system identified the problem, but the final decision rested with a human. This illustrates the current reality of autonomous driving: it can provide perception and warnings, but extreme scenarios still require human judgment. Different technical approaches exist—some emphasize algorithmic generalization, while others focus on sensor redundancy—yet all face the challenge of handling rare, edge-case scenarios.

Building public willingness to use the technology hinges on establishing trust. The aforementioned case is precisely part of that trust-building process. During the直播, He Xiaopeng reiterated the concept of "autonomous driving for the people," defining it as "autonomous driving that even your mother would feel comfortable using," emphasizing that the systems must return to the fundamentals of safety and ease of use. He noted that acceptance rates improved significantly after non-technical employees were invited to experience the technology.

The data is more direct: the second-generation VLA has reduced instances of hard braking by 99% and sharp acceleration by 98%. The significant reduction in passenger discomfort from such abrupt maneuvers marks a critical step from the technology being merely "usable" to being "desirable to use."

Liberated from reliance on high-definition maps, the second-generation VLA can cover more unstructured roads. However, Liu Xianming candidly admitted that the current version occasionally does not follow navigation instructions perfectly. He explained this as a necessary phase in the transition from rule-driven to reasoning-driven systems—the AI is learning to understand its environment and make independent decisions, but requires users to allow for adaptation time in certain scenarios.

He Xiaopeng also drew a clear line: scenarios like extreme weather or situations with no discernible path, which even challenge human drivers, are not recommended for autonomous driving use. This aligns with an industry consensus that these systems are aids, not replacements.

The competition in autonomous driving is evolving from a contest of parameters to a contest of trust. The past was about the breadth of functional coverage; the present is about which company can make users feel confident and willing to actually use the technology.

"Spending 300 million yuan monthly on this bet, continuously for over a year, even I felt incredibly anxious at times," He Xiaopeng unusually admitted during the live stream. Where is this "bet" placed? Liu Xianming's answer points to full-stack in-house development, spanning from chips and compilers to software architecture and data闭环. "You primarily have to dare to place the bet," he added.

Why make such a bold bet? He Xiaopeng presented a highly contentious viewpoint—that China's autonomous driving development should leap directly from L2 to L4. He argued that lingering at the L3 stage could lead to falling behind in the global competition. The core of this leapfrog logic lies in the division of responsibility. Liu Xianming explained that L4 requires the system to solve all problems definitively, without passing difficult decisions to the user. He Xiaopeng revealed that the second-generation VLA is indeed the first version built with L4 capabilities in mind, operating on the core principle of using AI, not rules, to solve problems.

The market is providing positive feedback. He Xiaopeng disclosed that since March 11, when XPeng opened test drives for the second-generation VLA across its 732 national stores, the test drive rate has doubled, with many users visiting specifically to experience the system. A more direct return is visible in the order structure: sales proportion of the Ultra trim models equipped with this system have increased significantly.

Regarding the rollout schedule, a gradual rollout began on March 19, prioritizing the XPeng P7 Ultra, followed by the G7 and X9 Ultra. Users of these three models will receive the update within the month. Rollout for more models will commence in April.

Addressing user concerns about version differences, He Xiaopeng provided a clear distinction: the Ultra version is developed for L4 capabilities, supporting full-scenario navigation. The Max version primarily covers high-frequency scenarios like highways and major urban roads.

Beyond immediate commercial returns, XPeng's sights are set globally. The company's autonomous driving team has set a goal for 2025 to compete head-to-head with Tesla. With recent media comparisons between the second-generation VLA and Tesla's FSD V13, He Xiaopeng offered his assessment.

"Looking at V13, we hold a relatively clear advantage. But I believe this is because XPeng is in China, our data is from China, and we are more familiar with Chinese road conditions," he emphasized. He stated that the second-generation VLA handles "vehicle-pedestrian interactions" more effectively, such as dealing with delivery workers, pedestrians, and narrow lanes. "This is not just a Chinese characteristic; Europe has many small roads, Southeast Asia too. As we enter more countries, XPeng might hold an even greater advantage."

Liu Xianming was more cautious: "Actually, we don't know how Tesla does it. It's more like crossing the river by feeling the stones—we've stumbled into many pitfalls and wasted a lot of money. But we believe the ultimate solutions might converge, despite different paths."

He Xiaopeng believes that China and the US are both in the top tier of autonomous driving development, but he contends that the complexity of Chinese roads is ten times that of American roads—encompassing not just highways and city streets but also rural paths outside third- and fourth-tier cities where one might encounter livestock.

"Autonomous driving is a comprehensive contest of hardware, software, engineering capability, and scalability. Currently, China and the US are in the first tier. But China's roads are more complex. Only by tackling these difficult challenges first to enhance the generalization ability of our AI models can the second-generation VLA truly achieve global implementation," He Xiaopeng said.

Any technology must undergo user validation. The industry's current assessment is that the inflection point for autonomous driving has arrived. The true measure of success, however, will be whether users who have experienced it are willing to recommend it to others.