Beijing traffic is notorious for its chaotic nature, where lane markings often seem irrelevant, and aggressive driving is the norm. In such an environment, any moment of hesitation can lead to three cars filling in the space intended for one. This makes for an intense testing ground for electric vehicles claiming to offer autonomous driving capabilities.
Last week, I traveled to Beijing to put the new XPeng VLA 2.0 to the test amidst this frenetic traffic. The findings indicate that the self-driving landscape is evolving, with Tesla no longer standing as the sole serious contender.
XPeng began rolling out the VLA 2.0 software via over-the-air updates in March 2026, specifically for its "Ultra" variants of the XPeng P7, XPeng G7, and XPeng X9. The name VLA, which stands for Vision-Language-Action, signifies a transformative leap in how these vehicles perceive and respond to their environment.
Prior iterations of the system resembled a disjointed group of individuals; one component monitored the road, another plotted the route, and yet another controlled steering. This separation sometimes resulted in delays or abrupt maneuvers. In contrast, VLA 2.0 employs a single, integrated system that interprets camera inputs and makes swift decisions, streamlining the driving experience significantly.
To achieve this, XPeng developed a powerful custom chip called Turing AI, boasting 2,250 TOPS of computational capability. The system underwent training using 100 million video clips showcasing challenging driving scenarios, enabling it to function more like a seasoned human driver rather than a rigid machine.
During a captivating 40-minute test drive through downtown Beijing, the system demonstrated remarkable capabilities, operating without any human intervention. In a city with near-constant interruptions from erratic drivers, the vehicle navigated bustling intersections efficiently, resembling a leisurely drive through a park.
This impressive performance wasn't an isolated incident; I conducted two additional drives totaling over four hours. One route remained within Beijing's busy streets, while the other ventured off into rural areas. The vehicle faced narrow, unmarked roads with unpredictable drivers, yet there was never a need to intervene or hit the brakes. Such reliability is exceedingly rare among contemporary autonomous systems.
A particularly striking moment during the drive involved merging into a narrow gap in fast-moving traffic. Standard driver-assist systems typically struggle in such situations, hesitating until a large opening appears or failing altogether and prompting human intervention. The XPeng P7 Ultra, however, confidently assessed the small gap, calculated the speed of adjacent vehicles, and smoothly moved in decisively, mirroring the behavior of an experienced local driver.
Another critical test occurred when the car emerged from a narrow rural road onto a busy five-lane street, contending with traffic from multiple angles and a swarm of scooters. The XPeng P7 edged forward, adeptly positioning itself in the innermost lane, then accelerating seamlessly into a gap to merge with ongoing traffic.
As we left Beijing, the vehicle prepared for a blind turn, slowing down early and positioning itself closer to the road's edge to better assess the situation. This maneuver is characteristic of an adept human driver seeking to maximize visibility and maintain safety. Once the road was clear, the car surged forward, providing ample space for the scooters and completing the overtaking in mere seconds.
This experience felt extraordinary, akin to observing a veteran driver navigate the streets with poise—no hesitation, full situational awareness, and generous accommodations for other road users. It was truly impressive.
Another notable instance was when the car intelligently adjusted its position on unmarked narrow roads with uneven edges. Instead of strictly sticking to the right, it veered away from potential bumps and only returned to the correct side when approaching oncoming traffic, demonstrating advanced situational judgment to avoid hazards.
This impressive human-like awareness stems from a generative world model—a virtual reality environment enabling XPeng to rigorously test its software without risk on actual roads. The company has scaled its testing efforts from 30,000 scenarios annually to 500,000; daily, the system simulates driving the equivalent of 30 million kilometers in this digital realm, facilitating error correction without real-world consequences. While Tesla largely relies on real-world data, XPeng embraces extensive simulations to shape its systems.
The rivalry among automotive firms has reached a new level of intensity. Recently, XPeng CEO He Xiaopeng tested Tesla's FSD v14.2 in Silicon Valley, spending five hours navigating San Francisco, where he surprisingly praised the system's performance.
Setting high expectations, He challenged his own team to ensure the VLA 2.0 achieved quality levels comparable to Tesla's finest by August 30, 2026. He humorously claimed that if they fail, Liu Xianming, his head of autonomous driving, must run across the Golden Gate Bridge without clothing. While likely jest, this illustrates the formidable pressure faced by companies in the race for supremacy.
Tesla currently faces challenges in China, with local drivers using outdated software, while North America users enjoy the latest innovations. Slow government approvals for advanced self-driving features have hindered Tesla, providing local brands like XPeng an advantage.
Other manufacturers, including BYD with its "God's Eye" system, are investing heavily in similarly sophisticated technology. Additionally, tech behemoth Huawei is pouring over €9 billion into its own autonomous driving software, and Xiaomi has joined the fray with its new SU7 series. The market is rapidly becoming saturated, with technology evolving faster than many can adapt.
On the commercial front, Tesla charges a monthly fee of €85 for its full self-driving capabilities in the U.S., while XPeng currently offers its VLA 2.0 as part of its vehicle purchase price. This pricing strategy creates challenges for Tesla to substantiate its subscription model when competitors provide similar features at no additional cost. Volkswagen has recognized XPeng's promise, recently signing a deal to incorporate VLA 2.0 into its forthcoming electric SUV tailored for the Chinese market, a significant endorsement from a well-established automaker.
The VLA 2.0 is particularly noteworthy for its ability to anticipate future maneuvers, distinguishing it from many systems that react solely to immediate obstacles. Unlike its competitors, it assesses traffic flow and prepares for turns or merges in advance, effectively eliminating the uncertainty often associated with driver-assist technology.
Many vehicles appear to be making educated guesses about their next moves, but the VLA 2.0 exudes a sense of purpose. Even when minor errors occur, such as following too closely or hesitating at a turn, it adeptly corrects them quickly and safely.
The automotive industry stands at a pivotal moment. Tesla once dominated the market for "smart" cars, but now, models like the XPeng P7, XPeng G7, and XPeng X9 are demonstrating that other manufacturers can create equally advanced systems. The anticipated XPeng GX could elevate the bar even higher, with potential Level 4 autonomy on the horizon. While definitive comparisons between XPeng and Tesla require further evaluations, one thing is clear: the VLA 2.0 system feels like a highly competent driving partner in the chaotic Beijing traffic. If it can excel here, it should manage similar challenges globally, with perhaps the exception of India's unique conditions. The gap between leading brands is narrowing, making the competition for the future of autonomous driving increasingly fierce.
