XPeng Unveils In-House AI at CVPR; Gen 2 VLA ADAS Now in Vehicles

On June 5, at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) held in the United States, XPeng Motors participated as the only Chinese automaker alongside companies like Tesla and NVIDIA in technical exchanges. The company revealed at last year’s CVPR that it was developing its own foundational large model; by March this year, it had already deployed its second-generation Vision-Language-Action (VLA) advanced driver-assistance system—based on this model—on mass-produced vehicles. According to data released by XPeng, in vehicles equipped with this system, user-enabled intelligent driving mileage has exceeded 50% of total mileage, making it the industry’s first ADAS to achieve such a usage rate. Many users have already formed the habit of using this feature daily. At this year’s conference, Liu Xianming, XPeng’s head of technology, unveiled for the first time the technical roadmap of the “XPeng World Model.” This model features capabilities in active reasoning, controllable generation, and long-horizon prediction, forming—alongside the second-generation VLA system—the foundation of XPeng’s intelligent driving technology. Specifically, the VLA system learns standard driving behaviors from real-world driving data, while the World Model predicts traffic condition changes and simulates potential risks to enhance driving safety. He Xiaopeng, founder of XPeng Motors, stated that the second-generation VLA has demonstrated the potential of physical AI and believes that the scaling laws governing AI development also apply to the field of autonomous driving. The company plans to continuously iterate both technological systems to further improve the safety and performance of future intelligent driving versions.