Xpeng Unveils VLA2.0 and World Model at CVPR 2026

Xpeng showcased advancements in its Physical AI technology at CVPR 2026, marking its third appearance at the conference. The company announced in a press release that its VLA2.0 foundation model has entered mass production after being first revealed at CVPR 2025.

Dr. Xianming Liu, Head of the General Intelligence Center at Xpeng, presented the technical evolution of VLA2.0, which achieved over 50 percent assisted driving mileage share within its first month of deployment. The model completes a full cycle from research to commercial application in autonomous driving.

Xpeng also introduced its world model, designed to enable deliberative reasoning, controllable generation, and long-horizon forecasting. The company explained that VLA2.0 learns driving behavior from humans, while the world model predicts environmental changes after each action, together forming Xpeng’s Physical World Foundation Model.

The company reported a 1,010 percent increase in per-GPU training efficiency and a 4,360 percent gain in single-job training efficiency over the past year. Xpeng plans to expand deployment of its Physical AI applications, including the VLA2.0 system, its Robotaxi program, and the IRON humanoid robot, with mass production of the robot expected by the end of 2026.