XPENG was invited to CVPR for the third time in June 2026. Dr Xianming Liu, head of the company’s centre for general intelligence, spoke at the first workshop on deploying foundation models for embodied AI, together with Tesla, Nvidia and Waymo and researchers from the University of California and the University of Toronto. The talk was titled Building the World Model for Autonomous Driving, and an XPENG paper on generating driving scenes was accepted at the conference.
The division of labour
XPENG describes it this way: VLA 2.0 learns from how people drive, while the world model learns the laws of physics by predicting what happens next. VLA 2.0 teaches the model how to act, and the world model gives it an understanding of how the surroundings develop after each action.
The world model’s three abilities are called reasoning, controlled generation and prediction over long sequences. XPENG’s researchers have published three papers: X-World, which produces physically plausible future video from given motion inputs and is used in closed-loop simulation and to generate training data; X-Foresight, which is built into VLA 2.0 and predicts both future images from several angles and the car’s own actions; and X-Cache, which according to XPENG cuts about 70 percent of the redundant computation and makes the engine up to about 2.7 times faster. A fourth, X-Mind, has been announced but not released.
Training data and model performance
VLA 2.0 is in formal mass production, and XPENG states, without bounding fleet, market or method of measurement, that in its first month after the OTA rollout the system accounted for more than 50 percent of the kilometres driven with the assistance system switched on. The company calls the figure a new benchmark for the Chinese industry.
The model has billions of parameters, is trained on hundreds of millions of video clips and uses more than four trillion tokens per model iteration. Without giving the absolute starting points, XPENG states that in the twelve months to March 2026 training efficiency rose 1,010 percent per GPU and 4,360 percent for a single job, while hardware utilisation went from 40 to 90 percent.
The robotaxi, built on the GX platform, has come off the line with 3,000 TOPS on board.
The date for the robot
IRON is, according to the release, to enter the phase where hardware and software are put together, with formal mass production by the end of 2026. The robots are then to start working as shopping guides in XPENG’s own stores from the first quarter of 2027.
The release gives no timetable for VLA 2.0, the robotaxi or IRON in Europe. Neither did the announcement of Master Agent and Turing AI in August.
Images: XPENG
Sources
XPENG: the world model presented at CVPR in Denver · 3 June 2026