Xpeng Releases Autonomous Driving World Model X-World: Domestic Data Can Be Zero-Sample Migrated to Overseas

Xpeng Motors' GWM team published a paper on arXiv proposing the autonomous driving world model X-World. Based on historical footage from 7 surround-view cameras and future action commands, this model can generate multi-view future driving videos, finely control scene elements like surrounding vehicles, pedestrians, and lane markings, and adjust weather and time periods via text prompts.

X-World uses a streaming autoregressive approach to generate frames frame by frame, supports real-time interaction, and can be used for closed-loop evaluation of end-to-end autonomous driving and online reinforcement learning training without relying on high-cost real-road tests. Its "zero-sample style transfer" capability can convert domestic driving data into overseas training resources, saving the cost of large-scale local data collection. Currently, it is used as a reinforcement learning training environment and generative data factory, capable of generating long-tail scenarios such as extreme weather and rare vehicle models.

Xpeng previously released the second-generation VLA model, achieved its first quarterly profit in Q4 2025 (net profit of 380 million yuan), and plans to increase its physical AI R&D budget to 7 billion yuan in 2026.