XPENG Releases World Model Technical Report Backing VLA 2.0 Autonomous Driving
XPENG has released a technical report for X-World, a generative world model that creates real-time multi-view simulations for autonomous driving, now integrated into VLA 2.0 development and validation.
- 30 million kilometers
- Over 500,000
What Happened
Traditional simulation testing techniques based on 3D Gaussian Splatting can reproduce real-world scenes but fail when the autonomous driving model deviates from original trajectories, such as sharp lane changes. This has forced the industry to rely on expensive real-world road testing with limited scenario coverage. XPENG's X-World, a generative world model built on video diffusion technology, addresses this by generating future multi-camera video streams from historical video and driving actions, enabling controllable, real-time simulation.
500,000+scenarios
up from 30,000 a year ago; daily simulated test mileage equivalent to 30 million km
- Strong cross-view consistency across seven surround-view cameras
- Strict action following: generated scenes match specified ego-vehicle behavior
- Long-horizon video simulation for stable predictions over extended time spans
X-World supports XPENG's VLA 2.0 in closed-loop evaluation, online reinforcement learning, and large-scale data generation. It enables evaluating safety metrics like collision rate and ride comfort in a virtual environment, and can generate missing long-tail scenario data for corner cases.
Why this matters
X-World reduces the need for expensive real-world testing by generating realistic driving simulations, helping XPENG develop safer autonomous driving more efficiently.
Terms in This Story
- World model
- A generative AI model that can simulate and predict future states of a driving environment.
- VLA
- Vision-Language-Action model, an AI system for autonomous driving that processes visual data, language instructions, and driving actions.
- DiT
- Diffusion Transformer, a neural network architecture used for generating high-quality video sequences.
Summarised from the linked release; details can be imperfect — always verify against the original source.