World Engine: Towards the Era of Post-Training for Autonomous Driving

Tianyu Li    Li Chen    Caojun Wang    Haochen Liu    Kashyap Chitta    Zhenjie Yang    Yuhang Lu    Naisheng Ye    Yihang Qiu    Yufei Wang    Luoxi Zou    Jiaxin Peng    Jin Pan    Zhaoyu Su    Andrei Bursuc    Shengbo Eben Li    Andreas Geiger    Peng Su    Hongyang Li

preprint 2026

Abstract

Autonomous vehicles must operate safely in the real world, where errors can have severe consequences. Although modern end-to-end driving policies excel in routine scenarios, their reliability is limited by the scarcity of safety-critical "long-tail" events in real driving datasets. These rare interactions define the practical safety boundary of the learned policy, yet they are difficult to collect at scale in the real world. Here we show that this fundamental limitation can be addressed by post-training pre-trained driving models on synthesized high-stakes interactions. We introduce World Engine, a generative framework that reconstructs high-fidelity interactive environments from real-world logs and systematically extrapolates them into realistic safety-critical variations. This paradigm enables reinforcement-based post-training to align policies with safety constraints, circumventing the physical risks inherent in real-world exploration.

BibTeX

@article{li2026worldengine,
  title  = {World Engine: Towards the Era of Post-Training for Autonomous Driving},
  author = {Li, Tianyu and Chen, Li and Wang, Caojun and Liu, Haochen and Chitta, Kashyap and Yang, Zhenjie and Lu, Yuhang and Ye, Naisheng and Qiu, Yihang and Wang, Yufei and Zou, Luoxi and Peng, Jiaxin and Pan, Jin and Su, Zhaoyu and Bursuc, Andrei and Li, Shengbo Eben and Geiger, Andreas and Su, Peng and Li, Hongyang},
  journal = {arXiv preprint arXiv:2606.19836},
  year   = {2026}
}