InfGen: Scenario Generation as Next-Token-Group Prediction

Zhenghao Peng , Yuxin Liu , Bolei Zhou

University of California, Los Angeles

Code | Paper
Image

InfGen is a unified transformer that treats the entire driving scene—map, lights, agent states and motions—as one long token stream and rolls it out autoregressively.

    :traffic_light: Generates both initial states, traffic light signals, and traffic participant trajectories in a single model.

    :mag: Supports dynamic agent injection for open-world, long-horizon simulation.

    :rocket: Autonomous driving planners trained in InfGen scenarios show ↑ robustness & generalization over log-replay baselines.

@article{peng2025infgen,
  title={InfGen: Scenario Generation as Next Token Group Prediction},
  author={Peng, Zhenghao and Liu, Yuxin and Zhou, Bolei},
  journal={arXiv preprint arXiv:2506.23316},
  year={2025}
}