ForeSight

See Tomorrow, Act Today:
Foresight-Driven Autonomous Driving

Bozhou Zhang^1,2,*, Nan Song^1,2,*, Yuang Wang¹, Jiankang Deng³, Xiatian Zhu⁴, Li Zhang^1,2,†

¹ Fudan University   ·   ² Shanghai Innovation Institute
³ Imperial College London   ·   ⁴ University of Surrey

_{* Equal contribution   ·   † Corresponding author}

Imagine the future. Plan the next move.

---

Overview

ForeSight studies planning through imagined futures. A foundation world model supplies future visual representations; a lightweight current-scene encoder adds present context. WM-QFormer compresses the future features, and a state-based decoder combines both sources to predict driving trajectories.

_{Architecture overview from Figure 2 of the paper.}

Quick start

See the NAVSIM guide for environment setup, data layout, training and evaluation commands. Supports loading the original Epona NuPlan checkpoint.

conda create -n foresight python=3.10 -y
conda activate foresight
bash scripts/setup_env.sh
bash scripts/test.sh

Citation

If you find this work useful, please cite:

@inproceedings{zhang2026foresight,
  title     = {See Tomorrow, Act Today: Foresight-Driven Autonomous Driving},
  author    = {Zhang, Bozhou and Song, Nan and Wang, Yuang and Deng, Jiankang and Zhu, Xiatian and Zhang, Li},
  booktitle = {CVPR Findings},
  year      = {2026}
}

Acknowledgements

We thank the authors and contributors of NAVSIM, DiffusionDrive and Epona for their open-source work.

License

See LICENSE.md for the repository license. Third-party components retain their respective licenses.