DriveWAM: Video Generative Priors Enable Scalable World-Action Modeling for Autonomous Driving

Paper | Project Page

DriveWAM: Video Generative Priors Enable Scalable World-Action Modeling for Autonomous Driving
Chen Shi*, Jinrui Xu*, Shaoshuai Shi, Kehua Sheng, Bo Zhang, Li Jiang†
The Chinese University of Hong Kong, Shenzhen & Voyager Research, Didi Chuxing
*Equal Contribution, †Corresponding Author

Highlights

• Unified Video-Action Policy: Adapts a pretrained video diffusion transformer (Wan2.2-TI2V-5B) into an end-to-end driving policy via joint flow-matching over video and action tokens.
• Scene-Evolving Driving Guidance: A frozen VLM (Qwen3-VL-8B) generates chunk-specific semantic intent injected via temporally localized cross-attention.
• Selective KV Memory: Training-free modality-aware cache selection achieves 12x memory reduction for 300s rollouts with minimal accuracy loss.
• Strong Performance: 90.1 PDMS on NAVSIM v1 (single front-view camera, simple regression head) and 0.83m ADE@4s on PhysicalAI-Autonomous-Vehicles benchmark.

News

• [2025/27] Code will be released soon. Stay tuned!

TODO

• Release inference code and pretrained checkpoints
• Release data preparation scripts
• Release training code

Citation

If you find this work useful, please consider citing:

@article{shi2025drivewam,
  title   = {DriveWAM: Video Generative Priors Enable Scalable World-Action Modeling for Autonomous Driving},
  author  = {Shi, Chen and Xu, Jinrui and Shi, Shaoshuai and Sheng, Kehua and Zhang, Bo and Jiang, Li},
  journal = {arXiv preprint arXiv:2605.28544},
  year    = {2025}
}

Acknowledgements

DriveWAM is built upon LinBotVA, Wan2.2, and Qwen3-VL. We thank the authors for their great work.