DreamStream: Towards Policy-Oriented Generative Simulation for End-to-End Driving

CoRL 2026 · Paper · Project Page · Code

DreamStream is a generative closed-loop simulator for end-to-end driving policies. A physics simulator executes the policy's plan and renders traffic layouts; a few-step autoregressive video model turns those layouts into camera frames that the policy consumes in the next step.

This repository hosts the checkpoints. Code, setup and usage instructions are on GitHub.

Contents

Path	Description
video_model/wan2.2-5b-nuplan-train-v1/ema_model.pt	DreamStream video generation model (Wan2.2-Fun-5B-Control backbone, 1280×704), trained on nuPlan
policy/	Front-camera end-to-end driving policies (DiffusionDrive, DiffusionDriveV2, DrivoR, LTF, SparseDriveV2) used for closed-loop evaluation in BridgeSim and for the FDπ metric
policy/aux/	Auxiliary files required by the policies (trajectory k-means anchors)

The base model (alibaba-pai/Wan2.2-Fun-5B-Control) is not redistributed here; the inference server downloads it automatically.

Usage

The video model checkpoint is downloaded automatically by the inference server; see the GitHub READMEs for where the policy checkpoints go. To fetch everything manually:

hf download UCLA-VAIL/DreamStream --local-dir DreamStream-ckpts

See the GitHub repository for how to run the video model server, closed-loop evaluation in BridgeSim, and FDπ.

License

The checkpoints are released for non-commercial research use only. See the GitHub repository for code licenses and third-party notices.

Citation

@inproceedings{leng2026dreamstream,
  title     = {{DreamStream: Towards Policy-Oriented Generative Simulation for End-to-End Driving}},
  author    = {Leng, Ziyang and Mo, Sicheng and Zhao, Seth Z. and Cai, Haoyuan and Zeng, Yu and McAllister, Rowan and Zhou, Bolei},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}