StressDream: Steering Video World Models
for Robust Policy Evaluation and Improvement

Junwon Seo, Sushant Veer, Ran Tian, Wenhao Ding, Apoorva Sharma, Karen Leung, Edward Schmerling, Marco Pavone, Andrea Bajcsy

---

StressDream is an inference-time method that optimizes diffusion noise to steer video world models toward plausible, high-impact outcomes - enabling robust evaluation and improvement of robotic policies.

---

Modules

Module	Domain	World Model	Reward
dubins/	Synthetic Dubins car	Image-based 3D Dubins Car	Safety score (failure set)
vista/	Driving	Vista	X-CLIP (+ optional Qwen2.5-VL)
ctrl_world/	Robot manipulation (DROID)	Ctrl-World	Qwen3-VL

Run StressDream

All scripts are run from the StressDream/ repo root. Each module has its own environment and detailed README.

🚗 Dubins Car

Start with dubins/demo.ipynb

See dubins/README.md for details.

conda create -n stressdream-dubins python=3.10 && conda activate stressdream-dubins
pip install -r dubins/requirements.txt
jupyter notebook dubins/demo.ipynb

CLI steering is also available:

python dubins/run_steering.py

🛣️ Vista Driving

Start with vista/demo.ipynb

See vista/README.md for details.

conda env create -f vista/environment.yml
conda activate stressdream-vista
jupyter notebook vista/demo.ipynb

CLI steering is also available with python vista/run_steering.py.

🦾 Ctrl-World Manipulation

See ctrl_world/README.md for details.

conda env create -f ctrl_world/environment.yml
conda activate stressdream-ctrlworld
python ctrl_world/run_steering.py \
    --hdf5_path ctrl_world/example_data/traj_0001.hdf5

❤️ Acknowledgements

StressDream builds on two open-source video world models:

• Vista — Vista: A Generalizable Driving World Model with High Fidelity and Versatile Controllability (OpenDriveLab/Vista)
• Ctrl-World — Ctrl-World: A Controllable Generative World Model for Robot Manipulation (ctrl-world.github.io), built on Stable Video Diffusion

⭐ Citation

If StressDream helps your research, please consider citing:

@article{seo2026stressdream,
  title={StressDream: Steering Video World Models for Robust Policy Evaluation and Improvement},
  author={Seo, Junwon and Veer, Sushant and Tian, Ran and Ding, Wenhao and Sharma, Apoorva and Leung, Karen and Schmerling, Edward and Pavone, Marco and Bajcsy, Andrea},
  year={2026}
}