Simulation Distillation: Pretraining World Models in Simulation for Rapid Real-World Adaptation

Jacob Levy^*1, Tyler Westenbroek^*2, Kevin Huang², Fernando Palafox¹, Patrick Yin², Shayegan Omidshafiei³, Dong-Ki Kim³, Abhishek Gupta^†2, David Fridovich-Keil^†1
1 UT Austin  ² UW  ³ FieldAI  ^* Equal Contribution  ^† Equal Advising

RSS 2026 (Accepted)

This project implements Simulation Distillation (SimDist), a scalable framework the distills structural priors from a simulator into a latent world model and enables rapid real-world adaptation via online planning and supervised dynamics finetuning.

📝 TODO

• Quadruped code release
• Manipulation code release (coming soon)

📚 Usage

• 💾 Installation
• 🧠 Pretraining (Go2)
  • Expert Policy Training
  • Data Generation
  • World Model Pretraining
  • Deployment (Simulation)
• 🤖 Deployment (Real-World Go2)
  • Hardware Setup
  • Simulation Setup (Optional)
  • Startup
  • Running the Robot
  • Shutdown
• ⚙️ Adaptation
  • Processing Real-World Data
  • Finetuning the World Model

Acknowledgements

For SLAM, we use the version of "point_lio_unilidar" from autonomy_stack_go2.

Citation

If you find our work helpful, please cite:

@article{2026simdist,
  title={Simulation Distillation: Pretraining World Models in Simulation for Rapid Real-World Adaptation},
  author={Levy, Jacob and Westenbroek, Tyler and Huang, Kevin and Palafox, Fernando and Yin, Patrick and Omidshafiei, Shayegan and Kim, Dong-Ki and Gupta, Abhishek and Fridovich-Keil, David},
  journal={arXiv preprint arXiv:2603.15759},
  year={2026},
  url={https://arxiv.org/abs/2603.15759}
}