Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

Sangoh Lee, Sangwoo Mo, Wook-Shin Han (POSTECH)

INDI (Intention Distillation) distills behavior-level intent into the VLA action decoder. A frozen teacher model interprets demonstration segments into intent representations of what the behavior is trying to achieve, and the action decoder is trained to recover this intent at an intermediate layer alongside action prediction. At deployment, no teacher is needed: the policy recovers the intent on its own and acts on it.

🚧 Code Coming Soon

We are preparing the code for release. Stay tuned!

In the meantime:

• 📄 Paper (arXiv)
• 🌐 Project page with real-robot demos

Citation

@misc{lee2026actintentdistillingbehavior,
      title={Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models},
      author={Sangoh Lee and Sangwoo Mo and Wook-Shin Han},
      year={2026},
      eprint={2608.23478},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2608.23478},
}