DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation

Researchers at UC Berkeley, UIUC, and Northwestern extend a pretrained video world model to multi-finger touch for contact-rich dexterous manipulation.

21 Sep 2026

DexTacWAM overview: the training recipe, the visuo-tactile world-action model, and real-robot results on six contact-rich tasks, averaging 70.6 against 38.0 for the strongest baseline.

Haoran Yuan^1,*,‡, Zekai Wang², Boning Shao², Haoran Lu^3,*, Trevor Darrell², Ismini Lourentzou^1,†, Wei Zhan^2,†

¹University of Illinois Urbana-Champaign
²University of California, Berkeley
³Northwestern University

* Work done during a visit to UC Berkeley; ‡ Project lead; † Equal advising, co-corresponding authors

arXiv preprint arXiv:2609.24976

PDF arXiv Project Page Code Weights & Data

Citation

@article{dextacwam2026,
  title         = {DexTacWAM: A Visuo-Tactile World-Action Model for Dexterous Manipulation},
  author        = {Yuan, Haoran and Wang, Zekai and Shao, Boning and Lu, Haoran and Darrell, Trevor and Lourentzou, Ismini and Zhan, Wei},
  journal       = {arXiv preprint arXiv:2609.24976},
  year          = {2026}
}