The authors built DexTacWAM, a world-action model that jointly predicts future sensory states and generates robot actions.
Touch-aware world models improve contact-rich dexterous robot manipulation
DexTacWAM jointly predicts visual and fingertip tactile dynamics, substantially outperforming vision-centric and tactile-conditioned baselines across six manipulation tasks.
Independent
Haoran Yuan · Zekai Wang · Boning Shao · Haoran Lu · Trevor Darrell · Ismini Lourentzou · +1 more
Research Digest··2 min read
Yuan et al.
Why this paper
Independent · Part of World Model Planning for Agents, now 19 papers
In one line
Pretrained video world models can be efficiently extended to multi-finger tactile dynamics using fingertip encoding and continual vision-to-touch learning, achieving a 70.6 average score on six dexterous tasks.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§