Berkeley's DexTacWAM adds multi-finger touch to world models, wins all 6 dexterous tasks

berkeley_ai · x · 2026-09-30

Researchers from UC Berkeley, UIUC and Northwestern present DexTacWAM, a visuo-tactile world-action model for dexterous manipulation.

Core idea: treats touch as part of the predicted world state, extending a pretrained video world model to multi-finger tactile dynamics via lightweight tactile-encoder adaptation and continual vision-to-touch learning; a single forward pass feeds visuo-tactile features directly to the action expert.
Compression: a finger- and pose-aware tactile compressor maps 10 fingertip streams into 2 hand-level latents, retaining 89.4% contact recall, yielding 2.26× faster training and 1.29× faster inference.
Data efficiency: adapts with only 4 hours of tactile data, no video-model tactile midtraining required.
Results: best on all 6 tasks with an average score of 70.6 vs 38.0 for the strongest baseline; demos include tool-mediated force control, bowl unstacking, bottle-cap unscrewing and two-hand wiping.

Code, weights and data are released on the project website.

More from Embodied