Berkeley's DexTacWAM turns video world models visuo-tactile, averaging 70.6 vs 38.0 on dexterous tasks

berkeley_ai · x · 2026-09-30

UC Berkeley's Humanoid Intelligence Center presents DexTacWAM, a visuo-tactile world-action model for dexterous manipulation.

Adapts a pretrained video world model into a visuo-tactile one via continual vision-to-touch learning, for predictive multi-finger contact modeling and action generation
Achieves the highest score on all six contact-rich tasks, averaging 70.6 vs 38.0 for the strongest baseline
Key ablation: swapping the tactile world-model latent for direct tactile features drops the 4-task mean from 74.7 to 26.6
Takeaway: world-model-style tactile latents beat injecting tactile features directly into the policy
Motivation: multi-finger tactile data are scarce and expensive; vision transfer offers a scalable path

More from Embodied