A new AI model teaches robot hands to combine sight and touch, and the touch part does the heavy lifting.
Researchers behind DexTacWAM, described in a paper posted to arXiv on September 23, 2026, built a visuo-tactile world-action model that reads pressure and contact data from each fingertip separately, compresses that signal with a pose-aware encoder, and feeds it into a video diffusion model that predicts how a scene will unfold. Tested on six contact-rich manipulation tasks on a 22-degree-of-freedom bimanual robot platform, the model topped every single one, averaging a score of 70.6 against 38.0 for the best competing approach. The team also fine-tuned a pretrained vision model to understand touch in about four hours using roughly 100 demonstrations per task, without a separate touch-focused pretraining stage. That adaptation cost little: visual prediction quality stayed within 0.5 dB of vision-only versions, and the tactile compressor kept 89.4% of contact-detection accuracy while training 2.26 times faster and running inference 1.29 times faster.
Vision-only robot models routinely fumble tasks that hinge on how hard or where something is gripped, because a camera cannot see the moment a peg starts to slip or a fingertip loses friction. By folding touch into the same predictive world model that handles video, instead of bolting a tactile sensor onto vision as an afterthought, DexTacWAM points at a cheaper path to robots that handle fiddly, contact-rich work like assembly or food handling. The efficiency numbers matter as much as the accuracy ones: a four-hour adaptation and small demonstration counts are the kind of budget an actual lab can work with, not just a well-funded one.
One number is worth double-checking before anyone gets carried away: the ablation test showing performance collapse from 74.7 to 26.6 when tactile world modeling is removed is reported across only four of the six tasks used for the headline comparison, and the source gives no reason for the narrower scope. That is not necessarily a red flag, but until the authors explain what happened to the other two tasks, the ablation reads as a subset chosen for effect rather than the full six-task story the top-line score implies.