Tactile sensing is crucial for dexterous manipulation, but collecting dense touch data on physical robots is slow and tied to specific sensor hardware. Two new arXiv papers tackle this bottleneck from different angles. TouchSight uses bare-handed egocentric video and generative visual augmentation to predict tactile signals, potentially letting a robot anticipate contact without a dedicated tactile sensor at inference time. DexTouch-WM instead learns an action-conditioned tactile world model from human touch, aiming to transfer a sense of contact from human demonstrations to robot control.
The papers agree that tactile data scarcity is a central obstacle and that predictive models can help overcome it. Where they differ is the source of training signal: TouchSight derives touch from visual observations, while DexTouch-WM builds its model from direct human tactile interaction. Both approaches remain at the research stage, but they point toward a future where robots gain dexterity without exhaustive, embodiment-specific data collection.