The world model has gained a sense of touch! 500,000 hours of video trained the first implicit tactile world action model

BeingBeyond released Being-H0.8, an implicit tactile world-action model for robots.

Built on UniHand-3.0, it uses over 500,000 hours of first-person video to learn contact prediction, tactile encoding, and slow-fast control.

The model is designed to help robots anticipate and react to touch during fine manipulation tasks like grasping, writing, and squeezing.

In dual-arm robot tests, it showed a scalable path toward dexterous embodied AI without needing dense tactile labels.