What's this about?
The first large-scale world-action model trained to predict future *touch* alongside future vision, for contact-rich robot manipulation (28↑).
What exactly is "contact-rich manipulation"?
Tasks where success depends on feel, not sight: inserting a connector, wiping a surface, handling deformable objects. Vision saturates — the missing signal at the moment of contact is force and texture.
Why does it matter?
Robot learning has been vision-first because cameras are cheap and touch sensors weren't scaled. A tactile-native foundation model is a bet the next capability jump in robotics comes from the fingertips.