Teaching robots to feel what they see through tactile prediction

Robots struggle to anticipate touch without direct contact. This work introduces Mirror Touch Net, which aligns visual and tactile representations through semantic, distributional, and geometric constraints, allowing a robotic hand to predict millimetre-scale pressure across 1,140 sensors from RGB images alone. The approach extends to predicting how a robot would respond to observed human touch. Manifold analysis shows the training reshapes visual features into geometry matching tactile structure, simplifying cross-modal prediction. Code is released, making this immediately useful for roboticists building anticipatory or empathic systems.