Robots handling objects need to sense forces, but touch sensors (optical, magnetic, etc.) produce different data formats, hindering knowledge sharing.
Researchers built a setup recording touch readings and wrench measurements (force and torque) from 24 sensors, collecting over 100 million pairs. They trained a model to predict wrenches from any sensor's data, learning a common physical representation.
This improved force estimation accuracy by 17.1% and cross-sensor matching by 13.11 percentage points. A robot using these features succeeded more often in manipulation tasks, showing that grounding touch in physics aids generalization across hardware.
