Can one glove teach robots fine-grained hand skills?

Dexterous robot manipulation fails when human demonstrations get lost in translation—retargeting discards fine control, while robot-specific teleoperation doesn't scale. RealDexUMI uses a shared dexterous hand module (light robotic hand + in-hand camera + tactile sensors) with an isomorphic glove controller, letting operators guide robots with their own fingers in real time. This preserves all the contact details and hand actions directly from collection to deployment. Policies trained on 8 tasks—fine pinching, long sequences, two-handed work—reach 88.75% success and transfer across different robot bodies.