Teaching robots to handle objects usually forces people to act through a robot's body, which limits natural skill. Copying only hand positions is not enough: if fingers do not press with the right force, objects slip, crush, or are missed.
The authors built CARD, which converts video, motion-capture, or touch-sensing glove demonstrations into robot commands that contain both hand poses and force targets for each finger region. It uses no object model, so rigid, squishy, and articulated objects need no special handling.
CARD kept contact on 43 to 78 percent of frames where human contact happened, far more often than methods that only copied hand positions. This makes human demonstrations a more scalable way to teach robots dexterous skills.