UniDex trains a single vision-language-action policy across eight different dexterous robot hands, ranging from 6 to 24 degrees of freedom. The training data — over 50,000 trajectories — comes from egocentric human video, retargeted to each hand morphology. The result: a unified policy that generalizes across embodiments, performing dexterous manipulation with hands it has never seen.
The key insight is that embodiment diversity is not a complication to be managed. It is the source of generalization. A policy trained on one hand learns that hand's idiosyncrasies — its joint limits, its finger lengths, its particular way of curling around an object. A policy trained on eight hands cannot memorize any of them. It is forced, by the diversity of the training set, to discover the abstract structure of manipulation itself: the contact geometries, the force patterns, the grip topologies that all hands implement differently but that all successful grasps share.
This is a general principle disguised as a robotics result. You cannot learn what is invariant from a single instance. Invariance is, by definition, the thing that survives variation. A child who has only ever seen one dog does not know what “dog” means — they know what that dog looks like. Concepts require contrasts. The hand is the wrong unit of analysis. Manipulation is the right one, and you reach it only by varying the hand.