Robotic hands can have very different numbers of joints yet face similar problems: hold an object, turn it and coordinate both wrists and fingers. Xiaoyu Yang and Nan Wu propose AcrossDEX 1.0, a pipeline combining human-motion pretraining, simulation^[1]-based expert learning, robot demonstration collection and visual feedback^[2] during execution.

During pretraining, the system predicts one second of bimanual motion from five images and thirty annotated hand states recorded over the preceding second. Shared wrist representations are retained when adapting to robots, while finger control accounts for the mechanics of each hand. The researchers tested six-channel Inspire hands and sixteen-joint Allegro hands, with control predictions revised from new camera observations.

In their evaluations, direct regression achieved 40.6% success on opening and returning a lid, 100% on Inspire object reorientation and 93.3% on Allegro reorientation. Flow matching achieved 90% on Allegro reorientation. These figures describe different tasks and conditions, not an overall ranking of robotic hands. Constraining new plans with previous plans can improve continuity.

The most revealing question is the difference between repeating a learned example and behaving well in a genuinely new situation. Generalization^[3] means coping with conditions that were not identical to the training examples. One striking video cannot tell us how many attempts failed. It helps to ask what the held-out tests looked like, whether objects and lighting changed, who selected the demonstrations and which baselines were compared. Those questions turn a fascinating result into a testable claim rather than a promise of a robot that can already do everything. Shared wrist representations and hand-specific finger control support transfer studies. Results still depend on the object and control method.