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 a pipeline combining human-motion pretraining, simulation-based expert learning, robot demonstration collection and visual feedback 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 the first hand hands and sixteen-joint the second hand 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 the first hand object reorientation and 93.3% on the second hand reorientation. Flow matching achieved 90% on the second hand reorientation. These figures describe different tasks and conditions, not an overall ranking of robotic hands. Constraining new plans with previous plans can improve continuity.

A useful distinction is between doing something correctly once and doing it reliably in many situations. Imagine finding your way through a new school: memorizing one landmark may not be enough when the route changes. Robot learning also needs tests with unfamiliar objects and slightly different surroundings. The numbers in a paper describe its particular experiment. Asking where the robot fails, when a person helps and what was changed during testing makes it easier to understand what the invention can really do today. Shared wrist representations and hand-specific finger control support transfer studies. Results still depend on the object and control method. For reference, the developers named this system AcrossDEX 1.0.