AcrossDEX 1.0 is a unified learning pipeline that connects human motion pretraining, simulation expert learning, robot demonstration collection, and visual closed‑loop execution. The human pretraining model predicts one second of bimanual motion from five images and thirty annotated hand states over the preceding second. Robot adaptation keeps the visual encoder, temporal backbone, and wrist parameters and adds separate finger fields and validity masks for the six‑channel Inspire hands and sixteen‑joint Allegro hands. The policy outputs sixty control targets from frontal images, command history, and task conditions and commits to a prefix of K steps before replanning.

Three generative action models share the same backbone: direct regression (behaviour cloning), diffusion, and flow matching. For the lid opening and return task with Inspire hands, direct regression achieved 40.6 % success, diffusion 18.8 %, and flow 9.4 %. For closed‑lid reorientation, direct regression reached 100 % (Inspire) and 93.3 % (Allegro); flow matching attained 79.3 % (Inspire) and 90.0 % (Allegro). Replanning with a temporal continuity constraint (RTC) lifted flow success from 0 % to 18.8 % and diffusion from 3.1 % to 9.4 % on opening; on reorientation, flow rose from 51.7 % to 79.3 % and diffusion from 41.4 % to 62.1 %.

The opening task, which lasts about 3.33 s and spans two hundred control steps, is more difficult than reorientation because it requires longer horizons and more complex contact transitions. Generative policies benefit from temporal constraints inherited from the previous plan, indicating that continuity is a key challenge.

Current constraints are: camera pose is fixed and images are calibrated fisheye converted to pinhole views; reorientation demonstrations use short closed‑lid segments, limiting contact diversity; evaluation focuses on task completion rate and orientation error, omitting other failure modes such as overheating or slippage.