AW-0 combines code-based planning, future video and robot control

AW-0’s researchers say the model generates skill programs, predicted video and robot actions together. Across eight real-world tasks, they report an average rubric score of 71.1, compared with 46.0 for the strongest baseline. In visually randomized RoboTwin scenes, they say a hybrid mode that can execute generated code raised success from 20.6% to 29.2%; it performed worse than the default mode in clean scenes.