Kairui Hu

Robot policies no longer keep a model in the control loop. Vision-language-action models hid the state in their weights, and agent harnesses kept memory in a model context, so how an action was derived could not be fully inspected. Here the robot and environment state is written as code variables, the code measures it from camera images and proprioception, and every decision is computed from it; one program runs across all episodes of a task and reaches 70.24% success on RoboDojo's 42 bimanual tasks, 38.9 points above the best leaderboard method.

Robot policies no longer keep a model in the control loop. Vision-language-action models hid the state in their weights, and agent harnesses kept memory in a model context, so how an action was derived could not be fully inspected. Here the robot and environment state is written as code variables, the code measures it from camera images and proprioception, and every decision is computed from it; one program runs across all episodes of a task and reaches 70.24% success on RoboDojo's 42 bimanual tasks, 38.9 points above the best leaderboard method.