Topological Necessities Transfers Robot Subgoals Across Embodiments

This is a sharp reframing of robot planning: learn the task’s structural necessities instead of inheriting subgoals from one controller. The results are compelling, though the transfer claim depends on preserving the environment’s topology.

• –H0 captures bottleneck order, while H1 identifies route forks, producing certified, recursively organized subgoals.
• –A PointMaze-discovered gate set reaches a reported 96.1 Humanoid aggregate and 90.9 on AntMaze-giant without retraining.
• –Kitchen experiments suggest the method can uncover nested commitment structure beyond simple maze navigation.
• –The strongest contribution is executor independence, potentially enabling one strategic planner to coordinate different robot bodies.
• –The fixed, isomorphic free-space assumption leaves open how well the approach handles changed layouts, contacts, or genuinely different task geometry.