Robots often use a hierarchy: a high-level planner sets long-range targets, while low-level controllers handle fast actions. In dynamic tasks like carrying objects, this separation can fail when the robot needs to recover balance or handle unexpected contacts faster than the planner updates. The fixed mapping from planner targets to controller goals does not adapt to rapid physical changes.
This work proposes EPO, an energy-based orchestrator that, at runtime, refines each controller's goal and selects the best controller-goal pair. For each low-level controller, EPO learns an energy function that scores how feasible and accurate a given goal is for realizing the planner's target. The pair with the lowest energy is executed.
In simulation, EPO improved the trade-off between precision and robustness. It also allowed adding a new controller without retraining existing ones. The same frozen system was deployed on a real humanoid robot for tasks like box pick-up and door opening without fine-tuning.