Energy-Based Policy Orchestration

This research introduces EPO, an energy-based policy orchestrator that enables runtime selection of feasible policy-goal pairs for modular goal-conditioned policy families. By learning composite energy models specific to each policy, EPO refines goals and selects the lowest-energy pair without retraining existing policies or a centralized scheduler. The method improves the precision-robustness trade-off in simulation and is demonstrated with sim-to-real deployment on real humanoid loco-manipulation tasks using frozen policies and energy models.