Decomposing Gait Robustness in Demonstration-Free Humanoid Locomotion: A Predictive Diagnostic for ROM-Guided Reinforcement Learning

We propose a diagnostic decomposition for reinforcement learning locomotion controllers, partitioning action determination into External Perception (E), Constraint Enforcement (C), and Policy Residue (R), normalized to E+C+R=1. Applied to the ROM-GRL framework (Liu & Lau, 2025), the decomposition provides a vocabulary for distinguishing sensor dependence, reward overfitting, and autonomous gait mastery. Four experiments with falsification criteria are proposed. No experimental results are presented.