Fault-Diagnosis-Aware Adaptive Policy Switching for Robust Map-Free Mobile Robot Navigation Under Sensor and Control Degradation

Abstract

Heterogeneous sensor and control degradation can destabilize direct sensor-to-action navigation. This study presents a fault-diagnosis-aware adaptive policy-switching framework that combines a calibrated seven-class temporal health-state classifier with hysteretic routing among nominal, robust-fallback, and six fault-specialized proximal policy optimization (PPO) policies. Unlike monolithic robust reinforcement learning, the architecture makes the inferred degradation mode an explicit and auditable control variable and separates health-state inference from behavior selection. Across a large held-out simulation evaluation with three independently trained policy seeds, diagnosed switching improved the central severity-0.55 static benchmark and within-episode transition outcomes relative to robust PPO, while the broader severity sweep showed regime-dependent behavior. Diagnosed switching also retained an aggregate advantage under interaction-matched robust training. Mechanism controls showed that the gain depended on fault-specific specialist competence and correct fault-to-specialist correspondence rather than policy-bank capacity alone. Single-thread inference remained well below the 200 ms control interval. Within the evaluated simulation scope, the results support explicit online diagnosis and behaviorally matched policy selection as a practical strategy for robust map-free navigation under heterogeneous degradation.