HERMES: Experience replay and operator plasticity in evolutionary UAV path planning

Abstract

Evolutionary algorithms are effective for UAV path planning in complex environments, yet they typically operate as amnesic search processes. Even when coupled with local heuristic repair, they repeatedly reconstruct similar obstacle-avoidance maneuvers without retaining reusable local geometric structure, leading to redundant search effort. We propose HERMES ( H omeostatic regulation and E pisodic R eplay with selective M emory consolidation for E volutionary path planning S ystems), a neuro-inspired framework that augments an A*-guided multi-objective evolutionary algorithm (MOEA)-based planner with three operator-level mechanisms: feedback-regulated mutation, replay-guided local segment reuse, and selective memory consolidation. Homeostatic regulation broadens mutation when infeasibility or hypervolume stagnation indicates an unproductive search state; replay substitutes stored local motion templates for repeated local A* reconstruction; and consolidation filters memory to retain only geometrically informative segments. Experiments on a benchmark suite spanning urban, suburban, and mountainous environments show that HERMES achieves the strongest overall performance among the compared methods. Ablation and telemetry analyses further show that replay is the principal driver of reduced local A* burden, while consolidation stabilizes memory reuse without sacrificing Pareto-front quality. These results show that, once local heuristic search is embedded in evolutionary variation, further gains can be obtained by regulating when variation should broaden and by retaining only those local experiences that remain computationally reusable.