UC-MESL: A MAP-Elites Skill Library with Hierarchical Policy Switching for Robust Rescue Robotics Under Sensing Uncertainty

Abstract

Robots deployed in disaster scenarios—such as collapsed buildings, flooded tunnels, or conflict-damaged urban areas—must navigate without GPS, operate under poor sensing conditions (dust, smoke, darkness), and adapt quickly as the situation changes. Most existing learning-based approaches train a single navigation policy, which often fails when the environment or mission context shifts in unexpected ways. This paper presents UC-MESL (Uncertainty-Conditioned MAP-Elites Skill Library), a system that instead learns a library of specialized navigation behaviors—called “skills”—and switches between them in real time based on what the robot currently knows about its environment. Each skill is tailored to a specific operating condition, characterized by three interpretable traits: how much risk it tolerates, how actively it seeks out unknown areas, and what movement pattern it follows. A lightweight selector reads live uncertainty estimates from the robot's onboard map and chooses the most appropriate skill at each moment. We evaluate UC-MESL across three simulated rescue environments (collapsed rubble, flooded tunnels, and war-damaged urban blocks) under four levels of sensor degradation and realistic communication outages. Compared to the best single-policy baseline, UC-MESL finds 18.4% more victims within the mission time budget, reaches the first victim 31.2% faster, reduces hazard exposure by 24.6%, and loses only 8.3% of performance under severe sensor noise—versus 29.7% for a single-policy NEAT baseline. These results show that maintaining a diverse set of specialized skills, and choosing among them intelligently, produces more reliable rescue autonomy than optimizing a single all-purpose policy.