Learning-Guided Task Allocation with Navigation-Aware Optimization for Multi-UAV Missions
Abstract
Multi-agent systems operating in complex environments require tight integration between high-level decision making and low-level navigation. While task assignment allocates robots to goals under mission constraints, navigation policies must account for environmental hazards such as obstacles and unsafe regions. Although deep reinforcement learning (DRL) enables effective navigation learning for Unmanned Aerial Vehicles (UAVs), learned approaches in multi-agent scenarios often lack interpretability and struggle to incorporate mission-level constraints such as priority among goals. This paper proposes a Mission Supervisor (MS) framework that explicitly couples task assignment with DRL-based navigation policies using a normalized reward formulation, enabling comparison of performance across different scenarios. The MS integrates UAVs' capabilities, goal requirements, and environmental factors, such as collision risk and navigation efficiency estimated via the learned $Q(s, a)$, to guide task assignment decisions. Three MS variants are proposed and evaluated across multiple scenarios with varying numbers of robots and goals. Results, including a controlled experiment in an indoor setting, demonstrate that incorporating navigation-aware metrics into task assignment leads to improved mission performance and robot safety, while maintaining interpretability and modularity.