Coverage Path Planning for Unmanned Aerial Vehicles in Complex 3D Environments with Deep Reinforcement Learning

Coverage path planning (CPP) for unmanned aerial vehicles (UAVs) defines a vital role in the automation process of UAV-supported disaster management. While multiple algorithms exist to solve the CPP problem for planar areas, the proposed algorithm is the first to handle complex three-dimensional environments and also account for power constraints and changing environments. By applying proximal policy optimization to an advantage-based actor-critic deep reinforcement learning model, the proposed framework enables an agent to efficiently cover the target area (TA), considering the orientation of the observation sensor, avoiding collisions as well as no-flying zones (NFZ) and reacting to changing environments. Furthermore, a safe landing mechanism, based on the Dijkstra algorithm, expands the framework to guarantee a successful landing in the respective start and landing zone (SLZ) within the power constraints. The model is trained on real data to learn the optimal control policy. Additionally, the framework was tested and validated on hardware in a drone lab to confirm its effectiveness and capability to perform real-time path planning.