Multi-Objective Reinforcement Learning for Autonomous Drone Navigation in Urban Area

Unmanned aerial vehicles (drones) are widely used for reality capture, data collection, and aerial logistics in multiple industries. Existing research in autonomous drones focuses much on path planning and automated obstacle avoidance. However, in complex environments such as urban areas, a safe and efficient drone navigation depends on many more dynamic constraints, such as layout of buildings, local aerodynamics (e.g., side winds), signal coverage, and malicious attacks. And many of these factors are interdependent, such as the wind field being affected by the geometric features of nearby buildings. This paper proposes a multi-objective reinforcement learning algorithm for the drone's path planning while satisfying a variety of interdependent constraint conditions. The drone will be able to sense multiple environmental and physical conditions of the near field, and develop dynamic policies for prioritizing the navigational decision to optimize the path while minimizing the negative environmental impact. A policy network will be trained to set priority weights for specific environment and physical factors at a given point. A simulation study is performed to test the proposed algorithm in an urban aerial logistics case.