Development of a deep reinforcement learning model for UAV motion planning during windy conditions.

Abstract

Over the past years, UAVs have garnered high application in the civilian space as they have been used in commerce and package delivery, wildlife monitoring, disaster management, among other things. This increase in use has in turn led increase in quantity of drones in the air space, this with time lead to use of semi-autonomous and autonomous UAVs. The introduction of autonomous and semi-autonomous UAVs will pose a huge risk of collisions as well as battery or fuel depletion during flight, especially when flying in unfavorable weather conditions. This could result in the injury of people as well as damage to property.

In this research project, we utilized deep reinforcement learning to enable a UAV carryout autonomous flight in an area with static objects during windy conditions. In the simulation environment, the UAV will be tasked with moving to a target location in an environment with variable wind speed and has static objects. The drone is equipped with a front camera that can continually take 640 x 480 pixels, that will be enable the UAV identify the objects in the environment, hence taking favourable actions to avoid collisions. In addition to the UAV image, drone battery, heading angle of the UAV, and distance of the destination from the position of the UAV was added to the state, hence requiring a concatenated neural network for the deep reinforcement learning algorithm with resnet-50 as the image extractor. The key finding of the research is that for battery-sensitive applications, our RL while, for time-sensitive applications, the constant airspeed strategy would be the best solution.