Machine Learning Approach to Real-Time 3D Path Planning for Autonomous Navigation of Unmanned Aerial Vehicle

Abera Tullu, Bedada Endale, Assefinew Wondosen, Ho-Yon Hwang

The need for civilian use of Unmanned Aerial Vehicles (UAVs) has drastically increased in recent years. Their potential applications for civilian use include door-to-door package delivery, law enforcement, first aid, and emergency services in urban areas, which put the UAVs into obstacle collision risk. Therefore, UAVs are required to be equipped with sensors so as to acquire Artificial Intelligence (AI) to avoid potential risks during mission execution. The AI comes with intensive training of an on-board machine that is responsible to autonomously navigate the UAV. The training enables the UAV to develop humanoid perception of the environment it is to be navigating in. During the mission, this perception detects and localizes objects in the environment. It is based on this AI that this work proposes a real-time three-dimensional (3D) path planner that maneuvers the UAV towards destination through obstacle-free path. The proposed path planner has a heuristic sense of A⋆ algorithm, but requires no frontier nodes to be stored in a memory unlike A⋆. The planner relies on relative locations of detected objects (obstacles) and determines collision-free paths. This path planner is light-weight and hence a fast guidance method for real-time purposes. Its performance efficiency is proved through rigorous Software-In-The-Loop (SITL) simulations in constrained-environment and preliminary real flight tests.