3D Autonomous Navigation of UAVs: An Energy-Efficient and Collision-Free Deep Reinforcement Learning Approach

Abstract

Energy consumption optimization is crucial for the navigation of Unmanned Aerial Vehicles (UAV), as they operate solely on battery power and have limited access to charging stations. In this paper, a novel deep reinforcement learning-based architecture has been proposed for planning energy-efficient and collision-free paths for a quadrotor UAV. The proposed method uses a unique combination of remaining flight distance and local knowledge of energy expenditure to compute an optimized route. An information graph is used to map the environment in three dimensions and obstacles inside a pre-determined neighbourhood of the UAV are removed to obtain a local as well as collision-free reachable space. Attention-based neural network forms the key element of the proposed reinforcement learning mechanism, that trains the UAV to autonomously generate the optimized route using partial knowledge of the environment, following the trajectories from which, the UAV is driven by the trajectory tracking controller.