Development of Data-Driven Real-Time Urban Air Mobility(UAM) Operation Methods

Abstract

Recently, the demand for unmanned aerial vehicle (UAS) is rapidly increasing, and its shape is also evolving in a wide variety. With the operation of the increasing number of drones, the question of how to manage traffic is emerging. Accordingly, research on autonomous intelligent operation technology is being actively conducted. This paper introduces the study of AI(Artificial Intelligence) and Data-Driven unmanned aerial system operation techniques that can be applied and expanded to various aircraft by type-independent, away from the flight control perspective route plan that requires accurate modeling in existing unmanned aerial vehicles. In the first part of this paper, we attempt to model trajectories only with flight path data over time, and based on this, we deal with the research of techniques for predicting the paths of heterogeneous drones. Using the Expectation-Maximization (EM) algorithm, which is widely used in unsupervised learning of machine learning, trajectory data were classified into different clusters according to the similarity of patterns. Using a stochastic Gaussian Mixture Model (GMM), we perform simulations to predict current and near future paths from historical trajectory data. The second part of this paper introduces an optimal path planning technique based on Deep Reinforcement Learning(DRL) that enables the operation of multiple drones capable of autonomous collision avoidance. Among reinforcement learning techniques, continuous action was performed using TD3 techniques that showed excellent performance in necessary fields such as robotics. The Dynamic A* algorithm was developed by applying the existing A* technique at small time intervals in a dynamic environment, and the optimality and computation cost were compared to show the excellence of this research and computation efficiency. Based on only one DRL pre-learning model that avoids such randomly moving obstacles, it is confirmed by simulation that multiple UASs can be operated.