Reinforcement Learning Model-oriented Autonomous Drone Navigation and its Applications

Abstract

Mobile robots, including unmanned aerial vehicles (drones), are utilized for surveillance, tracking, and data collection
in structures, infrastructure, and various surroundings. The significance of precise and comprehensive monitoring is
widely recognized for the early identification of issues and preventing their escalation. This necessitates the
development of versatile, autonomous, and robust decision-making mobile robots. Those systems must possess the
capability to learn by integrating data from numerous sources. Until relatively recently, they were specialized in tasks.
This study presents a generic navigation method that utilizes data from onboard sensors to direct the drone to the issue
location. In dangerous and safety-critical scenarios, the precise and swift identification of issues is essential. The
research employs the proximal strategy optimization Deep Reinforcement Learning (DRL) system with incremental
curricula learning and long-term memory neural networks to develop a versatile and flexible navigation system. The
study assesses several configurations compared to a heuristic method to illustrate its precision and efficacy. The study
evaluates the assurance of the drone's security by analyzing its performance using its navigation system in real-world
situations.