Real-Time UAS Path Planning and Autonomous Target Tracking Methods using Deep Reinforcement Learning and Imitation Learning, and Validation with Flight Test

Abstract

Unmanned Aerial Systems (UAS) are widely utilized in civilian and military sectors for diverse tasks. Research to improve UAS autonomy, especially using AI technologies, is ongoing. This study applies Deep Reinforcement Learning (DRL) to real-time path planning, collision avoidance, and autonomous tracking. The paper addresses two main tasks: 'Part 1: Real-Time UAS Path Planning' and 'Part 2: Autonomous Target Tracking,' utilizing DRL in both contexts. In Part 1, DRL's performance is evaluated against the conventional A* algorithm and further verified through flight tests. Additionally, a Vision-based DRL approach is introduced to enhance responsiveness to obstacles. Part 2 concentrates on the real-time autonomous tracking of illegal drones, demonstrating the efficacy of combining DRL with camera control techniques for effective detection and tracking, as corroborated by flight test results.