QUAV flight control based on axially symmetric DRL

Abstract

Deep reinforcement learning (DRL) has emerged as a prominent technique for advancing flight control systems. However, existing research often neglects the inherent symmetry present in quadrotor unmanned aerial vehicle (QUAV) dynamics, resulting in the drawback of instability and inefficiency. To tackle these problems, we propose an axially symmetric network to enhance flight control performance. To be specific, a converting module is proposed to fuse the vertical state and horizontal state to realize the stable and axially symmetric control performance. Furthermore, the proposed method exhibits generality and it could be validated using various DRL algorithms. Through a series of comparative experiments, we validate the superiority of the proposed controller, demonstrating notable improvements in both the efficiency and robustness of flight control operations.