Oscillation Mitigation for Multi-QUAV Navigation under Constrained Environment based on DRL and High-order Information of APF

Quadrotor unmanned aerial vehicle (QUAV) faces a significant challenge in navigating constrained environments, which remains a central focus in this field. The artificial potential fields (APF) is commonly used to facilitate obstacle avoidance during QUAV navigation. However, existing methods often struggle with considerable oscillation issues. To address this problem, this paper combines APF with deep reinforcement learning (DRL), utilizing higher-order information to reduce oscillations. Initially, a cascade DRL controller is trained in an unconstrained environment. Following this, the gradient of the repulsive APF is incorporated into the QUAV’s state to ensure effective obstacle avoidance. Moreover, the second-order information of the APF, represented by the gradient’s time derivative, is incorporated to regulate velocity as a damping factor. Simulation results demonstrate the proposed method’s ability to reduce oscillations, showcasing an 84.4% decrease in trajectory curvature.