摘要
针对追捕条件下速度劣势但加速度优势的旋翼无人机逃逸问题,结合无人机飞行过程中动力学约束,提出一种基于深度Q网络的无人机逃脱方法。该方法基于人工势场法改进无人机在强化学习过程中的奖励函数,通过合理的单步奖励,解决稀疏奖励问题。建立追捕者与逃逸旋翼无人机的仿真环境,通过模型仿真追捕环境下无人机的逃脱过程,使无人机不断学习逃脱动作决策方法。实验表明,运用改进的奖励函数的训练效果与普通奖励函数相比学习效果更好,旋翼无人机能够通过学习获得应对追捕的逃脱方法。
To enhance the escape ability of multicopters with high acceleration but low speed,a Deep Q Network(DQN)based escape method is proposed in this study,incorporating dynamic constraints during flight.The reinforcement learning process utilizes an improved reward function derived from the artificial potential field method to address the sparse reward problem.A simulation environment is established for a pursuer and an escape multicopter to facilitate the multicopter’s learning of the escape action decision method.Results indicate that the improved reward function yields better training effects compared to conventional reward functions.Furthermore,through learning,the multicopter acquires the ability to escape from capture.
作者
方俊逸
陈国良
FANG Junyi;CHEN Guoliang(School of Mechanical and Electronic Engineering,Wuhan University of Technology,Wuhan 430070,China)
基金
国家自然科学基金资助项目(61672396)
关键词
旋翼无人机
逃脱算法
深度强化学习
改进人工势场
multicopter
escape algorithm
deep reinforcement learning
improved artificial potential field
作者简介
方俊逸(1996-),男,湖北枣阳人,武汉理工大学机电工程学院硕士研究生.
