计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (8): 288-296.DOI: 10.3778/j.issn.1002-8331.2112-0528

• 工程与应用 • 上一篇    下一篇

深度确定性策略梯度学习的火星无人机控制

孙丹,郑建华,高东,韩鹏   

  1. 1.中国科学院 国家空间科学中心,北京 100190
    2.中国科学院大学,北京 100049
  • 出版日期:2023-04-15 发布日期:2023-04-15

Mars Unmanned Aerial Vehicles Control with Deep Deterministic Policy Gradient

SUN Dan, ZHENG Jianhua, GAO Dong, HAN Peng   

  1. 1.National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China
    2.University of Chinese Academy of Sciences, Beijing 100049, China
  • Online:2023-04-15 Published:2023-04-15

摘要: 为了降低控制器设计对火星无人机动力学模型的依赖,提高火星无人机控制系统的智能化水平,结合强化学习(reinforcement learning,RL)算法,提出了一种具有自主学习能力的火星无人机位置姿态控制器。该控制器由神经网络构成,利用深度确定性策略梯度(deep deterministic policy gradient,DDPG)算法进行学习,不断优化控制策略,最终获得满足控制要求的策略。仿真结果表明,在没有推导被控对象模型的前提下,基于DDPG算法的控制器通过学习,自主将火星无人机稳定控制到目标位置,且控制精度、调节时间等性能优于比例-积分-微分(proportion integration differentiation,PID)控制器的效果,验证了基于DDPG算法的控制器的有效性;此外,在被控对象模型改变或存在外部扰动的情况下,基于DDPG算法的控制器仍然能够稳定完成任务,控制效果优于PID控制器,表明基于DDPG算法的控制器具有良好的鲁棒性。

关键词: 火星无人机, 强化学习, 自主控制, 深度确定性策略梯度, 策略优化

Abstract: In order to reduce the dependence of controller design on Mars unmanned aerial vehicle(UAV) dynamic models and improve the intelligence level of Mars UAV control system, a reinforcement learning-based controller for Mars UAV is proposed. The controller consists of neural networks and is trained by deep deterministic policy gradient(DDPG) algorithm. Finally, it obtains a control strategy to meet the control requirements according to current states and targets. The simulation results demonstrate that the controller based on DDPG is able to control the Mars UAV to a specified position autonomously without the derivation of UAV dynamic model. Mean-while, the performance such as control precision and adjustment time reaches the effect of proportion integration differentiation (PID) controller, which verifies the effectiveness of DDPG-based controller. In addition, when the controlled object model changes or there is external disturbance, the controller based on DDPG still completes the task stably, and the control effect is better than PID controller, indicating that the controller based on DDPG has good robustness.

Key words: Mars unmanned aerial vehicle(UAV), reinforcement learning, autonomous control, deep deterministic policy gradient, strategy optimization