Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (14): 1-13.DOI: 10.3778/j.issn.1002-8331.2312-0256

• Research Hotspots and Reviews • Previous Articles     Next Articles

Research Review on Deep Reinforcement Learning for Solving End-to-End Navigation Problems of Mobile Robots

HE Li, YAO Jiacheng, LIAO Yuxin, ZHANG Wenzhi, LU Zhaoqing, YUAN Liang, XIAO Wendong   

  1. 1.School of Intelligent Manufacturing and Modern Industry (School of Mechanical Engineering), Xinjiang University, Urumqi 830017, China
    2.College of Information Sciences and Technology, Beijing University of Chemical Technology, Beijing 100029, China
  • Online:2024-07-15 Published:2024-07-15

深度强化学习求解移动机器人端到端导航问题的研究综述

何丽,姚佳程,廖雨鑫,张文智,卢赵清,袁亮,肖文东   

  1. 1.新疆大学 智能制造现代产业学院(机械工程学院),乌鲁木齐 830017
    2.北京化工大学 信息科学与技术学院,北京 100029

Abstract: Autonomous navigation is the prerequisite and foundation for mobile robots to accomplish complex tasks. Traditional autonomous navigation systems rely on the accuracy of maps and cannot adapt to highly complex industrial and service scenarios. End-to-end navigation methods for mobile robots that do not rely on a priori map information and are able to make autonomous decisions through deep reinforcement learning, and environment interaction learning have become a new research hotspot. Most existing classifications cannot comprehensively summarize the challenges and opportunities of end-to-end navigation problems. Based on the characteristics of end-to-end navigation systems, the challenges of the navigation problem are attributed to the key issues of poor perception ability of navigation agents, ineffective learning and poor generalization ability of navigation strategies. The research status and development trends of end-to-end navigation systems are described. Representative research results in recent years addressing these key issues are detailed respectively, and their advantages and shortcomings are summarized. Finally, the future development trends of end-to-end navigation for mobile robots are prospectively envisioned in aspects such as visual language navigation, multi-agents collaborative navigation, end-to-end navigation for fusion super-resolution reconstructed images and interpretable end-to-end navigation, providing certain insights for the research and application of end-to-end navigation for mobile robots.

Key words: end-to-end navigation, deep reinforcement learning, perception ability, learning efficiency, generalization ability

摘要: 自主导航是移动机器人完成复杂任务的前提和基础,传统的自主导航系统依赖于地图的精度,无法适应高度复杂的作业和服务场景。移动机器人不依赖先验地图信息,通过深度强化学习与环境交互学习能够自主决策的端到端导航方法成为新的研究热点。大多数现有的分类方法不能全面地总结端到端导航问题的挑战和机遇,根据端到端导航系统的特点,将导航问题的挑战归结为导航智能体感知能力差、学习效率低和导航策略泛化能力弱等关键问题,阐述了端到端导航系统的研究现状和发展趋势,分别详细介绍了近年来针对这些关键问题的代表性研究成果,并对其优势和不足进行了归纳总结。最后,从视觉语言导航、多智能体协同导航、融合超分辨率重建图像的端到端导航和可解释性端到端导航等方面展望了移动机器人端到端导航的未来发展趋势,为移动机器人端到端导航的研究和应用提供一定的思路。

关键词: 端到端导航, 深度强化学习, 感知能力, 学习效率, 泛化能力