基于深度强化学习的果园巡检机器人导航研究
2021-04-03户高铭



摘 要:智能农业机器人如何实现准确移动仍是开发者们面临的一个挑战。传统的导航主要是通过全球定位系统(Global Position System, GPS)的定位来完成导航任务,弊端是其定位精度易受GPS信号强弱的影响。为此,文章提出采用深度强化学习算法SAC(Soft Actor-Critic)来解决果园场景下的导航问题,通过有序随机的课程学习训练方式引导智能体训练。实验结果表明,该方法能够在不使用GPS的情况下很好地完成果园场景下的定点导航任务。
关键词:果园;巡检机器人;深度强化学习;导航;课程学习
中图分类号:TP242 文献标识码:A文章编号:2096-4706(2021)19-0154-04
Research on Navigation of Orchard Inspection Robot Based on Deep
Reinforcement Learning
HU Gaoming
(School of Information Engineering, Dalian Ocean University, Dalian 116023, China)
Abstract: How to realize the accurate movement of intelligent agricultural robot is still a challenge for developers. The traditional navigation mainly completes the navigation task through the positioning of Global Positioning System. The disadvantage is that its positioning accuracy is easily affected by the strength of GPS signal. Therefore, this paper uses deep reinforcement learning algorithm SAC (Soft Actor-Critic) to solve the navigation problem in orchard scene, and guide agent training through orderly and random course learning and training. Experimental results show that this method can well complete the fixed-point navigation task in orchard scene without using GPS.
Keywords: orchard; inspection robot; deep reinforcement learning; navigation; course learning
0 引 言
对智能农业装备的研究改进是推动智慧农业发展的重要内容,其中智能农业机器人导航技术已被广泛应用于各种作业情景,例如果园除草、果园巡检、作物施肥等作业环节的应用,该技术的引入不仅能为智能农业装备提供精准的导航,還有助于节省人力物力以及避免遗漏作业区域,提高作业效率和作业质量[1-3]。
然而,智能农业机器人如何实现准确移动仍是一个不小的挑战。果园占地面积大,其环境通常是非结构化的,在这种情况下寻求一条高覆盖率的安全作业路径并非易事,正是当下的重点研究内容。近年来,众多学者在果园巡检、小麦收割机等自主导航系统的研发上取得了一定成果。夏友祥等人[4,5]提出基于GNSS全局路径规划方法优化导航路径,减少重复作业和地头转弯路径,提高作业效率。然而基于全局路径规划的方法需要事先熟知环境信息,有赖于先验知识。……
