基于图卷积网络和多头注意力机制的采摘机器人路径规划算法研究

打开文本图片集

中图分类号：S126；TP242 文献标识码：A 文章编号：2095-5553（2026）02-0113-08

Abstract：Inordertosolvethecomplex farmlandenvironmentanduncertainexternalinterferencefacedbypathplanning of pickingrobots，apath planning model of picking robots basedongraph convolutional networksand multi-head attntion mechanism wasproposed，whichefectivelyutilized the spatiotemporalcharacteristicsandcorrelationof thenodedataof picking robots togenerate high-quality path values.Theweight distribution method of graph convolutional networks is redefined，andthespatiotemporalpropertiesof nodedataareconsidered，sothatthe graph convolutionalnetworks can bettercapture thespatiotemporal dependence ofnodedata.The multiorder nearest neighbor connection method is used to enlargethereceptive fieldof thegraphconvolutional network toenhancetherepresentationcapabilityof thegraph convolutionalnetwork．Inthedecodingstage，anattentionfiltermoduleisaddedtofiteroutirelevantormisleading atentioresults bydesigninganatentiongate，thus improving thequalityof decodeddata.Themulti-headatention mechanism is used todecodethe nodedataofthe picking robots and obtainthepath value.Theexperimentsconducted on real farmland data sets show that the path cost of the proposed model is reduced by 19.69% ， 16.51% and 14.12% compared with ant colony algorithm and by 16.35% ， 14.29% and 12.12% compared with genetic algorithm on farmland nodesof diferent sizesat15，3Oand5O，respectively.Thereductionofreasoning timewasmoresignificant，which decreased by 58.33% ， 46.88% and 46.88% ，respectively. These results show that this model is superior to the existing algorithms in both effciency and speed of path planning，and provides a more eficient solution.

Keywords：picking robots；path planning;graph convolutional network;multi-head atention mechanism;attention filtering

0 引言

农业是人类生存和发展的基础，也是国民经济的重要支柱1。（剩余11205字）