基于強(qiáng)化學(xué)習(xí)的機(jī)械臂避碰研究.pdf

1、華中科技大學(xué)碩士學(xué)位論文基于強(qiáng)化學(xué)習(xí)的機(jī)械臂避碰研究姓名：張尚煒申請學(xué)位級別：碩士專業(yè)：工業(yè)工程指導(dǎo)教師：李世其付艷20070115華中科技大學(xué)碩士學(xué)位論文華中科技大學(xué)碩士學(xué)位論文IIAbstractThispaperreptsontheobstacleavoidanceproblemfroboticmanipulats.Machinelearning(ML)hasbecomeanimptantmeanstoenhancingthein

2、telligenceftherobotsastheincreasingrequirementinpracticalapplication.Afteraperiodoflearningreinfcementlearning(RL)robotsagentwithoutpriknowledgecanavoidobstaclesjustdependontheexplationtheenvironmentalresponse.TheRLmetho

3、dwasappliedtothisdomainbecauseitreliesonthesenss’perceptiononenvironmentnottheneedofaccuratemodelingenvironmentrobotitself.Amultiagentobstacleavoidancesystemwasbuiltfa3DOFplanarmanipulat.Thesystemcombinesarepellinginflue

4、ncerelatedtothedistancebetweenmanipulatnearbyobstacleswiththeattractinginfluenceproducedbytheangulardifferencetodrivethemanipulatmoving.AccdingtherealtimedemofmanipulatcontroltheSarsa(λ)algithmwhichisamajmethodofRLwaseda

5、sabasiccontrolstrategyfitsonpolicyfeatureefficiency.TheimplementprocessofthealgithmwasgivenintheendofthispaperasimulationexperimentshowedtheRLmethod’sfeasibilityavailability.Astheobstacleavoidanceproblemfroboticmanipulat

6、shascontinuousstatespacethestatespacepartitioningisaimptantfacttoimprovetheapplicabilityefficacyofreinfcementlearningalgithms.Thekmeansclusteringalgithmisusedtopartitionastatespace.Aseriesofsimulationsareprovidedtodemons

7、tratethepracticalvaluesperfmanceoftheproposedalgithmsinsolvingrobotmotionplanningproblems.Asimulationplatfmwasdevelopedftheobstacleavoidanceproblemofa3DOFplanarmanipulatonthe.Itisusedtoshowthesimulationtrialresultsanalyz

8、etheobstacleavoidancealgithm.Aseriesofexperimentsdemonstratethesystemhavingstrongcapacityfcollisionavoidanceeveninsomecomplexenvironments.Keywds:reinfcementlearningobstacleavoidanceagentkmeansclusteringstatespacepartitio