Hybrid learning mechanisms under a neural control network for various walking speed generation of a quadruped robot

Abstract

Leggedrobotsthatcaninstantlychangemotorpatternsatdifferentwalkingspeedsareusefuland canaccomplishvarioustasksefficiently.However,state-of-the-artcontrolmethodseitheraredifficult todeveloporrequirelongtrainingtimes.Inthisstudy,wepresentacomprehensibleneuralcontrol frameworktointegrateprobability-basedblack-boxoptimization(PIBB)andsupervisedlearningfor robotmotorpatterngenerationatvariouswalkingspeeds.Thecontrolframeworkstructureisbased onacombinationofacentralpatterngenerator(CPG),aradialbasisfunction(RBF)-basedpremotor networkandahypernetwork,resultinginaso-calledneuralCPG-RBF-hypercontrolnetwork.First,the CPG-drivenRBFnetwork,actingasacomplexmotorpatterngenerator,wastrainedtolearnpolicies (multiplemotorpatterns)fordifferentspeedsusingPIBB.Wealsointroduceanincrementallearning strategytoavoidlocaloptima.Second,thehypernetwork,whichactsasatask/behaviortocontrol parametermapping,wastrainedusingsupervisedlearning.Itcreatesamappingbetweentheinternal CPG frequency (reflecting the walking speed) and motor behavior. This map represents the prior knowledgeoftherobot,whichcontainstheoptimalmotorjointpatternsatvariousCPGfrequencies. Finally,whenauser-definedrobotwalkingfrequencyorspeedisprovided,thehypernetworkgenerates thecorrespondingpolicyfortheCPG-RBFnetwork.Theresultisaversatilelocomotioncontrollerwhich enablesaquadrupedrobottoperformstableandrobustwalkingatdifferentspeedswithoutsensory feedback.Thepolicyofthecontrollerwastrainedinthesimulation(lessthan1h)andcapableof transferringtoarealrobot.Thegeneralizationabilityofthecontrollerwasdemonstratedbytesting theCPGfrequenciesthatwerenotencounteredduringtraining.