Learned parametrized dynamic movement primitives with shared synergies for controlling robotic and musculoskeletal systems

Asalientfeatureofhumanmotorskilllearningistheabilitytoexploitsimilaritiesacrossrelatedtasks.Inbiologicalmotorcontrol,ithasbeenhypothesizedthatmusclesynergies,coherentactivationsofgroupsofmuscles,allowforexploitingsharedknowledge.Recentstudieshaveshownthatarichsetofcomplexmotorskillscanbegeneratedbyacombinationofasmallnumberofmusclesynergies.Inrobotics,dynamicmovementprimitivesarecommonlyusedformotorskilllearning.Thismachinelearningapproachimplementsastableattractorsystemthatfacilitateslearninganditcanbeusedinhigh-dimensionalcontinuousspaces.However,itdoesnotallowforreusingsharedknowledge,i.e.,foreachtaskanindividualsetofparametershastobelearned.Weproposeanovelmovementprimitiverepresentationthatemploysparametrizedbasisfunctions,whichcombinesthebenefitsofmusclesynergiesanddynamicmovementprimitives.Foreachtaskasuperpositionofsynergiesmodulatesastableattractorsystem.Thisapproachleadstoacompactrepresentationofmultiplemotorskillsandatthesametimeenablesefficientlearninginhigh-dimensionalcontinuoussystems.Themovementrepresentationsupportsdiscreteandrhythmicmovementsandinparticularincludesthedynamicmovementprimitiveapproachasaspecialcase.Wedemonstratethefeasibilityofthemovementrepresentationinthreemulti-tasklearningsimulatedscenarios.First,thecharacteristicsoftheproposedrepresentationareillustratedinapoint-masstask.Second,incomplexhumanoidwalkingexperiments,multiplewalkingpatternswithdifferentstepheightsarelearnedrobustlyandefficiently.Finally,inamulti-directionalreachingtasksimulatedwithamusculoskeletalmodelofthehumanarm,weshowhowtheproposedmovementprimitivescanbeusedtolearnappropriatemuscleexcitationpatternsandtogeneralizeeffectivelytonewreachingskills.