TOWARDSPROPRIOCEPTIVEGRASPINGWITHSOFTROBOTICHANDSByThassyodaSilvaPintoADISSERTATIONSubmittedtoMichiganStateUniversityinpartialfulfillmentoftherequirementsforthedegreeofElectricalEngineering—DoctorofPhilosophy2021 ABSTRACTTOWARDSPROPRIOCEPTIVEGRASPINGWITHSOFTROBOTICHANDSByThassyodaSilvaPintoVariousrobotichands,gloves,andgrippershavebeendevelopedformanufacturing,prosthet-ics,andrehabilitation.However,theuseofrigidlinksandjointspresentschallengesincontrolandsafeinteractionswithhumans.Theemergingfieldofsoftroboticsseekstocreatemachinesthataresoft,compliant,andcapableofwithstandingdamage,wearandhighstress.Thisdisser-tationisfocusedonadvancingsoftactuators,softsensors,andperceptionforultimatelyrealizingproprioceptivegraspingwithsoftrobotichands.Inthiswork,severaltypesofsoftpneumaticactuators(SPAs)havebeentested,fabricated,andtested,includingoneembeddedwith3D-printedconductivepolylacticacid(CPLA)layercapableofstiffnesstuningandshapemodulation.Agrippermadeoftwosoftactuatorshasbeenprototypedtodemonstrategraspingofobjectsofdifferentsizesandshapes,withdesiredposture-holdingcapabilities.Carbonnanotube(CNT)-basedflexiblesensorarrayshavebeendesigned,fabricated,andin-tegratedtoSPAstoprovidedistributedstrainmeasurements.Thepresentedapproachallowscus-tomizeddesignofstretchablesensorarrayswithvariedsizeandshape.Simulationandexperimen-tationhavebeenperformedinordertoanalyzethesoftactuatordeformationduringbending,andtoconfirmthecapabilityoftheintegratedsensorarrayforcapturingtheactuatordeformation.3Dprintingoftouchandpressuresensorshasbeenfurtherinvestigatedforpotentialuseinrobotichands.Inparticular,anovelprocesshasbeenintroducedforproducingsoftconductorsandpressuresensors,involvingfirst3D-printingmicrochannelsinsoftsubstratesandthenfillingthechannelwithliquidmetal.WithaPolyJetprinter,functionalstraightmicrochannelshavebeenfabricatedwithsizesdownto150×150micrometersinthecross-sectionarea.Inaddition,spiral-shapedpressuresensorshavebeendevelopedwithacross-sectionsizeof350×350micrometersandoverallthicknessof1.5mm(50Aand70AShoreHardness).Althoughthesensorsrequire arelativelylargepressurethresholdtooperate,theyhaveshowntheabilitytowithstandhighpressuresupto1MPaandthushavepotentialtobeusedinindustrialapplicationsamongothers.Finally,preliminarycomputationalexplorationofintelligentgraspinghasbeenperformed.Inparticular,theclassificationofsoftgraspedobjectshasbeenexaminedthroughaneuroevolutionprocessforartificialbrains.SimulationwithSOFA(SimulationOpenFrameworkArchitecture)hasbeenconductedtoproducetheemulatedcontactforcemeasurements,whichhavebeenusedtotrainartificialneuralnetworks,includingMarkovBrainsfromtheModularAgent-BasedEvolver(MABE)platform,toproperlyclassifytheshapeandstiffnessofthegraspedobjects. iv Thisdissertationisdedicatedtomyparents,tomybrotherandtomysistersforlovingmeunconditionally,forunderstandingandsupportingmeinpursuitofmydreamsandcareeraspirations. v ACKNOWLEDGEMENTSIwouldliketothanktheDepartmentofElectricalandComputerEngineering,theDepartmentofMechanicalEngineering,theDepartmentofMicrobiologyandMolecularGenetics,theDe-partmentofIntegrativeBiology,andtheInstituteforQuantitativeHealthScience&EngineeringatMichiganStateUniversity.Also,IthankthesupportoftheSmartMicrosystemsLab,AdamiLab,andHintzeLabmembersthatcontributedtothisdissertationwork.Thisresearchwassup-portedbyCoordenac¸˜aodeAperfeic¸oamentodePessoaldeN´ıvelSuperior(CAPES)undertheSci-enceWithoutBordersprogram(BEX-13404-13-0),NationalScienceFoundation(DBI-0939454,ECCS-1549888,CMMI-1940950),andMSUFoundationStrategicPartnershipGrantProgram(16-SPG-Full-3236). vi PREFACEWhetheritsprimitivepurposewasforthrowingrocksorprovidingenhancedlocomotion,thehumanhandevolvedinawaythatallowsustoperformverycomplicatedtaskssuchasplayingin-struments,cooking,craftingandpainting.Itistruethatittakesquitesometimetolearnfinemotorskills,butthehand’scomplexfeedbacksystemplaysanimportantrolewhentrainingourbrainfordifferentmanipulationtasks.Moreover,theinherentcomplianceinhumanhandsprovidesadap-tivegraspingofobjectswithdifferentstiffnessesandshapes.Althoughautomatedmachinesarenowcapableofexecutingrepetitivemanipulationtaskswithoutexhaustion,theycanonlyworkwithwell-definedsettingsandinveryrestrictedenvironments.Ifitisdesiredforrobotsonedaytoleavethemanufacturingsitesandlaboratoriestoworkalongsidehumansathomeorworkplaces,advancedmanipulationskillswillbecrucial.Thehumanlifehasthrivedinvaryingandcom-plexenvironments,sorobotsinteractingwithitwillneedtohavethesamedexterousandflexiblegraspingthathumansdo. TABLE OF CONTENTS LIST OF TABLES . . LIST OF FIGURES . . . . . . . LIST OF ALGORITHMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix x . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv KEY TO SYMBOLS AND ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . xv . 1 Chapter 1 - Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Biological Sensory Signals and Natural Grasping . . . . . . . . . . . . . . . . . . . . 1.2 Robotic Grasping and Soft Robotic Grippers . . . . . . . . . . . . . . . . . . . . . . . 1.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Artificial Intelligence and Evolutionary Robotics 1.5 Contributions and Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Integration of Flexible and Stretchable Sensors . . 1 2 4 6 7 9 . . . . . . . . . . Actuators . 2 Chapter 2 - Materials, Design, Fabrication, Simulation and Controllers for Soft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.1 Elastomeric Materials for Soft Robots . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2 Soft Actuator Design Parameters and Fabrication Processes . . . . . . . . . . . . . . . 15 2.3 Soft Actuator Simulation . . 17 2.4 Pneumatic System Hardware and Control Software . . . . . . . . . . . . . . . . . . . 18 2.5 Soft Actuator Stiffness Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Chapter 3 - Fabrication, Characterization and Integration of CNT-based Flexible Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3.1 Screen-Printing-based Sensor Fabrication . . . . . . . . . . . . . . . . . . . . . . . . 27 3.2 Sensor Array Measurement and Characterization . . . . . . . . . . . . . . . . . . . . 30 3.3 Soft Actuator with Integrated Sensors Testing . . . . . . . . . . . . . . . . . . . . . . 34 . . . . . . . . . . sure Sensors 4 Chapter 4 - 3D-Printing of Liquid Metal-Based Stretchable Conductors and Pres- . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 4.1 Microfluidic Structure Design and Fabrication . . . . . . . . . . . . . . . . . . . . . . 39 4.2 Pressure Sensor FEM Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 . 47 4.3 Experimental Setup, Sensor Measurements and Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . Robots . 5 Chapter 5 - Computational Evolution of Control and Tactile Perception for Soft . 59 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Evolutionary Robotics with Soft Robots . . . . . . . . . . . . . . . . . . . . . . . . . 59 5.2 Evolving Markov Brain Controllers for Robotic Grasping . . . . . . . . . . . . . . . . 63 5.3 Discussions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 . . . . . . . . . 6 Chapter 6 - Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 83 vii APPENDIX . . . . . BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 viii ix LISTOFTABLESTableA1:Hyperelasticconstitutivemodelsfordescribingthemechanicalbehaviorofincompress-iblerubbermaterials...................................................................87TableA2:Mechanicalpropertiesofrubber-likematerialsusedinthefabricationofsoftactuators,flexiblesensorsandstructuresforsoftrobots.............................................87TableA3:ThedimensionsofpartsinthefabricatedSPAwithembeddedCPLA.............88TableA4:StiffnessvaluesofeachcomponentintheSOFAsimulation......................88TableA5:Thebinaryvalueofeachcorrespondentworkspaceconditioninthesimulationenviron-mentofSOFA........................................................................89 x LISTOFFIGURESFigure1.1:Someofthesensorsintheskinofahandandtheirrespectiveinterpretedstimulisuchastouch,pressure,andheat(Adaptedfrom[12])...........................................3Figure1.2:TheUtah/M.I.T.DexterousHandwithtendon-drivenactuation[16]...............4Figure1.3:Artificialevolutionofapopulationofrobotswithouthumanintervention..........8Figure2.1:Youngsmoduliofengineeredandbiologicalmaterials.(a)Elongationofahomogenousprismaticbar,and(b)theelastic(Young’s)modulusscalesforvariousmaterials(Adaptedfrom[62]).................................................................................13Figure2.2:Custom-builtpneumaticartificialmuscles(PAM).(a)FabricatedPAMswithdifferentdimensionsandmaterialsforlinearmotion,and(b)elongationtestingusingweightplates....15Figure2.3:Designapproachforthesoftactuatorandmoldparts.(a)Cross-sectionoftheSPAwithrepresentativedimensions;(b)coreandcavitysidesofthe3D-printedmold;(c)inflationstateoftheSPA............................................................................16Figure2.4:ThefabricatedSPAmadeofsiliconerubber...................................17Figure2.5:FiniteelementsimulationoftheSPA.(a)ContourplotoftheSPAFEMmodel(Drag-onSkin/PDMS),anditssequentialdeformationaspressureincreases.(b)Graphofthestrainvaluesateachnodecorrespondingtoapotentialsensorlocation(S1,S2,andS3in(a))..............18Figure2.6:Acontrolplatformforfluidicsoftroboticcomponents.Theminiaircompressoristheenergysourceforfluidicsoftactuators...................................................19Figure2.7:ControlsoftwarewithuserinterfacedevelopedinLabVIEW.Thesoftwareallowsmanualandautomaticcontrolofalloutputpressuresinthepneumaticboardforstaticorcontinuoussetpointvalues........................................................................21Figure2.8:FabricationoftheSPAwithaCPLAsheetintegratedtoitsbottomlayer.(a)3DprintedmoldpartsforfabricatingtheSPAcomponents;(b)Followingcuring,theupperandbottompartsoftheSPAarebondedtogether;(c)TheCPLAis3D-printedusingaFDM3D-printer;(d)ThincopperwiresaregluedtotheCPLAusingasilverpaste;(e)Ananti-slipfeaturetopreventslippageduringgrasping;(f)TheCPLAisencapsulatedwithuncuredsilicone;(g)TheSPAandtheencapsulatedCPLAarebondedtocompletethefabrication;(h)ThefinalSPA-CPLAdeviceaftercuringtime23Figure2.9:Afabricatedsoftpneumaticactuator(SPA)prototypewithanembeddedCPLAlayer24Figure2.10:Graspingofmultipleobjectsusingdifferentgraspingmodes.Aplasticcontainerwasgraspedusing(a)scooping,(b)pinching,and(c)parallelgrabbing.Additionaltestswereconductedforgrasping(d)aplushyminifootball,and(e)acupfilledwithcandies...........24 xi Figure2.11:Metallicweights(50g-500g)usedduringthesingle-fingerholdingexperiment.25Figure2.12:TestingloadcapacityoftheSPAintegratedwithCPLA.Asingle-fingerSPA-CPLAholding(a)aminimumweightof50g,and(b)amaximumof800gat22psi(innerchambers)and12Vinputatallhinges.Load-carryingtestswithoutpressureinput(andhingevoltageinputsoff):(c)underaloadof50gwithoutpressureandelectricalinputs,(d)underaloadof800gwithoutpressureandelectricalinputs...........................................................26Figure3.1:FabricationstepsoftheflexibleCNT-basedsensorarray:(a)thesensormaskisat-tachedtothepolymericsubstrate;(b)theCNTconductiveinkisappliedoverthemasksurfacetocreatethedistributedstrainsensors;(c)thetracemaskisattachedtothesubstrate;(d)AgNWsolutionisappliedthroughthemaskgapsusingapipette;(e)thincopperwiresareattachedtotheendpointofeachtracewithasilverpaste;(f)encapsulationofthedevicewithPDMS;(g)andbondingofSPAandsensorarraywithuncuredsilicone....................................29Figure3.2:ThefabricatedCNT-basedsensorarray.(a)Sensorarraydevice,and(b)aSPAwithintegratedsensors.....................................................................30Figure3.3:ContinuousmeasurementoftheresistancechangeinS1duringconditioningphase.Thedepictedlinesshowthedifferenceinthemaximumrangeofthesensoratinitialstepandafterthreetimeframes:10,30,and60minofloadingcycle....................................32Figure3.4:Sensorarraymeasurementsafterinitialconditioning:(a)theamountofstrainappliedinthecontinuousstretchandreleasetest,(b)andtherelativechangeinresistance(∆R/R)forsensorsS1,S2,andS3.................................................................33Figure3.5:Gaugefactorofeachstraingauge.(a)Strainand∆R/Rrelationshipbasedondifferentsetsofdata(withmeanlines)obtainedthroughthecyclicstretchingofthesensorarraysubstrate,and(b)thestandarddeviationof∆R/Rfromthemeasuredsamples........................34Figure3.6:ThesequenceofimagesoftheSPAwithembeddedCNT-basedsensors,showingthecurvatureandsensorvaluesatdifferentappliedpressures.(a)CapturedimagesduringactivationoftheSPA,and(b)thestraingaugearraymeasurementatvariousconstantpressuresetpoints.35Figure3.7:PressureandresistancerelationshipforthefabricatedSPAwithembeddedsensorarray.Thevariationinresistance(withmeanlines)forbothinflation(loading)anddeflation(unloading)stepsareshownforsensorsS1(a)andS3(b),whenthepressuresetpointwasvaryingfrom0to60kPa.(c)AsequenceofframesshowingtheactuationstepsandthecurvatureoftheSPA......36Figure3.8:TherelationshipbetweenstrainandcurvatureoftheSPAbasedontherangeofresis-tancevaluescapturedbythesensorarray.TheactualvaluerepresentsthemeasuredcurvaturefromtheSPAinnerradius...................................................................37Figure4.1:3D-printedstraightmicrochannelsoverasoftsubstrate(Agilus30)with2mmoverallthickness.Theminimummicrochannelcross-sectionsizeidentifiedwasof150µm×150µm(height/width).........................................................................42Figure4.2:3D-printedmicrofluidicspiral-shapedsoftpressuresensorwithembeddedliquidmetal xii (EGaIn)..............................................................................43Figure4.3:Designandfabricationstepsofthe3D-printedpressuresensorembeddedwithLM.(a)Thesensorcomponents:abottomlayermadeofpureAgilus30,atoplayermadeofAgilus30andVeroClearmixture(70AShoreHardness),amicrochannelstructureforfillingwithliquidmetal(EGaIn),andtwoendterminalsencapsulatedwithconductiveepoxy;thefabricationstepswere:(b)3D-printingofthebottomlayerwithmicrochannelcavitiesofcross-sectionsize350×350µm,(c)manualdispersionoftheglycerol-IPAmixture,(d)3D-printingofthetoplayerwithoutletsateachend,(e)vacuum-basedremovalofliquidsacrificiallayer,(f)manualinjectionofEGaIn,and(g)encapsulationofbothterminalsandsolderingofcopperstrandedwireswithconductiveepoxy44Figure4.4:Contourplotsofthesimulatedsoftpressuresensorforanappliedpressureof1MPa.(a)SubstrategeometrymeshedwithhexahedralelementsoftypeC3D20RH;(b)cut-viewofthevonMisesstress;(c)spatialdisplacementatz-direction;(d)logarithmicstrainatz-direction..46Figure4.5:PressureversusstrainplotobtainedfromtheFEAsimulationresults.Thestrainvaluescorrespondtotheaveragestrainamongallnodesinsidetheappliedloadregion(172nodes)..47Figure4.6:Testrigformeasuringandcharacterizingthe3D-printedpressuresensors.Averticallymountedfluidiccylinderwithcustom-builtforceconcentrator(16mmdiameter)andcontrolledbyapneumaticpowersource..............................................................49Figure4.7:Stepresponsecollectedfromthepressuresensorforaconstantinputof0.6MPa.Thetopgraphshowstheinputpressureandthebottomgraphshowsthemeasuredrelativechangeinresistance.............................................................................50Figure4.8:Comparisonofastepresponseforboththeexperimentaldataandtheobtainedtransferfunctionthroughmodelfitting..........................................................51Figure4.9:Thepressuresensoroutputundersinusoidalsensorinputwithrangefrom0.2to0.977MPa,withfourdifferentfrequencies(0.1Hz,0.25Hz,0.5Hz,and1Hz).Thedashedlinesshowthemeancurveofthecontinuousmeasurements..........................................53Figure4.10:Sinusoidalresponseformultipleinputfrequencies(frameviewof20seconds).Eachmeasurementwascollectedfor>2000secondsat0.1Hz,0.25Hz,0.5Hz,and1Hz,withapressurerangeof0.2MPato0.977MPa.................................................54Figure4.11:Graphforeachsinusoidalinputfrequencyafterreachingthesteady-state........55Figure4.12:Therelativechangeinresistanceversusthepressureinputforastaircaseinputsignal.Theaveragevalueof∆R/Rwascomputedforarangeof100pointsalongthesteady-stateregime,withapressureinputfrom0to1MPa,incrementsizeof0.1MPa,anddurationof2000secondsperstep..............................................................................56Figure4.13:ThecorrelationbetweenthecomputedaveragestrainfromFEAsimulationatthesensoractiveareaandtherelativechangeinresistancemeasuredfromthephysicalpressuresensordevice................................................................................57 xiii Figure5.1:AnFEM-basedreal-timecurvaturefeedbackcontrolofaSPA.TheresultswereachievedusingaPIDcontrollerandFEM-basedsimulationinSOFAwithdifferentmaterialprop-erties.................................................................................63Figure5.2:SoftroboticgrippermanualcontrolusingHuman-MachineInterface(HMI)withSOFA.(a)Asoftwearableglovewithbendingsensorsconnectedtoamicrocontrollerforpro-cessingtheanalogsignals,(b)softgripperwiththreefingersinunactuatedstate;(b)softfingerstouchinganobjectwithcollisiondetection;(c)objectbeingliftedbythesoftgripperwithpinchinggraspingmode........................................................................64Figure5.3:Softfingermeshnodelocationsforprobingforcevaluesduringeachiterationstepinthesimulatedworkspace.ThefollowingmeshnodenumbersattheSPAtipwereselected:1,30,and59................................................................................65Figure5.4:Softfingersatfinalsimulationtimestep(maximuminnerchamberpressure)incontactwithacubicobjectwithdifferentstiffnessvalues.Singlesoftfingerincontactwith(a)soft,(b)medium,and(c)hardcube.Adualsoftfingergripperincontactwith(d)soft,(e)medium,and(f)hardcube.............................................................................66Figure5.5:Forcevalues(XYZdirections)attheSPAtipincontactwithacube.Probedmeshnodesforasinglefingerandadualfingergripperincontactwithasoft(a-b),medium(c-d),andhard(e-f)cubicobject.................................................................67Figure5.6:Softfingersatfinalsimulationtimestep(maximuminnerchamberpressure)incontactwithasphericalobjectwithdifferentstiffnessvalues.Singlesoftfingerincontactwith(a)soft,(b)medium,and(c)hardsphere.Adualsoftfingergripperincontactwith(d)soft,(e)medium,and(f)hardsphere........................................................................70Figure5.7:Forcevalues(XYZdirections)attheSPAtipincontactwithasphere.Probedmeshnodesforasinglefingerandadualfingergripperincontactwithasoft(a-b),medium(c-d),andhard(e-f)sphericalobject..............................................................71Figure5.8:Evolutionaryprogressionforeachbrainoutputneuron(averageofall12cases),witha95%confidenceagainstreplicates.Eachgraphshowstheclassificationperformanceofacertainenvironmentcondition:(a)objectshape,(b-d)objectstiffness,and(e)numbersofcontactfingers76Figure5.9:Fitnessvalueforeachevolvedartificialbraintype:ANNs,MarkovBrains,andCGPs.Theaveragescorewascomputedfromall12casesandtheirrespectivereplicates............79Figure5.10:Longerneuroevolutionofsoftgraspingclassificationtaskwithvariedworkspaceconditionsfor40,000generations.......................................................80 xiv LISTOFALGORITHMSAlgorithm1:FitnessEvaluationandGeneticOptimizationinMABE.......................90 xv KEYTOSYMBOLSANDABBREVIATIONSAgNWSilverNanowireAIArtificialIntelligenceANNArtificialNeuralNetworkCADComputer-AidedSoftwareCGPCartesianGeneticProgrammingCNTCarbonNanotubeEGaInEutecticGallium-IndiumEREvolutionaryRoboticsFEMFiniteElementMethodGAGeneticAlgorithmHMMHiddenMarkovModelIPAIsopropylAlcoholMABEModularAgent-BasedEvolutionplatformMBMarkovBrainPDMSPolydimethylsiloxanePIDProportional-Integral-DerivativePLAPolylacticAcidPWMPulseWidthModulationSPASoftPneumaticActuatorSOFASimulationOpenFrameworkArchitecture 1 Chapter1IntroductionGrippersarewidelyusedbyanimalsandmachinesinordertointeractwiththeirenvironment.Inbiologicalorganisms,differentgripperstructureswerenaturallyevolvedforspecificgraspingandmanipulationtasks,takingmultipleformssuchasmouth,hands,pincers,beaks,trunksandtentacles.Besidesusingtheirnaturalgrippersasorgansofaction,manyanimalsrelyonsuchmechanismsforexploringtheirsurroundings,whichprovidesthelinkbetweenalivingorganismandtheworldaroundit.Asthedemandforfast,preciseandcontinuousmanufacturingprocessesincreasedovertheyears,robotsweredevelopedfortasksinvolvinggraspingandmanipulationofobjectsindangerousorunpleasantworkingconditionsinplaceofhumanworkers.Traditionally,theseroboticsystemshavebeencomposedofrigidbodiesmadefromstiffmaterialssuchasmetals(steel,aluminum)andceramics.Theyarewidelyappliedinindustryandcanbepreprogrammedtoexecutespecifictaskswithefficiency,butwithconstrainedadaptability.Inordertomaintainasafeworkenvironmentinsideafactory,robotsareplacedinadifferentareathanhumanssincetheirrigidlinksandjointscanbeharmfulforhumaninteraction.Theadditionofcompliantmaterialsincommonactuationmechanismscanleadtothedevelopmentofsafersystemsandenablegraspingandmanipulationofunknowobjects.Softroboticsisanemergingandcontinuouslygrowingfield,whichallowsapplicationswhere 2 robotscaninteractwithhumanssafelyinfactories,inthefieldorathome[1].Asoftrobotispri-marilymadeoutofsoftandextensiblematerialssuchassiliconerubbers[2],syntheticfibers[3]orgels[4],enablinglargedeformationandabsorptionofenergygeneratedfromimpacts.Thehighlydeformablecomponentsallowtheroboticsystemtoexperiencetheoreticallyinfinitedegreesoffreedom.Inaddition,thecompliantstructuremitigatestheimpactoftheenvironmentuncertaintyandproducescomplexmotionwithsimplecontrolinputs.Incontrasttoitshard-roboticcounter-parts,softrobotscanbefabricatedwithsimplestepsandinexpensiveandlightweightmaterials,andpermitmimicryofbiologicalfunctionalitiesthoughtheirintrinsicelements[5][6].Techniquescommonlyusedinroboticsforkinematicanddynamicmodelingcannotbedirectlyappliedtosoftrobotsduetotheircontinuumstructureandhighlynonlineardeformations.Agreatchallengefordynamicmodelingofsoftrobotsremainsduetotheirheterogeneousbodycompositionsandcomplexboundaryconditions.Theinvestigationofnewfunctionalmaterialswithmodulatedstiffnesspropertiesandproprioceptivesensing,aswellasnovelcontrolschemesforsoftsensory-motordevices,willbecriticalinthedevelopmentofenhancedsoftroboticsystems.1.1BiologicalSensorySignalsandNaturalGraspingSensoryreceptorsarethebridgeswhichconnectsourmindstotheoutsideworld.Differentkindsofstimuli(light,sound,temperature,ormechanicaldeformation)generateelectricalsignalsintheformofsymbolswhichcanbeinterpretedbythecentralnervoussystem[7].Inaddition,agivensensoryreceptor,biologicalorengineered,forexample,hasawell-definedrangeofstimuliforaparticularresponsetype.Variousspecieshavesimilarunderlyingsensorymechanisms,butthereisawidevariationinthesensitivityspectrum.Previousstudiesincrustaceanshaveshownthatstretchreceptorneuronscanregisterthemusclelength[8].Whenthereceptormuscleisstretched,thedetritusbecomedeformedandtheirmembranepotentialisreduced.Besidessendinginformationtothenervoussystem,stretchreceptorshavedirecteffectsonmotoneuronsconnectedtothesamemusclegroup.Thisfeedbacksystemisimportanttoregulationofmusclemovement. 3 Thehumanhandandwristarecomprisedof27bones,48musclesand22degreesoffreedom[9].Itisaversatiletoolcapableofexecutingavarietyofbasicgraspswhichhavebeendiscussedinmedicalliterature[10].Moreover,itdisplaysconformabilityonbothlargeandsmallscales,sincethefingerscancurlaroundobjectsduringgrasping,whilelocalizeddeformationhappensatthefingerandpalmtissuestoadaptitssurface.Furthermore,ourhandsareequippedwithagreatsophisticatedsensorysystem.Itscomplexitylevelisevidentbasedonthenumberofnervesinthehandandtheportionofbraindedicatedtoinformationprocessingandmovementcontrol[11].Figure1.1showsseveralskinsensorspresentonthehand.Sensoryreceptorslocatedatthehandtissue(skin,muscle,orjoints)providesomaticsensationssuchastouch,pressure,temperature,pain,andposition.Thesesomatosensorymechanismsallowustodeterminethetextureandshapeofobjectsthatwepalpate.Figure1.1:Someofthesensorsintheskinofahandandtheirrespectiveinterpretedstimulisuchastouch,pressure,andheat(Adaptedfrom[12]).Manyotherneuralstructuresplayaroleinadditionalimportantaspectsofsomaticsensation.Theprimarymotorcortex(M1),whichislocatedinthebackofthefrontallobe,isconsideredasthemainsiteformotorcontrol.Studiesconcernedwithbrainactivitymeasurementthroughfunctionalmagneticresonanceimaging(fMRI)haveshownthattheactivationofM1[13]andtheleftinferiorparietallobule(IPL)[14]arecompulsoryforthesomaticperceptionofhand-objectinteractivemovement.Withtheselow-levelandhigh-levelfeedbackloops,thehandcanbesaidto 4 beanorganwithacertainamountofdistributedcontrol.Cluesabouthigheraspectsofsensoryandmotorintegrationinvolvedinhumanhandgraspingtaskswillnotonlycontributetosomatosensorystudies,butalsoleadtoinsightfuldesignandcontrolapproachesforanewgenerationofroboticgrippers.1.2RoboticGraspingandSoftRoboticGrippersAroboticgripperdesigncanrangefromaservo-drivenclawwithmultiplefingerstosuctioncupspoweredbyvacuum.Inindustry,grippersmaybepreciseandemployedindifferentworkspaces,buttheyaredesignedtoperformveryparticulartaskswithwell-definedparts[15].Theirflexibilityisverylimitedsinceamagneticgripperwillnotbeabletopickupthermoplastics,andasuction-basedgripperforglassplatesisuselessformanipulatingmachinedparts.Thisproblemcouldbesolvedbyhavingacollectionofgrippersforasetoftasks;however,thereareafewdownsidesforthisapproach,suchasthehighcosttokeepaninventoryofmultiplegrippers,interruptionofmanufacturingprocessforgripperchange,andlackofcompatibilityfortactileandforcesensorsacrossmultiplegrippers.Infact,tactilesensingiskeytodexterousmanipulationsinceitallowstheapplicationoffeedback-basedcontrolalgorithmsthatexploitthesensorsignalsfortechniquessuchasgraspstabilityestimation,tactileobjectrecognitionandforcecontrol[16](Figure1.2).Figure1.2:TheUtah/M.I.T.DexterousHandwithtendon-drivenactuation[16]. 5 Asanalternativetorigidmechanisms,gripperswithsoftstructurescanbedevelopedformoreversatileandadaptivegrasping.Itscomplianceallowshigherintrinsicrobustnesstouncertaintyinactuationandperception,reducesthenumberofelementstobecontrolled,andofferslargertoler-ancetoexternalperturbationswhenmanipulatingabroadrangeofobjectsandconformingtothestaticenvironment.Moreover,thefabricationofsofthandsisfasterandlessexpensivethantheirrigidcounterparts.Forinstance,auniversalgripperbasedonthejammingofgranularmaterials[17]canbeusedtopickupawiderangeofdifferentobjectsthroughvacuum-inducedreversiblejammingtransitionswithoutsensoryfeedback.Furthermore,ahydraulic-drivensoftgripperforunderwaterapplicationsenablesdelicatemanipulationandsamplingoffragilemarinespeciesonthedeepreef[18].Asoftroboticapproachcanalsobebeneficialinthedesignofwearablede-vicesforhumanassistivetechnologies.Softroboticglovesdrivenbypneumaticsystems[19]orcableactuation[20]canproviderobotic-assistedtherapyduringpost-strokehandrehabilitationexercises.Althoughcomplianceallowsadaptivegrasping,itisdifficulttoestimatethespecificconfigu-rationofasofthandorgripperduetoitsmalleability.Theinformationaboutaroboticgripperconfigurationcanbeusefulfordecisionmakingduringmanipulation.Notonlyitcandetermineifagraspissuccessful,butitcanalsodetectaparticularobject’sshapeaswellasifitispickedupinthedesiredpose.A3Dconvolutionalneuralnetwork(CNN)canbetrainedtoestimatesuitablegraspsposesforpreviouslyunseenobjectsbylearningfeaturesandaclassifierfrompointclouddata[21].Thecombinationofbendingandforcesensorsonasoftroboticgripper,alongwithasensormodelandgraspingalgorithm,canidentifyanobjectonasinglegraspaftercharacterizationandtraining[22].Therefore,proprioceptivefeedbackinsofthandsisessentialforcomplexmanip-ulationofobjects.Thiscollectionofsensorydatawillenablerobotstograsptoolsintendedforuseonacertainorientation(pose),ensureappropriategraspingforces,andplaninteractionbetweenthegraspedobjectanditsworkspace.Byachievingseamlessgraspinginteractionwithsimilardexterityandflexibilityashumanhands,robotswillbeabletorobustlyusetoolsindifferenttasksandscenarios. 6 1.3IntegrationofFlexibleandStretchableSensorsAnexistingchallengeistheintegrationofsensingdevicesinthestructureofasoftmechanism,forfeedbackcontrolofthesystem.Severalcontrolschemesforsoftroboticsystemshavebeenexploredbyresearchers[23][24][25];however,feedbackcontrolofsoftrobotsusingcompact,in-tegratedsensorsisgenerallylimited.Themostcommonmethodsforsensormeasurementinsoftroboticsareresistiveandcapacitivetechnologies[26].Thesesensorsareusuallyfabricatedusingnanoscaleconductivematerialsdepositedoveraflexiblesubstratesuchaspolydimethylsiloxane(PDMS),whichremainsoneofthebestchoicesforstretchablesensorsubstratesduetoitsflexibil-ityandresistancetohightemperaturesandchemicalprocesses[27].3DmicrofluidicsystemscanbeeasilycreatedinPDMSwithcomplexgeometriesandtopologies,bytakingadvantageofthefactthatPDMSisanelastomerandprovidesawayofcontrollingbending,weaving,andbraidingthroughmoldtemplates[28],makingthemsuitableformicro-scalesensorapplications.Opticalfibersandpiezoresistivematerialscanalsobeusedassoftsensors.Atactilesensorsleevebasedonfiber-opticlightmodulation[29]canbeincorporatedinthebodyofasoftmanipulatorforpres-suresensing;however,largeinstrumentationdevicesarerequiredformeasuringthelightintensityvariation.Commerciallyavailablesensorscanbeembeddedinasoftpneumaticactuatorskin[30],buttheirsizeandshapecannotbecustomizedforvariousapplications.Microfluidicdevicesareanothersensingmodalitythatcanbeembeddedintosoftroboticstruc-turesandactuationmechanismstoprovideaclosed-loopfeedbacksystem.Microfluidicsisthescienceandtechnologyofsystemsthatprocessormanipulatesmallamountsoffluids,usingchan-nelswithdimensionsoftenstohundredsofmicrometers[31].Severalmicrofluidicstructureshavebeenappliedextensivelyinmanysensingapplicationssuchasforcedetection[32][33],straingauges[34][35],flowratemeasurement[36],andnoninvasivehealthanalysis[37][38][39].Cur-rentresearchhasidentifiedthedeformabilityandmobilityofliquidmetals(LMs)indeformablemicrostructures,whichbringssignificantpotentialforsoftrobotsandmachines[40].LMalloyssuchasEGaIn(75.5wt%galliumand24.5wt%indium)[41]andGalinstan(68.5wt%gallium,21.5wt%indium,and10wt%tin)[42],havebeenextensivelyappliedinthedesignofsoftsen- 7 sorsduetotheircapabilitiesinsimultaneouslypresentingmanyuniquephysicalpropertiessuchaslowmeltingpoint,excellentliquidity,highelectricalconductivity,goodthermalconductivity,lowvaporpressure,andlowtoxicityincomparisontomercury.SomeexamplesofLM-infusedmi-crofluidicsensorsincludewearablesoftsensorsforhumangaitmeasurement[43][44],softglovesforhandmotiondetection[45],softtactilesensorsforforcefeedbackinmicromanipulation[46],andsoftpneumaticactuatorswithembeddedmicrofluidicsensing[47][48][49].Anumberofothermethodsforpatterningliquidmetalsin2Dand3Dallowthecreationofmetallicmicrostructures,stretchableconductorsandsacrificialtemplatesformicrofluidicchannels[50][51][52].Inthenearfuture,itisbelievedthatLM-basedsoftmachineswillplaymanyrolesinbiomedicineandsmallmechanicalsystems.1.4ArtificialIntelligenceandEvolutionaryRoboticsWhenmachinesarecapableofperformingtasksinanintelligentmanner,weusuallydefinethisabilityasArtificialIntelligence(AI).Avastnumberofresearchsub-fieldsinAIhavestudieddifferenttechniquesinachievingintelligentmachines,notonlyfordescribingaspectsofhumanintelligence,butalsowithaimstocreatemachineswithbetterperformancethanhumansinwell-definedproblemsettings(playingstrategygames,recognizevisualorauditorypatterns,anddrivingacaronacrowdedstreet)[53].AnAImodelcanlearnthroughtrainingdata(machinelearning)orevolvedesiredperformancebasedonfitnessfunctions(evolutionaryalgorithms).Inevolutionarysystems,amachineiscapa-bleofdevelopingothersubmachinesaccordingtotheirabilitytoperformtasksinanenvironmentsimulatingtherealword.ThetheoryofevolutionwasdevelopedinCharlesDarwin’sOntheOri-ginofSpeciesbyMeansofNaturalSelection(1859),whichbasicallystatesthatvariationsoccurduringreproduction,andtheywillbepreservedinsuccessivegenerationswithapproximatepro-portiontotheireffectonreproductivefitness.Apopularmethodinevolutionarycomputationistheapplicationofgeneticalgorithms(GA),whereapopulationofartificialchromosomes(genotype), 8 encodingthecharacteristicsofanindividual(phenotype),selectivelyreproducesthechromosomesofindividualswithhigherperformanceandapplyingrandomchangesforseveralcycles(genera-tions).Similarly,autonomousrobotscanbeautomaticallycreatedbasedonatechniquecalledEvo-lutionaryRobotics(ER),whereartificialorganismscandeveloptheirownskillswithouthumanintervention[54].ThegeneralideabehindERisthataninitialpopulationofartificialchromo-somes,encodingeitherthecontrolsystemorthemorphologyofarobot,arerandomlycreatedandplacedinavirtualenvironment.Thefittestrobotsareselectedbasedontheirperformanceinvarioustasks,andthenallowedtoreproducewhileapplyinggeneticoperationssuchasmuta-tion,crossover,andduplication.Multiplecyclesofthisprocessarerepeateduntilarobotwiththedesiredperformanceisborn(Figure1.3).Thistechniquehasbeenappliedinapplicationsrang-ingfromtheevolutionofself-organizingbehaviorinaswarmofrobots[55]toevolvinggaitsforleggedrobots[56].Figure1.3:Artificialevolutionofapopulationofrobotswithouthumanintervention. 9 ItisinterestingtonotethatthefieldofERisbasedonafundamentalconceptdefinedasembodiedcognition[57],which,insteadofviewingintelligenceasmerelyamatterofsymbolprocessingwithinthebrainofalivingorganismoracontrolsystemofarobot,perceivesthelatterassomethingarisingfromtheinteractionbetweenbrain,body,andenvironment.Inaddition,certainprocessesthatwouldoriginallybeperformedbythebraincanbeoutsourcedtothebodybymeansofmorphologicalcomputation.Inthiscase,asoftrobotichandcangripacomplexobjectwithlesscontroleffortduetoitsmaterialelasticity,providingautomaticadaptivityforgraspingunknowobjects.Furthermore,animportantprocessinERstudiesistheabilitytotransfersimulatedbrains(con-trollers)toreal-wordroboticsystems(transferability).Forinstance,asimulationtoolwithvoxel-resolutionsoft,actuatablematerials,canevolvesoft-bodiedrobotswithmultiplebody-plansandgaitsfordifferentfitnessfunctions,mimiccardiacelectrophysiology,generatesqueezingforsmallspaces,andachieveswimminglocomotion[58][59][60][61].However,thetransferabilityprocessinthesecomputationaltoolsislimitedorevenimpossibleduetounmatchedparametersbetweenthesimulatedenvironmentandreal-worldsoftrobots.Theabilityofcomputationalevolutiontodevelopunexpectednovelsolutions,combinedwithsimulationofrealisticmaterialproperties,mayfavorthecreationofsoftroboticgripperswithmorenaturalmobilityandenhancedsensory-motorcoordination.1.5ContributionsandOutlineOurvisionforachievingrobustandadaptableroboticgraspingisbymeansofbiologicalinspira-tion,sensoryfeedback(proprioception),softandstretchablematerials(morphologicalcomputa-tion),andcomputationalevolutionofcontrolsystems(embodiedintelligence).Herewepresentcontributionsforeachofthoseareas,tacklingspecificaspectsinvolvedintheoverallchallengeofgraspingandmanipulatingobjectswitharobotichand.Thisdissertationisorganizedasfollows: 10 •Chapter2:Materials,Design,Fabrication,SimulationandControllersforSoftActua-torsThischapterdescribesthefirstbuildingblockinoursystem.Weexplainthematerialsanddesignsusedinthefabricationofsoftpneumaticactuators(SPAs).Differentinvestigatedprocessesaredescribedsuchassoftlithographyand3D-printing,withvariedSPAshapesandactuationmodes.Inaddition,wediscussthesimulationofthesedevicesusingfiniteelementmethod(FEM),whichcanbeusedtostudytheirmechanicalbehaviorpriortofab-rication.Acustom-builtpneumaticplatformisusedforcontrollingtheSPAsviagraphicalandembeddedsoftwareinseveralexperiments.AnovelapproachformodulatingtheSPAstiffnessispresented,usinga3D-printableconductivePLA(CPLA)materialthatcanbeintegratedtotheSPAstructureandusedingraspingtasks.•Chapter3:Fabrication,CharacterizationandIntegrationofCNT-basedFlexibleSen-sorsTheintegrationofsensorswithinanSPAshouldallowthemeasurementofdistributedde-formationwithoutaffectingitsadvantageousmalleability.Inlinewiththisreasoning,wehavedevelopedanoveltypeofsensorarraysmadeofCNTstrainggaugesandAgNWelec-trodesviaascreen-printingprocess.Adetailedrecipeforfabricatingandcharacterizingthesesensorsarraysisprovided.Moreover,wehaveperformedexperimentsonSPAswithintegratedsensorarraysinordertomeasureeachsensordeformation(changeinresistance)andestimatetheactualSPAcurvatureduringbending.TheresultsshowagreementwiththeobservedSPAdeformationandprovideafirststepintheimplementationofsensorarraysforlocalizedmeasurementsandfeedbackcontrolwithproprioceptivesensorsinsoftroboticmechanisms.•Chapter4:3D-PrintingofLiquidMetal-BasedStretchableConductorsandPressureSensorsMicrofluidicdevicescontrolfluidsonthemicrometer-scaleandarecommonlyusedforlab- 11 on-chipapplications,suchassensors,micropumpsandbiologicalanalysis.Commonlyre-portedfabricationmethodsforachievingflexiblemicrofluidicstructuresarelabor-intensive,requiremanycumbersomesteps,andhavelimitedoptionsformaterials.Thischapterpresentsarapid-manufacturingtechniqueusingaPolyJet3D-printerforcreatingsoftmicrofluidicsubstratesembeddedwithliquidmetalstogeneratestretchableconductorsandpressuresen-sors.Byusingthisnovelmethod,severalspiral-shapedsoftpressuresensorswithmultimaterial-basedsubstratescanbe3D-printedsimultaneouslyinlessthansixminutes.Experimentshaverevealedthatthese3D-printedmicrofluidicpressuresensorsareveryrobust,capableofwithstandinghighpressuresupto1MPa.•Chapter5:ComputationalEvolutionofControllersforSoftRobotsComputationallyevolvedcontrollersforsoftrobotichandshavethepotentialtoidentifyobjectshapeandhardnessorprovideposeestimationfordecisionmakingtasks.Inthischapter,wepresentsomepreliminarystudiesfortheclassificationtaskofsoftgraspedob-jectsthroughneuroevolutionprocessesofartificialbrains.Acombinationof3Dphysicsenginesanddigitalevolutiontoolsprovidedasubstrateforinvestigatingtheclassificationofseveralgraspingconditionsfromemulatedcontactforceandpressuremeasurements.Iden-tifiedsolutionsfromdifferentevolvedAImodelsarecompared.Inaddition,aroadmapforachievingfullyproprioceptivesoftgraspingisdiscussed.•Chapter6:SummaryandDiscussionThischaptersummarizesourcontributionsandfuturelinesofworkpresentedinthisdis-sertation.Finally,weconcludewithsomethoughtsregardingtheremainingchallengesinproprioceptivegraspingwithsoftrobotichands. 12 Chapter2Materials,Design,Fabrication,SimulationandControllersforSoftActuators2.1ElastomericMaterialsforSoftRobotsInordertoachieverobotsthatrealizedeformationwithlessenergy,theirbodyshouldbefabricatedwithlowmodulusmaterialssuchaselastomers.AmethodformeasuringtherigidityofacertainmaterialisbydeterminingitsYoung’smodulus(modulusofelasticity),whichdefinestheratioofstresstopercentelongationofprismaticbars(homogenous)subjecttoaxialloadingandsmalldeformations[62].Althoughsoftrobotsaremadefromheterogeneousandirregularstructuresthatundergolargedeformations,theYoung’smodulusisvaluableinformationforcomparingtherigid-ityofmaterialsusedinthefabricationsoftroboticsystems(seeFigure2.1).Conventionalrobotsarebuiltoutofmaterialswithelasticmoduluswithintherangeof109to1012Pa(metalsorhardplastics).Incontrast,naturalorganismsareoftencomposedofmaterialswithintherangeof102to108Pa(skinormuscletissue),ordersofmagnitudelowerthantherangefortraditionalrobots.Thisgreatmismatchpreventsrigidrobotsfrominteractingdirectlywithhumansandbiologicallycompatiblesystems,whereasprovidessoftrobotsapromisingopportunitytoaddresstheseissues. 13 Figure2.1:Young’smoduliofengineeredandbiologicalmaterials.(a)Elongationofahomoge-nousprismaticbar,and(b)theelastic(Young’s)modulusscalesforvariousmaterials.(Adaptedfrom[62])Platinum-basedsilicone-rubbersareoftenchosenasthebasisinthefabricationofsoftcompo-nentsduetotheirlowmoduli(durometeraslowas05-00),allowinghighstrainsandcuringpro-cessesattheroomtemperature.Nevertheless,themodelingofsilicone-basedstructurespresentschallengesduetotheirnonlinearnature.Thenonlineartheoryofelasticity,whichconstitutesthetheoreticalbasisforthestudyofhyperelasticmaterials,usesastrain-energyfunctiontodescribeinenergetictermsthemechanicalbehaviorofthisclassofmaterials[63].Foranisotropicmaterial,thestrainenergycanbedescribedasΨiso=Ψ(I1,I2,I3)(2.1)whereI1,I2,andI3arestraininvariantsI1=3Xi=1λ2i,I2=3Xi,j=1λ2iλ2j,I3=3Yi=1λ2i,i6=j(2.2) 14 withλ1,λ2andλ3representingtheprincipalstretches.Thehyperelasticmaterialmodels(Ta-bleA1)aremoresuitedtoexplainthenonlinearbehaviorofelastomersundertheassumptionofincompressibility.Inthecaseofuniaxialtensionofanincompressiblematerialλ1λ2λ3=1,λ2=λ3=λ−1/2,λ1=λ>1(2.3)andthegradientofdeformationisdefinedasF=λ000λ−1/2000λ−1/2(2.4)withtherightC∼andleftb∼Cauchy-Greentensors,C∼=FTF=λ2000λ−1000λ−1,b∼=FFT=λ2000λ−1000λ−1,C∼=b∼(2.5)Theinvariants(2.2)canbenowbedefinedfromtheright(deformation)orleft(spatial)Cauchy-GreentensorsI1=tr(C∼)=tr(b∼),I2=12((tr(C∼))2−tr(C∼)2)=12((tr(b∼))2−tr(b∼)2),I3=det(C∼)=det(b∼)(2.6)InSection2.3,weuseanumericalapproachtosimulatethebehaviorofsoftactuatorswiththestrain-energyfunctionofdifferenthyperelasticmodels. 15 2.2SoftActuatorDesignParametersandFabricationProcessesManytypesofactuationmechanismscanbeappliedinsoftrobots[64].Themostwell-knownandwidelyusedsoftactuatoristhepneumaticartificialmuscle(PAM)developedbyJ.L.McKibbeninthe1950s[65].Theseartificialmusclesconsistofaninflatablebladderinsideabraidedmesh.Theactuatorcanproduceaxialcontractionandradialexpansionmovementwhenpressurizedbyairgas(Figure2.2).Inpresentdays,thistechnologyiscommercializedbymanycompanieswhichprovidefluidicmusclesindifferentconfigurationsanddimensionsforapplicationsthatcancloselyemulatebiologicalmovements.(a)(b)Figure2.2:Custom-builtpneumaticartificialmuscles(PAM).(a)FabricatedPAMswithdifferentdimensionsandmaterialsforlinearmotion,and(b)elongationtestingusingweightplates.Inourwork,wehavemostlyexploredthefabricationofsoftpneumaticbendingactuatorsconsistingoffluidicchannelsinanelastomer(Figure2.3).Whenfilledwithpressurizedfluid,thechannelsexpand,causingthesoftactuatortobendtowardsastrain-limitinglayer.Theinflatablestructureismadefromrubberorelastomer(TableA2),whilethematerialusedfortheinextensiblelayercanbepaper,fabric,plasticfilmoranotherelastomerwithhigherdurometer. 16 (a)(b)(c)Figure2.3:Designapproachforthesoftactuatorandmoldparts.(a)Cross-sectionoftheSPAwithrepresentativedimensions;(b)coreandcavitysidesofthe3D-printedmold;(c)inflationstateoftheSPA.Similarly,therearedifferentfabricationprocessesavailableforcreatingcompliantroboticcom-ponentssuchassoftlithography(casting)[66],3D-printing[67],andphotopatterning[68].Asoftlithographyprocesscanbedividedintothreemainsteps:materialpreparation,vacuumdegassing,andcuring.Thespecificshapeoftheactuatorisachievedthroughthedesignofcavityandcorepartsofamold.Inaspecificstudy[69],wedesignedamoldusingaCADsoftware(SolidWorks,DassaultSystemes)withasquarecross-sectionof25×20mmandlengthof100mm.AsshowninFigure2.3a,theSPAdesignincludedtopographicalfeatures,asimilarconceptasin[70],andwith70◦angletoallowhigherbending.Themoldwasmadeofpolylacticacid(PLA)thermo-plasticandfabricatedwitha3Dprinter(MakerBotReplicator,MakerBotIndustries).Inaddition,wedesigneda2-partmoldwithmolded-inassemblyfeaturestofacilitatetheremovalofthesoftactuatorafterthecuringprocess(Figure2.3b).Wefirstmixedatwo-partliquidsiliconerubber(DragonSkin30,Smooth-On)andfilledthemoldwiththeuncuredmaterial.Thesiliconemixturewasthendegassedinavacuumchamberat-95kPaandallowedtocurefor16hoursattheroomtemperature.Thecuringtimecanbeshortenedbyheatingupthepouredsolutionoverahotplateorinsidealabovenincompatibletemperatureranges.Oncethecuringprocesswascompleted,thesoftactuatorwasremovedfromthemoldandathinlayerofsiliconewasbondedtotheSPAbottomsurfaceforcoveringthehollowchambersandenablingfluidicactuation.Inordertosupplycompressedairtothesoftactuatorchambers(seeFigure2.3c),asilicone-basedtubingwithsimilar 17 durometerwasinsertedthroughthebottomlayeroftheactuatorstructure.ThefabricateddeviceisshowninFigure2.4.Figure2.4:ThefabricatedSPAmadeofsiliconerubber.2.3SoftActuatorSimulationSimulationofsoftactuatorsandflexiblesensorsisanimportantprocessinthestudyofsoftroboticsystems,inordertoanalyzetheirmechanicalbehavior,frequencyresponseanddesignperfor-mance.Ananalyticalmodelcanprovidesomeinsightsintotheresponseofsoftcomponentstoex-ternalforcesforagivengeometry.However,itcannotcapturetheinteractionsofinternalelementswithdifferentmaterials.FiniteElementMethod(FEM)modelsprovidethenonlinearresponseofthesystemaswellasvisualizationofthesoftcomponentdeformationandglobalperformanceinmultipleconfigurations.Wehavesimulatedseveralsoftbendingactuatorprototypes,usingafiniteelementanalysissoftware(Abaqus/CAE,DassaultSystemes).Thegeometriesofthesoftactuatorandthesensorsubstrateweredesignedassolidbodiesandmeshedusingsolidtetrahedralquadratichybridelements(C3D10Helementtype),with48,253nodesand30,144elements.TheDragonSkin30andPDMSmaterialsweremodeledasanincompressibleYeohmaterial(µ=2.38kPa)[71]andincompressibleNeo-Hookeanmaterial(µ=1.84MPa)[72],respectively.AsshowninFigure2.5a,thesimulatedsoftbendingactuatorachievedaquarterbendingwhenapressureof40kPawasappliedtoitsinnerchamber.Theuniaxialstrainsvaluesatnodes4455,4365and4275weremeasuredfromtheSPAbottomsurfaceintheconvergedsolution(Figure2.5b). 18 (a)(b)Figure2.5:FiniteelementsimulationoftheSPA.(a)ContourplotoftheSPAFEMmodel(Drag-onSkin/PDMS),anditssequentialdeformationaspressureincreases.(b)Graphofthestrainvaluesateachnodecorrespondingtoapotentialsensorlocation(S1,S2,andS3in(a)).2.4PneumaticSystemHardwareandControlSoftwareIngeneral,fluid-drivenactuatorsforsoftrobotscanbepoweredbycompressedatmosphericair[73],hydraulicsystems[74],combustion[75],microfluidiclogic[76],andthroughphasechang-ingoflowboilingpointfluids[77].Therearemanypneumaticenergysourcesavailabletopowersoftrobots,andsomeimportantaspectsshouldbeconsideredsuchasselectingtheappropriatecompressorconfiguration(single,parallel,series)andmatchingthesoftroboticprojectspecifica-tions(energydensity,heat,pressure/flowrate)[78].Acustom-builtpneumaticplatformhasbeenfabricatedbasedonanopen-sourcedesignavail-ablefromtheSoftRoboticsToolkit[79].Thisplatformallowsthecontrolofupto4independentpressurechannelsforawiderangeofsoftroboticapplications(Figure2.6).Thecontrolboardhasasitspressuresourcea28psidualheadminiaturecompressor(BTC-IIS,ParkerHannifin)withbothpositiveandnegativepressure(vacuum)operationmodes.Allchannelsareactivatedbyagroupof3-waysolenoidvalves(VQ110U-6M,SMC),receivingamplifiedcontrolsignalsfromtransistorsthatareconnectedtoamicrocontroller(Mega2560,Arduino).Inordertoallowpres-surefeedbackcontrol,pressuregaugesensors(ASDXAVX100PGAA5,Honeywell)withanalog 19 measurementcapabilitiesareattachedtoeachchanneloutput.Figure2.6:Acontrolplatformforfluidicsoftroboticcomponents.Theminiaircompressoristheenergysourceforfluidicsoftactuators.Usingthishardwareframework,differentcontrolstrategiescanbeimplementedviasoftware.AcommonlyappliedmethodforcontrollingtheoutputpressureinapneumaticsystemusesPWMsignals.ThePWMdutycyclepercentageincombinationwithitswavefrequency,cangenerateopen/closecommandstothesolenoidvalves,thusaffectingtheoutletpressurevalueatacertaintime.Althoughamanualapproachcanbeachievedforselectingoutputvaluesbyusingswitch-ers/potentiometers,anautonomousrobotrequiresautomaticcomputationofitsdesiredoutputinresponsetoitscurrentinputvalues.Therefore,aproportional-integral-derivative(PID)controllerisdesignedtominimizetheerrorofthepressureoutput(PWMdutycycle)atagivenoutletbasedonadesiredsetpointinbothstatic(constantpressure)anddynamic(cyclicwave,trajectory)con-ditions.APIDcontrollerisdefinedas 20 u(t)=Kpe(t)+KiZe(t)dt+Kddedt,e(t)=r(t)−y(t)(2.7)withKp,Ki,Kd,beingthecontrollergainsfortuningandr(t),y(t),e(t)beingthesetpointvalue,actualoutput,andtrackingerror,respectively.Multiplecontrolprogramsweredevelopedusingdifferentprogramminglanguages,eachforaparticularapplicationandcompatibilitywithadditionalcomponents.ALabVIEWvirtualinstru-ment(VI)wascreatedforeasyinteractionbetweentheexperimenterandthepneumaticplatformduetoitsuser-friendlyinterface(Figure2.7).Besidesbeingabletoselectamanualorautomaticcontrolmode,theprogramcanalsoplotreal-timedataaswellassavereportsaboutcollectedmeasurements.Inthismethod,themicrocontrollerbehavesasadataacquisitionboard(serialcommunication)andallhigher-levelcomputationprocessesareperformedbythecomputer.Asanalternative,thecontrolschemewasalsoencodedonasingleexecutablefile,C++,andasascript,Python,forcompatibilitywithLinux-basedsystems.However,alluserinteractionsareinsteadper-formedthroughaterminalwindow(commandline).Themainadvantagesoftheaforementionedmethodsarethehigh-speedcomputationforamultiple-inputmultiple-output(MIMO)systemandthecommunicationacrossmanydevicessuchascameras(imagetracking),allowedbycurrentcomputerarchitectures.Somedrawbacksarethecommunicationdelayandpotentialnetworknoises(USB,Bluetooth,Wi-Fi),alongwithparallelprocessingofcompetingservices.Inaddi-tion,anembeddedversionofthecontrolsoftwarewasdevelopedforlowcostprototypingboards.Thisallowsanincreaseinsensorsamplingrateanddataprocessingsinceacross-communicationbetweentwocomputerarchitecturesiseliminated(firmwareandsoftware),andinternalmicrocon-trollerregisterscanbemanipulateddirectly.Differentfilteringtechniqueswereimplementedviasoftwareduetothepressuresensorsensitivityandpotentialenvironmentnoise.Oneapproachwastouseafirst-orderlow-passfilter,whichhasthetransferfunction 21 H(s)=ωs+ω(2.8)withω=2πfc,andfcasthecut-offfrequency.Forprocessingsignalsinthesampledtimedomain,thelow-passfiltercanbediscretizedasy(k)=y(k−1)e−aT+x(k−1)(1−e−aT)(2.9)whereaisthecoefficientrepresentingthecut-offfrequencyinthesampledtimedomain(radianspersecond)andTisthesamplingtimeoftheoriginalsignal(seconds).Figure2.7:ControlsoftwarewithuserinterfacedevelopedinLabVIEW.Thesoftwareallowsmanualandautomaticcontrolofalloutputpressuresinthepneumaticboardforstaticorcontinuoussetpointvalues.2.5SoftActuatorStiffnessModulationAgreatchallengeinthesoftroboticsresearchareaisthevariabilityandcontrollabilityofthede-formabilityandcomplianceofsoftrobots.Severalstiffeningapproachesarefoundintheliterature, 22 suchastheuseofactiveactuatorsarrangedinanantagonisticmannerandtheuseofsemiactiveac-tuatorsthatcanchangetheirelasticproperties[80].Inparticular,electricallyconductivematerialswithstiffnesschangingproperties[81]canalsobeembeddedinsoftroboticdevicestoallowrigid-itymodulationinspecificscenarioswhererelativelyhighforcesandfixedposturearerequired.In[82],acollaborativeworkwithMohammedAl-Rubaiai,allowedtheinvestigationofintegrating3D-printableconductivepolylacticacid(CPLA)materialwiththeSPAstructure,usingitasaninextensiblelayeraswellastoactivelymodulatethesoftactuatorstiffnessatdesiredlocations.Thisstudyhasbeendividedintwoparts:(1)characterizationofthethermomechanicalpropertiesandFEMsimulationoftheCPLAmaterial;and(2)fabrication,integration,andexperimentalval-idationofSPA-CPLAdevices.Mycontributionsarerelatedtothesecondpartofthiswork.AsshowninFigure2.8,additionalsoftactuatordesignsaredevelopedwith3D-printedresin-basedmolds(Form2,Formlabs).AthinlayerofsiliconeisappliedtotheSPAbottomsurfaceandcuredforcoveringthehollowchambersandenablingfluidicactuation.InordertomodulatethestiffnessoftheSPA,aflatsheetmadeofCPLA(CDP11705,Proto-pasta)isintegratedwiththedevice.Localindentationsaredesignedintheflatsheetgeometrytofacilitatebendingatthehingeloca-tions.Sincethematerialissuppliedasafilament,afuseddepositionmodeling(FDM)3Dprinter(QIDITECHI,QIDITechnology)isusedtofabricatetheconductivesheet.Thincopperwiresaresolderedtoeachhingeusingsilverpaste(SilverConductiveWireGlue,Amazon),withoutaffectingthedeviceflexibility.TheCPLAsheetisencapsulatedthroughasiliconerubberbathtoallowadhesionwiththeSPA.Asinglerectangular(20mm×140mm)siliconesheet(2mmthickness)isplacedinsideaglasscontainerwiththeCPLAsheetlayingontop.Thesiliconemixtureispouredinsidethecontaineruptoamarginof2mmabovetheCPLAsheet.TheSPAandencapsulatedCPLAarebondedtogetherusinguncuredsilicone.Inordertopreventslippageduringgraspingexperiments,anadditionalanti-slipfeatureenabledwithsurfacetextureisincludedinthedesignoftheSPA-CPLA.ThiscomponentismoldedwiththesamesiliconematerialastheSPA.Figure2.9showsthepictureofafabricatedprototype.ThedimensionsofallpartsintheactuatorarelistedinTableA3. 23 Figure2.8:FabricationoftheSPAwithaCPLAsheetintegratedtoitsbottomlayer.(a)3D-printedmoldpartsforfabricatingtheSPAcomponents;(b)Followingcuring,theupperandbottompartsoftheSPAarebondedtogether;(c)TheCPLAis3D-printedusingaFDM3D-printer;(d)ThincopperwiresaregluedtotheCPLAusingasilverpaste;(e)Ananti-slipfeaturetopreventslippageduringgrasping;(f)TheCPLAisencapsulatedwithuncuredsilicone;(g)TheSPAandtheencapsulatedCPLAarebondedtocompletethefabrication;(h)ThefinalSPA-CPLAdeviceaftercuringtime. 24 Figure2.9:Afabricatedsoftpneumaticactuator(SPA)prototypewithanembeddedCPLAlayer.AdditionalexperimentshavebeenconductedtotesttheCPLA-embeddedSPAsinatwo-fingergripperconfiguration.Thestiffnessmodulationineachfingercanbecontrolledindependently,al-lowingthegenerationofdifferentbendinganglestosuittheshapeoftheobjectbeingmanipulated.Multipleobjectswithvariousshapesandhardnessaretested,wheredifferentgraspingmodes(e.g.,scooping,pinching,grabbing.)areexecuted.InFigure2.10(a),(b),and(c),thesoftgripperisholdingaplasticcontainerusingtheaforementionedmodes,whileinFigure2.10(d)and(e),asoftminifootballandaplasticcupfilledwithcandiesareliftedwithaconstantpressureof20psi.Notethatinthesetrials,CPLAenableslocalshapereconfigurationoftheactuatorsfortheexecutionofdifferentgraspingmodes.(a)(b)(c)(d)(e)Figure2.10:Graspingofmultipleobjectsusingdifferentgraspingmodes.Aplasticcontainerwasgraspedusing(a)scooping,(b)pinching,and(c)parallelgrabbing.Additionaltestswereconductedforgrasping(d)aplushyminifootball,and(e)acupfilledwithcandies. 25 InordertoevaluatethepayloadcapacityoftheSPAintegratedwithCPLA,severalweightsareplacedinsideaplasticcontainer,whichisheldbyasinglefinger.Themassesrangefrom50gto500g(Figure2.11),andthetotalcarriedweightisincreasedinincrementsof50g.Thetestsareperformedforaminimumof50gandamaximumof800g.TwodifferentSPAstatesareconsideredforevaluatingthefingerstrength:pressurizedandunpressurized.Forthepressurizedcase,aconstantpressureof22psiisappliedtotheSPAinnerchambers,andallhingesareactive(12Vinput)throughouttheentireexperiment.Ontheotherhand,fortheunpressurizedcase,theSPAisinitiallysuppliedwith24psipressure,whilehavingallhingesactivated,andthenthepositivepressureisremoved(slowdecreasedowntoatmosphericpressure)afterthevoltageinputisturnedoffandtheCPLAlayeriscompletelycooledtoroomtemperature.AsshowninFigure19,forbothstates,thesoftfingerisabletowithstandthemaximumpayloadof800gwithoutcausinganydevicefailureordroppingtheweights.Theabilitytoholdshapeandcarryweightwithoutpressureandelectricalinputs(andthusnopowerconsumption)isparticularlysignificantforapplicationsthatinvolveholdinggivenposturesforlongperiodsoftime.Figure2.11:Metallicweights(50g-500g)usedduringthesingle-fingerholdingexperiment. 26 (a)(b)(c)(d)Figure2.12:TestingloadcapacityoftheSPAintegratedwithCPLA.Asingle-fingerSPA-CPLAholding(a)aminimumweightof50g,and(b)amaximumof800gat22psi(innerchambers)and12Vinputatallhinges.Load-carryingtestswithoutpressureinput(andhingevoltageinputsoff):(c)underaloadof50gwithoutpressureandelectricalinputs,(d)underaloadof800gwithoutpressureandelectricalinputs. 27 Chapter3Fabrication,CharacterizationandIntegrationofCNT-basedFlexibleSensors3.1Screen-Printing-basedSensorFabricationAsthefirststeptowardsproprioceptivefeedbackinSPAs,weinvestigatednewmethodsforachiev-ingflexiblesensorsthatcandetectstrainsalongasoftactuatorbottomlayer[69].InFigure3.1,wedemonstratehowaCNT-basedsensorarraycanbefabricatedthroughsimplesteps.Thisdesignallowsthecreationofcustomizableflexiblesensorswithoutgeometricconstraints.AnarrayofCNT-basedstraingaugeswasfabricatedusingascreen-printingprocess.Apoly-imidefilm(Kapton,DuPont)wasusedfordesigningthepatternofthesensorarrayandcircuittraces.Inthefirstmask,wecutthreeequallyspacedrectangles(10×2mm)correspondingtotheareasforthedistributedsensorarrayalongthedevicestructure.Asecondmaskwascreatedtoassistintheapplicationofthetracematerial,witheachsensorconnectedtotwotracesonitsends.Thematerialusedforthesensorsubstratewaspolydimethylsiloxane(PDMS)(Sylgard184,DowCorning)witha10:1baseandagentmixratio.Thesubstrate(≈2mmthickness)wasfabricatedusingtwoheat-resistantborosilicateglasssheets(150×150mm)clampedtogetherandheatedoverahotplatefor10minat150◦C.Apolyesterfilmwasadheredtotheinnersurfaceofeach 28 glasssheettohelpintheremovalofthePDMSsubstratewithoutcausinganytearorwrinkles.Thesizeandshapeofthespacermaterialdirectlyaffectthesubstrateuniformity.Inthisprocedure,wehaveusedmicroscopeslideswithidenticalwidthandlengthtoseparatetheglasssheetsatoppositeedges.Afterfabrication,thesubstratewascutinarectangularshape(120×25mm)andtreatedwithelectricdischargetoconvertittoalesshydrophobicsurface(wetting).Thesubstrateandthefirstmaskwereattachedtogethertoallowtheapplicationofthesensormaterial(Figure3.1a).AsinglewalledCNT(SWCNT)conductiveink,1mg/mLCNTs(VC101,ChasmTechnologies),wasappliedoverthepolyimidefilm(Figure3.1b),allowingthedepositionoftheSWCNTsonthePDMSsubstrateonlythroughthesmallcuts.Smallvariationsoninkdispersioncancausedevia-tioninsensingcharacteristics.AspatulawasusedtoleveltheCNTpasteevenwiththepolyimidefilmtoobtainanidealdispersionateachsensorlocation.Thesubstratewasonceagainheated(60◦C)inordertoensureadhesionbetweenbothmaterials.Oncedried,thesecondmaskwasattachedtothesubstrate(Figure3.1c).AgNWsinwater(AgNwL50H2O,ACSMaterial),with50nmofdiameterand200umoflength,wasappliedthroughtheopencutsusingapipette(seeFigure3.1d).Thetracesweredriedat60◦Cfor1h.AsshowninFigure3.1e,thincopperwireswereattachedtotheendpointsofeachtracebyfixingthemwithasilverpaste,60%Ag(PELCOColloidalSilver,TedPella).AsecondlayerofuncuredPDMSwasappliedtothetopfaceofthesensorarraytoencapsulatethedevice(Figure3.1f). 29 Figure3.1:FabricationstepsoftheflexibleCNT-basedsensorarray:(a)thesensormaskisat-tachedtothepolymericsubstrate;(b)theCNTconductiveinkisappliedoverthemasksurfacetocreatethedistributedstrainsensors;(c)thetracemaskisattachedtothesubstrate;(d)AgNWsolutionisappliedthroughthemaskgapsusingapipette;(e)thincopperwiresareattachedtotheendpointofeachtracewithasilverpaste;(f)encapsulationofthedevicewithPDMS;(g)andbondingofSPAandsensorarraywithuncuredsilicone.Asanexample,thesensorarraywasfabricatedwiththreestraingaugeswhicharereferredtoasS1,S2,andS3,inthiswork,withS1,beingclosesttotheairinlet,andS3,beingclosetothedistalendoftheactuator.TheflexiblesensorarraywascharacterizedandbondedtothebottomsurfaceofthesoftactuatorusinguncuredPDMS(Figure3.1g).ThefabricateddevicesareshowninFigure3.2. 30 (a)(b)Figure3.2:ThefabricatedCNT-basedsensorarray.(a)Sensorarraydevice,and(b)aSPAwithintegratedsensors.3.2SensorArrayMeasurementandCharacterizationTheCNTconductiveinkhasamesh-likenanostructurewhichcanchangeitselectroniccharacter-isticswhensubjectedtoextensionorcompression.WhenthereisavoltagebetweenbothendsofanindividualCNT-basedsensor,thelatterbehavesasavariableresistorthatchangesitsresistancevalueduringmechanicaldeformation.Avoltagedividercircuitcanbeusedtomeasurethevoltageacrosstheflexiblestraingaugesensorandthusitsresistancevalue.ByusingOhm’sLaw,wehave:Vs=Is×Rs(3.1)Is=ViRref+Rs(3.2)whereVi,Vsaretheinputvoltageandthevoltageacrossthesensor,respectively,Isisthecurrentgoingthroughthesensor,andRsandRrefarethesensorresistanceandthereferenceresistance,respectively.Bycombining(3.1)and(3.2),wecanget,Rs=RrefVi/Vs−1(3.3)TheresistanceRrefisselectedbasedontheminimumandmaximumresistancevaluesobtained 31 fromeachindividualsensor(denotedasRs,minandRs,max,respectively).Thisrelationcanbeexpressedas:Rref≈pRs,min×Rs,max(3.4)Inordertoassociatethechangeinresistancewiththeamountofstrainappliedatasensorposition,multiplemeasurementsoftheresistancearerecordedasthesubstrateissubjectedtodifferentvaluesofstrain.Theaxialstrainformulaisgivenby:ε=∆LsLso=Ls−LsoLso(3.5)whereLsoandLsarethenominal(untensioned)lengthandthecurrentlengthofthesensorarray,respectively.ThefabricatedCNT-basedsensorarraywasplacedonaprogrammablestretchingdevicewithbothendsclamped(80mmactivelength).Aloadingcycleof≈10s(stretchandrelease)wasperformedfor60minfortheinitialconditioningstepofthedevice(Figure3.3).Thesensorarraywassubjectedtoastretchof5%(Ls=84mm).Thisisanimportantprocessforstabilizingthemicroscalestructuralchangeofthenanomaterials.Theconditioningphasewasperformeduntilthedeviceachievedrersibleandstableelectricalsignalsforadesiredrangeofstrain.Althoughthesensorarrayhadreacheditsstabilitymarginafter60min,thestretchingcyclewasrunupto≈180mintoensuremeasurementrepeatability.Adataacquisitionequipmentwasusedincombinationwithanintegrateddataflowsoftware(LabVIEW,NationalInstruments)forcollectingthechangeinresistanceofeachstrainsensor. 32 Figure3.3:ContinuousmeasurementoftheresistancechangeinS1duringconditioningphase.Thedepictedlinesshowthedifferenceinthemaximumrangeofthesensoratinitialstepandafterthreetimeframes:10,30,and60minofloadingcycle.Aftertheconditioningstep,eachindividualsensorwasmeasuredinasequence.InFigure3.4,weshowashortsampleofthecollecteddatafromsensorsS1toS3after96minofstretching.EachstraingaugeexperienceddifferentrangesofresistancechangesincetheconcentrationoftheCNTmaterialandthethicknessalongthesubstratecandirectlyaffecttheirmeasurementrange.Thesensorsfabricatedinthisworkweretestedonlyforstretchesbelow10%.Alargerstrainvaluecouldincreasethenumberofcracksinthesensors,causingmalfunctionandunreliablemea-surements.Otherdesignoptionsthatalleviatemechanicalstresssuchasahorseshoe-shapewouldallowsensorfunctionalityforlargerstretches. 33 Figure3.4:Sensorarraymeasurementsafterinitialconditioning:(a)theamountofstrainappliedinthecontinuousstretchandreleasetest,(b)andtherelativechangeinresistance(∆R/R)forsensorsS1,S2,andS3.Basedonthedatacollectedfromthestretchingprocess,weobtainedtheresistance-strainre-lation(gaugefactor)foreachindividualCNT-basedsensor(Figure3.5).Asisobserved,allthreesensorsachieveasaturationlevelatacertainstrainvalue.SincetheCNTbundleschangetheirarrangementwhenthesubstrateisunderstrain,theconductivepathwithinthenanomaterialismodified.Thechangeineachsensor’sshapeduringstretchingcanalsocontributetoasaturatedstrainbehavior.Forsmallstrainvalues,thegaugefactorofeachstraingaugeisdefinedas:GF=∆R/Rε=∆Rs/Rso∆Ls/Lso=∆RsLsoRso(Ls−Lso)(3.6) 34 Figure3.5:Gaugefactorofeachstraingauge.(a)Strainand∆R/Rrelationshipbasedondifferentsetsofdata(withmeanlines)obtainedthroughthecyclicstretchingofthesensorarraysubstrate,and(b)thestandarddeviationof∆R/Rfromthemeasuredsamples.3.3SoftActuatorwithIntegratedSensorsTestingTheperformanceoftheintegratedsoftactuator-sensorwasevaluatedinexperiments.Inthispro-cedure,thesensorS2wasnotincludedsincetheprocessofembeddingthesensorarraysubstratewiththeSPAcausedamalfunctionofthissensor.Apressuregaugesensor(ASDXSeries,Honey-well)wasusedfordetectingtheinnerpressureofthesoftpneumaticactuatorduringthebendingmotion.Thesoftactuatorinletwasconnectedtoa12VDCminiatureaircompressor(Pmax=28psi)throughapolyurethanetubing.Duringthistest,wemeasuredtheresistancechangeofS1andS3.Whenthesensorarraywascombinedwiththesoftactuator,theminimumandmaximumresistancevalueswerechangedduetothecontributionofresidualstrainfromthesoftstructure.BymeasuringRs1andRs3withamultimeterduringtheactuatorinflation,weregisteredthenewvaluesas:Rs1,min≈1.3kΩandRs1,max≈2.7kΩ;Rs3,minand1.8kΩandRs3,max≈3.6kΩ.From(3.4),weobtainedthevalueforthereferenceresistorsasRref,s1≈1.8kΩ,Rref,s3≈2.4kΩ, 35 andaseriescircuitwasbuiltusingtrimpotentiometers.Thesensoroutputs(pressureandstraingauges)weresenttoamicrocontroller(Mega2560,Arduino),whichwasconnectedtoacomputerusingUSBinterfaceforcommunicationwiththedataacquisitionsoftware.Allmeasurementswereobtainedwithasamplingtimeof25ms,andasolenoidvalveconnectedtoanairpumpwascontrolledusingpulse-modulationwidth(PWM)at50Hz.AsshowninFigure3.6a,thefabricatedsoftactuatorachieved90◦bendinganglewithaninternalpressureof60kPa(≈8.7psi),whichisclosetotheFEMsimulationresults.Thecurvatureofthesoftactuatorinnerradiuswascapturedwithadigitalcamera,recordingindividualframesthatcorrespondtoaconstantpressurevalue.ThepercentageofthePWMdutycyclewascomputedbyaPIDfunctionwithinLabVIEW,withtheprocessvariablebeingthemedianofthepressuresensoroutputafterthirtymeasurements.Varioussetpointsweretestedwiththecontrolsystem,andtherespectiveresistancevaluesineachstraingaugewerecaptured,asshowninFigure3.6b.Figure3.6:ThesequenceofimagesoftheSPAwithembeddedCNT-basedsensors,showingthecurvatureandsensorvaluesatdifferentappliedpressures.(a)CapturedimagesduringactivationoftheSPA,and(b)thestraingaugearraymeasurementatvariousconstantpressuresetpoints.TheSPAcanbeappliedinseveraltaskssuchasreaching,graspingandtouching.Inorder 36 tohaveanaccurateestimationabouttheactuatorpositionorcurvature,thesensorarraymustbeabletocapturethedeformationduringmotion.InFigure3.7a,b,weshowthemeasurementsofthedifferenceinthesensoroutputresponse(hysteresis)duringinflation(loading)anddeflation(unloading),witharangeofpressuresetpointsvaryingevery0.8ms.Thetotaloperatingtimeis16s,withthesetpointincreasingfrom0to60kPa,andviceversa(Figure3.7c).Figure3.7:PressureandresistancerelationshipforthefabricatedSPAwithembeddedsensorarray.Thevariationinresistance(withmeanlines)forbothinflation(loading)anddeflation(unloading)stepsareshownforsensorsS1(a)andS3(b),whenthepressuresetpointwasvaryingfrom0to60kPa.(c)AsequenceofframesshowingtheactuationstepsandthecurvatureoftheSPA.Withthegaugefactorobtainedduringthesensorarraycharacterizationprocess,weestimatedtheamountofstrainintheSPAatdifferentcurvatures.TheactualcurvatureoftheSPAwasmeasuredfromthecapturedimageframesduringbending.Animageprocessingsoftware(Vision 37 Assistant,NationalInstruments)wasusedtomeasuretheSPAinnerradiusofcurvatureforeachpressuresetpoint.Thearclengthofaspecificbendingshapewasmeasuredandinsertedin(3.5)(withLso=100mm)tofinditscorrespondentaxialstrainvalue.InFigure3.8,wecomparebothsensorandactualmeasurementswhenthecurvatureincreases.TheseresultsagreewiththecompressionobservedinthebottomlayeroftheSPAduringpositivepressure.Figure3.8:TherelationshipbetweenstrainandcurvatureoftheSPAbasedontherangeofresis-tancevaluescapturedbythesensorarray.TheactualvaluerepresentsthemeasuredcurvaturefromtheSPAinnerradius.Byintegratingmultiplestraingaugesalongthesoftactuatorstructure,thesensorarraycanpro-videangularmeasurement,contactdetectionorproprioceptivesensingforasoftroboticsystem.Thedistributedmeasurementsallowestimationoftheactuatordeformationandthelocationsandforcesofinteractionsbetweentheactuatorandforeignobjects.Sincethesensorsarefabricatedthroughscreenprintingprocess,customizationcanbeappliedaccordingtothesoftactuatorgeom-etryandstraindirectionsofinterest.Inaddition,thenumberofstrainsensorsandtheirpositioncanbeoptimallychosenbasedontheregionswhichexperiencelargedeformations.Thesensorarraydesignpresentssomeconstraintsregardingthenumberofsensorsthatcanbefabricatedinasinglesubstrate.Sincethetracewidthcanaffectthesensorreading,reducingitsdimensiontoincreasetheamountofstraingaugescanimpactthereliabilityofthesensormeasurement.In- 38 creasingthethicknessofeachtracemayallowthedesignofthinnerfeatures,butthisprocesswasnotinvestigatedinthiscurrentwork.ThemanualfabricationmethodappliedinthisworkbringschallengesforachievingauniformsubstratethicknessandCNTinkdispersion.Automatedmassproductionultraviolet(UV)curingandscreen-printingprocessescouldreducetheuncertaintiesandissuesidentifiedintheinvestigatedmethod. 39 Chapter43D-PrintingofLiquidMetal-BasedStretchableConductorsandPressureSensors4.1MicrofluidicStructureDesignandFabrication4.1.1ConventionalFabricationMethodsTraditionally,microchannelstructuresforsensingdevicesarefabricatedusinglabor-intensiveandcumbersomemethods.Theliteratureinitsmajorityhasreportedmicrochannel-basedsensorsbyfollowingfabricationtechniquessuchaslasermicromachiningtocreatemolds[83],vapordeposi-tionofhydrophobicmonolayersforeasydemolding[84],spincoatingofPDMStocreatethinelas-tomerfilms[85],cross-linkingofsiliconesthroughoven-curing[86],andoxygenplasmatreatmenttoconstructthemicrochannelcavities[87].Aliquidmetal-basedsoftartificialskinwascreatedusingsiliconecastingovera3D-printedmold[88].Besidesmanyadditionalfabricationsteps,thesiliconecuringprocesstookmorethanthreehours.CurvaturesensorswithmicrochannelsfilledwithEGaIncanbeproducedusingacombinationofphotolithographyandreplicamoldin[89].However,theentirefabricationprocessincludingvapordeposition,siliconecross-linking,oxy- 40 genplasmatreatment,andelastomerfilmbondingisapproximatelyfourhourslong.Furthermore,PDMSmicrochanneltilesindevicestailoredtolaseraxotomyandlong-termmicroelectrodearrays(MEA)cantakemorethantwodaysforfabricationwhenusingsoftlithographyprocedures[90].Inaddition,abroadrangeofapproachesfortheapplicationof3Dprintingtechnologytorapidlyprototypemicrochannelstructureshavebeenexplored,acceleratingtheresearchanddevelopmentofmicrofluidicsensorsanddevices[91][92].Ontheotherhand,somerapidmanufacturingtech-niquessuchasusingpolyethyleneglycol(PEG)asasacrificiallayerthroughink-jetprintingre-quireda10-hourlongcuringprocessofthePDMS-basedsubstrate[93].Modifiedphotocurablematerialshavealsobeenexploredfor3D-printingofsoftpressuresensors,butacustom-builtprint-ingsystemisrequiredtodevelopthesedevices[94].Thesepreviouslyreportedmethodspresentsomedisadvantagesandchallengessuchastime-consumingprocedures,limitedmaterialsselec-tion,removalofthesacrificiallayer,andmicrochanneldevicereplication.Consequently,thereisademandfornewmanufacturingtechniquestoaddresstheseissues.4.1.2Microchannel3D-PrintingThefabricationofthemicrofluidicstructureforthepressuresensingdevicewasinspiredbasedonthetechniquepresentedin[95].Thispreviousstudyconductedinvestigationon3D-printingen-closedmicrofluidicchannelswithoutphotocurablesupportsinrigidmaterials.Themethodutilizesaviscousliquidsupportinstead,requiringminimaltonopostprocessingtoformsealedchan-nels.Inthiswork,wesoughttoimprovethismethodbyexploringitonsoftsubstratesinordertoachieveflexibleandstretchable3D-printablemicrochannel-basedconductorsandsensors.APolyJet-based3D-printer(J750,Stratasys)andaUV-curedresinwithpost-curedrubber-likeprop-erties(Agilus30,Stratasys)wereusedtocreatethemicrochannelcavities.Thisfabricationprocessfollowedathree-stepprocedure:3D-printingofthebottomsubstratecontainingopenmicrochan-nelcavities,fillingthemicrochannelcavitieswithliquidsupportmaterial,followedby3D-printingatop-substratedirectlyontothebottomsubstratetoclosethemicrochannels.Thesestepscanberepeatedseveraltimesdependingonthenumberofmicrochannelorsubstratelayers. 41 Traditionally,3D-printingofdeviceswithenclosedhollowchannelsrequiresinitiallyprintingthechannelsothatitisfilledwithasacrificialphotocurablesupportmaterial.Thismaterialisthenmanuallyremovedinpost-processing,aprocedurethatcantakehourstodays,anditisimpossibletoremovefromsmallchannelswithcomplexgeometries(spiralsorserpentines).Themethodusedhereallowsfabricationofchannelswithoutanyphotocurablesupportmaterial.The3Dgeome-trydesignforthemicrofluidicsoftsubstratewasseparatedintwoparts:abottomlayerwiththemicrochannelcavities,andatopflatlayerwithholesineachendofthemicrochannelforremov-ingtheliquidsupportmaterial.Thisdesignprocesscanbeeasilystacked,althoughwedidnotinvestigatesensordesignswiththesecharacteristicsinthiscurrentwork.First,thesoftsubstratebottomlayerwas3D-printedoveratransparencyfilm(PremiumTransparency,Xerox)tofacilitatethefinalsubstrateremovalfromthe3D-printerbed.Next,oncethebottomlayerprintingprocesswasfinalized,aliquidsacrificiallayer,composedofglycerol(Glycerol99.5%,Sigma-Aldrich)andisopropanol(2-Propanol99.5%,Sigma-Aldrich)mixture(70:30v:v),wasmanuallydispersedoverthemicrochannelcavitiesusingan1mlplasticsyringe.Asmallflexiblespatulawasusedtodistributethemixtureevenlythroughouttheexposedmicrochannels.Inaddition,duringtheliquiddispersionphase,thetoplayerprintingprocesswasalreadyinitialized,withthe3D-printerheadperformingautomaticcalibrationoutsidetheprintbedfor30seconds.Thisisanimportantprocesstoavoidbeadingoftheliquidsupportmaterialbetweeneachlayer,whichcouldleadtocloggingorirregularcavities.Theentireprintingprocess,includingmanualliquiddispersion,tookapproximately6minutes.Thismethodwastestedforfabricatingupto6sensorsubstratessimultaneously.Asaninitialinvestigation,wefirstcreateda3DCADmodelofasubstratewith2mmover-allthicknessandmultiplestraightmicrochannels(SolidWorks,DassaultSystemes)todeterminetheminimumcavitycross-sectionheightandwidthfor3D-printingmicrochannelsintoasoftma-terial.Weselectedseveraldimensionsbasedonreportedresultsfromliteratureregardingliquidconductor-basedsensors[96][97].Thewidthofthestraightmicrochannelshadarangefrom300µmdownto100µm,andaheightrangefrom200µmdownto100µm.AsshowninFigure 42 4.1,thesmallestachievablemicrochannelhad150µm×150µmcross-section,withapplicablemicrofluidicsinasoftsubstrate.Figure4.1:3D-printedstraightmicrochannelsoverasoftsubstrate(Agilus30)with2mmoverallthickness.Theminimummicrochannelcross-sectionsizeidentifiedwasof150µm×150µm(height/width).4.1.3PressureSensorDesignParametersAmicrofluidicpressuresensorwasdevelopedbydesigningspiral-shapedmicrochannelsintoasoftsubstrate.Thismakesthesensorsuitableforpressuredetection,sinceitwillnotrespondtouniaxialstretchesduetocounter-balancedelectricalresistancechangeinperpendiculardirections[86].Althoughitwasfoundthattheminimalmicrochannelsizeforthefabricationtechniqueinthisworkis150µm×150µm,thisdimensionpreventstheremovaloftheglycerol/IPAmixturefromtheinnercavitieswhendesignedasaspiral.Wehaveidentifiedthatmicrochannelcross-sectionsizeshigherthan350µm×350µmarebettersuitedforthistypeofsensordesign(Figure4.2).Thepressuresensorwasdesignedwitha1.5mmthicknesssubstrate(30mm×25mmrect-angularshape),andthespiralmicrochannelwascenteredatthemiddle.Themicrochanneldesign 43 wascomprisedofa3-turnspiral(inwardsandoutwards)and1.3mmspacingbetweenchannels,withatotalsensoractiveareaof20mmindiameter.Amixratioof70AdurometerbetweenAg-ilus30andVeroClearmaterialswasselectedinordertobalancethesensorcompliance.Sincethebottomlayerandcavitystructureallcombinedhadatotalheightof925µm,averythinupperlayer(575µm)enabledfastsensorproduction.ThecompletefabricationprocessisexplainedinFigure4.3.Figure4.2:3D-printedmicrofluidicspiral-shapedsoftpressuresensorwithembeddedliquidmetal(EGaIn). 44 Figure4.3:Designandfabricationstepsofthe3D-printedpressuresensorembeddedwithLM.(a)Thesensorcomponents:abottomlayermadeofpureAgilus30,atoplayermadeofAgilus30andVeroClearmixture(70AShoreHardness),amicrochannelstructureforfillingwithliquidmetal(EGaIn),andtwoendterminalsencapsulatedwithconductiveepoxy;thefabricationstepswere:(b)3D-printingofthebottomlayerwithmicrochannelcavitiesofcross-sectionsize350×350µm,(c)manualdispersionoftheglycerol-IPAmixture,(d)3D-printingofthetoplayerwithoutletsateachend,(e)vacuum-basedremovalofliquidsacrificiallayer,(f)manualinjectionofEGaIn,and(g)encapsulationofbothterminalsandsolderingofcopperstrandedwireswithconductiveepoxy. 45 4.1.4LiquidMetalEmbeddingandEncapsulationThemicrofluidicelectronicsinthisstudyarefabricatedusingEGaIn,whichhasahighelectricalconductivity(σ=3.4×106Sm−1),aresistivityofρ=29.4×10−8Ωm−1,andlowtoxicity[98].Aftercompletionofthe3D-printingprocessforthesubstrate,asmalltubingconnectedtoavacuumpumpwasinsertedintooneofthemicrochannelportstoextracttheglycerolmixture.Theremovalprocessonlytakesapproximately3to5secondsforeachdevice.Oncealltheliquidsupportwasremoved,a1mLsyringewith22gaugeneedle(0.70mm)wasusedtoinjecttheliquid-phasealloy(EGaIn,Sigma-Aldrich),composedof≥99.9%tracemetalbasis,insidethemicrochannelcavities(Figure4.3f).UncuredAgilus30wasinitiallyusedtoencapsulatetheopenmicrochannelportswhichwereconnectedtothincopperwires.However,asitwasreportedinpreviousworks[89],motionofthewiresinterfacingtheLMinthemicrochannelswithmultimeterprobescausesscatteringduringpressuretesting.Therefore,wehaveexploredanalternativemethodbysealingthemicrochannelportswithconductiveepoxy(8331SilverConductiveEpoxyAdhesive,MGChemicals),andthengluingthickbraidedcopperwirestoeachelectrodewiththesameadhesive.Toexpeditetheman-ufacturingprocedure,ahotplatewasusedtospeeduptheadhesivecuringprocessto10minutesat70◦C.Thisallowedagreatinterfacebetweensoftandhardconductivematerials.4.2PressureSensorFEMSimulationFiniteelementsanalysis(FEA)ofthe3D-printedsoftpressuresensordevicewascarriedoutusingamultiphysicssoftware(Abaqus/CAE,DassaultSystemes).Thesimulationwasconductedtoun-derstandtherangeofcompressivestrainsexperiencedbythesensor,whichwillbeinstrumentalinderivingthegaugefactorofthesensorwhenthelatteristreatedasastraingaugesensor.Asimplegeometryrepresentingthesensorsubstratewascreatedfollowingthesamephysicaldimensions.Inthisstudy,tofacilitatecomputation,wehaveconsideredthemicrochannelcavitiesandtheliq-uidmetalassolidbutsoftmaterialswithsimilarpropertiesasthesubstrate.Thematerialproperty 46 wassetasAgilus30usingtheOdgenhyperelasticmodel(TableA1)withparametersµ1=0.2127MPa,α1=1.3212,µ2=0.0375MPa,α2=4.318,µ3=0.001,α3=1.0248whichweredeterminedinpreviouswork[99].Aquasi-staticuniaxialstrainwassimulatedbyapplyingapressureinputofupto1MPawithincrementsof0.1MPaperstep.Thesensorsubstratewasmeshedwith280quadratichexahedralelementsand2058nodes(C3D20RH).Afixedboundaryconditionwassetonthebottomsurfaceofthesensortopreventplanarrotationanddisplacement.Thepressureonthetopsurfacewasdefinedbycirclewithadiameterof16mmcenteredatthesensoractiveareainordertoreplicatetheexperimentalsetup.Figure4.4showstheobtainedcontourplotsforvon-Misesstress,displacementandlogarithmicstrainalongtheloadaxis.Theaveragestrainwascomputedamongallnodesinsidetheappliedpressureregion(Figure4.5).Figure4.4:Contourplotsofthesimulatedsoftpressuresensorforanappliedpressureof1MPa.(a)SubstrategeometrymeshedwithhexahedralelementsoftypeC3D20RH;(b)cut-viewofthevonMisesstress;(c)spatialdisplacementatz-direction;(d)logarithmicstrainatz-direction. 47 Figure4.5:PressureversusstrainplotobtainedfromtheFEAsimulationresults.Thestrainvaluescorrespondtotheaveragestrainamongallnodesinsidetheappliedloadregion(172nodes).4.3ExperimentalSetup,SensorMeasurementsandModeling4.3.1ExperimentalSetupAforcetestingdevicewasbuiltforconditioningandcharacterizingthemicrofluidicpressuresen-sors.Weusedapneumaticcylinder(1.06DPSR02.0,ParkerHannifin)withaborediameterof1.0625inches(≈27mm)androddiameterof0.3125inches(≈8mm)andmountedtoarigidframeinaverticalposition.Acustom-builtmetallicforceconcentrator(6061Aluminum)of16mmindiameter(80%ofsensoractivearea)wasthreadedtoitsrodendinordertodistributethe 48 appliedpressureoverthesensortopsurface.Thesensorswereadheredtoaflatsurfaceundertheaircylinderrodend,whilekeepingitconcentrictothemicrochannelspiralshape(Figure4.6).Aminiaturizedpneumaticcontrollerboardwasusedtocontrolthepressureofacompressedairpipelinesource.Thewalloutletmaximumpressurewassetto80psiwithapneumaticfilterregu-lator.Theairpressureatthepneumaticcylinderwascontrolledbyasolenoidvalve(VQ110U-6M,SMCUSA)connectedtoaMOSFETswitch,whichwasmodulatedviaaprogrammablemicro-controller(ArduinoMega2560,Arduino)usingaPIDcontrolleralgorithmforpressuresetpointtracking.Sensormeasurementswerecollectedusingavoltagedividercircuit(Rref=47Ω)andconnectedtothesameprogrammingboardusingitsanalog-to-digitalconverterpinswiththede-faultinternalvoltageof5V.Twoinsulatedcoppertestleadswithalligatorclipswereusedtoconnectthepressuresensorterminalstothevoltagedividercircuit.AllsensordatawasrecordedviaserialcommunicationusingaPythonscriptrunningonaworkstationcomputerduringbothconditioningandcharacterizationprocedures.Thetotalpressureatthesensortopsurface,Psurf,wasdeterminedbyPsurf=Pgauge×AboreApuck=Frod2.01×10−4m2(4.1)wherePgaugeisthetotalpressureinsidetheaircylinder,Aboreistheborearea,Apuckistheareaoftheforceconcentrator,andFrodistheforcegeneratedattherod. 49 Figure4.6:Testrigformeasuringandcharacterizingthe3D-printedpressuresensors.Averticallymountedfluidiccylinderwithcustom-builtforceconcentrator(16mmdiameter)andcontrolledbyapneumaticpowersource.4.3.2PressureSensorResponseTherelativechangeinelectricalresistanceofthe3D-printedmicrochannelfilledwithliquidcon-ductor,∆R/R,wasrecordedasafunctionofthegeneratedpressureattheforceconcentrator,Psurf.Thesensorresponsewasfirstevaluatedbyapplyingastepsignalof30psiatthegauge(Psurf=0.6MPa).Theobtainedexperimentalresultshowsthatthesensorcanquicklyrespondtothepressureinput,butthentakesabout500secondstoreachthesteady-state,with∆R/R≈7.2asshowninFigure4.7.Thisobservedcreepcanbepartlyexplainedbytheviscoelasticityofthe3D-printedresinmaterialasdiscussedin[100]. 50 Figure4.7:Stepresponsecollectedfromthepressuresensorforaconstantinputof0.6MPa.Thetopgraphshowstheinputpressureandthebottomgraphshowsthemeasuredrelativechangeinresistance.Furthercomputationalstudywasperformedtoidentifyamodelwiththeseintrinsiccharacter-istics.Thesensordataobtainedduringthestepresponseexperimentwasimportedinasoftware(MATLAB,Mathworks)toestimateatransferfunctionmodel.Equation4.2showstheestimatedtransferfunctionmodelfromthetime-domaindatawithafittoestimateddataof93.23%andrep-resentedasasecondordersystemwithtwopolesandonezero.InFigure4.8,wecanobserveacomparisonbetweentheexperimentaldataandobtainedtransferfunctionstepresponses.Thestepresponsecharacteristicsoftheidentifiedmodelwerearisetimetr=150seconds,asettlingtimets=272seconds,andapeaktimetp=512seconds. 51 G(s)=∆R/RPsurf=1.406s+0.07097s2+0.4516s+0.005858(4.2)Figure4.8:Comparisonofastepresponseforboththeexperimentaldataandtheobtainedtransferfunctionthroughmodelfitting.Toanalyzethesensorresponseatmultiplefrequencies,asinusoidalwavewasgeneratedwithfrequencyfs=0.1Hz,0.25Hz,0.5Hzand1Hz,biasof30psiandamplitudeof20psiatthegauge(0.2MPa1000seconds).Althoughtheelectricalcharacteristicsofthesensorhavenotbeeninvestigatedduringthesim-ulationanalysis,theresistanceofthesensoroutputwascollectedatdifferentpressurevaluesforfurthercorrelationalstudybetweensimulationandexperimentalresults.Inordertomeasurethesensorresistanceclosetosteady-stateregimeatmultiplepressures,astaircasepressuresignalwithincrementsizeof0.1MPaanddurationof2000secondsperstepwasappliedtothecontrolboard.Anaveragevaluefortherelativechangeinresistanceateachpressurestepwascomputedforarangeof100pointsalongthesteady-stateregime(Figure4.12).BycombiningthesimulationresultsfromFigure4.5withtheexperimentalresultsfromFigure4.12,wehaveachievedacorre-lationbetweenthepressuresensoraveragestraininsidetheloadregionandtheobservedrelativechangeinresistanceforagivenappliedpressurevalue(Figure4.13). 53 Figure4.9:Thepressuresensoroutputundersinusoidalsensorinputwithrangefrom0.2to0.977MPa,withfourdifferentfrequencies(0.1Hz,0.25Hz,0.5Hz,and1Hz).Thedashedlinesshowthemeancurveofthecontinuousmeasurements. 54 Figure4.10:Sinusoidalresponseformultipleinputfrequencies(frameviewof20seconds).Eachmeasurementwascollectedfor>2000secondsat0.1Hz,0.25Hz,0.5Hz,and1Hz,withapressurerangeof0.2MPato0.977MPa. 55 Figure4.11:Graphforeachsinusoidalinputfrequencyafterreachingthesteady-state. 56 Figure4.12:Therelativechangeinresistanceversusthepressureinputforastaircaseinputsignal.Theaveragevalueof∆R/Rwascomputedforarangeof100pointsalongthesteady-stateregime,withapressureinputfrom0to1MPa,incrementsizeof0.1MPa,anddurationof2000secondsperstep. 57 Figure4.13:ThecorrelationbetweenthecomputedaveragestrainfromFEAsimulationatthesensoractiveareaandtherelativechangeinresistancemeasuredfromthephysicalpressuresensordevice.4.3.3FurtherDiscussionsDispersionoftheliquidsacrificialmaterialisstillachallengeprocessinthefabricationtechniquepresentedhere,sinceaccidentalformationofbeadscancausecloggingorirregularstructuredesignofthemicrofluidicchannels.Moreover,themanualremovaloftheliquidsupportmaterialthroughsuctionwithavacuumpump,andthemanualinjectionofliquidmetalusingsyringeshavevariedfabricationtimeduetonon-uniformtoolmanipulationandmaterialhandling.Awaytoimproveourfabricationmethodwouldbetomaketheseproceduresautomatedbythesame3D-printingmechanism. 58 Theintrinsicviscoelasticyoftherubber-likephotocurablematerialhasshownsomeimpactonthesoftpressuresensorresponsetime,takingupto500secondsforthesensortoreachasteady-stateregime.Additionalinvestigationsonsubstratesmadeofdifferentmixingratiosbetweensoftandrigid3D-printablephotopolymersandoverallthicknessesarerequiredinordertoanalyzetheirimpactonthesensorperformance.Encapsulationofthepressuresensorinletswasperformedbyusingsilverepoxyasaninterfacebetweentheliquidconductorandcopperstrandedwires.OthermethodshavebeentestedinitiallysuchasdepositionofuncuredAgilus30onthesensorterminalswiththincopperwiresattachedateachend,andcuringprocessusingUV-lightflashlight.However,poorqualitydepositionorcuringcausedleakageoftheliquidmetalwhensubjectingthesensortoveryhighpressurevalues.3D-printingofanencapsulationlayerhasalsobeentested,buttheattachedthinwiresandtheliquidmetalexposedsurfacemadeitachallengingprocessduetoblockageorundesiredcontaminationoftheprinterhead.Althoughtheselectedsilverepoxyshowedgreatadhesionandencapsulationproperties,furtherstudyisneededtoanalyzeitseffectonthesensorcharacteristics.Also,thesurfaceoxideskinonaliquidmetalcanaffecttheeffectivesurfacetensionandviscosity(non-sphericaldropletsformation),whichcanreduceitscontactwithothermaterialsandpotentiallyimpactingitselectricalproperties. 59 Chapter5ComputationalEvolutionofControlandTactilePerceptionforSoftRobots5.1EvolutionaryRoboticswithSoftRobotsAlthoughmanystudiescanbeperformedwithinasimulatedworldaboutrobotswithevolvedcon-trolsystemsormorphologies,anevolvedAImodelcanpotentiallybetransferredtoareal-worldrobottoimproveitsphysicalfunctionalities.Thelimitationsofavirtualenvironmentusedintheevolutionofrobotcontrollersandbody-plansaredeterminedbythesimulatorcapabilities.Areal-worldrobotwithmultipledegreesoffreedomthatwillinteractwithdifferentobjectsorobstaclesonitsworkspacerequiresasimulatorthatcancomputealargenumberofvariablesinordertocloselyrepresentthecharacteristicsoftheworldaroundit.Physicsenginesarecommonlyusedinevolutionarysystemssincetheyprovideacollectionoflibrariesthatcanperformreal-timecom-putationofrigidbodykinematics/dynamics,collisiondetection,mass-springsystems,andfluidmechanicsin2Dand3Dspaces[101].Asanexample,onestudyapproachusesaphysicsen-gine(NvidiasPhysX)tocoevolvethematerialpropertiesandlocomotivegaitsofcomplexsoftbodiesbuiltoutoftetrahedralmeshes[102].Inaddition,evolvingsoftrobotswithmultiplema-terialsandgenerativeencodingsuchacompositionalpattern-producingnetwork(CPPN)canlead 60 tosoft-voxelbasedsystems(VoxCAD)withalargediversityofcomplex,natural,multi-materialcreatures[58][60][103].Resultshaveshownthatthisevolutionaryprocesscangeneratesoftrobotswithdifferentmorphologiesandgaits,squishycreaturesthatcanreachorsqueezethroughtightapertures,andmultiplelocomotionstrategiesforspace-exploration.AninvestigationusingCovari-anceMatrixAdaptationEvolutionaryStrategy(CMA-ES)demonstratedthatevolvingcontrollersforaknifefish-inspiredsoftrobotisfeasibledirectlyonthephysicalrobot,whichwasabletoout-performahand-designedcontrollerintermsofrobottravelspeed[104].Sincesoftroboticsisstillagrowingfield,newsimulationtechniquesarebeingdevelopedtoovercomesomeoftheongoingchallengesinreal-timecomputationofsoftactuationandsensingmechanisms.Anopen-sourcesimulatorwithmulti-modelrepresentationshowsgreatpotentialforsimulat-ingsoftrobotsandtheirinteractionswiththeoutsideworld.TheSimulationOpenFrameworkArchitecture(SOFA)isacomputationallibraryprimarilytargetedforsimulatingtissues,muscles,organs,andbonesinmedicalapplications[105].Sinceprogrammingtheinteractionofrigidanddeformablematerialsrequiresvarioustechniquesingeometricmodeling,computationmechanics,numericalanalysis,collisiondetection,andrendering,theSOFAarchitecturewasbuiltbasedonahighlymodularframeworkthatfacilitatescollaborationbetweenresearcherswhilebeingabletofocusonacertaindomainofexpertise.Eachsimulatedobjectcanbedecomposedintomultiplesubcomponentsdescribingthemodelfeaturessuchasmeshtopology,mass,forces,integrationscheme,andconstraints,inascenegraphdatastructure.Inparticular,anewmethodwasstud-iedtocontrolsoftrobotsforobjectmanipulationorrollingmotionusingfiniteelementmethod(FEM),frictioncontacts,andquadraticprogramwithcomplementarityconstraintsinSOFA[106].Moreover,ageometriccomputingframeworkwhichpredictsthedeformationofcontinuumsoftbodiesundergeometricactuationscanbetestedbothinphysicalandsimulatedrobots,allowingrelationshipstudiesbetweenmaterialpropertiesandshapeparameters[107].Thedevelopmentofanefficientsoft-handsimulatorbasedinSOFAenabledthefeasibilitystudyofco-designingmor-phologyofsofthandsandtheircontrolstrategiesforgrasping,withcomputationtimefastenoughtosimulatemorethanamilliongraspsperday[108]. 61 Asaninitialsteptowardstheevolutionofcontrollersforsoftrobots,weexploredtheintegra-tionofstandardneuralnetworks(ANN,RNN)withtheSOFAplatformforcomputingthedesiredbehaviorofaSPAinaspecifictaskspace(curvature,positioning,orgrasping).Neuralnetworkscanbedescribedasmathematicalmodelsdefiningafunction,ordependingonitsstructuresize(numberoflayers),acompositefunctionwithmanysubfunctionswithinit.Somecomponentsofaneuralnetworkareneurons,connections,weights,biases,andactivationfunctions.Thestructureofagivenneuralnetworkcanbecomputationallyevolvedusinggeneticalgorithms.Sincethemathematicalmodelingofsoftrobotspresentsdifficultyatmanylevels(nonlinearmaterialprop-erties,parameteruncertainties),anditisacrucialprocessfordesigningcontrollersinrobotics,aneuralnetworkcanpotentiallybeusedforconvergingasimulatedsoft-bodiedrobottoatar-getedbehavior.Inonestudy,atypeofrecurrentneuralnetworkcalledalongshort-termmemory(LSTM)networkwasusedforlearningthetimeseriesmappingofredundantandunstructuredsensortopologyembeddedinasoftactuator,enablingthemodelingofasoftcontinuumactuatorinrealtimewhilebeingrobusttosensornonlinearitiesanddrift[109].Robustobjectmanipulationinunstructuredenvironmentsisachallengingprobleminroboticsduetotheuncertaintyassociatedwithcomplexandunpredictableenvironments,andwithobjectshapeandhardness.Conventionalobjectgraspingtechniquesinvolvepriorknowledgeoftherobotworkspacesuchas3Dmodelsofobjectsandmanipulators,aswellastheirweightandfrictionproperties.Thesemethodsarelimitedtoapplicationswheretherobotichandmodelisknownandtheobjectshapeiswell-definedorsimplifiedbygeometricprimitives.Althoughrigidarticulatedfingerscanassistwithpredictingaspecificgraspingconfiguration,thesemechanismscanbeex-pensivetofabricateandcontrol.Inaddition,theyarenotsuitableforgraspingandhandlingsoft,fragileobjects.Asanalternative,compliantrobotichandscanbefabricatedusingsoftmaterials,allowingadaptivitywhenhandlingirregularlyshapedordelicateobjects,andincreasingthetoler-ancewithuncertaintiesinperceptionandactuation.However,specificconfigurationatagiventimeishardtoknowduetothehandcompliance,requiringadvancedinternalsensingmechanismsandnovelsensory-motorinterfacemodelsforsoftrobots.Bytrainingathree-dimensionaldeepcon- 62 volutionalneuralnetwork(3DCNN)forgraspingunknownobjectswithsofthands,asoftrobotichandwasabletoestimatesuitablegraspposesfrommultiplegraspingdirectionsandwristorienta-tions,with87%successfulgraspingonpreviouslyunseenobjects[21].Proprioceptivecapabilitiesforsoftrobotichandscanbeachievedbyusingaclusteringalgorithmforautonomousobjectidentificationandforce-controlledgraspingwithposeuncertainty[22].However,thisapproachislimitedbyconstrainedsensordesignsandrequiresacamerainordertodetectapproximateobjectlocations.Inordertoevolvecontrollersforsoftrobotichands,whichcanpotentiallyidentifyobjecthard-nessorprovideposeestimationfordecisionmakingtasks,itisnecessarytouseamodelthatcanfacilitatetheswitchingbetweenmultipledesiredstates.AMarkovBrain(MB)canbeusedasthesubstrate,wheretherobotcontrollersarenetworksbuiltfromindividualcomputationalcompo-nents[110].Thesecomponentsinteractwitheachotherandtheoutsideworldandcanbedefinedasdeterministicorprobabilisticlogicgates,thresholdingfunctions,timersandcounters.Inaddi-tion,anMBiscomposedofastatebuffer(statevector)whereeachelementrepresentsonenode.AnMBupdateappliesasetofcomputationsdefinedbytheMBcomponentsinthisbuffer.Mul-tipleinputscanbewrittenintoapartofthisbufferbeforetheMBupdate,andoutputscanberetrievedfromasubsetofthisbufferonceanMBcomputationiscompleted.ThestateofanMBcanbechangedmuchmoreeasily,whichmakesMBabetterapproachforfine-grainedbehavior,andwell-suitedfortheintegrationoftemporalinformation[111][112].Inthischapter,wehaveinvestigatedtheevolutionofMBsusingtheModularAgentBasedEvolver(MABE),whichisamodularandreconfigurabledigitalevolutiontoolforbiologyandengineeringresearch[113].Furthermore,multipleAImodelssuchasANNs,CartesianGeneticProgramming(CGPs),andMBshavebeenevolvedinMABEtoproperlyclassifytheshapeandstiffnessofgraspedobjectsaswellasthesofthandparameters.Theseartificialbrainscanpo-tentiallybetransferredtoreal-worldsoftrobotstoevaluatetheirtaskperformanceandfacilitateprototyping.Evolvedcontrollersforagivencoupledbody-plan/actuationmechanismwillcon-tributetobuildingmoreadaptiveandresilientsoftmachinesinthenearfuture. 63 5.2EvolvingMarkovBrainControllersforRoboticGraspingThisworkinvolvestheevolutionofproprioceptivesensing-capablecontrollersforsoftroboticgrippersusingMBsforpotentialtransferringtheevolvedcontrollerstoreal-worldsoftrobots.WefirstinvestigatedtheimplementationofanFEM-basedcurvaturefeedbackcontrolofaSPAusingaPIDcontrollerandtheSOFAsimulator.ThecontrolalgorithmrunsinparallelwiththeFEMsimulation,whichcomputesthesoftactuatormeshdeformationinrelationtoitschamberinnerpressure.Thecurvatureiscalculatedwithrespecttothefixed-endanddistal-endmeshnodes(centerline)alongtheSPAbottomsurface(inextensiblelayer).ThePIDgainsforthisexperimentweretunedbasedontrialanderror.AsshowninFigure5.1,thedesiredSPAcurvatureisachievedforthegivensetpointvalue.Figure5.1:AnFEM-basedreal-timecurvaturefeedbackcontrolofaSPA.TheresultswereachievedusingaPIDcontrollerandFEM-basedsimulationinSOFAwithdifferentmaterialprop-erties.WehavesimulatedmultiplesoftpneumaticgripperswithdifferentnumbersoffingersusingtheSOFTROBOTSplugin(DEFROSTTeam,INRIA)forSOFAandmanuallycontrolledthemwithaHuman-MachineInterface(HMI)(seeFigure5.2a-d).AllgeometrieshavebeensimulatedwithtetrahedronmeshesusingtheTetrahedronFEMForceFieldcomponentandcontactdetection 64 andhandlingthroughthedefaultcollisionpipeline.Inaddition,aPythonscriptwasdevelopedtocontrolthesoftgripperinreal-time,varyingthesimulatedinternalpressurebasedonthefeedbackfrombendingsensorsembeddedwithinasoftwearableglove.Thesensoranalogsignalswereprocessedbyamicrocontroller(ArduinoUnoRev3,Arduino)andconvertedtograspingactionsviaserial(USB)communication.(a)(b)(c)(d)Figure5.2:SoftroboticgrippermanualcontrolusingHuman-MachineInterface(HMI)withSOFA.(a)Asoftwearableglovewithbendingsensorsconnectedtoamicrocontrollerforpro-cessingtheanalogsignals,(b)softgripperwiththreefingersinunactuatedstate;(b)softfingerstouchinganobjectwithcollisiondetection;(c)objectbeingliftedbythesoftgripperwithpinchinggraspingmode.Eventhoughexternalinputsfromwearabledeviceswerecapableofcontrollingasimulatedsoftroboticgripperinreal-time,additionalinformationfromthesimulationenvironmentisre-quiredforachievingcontrollerswithproprioceptivefeedback.Forachievingsuchagoal,wehavedefinedaclassificationtaskchallengeforevolvingartificialbrainscapableofdeterminingdifferentconditionsfromthesimulationenvironment.Inordertodetectanobjectshapeandstiffness,force 65 valuesonspecifiedmeshnodes(seeFigure5.3)weremeasuredatthesoftpneumaticactuatortipandsavedtoaspreadsheetfileduringeachiteration.Thesimulationruntimewassetto1.25sec-onds,withatimestepof10millisecondsandanincrementsizeof0.01fortotalinternalsurfacepressure.Thefollowingvalueswererecordedinthefile:currentsimulationtime,softfingerinter-nalpressurevalue,andforcevaluesatXYZdirectionsforallthreeselectednodes.These11datatypeswereusedasinputsfortrainingartificialbrainscapableofperformingaclassificationtask.Figure5.3:Softfingermeshnodelocationsforprobingforcevaluesduringeachiterationstepinthesimulatedworkspace.ThefollowingmeshnodenumbersattheSPAtipwereselected:1,30,and59.Thefirstapproachwastotesttheclassificationofthreedistinctsoftgraspingcharacteristics:objectshape,objectstiffness,andthenumberofcontactfingers.Inordertofacilitatetheevolu-tionaryprocess,wehaveconstrainedtheproblemtotwodifferentobjectshapes(cubeorsphere),threestiffnesslevels(soft,medium,orhard),andtwotypesofcontacts(oneortwofingers).ThestiffnessvaluesforeachsimulatedcomponentarepresentedinTableA4.Figure5.4showsthesoftfingersinteractionforacubeobject,andtherespectiveemulatedforcemeasurementsarepresentedinFigure??.Similarly,Figure5.6showsthesoftfingersinteractionforasphereobject,andtherespectiveemulatedforcemeasurementsarepresentedinFigure??.WehaveusedtheModularAgent-BasedEvolver(MABE)platformtoevolvemultipleartificialbrainsusingdifferenttypes 66 ofconnectinglogicssuchasstandardANNs,MarkovBrains,andCartesianGeneticProgramming(CGPs).Theclassificationtaskwasevolvedoffline,withallforcemeasurementsstoredinspread-sheetsbeingloadedtotheartificialbrainforeachevaluationcase.Theloadedfilesfollowedanamingconventionaccordingtothesimulatedconditions(objecttype,numbersoffingers,andobjectstiffness),witheachconditionrepresentedbyabitvalue(TableA5).Thenumberofinputneuronsandoutputneuronsweredefinedbasedonthenumberofdatacolumnsinaspreadsheet(11inputs)andthenumberofconditions(5outputs),respectively.Ascoringwassetas1pointforeachcorrectoutput,withamaximumof5pointsinclassifyingeachcasecorrectly.Themaximumfitnessvalueachievedinasolutionwasdefinedas120,where:MaximumFitness=12cases×2runs×5conditions=120(5.1)Figure5.4:Softfingersatfinalsimulationtimestep(maximuminnerchamberpressure)incontactwithacubicobjectwithdifferentstiffnessvalues.Singlesoftfingerincontactwith(a)soft,(b)medium,and(c)hardcube.Adualsoftfingergripperincontactwith(d)soft,(e)medium,and(f)hardcube. 67 (a)(b)Figure5.5:Forcevalues(XYZdirections)attheSPAtipincontactwithacube.Probedmeshnodesforasinglefingerandadualfingergripperincontactwithasoft(a-b),medium(c-d),andhard(e-f)cubicobject. 68 Figure5.5(cont’d)(c)(d) 69 Figure5.5(cont’d)(e)(f) 70 Figure5.6:Softfingersatfinalsimulationtimestep(maximuminnerchamberpressure)incontactwithasphericalobjectwithdifferentstiffnessvalues.Singlesoftfingerincontactwith(a)soft,(b)medium,and(c)hardsphere.Adualsoftfingergripperincontactwith(d)soft,(e)medium,and(f)hardsphere. 71 (a)(b)Figure5.7:Forcevalues(XYZdirections)attheSPAtipincontactwithasphere.Probedmeshnodesforasinglefingerandadualfingergripperincontactwithasoft(a-b),medium(c-d),andhard(e-f)sphericalobject. 72 Figure5.7(cont’d)(c)(d) 73 Figure5.7(cont’d)(e)(f) 74 Inthisneuroevolutionprocedure,wehaveselectedTournamentSelectionmethodastheopti-mizertype,withatournamentsizeof25.Asaninitialtest,wehavedefinedapopulationsizeof100agentswithallcandidatesolutionscompetingagainsteachotherfor1,000generations.Theinitialgenomesizewasdefinedas5,000,withaminimumsizeof2,000andmaximumsize20,000.Copyanddeletionmutationshadaninsertrateof0.00002andthepointmutationratewassetas0.015(selectgenomesiteandrandomizeitsvalue).InMABE,theMBswereconfiguredwith8hiddennodesandastartingnumberofgatesof6.Inaddition,theirinternalgatesweresetasTRITtypes,whereoutputscanhavevaluesof1,0,or-1,andGPtypes,whereoutputsarecontinuousval-uesgeneratedviamathematicaloperationsandfunctions.Acustom-builtworld(interfacebetweenbrainsandsimulatedstates)wasdesignedinMABEtoimportthespreadsheetscorrespondingtoeachclassificationcaseandevaluatetheentirepopulationofagents(artificialbrains).Theevolu-tionaryprocessfollowedasequenceofiterationswherebraininputsweremappedtoworldstates(spreadsheetlines),followedbybrainupdates(inputbuffer).Aftercompletingthefilereadingforaspecificcaseascorewascalculated(outputbuffer).Onceallfileswereevaluated,afinalfitnessvaluewascomputedfortheevolvedartificialbrain.Finally,geneticalgorithmoperationswereperformedbasedonpresetmutationrateparameters.AnewsequenceofevaluationswasinitiatedwithanewpopulationobtainedfromtheTournamentSelectionoptimizationprocess(Algorithm1).Theevolvedsolutionsof50replicateswereexportedfromMABEwithcomputedscoresofthefittestindividualinthepopulationforevery10generations.Figure5.8showstheevolutionaryprogressionforeachbrainoutputneuron(5environmentconditions)inall12cases,witha95%confidenceagainstreplicates.TheevolvedstandardANNartificialbrainshaveachievedthehighestclassificationscoreforallconditions.AlthoughtheMBclassifiershavenotreachedhigherscoresthanANNs,somereplicateshaveapproachedsimilarsolutionsforsomeconditionssuchasdetectingbetweenoneortwocontactfingers(Figure5.9e).ThefitnessvalueduringeachupdatewascalculatedusingEquation5.1,anditsevolutionaryhistoryisshowninFigure5.10.Themaximumevolvedfitnessscorewas108pointsforasinglereplicateofANNbraintype.Sincebettersolutionsareusuallyidentifiedwithextensivecomputationalruns, 75 alongerneuroevolutionattemptwastakentofurtherinvestigatetheclassificationperformanceofallbraintypes.Inthisadditionaltrials,wehaverunthealgorithmfor40,000generationswithoutmodifyinganysimulationsettings.AsshowninFigure5.11,thefinalfitnessvaluewasMBw≈0.82,CGPw≈0.85,andANNw≈0.86. 76 (a)(b)Figure5.8:Evolutionaryprogressionforeachbrainoutputneuron(averageofall12cases),witha95%confidenceagainstreplicates.Eachgraphshowstheclassificationperformanceofacertainenvironmentcondition:(a)objectshape,(b-d)objectstiffness,and(e)numbersofcontactfingers. 77 Figure5.8(cont’d)(c)(d) 78 Figure5.8(cont’d)(e) 79 Figure5.9:Fitnessvalueforeachevolvedartificialbraintype:ANNs,MarkovBrains,andCGPs.Theaveragescorewascomputedfromall12casesandtheirrespectivereplicates. 80 Figure5.10:Longerneuroevolutionofsoftgraspingclassificationtaskwithvariedworkspaceconditionsfor40,000generations.5.3DiscussionsandFutureWorkWehavedemonstratedsomeinitialstepsaswellaspreliminaryresultsforachievingreal-timesimulationofsoftactuatorswithsatisfactorycomputationtimeandevolutionofartificialbrainsfor 81 classificationtasksinsoftgraspingusingdifferentAImodels.TheSOFAsimulationtoolallowedtheuseofaPIDalgorithmtocontrolthebendingofaSPAbasedonitsbottomsurfacecurvature.Moreover,wehaveshownthatitcanbeusedincombinationwithanHMIdevicesuchasasoftwearableglovetocontrolasoftroboticgripperforgraspingandmanipulatingobjectsinasim-ulatedenvironment.Inaddition,wehaveexportedforcemeasurementsfromthetipofmultiplesoftfingerswhenincontactwithobjectsofvariedshapesandstiffnessesandusedthisdatatotrainartificialbrainscapableofclassifyingdiverseconditionsinaroboticgripperworkspace.SimulateddatawasimportedintoMABEplatformforevolvingstandardANNs,MarkovBrains,andCGPsmodelstoclassifythreedifferentsimulatedconditions,objectshape,objectstiffness,andthenum-beroffingersincontactduringgrasping.Obtainedresultsthroughneuroevolutionshowedaccept-ableclassificationcapabilitiesforallevolvedbraintypes.Aneuroevolutionof40,000generationsreachedaleveloffitnessupto0.85(maximumvalueof1)forstandardANNs,whileMarkovBrainsmodelsfoundsolutionswithfitnessvaluearound≈0.8.Whiletheseresultspresentedgreatclassificationcharacteristicsinevolvedartificialbrainsusingofflinemethod(nointerfacewithphysicsengine),additionalinvestigationisneededinordertoachievefullyproprioceptivegrasp-ingwithreal-timefeedbacksensing.Anintelligentsoftgraspingtechniqueshouldbeabletorecognizeobjectshapesandstiffnessesbyinteractingwithcertainobjectsonstrategicpointsthatprovideenoughinformationaboutitsworkspace.Moreover,theserepresentationsofelementsoftheenvironmentwithinwhichthesoftgripperoperateswouldprovideadaptivebehaviorduringdecision-makingtasks.Aroadmapforachievingsuchgoalsshouldinvolveexhaustiveneuroevolutiontrainingforthousandsofgenera-tionswithacertainlevelofcontrolofthesoftgrippersbytheartificialbrains.First,aninterfacebetweenSOFAandMABEisrequiredtolinkoutputneuronstoagroupofactionslikeperforminggripperrotationsandindividualcontrolofthepressurelevelforeachsoftfinger.Second,variouspiecesofinformationfromtheworkspaceshouldbepassedtoinputneuronsfordetectingappliedforcesandchangesincurvaturewhilemanipulatingparticularobjectsineachsimulationiteration.Arobustartificialbrainshouldbeabletodecidewhethertoprovideafewpinchingactionsor 82 completelyenvelopetheobjectofinterestinordertoinferattributesaboutit.Furthermore,priorknowledgeofanobjectcharacteristicswouldopenuptheprospectforrobotstoefficientlyhandledelicateobjectsorusetoolsforspecifictasks.Futureworkshouldincludebothevolutionofcon-trollersforsoftrobotichandsinsimulatedenvironmentaswellastransferabilityofartificialbrainswithsuccessfulsolutionstoreal-wordsettings,bridgingthegapbetweensimulationandreality.Theseenhancedmanipulationskillswouldenablerobotstohavedexterityandadaptativegraspingincomplexworkspacesalongsidehumanbeings. 83 Chapter6SummaryandDiscussionSoftroboticdevicescanachievesuitablegraspingperformancewithlowcostmaterialsandsimpleactuationmechanisms,whileprovidingadaptivityforunstructuredobjects.Inthisthesis,wehaveinvestigatedthedesign,fabricationandsimulationofsoftactuatorswithdifferentgeometriesandmaterials.Ourgoalwastoimprovesoftroboticgraspingthroughacombinationofsoftfingers,stretchablesensors,andintelligentobjectclassification,inordertoreachastepclosertodexterity,stabilityandfinemanipulationinvariedscenarios.Anovelapproachtostiffnessmodulationofsoftactuatorswasachievedbyusinga3D-printedCPLA-basedcomponentintegratedtoaSPAforvariablestiffnesscontrolviaelectricalsignals.ThisallowedmodulationoftheSPAhardnessatdifferentjointswhilegraspingobjectswithir-regularstructure.Atwo-fingeredgripper,composedoftwoCPLA-embeddedSPAs,demonstratedtheabilitytochangethegraspingposturetosuittheshape,size,andtextureoftheobjectsbeinggrasped.Also,theSPAshowedthatitcouldbeeffectivelylockedinadesiredbendingconfigura-tionwhilecarryingaweight,upto800g,evenintheabsenceofpressureorvoltageinput.ThisworkcanleadtofeedbackcontrolofstiffnessmodulatedsoftgripperstoachieveamorearbitrarydesiredshapefortheSPAunderafixedpressureinput,andrealizeanevenbroaderrangeofshapechangesforhandlingdifferentobjects.Moreover,anoveltypeofCNT-basedstrainsensorswaspresented,whichcanbeembedded 84 inthestructureofasoftroboticcomponent.Thisarrayofstraingaugeswasusedasadistributedsensornetworkalongwiththecompliantmechanism.Thesensorsizeandshapecanbecustomizedformanydifferentapplications.WeperformedexperimentswiththeCNT-basedstrainsensorfordetectingdeformationatdifferentlocationsatthebottomsurfaceofanSPA.Bothsensormeasure-ments(straindata)andcamerarecording(imagedata)agreedwiththecompression(curvature)observedinthebottomlayeroftheSPAduringpositivepressure.Theresultsprovidedinitialstepsintheimplementationofasensorarrayformonitoringlocaldeformationonasoftroboticmechanism.Furthermore,wehavepresentednewmethodsforachieving3D-printedstretchablepressuresensorsandconductorsusingliquidmetalasacircuitcomponentforpotentialuseinrobotichands.APolyJet3D-printerwasusedtocreateamicrochannelstructureinsidethesoftsubstrateincombinationwithaviscousliquidmixtureforthesacrificiallayer.Functionalstraight-shapedmicrochannelswerefabricatedwithacross-sectionalareadownto150×150µm.Aspiral-shapedpressuresensorwasdesignedwith350×350µmmicrochannelcross-sectionandinjectedwithliquidmetalmanuallyusingsyringes.Experimentalresultsshowedthatthemultimaterial-basedsensorwiththemixtureofAgilus30andVeroClear(70AShoreHardness)andoverallthicknessof1.5mmwasabletowithstandhighpressuresupto1MPa.Thispressuresensorissuitableforapplicationsthatrequireresistancetoveryhighdeformationssuchasinmodernelectronicsforseveralfieldsandindustry,includingwearableorimplantabledevices,militaryandsoftrobotics.Finally,wehaveperformedpreliminarycomputationalexplorationofintelligentgraspingus-ing3Dphysicsenginesandtoolsforevolvingandanalyzingdigitalbrains.Inparticular,wehaveanalyzedaclassificationtaskofsoftgraspedobjectsthroughneuroevolutionprocessesofvariousartificialbrains.SimulationwiththeSOFAframeworkhasbeenconductedtoproducetheemu-latedcontactforcemeasurementsduringreal-timepressurecontrolofsoftfingers.TheobtaineddatawasusedtotrainartificialneuralnetworksbyusingtheMABEplatformtoproperlyclas-sifytheshapeandstiffnessofthegraspedobjects.SeveralAImodelshavebeencomputationallyevolved,suchasstandardANNs,MarkovBrains,andCGPs,toclassifythreedifferentsimulated 85 softgraspingconditions:object’sshape,object’sstiffnesslevel,andthenumberoffingersincon-tactduringgrasping.Allevolvedbraintypesshowedacceptableclassificationcapabilitiesusinganeuroevolutionofflinemethod,whichfoundsolutionshigherthan0.8fitnessvalueafter40,000generations.Theseinitialresultsprovidedguidelinestogeneratebrainsthatcanpotentiallystorerepresentationsofelementsfromthesoftgripperworkspaceandassistondexterityandadaptivebehaviorduringdecision-makingtasksincomplexenvironments.Weexpecttheinvestigationfromthisdissertationtocontributenewcontrolalgorithmsandfabricationmethodsthatsynergisticallyusetactilefeedback,multimodalobjectrecognition,andstablegraspestimation,toenhancetheperformanceofdexterousmanipulationbysoftrobotichandsandgrippersoperatinginthehighlyvaryingenvironmentsoftherealworld. 86 APPENDIX 87 APPENDIXTableA1:Hyperelasticconstitutivemodelsfordescribingthemechanicalbehaviorofincompress-iblerubbermaterials.HyperelasticmaterialmodelsStrain-energyfunctionsNeo-HookeanΨ=C1(I1−3)ShearModulus:µ=2C1YeohΨ=P3i=1Ci(I1−3)iShearModulus:µ=2C1Mooney-RivlinΨ=P2i=1Ci(Ii−3)ShearModulus:µ=2(C1+C2)OgdenΨ=PNi=1µiαi(λαi1+λαi2+λαi3−3)ShearModulus:µ=12PNi=1µiαiArruda-BoyceΨ=C1P5i=1αiβi−1(Ii1−3i)ShearModulus:µ=C1(1+35λ2m+99175λ4m+513875λ6m+4203967375λ8m)β=1λ2m;α1=12;α2=120;α3=111050;α4=197000;α5=1673750TableA2:Mechanicalpropertiesofrubber-likematerialsusedinthefabricationofsoftactuators,flexiblesensorsandstructuresforsoftrobots.MaterialDensityDurometerTensileStrengthTearStrengthElongationatBreak(g/cm3)(Shore)(N/mm2)(N/mm)(%)Smooth-OnEcoflex00-301.0730-001.386.65900Smooth-OnDragonSkin301.0830A3.4518.90340DowCorningSylgard1841.0344A7.102.00120 88 TableA3:ThedimensionsofpartsinthefabricatedSPAwithembeddedCPLA.ParameterDimensionSPAlength140mmSPAwidth20mmSPAwallthickness2mmCPLAlength135mmCPLAwidth16mmCPLAthickness3mmAnti-slipfeaturethickness2mmTableA4:StiffnessvaluesofeachcomponentintheSOFAsimulation.ComponentYoungModulus(MPa)SoftFingerBellows500SoftFingerInextensibleLayer1500GraspedObject=Soft1GraspedObject=Medium10GraspedObject=Hard100 89 TableA5:Thebinaryvalueofeachcorrespondentworkspaceconditioninthesimulationenviron-mentofSOFA.ConditionBinaryValueObjectType=Cube0ObjectType=Sphere1ObjectStiffness=Soft0,0,1ObjectStiffness=Medium0,1,0ObjectStiffness=Hard1,0,0NumberofContactFingers=10NumberofContactFingers=21 90 Algorithm1:FitnessEvaluationandGeneticOptimizationinMABERequire:SimulatedSPAmeasurementsfromSOFAsoftwareInput:SpreadsheetfileswithforcevaluesinXYZdirectionsateachmeshnodeOutput:Artificialbrainsgenomesequenceandtheirrespectivefitnessscores1initialization;2whilei≤maximumnumberofgenerations(updates)do3whilej≤maximumnumberofcases(files)do4getfile;5resetbrain;6whilenotendofj-filedo7readlineatj-file;8updatebrain;9end10calculateoutputscore;11end12computefitnessvalue;13runTournamentSelectionoptimizer;14end 91 BIBLIOGRAPHY 92 BIBLIOGRAPHY[1]D.RusandM.T.Tolley,“Design,fabricationandcontrolofsoftrobots,”Nature,vol.521,no.7553,pp.467–475,2015.[2]K.Suzumori,S.Iikura,andH.Tanaka,“Developmentofflexiblemicroactuatoranditsapplicationstoroboticmechanisms,”inProceedingsofthe14thInternationalConferenceontheSynthesisandSimulationofLivingSystems(ALIFE),1991,pp.1622–1627.[3]D.Trivedi,D.Dienno,andC.D.Rahn,“Optimal,model-baseddesignofsoftroboticma-nipulators,”JournalofMechanicalDesign,vol.130,no.9,p.091402,2008.[4]M.Otake,Y.Kagami,M.Inaba,andH.Inoue,“Motiondesignofastarfish-shapedgelrobotmadeofelectro-activepolymergel,”RoboticsandAutonomousSystems,vol.2-3,pp.185–191,2002.[5]A.D.Marchese,C.D.Onal,andD.Rus,“Autonomoussoftroboticfishcapableofescapemaneuversusingfluidicelastomeractuators,”SoftRobotics,vol.1,no.1,pp.75–87,2014.[6]E.W.Hawkes,L.H.Blumenschein,J.D.Greer,andA.M.Okamura,“Asoftrobotthatnavigatesitsenvironmentthroughgrowth,”ScienceRobotics,vol.2,no.8,p.eaan3028,2017.[7]S.W.Kuffler,J.G.Nicholls,andA.R.Martin,Fromneurontobrain:Acellularapproachtothefunctionofthenervoussystem,2nded.SinauerAssociates,1984.[8]C.EyzaguirreandS.W.Kuffler,“Processesofexcitationinthedendritesandinthesomaofsingleisolatedsensorynervecellsofthelobsterandcrayfish,”TheJournalofGeneralPhysiology,vol.39,no.1,pp.87–119,1955.[9]P.Rabischong,“Phylogenyofthehand,”TheHand,pp.3–7,1981.[10]R.Malek,“Thegripanditsmodalities,”TheHand,pp.469–476,1981.[11]V.B.Mountcastle,Thesensoryhand:neuralmechanismsofsomaticsensation,1sted.HarvardUniversityPress,2005.[12]B.Boom.Mechanoreceptorsintheskin.[Online].Available:https://biologyboom.com/wp-content/uploads/2014/07/b30m1022l.jpg[13]E.Naito,P.E.Roland,andH.H.Ehrsson,“Ifeelmyhandmoving:anewroleoftheprimarymotorcortexinsomaticperceptionoflimbmovement,”Neuron,vol.36,no.5,pp.979–988,2002.[14]E.NaitoandH.H.Ehrsson,“Somaticsensationofhand-objectinteractivemovementis 93 associatedwithactivityintheleftinferiorparietalcortex,”JournalofNeuroscience,vol.26,no.14,pp.3783–3790,2006.[15]M.R.Cutkosky,Roboticgraspingandfinemanipulation,1sted.KluwerAcademicPub-lishers,1985.[16]S.C.Jacobsen,E.K.Iversen,D.Knutti,R.T.Johnson,andK.B.Biggers,“DesignoftheUtah/MITdextroushand,”inRoboticsandAutomation(ICRA),1986IEEEInternationalConferenceon,1986,pp.1520–1532.[17]E.Brown,N.Rodenberg,J.Amend,A.Mozeika,E.Steltz,M.R.Zakin,H.Lipson,andH.M.Jaeger,“Universalroboticgripperbasedonthejammingofgranularmaterial,”Pro-ceedingsoftheNationalAcademyofSciences,vol.107,no.44,pp.18809–18814,2010.[18]K.C.Galloway,K.P.Becker,B.Phillips,J.Kirby,S.Licht,D.Tchernov,R.J.Wood,andD.F.Gruber,“Softroboticgrippersforbiologicalsamplingondeepreefs,”SoftRobotics,vol.3,no.1,pp.23–33,2016.[19]P.Polygerinos,Z.Wang,K.C.Galloway,R.J.Wood,andC.J.Walsh,“Softroboticgloveforcombinedassistanceandat-homerehabilitation,”RoboticsandAutonomousSystems,vol.73,pp.135–143,2015.[20]H.In,B.B.Kang,M.Sin,andK.-J.Cho,“Exo-Glove:awearablerobotforthehandwithasofttendonroutingsystem,”IEEERobotics&AutomationMagazine,vol.22,no.1,pp.97–105,2015.[21]C.Choi,W.Schwarting,J.DelPreto,andD.Rus,“Learningobjectgraspingforsoftrobothands,”IEEERoboticsandAutomationLetters,vol.3,no.3,pp.2370–2377,2018.[22]B.S.Homberg,R.K.Katzschmann,M.R.Dogar,andD.Rus,“Robustproprioceptivegraspingwithasoftrobothand,”AutonomousRobots,pp.1–16,2018.[23]C.Duriez,“Controlofelasticsoftrobotsbasedonreal-timefiniteelementmethod,”in2013IEEEinternationalconferenceonroboticsandautomation.IEEE,2013,pp.3982–3987.[24]A.D.Marchese,K.Komorowski,C.D.Onal,andD.Rus,“Designandcontrolofasoftandcontinuouslydeformable2Droboticmanipulationsystem,”in2014IEEEinternationalconferenceonroboticsandautomation(ICRA).IEEE,2014,pp.2189–2196.[25]V.Vikas,P.Grover,andB.Trimmer,“Model-freecontrolframeworkformulti-limbsoftrobots,”in2015IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems(IROS).IEEE,2015,pp.1111–1116.[26]N.LuandD.-H.Kim,“Flexibleandstretchableelectronicspavingthewayforsoftrobotics,”SoftRobotics,vol.1,no.1,pp.53–62,2014.[27]D.Qi,K.Zhang,G.Tian,B.Jiang,andY.Huang,“Stretchableelectronicsbasedonpdms 94 substrates,”AdvancedMaterials,p.2003155,2020.[28]H.Wu,T.W.Odom,D.T.Chiu,andG.M.Whitesides,“Fabricationofcomplexthree-dimensionalmicrochannelsystemsinPDMS,”JournaloftheAmericanChemicalSociety,vol.125,no.2,pp.554–559,2003.[29]S.Sareh,A.Jiang,A.Faragasso,Y.Noh,T.Nanayakkara,P.Dasgupta,L.D.Seneviratne,H.A.Wurdemann,andK.Althoefer,“Bio-inspiredtactilesensorsleeveforsurgicalsoftmanipulators,”in2014IEEEInternationalConferenceonRoboticsandAutomation(ICRA).IEEE,2014,pp.1454–1459.[30]C.Suh,J.C.Margarit,Y.S.Song,andJ.Paik,“Softpneumaticactuatorskinwithembeddedsensors,”in2014IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems.IEEE,2014,pp.2783–2788.[31]G.M.Whitesides,“Theoriginsandthefutureofmicrofluidics,”Nature,vol.442,no.7101,p.368,2006.[32]D.M.Vogt,Y.-L.Park,andR.J.Wood,“Designandcharacterizationofasoftmulti-axisforcesensorusingembeddedmicrofluidicchannels,”IEEEsensorsJournal,vol.13,no.10,pp.4056–4064,2013.[33]J.-B.Chossat,H.-S.Shin,Y.-L.Park,andV.Duchaine,“Softtactileskinusinganembeddedionicliquidandtomographicimaging,”JournalofMechanismsandRobotics,vol.7,no.2,p.021008,2015.[34]J.-B.Chossat,Y.-L.Park,R.J.Wood,andV.Duchaine,“Asoftstrainsensorbasedonionicandmetalliquids,”IEEESensorsJournal,vol.13,no.9,pp.3405–3414,2013.[35]B.E.SchubertandD.Floreano,“Variablestiffnessmaterialbasedonrigidlow-melting-point-alloymicrostructuresembeddedinsoftpoly(dimethylsiloxane)(PDMS),”RscAd-vances,vol.3,no.46,pp.24671–24679,2013.[36]M.H.Zarifi,H.Sadabadi,S.H.Hejazi,M.Daneshmand,andA.Sanati-Nezhad,“Non-contactandnonintrusivemicrowave-microfluidicflowsensorforenergyandbiomedicalengineering,”ScientificReports,vol.8,no.1,pp.1–10,2018.[37]A.Koh,D.Kang,Y.Xue,S.Lee,R.M.Pielak,J.Kim,T.Hwang,S.Min,A.Banks,P.Bastienetal.,“Asoft,wearablemicrofluidicdeviceforthecapture,storage,andcolori-metricsensingofsweat,”Sciencetranslationalmedicine,vol.8,no.366,pp.366ra165–366ra165,2016.[38]J.Choi,D.Kang,S.Han,S.B.Kim,andJ.A.Rogers,“Thin,soft,skin-mountedmi-crofluidicnetworkswithcapillaryburstingvalvesforchrono-samplingofsweat,”Advancedhealthcarematerials,vol.6,no.5,p.1601355,2017.[39]J.Choi,R.Ghaffari,L.B.Baker,andJ.A.Rogers,“Skin-interfacedsystemsforsweat 95 collectionandanalytics,”ScienceAdvances,vol.4,no.2,p.eaar3921,2018.[40]X.Wang,R.Guo,andJ.Liu,“Liquidmetalbasedsoftrobotics:materials,designs,andapplications,”AdvancedMaterialsTechnologies,vol.4,no.2,p.1800549,2019.[41]M.D.Dickey,R.C.Chiechi,R.J.Larsen,E.A.Weiss,D.A.Weitz,andG.M.Whitesides,“Eutecticgallium-indium(egain):aliquidmetalalloyfortheformationofstablestructuresinmicrochannelsatroomtemperature,”AdvancedFunctionalMaterials,vol.18,no.7,pp.1097–1104,2008.[42]M.D.Dickey,“Stretchableandsoftelectronicsusingliquidmetals,”AdvancedMaterials,vol.29,no.27,p.1606425,2017.[43]Y.Meng¨uc¸,Y.-L.Park,E.Martinez-Villalpando,P.Aubin,M.Zisook,L.Stirling,R.J.Wood,andC.J.Walsh,“Softwearablemotionsensingsuitforlowerlimbbiomechan-icsmeasurements,”in2013IEEEInternationalConferenceonRoboticsandAutomation.IEEE,2013,pp.5309–5316.[44]Y.Meng¨uc¸,Y.-L.Park,H.Pei,D.Vogt,P.M.Aubin,E.Winchell,L.Fluke,L.Stirling,R.J.Wood,andC.J.Walsh,“Wearablesoftsensingsuitforhumangaitmeasurement,”TheInternationalJournalofRoboticsResearch,vol.33,no.14,pp.1748–1764,2014.[45]J.-B.Chossat,Y.Tao,V.Duchaine,andY.-L.Park,“Wearablesoftartificialskinforhandmotiondetectionwithembeddedmicrofluidicstrainsensing,”in2015IEEEinternationalconferenceonroboticsandautomation(ICRA).IEEE,2015,pp.2568–2573.[46]F.L.Hammond,R.K.Kramer,Q.Wan,R.D.Howe,andR.J.Wood,“Softtactilesensorarraysforforcefeedbackinmicromanipulation,”IEEESensorsJournal,vol.14,no.5,pp.1443–1452,2014.[47]Y.-L.ParkandR.J.Wood,“Smartpneumaticartificialmuscleactuatorwithembeddedmicrofluidicsensing,”inSENSORS,2013IEEE.IEEE,2013,pp.1–4.[48]J.Morrow,H.-S.Shin,C.Phillips-Grafflin,S.-H.Jang,J.Torrey,R.Larkins,S.Dang,Y.-L.Park,andD.Berenson,“Improvingsoftpneumaticactuatorfingersthroughintegrationofsoftsensors,positionandforcecontrol,andrigidfingernails,”in2016IEEEInternationalConferenceonRoboticsandAutomation(ICRA).IEEE,2016,pp.5024–5031.[49]Y.Hao,T.Wang,Z.Xie,W.Sun,Z.Liu,X.Fang,M.Yang,andL.Wen,“Aeutectic-alloy-infusedsoftactuatorwithsensing,tunabledegreesoffreedom,andstiffnessproperties,”JournalofMicromechanicsandMicroengineering,vol.28,no.2,p.024004,2018.[50]C.Ladd,J.-H.So,J.Muth,andM.D.Dickey,“3dprintingoffreestandingliquidmetalmicrostructures,”AdvancedMaterials,vol.25,no.36,pp.5081–5085,2013.[51]J.W.Boley,E.L.White,G.T.-C.Chiu,andR.K.Kramer,“Directwritingofgallium-indiumalloyforstretchableelectronics,”AdvancedFunctionalMaterials,vol.24,no.23, 96 pp.3501–3507,2014.[52]D.P.Parekh,C.Ladd,L.Panich,K.Moussa,andM.D.Dickey,“3Dprintingofliquidmetalsasfugitiveinksforfabricationof3Dmicrofluidicchannels,”LabonaChip,vol.16,no.10,pp.1812–1820,2016.[53]S.RusselandP.Norvig,Artificialintelligence:Amodernapproach,3rded.PrenticeHall,2010.[54]S.NolfiandD.Floreano,Evolutionaryrobotics:Thebiology,intelligence,andtechnologyofself-organizingmachines,1sted.TheMITPress,2000.[55]M.Dorigo,V.Trianni,E.S¸ahin,R.Groß,T.H.Labella,G.Baldassarre,S.Nolfi,J.-L.Deneubourg,F.Mondada,D.Floreano,andL.M.Gambardella,“Evolvingself-organizingbehaviorsforaswarm-bot,”AutonomousRobots,vol.17,no.2-3,pp.223–245,2004.[56]J.Yosinski,J.Clune,D.Hidalgo,S.Nguyen,J.C.Zagal,andH.Lipson,“Evolvingrobotgaitsinhardware:theHyperNEATgenerativeencodingvs.parameteroptimization,”inProceedingsofthe11thEuropeanConferenceonArtificialLife(ECAL),2011,pp.890–897.[57]R.PfeiferandJ.Bongard,Howthebodyshapesthewaywethink:Anewviewofintelli-gence,1sted.TheMITPress,2006.[58]N.Cheney,R.MacCurdy,J.Clune,andH.Lipson,“Unshacklingevolution:evolvingsoftrobotswithmultiplematerialsandapowerfulgenerativeencoding,”inProceedingsofthe15thAnnualConferenceonGeneticandEvolutionaryComputation(GECCO),2013,pp.167–174.[59]N.Cheney,J.Clune,andH.Lipson,“Evolvedelectrophysiologicalsoftrobots,”inProceed-ingsofthe14thInternationalConferenceontheSynthesisandSimulationofLivingSystems(ALIFE),2014,pp.222–229.[60]N.Cheney,J.Bongard,andH.Lipson,“Evolvingsoftrobotsintightspaces,”inProceedingsofthe2015AnnualConferenceonGeneticandEvolutionaryComputation(GECCO),2015,pp.935–942.[61]F.Corucci,N.Cheney,H.Lipson,C.Laschi,andJ.Bongard,“Evolvingswimmingsoft-bodiedcreatures,”inProceedingsofthe15thInternationalConferenceontheSynthesisandSimulationofLivingSystems(ALIFE),2016,p.6.[62]C.Majidi,“Softrobotics:aperspectivecurrenttrendsandprospectsforthefuture,”SoftRobotics,vol.1,no.1,pp.5–11,2014.[63]R.W.Ogden,Non-linearelasticdeformations.CourierCorporation,1997.[64]P.Polygerinos,N.Correll,S.A.Morin,B.Mosadegh,C.D.Onal,K.Petersen, 97 M.Cianchetti,M.T.Tolley,andR.F.Shepherd,“Softrobotics:Reviewoffluid-drivenintrinsicallysoftdevices;manufacturing,sensing,control,andapplicationsinhumanrobotinteraction,”AdvancedEngineeringMaterials,vol.19,no.12,p.1700016,2017.[65]C.-P.ChouandB.Hannaford,“Measurementandmodelingofmckibbenpneumaticartificialmuscles,”IEEETransactionsonroboticsandautomation,vol.12,no.1,pp.90–102,1996.[66]A.D.Marchese,R.K.Katzschmann,andD.Rus,“Arecipeforsoftfluidicelastomerrobots,”SoftRobotics,vol.2,no.1,pp.7–25,2015.[67]T.J.Wallin,J.Pikul,andR.F.Shepherd,“3Dprintingofsoftroboticsystems,”NatureReviewsMaterials,vol.3,pp.84–100,2018.[68]J.C.Breger,C.Yoon,R.Xiao,H.R.Kwag,M.O.Wang,J.P.Fisher,T.D.Nguyen,andD.H.Gracias,“Self-foldingthermo-magneticallyresponsivesoftmicrogrippers,”ACSappliedmaterials&interfaces,vol.7,no.5,pp.3398–3405,2015.[69]T.Pinto,L.Cai,C.Wang,andX.Tan,“CNT-basedsensorarraysforlocalstrainmea-surementsinsoftpneumaticactuators,”InternationalJournalofIntelligentRoboticsandApplications,vol.1,no.2,pp.157–166,2017.[70]B.Mosadegh,P.Polygerinos,C.Keplinger,S.Wennstedt,R.F.Shepherd,U.Gupta,J.Shim,K.Bertoldi,C.J.Walsh,andG.M.Whitesides,“Pneumaticnetworksforsoftroboticsthatactuaterapidly,”AdvancedFunctionalMaterials,vol.24,no.15,pp.2163–2170,2014.[71]Y.Elsayed,A.Vincensi,C.Lekakou,T.Geng,C.Saaj,T.Ranzani,M.Cianchetti,andA.Menciassi,“Finiteelementanalysisanddesignoptimizationofapneumaticallyactuatingsiliconemoduleforroboticsurgeryapplications,”SoftRobotics,vol.1,no.4,pp.255–262,2014.[72]R.V.Martinez,J.L.Branch,C.R.Fish,L.Jin,R.F.Shepherd,R.M.Nunes,Z.Suo,andG.M.Whitesides,“Robotictentacleswiththree-dimensionalmobilitybasedonflexibleelastomers,”Advancedmaterials,vol.25,no.2,pp.205–212,2013.[73]M.T.Tolley,R.F.Shepherd,B.Mosadegh,K.C.Galloway,M.Wehner,M.Karpelson,R.J.Wood,andG.M.Whitesides,“Aresilient,untetheredsoftrobot,”Softrobotics,vol.1,no.3,pp.213–223,2014.[74]R.K.Katzschmann,J.DelPreto,R.MacCurdy,andD.Rus,“Explorationofunderwaterlifewithanacousticallycontrolledsoftroboticfish,”2018.[75]N.W.Bartlett,M.T.Tolley,J.T.Overvelde,J.C.Weaver,B.Mosadegh,K.Bertoldi,G.M.Whitesides,andR.J.Wood,“A3d-printed,functionallygradedsoftrobotpoweredbycombustion,”Science,vol.349,no.6244,pp.161–165,2015.[76]M.Wehner,R.L.Truby,D.J.Fitzgerald,B.Mosadegh,G.M.Whitesides,J.A.Lewis,andR.J.Wood,“Anintegrateddesignandfabricationstrategyforentirelysoft,autonomous 98 robots,”Nature,vol.536,no.7617,pp.451–455,2016.[77]A.Miriyev,K.Stack,andH.Lipson,“Softmaterialforsoftactuators,”Naturecommunica-tions,vol.8,no.1,p.596,2017.[78]M.Wehner,M.T.Tolley,Y.Meng¨uc¸,Y.-L.Park,A.Mozeika,Y.Ding,C.Onal,R.F.Shepherd,G.M.Whitesides,andR.J.Wood,“Pneumaticenergysourcesforautonomousandwearablesoftrobotics,”SoftRobotics,vol.1,no.4,pp.263–274,2014.[79]D.P.Holland,E.J.Park,P.Polygerinos,G.J.Bennett,andC.J.Walsh,“Thesoftroboticstoolkit:sharedresourcesforresearchanddesign,”SoftRobotics,vol.1,no.3,pp.224–230,2014.[80]M.Manti,V.Cacucciolo,andM.Cianchetti,“Stiffeninginsoftrobotics:areviewofthestateoftheart,”IEEERobotics&AutomationMagazine,vol.23,no.3,pp.93–106,2016.[81]M.Al-Rubaiai,T.Pinto,D.Torres,N.Sepulveda,andX.Tan,“Characterizationofa3d-printedconductiveplamaterialwithelectricallycontrolledstiffness,”inASME2017Con-ferenceonSmartMaterials,AdaptiveStructuresandIntelligentSystems.AmericanSocietyofMechanicalEngineers,2017,pp.V001T01A003–V001T01A003.[82]M.Al-Rubaiai,T.Pinto,C.Qian,andX.Tan,“Softactuatorswithstiffnessandshapemodulationusing3d-printedconductivepolylacticacidmaterial,”Softrobotics,vol.6,no.3,pp.318–332,2019.[83]S.PrakashandS.Kumar,“Fabricationofmicrochannels:Areview,”ProceedingsoftheInstitutionofMechanicalEngineers,PartB:JournalofEngineeringManufacture,vol.229,no.8,pp.1273–1288,2015.[84]J.K.Paik,R.K.Kramer,andR.J.Wood,“Stretchablecircuitsandsensorsforroboticorigami,”in2011IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems.IEEE,2011,pp.414–420.[85]C.Majidi,R.Kramer,andR.Wood,“Anon-differentialelastomercurvaturesensorforsofter-than-skinelectronics,”SmartMaterialsandStructures,vol.20,no.10,p.105017,2011.[86]Y.-L.Park,C.Majidi,R.Kramer,P.B´erard,andR.J.Wood,“Hyperelasticpressuresens-ingwithaliquid-embeddedelastomer,”JournalofMicromechanicsandMicroengineering,vol.20,no.12,p.125029,2010.[87]B.-H.Jo,L.M.VanLerberghe,K.M.Motsegood,andD.J.Beebe,“Three-dimensionalmicro-channelfabricationinpolydimethylsiloxane(PDMS)elastomer,”Journalofmicro-electromechanicalsystems,vol.9,no.1,pp.76–81,2000.[88]Y.-L.Park,B.-R.Chen,andR.J.Wood,“Designandfabricationofsoftartificialskinusingembeddedmicrochannelsandliquidconductors,”IEEESensorsJournal,vol.12,no.8,pp. 99 2711–2718,2012.[89]R.K.Kramer,C.Majidi,R.Sahai,andR.J.Wood,“Softcurvaturesensorsforjointan-gleproprioception,”in2011IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems.IEEE,2011,pp.1919–1926.[90]R.Habibey,A.Golabchi,S.Latifi,F.Difato,andA.Blau,“Amicrochanneldevicetai-loredtolaseraxotomyandlong-termmicroelectrodearrayelectrophysiologyoffunctionalregeneration,”Labonachip,vol.15,no.24,pp.4578–4590,2015.[91]A.K.Au,W.Huynh,L.F.Horowitz,andA.Folch,“3D-printedmicrofluidics,”AngewandteChemieInternationalEdition,vol.55,no.12,pp.3862–3881,2016.[92]R.Amin,S.Knowlton,A.Hart,B.Yenilmez,F.Ghaderinezhad,S.Katebifar,M.Messina,A.Khademhosseini,andS.Tasoglu,“3D-printedmicrofluidicdevices,”Biofabrication,vol.8,no.2,p.022001,2016.[93]A.Alfadhel,J.Ouyang,C.G.Mahajan,F.Forouzandeh,D.Cormier,andD.A.Borkholder,“Inkjetprintedpolyethyleneglycolasafugitiveinkforthefabricationofflexiblemicroflu-idicsystems,”Materials&design,vol.150,pp.182–187,2018.[94]M.O.F.Emon,F.Alkadi,D.G.Philip,D.-H.Kim,K.-C.Lee,andJ.-W.Choi,“Multi-material3dprintingofasoftpressuresensor,”AdditiveManufacturing,vol.28,pp.629–638,2019.[95]A.D.Castiaux,C.Pinger,E.A.Hayter,M.E.Bunn,R.S.Martin,andD.M.Spence,“Polyjet3D-printedenclosedmicrofluidicchannelswithoutphotocurablesupports,”Ana-lyticalchemistry,2019.[96]Y.-L.Park,B.-r.Chen,andR.J.Wood,“Softartificialskinwithmulti-modalsensingca-pabilityusingembeddedliquidconductors,”inSENSORS,2011IEEE.IEEE,2011,pp.81–84.[97]H.-S.Shin,J.Ryu,C.Majidi,andY.-L.Park,“Enhancedperformanceofmicrofluidicsoftpressuresensorswithembeddedsolidmicrospheres,”JournalofMicromechanicsandMi-croengineering,vol.26,no.2,p.025011,2016.[98]R.C.Chiechi,E.A.Weiss,M.D.Dickey,andG.M.Whitesides,“Eutecticgallium–indium(egain):amoldableliquidmetalforelectricalcharacterizationofself-assembledmonolay-ers,”AngewandteChemieInternationalEdition,vol.47,no.1,pp.142–144,2008.[99]F.F.AbayazidandM.Ghajari,“Materialcharacterisationofadditivelymanufacturedelas-tomersatdifferentstrainratesandbuildorientations,”AdditiveManufacturing,p.101160,2020.[100]D.M.Dykstra,J.Busink,B.Ennis,andC.Coulais,“Viscoelasticsnappingmetamaterials,”JournalofAppliedMechanics,vol.86,no.11,2019. 100 [101]F.Corucci,N.Cheney,S.Kriegman,J.Bongard,andC.Laschi,“Evolutionarydevelopmen-talsoftroboticsasaframeworktostudyintelligenceandadaptivebehaviorinanimalsandplants,”FrontiersinRoboticsandAI,vol.4,p.34,2017.[102]J.Rieffel,F.Saunders,S.Nadimpalli,H.Zhou,S.Hassoun,J.Rife,andB.Trimmer,“Evolv-ingsoftroboticlocomotioninphysx,”inProceedingsofthe11thannualconferencecom-panionongeneticandevolutionarycomputationconference:Latebreakingpapers.ACM,2009,pp.2499–2504.[103]G.Methenitis,D.Hennes,D.Izzo,andA.Visser,“Noveltysearchforsoftroboticspaceexploration,”inProceedingsofthe2015annualconferenceonGeneticandEvolutionaryComputation,2015,pp.193–200.[104]F.Veenstra,J.Jørgensen,andS.Risi,“Evolutionoffinundulationonaphysicalknifefish-inspiredsoftrobot,”inProceedingsoftheGeneticandEvolutionaryComputationConfer-ence,2018,pp.157–164.[105]INRIA.SOFA-SimulationOpenFrameworkArchitecture.[Online].Available:https://www.sofa-framework.org/[106]E.Coevoet,A.Escande,andC.Duriez,“Softrobotslocomotionandmanipulationcontrolusingfemsimulationandquadraticprogramming,”in20192ndIEEEInternationalConfer-enceonSoftRobotics(RoboSoft).IEEE,2019,pp.739–745.[107]G.Fang,C.-D.Matte,R.B.Scharff,T.-H.Kwok,andC.C.Wang,“Kinematicsofsoftrobotsbygeometriccomputing,”IEEETransactionsonRobotics,2020.[108]R.Deimel,P.Irmisch,V.Wall,andO.Brock,“Automatedco-designofsofthandmor-phologyandcontrolstrategyforgrasping,”in2017IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems(IROS).IEEE,2017,pp.1213–1218.[109]T.G.Thuruthel,B.Shih,C.Laschi,andM.T.Tolley,“Softrobotperceptionusingembeddedsoftsensorsandrecurrentneuralnetworks,”ScienceRobotics,vol.4,no.26,2019.[110]A.Hintze,J.A.Edlund,R.S.Olson,D.B.Knoester,J.Schossau,L.Albantakis,A.Tehrani-Saleh,P.Kvam,L.Sheneman,H.Goldsbyetal.,“Markovbrains:Atechnicalintroduction,”arXivpreprintarXiv:1709.05601,2017.[111]S.Chapman,D.Knoester,A.Hintze,andC.Adami,“Evolutionofanartificialvisualcortexforimagerecognition,”inArtificialLifeConferenceProceedings13.MITPress,2013,pp.1067–1074.[112]R.S.Olson,J.H.Moore,andC.Adami,“Evolutionofactivecategoricalimageclassifica-tionviasaccadiceyemovement,”inInternationalConferenceonParallelProblemSolvingfromNature.Springer,2016,pp.581–590.[113]C.BohmandA.Hintze,“MABE(ModularAgentBasedEvolver):Aframeworkfordigital