ABSTRACT
Mostanimalshavesignificantbehavioralexpertisebuiltinwithouthavingto explicitlylearnitallfromscratch.Thisexpertiseisaproductofevolutionofthe organism;itcanbeviewedasaverylong-termformoflearningwhichprovides astructuredsystemwithinwhichindividualsmightlearnmorespecializedskills orabilities.Thissuggestsonepossiblemechanismforanalogousrobotevolution bydescribingacarefullydesignedseriesofnetworks,eachonebeingastrict augmentationofthepreviousone,whichcontrolasix-leggedwalkingmachine capableofwalkingoverroughterrainandfollowingapersonpassivelysensedin theinfraredspectrum.Asthecompletelydecentralizednetworksareaugmented, therobot'sperformanceandbehaviorrepertoiredemonstrablyimprove.The rationaleforsuchdemonstrationsisthattheymayprovideahintastothe requirementsforautomaticallybuildingmassivenetworkstocarryoutcomplex sensory-motortasks.Theexperimentswithanactualrobotensurethatanessence ofrealityismaintainedandthatnocriticaldisablingproblemshavebeen ignored.
