Learning and Fast Object Recognition in Robot Skill Acquisition: A New Method

Invanant objecr recoglWon alms nl racognising an objccl indqxndcntly of its posiliorr. scale m d orietitat~on, ikis 1s lrnportnnt in robor skdl n~qw~jkhn during g~arping npcratiuns especially when workmg in unsrrucnrrad cnvirunmcnts. In Ih~s pqmr ue present a new approach 16 embed nianipuliitwe skills elk line uslng hrntied sensorial i n f m ~ l i o n The novel approach ttses dn ANN for i>hjwt tcarnmy and on-line recogn~tion trsing a dcscriprivc vector build on murrent prtteins of differelhi complex objscw Expcr~menlnl lrnming results uawg 8n IEEE 1394 h s t camcrs arc prcm[cd Canonical parts ale used for cmnparlnp ctiifcrcnl cc~nncdionlsl models (Baekpropagmon, Pcrcepltoti and FuzzyARTMAP) and camplcx and gcomctriditl figures are hen used for experime~its. rhc mc(hotl showcd an owerll 93% cficiency and for natso-complcx ubjccks 100% recognition and recogn~tidn limes of lws than one nullisecond Ke>words Xrhficinl h'eural Netvvorks, Assembly. Machinc V~siun, ht.anufacturmg Cell, Patteln Rccognlkiun, Kohol~cs. I . introduction The success of grasping and as.~enibly uperallbns usmg tndustrtal robots la currently ha.;cd on the accuracy of Ihc robot and the precise knowldgc of Lhc cnvironmcnl, 1.e. ~nfollnation about the geomctry uf [he ausen~bly par& nnd therr lotxilizalion ~n the workqxioc. lectlniques are wughl lo provide self-adaptat1011 for robots. 'nris dt~ilrnenl reports ;I nuvel neural-bas4 methodology Cor inv&ant ohjrd recognibon and npplied to self-adaptmg 1ndiatriaJ ruboh which con perTom asieinbly tasks, Ncw o h j ~ r t ~ can nlsu be learned quickly t f ccrtnm clues tlrc ~ I Y C D to the leamec, since the ~nctl~odology presented. uses only two onl~nc patterns for lesmuig complex objects. The archtwturt. is lirstly lralncd with clues represaitmg dlfPerent objects thrt thc rubor 1s likoly to encounter (and w~rh olher:, thul r c p r ~ ~ ~ n t complex objects) wthm the work~ng spncc lu form ils inilia1 knowledye b:w. The mam lden suggests that 11 is possible to gd bhl w d rcliitbls ~nformt~on from a st~nple h~t h c u x d analysis of what an object might show, The very itnportaot a.qxccx of the scene (we have called "clues''), can bt: used la te tu wtricve memw~ecd aspects of 1112 ~bjeci without havlnp to recf~I1 deraled features, By using ncursl nclwvrks 11 is possible 11, learn maoipukt~ue skdls which can bc used by 'in induhtfld n~anip~tlator In somewrp we humans do that prwcsr wec an object hns been seen and Icsmcti for the Grrt timc. The artwle describes a robust irtcthnd for ver? k t lexrnmg. perimeter ,md centtold cabulnt~onu. object furx~ions end pose tstimalron,