Neuro-Inspired Spike-Based Motion: From Dynamic Vision Sensor to Robot Motor Open-Loop Control through Spike-VITE

Senso s 2013, 13, 15805-15832; doi:10.3390/s131115805 senso s ISSN 1424-8220 www.mdpi.com/jou nal/senso s A icle Neu o-Inspi ed Spike-Based Mo ion: F om Dynamic Vision Senso o Robo Mo o Open-Loop Con ol h ough Spike-VITE Fe nando Pe ez-Peña 1,*, A u o Mo gado-Es e ez 1, Alejand o Lina es-Ba anco 2, Angel Jimenez-Fe nandez 2, F ancisco Gomez-Rod iguez 2, Gab iel Jimenez-Mo eno 2 and Juan Lopez-Co onado 3 1 Compu e A chi ec u e and Technology A ea, Uni e sidad de Cádiz, School o Enginee ing, Calle Chile, 1, Cadiz 11002, Spain; E-Mail: a u o.mo ga[email p o ec ed] 2 Robo ic and Technology o Compu e s Lab (RTC), Uni e sidad de Se illa, ETSI In o má ica, A d. Reina Me cedes s/n, Se illa 41012, Spain; E-Mails: [email p o ec ed] (A.L.-B.); [email p o ec ed] (A.J.-F.); [email p o ec ed] (F.G.-R.); [email p o ec ed] (G.J.-M.) 3 Au oma ion and Sys em Enginee ing Depa men , Poly echnic Uni e si y o Ca agena, Campus Mu alla del Ma , Ca agena, 30202, Spain; E-Mail: [email p o ec ed] * Au ho o whom co espondence should be add essed; E-Mail: [email p o ec ed]; Tel.: +34-956-015-705; Fax: +34-956-015-101. Recei ed: 5 Oc obe 2013; in e ised o m: 11 No embe 2013 / Accep ed: 13 No embe 2013 / Published: 20 No embe 2013 Abs ac : In his pape we p esen a comple e spike-based a chi ec u e: om a Dynamic Vision Senso ( e ina) o a s e eo head obo ic pla o m. The aim o his esea ch is o ep oduce in ended mo emen s pe o med by humans aking in o accoun as many ea u es as possible om he biological poin o iew. This pape ills he gap be ween cu en spike silicon senso s and obo ic ac ua o s by applying a spike p ocessing s a egy o he da a lows in eal ime. The a chi ec u e is di ided in o laye s: he e ina, isual in o ma ion p ocessing, he ajec o y gene a o laye which uses a neu oinspi ed algo i hm (SVITE) ha can be eplica ed in o as many imes as DoF he obo has; and inally he ac ua ion laye o supply he spikes o he obo (using PFM). All he laye s do hei asks in a spike-p ocessing mode, and hey communica e each o he h ough he neu o-inspi ed AER p o ocol. The open-loop con olle is implemen ed on FPGA using AER in e aces de eloped by RTC Lab. Expe imen al esul s e eal he iabili y o his spike-based con olle . Two main ad an ages a e: low ha dwa e esou ces (2% o a Xilinx Spa an 6) and powe equi emen s (3.4 W) o con ol a obo wi h a high numbe o DoF (up o 100 o a Xilinx Spa an 6). I also e idences he sui able use o AER as a communica ion p o ocol be ween p ocessing and ac ua ion. OPEN ACCESS Senso s 2013, 13 15806 Keywo ds: spike sys ems; mo o con ol; VITE; add ess e en ep esen a ion; neu o-inspi ed; neu omo phic enginee ing; an h opomo phic obo s 1. In oduc ion Human beings, and hei ances o s be o e hem, ha e e ol ed h oughou millions o yea s and ob iously hei sys ems o pe o m asks oo. Mos o hese asks a e commanded by he b ain. The e o e, enginee s, and specially he neu omo phic enginee ing communi y [1,2] ha e ixed as hei main goal o mimic he human sys ems which a e supposed o ha e an ex ao dina y beha io ca ying ou hei own asks. In pa icula , eaching mo emen s (planning and execu ion) ha e been o ages one o he mos impo an and s udied ones [3]. I we ake a close look in humans, we will ind ha he sys em in ol ed in hese asks is he cen al ne ous sys em (CNS). This sys em is a combina ion o he b ain and he spinal co d and, simpli ying, i consis s o neu on cells and uses spikes o g adua ed po en ials o ansmi on he in o ma ion ac oss he ana omy [3]. Nowadays, i is possible o in eg a e se e al housands o a i icial neu ons in o he same elec onic de ice ( e y-la ge-scale in eg a ion (VLSI) chip [4], Field-P og ammable Ga e A ay (FPGA) [5] o Field-P og ammable Analog A ay (FPAA) [6]); which a e called neu omo phic de ices. The e a e many Eu opean p ojec s ocused on building compu ing sys ems which exploi he capabili ies o hese de ices (B ain-inspi ed mul iscale compu a ion in neu omo phic hyb id sys ems (B ainScale; websi e: h p://b ainscales.kip.uni-heidelbe g.de/index.h ml), SpiNNake (websi e: h p://ap .cs.man.ac.uk/ p ojec s/SpiNNake /) and he Human B ain P ojec (HBP; websi e: h ps://www.humanb ainp ojec .eu/) as examples). One o he main challenges is which de ices and how o in eg a e hem o p oduce unc ional elemen s. One o he p oblems aced when we y o in eg a e and implemen hese neu al a chi ec u es is he communica ion be ween hem: i is no easy o dis inguish which neu on o wha de ice is i ing a spike. To sol e his p oblem, new communica ion s a egies ha e been exploi ed, such as he Add ess-E en - Rep esen a ion (AER) p o ocol [7]. AER maps each neu on wi h a ixed add ess which is ansmi ed h ough he in e connec ed neu onal a chi ec u e. By using he AER p o ocol, all neu ons o a laye a e con inuously sha ing hei exci a ion wi h he o he laye s h ough bus connec ions; his in o ma ion can be p ocessed in eal ime by a highe laye . AER was p oposed o achie e communica ion be ween neu omo phic de ices. I ies o mimic he s uc u e and in o ma ion coding o he b ain. Like he b ain, AER will le us p ocess in o ma ion in eal ime, by implemen ing simple spike-based ope a ion a he ime each spike is p oduced o ecei ed. Tha ’s one o he easons o using i : he in insic speed behind he spike-based philosophy. Ano he one is he scalabili y allowed by i s pa allel connec ions. The mo ion p oblem is s ill being widely s udied. One success ul app oach o hese con ol a chi ec u es, including isual eedback, dealing wi h mo ion p oblem was based on ― isual se oing‖ whe e a came a guides he a m mo emen compu ing complex algo i hms [8,9]. Nowadays, his sys em is s ill used in indus y due o i s eliabili y, bu hese sys ems we e based on high esou ce consump ion Senso s 2013, 13 15807 compu a ional models ins ead o low esou ce neu oinspi ed spikes-based models. In his wo k, we desc ibe he s eps aken owa ds a ully neu oinspi ed a chi ec u e. Once he i s app oaches o bio- inspi ed image senso s appea ed a ew yea s ago (ea ly 2000s) [10], and he ace o make a comple e sys em began. Un il hose days, he e we e some ad ances in desc ibing neu oinpi ed con ol algo i hms: in [11] a couple o hem we e shown: one o gene a e non-planned ajec o ies (Vec o In eg a ion To-End Poin —VITE) and he o he one o ollow hem by muscles (Fac o iza ion o LEng h and TEnsion—FLETE). Then, many ela ed wo ks using hem we e published [12–14]. All o hese wo ks lis ed we e based on simula ions, bu his wo k p esen s a eal spike-based ha dwa e implemen a ion o VITE. Recen wo ks show a ha dwa e implemen a ion o hese algo i hms: In [15] hey use he same amewo k (algo i hm and pla o m) as his a icle and hey deal wi h he p oblem o wo ames o e e ence, one o isual and he o he one o he obo by doing a mapping be ween hem. Then, in [16] a whole pseudoneu al a chi ec u e is designed and applied o an iCub obo [17]; one o he algo i hms selec ed was VITE [18]. In [19], hey include he join limi s using he Lag ange heo em. The g ea and simple con ol achie ed a leas shows he oppo uni ies o using he VITE algo i hm. Two o he nea es wo ks o his pape om close esea ch g oups a e in [20], whe e hey used bo h algo i hms bu wi h a PC o un he equa ions and ind he bes way, and also in [21], whe e a spike p ocessing P opo ional In eg al and De i a i e (PID) con ol is implemen ed; i is in his las a icle whe e he Pulse F equency Modula ion (PFM) modula ion o mo o unning appea s. Mos o he lis ed wo ks used a compu e o p ocess sepa a ely a leas one elemen o he dynamic sys em. Also, he compu ing mode was no spike-based. This p o okes delays, non- eal ime [22] and de ini ely i is no an en i ely bioinspi ed beha io , al hough ha was he o iginal hough . This a icle is ocused on eal ime planning, execu ion and mo ion con ol in a bioinspi ed way: o design a ully neu oinspi ed a chi ec u e om he e ina o he obo . We ha e se wo cons ain s: only spikes can low ac oss he sys em and only addi ion, sub ac ion and injec ion o spikes a e allowed. In his way, copying hese neu al cons ain s, we achie ed a neu o-inspi ed con ol. We suppo he spike p ocessing me hod o all he algo i hms used; ha is he main claim o his pape : design, de elopmen and implemen a ion o a spike-based p ocessing con ol a chi ec u e and o a oid using an ex e nal compu e o p ocessing, wi h ex emely low powe consump ion and AER communica ion. Fi s o all we ha e o de e mine he elemen s o in eg a e acco ding wi h he biological p inciples: image senso , a ha dwa e a chi ec u e whe e he CNS beha io is emula ed and a obo o execu e he mo emen s. The e a e many so s o p oblems o deal wi h in his selec ion: senso mus be a spiking e ina, he a chi ec u e has o keep as many CNS ea u es as possible, wi hin only addi ion, sub ac ion and injec ion o spikes, and inally, he obo ic pla o m will be made o mo o s which mimic he muscles. The i s elemen o he a chi ec u e is he image senso . We ha e chosen a silicon e ina, he dynamic ision senso de eloped by he Tobi Delb uck esea ch g oup [23]. I is a VLSI chip made o 128 × 128 analog pixel i ing spikes (wi h AER p o ocol) when a h eshold is eached. Applying se e al p ocessing laye s o hese e en s low [24], a single e en , which mee s he cen e o an objec , is isola ed. The e o e, his e en plays he ole o he a ge posi ion o he sys em, so he e ina will deli e he eaching posi ion o he a chi ec u e. The main pa o he sys em u ns a ound he VITE algo i hm [11]. I was selec ed because i is inspi ed by he biological mo emen and was designed o mimic i . I has been ansla ed in o he spikes Senso s 2013, 13 15808 domain using spike-based building blocks which add, sub ac o injec spikes like he human neu al sys em. This algo i hm gene a es a non-planned ajec o y and i needs a second algo i hm o p oduce and con ol he o ces applied o he mo o s which mimic he muscles. In his pape we a e ocused on he i s algo i hm and he e o e, no eedback is pe o med. Ou aim is o e alua e he iabili y o ansla ing he VITE comple ely in o he spikes domain and applying i o a eal obo ic pla o m in o de o enable he second algo i hm and o close he con ol loop. In o de o achie e he desc ibed goal, we ha e ans o med he algo i hm using exis ing spike p ocessing blocks de eloped o ou esea ch g oup [21,25,26] and pu hem in o MATLAB Simulink o es and adjus he blocks. A e wa ds, wo FPGA based boa ds we e used o alloca e he blocks and mimic he biological s uc u e (one o he b ain and he o he one o he spinal co d). The inal a chi ec u e esul s in he e ina connec ed o he p ocessing FPGA boa d, his one o he ac ua ion FPGA boa d and inally o a s e eohead obo ic pla o m. The main achie ed esul is ha i is possible o con ol he obo ic pla o m in an open loop way by mapping he in o ma ion ecei ed by he e ina wi h he expec ed mo emen a he obo . The e is high accu acy be ween he simula ed cu es and he signals ead om he mo o s encode . Finally, a oiding he use o he second algo i hm causes a new componen implemen ed in he second FPGA boa d. The ask o his componen was o adap he spikes in o a Pulse F equency Modula ion (PFM) modula ion o eed he mo o s. The es o he pape is s uc u ed as ollows. Sec ion 2 p esen s he esea ch inspi a ion and mo i a ion: how an in ended mo emen is p oduced om a biological poin o iew. Then, he backg ound o VITE algo i hm is desc ibed and he spike-based p ocessing is p esen ed. The nex sec ion ansla es he VITE algo i hm in o he spike-based pa adigm, he SVITE: spike-based VITE. Then, he implemen a ion o he SVITE in o he FPGA boa ds dealing wi h he ha dwa e ad an ages and disad an ages is p esen ed. In he nex sec ion, he esul s pe o med a e shown and he accu acy o ansla ion is discussed based on he esul s ob ained. The las wo sec ions a e de o ed o he discussion, which includes he connec ion be ween he elec onic algo i hm and he biological mo emen , and u u e di ec ions o his esea ch. 2. Biological Mo emen 2.1. In oduc ion In his sec ion we a e going o p esen a sho e iew o how an in ended mo emen is pe o med by humans. Jus he i s s age o he mo ion will be desc ibed: he one connec ed wi h he b ain; we lea e he passi e and e lex mo emen s aside because hey a e no execu ed om a sigh ed a ge . Ou s a ing-poin is he human cen al ne ous sys em (CNS) ha plays he ole o mo emen con olle . The CNS consis s o he b ain and he spinal co d [27]: • The b ain in eg a es he in o ma ion om he spinal co d and mo o co ex in o de o plan, coo dina e and execu e he desi ed mo emen s. This a icle co e s he unc ion de eloped by his pa . • The spinal co d ecei es in o ma ion om se e al senso y elemen s and includes he mo o neu ons in cha ge o in ended and e lec ed mo emen s and he ac s o he in o ma ion low. Senso s 2013, 13 15809 Focusing in he b ain, inside he ce eb al co ex, he e a e wo sys ems, wi h h ee a eas pe each, esponsible o p ocessing he senso y and mo o in o ma ion: senso and mo o sys em. The a eas a e called p ima y, seconda y and e ia y depending on he abs ac ion le el hey manage. The mo o sys em has o p ocess he in o ma ion abou he ex e nal wo ld p esen ed by he senso y sys em and p ojec i in o he neu al elemen s o ca y ou he mo emen . To ca y ou hese asks, mo o sys ems ha e a con inuous senso y in o ma ion low and a hie a chical o ganiza ion o h ee le els: he spinal co d, he b ains em and he mo o co ex a eas [27]. This o ganiza ion is also massi ely connec ed and has eedback a all o i s le els. A b ie desc ip ion o hese a eas is as ollows: • The spinal co d is he lowes le el in he hie a chy. I has neu al ci cui s o p oduce a wide ange o mo o pa e ns. The mo ion in his le el can occu e en when i is disconnec ed om he b ain [3]. • The b ains em is esponsible o d i ing he neu al sys ems. All connec ions be ween he spinal co d and he b ain go h ough he b ains em and ac oss wo pa allel ubes. • Rela ed o he mo o co ex, we shall conside hese h ee suba eas: p ima y mo o co ex (B oadmann a ea 4), p emo o co ex (pa o B oadmann a ea 6) and supplemen a y mo o co ex (ac ually i is pa o he p emo o co ex). This las a ea p ojec s di ec ly o he spinal co d. The es o he a eas p ojec o he spinal co d h ough he b ains em. Figu e 1 ep esen s he desc ibed beha io . 2.2. In ended Mo emen s The uppe elemen s o he mo o sys em (mo o co ex) a e esponsible o : mo ion planning acco ding o a ge and en i onmen in o ma ion. To do so, he e a e many p ojec ions be ween senso y a eas and mo o co ex (p ima y and p emo o ). The senso y a eas may in eg a e in o ma ion om di e en kinds o senso s and p ojec di ec ly o he p emo o co ex. I will ecei e in o ma ion o he eaching a ge and i has wo di e en a eas: • Supplemen a y mo o a ea: I plays an impo an ole in complex mo emen execu ions and in mo emen p ac ices. • P emo o co ex: I con ols he eaching mo emen s and i can p ojec o he p ima y mo o co ex and di ec ly o lowe mo o con olle ins ances. In he six ies, an ac i i y in he p ima y mo o co ex be o e he mo emen execu ion was obse ed. I was due o a planning o he mo emen in p og ess [28]. The e is one mo e subco ical s uc u e ype playing an impo an ole in mo ion con ol: he basal ganglia [29]. They ecei e inpu s om he neoco ex and p ojec o he b ains em con olling he mo emen in p og ess. Al hough hey play a ole, i is s ill la gely unknown and we ha e no in oduced i in ou model. Figu e 1 shows a block diag am o he in ended mo emen s’ p oduc ion: Senso s 2013, 13 15810 Figu e 1. Simpli ied block diag am o in ended mo emen s’ execu ion. The e a e wo pa s: one o he planning and p og amming (in ol ed in his a icle) and he second pa o he mo emen execu ion. In he diag am, he a ows ep esen he in o ma ion lows: solid lines a e used o ep esen ajec o y in o ma ion and do ed lines a e used o ep esen mo emen commands. In his a icle, he second pa has been implemen ed as a wi e; nei he eedback has been conside ed. 3. Vec o In eg a ion o Endpoin (VITE) Algo i hm 3.1. In oduc ion Any simple ac ion in ol es he use o coo dina ion o hund eds o biological elemen s. A join mo emen o example, will cause coo dina ion be ween wo o mo e muscles. A he same ime, each muscle consis s o se e al ibe s and senso y cells connec ed o e e en and a e en ne es coming down om he spinal co d. The VITE algo i hm [11] ies o model he human mo emen s keeping as many de ails o he neu al sys em in mind as possible. Essen ially, VITE gene a es he ajec o y o be ollowed by he join , bu in con as o app oaches which equi e he s ipula ion o he desi ed indi idual join posi ions, he ajec o y gene a o ope a es wi h desi ed coo dina es o he end ec o and gene a es he indi idual join d i ing unc ions in eal- ime employing geome ic cons ain s which cha ac e ize he manipula o . No ice ha VITE is he i s laye in ol ed in a planned a m mo emen . I does no in eg a e any eedback om he end obo . I gene a es he ajec o y ega dless o he o ces needed o de elop he mo emen . Thus, i eeds a heo e ical second laye commanded by ano he algo i hm [12]. This second laye con ac s wi h he end manipula o and manages he command ecei ed by he p e ious one. 3.2. Block Diag am and Equa ions The block diag am (Figu e 2) and he equa ions a e p esen ed in his subsec ion in hei simples o m o he algo i hm. The algo i hm will in eg a e he di e ence ec o a each ime in o de o upda e he p esen posi ion (Equa ion (2)). Bu i will no be upda ed un il he GO signal has a non-null alue (Equa ion (2)). Meanwhile (GO signal has a null alue) he di e ence ec o is p e-compu ed in o de o be eady o he shoo in he con ol signal. This ime is known as he ―mo o p iming‖. I a any ime while he mo emen is being done he GO signal goes ze o, he mo emen will be ozen in ha posi ion. The a ge posi ion can be upda ed du ing he mo emen . This change will cause jus an upda e in he di e ence ec o ega ding he new goal. Senso s 2013, 13 15811 All he posi ions poin ed in he algo i hm mus be e e ed o he same ame. The e o e, i spa ial posi ions a e conside ed, he in eg a ion o he p esen posi ion (PP) will be ma ched wi h he speed p o ile o he mo emen . I we make a compa a i e be ween his algo i hm and he classical con ol heo ies o indus ial applica ions (P opo ional, In eg al and De i a i e con olle s), his algo i hm would esul simila o classical in eg al con olle due o he inal in eg al componen bu i is no . This in eg al plays he ole o he end obo o eedback he ideal posi ion eached. The special componen , GO signal, ca ies ou a pseudo p opo ional playing he ole o a pseudo dis u bance. Figu e 2. Block diag am. Taken om [11]. 3.3. Some Conside a ions o he Algo i hm Applica ion 3.3.1. Synch onous Mo emen The algo i hm aces he p econcei ed heo ies ha alk abou a p ep og ammed ajec o y be o e a mo emen is done. Wi h his algo i hm, he mo emen s a e ca ied ou in a eal ime and i is possible o change he a ge du ing he mo emen wi hou dis u bing i . Also, in he in oduc ion we s a ed ha his algo i hm is in ended o co e a comple e mo emen in ol ing se e al muscles; a join o example. Figu e 3. Th ee mo emen s composed o wo join s a e ep esen ed. The s a poin s a e B1, B2 and B3 and he end poin is he same o all o hem: E. Solid lines ep esen he igh way o pe o m he mo emen and he do ed lines indica e a composi e o wo ac ions. This is a g aph aken om [11]. Senso s 2013, 13 15812 Thus, i ecei es an abs ac e e ence, i.e., a spa ial poin and i should gene a e a ajec o y o all he muscles in ol ed in he ac ion. In addi ion o his, he mo emen canno be a composi e o wo o mo e join mo emen s (do ed lines in Figu e 3); i mus be a ges u e o synch onous ac ion o a ious muscles (solid lines in Figu e 3). This concep is called syne gies and hey happen in a na u al and dynamic way. Thus, o pe o m a synch onous mo emen , each muscle g oup should con ac o expand a a di e en quan i y acco ding o he di e ence ec o compu ed o each one. F om his concep o synch onous mo emen comes up he need o pa e n and speed ac o iza ion. Wi h an independen speed con ol o each muscle g oup i is possible o adjus all o hem o each high accu acy in he synch oniza ion be ween all he muscles in ol ed in he mo emen . 3.3.2. Coding he Posi ion One impo an issue ega ding his algo i hm is how o code he posi ion o an isola ed muscle o a whole join . F om [11], wo me hods can be used o know he posi ion o he end muscle: he co olla y discha ge and he in low in o ma ion. The i s one is he command p o ided o he muscle and om he b ain. The second is he eedback in o ma ion om he muscle. The e o e, wi h he co olla y discha ges i is supposed ha he end e ec o a i es o he o de ed posi ion and wi h he in low in o ma ion i is possible o upda e he posi ion i a passi e mo emen is pe o med and also o check he posi ion in an in ended mo emen . The VITE algo i hm uses only he signal om he b ain o upda e he p esen posi ion and he e o e, o gene a e he ajec o y. Thus, i is supposed ha he end e ec o eaches he commanded posi ion. This is a ypical way o a epe i i e mo emen . Howe e , ega dless o whe he he in low in o ma ion exis s o no , i is necessa y o implemen ga es o inhibi o allow i . 3.3.3. GO Signal and Speed P o ile The GO signal is in cha ge o he mo emen speed con ol. I is also he ga e o ha mo emen . The implemen a ion o his signal causes a di e en speed p o ile in he global mo emen . The ypical signal used is a amp; he highe he slope, he as e he mo emen . Wi h a amp p o ile, he gene al speed p o ile achie ed is a bell-shaped one. A he beginning speed is low. When he a ge is being eached he speed inc eases. A he end o he mo emen he speed goes down o inc ease he accu acy. The symme y o his bell shaped p o iles a y wi h speed [30]. No ice ha his signal loses i s meaning when he a ge is eached. To sum up, i can be said ha i is no impo an how o each a a ge , bu jus o each i . So, he ajec o y does no ma e , excep o i ing he join angle cons ain s. 4. SVITE: Spike-Based VITE 4.1. Spike-Based P ocessing This sec ion p esen s a b ie desc ip ion o spike-based p ocessing. This way o p ocessing aims o mimic he beha io o he human ne ous sys em. The in o ma ion in his sys em is analogue and we y o ep oduce i , bu wi h digi al de ices. The design is made up o a ha dwa e desc ip ion language (HDL) o se e al blocks. These blocks p ocess he in o ma ion in he simples way: addi ion, sub ac ion and injec ion Senso s 2013, 13 15813 o spikes a e allowed as his is supposed o be in a biological neu onal p ocess. The in o ma ion is based on he i ing a e o he blocks ying o mimic he human neu ons analogue. The e a e only spikes lowing be ween hese blocks, being p ocessed while hey low, un il hey a e applied o he mo o s. This p ocessing way akes ad an age o he highe clock equency o hese digi al sys ems o achie e an equi alen p ocessing. The p incipal ad an age is i s simplici y; we do no need complex p ocesso s o sol e equa ions. Also he powe consump ion is an impo an poin : he e is a huge di e ence o wa s be ween he ypical p ocess compu e and he elec onic elemen s. Ano he p o i able ad an age is space: we use small elec onic de ices which could be alloca ed in a s and-alone way. The e is a published wo k [21] which e eals he powe o his spike-based p ocessing, whe e a spike PID con olle was designed and es ed. 4.2. T ansla ion in o he Spikes Pa adigm In a p e ious sec ion, he VITE algo i hm has been p esen ed and ho oughly desc ibed. This sec ion p esen s he ansla ion o he algo i hm in o he spikes pa adigm. We ha e called his new algo i hm SVITE, o spike-based VITE. The ansla ion is done in wo ways: keeping he in o ma ion in a spike-based sys em in mind and aking ad an age o he Laplace ans o ma ion o sol e he equa ions conside ing ze o ini ial condi ions. In hese spike sys ems, he in o ma ion has a ela ion wi h he in e spike in e al (ISI) and speci ically wi h he i ing a e which can be unde s ood as he equency. Tha is he eason why we i s go in o equency domain wi h Laplace (we a e no ma ching Laplace domain wi h i ing a e, i is jus an in e p e a ion o le us ansla e in o spike-based p ocessing pa adigm). The e o e, aking he equa ions o he algo i hm as ou s a ing-poin and using he Laplace ans o ma ion o sol e he equa ions, he main pa s o he algo i hm a e ansla ed ega dless o whe he he GO signal is used, because i we ansla e he equa ions in a s ic way, he p oduc be ween GO signal and he di e ence ec o will be ansla ed in o a con olu ion in he equency-domain and i will no be co ec because his GO signal was designed in o de o con ol he speed o he mo emen in he o iginal algo i hm. Wi h his a gumen and ega ding he in o ma ion inside a spike sys em, he ansla ion in o spike pa adigm o his p oduc will be an addi ion o wo spike ains in he spikes-p ocessing pa adigm. As a esul , he i ing a e (o equency) o he esul ing signal will be inc eased in any case. The nex sec ion deals wi h his idea. Thus, he ansla ion, s a ing wi h Equa ions (1) and (2), is as ollows: By unning Equa ions (5) and (6) i is easy o a ange hem in blocks (Figu e 4): (1) (2) (3) (4) (5) Senso s 2013, 13 15820 which is i ing a he cen e o he a ge objec . These add esses a e usually decomposed in (x, y) coo dina es o he e ina isual e e ence. By means o expe imen al indings, gene a o esolu ion is: (15) whe e he pa ame e NBITS is he numbe o bi s selec ed o implemen he spike gene a o ha supplies he a ge . Wi h his da a, we can ansla e he a ge in o ma ion om he e ina o a sui able e e ence o each algo i hm. The algo i hm is spli in o as many pa s as mo o s he e a e in he obo ic pla o m. In his wo k, he obo ic pla o m has ou deg ees o eedom: wo axes wi h wo mo o s pe axis. We use wo SVITE: one o he x-axis and ano he one o he y-axis. Each SVITE ou pu is sen o wo mo o s (same beha io ). Fu he mo e, his di ision is in ag eemen wi h he pu sued goal o p oducing an in ended bu synch onous mo emen . I we ha e wo algo i hms, i is possible o adjus each GO signal o succeed. One p oblem de i ed om he eplica ion o he algo i hm is ela ed o he spike p oduc ion. All he algo i hms i e nea ly a he same ime and al hough AER handshake p o ocol akes a sho pe iod o ime o communica e wi h he bus, a ew spikes could be los . Indeed, he access o he bus can c ash wi h he spike p oduc ion o he algo i hms, leading o spike loss. Thus, o a oid his loss wo op ions a e a ailable: Include a Fi s -In-Fi s -Ou (FIFO) bu e memo y ha is used o pallia e spo adic high speed p oblems in he AER communica ion a he ou pu o each algo i hm; he a bi e will access he memo ies o ca y ou he communica ion. A oid he use o an a bi e and communica e he ins ances by using jus a pai o wi es o ansmi he spikes (poin o poin communica ions om p ocessing o ac ua ion FPGAs). In he p esen ed a chi ec u e, bo h schemes ha e been s udied: communica ion wi h he nex laye by AER p o ocol (only one pai o add esses is needed; one pe each algo i hm) and wi h a pai o wi es a e conside ed. The di e ences a e shown in he esul s sec ion. 5.3. Ac ua ion Laye This laye will adap he in o ma ion ecei ed in o de o eed he mo o s o he obo ic pla o m. I ecei es add esses o bo h algo i hms and p oduces he spikes o he mo o s. I is implemen ed in he Spa an-3 Xilinx FPGA p esen in he AER-Robo boa d [21] (de eloped unde he SAMANTA-II p ojec (Mul i-chip AER ision sys em o obo ic pla o m II (SAMANTA-II), Oc obe 2006 o Sep embe 2009), which is able o d i e DC mo o s h ough op o-couple isola o s and ull L298N b idges o mo o ac ua ion. The obo ic pla o m is a s e eo- ision obo wi h ou deg ees o eedom powe ed by DC mo o s. Al hough, he mo o s ha e an isola ed mo emen , a his momen , hey a e coupled in pai s, one o each axis acco ding o he usage o jus one e ina. Thus, each axis is ed wi h one algo i hm, so we ha e one algo i hm o he pai o mo o s o he axis. The powe supply equi emen o he mo o s is 24 Vdc. The manu ac u e o he mo o s is Ha monic D i e and he model is RH-8D6006. The s uc u e o he obo ic pla o m is made so ha he mo o s o he y-axis a e c ossed o hei axis and ha e a ansmission bel o Senso s 2013, 13 15821 mo e he a m. Wi h ega d o his s uc u e, we ed he y-axis mo o s wi h he posi ion (because i is needed o hold he spike ansmission) and he x-axis mo o s wi h he speed p o ile. We p opose o use PFM o un he mo o s because i is in insically a spike-based solu ion almos iden ical o he solu ion ha animals and humans use in hei ne ous sys ems o con olling he muscles. Ne e heless, we need o adap he spikes because he digi al clock o he boa ds is ixed a 50 MHz esul ing in a spike wid h o 20 ns and his signal is e y as and he spikes oo sho o he mo o s model o he obo ic pla o m [21]. To compu e he maximum and minimum spiking a e allowed we mus look in de ail a he boa d componen s and he DC mo o , espec i ely. On he one hand, o maximum i ing a e, he powe s age o he boa d consis s o an op ical isola o and H-b idges, as we ha e men ioned. Bo h componen s ha e some ea u es ega ding he swi ching equency: o he H-b idge i is ixed a 40 KHz, bu i is ecommended o wo k a 25 KHz (minimum pe iod o 40 µs) o a oid mal unc ioning. As o he isola o s, he e is no maximum swi ching equency de ined, bu wo impo an empo al es ic ions mus be conside ed: 6 µs and 5 µs o aise and all, espec i ely. Me ging hese da a, i esul s in o a maximum i ing a e o 25 KHz (40 µs o minimum pe iod), and wi hin his max a e he spike wid h can be sol ed. Using he minimum pe iod o 40 µs and aking in o accoun he empo al es ic ions o he isola o s, i esul s in a ime pe iod o 29 µs as maximum wid h. We ha e chosen a secu e wid h o 25 µs o ma gin and o sp ead ou he spikes up o 750 clock cycles. De ini ely, wi h hese da a, he maximum swi ching equency will be 25 KHz and he spike wid h 750 clock cycles. On he o he hand, o compu e he minimum spiking a e allowed i is necessa y o analyze he a ge ac ua o (DC mo o in ou case). A DC mo o ac s as a low pass il e and he ans e unc ion can be calcula ed using he pa ame e s om he manu ac u e [33]. This unc ion and, pa icula ly, i s s ep esponse allow us o selec he mo o ’s minimum swi ching equency (maximum pe iod) sui able o ollow an inpu p ope ly. The s ep esponse calcula ed illus a es an app oxima e o al ime o 40 ms o ollow he inpu . The e o e, we a e going o selec a lowe o de alue wi h a li le ma gin: 1 ms o maximum pe iod, so a minimum equency o 1 KHz o he incoming spikes. These wo limi s will allow us o build up he empi ical able ha maps he ision e e ence sys em and he mo emen p oduced a he pla o m. To sum up, we ha e he ope a ing ma gin o he mo o s: om 1 KHz o 25 KHz and he spike wid h as 750 clock cycles. No ice ha , i we make he spike injec ion in he GO block lowe han en pe cen o slope, i would no cause any mo emen a all because he mo o will il e he spikes. In con as , a much highe slope could sa u a e he sys em wi hou a closed loop con ol. 5.4. Ha dwa e Resou ces Consump ion In gene al, o measu e he ha dwa e consump ion in a FPGA, wo poin s should be conside ed: he dedica ed esou ces included o build up complex de ices such as mul iplie s and he con igu able logic blocks (CLBs) o gene al pu pose. The algo i hm does no use any complex s uc u e. I jus needs coun e s and simple a i hme ic ope a ion esou ces. The e o e he measu emen s a e ocused in o he a ailable slices a he FPGA. Senso s 2013, 13 15822 We ha e syn hesized he algo i hm, including a spikes a e coded gene a o [35], a spikes moni o [36] and he in e ace wi h o he neu omo phic chips o a co ec debugging and a use ul in eg a ion. Table 1 p esen s he da a o he de ice wi h he epo ob ained. Table 1. Ha dwa e esou ces consump ion by he Spa an 6 1500 de ice. Numbe o Slices Max. Blocks in he De ice Use by One Block Algo i hm 238 96 1.033% Algo i hm plus moni o 533 43 2.31% Algo i hm plus in e ace 242 95 1.05% Algo i hm plus moni o and in e ace 537 42 2.33% In his able, he i s column desc ibes he elemen implemen ed o each case. The nex column shows he amoun o slices needed o syn hesize he uni s a he FPGA. The ollowing column ep esen s he maximum numbe o uni s ha could be alloca ed inside he FPGA. Finally, in he las column he o al capaci y o he de ice o all he syn hesis pe o med is shown. The esul s e idenced a low ha dwa e esou ce usage when an isola ed algo i hm is implemen ed, jus one pe cen . Also, i is ema kable ha he in e ace wi h o he neu omo phic chips almos does no p o oke an inc emen in he ha dwa e esou ces consump ion (only ou slices). Consequen ly he inal implemen a ion o a comple e a chi ec u e will consis o he algo i hm and he in e ace. Howe e , he design and es phases need he moni o in o de o check he igh beha io o he algo i hm. All he esul s p esen ed in his sec ion co espond jus o slice consump ion. The FIFO included wi hin implemen a ion uses dedica ed memo y blocks al eady p esen in he de ice, so i is no compu ed. I we compa e he maximum numbe o algo i hms ha can be alloca ed a he FPGA ( ha co esponds o he deg ees o eedom (DoF) con ollable in ou a chi ec u e) wi h he iCub Robo necessi y [18], i shows a g ea ad an age using ou app oach. We can con ol up o 95 DoF (wi hou moni o ) in compa ison wi h iCub pla o m which allows 53 DoF. 5.5. Powe Consump ion The powe consump ion o he design implemen ed can be di ided in h ee di e en pa s: he de ice s a ic, design s a ic and design dynamic powe consump ion. The de ice s a ic powe consump ion is also called he o -chip powe and i is e e ed o he powe consump ion o he boa d wi hou any con igu a ion. The design s a ic powe consump ion is he powe used when he design is jus p og ammed in o he boa d bu i is no unning. Finally, he dynamic powe consump ion is e e ed o he powe used by he design when i is unning. We ha e used he XPowe es ima o ool om Xilinx o ge he de ice s a ic and design dynamic powe consump ion. The esul s a e: 0.113 W o he de ice s a ic powe and 0.027 W o he design s a ic powe . The design dynamic powe is ob ained by compu ing he di e ence be ween he eal measu emen , when he algo i hm is unning, and he addi ion o de ice s a ic and design s a ic powe consump ion. The powe consump ion measu ed is 3.4 W, hus he design dynamic powe is 3.26 W. Senso s 2013, 13 15823 6. Resul s This sec ion p esen s se e al esul s o he whole design. These esul s aim o show he e olu ion om he o iginal VITE algo i hm design by G ossbe g [11], going h ough i s ansla ion in o spikes (SVITE), o a eal obo ic pla o m. Then, he wo op ions explained o he communica ion be ween he ac ua ion and p ocessing laye a e illus a ed o check he ansla ion done. Finally, we wan o show he pe o mance o he designed con ol sys em. A e wa ds, we p esen some discussions. These esul s ha e been achie ed by means o hese ools: on he one hand, wi h MATLAB and Xilinx Sys em Gene a o we ha e managed he heo e ical and simula ion scena ios o ge he simula ed esul s (pa o he Figu es 10–14); on he o he hand, he so wa e sui e om Xilinx was used o syn hesize in o he FPGA de ices o he boa ds a he ha dwa e se up o ge he unning esul s (Figu es 15–18). Figu e 12. Pe o mance achie ed co esponding o one pe cen age slope in GO signal. Do ed lines a e simula ed in on o measu emen solid lines. The bell shape p o ile signals ep esen he speed. The ipple in he spike-base beha io is due o he unc ion ha ans o ms he spikes in o a con inuous signal. The a ge is he same o bo h simula ed and measu emen s signals and i is ep esen ed as a i ing a e. I akes a o al o 17 s o each he a ge i we look h ough he posi ion. Figu es 12 and 13 show he expec ed beha io o he algo i hm ansla ed in o spikes pa adigm agains he beha io o he o iginal design o he algo i hm by G ossbe g. The do ed lines a e aken om simula ions o he o iginal VITE algo i hm; solid lines show he same da a bu measu ed in he boa ds wi h a special AER moni o [36,37]. The speed p o ile is aken be o e he in ege block ( he In eg a e and Gene a e block in Figu e 5) and he posi ion o de s a he ou pu . The accu acy o bo h signals is highly p ecise and i sugges s he oppo uni y o succeeding wi h a ully spike-based obo con olle . Figu e 14 shows simula ion da a o he speed p o ile achie able when he pa ame e slope_coun e in GO block goes h ough di e en alues causing 10, 50, 100, 500 and 1000 pe cen age slopes. The bell Senso s 2013, 13 15824 shape p o iles con i m he s udies in [27] whe e i is said ha as as e is he mo emen he highe asymme ic speed p o iles a e pe o med. A he desc ip ion o he laye s we p esen ed wo op ions o he communica ion be ween bo h: AER communica ion o jus a pai o wi es ca ying he spikes. In Figu e 15, bo h op ions a e shown ega ding speed p o ile. This g aph does no e eal any change a he i ing a e when an AER communica ion is applied be ween he laye s. This ac allows us o say ha AER communica ion is a good s a egy o connec spike-based p ocessing and ac ua ion wi hou any modi ica ion in he spiking a e ansmissions. These wo es s conce n only he x-axis which is he one ed by he speed p o ile. The slope used was 10%. Figu e 13. Pe o mance achie ed co esponding o en pe cen age slope in GO signal. Do ed lines a e simula ed in on o measu emen solid lines. The bell shape p o ile signals ep esen he speed. The ipple in he spike-base beha io is due o he unc ion ha ans o ms he spikes in o a con inuous signal. The a ge is he same o bo h simula ed and measu emen s signals and i is ep esen ed as a i ing a e. I akes a o al o 12 s o each he a ge i we look h ough he posi ion. Figu e 14. Speed p o iles achie ed by modi ying slope_coun e pa ame e o GO block. Making a compa a i e be ween slow and as mo emen s we can app ecia e ha he peak eloci y is eached la e o as e mo emen s i en i e leng h is conside ed. Senso s 2013, 13 15825 Figu e 15. Speed p o iles ead ou om mo o encode s. The blue p o ile is ead wi h an AER communica ion and he ed p o ile is om a pai o wi es communica ion. The inpu was an AER add ess ma ching coo dina e 125 o he x-axis. In Figu e 16, he posi ion (angle) eached o each ype o communica ion is shown. The e e ence o de ed o his mo emen was 75 deg ees. As we can see, wi h bo h ypes o ansmission, he a ge will be eached by he pla o m. The blue line ep esen s he posi ion eached when a wo wi es communica ion is pe o med and he ed line ep esen s he posi ion eached when an AER communica ion is used. The iny di e ence be ween bo h communica ion modes a e due o he ac ha wi hin AER, a handshake ook place o access he bus. The es also e eals a maximum speed o 83 deg ees pe second on he x-axis. Figu e 16. Posi ion eached by he mo o s o he x-axis when wo di e en communica ion s a egies a e used. Senso s 2013, 13 15826 Despi e using o no an AER communica ion o wo wi es, one o he mo e a ac i e i ems o he algo i hm was ha i is possible o gene a e synch onous mo emen s by con olling he GO signal independen ly o each mo o . Figu e 17 shows he eal measu emen s o he posi ion eached when he a ge is ixed a (125, 90) in he ame o e e ence o he e ina. This becomes an angle o 75 deg ees o he x-axis and 48 deg ees o he y-axis in he ame o e e ence o he obo ic pla o m. Mo eo e , he igu e shows he ansla ion o he a ge deli e ed by he senso . Since ou obo ic pla o m has a special a chi ec u e ha leads us o use he posi ion commands o he y-axis and he speed commands o he x-axis ( o hold he a ge posi ion a he end) i is no easy o p oduce synch onous mo emen s; indeed i is impossible because he posi ion commands a e slowe han he speed commands. Ne e heless, we ha e ed he y-axis also wi h he speed p o ile (al hough i does no hold he posi ion a he end) o check how synch onized a mo emen can be done wi h his algo i hm; he esul is ex emely accu a e i we compa e i o he a ge p o ide ep esen a ion. Tu ning o he slowe ajec o y and looking a i , he x-axis ge s he posi ion commanded in app oxima ely one second and hen s a s he mo emen in he y-axis. I we compa e he heo e ical signal deli e ed o he mo o s (Figu e 13) and he mo emen achie ed ( eads ou om he encode s o he mo o s and using jAERso wa e ool [38]) shown in Figu e 17, i e eals a couple o commen s ega ding ime eaching leng h. Fo he x-axis, he one commanded by he speed p o ile, i is qui e di e en , bu in bo h cases, hey ha e he same maximum i ing a e and o he y-axis, commanded by he posi ion, we ound nea ly he same leng h. The eason o he di e ence a ime leng h in he x-axis can be unde s ood wi h hese wo poin s: non- eedback used, which means once he mo o s a s unning we do no ha e any ine ia con ol, non- eedback om any senso . Howe e , he accu a e mo emen achie ed when bo h axis use he speed p o ile leads us o hink abou ge ing ges u es i all he mo o s we e able o be ed wi h he sui able eloci y p o ile o hold he posi ion. The es pe o med in Figu e 16 o check he communica ion s a egy e ealed a ime o each he op mo emen o 0.9 s ( he es was done wi h he maximum mo emen achie able by he obo : om 0 o 127 coo dina e o he e inas’ ame). I we conside his eaching ime and join i wi h he la ency shown in Equa ion (7), i esul s in a minimum o 1.1 s dis ance be ween swi ching he a ge . I will be he limi o he obo o be able o ollow in he igh way. Thus, in Figu e 18 we ha e pe o med a eal es o check he acking p ope ies o he obo . The es is done jus o he x-axis which is he one commanded by he speed p o ile. The angle ed by he e ina is calcula ed using jAER. Fo hese es s, he a ge has been deli e ed o he p ocessing laye by DVS senso wi h h ee di e ence ime dis ances: 2.6, 3.9 and 5.2 s o each one. I is possible o de ec he la ency o 0.1 s a he beginning. The p ocessing ime can also be calcula ed by compu ing he ime be ween he a ge deli e y and he s a o he mo ion minus he ixed la ency. I esul s 0.5 s. The di e ence be ween he angle eached by he obo and he mo ion ep esen ed is due o he esolu ion ob ained by he encode s o he mo o s and also o he cumula i e e o in an open-loop con olle wi hin many a ge s wi hou calib a ion be ween hem. Finally, empi ical es s e eal an accu a e acking by he obo when he dis ance be ween a ge deli e ies is a leas 2 s. Senso s 2013, 13 15827 Figu e 17. Angle s. ime eached o bo h axis wi h (125, 90) inpu . The e ina has 128 × 128 pixels. The communica ion be ween neu al boa ds was made by AER. The ed lines show he ajec o y ollowed by he obo when we used he speed p o ile o bo h axis and jus o he x-axis; he blue lines ep esen he mo ion deli e ed by he DVS senso . Figu e 18. Angle s. ime acked o he x-axis. The inpu is a go and e u n along he x-axis in he ame o e e ence o he e ina. The ed poin s show he angle ajec o y ollowed by he obo and he blue poin s show he a ge s deli e ed o he obo . Finally, o conclude he esul s sec ion, he e a e some impo an conside a ions e ealed by hem: • Lowe spiking a es: we ha e highe spiking a es a he simula ions han in eal esul s; o eal obo ic pla o ms, he i ing spiking a e should be adap ed o he mo o s in o de o succeed wi h he PFM con ol. Also, he a es a e limi ed by he elec onic componen s o he used boa ds. Senso s 2013, 13 15828 • Li e ime limi ed: he signals ead ou om he encode s a e sho e han hey we e supposed o be aking simula ion esul s. This i s wi h he non- eedback used om he p op iocep i e senso s; no mo o ine ia con ol a all. • Delay suppo : he same as he p imi i e algo i hm suppo s delay [39], ou ansla ed algo i hm could also suppo hem: i he GO signal is sho a e o be o e he a ge is submi ed, i will only cause a delay o a je k due o he highe s a ing speed, espec i ely. • We ha e o choose an AER o a couple o wi es communica ion. I he obo ic pla o m has many DoF, he AER p o ocol may be used. O he wise, a pai o wi es communica ion has an accu a e enough beha io up o eigh deg ees o eedom (16 bi s o AER bus). • The mapping unc ion was c ea ed wi h empi ical esul s. Thus, he able has an in insic limi a ion: i was designed wi h speci ic pa ame e s and i would cause he obo ic pla o m o mal unc ion i we change hem. The e o e, o check a igh synch onized mo emen as VITE desc ibed i is needed o close he loop using p op iocep ion senso s. Also, o make he algo i hm po able o any obo ic pla o m, his mapping unc ion should no be cons uc ed wi h empi ical esul s. 7. Discussion: Connec ion be ween SVITE and Biological Mo emen Al hough he physiological e idence o ma ch heVITE algo i hm wi h neu ological beha io is clea ly de ined a he li e a u e [40], in his sec ion a connec ion be ween he SVITE and he biological mo emen will be es ablished as long as we ha e added addi ional blocks and modi ied nea ly all o hem. The biological mo emen has wo di e en sou ces o senso y in o ma ion: one called p op iocep i e in o ma ion o sensing coming om join posi ion and muscle ension senso s and ano he low ega ding a ge si ua ion gi en by he e ina. The in o ma ion gi en by he e ina goes in o he algo i hm and s a s he Di e ence Vec o (DV) compu a ion; his p e ious compu a ion ma ches wi h ac i i y egis e ed in he p emo o co ex be o e he mo emen begins in biological mo ion by speci ic neu on cells [41]. Fu he mo e, he GO signal appea s be ween he p emo o co ex and he p ima y mo o co ex. This signal is he shoo ing o he mo emen . Also, he DV signal is no deli e ed s aigh o mo o s. The e a e h ee spike-based blocks in be ween. This ma ches he concep o ―p ojec o in e neu ons‖ a he han ―di ec ly o mo o -neu ons‖ [32]. As we ha e seen up o his poin , he mo emen is gene a ed di ec ly om he algo i hm’s ou pu o di ec ly om he b ain’s ou pu in a biological way. We sugges a eed o wa d model, wi hou ei he a p op iocep i e senso no in o ma ion om he e ina a his s ep o he p ocessing because hey will be included in he nex s ep: he second phase FLETE which includes eed-back p ocessing. In gene al e ms, he eedback allows a p ecise eaching mo emen and also upda ing he p esen posi ion du ing a passi e mo emen . In he p esen wo k i is assumed ha he obo eaches he posi ion commanded. This eed- o wa d model does no accep delays, noise and he mo o commands mus be highly p ecise [32]. The nex s ep (FLETE) shall include se e al eed-back loops: one sho local loop a he muscle (mimicked by he mo o s in a obo ) as an au oma ic gain con ol, a second sho local loop p o ided by he e ina in he isual senso y a ea and a las a la ge loop om p op iocep i e senso s. Senso s 2013, 13 15829 8. Conclusions A ully neu oinspi ed a chi ec u e has been p esen ed: om an AER e ina o PFM con olled mo o s. The sys em aims o keep as many ea u es in mind as possible om he biological in ended mo emen s: • Hie a chy sys em wi h a p ocessing and an ac ua ion laye like he b ain and b ains em. • Some ac i i y p e ious o he mo emen ( he la ency and compu ing o di e ence ec o ) like he p e ious ac i i y de ec ed a he p emo o co ex. • GO signal is in he edge o he p emo o and p ima y mo o co ex and i s ha dwa e implemen a ion based on spikes sa es compu a ional cos s because i is done as an addi ion ins ead o a signals mul iplica ion. This esea ch mee s i s goals wi h neu omo phic enginee ing ones, ha is: mimicking he neu onal sys em beha io in o de o de elop use ul applica ions. The con olle designed uses only 2% o he Spa an 6 FPGA and he powe consump ion is 3.4 Wa s excluding he mo o s. The esul s e eal he accu a e use o AER p o ocol o ac ua ion pu poses. The con olle can be eplica ed up o one hund ed imes o con ol complex obo ic s uc u es wi h such DoF. The open-loop neu o-con olle designed and implemen ed is able o each, in a synch onized way, any posi ion commanded by he ou pu o an accu acy acking done by a cascade a chi ec u e based on DVS senso . Fu he esea ch will p o ide a ull eedback a chi ec u e including passi e mo emen upda ing and p op iocep i e in o ma ion om muscles o ine uning. Wi h his ad ances we aim o use he sys em o imp o e accu acy in eal- ime unning obo s. Acknowledgmen s This wo k has been suppo ed by he Spanish g an (wi h suppo om he Eu opean Regional De elopmen Fund) VULCANO (TEC2009-10639-C04-02) and BIOSENSE (TEC2012-37868-C04-02). Con lic s o In e es The au ho s decla e no con lic o in e es . Re e ences 1. Tellu ide Cogni i e Neu omo phic Wo kshop. A ailable online: h ps://neu omo phs.ne (accessed on 15 May 2013). 2. Capo Caccia Cogni i e Neu omo phic Wo kshop. A ailable online: h p://capocaccia.e hz.ch (accessed on 15 May 2013). 3. She ing on, C.S. The In eg a i e Ac ion o he Ne ous Sys em; Yale Uni e si y P ess: New Ha en, CT, USA, 1906. 4. Camille i, P.; Giulioni, M.; Dan e, V.; Badoni, D.; Indi e i, G.; Michaelis, B.; B aun, J.; Del Giudice, P. A Neu omo phic VLSI Ne wo k Chip wi h Con igu able Plas ic Synapses. In P oceedings o he 7 h In e na ional Con e ence on Hyb id In elligen Sys ems, Kaise slau e n, Ge many, 17–19 Sep embe 2007.