Artif l.ife (1999) 3:1 14 ◎ ISARO13 1999 ORIGINAL ARTICLE

Masanori Sugisaka Design of an artificial brain for robots

Received: May 30, 1998 / Accepted: June 30, 1998

Abstract This paper proposes a new information pro- interfaces. The control was produced by the RN-2000 cessing and control system, which is called artificial brain neurocomputer, which is able to learn various control laws (ABrain), for robotics using neurocomputers ar-rd a Von- instantly, in order to track a moving object within a prede- Neumann-type microcomputer, and interfaces operating termined range o[ errors. these computers. We introduce three robotic systems with We have also developed a fast pattern recognition ABrains developed recently in our laboratory. One is an system based on moment invariants using the primitive ABrain to recognize the objects in a robotic system for ABrain.e'1o The fast pattern recognition system consists of a recognition and tracking. The others are an ABrain fbr CCD video camera, an image processing system named controlling the duty ratio in the robot system for recogni- FDM, a monitor, two stand lights, a NEC PC-9801 micro- tion and tracking, ancl an ABrain for controlling the computer, and a RICOH neurocomputer RN-2000 which is steering angle in an intelligent mobile vehicle. Based on considered as the primitive ABrain. Experimental studies our results, we present a realization of a general type to recognize five dynamic patterns of Japanese chestnuts of ABrain to recognize a moving pattc-nt and track it were performed. From the studies, a high speed of both simultaneously. and recognition has been achieved compared with the former pattern recognition system based on the soft- Key words ABrain Pattern recognition I'racking ware of artiflcial neural networks developed by us." In Neurocomputer additon, we have also developed an intelligent navigation technique for a mobile vehiclel2-Is for smooth running using the primitive ABrain. rs Based on the systems developed in our laboratory, the !ntroduction configuration of a new ABrain is proposed in this paper. First, we give the general concepts of the ABrain.rcre .Wc have alrcady developcd a prilnitive artincial brain Second, we consider how to make the ABrain. Third, we (denotcd ABrain),l which was cOnstructed by one neuro‐ give preliminary results obtained from our recent studies to C01mputer and a Von¨Neulnann― type NEC PC-9801■ licro― design a general type of ABrain. Then we propose a general COmputer, fOr traCking a mOVing ObiCCt,utiliZing Our structure for the ABrain to perform various behavioral pre,1。 us results.2 6rrhe systenl consists of one cc)I)vide。 functions such as learning, recognition, decisions, etc., can■era;two]〕 C nlotors with cncOders,one on the hOrizon‐ based on adaptation, evolution, self-organization, and the tal axiS aind One the vertical axis;a pulse― width lnodulatiOn emergence of robotics. (PWM)driving unit;a32‐ bit mi∝ ocomputer NEC PC_ 9801: the iRIC(DII neurocon■ puter R:N-2000;7,8 and their

What is the ABrain?

M. Sugisaka (f!) In this section, we show what the ABrain is. Before we Departmcnt of E,lectrical and Electronic Enginecring, Oita explain it, we state briefly the why the ABrain is University, 7[X) Dannohara, Oita 870-11, Japan necessary. If a new ABrain like a Tel. + 8 1 -975-.54-7U3 I I Fax + 8 1 -t)75,-54-784 I brain is developed c mltil: msugitc cc.6ila-u.at.jp and is imbedded into a robot or a machine, it will become friendly toward human beings. If emotion emerges from the 'l'his -llhird work was presented, in part, at the Intoruational Sympo- ABrain of a robot, the robot is able to behave like a human siunr on ancl Robotics, Oita, Japan, January l9--2i, l99g being. In other words, it means that we give artiflcial life to 8 'Ihis the robot by installing the Allrain. is a dream which The ABrain for a robot consists of three partial ABrain, many engineers have wanted to achieve for a long tin.re. as statcd above: (1) R-ABrain, (2) T,D-ABrain, (3) B- Let us focus our on the functions of the ABrain. Consequently, the ABrain in thc robot is able ABrain. What is the Allrain? Figure 1 shows the flow of to recognize objects, to think how to behave, to make a inforntation and lcarning in the Atsrain. Stimulus from a decision, and to take action. Thc structures or netrvork sensor is transl'erred into the neural network system for systems of the three partial ABrains arc illustrated in recognition, which is denoted the ABrain for recognition outline in Fig. 2. (the R-ABrain) in the brain's nervous system. 'l'he results processeci by the R-ABrain are transferred into the neural network system for thinking, and based on the rcsults pro- How to make an ccssed by the neural network system, a clecision will be ABrain made for what kind of behavior should be taken. This neu- ral nctwork system is called the ABrain for thinking Onc way for us to build an ABrain for a robot is by utilizing and decision (the'I',D-ABrain). The output proccssed by thc functions of the . The functions of a hurnan the T,D-ABrain is transferred into the neural network sys- brain are still not clarifled, and it is regarded as a complex tem for bchavior or action, which is called the ABrain for adaptive system. The human brain and the nervous system behavior (the B-ABrain). Finally, stirnulus processecl by the in the whole body consists of a huge set of neural network B-ABrain is transferred into the actuators ancl clrivers ir.r systems, which are an aggregation of neurcns. The ultra- order for the system or robot to achieve a predeterminecl structurc of tl'rc , which is a unit for information 22 objective. processing, is shown in Fig. 3.20 We nor,v illustrate an example of the functions of thc A new research field for building ABrain for robots has ABrain. When a human bcing sees a tiger, thc human being rccently been developcd by mixing brain science and engi- runs away in order to aviod danger. On the othcr hancl, neering such as VLSI harclware and computer . when a human being sees a cake, the human being tries tcr Realization of arr ABrain by hardrvare architecture with eat it. This example suggests both what the ABrain is ancl self-organizing functions as stated above, is a clream of rvhat we should make. human beings for the tlventy-first century. Our efforts to- A learning process is includccl in the ABrains (R- rvard this object should yield a better life f

The ABrain developed in our laboratory on February 2, lu9.5r (dcnr,rted ABrain- I ) was inslallccl into a recognil iorr and tracking system for moving objects.2 This ABrain is Stimulus able to learn how to control two DC motors for tracking a moving object in three-dirnensirxral space by itself instantly. a Thereafter, we devcloped an Atsrain on February 28. ru Recognition I995e (denoted ABrain-2) to recognize five dynamic pat- A priori terns of Japanese chestnuts. We developed an ABrain or-r a memory February 1.7, 199'lts (denoted ABrain-3) for controlling thc j angle of the steering wheels of a mobilc vehicle. We rvill Thinking explain each ABrain briefly below. Learning ABrain-1 Decision posteri rt memory ABrain-1r consists of a 32-bit Von-Neumann-type micro- computer NEC PC-9801[3A, a RICOH neurocomputer RN- Behavior 2000 (Yamato), the intcrlaces, the software (Nadeshiko) for both constructing a neural nctwork and learning a contrcll Exarnple: tiger > run away law by it, and the software for tracking a moving object. cake ' want to eat Stimulus ABrain-l for a recognition and tracking system lor mov- ing objects is shown in Fig. 4. Dctailed descriptions of the Fig. 1 Flowchart of in the ABrain system have been givcn previously.3-6 The solid line in Fig. 4 Fig.2a-c Component parts of moment the artificial brain. a Artificial invariants brain for recognition (R- ABrain), b artificial brain for NEC PC9801 A for thinking and decisions (T,D- Q sensor input NEURO― ABrain) (selfJearning and self-thinking), c arlificial brain COMPUTERI Q into artificial for behavior (B-ABrain) RN braln for thinking and 1/0 Q decision(TD― ABrain) a of

circuit

It of recogn RN-2000 and deo ision (input for NEC PC9801 β A B― AB b

|~llこ C9

NEURO― neural network for behavior 1/F 00MPUTER

RN input into driver, 1/0 -2000

C

・L 】. 賄 lra6

一一―く PC9801 A CCD Camera Video 1/F BOard E

Para‖e! 1/O Board M Axon Hillock Soma t n a p Para‖ e: r m i n ・ E M 1/O board

o cSp %

E d ” Axon n 。 輌 軍 r a t ・ 一 ‐l u r n Pla3ma ic‐e Motors Dendrits {N138: Msmblano NeurocOmputer Sub3tanCel Driving Circuit RN-2000 Fig. 3 Ultrastructure of the nei:ron

Prinlitive ABrain

Fig, 4 Arlificial brain (ABrain) for recognil ion and tracking. M, motor; Fi, encodcr 10

shorvs tl'rc conliguration of thc primitive ABrain. We suc- neural network in the RN-2000 neurocomputer are shown ceeded in tracking the desired values (sinusoidal, circular, in Table 1. In this table, the input data to the are and elliptical movements) of encoders on both the X- and the deviaLions or errors ofe, and e, which are equal to 1, X-axes, and a light from an electric lamp moved by hand. 2, . . . ,10. The teaching data are the duty ratios of the DC The structure of the neural network in the RN-2000 motors on both the horizontal and vertical axes, which are neurocomputer in the tracking experiments is shown in Fig. equal to 0.1,0.2,..., 1.0. The input dal-a are transformed 5. There are four neurons in the input laycr, 16 neurons in into four-bit digits by using a bit transformation technique. cach of the {irst ancl second intermediate layers, and one The teaching signals are normalized so that the maximum ncuron in the output layer. Thc inputs to thc neural value 1.0 and minimum value 0 in Table 1 correspond to 127 netr.vork are errors between the center of the CCD camera and 0, respectively. and the desired values of the encoders on the X- and y- In the tracking experiment, we used the neural network axes, or the point of highest intensity of the light from an shown in Fig. 5, which is constructed by Nadeshiko soft- clectric lamp movecl by hand, as shown in Fig. 6. The teach- ware. The conflguration of the neurocomputer using the ing signals are the duty ratios of the pulse width moclulation neural network is illustrated in Fig. 7. The duty ratios for the (PWM). The why we use four neurons in the input two DC motors on the X- and Y-axes are produced sequen- la1,g,' ... given in detail elscwhere.l tially from one neural network. In other words, at flrst the As the learning clr training data for the neural network error between the desired value of X (detoted Xddi,"din Fig. shorvn in Fig. 5 we used proportional control data. Both the 7) and the value of the encoder of the DC motor on the X- real training data and the corresponding data used for thc axis (denoted X-motor in Fig. 7) is processed by the neural network in order to get the duty ratio of the X-motor. Then the same process is repeated in order to get thel duty ratio of the Y-motor By taking account of the hardware specifications, the sampling time employed the ● ● in experiments was 100ms. The procedures described abovel* are repeated at each

● ● sampling time throughout the tracking experiments. We describe briefly how to detect the position of the point of highest light intensity from a lamp being moved by hand, lnput and then explain how to track it. The detection procedures (Errors are as follows: of x and y) Teaching signal (Duty Ratio) Table 1 Data for training ● ● L,rrors (deviations) Teaching signals Learnecl results 0 0000 0 0 1 0 1 3 0 1 0001 2 0 ハ 15 2 0()10 卸 “ 27 3 0011 Ю , 31 Fig. 5 Structure of tlre ncural netrvork in RN-2(XX.) 4 0100 53 5 0101 ” “ 0 7 6 63 6 011() 7 0 ” 72 7 0111 “ 0 2 85 8 1000 ” 1 4 108 Highest 9 1001 0 0 2 7 112 intensity Of light 10 1()10 12() Y-axis ●

~~ ′ |` E e「 o y や l o 日 X motor 、 V O X desired 」 X― axls ∽ C I O ■ _(― ――――――――)卜 | + 〓 e m Center Y desired

RN-2000

Fig. 7 Configuration of the neurocontroller (I and II shorv the orcler of Fig.6 Errors or deviations prooessing) 11 1. read the 6-bit image data of the light (15 x l2 : 1S0 Thesimulationof theproposedsystemwith ABrain-2is pixels) using a CCD camera; carrieed out in the following steps. The flrst step uses the 2. write the image data into the memory of the NEC PC- image scene, which is compressed into256 (16 x 16) pixels 9t301BA; by preprocessing from the original image scene to the input 3. detect the position of the point of highest intensity of the pattern. In the seecond step, seven moment invariants are moving light. comp.uted for each pattern. ln the third step, three moment tnvariants instead of seven are usecl as the inputs of the The tracking proceclures are as follows: artificial neural network constructed in the ncurocomputer 1. calculate thc duty ratio for the X- and Y-motors sequen- for all samples because of the hardware specilications of tially using the trained neural netwoi-k in the RN-2000 the neurocomputer. In the fourth step the training begins, neurocomputer; using the back-propagation method by the neurocomputer. 2. move the system in order to make the center of the CCD By ending the artihcial neural network training in the video camera coincide with the position of the point of neurocomputer, the output of the network is usecl to recog- highest intensity of the light using the duty ratios calcu- nize the dynamic patterns. lated above. We now show the moment invariants and the neural we show one of the results of tracking a light bcing nc'twork for recognition. The moment invariants Qr, Qr' Q, moved by hand in Fig. g, where the aistance"il.,rr""i are invariant under translation and rotation.er1 the moving light and the CCD vidco camera is approxi- 16 ib rnately 60cm. In this figure, the loci fiom 8s to 2js are Mnr:ZLi'i'p,, (1) illustrated. The solid line shows thc locus of the X and Y i:\ j=l coordinates, and the dotted line is the locus of the point of 16 16 highest intensity. The result obtained from rracking a light pr, : II(i- i)"(i j\" t,, (2) being moved by hand is satisfactory for two reasons. T'he i-t i=l neural network structure constructed in the RN-2000 er: llzo r voz (3) neurocomputer is simple, and the control law used to train the neural network is a simple proportional law. er: (vrrn * pnz) + +pf, (4) : * Allrai,-2 Ot (tt,o Frz) + (3Ltto - tttu)' (5) where P, is a digital 0 or 1 function, i and 7 have integer ABrain-2e'r') was clevelopecl for the pattern recognition sys- values, and i = MrJM* andi - MollMN. tem shown in trig. 9. It consists of a CCD video camcra, an The moment invariants above are used as the inputs for image processing system named FDM, a ntonitor, two stand the neural network in the RN-2000 neurocomputer. In our lights, an NEC PC-9801 microcomputer, and a RICOH RN- pattern recognition system, only the three moment invari- 2000 ncurocomputcr, where tl.rese two different types of ants Or, Qr, and Q, are used instead of seven moment computers constitute an artilicial brain. Expcrin-rcntal stud- invariants because of the hardware speciflcation in the ics to recognize five dynamic patterns of Japanese chcstnuts neurocomputer. The artiflcial neural network system con- were performed. Fron.r the studies, a high spccd in both structed in the neurocomputer is shown in Fig. 10. It con- learning and recognition has been achievecl. sists offour layers: the inputlayer with 15 neurons, the first hidden layer'with 16 neurons, the second hidden layer with 16 neurons, and the output layer with 5 neurons, which is equal to the number of patterns to be recognized. 20 The experimental studies on the learning of five dynamic 15 patterns of Japanese chestnuts were performed with the image processing systeem shown in Figs. 9 and 10. Figure L1 l0

5

5 -t0 -15 -20 -20 -10 0 t0 20 30 40 x Monitor FDM98 PC9801 CCD camera and its height

Fig. 8 Results of tracking a moving light (8,25s). I-oci of the ligtrt ancl Fig. 9 New pattern recognition systcm usilg an I1ICOH ItN-2000 X and Y coorclinatcs neurocomputer r 1500 リ driving and l ccD motor Q, t steering boards

r V,

l t Q, O O ∞

J I Y' Q. に

Fig。 10 1Neurai nCtヽVOrk c()nstructcd ill an ltN-2000 ncurocon〕 DClAC driving Putcr switch batteries inverter 1100 Patternl Pattern2 Pattern3

O

Pattern4 PatternS

Fig. 12 Thc mechanical configuration of the mobile vehicle

Fig. 11 Examples for testing recognition of flve diffcrent patterns Primitive artiicial brain Main switch OUT I PC-9821Ap/U7 DC/AC shows the flve different patterns that were used in thc su,=Ro、 1 32/32 RN-2000 learing process. f'hc results obtained confirm the effective- ness of the proposcd system and show an improvement of IN Gamera both the learning and recognition speeds, which is the rnain poinl of ABrain-2.'"' +12V DC/DC Motor driver for up and down movement Motor ofthe camera a ABrain-3

left and right movemsnt ABrain-3ls has been developed recently for controlling the a steering angle of the front wheels a of mobile vehicle called Encoder for t the neuro-mobile vehicle, and shown in Fig. 12.12-ts up and down movement The of the camera configuration of the electricalpart of the NMV is shown in Fig. 13,_ where the dotted line indicates the primitive Encoder for left and right movement ABrainls which belongs to the same category of B-ABrain ofthe oamera as ABrain-1. The neural network used in ABrain-3 consists Motor of four layers: the input layer with 8 neurons, the flrst for driving hidden layer with 16 neurons, the second hidden layer with 16 neurons, and the output layers with 1 neuron, The input Sensor board for to the neural network is the error of the angle between the bottom of the centerline of the CCD camera and the center Motor driver of gravity of the object detected. The output from the neu- ral network is the number of pulses of the stepping motor to, Encoder lor steering control the steering angle.ls The experimental results are' satisfacLory. Fig. 13 'l'he electrical configuration of the mobile vehicle 13 In this section, we explained two dilferent types of ABrain. One is the R-ABrain and the other is the B- ABrain. We do not show the structure of the o T,D-Atsrain c because at this stagc o its functions are perfrtrmed using the ∽ E c inlormation obtained from experiments by skilled engi- o 〇 Deoision 一り necrs. However, 一o it is itxportant to dcvelop the structure of Circuit C irouit o l ∩ the neural network in the T,D-ABrain orcler

General strudure Of an ABrain Circuit for Generating lnternal Performanoe Criterion In this section, we propose the general structure ol

Fig. I4 'I'he configuration of the artificial brain conliguration of a general ABrain lbr robots is shown in Fig. 14.'?r 'Ihis conhguration has a general structure for the ABrain used for robots and intelligent machines or Gonclusions computers. f'he ABrain consists ol the following artificial organs: We hrive considered how to de sign the general structure ol' an ARrain, wliich consists of three types of primitive l. various sensors which receive dilfercnt intbrmation; ABrain (R-ABrain, T,D-ABrain; B-ABrain). Also, we ex- 2. a self-organizing circuit which organizes the structure of plained both the R-ABrainer(' and the B-ABrain,r.15 which the artificial neural nelworks and tl-re nurnbers of both were recently developed in our laboratory. Since we hzrve neurons ancl neurocomputers in the ABrain; obtainecl guidelines 1or designing the general ABrain hascd 3. several parallel neurocomputers and a Von-Neumann- on the results developed by our group, wc typc contputer; will ap1>ly these guiclelines to constructing ABrain for intelliger-rt rnachines zi. a circuit for generating internal performance criteria in and robots in various order to optimize the engineering lields related to humarr total neural network systems, beings. inclucling the piirallel neurocomputers and a Von- Neumann-type compu ter'; -5. decision circuit, which delermine ancl controls the actions of drivers or actuators. References

'I'he general harclware structure of the ABrain shown in l. Sugisaka lvt (i997) Neurocolnputer control in an ar.tiflcial brain lor Fig. 1,1 is able to perform the functions of information lracking moving ol'rjccts. Artif Lif'e Robotics l:.17 .51 2. Sugisaka NI, Tonoya N (l99fi) Ncuro control for recognition procesing ancl control stated above. In other words, each and tracking system ol rnoving obiect. Syst Sci 22(2):5()-61 ARrain, i.e., t.he R-ABrain; T,D-ABrain; and R-ABrain, is 3. Sugisaka M (1992) Recognition and tr.acking systenr of rnoving enrbedcled into this gcneral ABrain harclware. objects (in Japanese). Patent'l-cikkaihei 5-2333t)62 4. Sugisaka ll'he R-ABrain cclnsists of the neural netwclrk oI the M, Kaita H, Hara N,l et al. (1992) I{ecogni1i(}n and track_ ing system oI rnoving objects neurocomputer in which the inputs based on artificial neural networ.k arrcl to thc neumns in the PWI\I cont(rl. I)rocecdings ol'the IFToNIVI-jc lntenrational Syllr- ir-rput layer are fcatures (e.g., mon-rent invariants) of posiutr on Theory ol Machines arrd Mechanics, Nagoya, Jaltan, the object. The T,D-ABrain consisls of the neural nctwork Septemtrer' 21-26, v<'tl 1, 1992, pp 104 107 5. Sugisaka M (1992) Recognition ol'the neurocompltter in which the inpuls to 1he neurons ancl tracking systent ol'rnoving objects basecl on artilicial ncural rrctwork. l,roceedings of thc 1 992 in the input layer iire genetic inlblntation and/or knowledge Korean Autonratic (lontro[ ('un[crcuuc, Suoul, Korea, Oclober obtained lrorn the experirnent and liont learning. The 19 21 , lL)t)2, pp 573 5721 B-ABrain consists o1 the neural network of the neuro- 6. Sugisaka M, Har-a M, Tirnoya N (199.1) Neuro-fuzzy control for lccognition and tracking systenr ol ruovitg obiects. l)r.occcclings conlputer in which 1he inputs to the neuror.rs of in the input thc 3rd lnl.ernational Workshop on Advanced Motion (ilitrol, layer are lhe errors ol difTerences betwcen the present Berkclcy, [JSA, March 20-23, t9()1, pp l09l-1098 states and the clesired stales. The ir.rformatiou receivecl 7. IJguclri H, Furuta'l', Horiguchi H et nl. (1991) Neural nelwork hardware with learning from sensors is processecl sequcntitrlly or in parallel in lunction utilizing pulsc-density nrodula- tion. 'l'rans Inst Elcctron Inf Comnrun Eng J7.l-Cl II:369,376 the gcneral ABrain, which consists ol' the three pr.imitive li. ()teki S, F{ashimoto A, F-urula'l', cr al. (1993) A digiral neural ABrains. T'lie problem of how 1o design the T,D-ABrain network VLSI with on-chip learning using stochastic pulse cncocl- has yet to be solvecl. The dcsigns ol' both the general ing. Procoeding of the 1993 Internutional Joint Cirnlerence on Neural Networks, Nagoya. I.rpan, 1993, pp 3039 30;15 ABrain hardware and thc operating software are no\ / 9. Sugisaka M (199fi) Patter! r..cognition using ueurocorrrltuter. Iri: uncler investigation. Sugisaka M (ccl), Plocecclings of the [nternational Synrposiunt on