Active Teaching in Robot Programming by Demonstration

Transcription

1 IEEE Inernaional Symposium on Robo and Human Ineracive Communicaion (RO-MAN 7) Acive Teaching in Robo Programming by Demonsraion Sylvain Calinon and Aude Billard Learning Algorihms and Sysems Laboraory (LASA), Ecole Polyechnique Fédérale de Lausanne (EPFL) CH-5 Lausanne, Swizerland {sylvain.calinon, Absrac Robo Programming by Demonsraion (RbD) covers mehods by which a robo learns new skills hrough human guidance. In his work, we ake he perspecive ha he role of he eacher is more imporan han jus being a model of successful behaviour, and presen a probabilisic framework for RbD which allows o exrac incremenally he essenial characerisics of a ask described a a rajecory level. To demonsrae he feasibiliy of our approach, we presen wo experimens where manipulaion skills are ransferred o a humanoid robo by means of acive eaching mehods ha pu he human eacher in he loop of he robo's learning. The robo rs observes he ask performed by he user (hrough moion sensors) and he robo's skill is hen rened progressively by embodying he robo and puing i hrough he moion (kinesheic eaching). I. INTRODUCTION In a Robo Programming by Demonsraion (RbD) framework, a robo learns new skills hrough he help of a human insrucor []. Tradiional approaches in RbD end o consider he human user as an exper model who performs a ask while he robo observes passively he demonsraion, see e.g. [], []. However, in humans, eaching is a social and bidirecional process in which eacher and learner are boh acive. Insead of considering he eacher solely as a model of successful exper behavior, a recen body of work suggesed o refer o he eacher-learner couple as a eam ha engages in join problem solving [4], and o use acive eaching mehods o pu he human eacher in he loop of he robo's learning [5]. In previous work, we developed a probabilisic framework for exracing he relevan componens of a ask by observing muliple demonsraions of i [6]. The sysem is based on Gaussian Mixure Models (s) encoding a se of rajecories colleced by he robo hrough is sensors. In his work, we exend he use of our RbD framework by adoping he perspecive ha he eacher is no only a model of exper behavior bu becomes an acive paricipan in he learning process. Firs, we sugges o follow an incremenal learning approach which allows he eacher o wach he robo's reproducion aemps afer each demonsraion, and hus helps him/her assess he robo's curren undersanding of he skill and prepare he following demonsraion accordingly [7], [8]. This scaffolding process was previously proposed in roboics o le he robo gradually generalize he skill for an increasing range of conexs [4], [9]. We hen sugges o use differen modaliies o produce Fig.. Differen modaliies are used o convey he demonsraions and scaffolds required by he robo o learn a skill. The user rs demonsraes he whole movemen while wearing moion sensors (lef) and hen helps he robo rene is skill by kinesheic eaching (righ), ha is, by grasping he robo's arms and moving hem hrough he moion. he demonsraions, similarly o he eaching process where a human eacher would rs demonsrae he complee skill o he learner, followed by pracice rials performed by he learner under he supervision of he eacher (Fig. ). The learner can hen experience and adap he skill for his/her paricular body capaciies, as explored in developmenal psychology sudies []. In our RbD framework, we follow a similar sraegy by providing moion sensors and kinesheic eaching faciliies o help he user ac pedagogically wih he robo. Indeed, o become a good eacher, he user does no have o simply use his/her knowledge; he mus engage in an aciviy ha benes he learner. We ake he perspecive ha unlike observaional learning, pedagogy requires a special ype of communicaion o manifes he relevan knowledge of a skill. As discussed by [], he eacher rs needs o analyze his/her knowledge conen o emphasize in his/her demonsraions he aspecs ha are relevan for he learner. A. Sysem overview II. EXPERIMENTAL SETUP We assume ha he essenial characerisics of a skill can be exraced by observing muliple demonsraions of i. The generalizaion of a skill o differen conexs was previously explored in roboics [9], [] [7], bu mos of he approaches proposed represen he consrains a a symbolic level, which requires o segmen he skill in a preprocessing phase and, hus, o use pre-deermined conrollers o le he robo reproduce he skill. Here, we consider

2 he mos general case of represening he consrains a a rajecory level, which allows o combine several ses of consrains in a coninuous form (e.g., by considering consrains on differen modaliies or on differen objecs), and where he consrains can gradually change during he moion. Thus, our framework allows o auomaically nd a conroller for he robo ha reproduces smooh generalized rajecories saisfying several consrains simulaneously. The robo gradually builds a model of he skill by observing several demonsraions performed in various conexs (e.g., by demonsraing a manipulaion skill wih differen iniial posiions of objecs). Afer each demonsraion, he robo reproduces a generalized version of he ask by combining probabilisically he differen consrains exraced. The model is rened incremenally afer each demonsraion, and he user sops he ineracion when he robo has correcly learned he skill. B. Hardware The experimens are conduced using a Fujisu HOAP- humanoid robo wih 8 degrees of freedom (DOFs), of which 6 DOFs of he upper orso are used in he experimens. The iniial posiions of he objecs are recorded hrough a moulding process where he eacher grabs one of he robo's arm, moves i oward he objec, pus he robo's palm around he objec and presses is ngers agains he objec o le he robo feel ha an objec is currenly in is hand. When he objec ouches he palm, a force sensor inside he robo's palm is used o regiser he objec posiion, i.e., when he force sensor rerieves a value over a given hreshold, he robo briey grasps and releases he objec while regisering is posiion in D Caresian space. The user's movemens are rs recorded by 8 X-Sens moion sensors aached o he orso, upper-arms, lower-arms, hands (a he level of he ngers) and back of he head. Each sensor is used o exrac join angle rajecories by inegraing he D rae-of-urn, acceleraion and earh-magneic eld a a rae of 5 Hz and wih a precision of.5 degrees. We hen use he moor encoders of he robo o record informaion while he eacher moves he robo's arms. The eacher rs selecs he moors ha he/she wans o conrol manually by slighly moving he corresponding moors before reproducing he skill. The seleced moors are hen se o a passive mode, which allows he user o move freely hese corresponding DOFs while he robo execues he res of he ask. In his way, he eacher provides parial demonsraions while he robo acquires propriocepive informaion when performing he ask in is own environmen. The moion of each join is recorded a a rae of Hz, and each rajecory is resampled o a xed number of poins T =. The robo is provided wih moor encoders for every DOF, excep for he hands and he head acuaors. C. Probabilisic model We consider a se of rajecories {, ξ}, where each daapoin consiss of a emporal value and a spaial vecor ξ describing any kind of sensory informaion (e.g., a posure dened by join angles or a posiion in Caresian space). The spaial vecor ξ can be described eiher in he original daa space or in a laen space of moion (e.g., by projecing linearly he daa in a subspace of reduced dimensionaliy). To exrac consrains from his se of rajecories, we rs model he join probabiliy p(, x) in a Gaussian Mixure Model (), rained incremenally by a modied version of Expecaion-Maximizaion (EM) [7]. A generalized version of he rajecories is hen compued by esimaing E[p(ξ )], wih associaed consrains deermined by cov (p(ξ )). If muliple consrains are considered (e.g., considering acions ξ () and ξ () on wo differen objecs), he resuling consrains are compued by rs esimaing p(ξ ) = p(ξ () ) p(ξ () ) and hen esimaing E[p(ξ )] o reproduce he skill. For a complee descripion of he algorihms, he ineresed reader is referred o [6] for a deailed descripion of he exracion of consrains process and o [7] for he incremenal learning process. Regression is used o generalize over he se of demonsraions and o reproduce a smooh rajecory by esimaing he condiional probabiliy p(ξ ) a each ime sep. Several regression echniques based on Locally Weighed Regression (LWR) were previously proposed in roboics o generalize over a se of demonsraions [8], [9]. Our approach follows a similar sraegy by using s o represen he join disribuion of he daase {, x}, and using Gaussian Mixure Regression () o esimae p(x ), as proposed in []. By using, i is hen possible o deal wih encoding, recogniion and reproducion issues in a single framework. Apar from rerieving generalized rajecories from he demonsraions, he variaions (and correlaions) allowed around he generalized rajecories are also rerieved by he model, which are used by he robo o exend he learned skill o differen conexs and o nd an opimal conroller saisfying several consrains simulaneously. III. EXPERIMENTS We presen wo experimens o show ha he mehod is generic and can be used wih daases represening differen modaliies. In he rs experimen, he skill is represened in join space, where he / process is performed in a laen space of moion exraced by Principal Componen Analysis (PCA). In he second experimen, he skill is represened in ask space, where he / process acs on he posiion of he righ hand relaive o differen objecs in he scene. A. Learning bi-manual gesures This experimen shows how a bimanual skill can be augh incremenally o he robo in join space using differen modaliies (observaional learning and scaffolding). The ask consiss of grasping and moving a large foam die (Figs. and ). Saring from a res posure, he lef arm is moved rs o ouch he lef side of he die, wih he head following he moion of he hand. Then a symmerical gesure is The noaions E[ ] and cov( ) are used respecively o express expecaion and covariance.

3 Fig.. Illusraion of he use of moion sensors o acquire gesures from a human model. A simulaion of he robo is projeced behind he user o show he gesure observed by he robo. The gesure depiced here is similar o he one used in he experimen, excep ha in hese snapshos he user does no face he robo and only mimics he grasping of he objec. Fig. 4. Snapshos of he sixh reproducion aemp. Firs aemp Third aemp Sixh aemp Fig. 5. Lef: Experimenal seup wih frame of reference. Righ: Deniion of a posure by he posiion of he hand and an angular parameer α. Fig.. Reproducion of he ask afer he rs, hird and sixh demonsraion. In he rs aemp, he robo his he die when approaching i, making i fall. In he hird aemp, he robo's skill ges beer bu he grasp is sill unsable. In he sixh aemp, he die is grasped correcly. The rajecories of he hands are ploed in he rs row. The second row shows he corresponding snapshos of he reproducion aemps. performed wih he righ arm. When boh hands grasp he objec, i is lifed and pulled back on is base, wih he head urned oward he objec (Fig. ). The eacher performs he rs demonsraion of he complee ask while wearing moion sensors. He can hus demonsrae he full gesure by conrolling simulaneously he 6 join angles, which are hen projeced ino a subspace of lower dimensionaliy deermined by PCA. Afer observaion, he robo reproduces a rs generalized version of he moion. This moion is hen rened by physically moving he robo's limbs during he reproducion aemp, ha is, by embodying he robo and puing i hrough he moion. Noe ha he gesure can only be rened parially by guiding he desired DOFs while he robo conrols he remaining DOFs. Indeed, he eacher can only move a limied subse of DOFs by using his or her wo arms. Thus, as discussed in [8], he combinaion of observaional learning (hrough moion sensors) and kinesheic eaching echniques allow o leverage he respecive drawbacks of he wo mehods. Resuls of he experimen are presened in Figs. and 4, where he resuling hand pahs are similar o he ones demonsraed by he user (Fig. ). The sysem nds auomaically ha ve principal componens and ve Gaussian componens are sufcien o encode he moion. Afer he rs demonsraion, he robo can only reproduce a smoohed version of he gesure produced by he user. Because he user's and robo's bodies differ (he robo is smaller han he user, bu he size of he die does no change), he robo rs his he die by moving is lef hand rs, making he die fall before moving is righ hand. Observing his, he eacher progressively renes he robo's skill by providing appropriae scaffolds, ha is, by conrolling he shoulders and he elbows of he robo while reproducing he skill so ha i may grasp he die correcly. In he hird reproducion aemp, he robo lifs he die awkwardly. In he sixh aemp, he robo skillfully reproduces he ask by iself (Fig. 4). Therefore, he user decides o sop he eaching process. B. Learning o move objecs In his experimen, he robo learns how o move differen chess pieces by considering he hand-objecs relaionships a a rajecory level. The experimen draws insighs from he eaching processes described by [] in developmenal psychology. I shows ha eaching manipulaion skills o he robo can be achieved hrough a scaffolding process where he user gradually highlighs he affordances of differen Noe ha he skill is represened as join angle rajecories projeced in a laen space of moion of lower dimensionaliy.

4 Demo Demo Demo Demo 4 Demo 5 Demo 6 Knigh Bishop Rook Fig. 6. Trajecories for he six consecuive kinesheic demonsraions and for he hree differen chess pieces (only x and x are represened, corresponding o he plane of he chessboard). The cross and he plus sign represen respecively he chess piece o grasp and he opponen chess piece. Consrains for he Rook/King Consrains for he Bishop/King Consrains for he Knigh/King Daa Daa Daa Relaive o objec x () x () x () x () x () x () ˆx () ˆx () ˆx () x () x () x () x () x () x () ˆx () ˆx () ˆx () x () x () x () x () x () x () ˆx () ˆx () ˆx () Relaive o objec Daa ˆ ˆ ˆ Daa ˆ ˆ ˆ Daa ˆ ˆ ˆ Fig. 7. Exracion of he hand-objecs posiion consrains (posiion x () of he righ hand relaive o he Rook/Bishop/Knigh and posiion of he righ hand relaive o he King). In each cell, he rs column shows he rajecories, he second column shows he and he las column shows he exracion of he generalized rajecories and associaed consrains hrough. The pars of he moion showing a hin envelope around he generalized rajecories are locally highly consrained, while he pars wih a large envelope allow a loose reproducion of he rajecories. objecs in is environmen and he effeciviies of he body required o perform acions on hese objecs. A chess game is used in his experimen o explore how a robo can learn he affordances of differen chess pieces (he paricular relaions and rules associaed wih he objecs) and he associaed effeciviies (how he robo should use is body o manipulae and displace hem wihou hiing oher chess pieces). The chess paradigm was also used by [] o explore he correspondence problem in imiaion learning, ha is, o know how o reproduce a paricular moion on he chessboard by considering differen chess pieces (differen embodimens). The seup of he experimen is presened in Fig. 5. Six consecuive kinesheic demonsraions are provided o show how o grasp a Rook, a Bishop or a Knigh and bring i o he King of he adversary (Figs. 6). Three differen models are creaed for he Rook, he Bishop and he Knigh. The seup is simplied by using only wo chess pieces a he same ime (aenion scaffolding is used o le he robo recognize only

5 hese wo chess pieces), and we do no ake ino accoun he generalizaion across differen frames of reference. I means ha in our seup, he Rook is moved only in a forward linear direcion, he Bishop is moved only in a forward-lef diagonal direcion, and he Knigh is moved only wo squares forward followed by one square o he lef, forming an inverse L shape. Through he eacher's suppor, he robo exracs and combines he consrains relaed o differen objecs. Afer each demonsraion, he robo ries o reproduce he skill wih he chess pieces placed in a new conguraion ha has no been observed during he demonsraions. The user can hus es he abiliy of he robo o generalize he skill over differen siuaions. The consrains exraced afer he sixh demonsraion are presened in Fig. 7, showing he displacemen consrains wih respec o he rs objec (respecively he Rook, he Bishop and he Knigh) and he second objec (he King). For he Rook, we see ha he rajecory is highly consrained for x () from ime sep while he hand grasps he Rook, i.e., he hand-objec relaionship allows only low variabiliy during his par of he skill. The Rook is hen moved in a sraigh line, i.e., he direcion in x remains consan afer grasping of he Rook. However, is nal posiion can change in ampliude, i.e., he Rook is moved along a sraigh line bu is nal posiion on his sraigh line can vary, which is reeced by he consrains exraced for x () (larger envelope afer ime sep 7). For he Bishop, we see ha he generalized rajecories (relaive o Objec ) follows a diagonal. The direcion is highly consrained bu he nal posiion is no. For he knigh, he generalized rajecories (relaive o Objec ) are more consrained. Indeed, for a given iniial posiion of he Knigh, only one nal posiion is allowed in he proposed seup. This is reeced by he consrains for x () and ), where he and x () (and complemenary for pah wih respec o he iniial posiion of he Knigh is highly consrained (he followed pah is quasi invarian across all demonsraions). For he hree chess pieces, he consrains for he verical axis x () share similariies, showing ha he user grasps he chess piece from above, and displaces he chess piece following a bell-shaped rajecory in a verical plane. We observe ha for each chess piece he consrains for he rs objec are correlaed wih he consrains for he second objec. Indeed, he posiions of he wo objecs have imporan dependencies for he skill (i is imporan o reach he King of he opponen wih he chess piece). By considering he consrains relaive o he wo objecs, he reproducion of he hand pah is hen compued for new iniial posiions of he objecs. To do so, he absolue consrain for each objec is compued by adding he new iniial posiion of each objec o he corresponding relaive consrain (represened as a varying mean and associaed covariance marix along he moion). The nal hand pah used for reproducion is hen compued by muliplying a each Noe ha he basic rules remain he same and ha he direcions only depend on he frame of reference considered. α Rook Bishop Knigh α α Fig. 8. Exracion of he consrains on he gesure (modeled by angle α) used o move he differen chess pieces. Rook Bishop Knigh Firs aemp Third aemp Sixh aemp Fig. 9. Reproducion of he ask afer observaion of he rs, hird and sixh demonsraion. The cross and he plus sign represen respecively he chess piece o grasp and he opponen chess piece. ime sep he wo Gaussian disribuions characerizing he absolue consrains for he wo objecs (i.e., by compuing he probabiliy of wo evens in conjuncion). This hand pah is convered o join angles by using a geomerical inverse kinemaics algorihm parameerized by he angle α beween he elbow and a verical plane (Fig. 5). A generalizaion of he angle α observed during he demonsraions is used o reproduce he ask (Fig. 8). This allows o reproduce naurally-looking gesures by providing an addiional consrain o he redundan inverse kinemaics problem. For he hree chess pieces, we see ha he gesure used o reach for he chess piece share similariies. Indeed, he α rajecories sar wih a negaive value and progressively converge o zero. I means ha he elbow is rs elevaed ouward he body and is progressively lowered, approaching he body unil he arm and forearm are almos in a verical plane. This allows he user (and he robo) o approach he chess piece carefully wihou hiing he oher chess piece. Indeed, when experiencing he skill ogeher wih he robo hrough a scaffolding process, he user quickly noices ha when he robo is close o he chess piece, is elbow has o be lowered o grasp he chess piece correcly. This is mainly due o he missing DOF a he level of he wris, which consrains he grasping posure of HOAP- o a value of angle α close o zero. When he chess piece is grasped, we see ha wo differen movemen sraegies are

6 adoped depending on he pah o follow. As he Rook is moved forward, he angular conguraion does no need o be changed afer grasping. As he Bishop is moved in a diagonal (forward-lef), he user helps he robo adop a correc posure o avoid hiing is own body when performing he move (i.e., learning effeciviies). This is reeced by he decreasing negaive value of angle α. A similar sraegy is employed for he Knigh, bu he ampliude of change in angle α is lower due o he shorer pah followed by he Knigh. We also see ha here is a sligh endency o rs keep he arm in a verical plane (α close o zero), and o nally use a posure wih a sligh elevaion of he elbow (negaive α value). This behaviour is probably due o he inadveren decomposiion of he L shape by he user when helping kinesheically he robo displace he Knigh. Finally, Fig. 9 shows differen reproducion aemps for iniial conguraions ha have no been observed by he robo during he eaching process. We see ha he abiliy of he robo increases, i.e., each demonsraion helps he robo rene is model of he skill and abiliy o generalize across differen siuaions. IV. CONCLUSION We presened a probabilisic RbD framework exracing incremenally he consrains of a ask in a coninuous form, which allows o generalize and reproduce he ask in new conexs. We used wo experimens o highligh he advanages of designing a human-robo ineracion scenario mimicking he human process of eaching and aking advanages of he eaching abiliies of he user. The rs experimen showed he imporance of having mulimodal cues o srucure he demonsraed asks and o reduce he complexiy of he skill ransfer process. Indeed, we showed ha while observaional learning is a useful sraegy o demonsrae a skill in is inegriy, he use of kinesheic eaching gives he opporuniy o see and feel he soluions o he correspondence problem (deecing he mach beween self-generaed and oher moion). The second experimen showed ha eaching manipulaion skills o he robo can be achieved hrough a scaffolding process, where he user gradually highlighs he affordances of differen objecs in is environmen and he effeciviies of he body required o perform grasping and moving acions on hese objecs. ACKNOWLEDGMENT The work described in his paper was suppored in par by he Secréaria d'ea a l'educaion e a la Recherche Suisse (SER), under Conrac FP6-, Inegraed Projec Cogniron of he European Commission Division FP6-IST Fuure and Emerging Technologies, and was suppored in par by he Swiss Naional Science Foundaion, hrough gran of he SNF Professorships program. REFERENCES [] A. Billard and R. Siegwar, Special issue on robo learning from demonsraion, Roboics and Auonomous Sysems, vol. 47, no. -, pp , 4. [] K. Ikeuchi and T. Suchiro, Towards an assembly plan from observaion, par I: Assembly ask recogniion using face-conac relaions (polyhedral objecs), in Proceedings of he IEEE Inernaional Conference on Roboics and Auomaion (ICRA), vol., no. -4, May 99, pp [] Y. Kuniyoshi, M. Inaba, and H. Inoue, Learning by waching: Exracing reusable ask knowledge from visual observaion of human performance, IEEE Transacions on Roboics and Auomaion, vol., no. 6, pp , 994. [4] C. Breazeal, A. Brooks, J. Gray, G. Hoffman, C. Kidd, H. Lee, J. Lieberman, A. Lockerd, and D. Chilongo, Tuelage and collaboraion for humanoid robos, Humanoid Robos, vol., no., pp. 5 48, 4. [5] K. Dauenhahn, The ar of designing socially inelligen agens: Science, cion, and he human in he loop, Applied Aricial Inelligence, vol., no. 7-8, pp , 998. [6] S. Calinon, F. Guener, and A. Billard, On learning, represening and generalizing a ask in a humanoid robo, IEEE Transacions on Sysems, Man and Cyberneics, Par B. Special issue on robo learning by observaion, demonsraion and imiaion, vol. 7, no., pp , 7. [7] S. Calinon and A. Billard, Incremenal learning of gesures by imiaion in a humanoid robo, in Proceedings of he ACM/IEEE Inernaional Conference on Human-Robo Ineracion (HRI), March 7, pp [8], Wha is he eacher's role in robo programming by demonsraion? - Toward benchmarks for improved learning, Ineracion Sudies. Special Issue on Psychological Benchmarks in Human-Robo Ineracion, vol. 8, no., 7. [9] J. Saunders, C. Nehaniv, and K. Dauenhahn, Teaching robos by moulding behavior and scaffolding he environmen, in Proceedings of he ACM SIGCHI/SIGART conference on Human-Robo Ineracion (HRI), March 6, pp [] P. Zukow-Goldring, Caregivers and he educaion of he mirror sysem, in Proceedings of he Inernaional Conference on Developmen and Learning (ICDL), 4, pp. 96. [] G. Gergely and G. Csibra, The social consrucion of he culural mind: Imiaive learning as a mechanism of human pedagogy, Ineracion Sudies, vol. 6, pp , 5. [] S. Muench, J. Kreuziger, M. Kaiser, and R. Dillmann, Robo programming by demonsraion (RPD) - Using machine learning and user ineracion mehods for he developmen of easy and comforable robo programming sysems, in Proceedings of he Inernaional Symposium on Indusrial Robos (ISIR), 994, pp [] M. Nicolescu and M. Maaric, Naural mehods for robo ask learning: Insrucive demonsraions, generalizaion and pracice, in Proceedings of he Inernaional Join Conference on Auonomous Agens and Muliagen Sysems (AAMAS),, pp [4] J. Seil, F. Röhling, R. Haschke, and H. Rier, Siuaed robo learning for muli-modal insrucion and imiaion of grasping, Roboics and Auonomous Sysems, vol. 47, no. -, pp. 9 4, 4. [5] A. Alissandrakis, C. Nehaniv, K. Dauenhahn, and J. Saunders, An approach for programming robos by demonsraion: Generalizaion across differen iniial conguraions of manipulaed objecs, in Proceedings of he IEEE Inernaional Symposium on Compuaional Inelligence in Roboics and Auomaion, 5, pp [6] S. Ekvall and D. Kragic, Learning ask models from muliple human demonsraions, in Proceedings of he IEEE Inernaional Symposium on Robo and Human Ineracive Communicaion (RO-MAN), Sepember 6, pp [7] M. Pardowiz, R. Zoellner, S. Knoop, and R. Dillmann, Incremenal learning of asks from user demonsraions, pas experiences and vocal commens, IEEE Transacions on Sysems, Man and Cyberneics, Par B. Special issue on robo learning by observaion, demonsraion and imiaion, vol. 7, no., pp., 7. [8] S. Schaal and C. Akeson, Consrucive incremenal learning from only local informaion, Neural Compuaion, vol., no. 8, pp , 998. [9] S. Vijayakumar, A. D'souza, and S. Schaal, Incremenal online learning in high dimensions, Neural Compuaion, vol. 7, no., pp. 6 64, 5. [] Z. Ghahramani and M. Jordan, Supervised learning from incomplee daa via an EM approach, in Advances in Neural Informaion Processing Sysems, J. D. Cowan, G. Tesauro, and J. Alspecor, Eds., vol. 6. Morgan Kaufmann Publishers, Inc., 994, pp. 7. [] A. Alissandrakis, C. Nehaniv, and K. Dauenhahn, Imiaion wih ALICE: Learning o imiae corresponding acions across dissimilar embodimens, IEEE Transacions on Sysems, Man, and Cyberneics, Par A: Sysems and Humans, vol., no. 4, pp ,.