A Lower-Limb Power-Assist Robot with Perception-Assist

Transcription

1 20 IEEE International Conerence on Reabilitation Robotics Reab Wee Zuric, ET Zuric Science City, Switzerland, June 29 - July, 20 A Lower-Limb Power-Assist Robot wit Perception-Assist Yosiai ayasi Dept. Advanced Tecnology Fusion, Saga University, Saga, Japan ayasi@me.saga-u.ac.jp Kazuo Kiguci Dept. Advanced Tecnology Fusion, Saga University, Saga, Japan iguci@ieee.org Abstract In order to assist te motion in te daily lives o pysically wea persons suc as elderly persons, many inds o power-assist robots ave been developed. In te case o some pysically wea persons, te ability to perceive te environment is sometimes deteriorated also. A metod o perception-assist as been proposed to assist not only te user s motion but also te user s interaction wit an environment, by applying te modiication orce to te user s motion i it is necessary. In tis paper, te perception-assist or a lower-limb power-assist exoseleton robot is proposed. In te daily lie, te waling is very important or persons to acieve desired tass. Basically, te robot assists te user s muscle orce according to te user s motion intention wic is estimated based on EMG signals. I te robot as ound problems wic migt lead te user to dangerous situation suc as te alling down, te robot tries to modiy te user's motion in ition to te ordinal power-assists to mae te user wal properly. Since te user migt all down by te eect o te itional modiication orce o te perception-assist, te robot automatically prevents te user rom alling down by considering ZMP (Zero Moment Point). Te eectiveness o te proposed metod as been evaluated by perorming experiments. Keywords- perception-assist, all prevention, power-assist. I. INTRODUCTION Some persons ave diiculties in living an independent lie because o a disease or a decline o muscle orce. Power-assist robots are useul to elp pysically wea persons suc as elderly persons live an independent lie and are also expected to be applied to various ields []-[5]. In te power-assist robots or pysically wea persons, it is important to assist te user s motion automatically according to te user s intention. Te power-assist robots are able to estimate te user s motion intention based on bio-logical signals lie electromyogram (EMG) [7, 8]. owever, in te case o some elderly or pysically wea persons wose motor ability is deteriorated, te ability to perceive te environment is sometimes deteriorated also. In suc case, even toug te user expands te motor activity range by using te power-assist robot, te user migt not be able to perceive te riss associated wit teir own motions correctly and be involved in an unexpected accident. Tereore, it is sometimes necessary or te powerassist robot to ave te ability o te perception-assist as well as te ability o te ordinary power-assist. Te power-assist robot wit perception-assist as been proposed to eep monitoring te interaction between te user and te environment using its own sensors, and tries to modiy te user s motion automatically by ing itional orce only i any problems are ound in te user s motion. For example, in te case o upper-limb power-assist robots, te user acieves tass using various tools. I te robot judges te tool motion is wrong or dangerous, te robot automatically tries to modiy te tool motion by ing itional orce to te user s and [6]- [20]. In te perception-assist, sensors suc as stereo cameras and/or laser range inders are necessary to monitor te interaction between te user and te environment. In tis paper, te perception-assist or a lower-limb powerassist exoseleton robot is proposed to mae elderly and/or pysically wea persons wal properly and saely in a daily living environment. I a person wose ability to perceive te environment is deteriorated misidentiies a bump, te person migt stumble on te bump and all down because o te spatial misconception, wic could result in serious damage. To prevent a user away rom an unexpected accident, te proposed power-assist robot observes te interaction between te user and te environment in ront o te user and modiies te user s motion as necessary by ing te modiication orce automatically, altoug te robot always perorms te ordinary power-assist normally as well. owever, te user migt lose is/er balance and all down by te eect o te itional modiication orce o te perception-assist because it is automatically generated regardless o te user's motion intention. Consequently, te perception-assist is carried out considering ZMP (Zero Moment Point) [2] to eep stable waling in te proposed metod. In tis study, te proposed perception-assist metod is applied to a 2DOF lower-limb power-assist exoseleton robot and te eectiveness o te proposed metod is evaluated by perorming experiments. II. LOWER-LIMB POWER-ASSIST ROBOT Te lower-limb power-assist exoseleton robot used in tis study is sown in Fig.. Eac limb o te robot consists o two DC motors wit encoders to measure te ip and nee joint angles, An encoder to measure te anle joint angle, two tactile switces in te eel and te toe to now weter te sole contacts wit te ground, lins or tig and san, and a oot older. Te robot is attaced to te torso, tig, san, and oot //$ IEEE 73

2 Fig. Lower-limb power-assist exoseleton robot. Table I Range o motions o lower limb Type o motion uman [deg] robot [deg] ip lexion ip extension Knee lexion 40 0 Knee extension 0 0 o te user. ip lexion/extension motion and nee lexion/extension motion o te user can be assisted by te robot. Te range o motion o eac joint is sown in Table I. As sown in Table I, te range o motion o eac joint o te robot is sligtly smaller tan te pysiological range o motion in order to ensure te saety. Te laser range inder (URG-04LX, OKUYO AUTOMATIC CO., LTD.) is attaced near te nee joint o te robot. In ition, te encoder is installed to measure te rotation o te axis o te laser range inder. Te laser range inder is used to renize te environment in ront o te user in x-z plane as sown in Fig.. III. POWER-ASSIST Te controller o te robot consists o te power-assist part and te perception-assist part. Eac motor torque is te sum o calculated torques in eac part. In tis section, te power-assist part is explained in detail. Te sin surace electromyogram (EMG) is used to predict te user s motion intention and assist te user s motion in realtime in accordance wit te user s motion intention, since it directly relects te user s motion intention [7]. Seven cannels o EMG signals per leg are measured and used as input signals to te robot. Eac cannel mainly corresponds to one muscle. Te locations o EMG electrodes are sown in Fig. 2. Since te raw EMG signal is not suitable or te control input, a eature o te EMG signal is extracted. We cose to use te Root Mean Square (RMS). Te RMS calculation is written as RMS = N 2 v i N were N is te number o samples in a segment, v i is te voltage at i t sampling point. Te number o samples N is set to be 300 and te sampling time is 500μsec (te sampling requency is 2 z) in tis study. Te relationsip between te seven RMS values and te generated vector o joint torques is implemented as sown in eq. (2). () Fig. 2 Location o te EMG Electrodes. Fig. 3 Fuzzy-Neuro Modiier. τ, pow = w τ, pow w w2 w 2 L L c c w 2 6 w7 M w6 w7 c 6 c7 were τ,pow and τ,pow are torques or ip lexion/extension and nee lexion/extension, respectively. c j represents te RMS value o te EMG signal measured in cannel i. w j is te weigt value or j t EMG signal to estimate te ip lexion/extension motion and w j is te weigt value or j t EMG signal to estimate te nee lexion/extension motion. Te relationsip between eac joint torque and te RMS values o te EMG signals is canged depending on te user s lowerlimb posture. Tereore, te weigts o te matrix in eq. (2) are adjusted in real time according to te user s lower-limb posture by using a neuro-uzzy modiier wic is sown in Fig. 3 [7]. Tis adaptation tecnique wit neuro-uzzy modiier enables te controller to adjust itsel to be suitable or any posture o any person. w j and w j are adjusted as ollows by using neurouzzy modiier as sown in Fig. 3. wj = w0, j CWj (3) w = w CW j 0, j were w 0j and w 0j are te initial weigt values. CW j and CW Kj are canged according to te user s lower-limb posture in order to adjust te relationsip between eac joint torque and te RMS values. Kj (2) 732

3 IV. Fig. 4 Bump detection. PERCEPTION-ASSIST In te perception-assist part, te robot judges weter te user s motion is in danger at irst. Te user is able to move is/er body according to is/er motion intention by te assist o te robot. owever, i te user as problems in is/er ability to perceive te environment, te user migt not be able to renize te danger o own motion correctly. Tereore, i te robot judges te user s motion is dangerous, te robot needs to modiy te user s motion automatically. In order to realize te perception-assist, te robot is required to understand not only te motion o te user but also te interaction between te user and te surrounding environment. In te case o lowerlimb power-assist robots, te robot needs to prevent te user rom alling down during waling, ascending, descending, or standing up. Speciically, a user wo as problems in te ability to perceive te environment migt overloo or misunderstand a bump and stumble on it. Altoug a ealty person migt be able to recover te own balance, a user wo needs te elp o te power-assist robot migt all down. For tis reason, te robot needs to be careul wit te user s interaction wit te environment. In tis paper, te robot detects bumps and prevents te user rom alling down and stumbling by te perception-assist during waling. owever, altoug te robot tries to modiy te user s motion in ition to te ordinary power-assists to wal properly, te motion modiication is automatically generated regardless o te user s intention. For tis reason, te user migt lose is/er balance and all down by te eect o te itional modiication orce o te perception-assist. To prevent suc case, te robot taes into account ZMP in te perception-assist algoritm. A. Environmental sensing Te laser range inder is used to renize te environment in ront o te user. A bump in ront o te user is detected by te laser range inder and te distances rom te user s toe to te bump and te eigt o te bump are calculated. Te robot is careul wit te user s motion so tat te user does not stumble on te bump. Te measured points rom te laser range inder are sown in Fig. 4. As sown in Fig. 4, eac vector is calculated based on tese adjacent measured points. Ten, θ wic is te angle between tese vectors and te orizontal vector can be calculated. Te robot estimates weter tere is a bump or not based on θ. Wen te absolute value o θ is small enoug, te robot judges tat te ground is lat. On te oter and, wen te absolute value o θ is large enoug, te robot judges tat tere is a bump in ront o te robot. Wen te absolute value o θ is middle, te robot judges tat te ground is slope. Tese tresold values are deined experimentally. I te robot detects a bump, te distance rom te user s toe to te bump and te eigt o te bump are recorded. Te scanning time and te angular resolution o te laser range inder are 80 msec and 0.36 deg, respectively. B. ZMP In order to prevent te user rom losing is/er balance and alling down by te eect o te itional modiication orce by te perception-assist wic is given regardless o te user s intention, te robot taes into account ZMP. Te user s body is modeled as sown in Fig. 5. Te supporting leg is ound by tactile switces located in bot soles. Ten, te posture o te user s body region is c calculated using te joint angles o te supporting leg as ollows. θ = θ θ + (7) b θ a were θ b is te angle o upper body region, θ, θ and θ a are te angles o ip, nee and anle joint o te support leg, respectively. θ, θ and θ a are te angles o ip, nee and anle joint o te swing leg. Te positions o eac joint are calculated as ollows x z x z x z = x = z Fig. 5 User posture. = Ll sinθ = L cosθ = x = z l a a L sinθ + L cosθ + L sinθ L cosθ x a = x + Ll sin( θ θ ) z a = z Ll cos( θ θ ) (8) (9) (0) () were L l and L are te lengts o te san region and te emoral region, respectively. x and z are te positions o eac joint. Te sp represents te supporting leg, sw represents te swing leg, a, and represent anle, nee and ip joint, respectively. In ition, θ = θ - θ b, θ = θ - θ b. Te positions and accelerations o te center o gravity (COG) o upper body, tig and san are calculated rom (7) to (). ZMP is calculated is written as 733

4 Fig. 6 Perception-assist area based on ZMP. Fig. 8 Flowcart o perception-assist. xzmp = Fig. 7 Virtual wall area. n mi (&& zi n + g) x i mi (&& zi n + g) mi && xi zi (2) were x zmp is te coordinate o ZMP in te orizontal direction, m i is te mass o eac uman body part and robot, x i is te position o te COG o eac part in te orizontal direction, z i is te position o te COG o eac part in te vertical direction and g is te acceleration o gravity. Te weigt and te position o te COG o eac body part are decided based on [22]. Te perception-assist is perormed considering x zmp. C. Fall prevention assist Te robot perorms all prevention assist based on te results o environmental sensing and calculated ZMP. Wen te robot estimates tere is no bump in ront o te user, te robot evaluates te user s waling based on x zmp in order to prevent te user rom alling down in ition to te ordinary power-assist. I x zmp is located in an undesired area, te robot automatically generates itional ip joint torque to cange te location o x zmp to te sae area by activating te upper body o te user. Te area in wic te robot generates te itional ip joint torque in ition to te power-assist is deined as sown in Fig. 6 so tat x zmp is moved to center o te support polygon. Because te mass o te upper body region is larger tan oter regions and te movement o te mass o te upper body is te most eective way to cange ZMP, te itional ip joint torque (i.e., te perception-assist torque) is generated at te ip joint motor o te support leg. In Fig. 6, τ max is te amount o te maximum ip joint torque wic is deined to protect te user rom te injury. Te perceptionassist torque or waling-assist is calculated using eq. (3). 0, ( Δxzmp < d sae ) τ max ( Δxzmp Sgn d sae ), τ wal = dcri (3) ( d sae Δxzmp d sae + dcri ) Sgn τ, ( + ) max Δxzmp d sae dcri Δxzmp = xc sgn(, ZMP xzmp Sgn = Δxzmp ) were x c,zmp is te center position o te support polygon, d sae and d cri are sown in Fig. 6. τ max, d sae and d cri are experimentally deined. Wen te robot detects a bump, a virtual wall area is generated in ront o te bump as sown in Fig. 7(a). It assumes tat te irst point wic is judged as a bump by te environmental sensing is te origin. Te virtual wall area is sown by te trapezoid represented by Δl, Δ and in Fig. 7 (a). ere, means te eigt o te bump, Δl, Δ are te experimentally deined parameters as sown in Fig. 7(a). In te experiments to deine Δl and Δ, some subjects wal at normal speed (about.4m/sec) [23]. Ten, we deined Δl and Δ as minimum values in wic te subjects could overcome te bump. Te location o te toe o te user s swing leg is calculated using eq. (). I te user s oot enters te virtual wall area as sown in Fig. 7(a), te robot judges te user migt stumble on te bump and generates te itional motion modiication orce along te virtual wall in ition to te power-assist orce to avoid te collision wit te bump. Te itional motor torque generated by te eect o te itional motion modiication orce (i.e., te perception-assist torque) is written as ollows. τ τ, = J T, (4) were τ, is te itional ip joint torque, τ, is te itional nee joint torque, J is Jacobian matrix and is te itional motion modiication orce generated along te virtual wall. Te amount o is related wit Δl and Δ. Te amount o was deined based on te cange o ZMP and te eeling o te subjects at te same time wen Δl and Δ are deined. Note tat τ, and τ, are generated witout te user s intention. Tereore, anoter ip joint torque must be 734

5 (a) ZMP wen te robot generated te perception-assist orce witout taing into account ZMP (bump : widt o about 300mm, eigt o about 200mm ) Fig. 9 Experimental condition. ed to cancel te undesired eect o te ZMP cange caused by te itional orce so tat te user does not lose is/er balance and all down by te eect o te perception-assist. Based on eq. (2), τ cancel wic is te torque to cancel te eect o te cange o ZMP is calculated as ollows. body τ d τ = d, l,, t, xl, xt, sinθ z x sinθ z sin( θ θ ) x cancel b b t, + zl, sinθ z cosθ cosθb cos( θ cosθ body d b θ ) (5) τ = (6) were d t, is te lengt between ip joint and te COG o te emoral part, d l, is te lengt between nee joint and te COG o te leg part and d are te lengt between ip joint and te COG o te body part. x p, and z p, are te position o eac body part (p=t:emoral part, l:leg part, b:body part). Ater te user s oot is lited over te bump, te robot estimates weter te top surace o te bump as enoug space to place te user s oot on it. I te top surace o te bump as enoug space to place te user s oot on it, te robot taes te itional perception-assist torque away rom te user and perorms te normal power-assist. In te case tat te top surace o te bump is too narrow, i te robot renizes te ground aead on te bump, te robot perorms anoter perception-assist to overcome te bump as sown in Fig. 7(b). Meanwile, i te robot cannot renize te ground aead on te bump, te robot tries to turn bac te user s oot because te robot cannot ind te space to place te user s oot on it saety. Tus, te robot taes into account ZMP to prevent te user rom losing is/er balance and alling down in any case. Te lowcart o te perception-assist is sown in Fig. 8. Te robot perorms te perception-assist in accordance wit te lowcart sown in Fig. 8 in ition to te power-assist represented in eq. (2). V. EXPERIMENTS In order to veriy te eectiveness o te proposed metod, experiments were carried out. In te experiments, a young male subject waled on te treadmill as sown in Fig. 9. For te irst experiments, a bump (widt o about 300mm, eigt o about 200mm) was prepared in ront o te subject. (b) ZMP wen te robot generated te perception-assist orce wit taing into account ZMP (bump : widt o about 300mm, eigt o about 200mm ) (c) ZMP and joint torques wen te robot generated te perception-assist orce wit taing into account ZMP (bump : widt o about 30mm, eigt o about 40mm ) Fig. 0 Experimental results. Ater te subject waled on te treadmill 0 seconds, te subject moved is oot to stumble on te bump on purpose in order to veriy te eectiveness o te proposed metod. Te robot generates te perception-assist torque to prevent te subject rom stumbling on te bump wen te subject s oot enters in te virtual wall area. In te experiments, te robot did not generate te perception-assist orce during 0 seconds rom te beginning. Te experimental result wit te perception-assist orce witout taing into account ZMP is sown in Fig. 0(a) and tat wit taing into account ZMP is sown in Fig. 0(b). In te bot cases, te subject was able to place is oot on te bump witout stumbling. owever, in te case o Fig. 0(a) in wic te robot assisted te subject witout taing into account ZMP, te subject lost is balance by te unintended itional orce rom te perception-assist, and ZMP moved bacward opposite to te direction o movement at 4-5 seconds. Meanwile, in te case o Fig. 0(b) in wic te robot assisted te subject wit taing into account ZMP, te subject ept is balance even toug tere was te itional orce rom te perception-assist, and ZMP was stable. In te experiments, te robot generated larger itional orce by te perception-assist tan usual ( wic was deined experimentally) in order to 735

6 veriy te eectiveness o te proposed metod because te subject was a young ealty male as remared above. For te second experiments, anoter bump (widt o about 30mm, eigt o about 40mm) was prepared in ront o te subject. Tis bump was narrower and lower tan te bump in te irst experiments. Te irst experiment situation is te case tat te user cannot lit up is oot up to a eigt o a bump altoug te user can renize te bump. Te second experiment situation is te case tat te user migt not be able to renize a bump because te bump is small. In tis experiment, ater te subject waled on te treadmill 0 seconds, te subject placed te rigt oot on te bump on purpose. Altoug te subject stepped on te bump during waling, te robot avoided it by te itional orce by te perception-assist. Also in tese experiments, te robot did not generate te perception-assist torque during 0 seconds rom te beginning. Te experimental result is sown in Fig. 0(c). In tis case, te robot renized te bump at about 2 seconds, and ten generated te perception-assist orce to lit up te subject s oot. In ition, te robot renized te widt o te bump is narrow to place te oot on it ater about 5 seconds, and ten te robot generated te anoter itional orce so tat te subject stepped over te bump. During te experiments, te robot too into account ZMP to prevent te subject rom alling down. Tese experimental results sow tat te proposed perception-assist wit taing into account ZMP can prevent te user rom alling down. VI. CONCLUSION In tis paper, te perception-assist or a lower-limb powerassist exoseleton robot is proposed to assist te waling o te elderly or pysically wea persons. In te proposed perceptionassist, te power-assist robot observes te environment in ront o te user and modiies te user s motion as necessary to prevent a user rom alling down, wile te robot perorms te ordinary power-assist simultaneously. In ition, te proposed metod perorms te perception-assist taing into account ZMP to eep stable waling. Te experimental results wic were carried out by te ealty young man sowed te eectiveness o te proposed metod. ACKNOWLEDGMENT Te autors would lie to tan Saga Arizono Prostetic- Ortotic Company or teir cooperation to mae te lowerlimb power-assist robot used in tis study. REFERENCES [] A.B. Zoss,. Kazerooni, and A. Cu, Biomecanical Design o te Bereley Lower Extremity Exoseleton (BLEEX), IEEE/ASME Transactions on Mecatronics, (2), 2006, pp [2] S. Lee and Y. Sanai, Power Assist Control or Waling Aid wit AL-3 Based on EMG and Impedance Adjustment around Knee Joint, Proc. o IEEE/RSJ International Con. on Intelligent Robots and Systems, 2002, pp [3] K. Nagai and I. Naanisi, Force Analysis o Exoseletal Robotic Ortoses or Judgment on Mecanical Saety and Possibility o Assistance, Journal o Robotics and Mecatronics, 6(5), 2004, pp [4] L. Lucas, M. DiCicco, and Y. Matsuoa, An EMG-Controlled and Exoseleton or Natural Pincing, Journal o Robotics and Mecatronics, 6(5), 2004, pp [5] J. Rosen, M. Brand, M. Fucs, and M. Arcan, A Myosignal-Based Powered Exoseleton System, IEEE Trans. on System Man and Cybernetics, Part A, 3(3), 200, pp [6] N.G. Tsagaris and D.G. Caldwell, Development and Control o a Sot-Actuated Exoseleton or Use in Pysioterapy and Training, Autonomous Robots, 5(3), 2003, pp [7] K. Kiguci and Q. Quan, Muscle-Model-Oriented EMG-Based Control o an Upper-Limb Power-Assist Exoseleton wit a Neuro-Fuzzy Adjuster, Proc. o IEEE World Congress o Computational Intelligence, 2008, pp [8] R.A.R.C. Gopura and K. Kiguci, SUEFUL-7: A 7DOF Upper-Limb Exoseleton Robot wit Muscle-Model-Oriented EMG-Based Control, Proc. o IEEE/RSJ International Con. on Intelligent Robots and Systems, 2009, pp [9] A.B. Zoss,. Kazerooni, Biomecanical Design o te Berley Lower Extremity Exoseleton (BLEEX), in Proc. o te 2006 IEEE/ASME Transactions on Mecatronics, Vol., No. 2, 2006, pp [0] C. Fleiscer and G. ommel, Embedded Control System or a Powered Leg Exoseleton, Embedded Systems - Modeling, Tecnology, and Applications, 2006, pp [] A.M. Dollar and. err, Design o a quasi-passive nee exoseleton to assist running, Proc. o IEEE/RSJ International Con. on Intelligent Robots and Systems, 2008, pp [2].K. Kwa, J.. Noorden, M. Missel, T. Craig, J.E. Pratt, P.D. Neuaus, Development o te IMC Mobility Assist Exoseleton, Proc. o IEEE International Con. on Robotics and Automation, 2009, pp [3] S. Lee, Y. Sanai, Te natural requency-based power assist control or lower body wit AL-3, Proc. o InternationalConerence on Systems, Man and Cybernetics, 2003, pp [4] K. Suzui, G. Mito,. Kawamoto, Y. asegawa,y. Sanai, Intention- Based Waling Support or Paraplegia Patient wit Robot SuitAL, Climbing and Waling Robots, 200, pp [5] K. Kong and D. Jeon, Fuzzy Control o a New Tendon-Driven Exoseletal Power Assistive Device, Proc. o IEEE/ASME International Conerence on Advanced Intelligent Mecatronics, 2005, pp [6] K. Kiguci and M. Liyanage, A Study on a 4DOF Upper-Limb Power- Assist Exoseleton wit Perception-Assist, Proc. o International Con. on Biomedical Electronics and Devices,, 2008, pp [7] K. Kiguci and M. Liyanage, A Study o a 4DOF Upper-Limb Power Assist Intelligent Exoseleton wit Visual Inormation or Perception- Assist, Proc. o IEEE International Con. on Robotics and Automation, 2008, pp [8] K. Kiguci, M. Liyanage and Y. Kose, Perception Assist wit Active Camera or an Upper-Limb Power-Assist Exoseleton, Proc. o 3rd Asia International Symposium on Mecatronics, 2008, pp [9] K. Kiguci, M. Liyanage and Y. Kose, Intelligent Perception Assist Wit Optimum Force Vector Modiication For An Upper-Limb Power- Assist Exoseleton, Proc. o IEEE International Con. on Biomedical Robotics and Biomecatronics, 2008, pp [20] K. Kiguci, M. Liyanage, Y. Kose, Y. ayasi, and D. Jeon, 3D Perception-Assist or an Upper-Limb Power-Assist Exoseleton, Proc. o te 4t IASTED International Con. on Robotics and Applications, 2009, pp [2] A. Tsuaara, R. Kawanisi, Y. asegawa, Y. Sanai, Sit-to-Stand and Stand-to-Sit Transer Support or Complete Paraplegic Patients wit Robot Suit AL, Advanced Robotics Vol.24, No., 200, pp [22] K. Isii Kiaci, T. Nisiyama, Textboo series or pysical education, Sports and ealt Sciences, Icimura Publising ouse Co., (in Japanese) [23] W. Naano, Y. Oasi, Te Inluence o Age and Waling Speed on te Adjustment o Step Lengt or Stepping Over Obstacles, Pysical Terapy Japan, Vol. 37, No. 3, 2002, pp (in Japanese) 736