HUMANOID robots continue to be an active topic for

Transcription

1 1050 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 59, NO. 2, FEBRUARY 2012 Dependable Humanod Navgaton System Based on Bpedal Locomoton Yeonsk Kang, Hyunsoo Km, Soo-Hyun Ryu, Nakju Lett Doh, Assocate Member, IEEE, YongHwan Oh, and Bum-jae You, Member, IEEE Abstract In ths paper, a dependable humanod navgaton system s proposed by consderng many dffcultes n humanod navgaton based on bpedal locomoton n an uncertan envronment. In partcular, we propose a layered archtecture to resolve complcated problems through a herarchcal manner. Wthn the proposed software archtecture, a walkng path planner, a walkng footstep planner, and a walkng pattern generator are ntegrated n a herarchy to create a relable moton that overcomes foot slppage and localzaton sensor nose. Each layer s desgned to overcome dffcultes orgnatng from bpedal locomoton such as unstable dynamcs, ncluson of a snusodal nose component n the localzaton sensor measurement, and dsturbance regardng dscrete footsteppng. The desgned navgaton system s mplemented on a human-szed expermental humanod platform and s tested for the evaluaton of ts relablty and robustness n varous tasks. Index Terms Bpedal locomoton, humanod navgaton, localzaton, software archtecture. I. INTRODUCTION HUMANOID robots contnue to be an actve topc for research, ncludng dverse research subjects such as machne vson, artfcal ntellgence, manpulaton, and balancng of robot body postures. These are fundamental capabltes n realzng a humanod robot system that can provde servces for elderly people or patents. We have developed a humanod robot platform that s used to mplement results of research actvtes on task executon related to home servces. So far, many dfferent types of untary acton have been mplemented, such as balanced walkng, vsual servong, star clmbng, object recognton, etc. [1] [6] In the future, home servces such as house cleanng, garbage collecton, meal preparaton, and delvery are expected to be provded by the developed humanod robot. Bpedal locomoton s one of the most dstnctve features of the developed humanod robot, whch enables the robot to share the same workspace wth humans wthout extra modfcaton of the envronment. It s usually dffcult for a humanod robot Manuscrpt receved November 4, 2010; revsed May 7, 2011; accepted May 19, Date of publcaton July 22, 2011; date of current verson October 18, Y. Kang s wth the Department of Automotve Engneerng, Kookmn Unversty, Seoul , Korea (e-mal: ykang@kookmn.ac.kr). H. Km, Y. Oh, and B. You are wth the Korea Insttute of Scence and Technology, Seoul , Korea (e-mal: scwtbn@hanmal.net; oyh@kst.re.kr; ybj@kst.re.kr). S.-H. Ryu and N. L. Doh are wth the School of Electrcal Engneerng, Korea Unversty, Seoul , Korea (e-mal: okrsh@korea.ac.kr; nakju@korea.ac.kr). Color versons of one or more of the fgures n ths paper are avalable onlne at Dgtal Object Identfer /TIE to walk n a stable manner whle achevng proper moton wthout fallng down because a robot has many degrees of freedom, and hence, control of ts moton s a complex task. In mplementng proper walkng moton, ths paper employs a walkng pattern generaton method whch s based on the lnear nverted pendulum model (LIPM) [7]. There are numerous exstng walkng pattern generaton methods. In partcular, a set of partcles s proposed n descrbng the humanod robot, and Fourer seres s used to generate stable walkng n [8]. In [9], a walkng moton wth a neural oscllator s mplemented, whch mtates the nervous system of the human body. An analytcal method for real-tme gat plannng for humanod robots usng a cubc polynomal of zero moment pont (ZMP) s ntroduced n [10]. Reference [11] also proposed an analytcal method usng lnear polynomals n whch the nonmnmum phase property of the LIPM s consdered. In [12], a walkng pattern generaton method s proposed usng prevew control. A control scheme that conssts of a feedback controller and a feedforward controller for a stable walkng pattern s proposed n [13]. Dfferent reference generaton methods are proposed based on the LIPM n [14] and [15]. In [16], a walkng trajectory generaton method based on three-mass model s proposed. Although there have been substantal mprovements regardng humanod balancng and control, untl recently, the navgaton of humanod robots has drawn the attenton of only a lmted number of research groups such as [17] [19]. The developed methods are based on footstep plannng gven a precse map of the workspace wthout consderng uncertantes n localzaton and foot slppage. In order to overcome the localzaton error, methods based on vsual trackng [20] or partcle flters [21], [22] are proposed. However, these methods requre heavy computatonal capabltes for fast vson mage processng or partcle flter processng. The man research nterest of ths paper s on developng a robust navgaton control strategy for a humanod robot that remans relable on the face of foot slppage and uncertanty n localzaton. There exst several dffcultes that need to be consdered for robust humanod navgaton. The frst one s ts nherent unstable dynamcs smlar to the nverted pendulum control problem. Snce foot moton s coupled wth body posture stablzaton, a robot has to plan ts future footsteps n correspondence wth the mass center of ts whole body, whch makes t dfferent from the conventonal moble robot navgaton problem. The second dffculty also orgnates from the nature of bpedal locomoton. As the humanod robot moves forward, backward, or laterally, the movement of the robot s center of mass (CoM) nduces snusodal oscllaton of ts body. Although the oscllaton s /$ IEEE

2 KANG et al.: DEPENDABLE HUMANOID NAVIGATION SYSTEM BASED ON BIPEDAL LOCOMOTION 1051 unform and known n advance, t makes the localzaton of the humanod robot complcated especally when the localzaton sensor accuracy s poor. The thrd one s the avodance of foot collson, whch results n a nonzero turnng radus of the robot walkng trajectory durng path plannng. The generated path should take nto account ths constrant and should prevent collson between feet. Fnally, the control nput delay of the footstep nput should be consdered. Because the foot moton s coupled wth body posture stablzaton, an mmedate change n the footstep plan s not allowed, and the control nput needs to be stored n an nput command buffer before t s executed. In ths paper, we develop a dependable control strategy that takes nto account the aforementoned challenges. Frst, a trackng control algorthm based on a dscrete walkng model of humanods s proposed. The walkng model ncludes foot slppage errors n both translaton and rotaton, whch many prevous studes on footstep plannng do not take nto account. Next, a localzaton method s developed to estmate the central walkng path out of oscllatory poston measurements resultng from the aforementoned snusodal CoM and ZMP pattern for balanced walkng. Fnally, a navgaton software archtecture s proposed, whch ncludes multple software modules for walkng path planner, walkng footstep planner, walkng pattern generaton, localzaton, etc. These modules are seamlessly ntegrated through the developed software archtecture and allow the humanod to follow the desgned path to the goal poston. The overall navgaton software archtecture s mplemented and tested through an expermental humanod hardware platform as well as realstc smulatons that ncorporate the knematcs of the humanod s whole-body moton. In Secton II, the overall control archtecture of the developed navgaton software module s presented. Secton III descrbes the path-plannng algorthm based on a vsblty graph (VG). The modelng and controller desgn of dscrete footsteps plannng of the humanod are explaned n Secton IV. In Secton V, the walkng pattern generaton procedure s explaned, followed by Secton VI whch descrbes the proposed method for estmatng the humanod robot s pose. In Secton VII, the developed footstep plannng and control method s verfed by a numercal smulaton that solves the knematcs of the humanod whole-body moton and the resultng footstep trajectores. In Secton VIII, the performance of the developed navgaton archtecture s nvestgated through executon on a human-szed robot hardware platform. The concluson s presented n Secton IX. II. NAVIGATION SOFTWARE ARCHITECTURE In order to realze relable navgaton of the humanod robot n an uncertan envronment, we developed a layered archtecture for the humanod navgaton software, as shown n Fg. 1. Above the navgaton software layers, there exsts a msson management layer, whch manly determnes the current acton unt based on the status of the gven task [23]. Therefore, the navgaton software layer s only actvated when the current acton requres the robot to relocate to another poston. It takes charge of the entre autonomous navgaton process based on bpedal locomoton to start from any locaton wthn a gven Fg. 1. Navgaton software archtecture. map to reach the destnaton gven from the hgher level msson management layer. There exst three layers wthn the proposed navgaton software. The uppermost layer s the walkng path planner layer n whch a safe path s generated based on the gven map of the envronment and obstacle nformaton. Startng from the robot s current poston, the output path s descrbed by multple lnebased paths accordng to the global coordnate frame. Gven a safe path leadng to the destnaton of the humanod robot, the walkng footstep planner contnuously determnes the footstep locaton wth respect to the robot s body fxed coordnates. The desgned path s tracked by compensatng the error from the path to the robot s current poston estmates. After the footstep locaton s determned, fnally, a walkng pattern s generated accordng to the stablzng CoM/ZMP trajectores. Fnally, the generated footstep locaton and the correspondng CoM/ZMP trajectores are delvered to a real-tme low-level control layer, whch runs on a separate embedded computer runnng real-tme operatng system to manage jont angle control processes as well as communcaton between processes. Snce ths moton control process s very mportant to guarantee relable robot moton, most of the hgh-level processes requrng heavy computaton such as decson makng or machne vson are mplemented on a dfferent computer, although they communcate through a TCP/IP-based Ethernet connecton. In Fg. 2, the humanod navgaton system s dsplayed to vsualze the layout of the developed hardware components ncludng computers and sensor modules. In addton to the three man layers mentoned prevously, the humanod navgaton software ncludes software modules such as the management of nterfaces to the envronment database and localzaton module. Event management related to navgaton such as path replannng trggered by the detecton of unknown obstacles s carred out as well. A detaled explanaton of the three man layers wll be presented n the followng sectons.

3 1052 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 59, NO. 2, FEBRUARY 2012 Fg. 3. Lne-based path for a humanod robot. Fg. 2. Humanod robot navgaton system layout. Although the proposed software archtecture may be smlar to that of conventonal moble robot control software, each layer s desgned to take nto consderaton unque characterstcs of the humanod robot, such as foot collson checks and balanced walkng. The proposed archtecture also takes nto account the executon of the developed system on dverse workng envronments wthout sgnfcant modfcaton of the surroundngs. The only modfcaton that we have assumed throughout ths paper s the use of prnted markers on the celng, whch can easly be attached. The mappng of the marker locaton can be processed onlne. However, ts measurement ncludes errors of around 5 cm dependng on the relatve poston from the marker, whch needs to be fltered approprately to produce better estmates of the robot s pose. The localzaton module, ncludng ths flterng process, wll also be explaned n detal n Secton VI. III. WALKING PATH PLANNING In ths secton, a global path planner for humanod locomoton s descrbed. There are varous approaches for humanod path plannng such as [19] and [24] [26]. As already explaned n Secton I, the path-plannng problem of a humanod s dfferent from that of a conventonal moble robot because of the addtonal constrants and flexblty orgnatng from bpedal moton. In ths paper, for the sake of ncludng constrants on humanod walkng such as foot collson avodance and the ablty of lateral steppng, a lne-based path s desgned, whch yelds a path composed of straght lnes and curves, as shown by the dotted lne n Fg. 3. In ths knd of path, the robot tracks the lne and stops at a turnng spot, where the turnng radus s ncluded n the path desgn to avod collson between feet. As a basc tool for path plannng, we used the VG whch yelds the shortest polygonal path. The fundamental assumpton of the VG s that a robot can change ts orentaton, whle ts poston s fxed. However, ths assumpton does not hold for a humanod robot whch needs to change ts poston to conduct a turnng moton as n Fg. 3. To use the VG for humanod robots, our path planner augments the tradtonal confguraton space by area A marked n Fg. 3. Then, a modfed VG Fg. 4. Shortest dstance path has multple turnng spots whch can be reduced by the constructon of an extended VG, as shown n the dotted lne. s generated based on ths augmented confguraton space on whch we can fnd a polygonal path adequate for humanod robots. Furthermore, we perform postprocessng to reduce the trackng tme. In the generated path, there are multple turnng spots (ponts A C n Fg. 4), whch force the robot to decelerate and accelerate, whch, n return, ncreases the trackng tme. Thus, for a humanod robot, t s more natural to desgn a tmeoptmal path rather than the shortest path. However, fndng a tme-optmal path for a humanod robot s an NP-hard problem because there could be an nfnte number of combnatons. Thus, our path planner ams to fnd a tme-reducng path n two steps. The frst step s to reduce the number of turnng spots. For that purpose, we extended each lne segment to add addtonal edges and nodes and constructed an extended graph, as shown by the dotted lnes n Fg. 4. The second step s to add temporal nformaton that dfferent steps (forward, backward, and sde-walk) requre and to run a search algorthm based on the tme ndex. Fg. 5 shows the paths generated by the modfed VG and the proposed path planner. Note that the modfed VG s the same as the tradtonal VG except for the sze of the confguraton space. Here, we can verfy two ponts. One s that a lne-based path that s adequate for the humanod robot can be generated by the modfed VG. The other s that the proposed path planner yelds a path whose trackng tme s less than that of the modfed VG. However, t should be noted that the algorthmc complexty ncreases by the constructon of the extended VG. IV. WALKING FOOTSTEP PLANNER FOR PATH TRACKING Havng the walkng center of the humanod robot calculated by the localzaton module, the footstep-plannng layer determnes the desred poston and angle of the feet to make the humanod robot track the desred path. Snce a footstep s a

4 KANG et al.: DEPENDABLE HUMANOID NAVIGATION SYSTEM BASED ON BIPEDAL LOCOMOTION 1053 Fg. 7. Overall block dagram for walkng pattern generaton. Fg. 5. Paths generated by the (a) modfed VG and (b) proposed path planner. The error of the headng drecton ψ e s gven by defnton ψ e = ψ d ψ(k) (4) where ψ d represents the drecton of the lne-based path defned by (2). As the lateral and headng errors are defned, the control nputs d(k) and r(k) are computed as outputs from the footstep planner layer. At frst, the footstep dstance d(k) s computed by d(k) = mn[d max,d goal ] (5) where d max represents the maxmum footstep dstance allowed for the humanod and d goal represents the dstance to the goal poston. Because of the delay n estmatng the locaton of the robot, the footstep dstance s calculated n advance by consderng the delay and s stored n a buffer before actual executon. In order to regulate the lateral error δ e and headng error ψ e, r(k) s determned by r(k) =K δ δ e + K ψ sn(ψ e ) (6) Fg. 6. World coordnate and error defnton. dscrete measure, the walkng center of the humanod evolves accordng to the followng dscrete equatons: x(k +1)=x(k)+d(k)cos(ψ(k)) + w x (k) y(k +1)=y(k)+d(k)sn(ψ(k)) + w y (k) ψ(k +1)=ψ(k)+r(k)+w ψ (k). (1) In the aforementoned equaton, {x(k),y(k),ψ(k)}, respectvely, represent the X and Y coordnates of the humanod poston and ts headng at the kth footstep. Furthermore, d(k) and r(k) represent the footstep dstance and rotaton angle, respectvely. These two values are control varables that are used to make the robot follow the desred path. Also, w x (k), w y (k), and w ψ (k) are external dsturbances caused by slppage on each foot. Fg. 6 shows the coordnate system and error defntons n the path trackng control algorthm. Supposng that the desred path of the humanod s gven as a lne descrbed by ax + by + c =0 (2) then, the lateral error δ e from the desred trajectory to the humanod robot s determned by δ e = ax(k)+by(k)+c. (3) a2 + b 2 where K δ and K ψ are the control parameters to be adjusted to take nto account the tme delay of the footstep control commands. The output footstep dstance d(k) and the rotaton angle r(k) are gven to the walkng pattern generaton layer to produce the correspondng CoM/ZMP moton. V. WALKING PATTERN GENERATION PROCEDURE In ths secton, we shall explan how the walkng pattern of the foot, ZMP, and CoM can be generated from external nformaton on the walkng path. The output from the footstep planner layer s nspected wth the ad of two procedures before a walkng pattern s generated. These two procedures should be consdered because of the knematc constrants of the robot tself. We use an ellptcal boundary adjustment and a collson detecton method as the two procedures. The walkng pattern for a humanod robot can be generated wth the walkng pattern generaton method advanced pole-zero cancelaton by seres approxmaton (APZCSA) usng the nspected walkng path [13]. Fg. 7 shows the overall block dagram for the walkng pattern. We shall explan each procedure of Fg. 7 n detal n the followng sectons. A. Ellptcal Boundary Adjustment The local footstep commands from the walkng footstep planner consst of walkng values of the sagttal and frontal planes, and the turnng angle wth whch the humanod robot should walk wth reference to the current local frame. The

5 1054 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 59, NO. 2, FEBRUARY 2012 Fg. 8. Walkng path and foot postons. Fg. 10. Checkng the collson and modfyng the swng foot poston. (a) Orgnal swngng foot poston. (b) Checkng the collson between feet. (c) Modfyng the swngng foot poston. The center of the ellpse s the center of the swngng foot poston when the robot s standstll. The ellpse s represented by x 2 lsslmax 2 + y 2 lfslmax 2 =1 (7) Fg. 9. Adjustng the swngng foot appled to the ellptcal area n the case where the swng foot s the left foot. swngng foot poston s defned by these walkng values. It s located on the vertcal lne of the walkng path and s selected by consderng the currently supported foot, as shown n Fg. 8. The swngng foot poston, however, has to be checked aganst the constrants regardng the knematc lmts and the possblty of collson between feet before t can be used n generatng the walkng pattern. We shall, therefore, deal wth knematc lmts n ths secton. For the case n whch the robot needs excessve dstance for one step, the foot poston could be outsde the workspace. To avod ths case, the swngng foot postons are constraned nsde a specfc area, whch s taken as ellptcal n shape for ths paper. The ellptcal area n Fg. 9 s consttuted wth the maxmum sagttal step length (l SSLmax ) and the maxmum frontal step length (l FSLmax ), where l SSLmax and l FSLmax mean the maxmum step length for one step at the sagttal and frontal planes, respectvely. Ths ellptcal area represents the area whch can be reached by the swngng foot despte the knematcal constrants of the robot. Fg. 9 shows the foot adjustment for foot postons outsde the ellptcal area. In Fg. 9, the superscrpt or descrbes the orgnal swngng foot poston and the superscrpt mod represents the modfed swngng foot poston that s moved to the boundary of the ellptcal area. where x and y denote the poston of the swngng foot wth respect to the center of the ellpse. If the swngng foot poston s nsde the ellptcal area, the foot poston can be utlzed for the walkng pattern. However, f the swngng foot poston s outsde the ellptcal area, the swngng foot poston should be modfed by placng t on the boundary of the ellpse. The modfed poston s gven by the pont of ntersecton n Fg. 9 on the boundary of the ellpse. The modfed swngng foot postons are represented as shown n the followng: x mod = y mod = l SSLmax l FSLmax x or (lfslmax x or ) 2 + ( lsslmax y or l SSLmax l FSLmax y or (lfslmax x or ) 2 + ( lsslmax y or ) 2 (8) ) 2. (9) These modfed swngng foot postons can satsfy the robot s knematc lmts. B. Foot Collson Detecton Besdes the knematc lmts, a robot should also consder the possblty of collson between both feet n the context of walkng pattern generaton. The four corners of each foot are checked, as shown n Fg. 10, wth reference to the other foot frame to detect the collson of the feet n a smple manner. The corners of each foot are shown by red crcles n Fg. 10(b). If a collson s detected, the swngng foot poston should be shfted by some dstance from the supportng foot, as shown n Fg. 10(c). The shftng dstance s selected as a dstance that s suffcent enough to avod foot collson.

6 KANG et al.: DEPENDABLE HUMANOID NAVIGATION SYSTEM BASED ON BIPEDAL LOCOMOTION 1055 Fg. 11. Block dagram for generatng the walkng pattern wth APZCSA. C. Generatng the Walkng Pattern In generatng the walkng pattern, we use the cart-table model as a smplfed humanod model for the purpose of smple calculaton [12]. The cart-table model s consttuted of the ZMP and CoM whch, respectvely, represent the stablty and moton of the humanod robot. If the cart-table model s analyzed n the frequency doman, t can be establshed that ths model s margnally stable and has the nonmnmum phase property [13]. These characterstcs make the generaton of a walkng pattern more dffcult. Therefore, ths paper makes use of the APZCSA n overcomng these characterstcs [13]. The APZCSA conssts of feedforward and feedback controllers, as shown n Fg. 11. The feedback controller makes the cart-table model more robust by movng the poles to more stable regons. The pole placement method s used for the feedback controller. The feedforward controller reduces the nonmnmum phase property and mproves the trackng performance. The nonmnmum phase occurs due to the unstable zero, whch s located out of the unt crcle n the Z doman. We utlze the polezero cancellaton by a seres approxmaton. Ths controller can approxmate the nverse of an unstable zero by usng the seres approxmaton. However, the feedforward controller wth a seres approxmaton has a steady-state error. Ths steadystate error can be reduced by addng 1/(zu Norder 1),asshown n Fg. 11. Stable patterns of the ZMP and CoM are generated by these two controllers. Fnally, the swngng foot trajectory s generated by the snusodal functon usng the swngng foot poston and prevous foot poston. Fg. 12. Snapshots of the humanod walkng taken from the celng. Fg. 13. (a) Stargazer localzaton sensor module wth a marker. (b) Sensor module mounted on the robot. VI. INDOOR LOCALIZATION Localzaton of the humanod robot s dffcult for several reasons. Frst, humanod footsteps suffer contnuous slppage, whch results n large error accumulaton for the dead-reckonng-based localzaton method. The use of an external localzaton sensor s preferred for ths reason. However, constant ptchng and rollng moton durng walkng make t dffcult to accurately measure the poston of the humanod. In addton, there exsts repeated lateral oscllaton of the robot due to the snusodal CoM/ZMP profle, as shown n Fg. 12. The red crcles represent consecutve poston of the humanod robot s center, as seen from the celng. Although the robot walks straght n the forward drecton, the center poston alternates between left and rght, whch makes t dffcult to judge whether the robot s accurately followng the desred path. Even though the oscllatory CoM profle s known n advance, t s stll dffcult to subtract the oscllatory component Fg. 14. Effect of the movng-average flter on the sensor measurement. from the robot s current poston and to calculate the walkng center of the robot because of the tme delay and nose n the localzaton sensor measurement. Fg. 13 shows the stargazer sensor hardware module mplemented on the robot s head. The nfrared camera on the sensor module takes the mage of the celng, where markers are attached, to localze the robot s current poston. Each marker s prnted wth a unque pattern of materal that reflects the nfrared lght emtted from the LEDs surroundng the camera. Therefore, the sensor s less senstve to ambent lght condtons. However, due to the mage

7 1056 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 59, NO. 2, FEBRUARY 2012 Fg. 15. (a) Humanod trackng the y = x lne. (b) Humanod trackng the y = x lne. Fg. 17. Robot trajectory durng the straght lne followng the experment usng the proposed path trackng controller. actual moton s affected by the whole-body balancng control effort aganst external dsturbances. Although the pattern of the component s unknown, the perod of oscllaton s determned by the walkng footstep planner, and t s less affected by the balancng controller. The elmnaton of perodc nfluences wth known perods s a characterstc of the movng-average flter. Therefore, t s decded to apply a movng-average flter for the stargazer sensor measurement, whch s defned by z(k) ={z x (k),z y (k),z ψ (k)}. (10) Then, the flter estmates are calculated by the cumulatve sum of the measurements dvded by the number of data Fg. 16. (a) Robot at the ntal poston and the path-plannng result. (b) Robot s moton toward the frst waypont. (c) Robot at the frst waypont. (d) Robot s moton toward the second waypont. (e) Robot at the second waypont. (f) Robot at the fnal destnaton. processng tme of the on-board dgtal sgnal processng unt, the sensor update rate s lmted to around 5 10 Hz, whch may vary dependng on the mage qualty and buffer condton for packets n seral communcaton. The shape of the oscllatory component s not an exact snusodal functon. Although ts shape s decded by the current CoM profle generated from the pattern generaton layer, the z MA (k) = 1 =p z(k ) (11) p =0 where p s the sze of the movng-average flter determned by the number of sensor measurements between each footstep. Fg. 14 shows the effect of the movng-average flter n terms of elmnatng the oscllatory component, whle the humanod robot walks for 12 steps. The thck dotted lne shows the deadreckonng-based estmaton of the robot s head poston, the sold lne wth stars represents the stargazer measurement, and fnally, the red sold lne represents the movng-average flter output. Although the robot s walkng center poston shfts to the rght for about 4 cm based on the stargazer measurement, the dead-reckonng-based estmaton does not nclude ths drft. The fltered output shows no oscllatory behavor durng the forward walkng moton, and the robot s walkng center s well represented by ths output. However, poston estmaton s delayed for one perod of walkng as an effect of the flter. Ths delay s taken nto account durng the footstep-plannng stage.

8 KANG et al.: DEPENDABLE HUMANOID NAVIGATION SYSTEM BASED ON BIPEDAL LOCOMOTION 1057 Fg. 18. (a) (f) Snapshots of the trajectory of an expermental humanod robot that follows a desgned path. (g) (l) Snapshots of the expermental humanod robot that follows a desgned path. VII. SIMULATION RESULT A. Walkng Footstep Planner The smulaton envronment s provded from the developed real-tme control software based on a humanod knematc model. Snce the hgh-level control nterface s the same for both the expermental mplementaton and the smulaton envronment, the development process s greatly smplfed, and a smlar performance s guaranteed. However, the smulaton envronment stll does not nclude communcaton delays and hardware uncertantes such as foot slppage and sensor nose, whch have to be generated artfcally durng smulaton. In the frst smulaton, t s assumed that the walkng path planner layer outputs the desred lne as the y = x lne, and the performance of the footstep planner as explaned n Secton IV s evaluated. In Fg. 15(a), the trajectory of the humanod robot trackng the y = x lne s plotted wth a green lne. The fnal poston of the humanod robot s plotted wth a red box, and the ntal pose of the humanod robot s at the orgn facng rght. As the desred trajectory s set as the y = x lne, the robot gradually changes ts headng toward the desred drecton and quckly converges to the desred lne. The trajectory shows snusodal oscllaton around the center of the walkng path because the smulaton ncludes the whole-body moton of the humanod robot, ncludng the CoM movements. In Fg. 15(b), the trajectory s the y = x lne, and as the controller detects the headng error, the footstep planner generates the desred footsteps that converge to the desred lne smlar to the prevous result. B. Integrated Navgaton Software In ths secton, the entre humanod navgaton software s ntegrated, ncludng three layers and nterfaces wth low-level control layers. It s assumed that the humanod robot s gven a map of the obstacles, and the fnal destnaton s gven from the upper-level msson management software layer. The sze of the map s 5 6 m. The localzaton sensor measurement s replaced wth the smulated sensor measurements based on dead reckonng. In Fg. 16, the sequence of plots represents the snapshots of a smulaton durng whch the robot generates a safe path to the desred goal poston and follows the path whle avodng the vrtual obstacles n the map. The black lnes represent the layout of the obstacles n the map, and the blue crcles and lnes represent the safety space from the obstacle. The red dots represent the smulated measurements of the laser scanner, whch s assumed to be nstalled on the robot. In Fg. 16(a), the robot s set at ts ntal poston, and the VG-based walkng path planner generates a safe path to the goal poston. Fg. 16(b)

9 1058 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 59, NO. 2, FEBRUARY 2012 Fg. 19. (a) (k) Snapshots of the expermental results wth path replannng. (l) (v) Snapshots of the vdeo recorded durng the path-replannng experment.

10 KANG et al.: DEPENDABLE HUMANOID NAVIGATION SYSTEM BASED ON BIPEDAL LOCOMOTION 1059 shows that the robot starts movng toward the frst waypont marked by a green box where the frst lne trajectory ends. The robot takes a rght turn maneuver to head toward the second waypont shown n Fg. 16(c). Fg. 16(d) and (e) shows a smlar procedure of the robot movng toward the next waypont and reachng the goal poston n Fg. 16(f). Snce ths smulaton result does not nclude any error n localzaton or foot slppage, the robot does not feature any error-correctng behavor. VIII. EXPERIMENTAL RESULT In ths secton, the expermental results of the overall navgaton system mplemented on a humanod robot hardware platform are presented. Fg. 17 shows the trajectory of the robot that follows a straght dotted lne. The red thn lne represents the stargazer sensor measurements, and the black thck lne represents the movng-average estmates of the robot pose. Intally, the robot pose has some error from the desred path. The trajectory startng from a poston around (x, y) =(1.5, 1.3) slowly converged to the desred path due to the effect of the trackng controller mplemented n the walkng footstep planner layer. Fg. 18(a) (f) shows snapshots of the experment, n whch the stargazer measurements are plotted wth a green lne and the current poston of the expermental humanod robot s plotted wth red boxes. The red dots are the smulated measurements from the laser scanner because the scanner s turned off for ths experment. The experments n a 5 6 map yelded smlar results to the smulaton results n Secton VII. The robot desgned a safe path to the goal poston defned wth respect to the global coordnate frame of the stargazer sensor measurement. Fg. 18(g) (l) shows photographs taken durng the experments. On the floor, t s shown that the future desred trajectory of the humanod s marked wth black tapes so that the trackng performance can be verfed. The robot s performance n terms of accurately followng the desgned path s verfed by both localzaton sensor measurements and marks on the floor. Fg. 19(a) (k) shows the second set of expermental results wth an unknown obstacle. The laser scanner s turned on, and ts measurement data are dsplayed wth red and blue dots, where red dots represent an nvald return from the floor or empty space, whle blue dots represent actual measurements from the obstacles. In Fg. 19(a), the robot has desgned an ntal path based on the obstacle confguraton n the map. As the robot approaches the frst waypont, the unknown obstacle s detected, and ts nformaton s updated on the map, as shown n Fg. 19(c). The robot replans the path by avodng a new popup obstacle and follows the replanned path. The confguraton of the second experment s shown n Fg. 19(l). A table covered wth a cloth s the unknown obstacle that dd not exst n the ntal map. As the robot approaches the unknown obstacle, the obstacle s detected n Fg. 19(n). In Fg. 19(o) (v), snapshots of the robot followng the replanned path are shown. From the expermental results, the overall performance of the developed navgaton system s verfed. The trackng accuracy consttutes an error of approxmately cm, whch depends on the localzaton sensor accuracy as well as dsturbances such as foot slppage. IX. CONCLUSION In ths paper, we have proposed a dependable navgaton software archtecture that features a localzaton sensor module wth reasonable accuracy and a relable humanod hardware platform that executes several stable untary walkng motons, such as walkng forward and backward, left and rght, and turnng. Although these untary walkng motons can be exploted to desgn a footstep plan that can lead to the desgned goal poston, uncertantes such as sensor nose and foot slppage have to be overcome. For example, no matter how accurately the robot s controlled, t s dffcult for the humanod robot wth such a large degree of freedom to mantan a stable straght path for more than 10 m wth errors of less than 5 cm. Therefore, the localzaton nformaton s provded from an external sensor. The flexblty and robustness of the navgaton software archtecture s ensured to ncorporate error compensaton relatve to the desred path durng walkng. In partcular, the proposed navgaton archtecture employs a trackng control algorthm between the walkng path planner and the actual command of the footsteps to nclude addtonal degrees of freedom to regulate error caused by uncertantes and dsturbances. Its performance s verfed through expermental results on a human-szed humanod robot, as well as realstc smulatons. The proposed dependable navgaton system for humanod robots wth hgh relablty enables humanod robots to provde home servces. By combnng the navgaton system wth vsual servong, graspng, and delvery of objects, a fnal ntegrated humanod robot system wll meet the requrements n provdng meanngful home servces such as garbage collecton and delvery. REFERENCES [1] S. Km, C. Km, B. You, and S. Oh, Stable whole-body moton generaton for humanod robots to mtate human motons, n Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2009, pp [2] J. Lee, M. Je, S. Km, M. Jung, C. Km, and B. You, Balloon burster: A CORBA-based vsual servong for humanod robot n a dstrbuted envronment, n Proc. Annu. Conf. SICE, 2007, pp [3] S. Km, S. Hong, and D. Km, A walkng moton mtaton framework of a humanod robot by human walkng recognton from IMU moton data, n Proc. IEEE-RAS Int. Conf. Humanod Robots, 2009, pp [4] K. C. Tan, Y. J. Chen, K. K. Tan, and T. H. Lee, Task-orented developmental learnng for humanod robots, IEEE Trans. Ind. Electron., vol.52, no. 3, pp , Jun [5] C. Fu and K. Chen, Gan synthess and sensory control of star clmbng for a humanod robot, IEEE Trans. Ind. Electron., vol. 55, no. 5, pp , May [6] T. Sato, S. Sakano, E. Ohash, and K. Ohnsh, Walkng trajectory plannng on stars usng vrtual slope for bped robots, IEEE Trans. Ind. Electron., vol. 58, no. 4, pp , Apr [7] Y. Cho, D. Km, Y. Oh, and B. J. You, Posture/walkng control for humanod robot based on the knematc resoluton of CoM Jacoban wth embedded moton, IEEE Trans. Robot., vol. 23, no. 6, pp , Dec [8] J. Yamaguch, E. Soga, S. Inoue, and A. Takansh, Development of a bpedal humanod robot-control method of whole body cooperatve dynamc bped walkng, n Proc. IEEE Int. Conf. Robot. Autom., 1999, pp [9] W. Yang, N. Y. Chong, C. Km, and B. J. You, Self-adaptng humanod locomoton usng a neural oscllator network, n Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Oct. 2007, pp [10] K. Harada, S. Kajta, K. Kaneko, and H. Hrukawa, An analytcal method on real-tme gat plannng for a humanod robot, n Proc. IEEE/RAS Int. Conf. Humanod Robots, 2004, pp

11 1060 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 59, NO. 2, FEBRUARY 2012 [11] Y. Oh, K.-H. Ahn, D. Km, and C. Km, An analytcal method to generate walkng pattern of humanod robot, n Proc. IEEE Annu. Conf. Ind. Electron., Nov. 2006, pp [12] S. Kajta, F. Kanehro, K. Kaneko, K. Fujwara, K. Harada, K. Yoko, and H. Hrukawa, Bped walkng pattern generaton by usng prevew control of zero-moment pont, n Proc. IEEE Int. Conf. Robot. Autom., 2003, pp [13] S. Hong, Y. Oh, D. Km, and B. J. You, A walkng pattern generaton method wth feedback and feedforward control for humanod robots, n Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., St. Lous, MO, Oct. 2009, pp [14] N. Moto, T. Suzuk, and K. Ohnsh, A bpedal locomoton plannng based on vrtual lnear nverted pendulum mode, IEEE Trans. Ind. Electron., vol. 56, no. 1, pp , Jan [15] K. Erbatur and O. Kurt, Natural ZMP trajectores for bped robot reference generaton, IEEE Trans. Ind. Electron., vol. 56, no. 3, pp , Mar [16] T. Sato, S. Sakano, and K. Ohnsh, Real-tme walkng trajectory generaton method wth three-mass models at constant body heght for threedmensonal bped robots, IEEE Trans. Ind. Electron., vol. 58, no. 2, pp , Feb [17] J.-S. Gutmann, M. Fukuch, and M. Fujta, Real-tme path plannng for humanod robot navgaton, n Proc. Int. Jont Conf. Artf. Intell., 2005, pp [18] J. Chestnutt, J. Kuffner, K. Nshwak, and S. Kagam, Plannng bped navgaton strateges n complex envronments, n Proc. IEEE Int. Conf. Humanod Robots, 2003, pp [19] K. Sabe, M. Fukuch, J.-S. Gutmann, T. Ohash, K. Kawamoto, and Y. Yoshgahara, Obstacle avodance and path plannng for humanod robots usng stereo vson, n Proc. IEEE Int. Conf. Robot. Autom.,2004, pp [20] P. Mchel, J. Chestnutt, S. Kagam, K. Nshwak, J. Kuffner, and T. Kanade, GPU-accelerated real-tme 3D trackng for humanod locomoton and star clmbng, n Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., San Dego, CA, 2007, pp [21] S. Thompson, S. Kagam, and K. Nshwak, Localsaton for autonomous humanod navgaton, n Proc. IEEE-RAS Int. Conf. Humanod Robots, Dec. 2006, pp [22] R. Tellez, F. Ferro, D. Mora, D. Pnyol, and D. Facont, Autonomous humanod navgaton usng laser and odometry data, n Proc. IEEE-RAS Int. Conf. Humanod Robots, Daejeon, Korea, Dec. 2008, pp [23] H. Lm, Y. Kang, J. Lee, J. Km, and B. J. You, Software archtecture and task defnton of a multple humanod cooperatve control system, Int. J. Humanod Robot., vol. 6, no. 2, pp , Jun [24] J. Kuffner, K. Nshwak, S. Kagam, Y. Kunyosh, M. Inaba, and H. Inoue, Onlne footstep plannng for humanod robots, n Proc. IEEE Int. Conf. Robot. Autom., 2003, pp [25] J. Chestnutt, M. Lau, G. Cheung, J. Kuffner, J. K. Hodgns, and T. Kanade, Footstep plannng for the Honda ASIMO humanod, n Proc. IEEE Int. Conf. Robot. Autom., 2005, pp [26] P. Mchel, J. Chestnutt, J. Kuffner, and T. Kanade, Vson-guded humanod footstep plannng for dynamc envronments, n Proc. IEEE/RAS Int. Conf. Humanod Robots, 2005, pp Hyunsoo Km receved the B.S. and M.S. degrees n nano/it engneerng from Seoul Natonal Unversty of Technology, Seoul, Korea, n 2007 and 2009, respectvely. He worked as a Research Scentst wth the Center for Cogntve Robotcs Research, Korea Insttute of Scence and Technology, Seoul, n Soo-Hyun Ryu receved the B.S. and M.S. degrees n electrcal engneerng from Korea Unversty, Seoul, Korea, n 2008 and 2011, respectvely, where he s currently workng toward the Ph.D. degree n electrcal engneerng. Hs research nterests are underwater SLAM, humanod path plannng, etc. Nakju Lett Doh (A 03) receved the B.S., M.S., and Ph.D. degrees n mechancal engneerng from the Pohang Unversty of Scence and Technology, Pohang, Korea, n 1998, 2000, and 2005, respectvely. He s currently an Assocate Professor wth the School of Electrcal Engneerng, Korea Unversty, Seoul, Korea. Hs research nterests are the 3-D SLAM, structure from moton, robot path plannng, etc. YongHwan Oh receved the B.S., M.S., and Ph.D. degrees n mechancal engneerng from Pohang Unversty of Scence and Technology, Pohang, Korea, n 1991, 1993, and 1999, respectvely. He s a Prncpal Research Scentst wth the Interacton and Robotcs Research Center, Korea Insttute of Scence and Technology, Seoul, Korea. In 1999 and 2000, he was a JSPS Postdoctoral Fellow wth the Prof. A. Takansh Laboratory, Department of Mechancal Engneerng, Waseda Unversty, Shnjuku, Japan. Hs research nterests nclude mechancal desgn of humanod robot, sensor-based bpedal walkng, robust moton control of mechancal system, and moton/nteracton control of redundant manpulators. Yeonsk Kang receved the Ph.D. degree n mechancal engneerng from the Unversty of Calforna (UC), Berkeley, n He was a Postdoctoral Researcher wth the Center for Collaboratve Control of Unmanned Vehcles, UC Berkeley, n He worked as a Senor Researcher Scentst wth the Center for Cogntve Robotcs Research, Korea Insttute of Scence and Technology, Seoul, Korea, n Snce 2010, he has been a Professor wth the Department of Automotve Engneerng, Kookmn Unversty, Seoul. Hs research nterests nclude navgaton, nonlnear control, and target trackng. Bum-jae You (S 88 M 88) receved the B.S. degree n control and nstrumentaton engneerng from Seoul Natonal Unversty, Seoul, Korea, n 1985 and the M.S. and Ph.D. degrees n electrcal and electronc engneerng from the Korea Advanced Insttute of Scence and Technology, Daejon, Korea, n 1987 and 1991, respectvely. He was the Head of the Robotcs Dvson, TurboTek Co., Ltd., Korea, from 1991 to 1994, conductng the development of a hgh-speed mage processor and moton controller for robots. He joned the Korea Insttute of Scence and Technology, Seoul, n 1994, where he s the Center Drector of the Center for Cogntve Robotcs Research. He was a Vstng Fellow wth the Department of Computer Scence, Yale Unversty, New Haven, CT, for three months n Hs research nterests nclude humanods, vson-based robotcs, vson-based control, real-tme computer vson, ntellgent servce robots, and dgtal sgnal processor applcatons.