A Robot Simulator Based on the Cross Architecture for the Development of Cognitive Robotics

Transcription

1 2016 XIII Lain American Roboics Symposium and IV Brazilian Roboics Symposium A Robo Simulaor Based on he Cross Archiecure for he Developmen of Cogniive Roboics Danilo H. Perico, Thiago P. D. Homem, Aislan C. Almeida, Isaac J. Silva, Claudio O. Vilão Jr., Vinicius N. Ferreira, Reinaldo A. C. Bianchi Elecrical Engineering Deparmen Cenro Universiário FEI São Bernardo do Campo, São Paulo, Brazil {dperico, rbianchi}@fei.edu.br Deparmen of Informaics Insiuo Federal de São Paulo São Paulo, São Paulo, Brazil Absrac This paper presens a new 2D robo simulaor based on he Cross Archiecure for RoboCup Soccer Humanoid League domain. A simulaor is an imporan ool for esing cogniive algorihms in robos wihou he need of handling wih real robo problems; moreover, a simulaor is eremely useful for allowing reproducibiliy of any developed algorihm, even if here is no robo available. The proposed simulaor allows he direc applicaion of he algorihms developed in he simulaor ino a real robo ha works wih he Cross Archiecure. Besides, he simulaor is freely available, i is open-source and has low compuaional cos. Eperimens were conduced in order o analyze he porabiliy of a decision code developed in he simulaor o a real robo. The resuls allowed us o conclude ha he simulaor can be used o es new algorihms, since he decisions performed by he robo in simulaion and in he real world were quie similar. I. INTRODUCTION The Cross Archiecure [1] is a sofware archiecure composed by independen modules, ha was proposed in order o allow a humanoid robo, equipped wih a compuer, o perform several concurren asks wihou he need of microconroller boards and wihou performance loss. Several modules of he Cross Archiecure are dedicaed o he developmen of he Arificial Inelligence algorihms, for insance, robo s decision, localizaion and planning. Neverheless, one of he main problems of developing cogniive algorihms in real humanoid robos are he physical issues, such as backlashes, broken pars, bad connecions and so on. Anoher problem of developing cogniive algorihms in unique humanoid robos, i.e. hose ones ha are no sandard, is he lack of reproducibiliy of any proposed code. Reproducibiliy is very imporan in order o allow he spread of knowledge. Thereby, a simulaor is an imporan ool for debugging and esing high-level algorihms wihou he need of dealing wih real robo problems. Moreover, i allows he reproducibiliy of any proposed algorihm even if here is no robo available. Hence, he goal of his work is he proposal of a 2D humanoid soccer simulaor ha allows he absracion of real world problems during he developmen of high-level algorihms ha are based on he Cross Archiecure. Thus, he simulaor s domain is he RoboCup Soccer Humanoid League. There are already several kinds of robo simulaors available, such as hose depiced in Secion III. However, he cogniive algorihms developed in he proposed simulaor can be direcly applied in he humanoid robos ha use he Cross Archiecure as heir sofware archiecure. This work is srucured as follows: Secion II briefly describes he Cross Archiecure. Secion III depics some usual robo simulaors. Secion IV shows he proposed simulaor. In Secion V a case sudy is presened and Secion VI provides he conclusions and indicaes avenues for eend his work. II. CROSS ARCHITECTURE Sofware archiecure is essenial o he developmen of successful applicaions in roboics, specially considering he comple archiecure requesed by humanoid robos. Sofware archiecure can be described as a guidance for sofware design and developmen, acing as a high-level modeling, i represens he fundamenal conceps, properies or resricions of a sysem [2]. Following he sofware archiecure conceps, he Cross Archiecure was developed wih focus in humanoid robo soccer players. Mos common paradigms for sofware archiecure used in roboics are [3]: hierarchical, reacive [4] and hybrid [5]. The hierarchical paradigm is based in he sequenial and orderly relaion beween he hree primiives of roboics: Sense Plan Ac. On he oher hand, he reacive paradigm uses a verical decomposiion, where he robo has a Sense Ac ype of organizaion wih some primiive behaviors [3]. Due o he srong accepance of reacive and hierarchical paradigms in roboics, a new paradigm, ha combined hese las wo, was proposed. This new paradigm was named hybrid and i aims he creaion of a sysem capable o plan as long as i senses and reacs [6]. The hybrid paradigm argues ha a robo has o deliberae while i senses and acs, and i is done by decomposing a ask ino sub-asks. Tasks ha are reacive, such as proecing a robo from a falling or adaping he walking o sloped errains [7], do no need o be planned or deliberaed. When one of hose siuaions occurs, he robo should simply sense and /16 $ IEEE DOI /LARS-SBR

2 reac, leaving comple behaviors, like deciding wha o do or localizing iself, o he hierarchical srucure. The Cross Archiecure is based on he hybrid paradigm. Due o he blackboard concep [8], ha is a shared memory area ha allows iner-process communicaion, he Cross Archiecure allows an absracion layer beween he processes required for a humanoid robo soccer player. Addiionally, his archiecure also allows a humanoid robo o work properly wihou any dedicaed sub-conroller implemened in lowlevel hardware (e.g., a microconroller board is no needed for conrolling he moors). The Cross Archiecure saes ha independen processes can communicae beween each oher in order o achieve an inelligen sysem. I is composed by 8 processes: Vision, Localizaion, Decision, Communicaion, Planning, Inerial Measuremen Uni (IMU), Conrol and Managemen [1]. Figure 1 depics how he processes are inerconneced and how hey access he blackboard (bkb) o echange daa among hem, considering a robo n. For eample, he Decision process, in order o reason abou he ne acion, uses daa from Vision, Localizaion and Communicaion. Then, i may se he Vision o coninue searching for he ball or i may send o Planning a command o compue a differen rajecory, which mus send o Conrol a message o sar a movemen. There are some researches wih humanoid robos ha already use he Cross Archiecure as heir basic archiecure for he developmen of cogniive behaviors, for eample [7], [9] and [10]. III. SIMULATORS Robo simulaors are commonly used for he developmen of algorihms wihou he need of having an acual robo. The usage of simulaors allows he performance of a large number of eperimens, wihou he risk of damaging he robos. The mos used robo simulaors have some disadvanages. For eample, he robo s behavior can differ beween simulaion and real world. This happens because he simulaion Fig. 1. The Cross Archiecure. Adaped from [1]. Gray boes are he simulaed processes. model is commonly simplified and, hence, imperfec. So, he simulaion migh no reproduce he same resuls as he real world, even using a quie good physics engine [11]. Beyond ha, he 3D simulaors like Gazebo [12], Webos [13], V-REP [14] and RoboCup Soccer 3D [15] ofen have high compuaional cos and hey need o have he specific desired robo model, which is no always available by defaul. In he 2D simulaors, such as he RoboCup Soccer 2D [15] and TauraSim [16], he grea difficuly perceived was he inegraion of he Cross Archiecure in hese environmens and he direc porabiliy of he developed algorihms o real robos. The quoed simulaors are briefly described hereafer. A. Gazebo Gazebo is an open source roboic simulaion package, ha uses he open source OGRE rendering engine, which produces high-qualiy graphics [17]. Gazebo offers he abiliy o simulae robos in indoor and oudoor environmens. The environmen can be made using simulaed ligh and saic objecs. The physics engines i suppors are: ODE, Bulle, Simbody, and DART (ODE is defaul engine). Gazebo is available for Mac and Linu [18] and i is currenly being mainained by he company Open Source Roboics Foundaion (OSRF), which is he same company ha mainains he ROS. B. Webos Webos is a 3D simulaion environmen used o model, program and simulae mobile robos. Webos can runs naively on Windows, Mac and Linu, so he world files and API funcions can easily be used in differen operaing sysems. A simulaion can have many robos, limied only by compuer memory and CPU power, and each robo may run wih is own specific conroller code. The robo conrollers can be programmed in C, C++, Java, Pyhon and Malab, or use a ROS node. The robo behavior can be esed in physically realisic worlds by using he Open Dynamics Engine (ODE), physics engine library [18]. C. V-Rep V-REP (Virual Robo Eperimenaion Plaform) is available for Windows, Mac and Linu. The conrollers can be programmed in C/C++, Pyhon, Java, Lua, Malab, Ocave or Urbi. The physics engines are: Bulle Physics, ODE, Newon and Vore Dynamics. D. RoboCup Soccer 3D In Robocup, 3D soccer simulaion is performed in he RoboCup Simulaed Soccer Server 3D (RCSSServer3D) since 2008, and runs on Linu, Windows and Mac [19]. RC- SSServer3D is based on SimSpark. SimSpark uses he ODE library for simulaing rigid body dynamics. The robo agens in he simulaion are he Aldebaran NAO robo. E. RoboCup Soccer 2D The RoboCup Soccer Simulaor 2D is he official simulaor of RoboCup 2D simulaion league. In his sofware, each agen connecs o he server, simulaing a soccer game and an independen monior shows he saus of he game. 318

3 F. TauraSim TauraSim is a simulaor developed by Taura Bos eam for developing high-level sraegies in RoboCup Soccer Humanoid League. This simulaor works in wo modules: he firs one handles he 2D simulaion and he second one is an inerface for he robo s inelligence. The communicaion beween he modules is made via socke in JSON forma. IV. ROBOFEI-HT SIMULATOR RoboFEI-HT Simulaor (Figure 2) is inspired by TauraSim, since i is also a 2D simulaor made in Pyhon wih he suppor of he graphical library Pygame [20]. However, he proposed simulaor is buil upon he Cross Archiecure, using he blackboard concep for communicaion beween processes. The proposed simulaor aims o simulae hree processes from he Cross Archiecure: Vision, Conrol and IMU. Figure 1 depics he simulaed processes in gray boes. All he processes linked o he cogniive algorihms of he robo, such as Decision, Localizaion, Planning and Communicaion runs in boh plaforms: real or simulaed robos. As already menioned, his absracion is possible due o he blackboard. So, from he poin-of-view of cogniive processes, i does no maer wheher he robo is real or virual; decisions will be aken following he daa published in he blackboard. The proposed simulaor suppors muli-robo sysems, where he communicaion beween he robos is made via UDP, following he same procedure used by real robos. RoboFEI-HT Simulaor brings some conribuions in relaion o he oher simulaors, shown in Secion III. The firs one is he direc porabiliy of he high-level algorihms from he simulaion o a real robo ha uses he Cross Archiecure. Anoher advanage is he fac ha he proposed simulaor is free, open-source, and does no require a powerful compuer o work, which guaranee he reproducibiliy of any proposed cogniive algorihm for humanoid robos. Fig. 2. RoboFEI-HT Simulaor. A. Physics Simulaion The purpose of his simulaor is o simulae he physical environmen of he soccer compeiion in order o es he highlevel algorihms developed by he eam. The simulaed world has wo dimensions, allowing he robos o move forward, backward, o he lef, o he righ and o roae around is own ais, represening he movemens of real robos. The objecs movemens are simulaed using he concep of speed. For robos, i is applied a roaion model, which changes he orienaion θ of he robo by summing a consan urning facor Δθ o he previous robo s orienaion θ 1, as shown in equaion 1. The speed vecor ˆv is compued by applying he curren orienaion θ o he previous speed ˆv 1, as shown in equaion 2, where a f and a s are he forward acceleraion and he sideways acceleraion respecively. Then, he curren posiion ŝ of he robo is compued by he sum of he previous robo s posiion ŝ 1 and he curren robo s speed ˆv,asin equaion 3. The simulaor runs 60 frames per second, however his value can be easily changed if needed. v v y v = 1 v y + 1 s s y θ = θ 1 +Δθ. (1) cos(θ ) sin(θ ) af. (2) sin(θ ) cos(θ ) a s s = 1 v s y + 1 v y. (3) Robos movemens are affeced by errors. These errors are simulaed by normal disribued random numbers, ha are added o he movemens speed. There are hree kinds of errors: speed error, which causes he robo o move forward or backward; drifing error, which causes sideways movemens; and roaion error, which causes robos o roae while moving. For he ball, fricion is used for reducing is speed wih ime. Fricion is represened by a facor f which, in each ieraion, is applied o he speed ŝ 1, resuling in a new speed ŝ. Noe ha he facor f ranges from 0 o 1, as shown in equaion 4. Then, he ball s posiion is updaed following Equaion 3. v = v y f 0 v 1 0 f v y. (4) 1 Collisions are ineracions beween objecs, which can be classified in four caegories: ineracions beween robos; ineracions beween robo and ball; ineracions beween robo and saic objecs; and ineracions beween ball and saic objecs. All ineracions beween robos, ball and goal poles were approimaed by an algorihm of collisions beween circles. To preven robos and he ball o move beyond he limis of he simulaion, he borders of he simulaion were considered as walls. 319

4 B. Movemen Conrol Two processes of he Cross Archiecure are responsible for he robos movemen: Decision, which chooses he acion performed by he robo, and Conrol, which eecues he movemen. They communicae wih each oher hrough he blackboard. So, Conrol reads a variable published by Decision in he blackboard and ac following his received insrucion. In order o simulae his behavior, each simulaed robo has is own Conrol process, which reads he same variables from he shared memory ha he Conrol of he real robo would read, and eecues he same acions he real robo would eecue. C. Vision The Vision process is he main robo sensor and i is responsible for recognizing objecs in he field, like he ball, eammaes and opponens. The field-of-view of he robo is se o 70 degrees and he pan is limied o 270 degrees, following he rules of he RoboCup Humanoid League. The search for objecs in he soccer field is conrolled by he Decision process, ha defines wha kind of objec should be searched and deermines he beginning and he end of he process. During he search, objecs posiioned in he field-ofview and far up o 3 meers from he robo can be idenified. Finally, he calculaed disance value (in meers) and angle value (in degrees) wih respec o he robo are published in he blackboard. Normal disribued random numbers can be also added in hese calculaed values in order o simulae vision errors. V. A CASE STUDY Aiming he evaluaion of he porabiliy beween he proposed simulaor and a real robo, some high-level sraegies, previously developed in he simulaor, were applied in an acual robo. A. Sofware and Hardware Seup A humanoid robo inspired on DARwIn-OP [21] was used in he eperimens. This robo uses an Inel Core i5 1.66GHz compuer wih 8GB SDRAM, running Ubunu LTS. The same compuer was also used in he simulaed eperimens. For reproducibiliy reasons, he simulaor proposed in his work along wih he decision algorihms used during he eperimens are available a hp://fei.edu.br/ rbianchi/sofware. hml, as well as he documenaion ha describes is insallaion and usage. B. Eperimens This secion illusraes he use of he Decision process on he simulaor and is eension o real robos. The developmen and analysis of he Decision was chosen considering ha he Conrol, IMU and Vision are also implemened and available for he real robo. The proposed simulaor allows he scienific communiy o develop specific cogniive algorihms, such as sraegies and robo localizaion, wihou dealing wih physical robo problems. (a) Simulaion (b) Real robo Fig. 3. Firs proposed scenario for boh domains. (a) Simulaion (b) Real robo Fig. 4. Second proposed scenario for boh domains. In order o validae he ransference of he developed source code from he simulaor o a real robo, he Decision process consised of a basic approach, named naïve, in which he robo searches for he ball, walks oward i, aligns iself o he opponen s goal and kicks he ball. This work focuses in he cogniive processes of he robo, herefore he analysis of he eperimens were qualiaively made disregarding he physics of he simulaor, aiming o compare he robo s behavior beween simulaed and real environmens. Two scenarios were considered for he eperimens. Figure 3 shows he firs scenario, in which he ball is posiioned a he penaly mark and he robo is posiioned near he ball. Figure 4 presens he second scenario in which he ball is posiioned o he lef of he goal and he robo mus walk oward i, aligns o he opponen s goal and kicks he ball. In order o analyze he porabiliy of he algorihms beween simulaion and real world, he rajecory of he robo was sored and compared in boh domains. The resuls demonsrae ha he acions performed by he real robo, in boh eperimens, are similar o he acions performed by he simulaed robo, as shown in Figures 5 o 8, ha describe he rajecories of he robo and he ball for each environmen 1. Figures 5 and 6 show he sraigh pah performed by he robo in boh environmens, in which he robo is already oriened oward he opponen s goal, i walks forward and 1 The video of he eperimens is available a: hps://youu.be/w4rhseu68z4. 320

5 (a) (b) (c) (d) Fig. 5. Simulaion: pah performed in he firs scenario. (a) (b) (c) (d) Fig. 6. Real world: pah performed in he firs scenario. kicks he ball. Figure 7 and 8 show he curved pah, in which he robo walks oward he ball, urns lef or righ, according o is orienaion, aligns o he opponen s goal and kicks he ball. The eperimens indicae ha a high-level decision algorihm, implemened in he simulaor, can be reproduced in real robos wih minor changes. This is possible because he oupu of he simulaed processes (IMU, Conrol and Vision) are very similar o he oupu of hese same processes in real robos. VI. CONCLUSION This paper describes he RoboFEI-HT Simulaor developed for he RoboCup Humanoid Soccer domain, which allows cogniive algorihms o be implemened, simulaed and ransferred o real humanoid robos. The simulaor has some advanages when compared o oher eising simulaors: he usage of he Cross Archiecure, allowing he ransference of reasoning, learning and localizaion algorihms direcly o real robos ha uses his archiecure; and he fac ha he sofware is open-source, allowing he reproducibiliy of any new proposed code, even when here is no a robo available. The performed eperimens showed ha high-level algorihms developed in simulaion can be eended o he robos wihou many changes. For insance, during he eperimens, he same decision algorihm implemened in he simulaor was eended o he real robo wihou many changes and he robo performed he same acions as running in simulaion. As fuure work, new cogniive acions such as reasoning, learning, and self-localizaion can be implemened and esed in order o evaluae he simulaor and he porabiliy of some oher high-level Arificial Inelligence algorihms o real robos. ACKNOWLEDGMENT Danilo H. Perico, Thiago P. D. Homem, Isaac J. Silva, Claudio O. Vilão Jr. and Vinicius N. Ferreira acknowledge suppor from CAPES; Aislan C. Almeida acknowledges suppor from CNPq. REFERENCES [1] D. H. Perico, I. J. Silva, C. O. Vilão Junior, T. P. D. Homem, R. C. Desro, F. Tonidandel, and R. A. C. Bianchi, Roboics: Join Conference on Roboics, LARS 2014, SBR 2014, Roboconrol 2014, São Carlos, Brazil, Ocober 18-23, Revised Seleced Papers. Berlin, Heidelberg: Springer Berlin Heidelberg, 2015, ch. Newon: A High Level Conrol Humanoid Robo for he RoboCup Soccer KidSize League, pp [Online]. Available: hp://d.doi.org/ / [2] C. Yang, P. Liang, and P. Avgeriou, A sysemaic mapping sudy on he combinaion of sofware archiecure and agile developmen, Journal of Sysems and Sofware, vol. 111, pp , [3] R. Murphy, Inroducion o AI Roboics. Cambridge, MA: MIT Press,

6 (a) (b) (c) (d) Fig. 7. Simulaion: pah performed in he second scenario. (a) (b) (c) (d) Fig. 8. Real world: pah performed in he second scenario. [4] R. A. Brooks, A robus layered conrol sysem for a mobile robo, Roboics and Auomaion, IEEE Journal of, vol. 2, no. 1, pp , [5] R. C. Arkin and T. Balch, Aura: Principles and pracice in review, Journal of Eperimenal and Theoreical Arificial Inelligence, vol. 9, pp , [6] R. C. Arkin, Behavior-based roboics. MIT press, [7] I. J. Silva, D. H. Perico, T. P. D. Homem, C. O. Vilão, F. Tonidandel, and R. A. C. Bianchi, Using reinforcemen learning o improve he sabiliy of a humanoid robo: Walking on sloped errain, in h Lain American Roboics Symposium and rd Brazilian Symposium on Roboics (LARS-SBR), Oc 2015, pp [8] B. Hayes-Roh, A blackboard archiecure for conrol, Arif. Inell., vol. 26, no. 3, pp , Aug [9] C. O. Vilão, D. H. Perico, I. J. Silva, T. P. D. Homem, F. Tonidandel, and R. A. C. Bianchi, A single camera vision sysem for a humanoid robo, in Roboics: SBR-LARS Roboics Symposium and Roboconrol (SBR LARS Roboconrol), 2014 Join Conference on, Oc 2014, pp [10] D. H. Perico, R. A. C. Bianchi, P. E. Sanos, and R. L. de Mánaras, Collaboraive communicaion of qualiaive spaial percepions for mulirobo sysems, in Proc. 29h Inernaional Workshop on Qualiaive Reasoning (IJCAI), New York, NY, USA, 2016, pp [11] A. Farchy, S. Barre, P. MacAlpine, and P. Sone, Humanoid robos learning o walk faser: From he real world o simulaion and back, in Proceedings of he 2013 inernaional conference on auonomous agens and muli-agen sysems. Inernaional Foundaion for Auonomous Agens and Muliagen Sysems, 2013, pp [12] Gazebo - (robo simulaion made easy), hp://gazebosim.org, 2016, las Accessed in: 19/05/2016. [13] Webos, hp:// 2016, las Accessed in: 19/05/2016. [14] Virual robo eperimenaion plaform, hp:// 2016, las Accessed in: 19/05/2016. [15] Robocup soccer simulaion. RoboCup. [Online]. Available: hp: //wiki.robocup.org/wiki/soccer Simulaion League [16] F. J. C. Monenegro. (2015) Taurasim. Taura Bos eam, UFSM. [Online]. Available: hps://gilab.com/splinerfm/taurasim [17] I. Afanasyev, A. Sagiov, and E. Magid, Ros-based slam for a gazebosimulaed mobile robo in image-based 3d model of indoor environmen, in Advanced Conceps for Inelligen Vision Sysems. Springer, 2015, pp [18] A. Saranowicz and G. L. Marioini, A survey and comparison of commercial and open-source roboic simulaor sofware, in Proceedings of he 4h Inernaional Conference on PErvasive Technologies Relaed o Assisive Environmens. ACM, 2011, p. 56. [19] Y. Xu and H. Vaankhah, Simspark: An open source robo simulaor developed by he robocup communiy, in RoboCup 2013: Robo World Cup XVII. Springer, 2013, pp [20] Pygame. [Online]. Available: hp:// [21] I. Ha, Y. Tamura, H. Asama, J. Han, and D. W. Hong, Developmen of open humanoid plaform DARwIn-OP, in SICE Annual Conference 2011, 2011, pp