Two Humanoid Simulators: Comparison and Synthesis

Transcription

1 Two Humanoid Simulators: Comparison and Syntesis Nima Safii, Luis Paulo Reis, Rosaldo J. F. Rossetti Artificial Intelligence and Computer Science Laboratory Department of Informatics Engineering Faculty of Engineering of te University of Porto - FEUP Porto, Portugal {nima.safii, lpreis, rossetti }@fe.up.pt Abstract In tis paper, an overview of two umanoid simulation platforms is provided: Simspar, a 3D Robocup simulator and te robotics simulator SimTwo. Altoug tese two simulators ave different bacground, today tey sare te same umanoid robot model, namely te Albaderan NAO robot. According to te fact tat developing reliable and robust biped locomotion and low level umanoid beaviors are still callenging tass, simulation as an important role in improving umanoids movement development approaces. In tis paper te two umanoid robotic simulators will be compared in face of simulating Humanoid low level beaviors. Te comparison is based on identifying te same role and locomotion approaces. Te results sow tat Simspar is closer to reality tan SimTwo. Keywords-Bipedal Locomotion; Humanoids Simulation; Soccer Humanoid Robots. I. INTRODUCTION Humanoids robots try to mimic uman-lie beaviors and movements. Wile weeled robot locomotion is not adapted to many uman environments, suc as stairs and areas littered by many obstacles, umanoid robots are able to avoid different sapes of obstacles, and attain postures tat are more desirable. Terefore, by using biped locomotion, umanoid robots can function and perform teir tass easier tan weeled robots in areas designed for people. According to te fact tat biped locomotion is similar to uman movement, tis similarity causes people to interact wit umanoid robots easier as well. Altoug umanoid robots ave lots of motivations beind tem, te control of umanoid waling and running is still callenging and as not been solved [1]. Beside, biped locomotion is nown as one of te most complicated robotics tass in te field. Te idea beind a umanoid simulator is to develop a virtual agent capable of tining and acting. Terefore, acquired nowledge from simulation can be transferred to real robots. Simulation is an easier way to develop biped locomotion metodologies. It maes performing gait optimization [2] and learning pase [3] easier and more efficient. It is also not reasonable to test te algoritm directly on te real umanoid robot since learning needs lots of iterations and umanoids always are very expensive and ard to troublesoot in te case of falling or breaing. Terefore, Humanoid simulators must emulate te real world in te best possible way. To mae tis feasible, it is necessary to construct accurate and reliable models of te real robots. On te oter and, pysical Simulators always contain some simplifications wen it intended to model te real world, meaning te result of simulation and reality is usually not te same [4]. Tis problem is called te reality gap in te field of evolutionary robotics [5]. Researcers always try to find a simulator wic as te least differences between its result and reality [6]. In tis case of study, some researcers try to compare different simulators and find te differences between simulators in order to find exact causes of differences between tem. It can also lead tem to obtain te best model of a simulator tat is closer to te reality. Performance of umanoids simulation depends on many factors related to modeling te pysical environment and eac simulator uses its specific arcitecture suc as pysics engine and metods for discretizing time. A clear approac to compare efficiency of simulating umanoid does not exist yet. In 2009, Sivaram and Stone compared tree umanoid robotics platforms [7]. Tey compared te simulators by analyzing results of robots waling wit te same robots model and same approac of waling. For example, tey compared te speed of waling in simulators wit te speed of waling in reality, so as to conclude wic simulator is closer to te real world. In tis paper we coose two different umanoid simulators, simtwo [8] and Simspar (RoboCup 3D league Simulator) [9], wic simulate te same model of umanoid robot, NAO [10]. Tey will be overviewed, ten; a metodology will be presented to compare tem in order to guide us to find out te better simulator. By implementing te metodology on tese simulators, we intend to generate some results tat will allow us to carry out appropriate comparison. Finally, comparing tese results can lead us to find te exact causes of differences to find out wic of tese simulators is better to use as a umanoid test bed. Te rest of te paper is organized as follows. Section 2 presents an introduction to te umanoid robot model and platform. Section 3 and 4 present respectively te RoboCup 3d league simulator and te SimTwo simulator. Section 5 presents our metodology for comparing te simulators. Section 6 presents te implementation and results. Te paper concludes wit some final remars and pointers for future wor. II. HUMANOID ROBOT MODEL AND PLATFORM A umanoid robot is a robot wit its overall appearance based on a uman body and tat is able to stand and move on its own two legs. Te number of teir joint actuators indicates

2 te number of Degree of Freedom (DOF). Lie umans, umanoid's body moves in tree planes, including transverse (axial), frontal (coronal) and sagittal. Sagittal plane indicates te vertical plane running from front to bac and divides te body into left and rigt sides. Frontal or Coronal plane is a plane perpendicular to te sagittal plane, running from side to side and dividing te body into front and bac. In te NAO robot model, 12 DOFs are embedded in its legs. Tere are tree DOFs in eac legs moving in te sagittal plane: one in te ip, one in te anle and one at te nee. Tere is also one actuator in te ip, wic can turn and move te leg in transverse plane. In addition to DOFs of legs, te model as 10 DOFs in te upper parts of its body. Fig. 1 illustrates a scematic view of a umanoid robot. As it was described, DOFs 1, 3, and 4 move on Sagittal plane, DOFs 2 and 5 move on Frontal plane and DOF 6 moves on transverse plane. Figure 1. III. A. Human body planes, B. NAO joint's configuration ROBOCUP 3D LEAGUE SIMULATOR Te Robocup Simulator is a generic simulation platform for pysical multi-agent simulations and it is used in te RoboCup 3D simulation league currently. Tis simulator as been developed by te RoboCup community since It is designed as a flexible application framewor and intends to be a generic simulator, capable of simulating anyting, from te launc of a projectile for academic purposes to a big soccer game for scientific researc purposes. Te framewor facilitates excanging single modules and extending te simulator [11]. Te Simulator consists of two important parts: Server, and Monitor. A. Server Te server is responsible to andle connections from te agents; to receive and to process messages, send reply messages to te agents. Te server arcitecture is illustrated in Fig. 2. simulation engine, wic allows simulating te system s dynamics and te pysical properties of te simulated objects. It provides advanced joint types and integrated collision detection wit friction. ODE is particularly useful for simulating objects in virtual reality environments. It is crossplatform and provides a user-friendly C/C++ Application Programming Interface (API). RoboCup simulator is very well-nown because it can be seen as a multi-agent environment enabling different types of experiments. Several agents can connect to te server simultaneously. Its object and memory management is based on Zeitgeist [11]. It is a framewor for andling data objects and functional components of a system in a uniform way, wic strictly follows te object-oriented paradigm using C++ programming language. Te core of te simulator is a simulation server. It receives te messages wit actions from te agents, performs te simulation operations and sends a reply message bac to te agent wit te environment information. Te simulation engine acts as a server, andling te messages of te agents and replying on to oter messages. In Eac cycle of te simulation, agents send a message to te server containing information about teir effectors (e.g. joints). Te message from te server to te agent contains temporal information and specific information from te application domain (wic is soccer). Tis information includes game state (play mode, time and current result), and information of te preceptors of te robot (e.g. joints, gyroscopes, foot sensors, vision information). Te messages are constructed using a LISP-lie format. B. Monitor Te monitor provides a simple grapical interface tat allows te user to watc a simulation. Simulations may be watced in real-time, but it is also possible to play simulation log files offline, for post processing assesment. In te particular case of umanoid soccer simulation, it provides additional information suc as te team names and te game time, play mode and results. Several sortcut eys can be used to cange camera views, to drop te ball, and carry out oter useful operations. IV. SIMTWO ARCHITECTURE In 2006, SimTwo [8] was developed by Dr. Paulo Costa in te University of Porto. It focused on preparing a simulated environment to test and develop mobile robot approaces and applications. Figure 2. Robocup Simulator Server Arcitecture [11] Te simulator uses te Open Dynamics Engine (ODE) [12] to simulate te pysical environment. ODE is a pysical Figure 3. Snapsot from SimTwo environment

3 Te SimTwo system is a realistic simulation were anyone can implement various types of mobile robots wit different configurations including te different body mass and sape and any types of joints, wic can be described wit a mixture of classic joints and weels. Terefore, in addition to simulate weeled robots, umanoid robots can also be simulated. Recently, te model of NAO robots was also added to tis simulator. In Fig. 3, te environment of SimTwo is sown Te realism of te dynamics implemented in SimTwo is acieved by dividing a robot in to a system of rigid bodies and electric motors. Te "mecanics" associated wit bodies is numerically simulated considering teir pysical sape, and mass moments of inertia, friction and elasticity of te surfaces. Lie te Robocup simulator te Pysical simulation is performed by using ODE, wic was described in te previous section. Certain joints are designed as saft gimbals typically and as pipeline explicitly. Eac of tem can ave its associated drive system and sensors. Te drive system may consist of a DC motor and gearbox. Te DC motor model contains several elements suc as nonlinear saturation of te applied voltage, current limit and friction. To define te environmental configurations in te simulator, SimTwo can design any types of obstacles wic as different sape and role suc as slops, stairs, cubes and so on. Obstacles and robots can also be moved easily during te simulation by dragging mouse on tem. To define and implement te environmental configuration and robot's model, SimTwo as different tools, wic allows developers to import teir configuration as XML scripts. In order to sow te results of te simulation, a grapical viewer based on OPEN-GL is also implemented. A debugger tool is also developed as a grapical viewer, wic can draw any graps and trajectories. In Figure bellow tool-boxes of te SimTwo can be found. Te two windows on te rigt.sow te XML-based environment as well as te visual viewer and debugger on te left. Figure 4. View of SimTwo tool boxes SimTwo as good capabilities to simulate actuators and to define obstacles and environment features, but it does not ave adequate arcitecture to manage several agents to connect to te simulator as a multi-agent environment. V. METHODOLOGY TO COMPARE THE SIMULATORS Suc as in [7], our comparison approac is based on a bencmar metod. In tis metod, results and comparisons of running a computer program are acieved by running a number of standard tests and trials against it, in order to assess te relative performance of an object. To coose our standard tests, we run te same umanoid scenarios in different simulators. Terefore, a umanoid scenario must ave te same dependencies and caracteristics, including: Same approac for modeling agents beavior cycle; Same approac for controlling robots; Same macine specification to run te simulation. In tis paper two umanoid scenarios will be proposed, wic are called Vibrating and Waling. Following te caracteristics and dependencies of tese scenarios will be explained. A. Modeling of Agent's Beavior Cycle Performing agent's tass and scenarios are related to agent's beavior arcitecture, wic is implemented inside te agent. Tis arcitecture must be defined as a cycle, since agents must receive sensor data from te simulator, ten process it and try to mae decision based on it and after tat, send an action bac to te simulator. Te figure below sows te arcitecture of our agent beavior cycle. Figure 5. Agent beavior cycle arcitecture Sensors from te simulator are received as a string. So, a message parser unit is needed to interpret it and prepare needed information for producing world state. According to te fact tat our agent will be implemented in two different simulators, terefore te sensor's string and message parser unit will be different. Because our comparison needs to use te same decision-maing structure, it is very important to prepare te same information from sensory input. World state unit processes all information from te world wic is needed for agents to perform its decision. In different simulators, te world state of te agent must also be te same. For example coordinate system of joint's angle is not te same in different simulators, terefore tese inerent differences must be solved and considered by te world state unit. In tis study, umanoids low-level movements ave been cosen as our scenarios, we divided decision-maing unit in two separate units: gait generator and joint control. Following related to eac scenarios, te model of tese two units will be explained. B. Modeling of robots waling (First Scenario) Based on te wor referenced in [7], te main and first scenario wic is presented for comparison between simulators is waling, since waling uses lots of actuators and it is nown as a very complex motion. According to te fact tat all of te pysical futures can influence te performance of its simulation

4 directly, it is very ard to simulate biped locomotion, especially some features lie friction between surface and robots. A Biped locomotion approac based on uman motion captured data and Truncated Fourier Series is used in tis study. In 2007 Truncated Fourier Series Formulation (TFS) metod was used as a gait generator in bipedal locomotion [13]. Recently, an optimized gait generator based on TFS was implemented on a simulated umanoid robot and TFS less parameters were also reduced by 2 dimensions (down to 6 dimensions) [14] [15]. According to te fact tat tis model ad least parameters and more simple approac compare to oter newer approaces, it is considered for te first scenario. In te following its detail will be explained more. In tis approac, movements of tree DOFs in eac leg, wic are moved in sagittal plane, will be calculated. Foot in sagittal plane was also ept parallel to te ground by using anle joint in order to avoid collision. Terefore anle trajectory could be calculated by ip and nee trajectories and anle DOF parameters were eliminated. In tis model, legs joint angular trajectories in sagittal plane are divided in two parts; te upper portion and te lower portion. Te TFS for generating eac portion of ip and nee trajectories are formulated below. 2π = + = T n θ B i.sin ( iw t ) c, w i = 1 2π = + = T n θ + A i.sin ( iw t ) c, w i = 1 θ n + = C i.sin ( iw t ) + c, w = w i = 1 θ = c 0 In tese equations, C is offset of ip trajectory and C is offset of nee trajectory. Te plus (+) sign represents te upper portion of waling trajectory and te minus (-) sows te lower portion. i=1 and A i, B i, C i are constant coefficients for generating signals. Te and index stands for ip and nee respectively. C, C are signal offsets and T is assumed as a period of ip trajectory. Considering te fact tat all joints in waling motion ave equal movement frequency and stride rates is statistically equal, te equation w =w = can be concluded. Wile Left leg is considered as supporting leg, te variation of its nee angle is so minute tat can be assumed fixed, Tis duration of waling is named nee loc pase and θ = c 0. Te amount of sift pase of te two leg trajectories signal is one alf of te period of eac signal so by producing trajectory of one leg te oter leg's trajectory can be calculated. Te trajectories for bot legs are identical in sape but are sifted in time relative to eac oter by alf of te waling period. Te waling scenario must be implemented as te gait generator unit, in te same way. For generating proper waling trajectory, a robot must learn and find te best value of te parameters wic leads a robot to wal more stably and (1) robustly. According to te fact tat gait generator will produce angular trajectories, Control unit must control te joint actuators along teir references. In te next section we will introduce our optimization metod wic is used in tis project. C. Modeling of Vibration sill (second scenario) In tis scenario, in order to detect differences of actuators simualtion, we ave cosen te elbow joint, wic as te least friction wit surface. Terefore, environmental factors ave less influent on simulating its movements. After te, in order to find te differences better, a scenario wit te biggest cange of power of elbow's actuator is presented. Te algoritm of te scenario will be presented below. It commands a robot to move its elbow to reac 50 degree and command to go to te opposite side, -50 degree in te next simulation step and after tat repeats te scenario again. Due to a quanta of simulation step is reasonably small, 0.04 and 0.02 second in SimTwo and Simspar respectively, to reac to tis angle a robot must utilize te igest power of its elbow's actuator. if cangestep=true ten begin SetAxisPosRef (0, GetAxisIndex (0,'rigt_elbow', 0),-50); cangestep:=false; end else begin SetAxisPosRef (0, GetAxisIndex (0,'rigt_elbow', 0),50); cangestep:=true; end; Tis scenario will be implemented as te trajectory generator unit of te second scenario wit te same configuration on bot simulators. As a controlling unit, a PD controller wit te same specification will be used. VI. IMPLEMENTATION AND RESULTS To compare SimTwo and Simspar bot of tem were tested on macines wic ave te same macine specification. Te specification of macines are Pentium IV 3 GHz Core 2- Duo wit 2 GB of pysical memory. Scenarios considering te same specification are implemented on te simulators. As it was mentioned before, joints angular trajectories produced by eac scenario are used by a simulated robot to test it. In order to utilize tem, all individual robot's joints sould attempt to drive towards teir target angles using proportional derivative (PD) controllers. Same configuration of te PD must be applied for eac different simulator, In tis project, te P factor was assumed to be 200 and te D was assumed to be 2. In te following, implementation result of two scenarios in different approaces will be discussed. A. waling (First Scenario) According to first scenario wic was sown in te previous section, te best parameters to generate joints angular trajectories for bipedal locomotion must be found. According to [14], for tis ind of optimization problem, Genetic Algoritms (GA) can be used. Terefore on RobocupSim, first Robots try to find te best Parameters by using GA. Ten results of it will be used for te waling robot wic is

5 simulated on te SimTwo. By extracting te relation between motor's command and Joint's angle, and comparing speeds of waling and joints angular trajectories extracted from simulation, comparison can be done. 1) Simspar Results For using GA as an optimizer in Simspar, te cross over rate and mutation rate are set to 0.8 and 0.06, respectively. Population for eac generation is 100 and selection metod is roulette weel. Termination condition is to ave a generation counter greater tan 28. Terefore te GA requires 2800 trials to find appropriate TFS parameters. After 9 ours, since starting GA on te macine, generation exceeded to 28 and te robot could wal 34.5m in 77 s wit average body speed of around 0.45 m/s and time period of eac step was about 0.41s. GA led te robot to learn ow to wal straigt, Fig. 6 exibit biped locomotion is obtained from GA searc and in Fig. 7 angular trajectory generated by TFS after learning process is sown. Followed trajectories based on te learned trajectory is also sown in tis figure. same waling trajectories and te results of te waling must be te same in bot simulators. But, after implementing first scenario on SimTwo, te umanoid could not wal. Tis result sows differences between tese two simulators. Fig. 9 sows te umanoid robot felt down during its waling in te SimTwo environment. As an additional test, we allowed te umanoid robot to wal above te ground. Tis scenario can allow a robot to test its waling witout environmental friction and ground force. Fig 9 also sows te robot wic is waling above te ground. Results were very interesting since sowing tat te robot can generate and follow te exact joints angular trajectories, wic were produced by te model on Simspar. Fig. 10 sows te angular trajectory of te left ip wic is performed by SimTwo wen robot is waling scenario above te ground. Figure 8. B) Robot waling above te ground A) Robot felt in during waling Figure 6. Waling is obtained by GA on te Simspar Figure 9. Hip Angular Trajectories, generated by waling above te ground In conclusion, te results of testing first scenario on bot simulators were different. It as also proved tat different approaces are used for modeling and simulating actuators especially in te case of facing surface friction and ground forces, because witout tese environmental elements, generated angular trajectories were te same in bot. B. Vibration Scenario Figure 7. Angular Trajectory generated by learned TFS (Desired angle) and followed one by controller (Followed angle) for left ip and left nee of simulated Nao robot 2) SimTwo Results Te values of te waling parameters acieved by RobocupSim are used for te waling scenario, wic is simulated in SimTwo. According to te fact tat te Humanoid in SimTwo uses te same controller's configuration and gait generator model, terefore te robot's model must produce According to te fact tat bot simulators acieved different results in te manner of simulating actuators, vibration scenario was very useful in order to study te differences between simulators. It uses te biggest cange of actuator eac time. It as been implemented on bot umanoid simulators wit te same controller unit caracteristics. In SimTwo, simulation results are not very realistic, since it sows te elbow actuator is very powerful an it can move te elbow to reac 10 degree in duration of one step and in te next simulation step (0.04 s) it can reac -10 degree. Terefore, in tis situation te simulated actuator of elbow can create a big torque, wic can rotate te elbow 500 degree/sec. It means

6 tat te simulated actuator in SimTwo needs to use ig power or voltage to produce tis uge torque. Terefore, it seems not to be a real simulation since NAO robot as a small litium battery to prepare tis ig power. Fig. 10 sows a snapsot from SimTwo wile robot is executing te second scenario. Fig. 11 also sows te angular trajectory of elbow simulated on SimTwo. tat te 3D league RoboCup simulator is closer to reality tan SimTwo. According to te fact tat SimTwo is very interesting in wat concerns its tools and global facilities, turning it into a more realistic simulator may be a ey point for its more global use. In addition, it can be interesting to use te results of tese two simulation studies on a real robot to find out wic of tem is more near to reality troug a new set of experiments. ACKNOWLEDGEMENTS Te first autor is supported by FCT under grant SFRH/BD/66597/2009. Tis wor as been partially funded by FCT Project ACORD - Adaptative Coordination of Robotic Teams (PTDC/EIA/70695/2006). Figure 10. NAO executes te Vibration on SimTwo. It is also sown te elbow moves wit big variations. Figure 11. Angular trajectory of elbow wit in duration of 1 sec. simulated in SimTwo However, variation of degree on SimTwo was very big, but implementation of te vibration scenario on Simspar sows te angle of elbow canges varied less tan four degree for eac step. It means tat its actuator can produce power for moving te elbow wit te maximum speed of 100 degree/second and it is not utilizing te big power. Terefore, simulated actuator by Simspar is closer to reality tan SimTwo. However in te time tat tis paper as been written, group tat developed te SimTwo project was woring on it and already acieved better simulation results. So, it is expected tat te new SimTwo version may acieve results similar to tose acieved today by te RoboCup simulator. VII CONCLUSION In tis paper, te RoboCup 3D league Simulator and SimTwo were reviewed concerning teir capabilities to simulate umanoid robots. In order to compare tem, two different scenario based on bencmar approaces were presented. After implementing tese scenarios, te results sowed differences in teir beaviors mostly in te case of simulating actuators. Analyzing te results led us to conclude REFERENCES [1] E. Westervelt, As Te Experts: Biped Waling, IEEE Control System Magazine, pp , [2] J. E. A. Bertram, Constrained optimization in uman waling: cost minimization and gait plasticity, Te Journal of Experimental Biology, Vol. 208, pp , [3] K. Wolff, J. Pettersson, A. Heralic, and M. Wade, Structural evolution of central pattern generators for bipedal waling in 3d simulation, Proceedings of te 2006 IEEE International Conference on Systems, Man, and Cybernetics (SMC 2006), pp , [4] J. C. Zagal, J. Ruiz, Combining Simulation and Reality in Evolutionary Robotics, Journal of Intelligent and Robotic Systems, Vol.50(1), pp , [5] S. Koos, J.-B. Mouret, Crossing te reality gap in evolutionary robotics by promoting transferable controllers, Proceedings of te 12t annual conference on Genetic and evolutionary computation, pp , [6] T. Laue, M. Hebbel, Automatic Parameter Optimization for a Dynamic Robot Simulation, L. Iocci et al. (Eds.), RoboCup 2008, LNAI 5399, pp , [7] S. Kalyanarisnan, T. Hester, M. Quinlan, Y. Bentor,P. Stone, Tree Humanoid Soccer Platforms: Comparison and Syntesis, Proceedings of te RoboCup International Symposium, LNCS/LNAI Springer, pp , [8] J. Boedecer and M. Asada, SimSpar Concepts and Application in te RoboCup 3D Soccer Simulation League, Autonomous Robots, 2008, pp [9] J. Boedecer, Humanoid Robot Simulation and Waling Beaviour Development in te Spar Simulator Framewor, Artificial Intelligence Researc University of Koblenz, [10] NAO Albedaran, available at: ttp:// [11] O. Obst, M. Rollmann, Spar - a generic simulator for pysical multiagent simulations. In: LNCS. V. 3187, G. Lindemann, I. J. Timm, R. Unland, 2004, Springer. [12] R. Smit, ODE _ Open dynamics engine, [13] L. Yang, C. M. Cew, A.N. Poo, Adjustable Bipedal Gait Generation using Genetic Algoritm Optimized fourier Series Furmulation, In Proceedings of te 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp , [14] N. Safii, M.H. Javadi, B. Kimiagalam, A Truncated Fourier Series wit Genetic Algoritm for te control of Biped Locomotion, In Proceeding of te 2009 IEEE/ASME International Conference on advanced intelligent Mecatronics, pp , [15] N. Safii, L.P. Reis, and N. Lau, Biped Waling using Coronal and Sagittal Movements based on Truncated Fourier Series, RoboCup 2010: Robot Soccer World Cup XIV, Javier Ruiz-del-Solar, eds., 2011, pp