Home

RoboSoc a System for Developing RoboCup Agents for

1. The Sensors The Controller The Actuators gt pars ng control data results raw sensor server commands ata The Server Interface UDI IP messages The Soccer Server Figure 3 3 The data flow in the basic system of RoboSoc The basic system consists of the server interface the controller the sen sors the sensor interface the actuators the actuator interface and the de cision maker The rest of this section is devoted to describe these different subparts in detail 3 3 1 The server interface The server interface takes care of the lowest level of communication with the server by providing an interface to the socket implementing the actual server communication The controller uses the interface to connect to the server when the agent is started receive the server messages when they arrive and send the commands back to the server 3 3 2 The controller One of the most important units in RoboSoc is the controller It is respon sible for the timing of the agent and the synchronization with the server Its task is to keep track of the current game time receive messages as soon as they arrive and control when to act It should also dispatch the messages from the server to the sensors and guide the decision maker by generating events To complete its task the controller uses an internal clock an IO handler and three control algorithms The internal clock will generate an interrupt 27 every recv s
2. The ball view The ball view keeps track of the ball its absolute position relative distance and direction and a very rough estimation of its speed In Table 3 8 all the methods are described All relative angles are relative to the facing direction of the agent taken from the agent view The player view The player view keeps track of the other players It will give access to three Structures teammates opponents and unknown players Each will give information about the absolute position relative distance and direction body direction face direction and uniform number not available for the unknown players In Table 3 9 all the methods are described All relative angles are relative to the facing direction of the agent taken from the agent view 39 Table 3 9 Functionality provided by the player view Get ClosestPlayer hist A pointer to a player object Get the closest seen ME Aponia to ap arer ON player hat eydor ago Get ClosestOpponent hist A pointer to a player object Get the closest seen oppo ER A polmer foa plaat ON nent seen it eyle ago Get ClosestTeammate hist A pointer to a player object Get the closest seen teammate seen hist cycles ago Get Opponents hist A vector of player objects Get the opponents seen puer eS y Get Teammates hist A vector of player objects Get the teammates seen pou eee ce e GetUnknownPlayers hist A vector of player objects Get the unknown players e PN pem ist eyeter age
3. which returns the observation time SetObserva tion Time Time which sets the current observation time and Get CyclesS ince Time which returns the number of cycles between observation time and time There is a compiler directive called PESSIMISTIC_OBSERVATIONS_ which controls how observation times are updated when doing arithmetic with modified values If it is set then the earliest observation time value is taken otherwise the latest The standard modifiers WithUnknown c type WithConfidence type and WithObservationTime lt type time does not work with the templates for the angles and coordinates there are special versions of the modifiers available called AngleDegUnknown type for unknown AngleDegUC type for both unknown and confidence and AngleDegUCO type for unknown confidence and observation time and similar for AngleRad CoordVector and CoordPoint A 3 The basic system This section describes the basic system of RoboSoc and how to use it A 3 1 How to create an agent To create an agent using the RoboSoc basic system without the framework you have to create a main function where you start the different parts In Example A 1 you can see how the code for a basic agent might look like The function rs_debug is used for debugging and is only needed if the com piler directive USE DEBUG defined in library compiler directives h The constructors for the different units are described in the following
4. 10 100 150 if controller Init std cout lt lt Error in Init nExit n return 1 controller Start return 0 63 Example A 2 An example of a main loop void CBasicDecision Start i while serverAlive if pause 1 amp amp errno EINTR Std cerr lt lt Something went wrong in pause Mn 64 Example A 3 An example of a basic decision maker CCommand CBasicDecision InitAgent f return new CInitCommand RoboSoc 5 false void CBasicDecision Decide 1 if ballDir IsUnknown ballDir lastTurn lastTurn 0 if sensorInterface gt SawBall ballDir sensorInterface gt GetBallDirection ballDist sensorInterface gt GetBallDistance counter 0 else counter if counter gt 4 ballDir MakeUnknown ballDist MakeUnknown counter 0 if ballDir IsUnknown actuatorInterface AddPlan new CTurnCommand CAngleDeg 30 else if ballDist lt 1 actuatorInterface AddPlan new CKickCommand 100 CAngleDeg 0 else if Abs ballDir lt 5 actuatorInterface gt AddPlan new CDashCommand 50 else 1 actuatorInterface AddPlan new CTurnCommand ballDir GetDeg void CBasicDecision nActuatorSensorData const CCommand const cmd sensorInterface gt GetLastCommand if cmd NULL amp amp cmd gt GetType CMD Turn i lastTurn static_cast lt const CTurnCommand gt cmd gt G
5. 180 turn neck Moment Moment minneckmoment maxneckmoment degrees default 180 180 turn neck is relative to to the direction of the body and the resulting angle must be between minneckang and maxneckang degrees default 90 90 Can be invoked at the same cycle as a turn dash catch or kick 2 1 5 The command cycle The command cycle ie the cycle within which the agent must reason decide on action and send it to the server is actually composed of four different cycles running in parallel The four cycles are e The simulator cycle controls the time in the simulation The length is determined by the parameter simulator step its default value is 100 milliseconds e The send physical sensor data cycle controls when to send the current physical sensor data to the agent The length is determined by sense body step its default value is 100 milliseconds e The send visual sensor data cycle controls when to send the cur rent visual sensor data to the agent The length is determined by send step its default value is 150 milliseconds e The receive commands from the clients cycle controls when the server receives commands from the clients The length is determined by recv step its default values is 10 milliseconds The four cycles are controlled by the same internal clock inside the server and due to the new implementation of the server they are updated quite accurately But they are still asynchronous as shown in Fig
6. Table 3 10 The truth tables used by the predicates JW ne opu pv Ope E uj pt f jp jo pe go pu Jue jp qp De lj g t j f f f Jf jt jf u u u ju qu f ju ju jt fu fua 3 4 3 The predicates The purpose of the predicates is to encapsulate the concept of a predicate in a three valued logic definition of the concept In other words one has a static feature the predicate should describe like is the ball on our half of the field with or without parameters that can either be true false or unknown The truth tables used comes from Kleene 15 and are shown in Table 3 10 The predicates are not automatically updated when new information is available but they can use information stored in the views and thereby get access to historic data 3 4 4 The skills The skills encapsulate the concept of acting or rather short term specialized planning or to use the current world model as provided by the views and the predicates to derive a sequence of primitive actions that will make the agent perform a certain task or take it towards an intended goal state Every skill should do two things first of all it should be able to deter mine if it is applicable at a given moment or if the goal state is already reached or some of its preconditions are not fulfilled It should also be able to generate a plan a sequence of primitive actions consisting of at least one primitive action The primitive actions are the actio
7. sent in the two cycles the first to arrive will be heard The physical sensors Each sense_body_step the agent will receive physical sensor data In Ta ble 2 3 the data given by the physical sensor is described 2 1 4 The available commands This section will discuss what basic actions the agent can tell the server to perform Table 2 4 contains a short description of the syntax of the commands the range of the parameters and how often the actions can be sent The detailed description of the effects of the actions below is taken from the Soccer server manual e turn The turn moment must be between minmoment and maxmoment 180 degrees and 180 degrees by default However there is a concept of inertia that makes it more difficult to turn when the agent is moving The actual angle the player is turned is calculated as follows ActualAngle Moment 1 0 inertia_moment x player_speed Note that player speed is the amount of the player s velocity vec tor and is therefore always positive inertia moment is a parame ter with default value 5 0 Therefore with default values when the player is at max speed 1 0 the maximum effective turn it can do 12 is 30 However because the agent can not dash and turn in the same cycle the fastest a player can be going when executing a turn is player speed max x player decay which means the effective turn with default values is 60 turn neck Each client has a neck which ca
8. C on a UNIX platform 3 1 Overview The design goal of RoboSoc is to create a system for developing RoboCup agents especially but not only for students which is as general open and easy to use as possible and that encourages and simplifies the modification extension and sharing of RoboCup agents and parts of them As stated in the introduction this is a very open problem and some assumptions about what the user wants from the system have to be made I assume the user wants 1 accurate complete and consistent data in that order of importance 2 use as much time as possible for the decision making 3 rather act on incomplete information than not act at all 4 soccer objects the agent can manipulate like representations of the ball and the players 22 FRAMEWORK amp USER BASIC SYSTEM amp USER dX Decision Maker View Manager Sensor Interface Sensors Actuator Interface r i Actuators BASIC SYSTEM Ii Controller 4 1 Server Interface y UDP P m Soccer Server Object Collection of objects Program Figure 3 1 The RoboSoc architecture The design goal together with the four assumptions above and the four essential problems discussed in section 2 2 has lead to a system consisting of three major parts e a library of utili
9. The commands and their properties will be further discussed in section 2 1 4 Since the UDP IP protocol is unreliable data to and from the server will be lost in the network This together with the noisy sensors and the fact that players can act as often as 10 times a second makes the simulation highly dynamic and uncertain This also makes the task of developing RoboCup agents very interesting Apart from the simulation server there is also a visualization tool called the Soccer monitor where you can watch the games as they are played In Figure 2 1 you can see what it looks like The players are the bigger circles 4 Figure 2 1 The Soccer monitor and the small circle next to the dark player below the middle circle is the ball The players are facing the direction of the light half of the circle The black bars on the left and right side of the field are the goals The text in the three shaded areas on top of the screen is from the left the name of the team playing on the left side and its score the current play mode and the current time and the name of the team playing on the right side and its score In this example FCFoo is playing on the left side and has scored zero goals the current play mode is play on and the current time is 482 the team on the right side is CMUnited and they have also scored zero goals The rest of this section will describe important parts of the simulation in more detail Most of the material comes
10. a new RoboCup team from scratch usually takes at least six months often more Since we want students to be able to use RoboCup to implement and test different AT algorithms we need some way to make the students start from a much higher level so they can spend their time implementing the AI parts and skip most of the tedious low level programming needed to create a functional RoboCup team Therefore a platform for creating new teams is needed This platform can also be used by non students creating their first team but I think the educational aspects put higher demands on the ease of use clean design and sharability so that the teams can improve from year to year even with different students and different approaches The platform should first of all take care of the most basic tasks of a RoboCup agent namely receiving data from the server interpret this sensor data and sending commands back to the server To be able to do this the platform must have a well defined work cycle since timing is very important and events are asynchronous It also needs data structures for storing the results from the interpretation of the sensor data Finally it needs some basic actions or skills the agent can perform Since different teams use different theories and models the platform must be able to work with all kinds of different agent architectures and cope with other user defined design issues The platform should also support encourage and promote the sharing of
11. estimated time since clock was stopped Reset ticks since last physical sensor data case Visual sensor Reset ticks since last visual sensor data end end 3 3 3 The sensors The purpose of the sensors are to parse the messages sent by the server and store the intermediate result for further processing either by some user defined unit or if the framework is used by the views The parser is an adapted version of the parser used in CMUnited 98 34 3 3 4 The sensor interface To hide the implementation of the sensors there is a sensor interface There are actually two different sensor interfaces the first called the basic sensor interface should be used when only using the basic system and the second called the sensor interface should be used when the information processing framework is used 30 The basic sensor interface contains methods for accessing all the raw data produced by the sensors If the information processing framework is used then the sensor interface is also responsible for telling the views when new data is available from the sensors How this is done is described in section 3 4 2 3 3 5 The actuators The actuators are designed to provide an interface between the agent and the server when it comes to acting It takes special command objects there is one class for each type of command available to the agent as input and converts them to something executable by the server The actuators are called by the control
12. extended to support more advanced functionality like making sure the arguments are legal according to the simulation and to assist the updating of the world model The available commands and their arguments are shown in Table 3 1 in section 3 3 5 3 3 The basic system The purpose of the basic system is to take care of the basic server commu nication and the timing as described in section 2 2 To accomplish this the task is to receive the messages from the server as soon as they arrive and send them to the sensors which parse the messages and store the raw data Concurrently the controller keeps track of the current game time and when it is time to send commands back to the server When the time comes the controller gets the next command from the actuators which are responsi ble for queuing the commands from the decision maker and sends it to the server for execution The type of data and how it flows through the basic system is shown in Figure 3 3 It is possible to use only the basic system without the information pro cessing and skills framework on top of it The only difference is that the user must use simpler versions of the sensor and actuator interfaces all the functionality of the basic system are still available to the user 26 The Sensor Interface The Actuator Interface parsed sensor data command objects processed raw sensor commands data F8
13. game and where you can step through the execution of the agent and the whole game The major problem for the interactive analysis is the support for the step feature within the current Soccer server The whole area of developing support for debugging of real time multi agent systems is very interesting A more realistic debug feature could be view servers used by RoboLog 23 and in some sense also the Headless chickens from Link ping 7 A view server is a server you connect a running agent to which then use the data sent to it by the agents to visualize the current state of the agent or use it for inspection or debugging purposes I hope to be able to continue working on RoboSoc and add some of the improvements discussed above The next event will be to create a team with RoboSoc that will compete in the Swedish and European championships in May 46 Appendix A The user s manual A 1 Introduction This is a preliminary version of the user s and the reference manual for the RoboSoc system for developing RoboCup agents It consists of four sections This first section contains a short introduction to the software and the following three sections described each of the three modules that make up the complete RoboSoc system For the latest information about the current release of RoboSoc look at the README file included in the latest release For information about how to install the software look in the INSTALL file included in the
14. game view which keeps track of the current game Each of the views mentioned above are described in the following sections In Figure 3 4 the dependency between the views are shown To overcome the problem with circular references some of the views use the previous cycle s data instead of the current data This is possible since each of the views keep historic data from the last cycles in case a need for them are found They could for example be used for calculating a trajectory of an object based on the previous observations of it The parameter view The parameter view keeps track of all the Soccer server parameters de scribed in section 2 1 1 and a few other important user defined parameters like the team name and whether the agent is the goalie or not It is also pos sible for the user to define her own parameters and let the parameter view or rather an extended version of the parameter view take care of them The values for the parameters can either be supplied from text files containing them or from the command line If no value for a parameter is supplied a default value is used It is possible for the user to change the parameters during run time even though it is not likely to be necessary Since the pa rameter view only contains parameters that are independent of the rest of the agent nothing is done in the event handlers 35 Player View Marker View H Agent View Ball View Command View Game Vi
15. library compiler directives h The constructors for the different units were described in the last sections The initialization of the agent is done in the CDecision class and since the base class CDecision does not send any init message to the server you have to write your own decision class How this is done see section A 3 9 A 4 2 The views The view is the basic component of the information processing in RoboSoc With information processing we mean transforming the data from the sen sors together with information from other views to more detailed or special ized information and storing it Each view should but does not have to be specialized in some area like the ball a certain skill or some other subject needed by the decision process or other parts of RoboSoc An important issue is to minimize the redundancy of information To make this possible each view can have links to other views containing already processed and stored information To solve the problem with dependencies each update contains a unique number so that the view knows if it has been updated already this also makes it possible to discover circular dependen cies How do they work The purpose of the views is to process and store information To make this possible they need to know when new information has arrived and they need to be able to access the raw data from the server and information stored in other views The data from the sensor interface is available f
16. processed and stored information To solve the problem with dependencies each update contains a unique number so that the view knows if it has been updated already this also makes it possible to discover circular dependen cies as described above The views are like the decision maker controlled by events The sen sor interface will generate events whenever new sensor data is available or something else occurred that the views should know about like a new cycle started All the events are described in Table 3 3 34 Table 3 3 The events the views needs to handle After a new cycle has started After the agent has been initiated After visual sensor data has been received UpdateAfterHear After aural sensor data has been received UpdateAfterSense After physical sensor data has been received UpdateAfterCommand After feedback from the actuators has been received Information processing To provide the basic information processing needed by most RoboCup agents RoboSoc comes with a set of views that will build up a world model consist ing of a model of the ball a model of the agent a model of the other players and a model of the markers on the soccer field Each of these models are put in its own view There are also three more views the parameter view which keeps track of the server parameters and optional client parameters the command view which is used to calculate the total effects of the actions executed by the agent and the
17. sec tions The initialization of the agent is done in the CDecision class and since the base class CDecision does not send any init message to the server you have to write your own decision class How this is done is described in section A 3 9 53 A 3 2 How to use the basic system with the framework The only difference is that you have to use the class CSensorInterface instead of CBasicSensorInterface when you start the sensor interface in Example A 1 You also need to start the view manager and use a different decision class but that is discussed in section A 4 A 3 3 The controller The constructor is CController CServerInterface const CSensors const CActuators const CDecision const const t uint i6 cycle length const t uinti6 recv step const t uinti6 sense step const t uinti16 visual step where CServerInterface CSensors CActuators CDecsion is pointers to the current server interface sensors actuators and decision ob jects The cycle length it the length of a cycle in milliseconds recv step is the interval between the server polls the sockets sense step the interval between sense body messages and visual step the interval between see messages recv step sense step and visual step can be found in server conf A 3 4 The sensors The constructor is CSensors CBasicSensorInterface where CBasicSensorInterface is a pointer to the sensor interface object A 3 5 The sensor interface The constructor is
18. side GetOurSide t side Get TheirSide t_uint GetUniformNumber t_ play_mode GetPlayMode t_uint GetOurScore t uint GetTheirScore string GetCoachMessage string GetPlayerMessage CCommand GetLast Command 0 0 56 object ayer void fn const player object Called by the controller when a new Called by the controller when a new Get the number of markers in the last see message Apply the function fn on each line Get the number of lines in the last see message Apply the function fn on each player Get number of players in the last see Return true if the ball was seen in the Return the distance to the ball Return the direction to the ball Return the distance change for the Return the direction change for the Get the last command sent to the server 0 Table A 8 Public methods in class CActuatorInterface Name Description Init CActuators Tells the interface what actuators to use called by the actuators bool AddPlan CCommand Add a command to the send queue A A The framework A 4 1 How to create an agent with the framework To create an agent using the RoboSoc framework you have to create a main function where you start the different parts In Example A 4 you can see how the code for a basic agent using the framework might look like The function rs_debug is used for debugging and is only needed if the com piler directive USE DEBUG is defined in
19. state The whole RoboSoc system implemented in C in the Solaris Unix environment is working and have been used in a course on Al progamming Contents Introduction EV Ehe Problem 9E REI dE ek ARE ede 1 2 Phe Solution 4 2 ercRe Xa Shoe eR A ROT ex SOR RU 1 3 The structure of the report ln Background 2 1 Phe Soccer Server e d ical e ede ee at a ee ias 2 1 1 The server parameters 0 4 2 1 2 Thesimulation 2 2 snl 2 1 3 The sensor inputs 2 1 4 The availablecommands 2 1 5 The command cycle 2 2 Educational aspects saaa a a 2 3 Related work eR e Ree gag Ro E D a dew ac og 2 3 1 What is an agent 2 3 2 Agent architectures aooo 2 3 8 Other RoboCup libraries RoboSoc S l OVERVIEW excu cae PEE AE Cad ewe pui 3 2 The library 4 3 s o wae yo mre Rp RR Eg 3 2 T Datatypes oo keg e ome ea 39 2 2 UTES 4 uui ae eae WA d re cds 3 23 Geometric objects leen 3 2 4 Game objects i sopa eee ee so o VUE Res aS 3 2 5 Command objects ln 3 9 The basie syst m s ciu lese dec Agee be ah cs i 3 3 1 The server interface 2l len 3 39 24 The controller 5e exeo Rp RR 3 3 3 Ehe sensots luce dece hoe FL C ode aso ipte Eg 3 3 4 The sensor interface a 3 3 5 Theactuators e a E e i E 3 3 6 The actuator interfa
20. the object seen pec is another player 2 1 3 The sensor inputs As mentioned before there are three types of sensor inputs providing visual aural and physical sensor data Each of the sensors and what type of data they send to the agent will be discussed in this section The visual sensors Each send step the agent receives visual sensor data In Table 2 2 the data given by the visual sensors are described Some of the data is only available if the agent is close enough to the object the quality of the data is also dependent on the distance to the object For a detailed description of the exact details see the Soccer server manual 8 The aural sensors The agent will receive data from the aural sensor when it is available Data is available when one of the players the coaches or the referee says something 11 Table 2 3 Data given by the physical sensor ViewQuality The current setting of the agent s ViewQuality affects the ps E amount and quality of the visual data sent to the agent ViewWidth The current setting of the agent s ViewWidth affects the ONY ammount ama quii of the ial data sont to he agent 8 The current effort of the agent AmountOfSpeed The amount of the agent s current speed vector and the distance to the speaker is less than audio_cut_dist The agent will hear all messages from itself the referee and the coaches but only one message from another teammate in two cycles If more than one message is
21. top shows the sign of the coordinates for each team and the two at the bottom shows the directions of each team coordinate system and the directions are on the other hand dependent on what side the team is playing on see Figure 2 3 The origo is in the center of the field negative x values are always found on the team s own side of the field and positive on the opposite side negative y values are found on the team s left part of the field and positive on the right when standing in origo facing the opponents goal Directions are also based on the side of the team with straight ahead 0 degrees from the origo towards the opponents goal Positive angles are to the right of the agent standing in origo facing the opponents goal and negative angles are to the left All movable objects on the field i e the players and the ball are treated as circles All distances and angles are to the center of the circles The referee The simulated referee controls the play mode of the game It decides when a team scores a goal when the ball is out of bounds and when to call an offside There are also rules which the simulated referee can not enforce since they concern the intention of the players Some of these rules are surrounding the ball not putting the ball into play and intentionally block ing the movement of other players Therefore a human referee also has the possibility to call free kicks Movements of objects At the end of e
22. turn towards the ball else dash with 5096 of the maximum power Limitations It does not take obstacles or the movement of the ball into account MoveTo Purpose To move to the absolute position X Y Arguments X and Y Preconditions The position of the agent is known Algorithm if allowed to use the move command 41 then Move X Y else if the direction to the point X Y is greater than 5 degrees then turn towards the point else dash with 5096 of the maximum power Limitations It does not take obstacles into account Score Purpose To kick the ball into the opponents goal Arguments None Preconditions The ball is kickable Algorithm Kick the ball as hard as possible towards the center of the opponents goal Limitations It does not take obstacles nor the agent s body into account when kicking TrackBall Purpose Follow the motion of the ball Arguments None Preconditions The direction to the ball is known Algorithm Turn towards the ball Limitations It does not take motion of the ball into account 3 5 Educational value As stated previously in this chapter the RoboSoc system have been used in the AI programming course given by the department of computer and information science at Link ping university For more information about the course look at its web page 1 The main goal of the course is for the students to create their own RoboCup teams and then compete against each other at the end Previous year
23. CBasicSensorInterface string my team name where my_team_name is the name of the team The string is used to infere which team a player belongs to A 3 6 The actuators The constructor is CActuators CActuatorInterface where CActuatorInterface is a pointer to the actuator interface object A 3 7 The actuator interface The constructor is CActuatorInterface A 3 8 The decision maker The constructor is CDecision CActuatorInterface const actuator interface CBasicSensorInterface const sensor interface where C ActuatorInterface is a pointer to the actuator interface object and CBasicSensorInterface is a pointer to the sensor interface object 54 Table A 6 The events available to the decision maker When it is time to send a command before it is too late DelayedActions When a sense_body indicates that more actions were per bin i AR formed than was sent the previous cycle Missing Actions When a sense_body indicates that less actions were per Pore formed than was sent the previous cycle A 3 9 How to do the decision making To add the decision making you have to the following Create a new class that inherits from CDecision In that class you have to start the main loop of the program for an example see Example A 2 Then you have to implement some of the event handlers that the controller will call The event handlers are described in Table A 6 For an example implementation of a basic agent look at the basic decision c
24. RoboSoc a System for Developing RoboCup Agents for Educational Use Fredrik Heintz frehe ida liu se March 23 2000 Abstract This report describes RoboSoc a system for developing RoboCup agents designed especially but not only for educational use RoboSoc is designed to be as general open and easy to use as possible and to encourage and simplify the modification extension and sharing of RoboCup agents and parts of them To do this I assumed four requirements from the user she wants the best possible data use as much time as possible for the decision making rather act on incomplete information than not act at all and she wants to manipulate the objects found in the soccer environment RoboSoc consists of three parts a library of basic objects and utilities used by the rest of the system a basic system handling the interactions with the soccer server and the timing of the agent and a framework for world modeling and decision support The framework defines three con cepts views predicates and skills The views are specialized information processing units responsible for a specific part of the world model like mod eling the ball or the agent controlled by events generated by the basic system The predicates can be used either by the decision maker or the skills to test the state of the world The skills are specialized short term planners which generate plans for what actions the agent should do in order to reach a desired goal
25. Unknown Goal Unknown Flag 50 Table A 4 Data types for RoboSoc concepts available in RoboSoc t side S Unknown A type for representing the sides S Left of the field S Right team T Unknown A type for representing the T Our Team teams T Their Team t play mode PM_Unknown A type for representing the pos M Before Kick Off sible game modes M Time Over M Play On M Drop Ball M Offside Kick M Our Offside Kick M Their Offside Kick M Half Time M Time Up M Extended Time M Kick Off M Our Kick Off M Their Kick Off M Kick In M Our Kick In M Their Kick In M Free Kick M Our Free Kick M Their Free Kick M Corner Kick M Our Corner Kick M Their Corner Kick M Goal Kick M Our Goal Kick M Their Goal Kick M Goal M_Our_Goal M_Their_Goal M Goalie Got Ball M Our Goalie Got Ball M Their Goalie Got Ball P P P P P P P P P P P P P P P P P P P P P P P P P P P P P P P 51 Table A 5 Named instatiations of templates available in RoboSoc They add the feature of unknown values a confidence factor and an obser vation time respectively They are implemented as templates but are only tested for basic C types and other modifiers instantiated with basic types The basic idea is to create a new class with an attribute to store a value of the type the template was instatitated with an
26. able margin there is no effect The one impor tant difference between dashes and kicks is how the kick power rate is figured Let dir diff be the absolute value of the angle of the ball rel ative to the direction the player s body is facing if the ball is directly ahead this would be 0 Let dist_ball be the distance from the center of the player to the center of the ball Then the kick power rate is figured as follows kick power rate X 1 180 kickable margin Basically this means that the most powerful kick can be done when the ball is directly in front of the player and very close to it and drops off as both distance and angle increase 13 25 x dir_diff 25 x dist_ball player size ball size Table 2 4 The syntax of the basic actions the agent can perform 8 catch Direction Direction minmoment maxmoment degrees default 180 180 change view Width Quality Width narrow normal wide Quality high low dash Power Power minpower maxpower default 100 100 kick Power Direction Power minpower maxpower default 100 100 Direction minmoment maxmoment degrees default 180 180 move X Y X 54 5 54 5 Y 34 34 move only works in before kick off mode and 5 seconds after a goal is scored say Message Message a string of at most say msg size characters default 512 turn Moment Moment minmoment maxmoment degrees default 180
27. ach time step the server updates the current state of all the objects in the world To quote the the Soccer server manual the movement of each object is calculated in the following manner 8 us lut t ut 5 aj aj accelerate p ups pi Py ED ugt ae move vitt vtt decay x uZ uitt decay speed af i 1 0 0 reset acceleration where pf 95 and vt vt are position and velocity of the object in timestep t decay is a decay parameter specified by ball decay or player_decay depending on what object is being updated a af is the acceleration of the object which is derived from the Power parameter in dash in case the object is a player or kick in case the object is the ball commands in the following manner ab ay Power x power rate x cos 6 sin 6 where power rate is either dash power rate or is calculated from kick power rate as described in section 2 1 4 and 6 is the di rection of the object in timestep t In the case of a player 6t is the direction the player is facing In the case of the ball the direction is Oball Oki dert Direction where OL and Decker are the directions of the ball and the kicking player at time and Direction is the second parameter of the kick command Collisions At the end of the simulation cycle if two objects overlap then a collision occurs and the objects are moved back until they do not overlap Then the velocit
28. ameterView SP port parameterView SP buffer size CMyDecision decision amp actuator interface amp sensor interface amp view manager CController controller amp server interface amp sensors amp actuators amp decision parameterView SP simulator step parameterView SP recv step parameterView SP sense body step parameterView 5SP send step controller Start return 0 66 Bibliography Al Programming course web page http www ida liu se TDDA14 March 2000 Minoru Asada and Hiroaki Kitano editors RoboCup 98 Robot Soccer World Cup I Springer Verlag Berlin 1999 M E Bratman Intentions Plans and Practical Reason Harvard University Press 1987 Hans Dieter Burkhard Markus Hannebauer and Jan Wendler AT Humboldt development practice and theory In Hiroaki Kitano edi tor RoboCup 97 Robot Soccer World Cup I pages 357 372 Springer Verlag Berlin 1998 Hans Dieter Burkhard Jan Wendler Pascal Gugenberger Kay Schr ter and Ralf K hnel AT Humboldt in RoboCup 98 In Minoru Asada and Hiroaki Kitano editors RoboCup 98 Robot Soccer World Cup I Springer Verlag Berlin 1999 Silvia Coradeschi and Jacek Malek How to make a challenging AI course enjoyable using the RoboCup soccer simulation system 1999 Silvia Coradeschi and Paul Scerri A User Oriented System for Develop ing Behavior Based Agents In Minoru Asada and Hiroaki Kitano edi tors RoboC
29. ary ones for creating a functional RoboCup agent But adding other frameworks is possible I can think of at least four other useful frameworks a role framework a communication framework a coach framework and a machine learning framework The role framework is probably the most wanted one since most teams use some sort of roles to divide the tasks between the agents today Otherwise it is interesting to wait and see where the development of RoboCup is going before taking a concept on as one of general interest It is also possible to create completely separate frameworks that do not use the current ones for example support for genetic programming 45 Development tools and debugging One direction of development which could be more fruitful than to add more frameworks is to provide a set of development tools that can help the developers with the creation of the agents Tools that are of interest are for example the layered disclosure tool provided by CMUnited 32 and tools that graphically show what the agent currently believes in Regarding debugging I see two possible types of debugging The first type is the retrospective analysis like the logplayer from CMU where you first store data from a game and then watch the game and inspect the actions of the agents afterwards The second type is the interactive analysis and maybe even pro active analysis The interactive analysis is when the current state of the agent can be inspected during a
30. asic system it is possible to use the framework designed to support and modularize the world modeling and the support for decision making discussed in section 2 2 The purpose is to provide building blocks for the user who either can use the existing blocks as they are or extend them with new functionality It is also possible to create completely new blocks There are three kinds of building blocks views skills and predicates They each encapsulate an important concept used in the information processing and the decision making A view is a way of looking at and extracting information focusing on some special property or object from a collection of data which is dependent either on the current time or on the history of the agent A skill is a complex action combined of several primitive actions that will take an agent towards a certain goal state A predicate works like a predicate defined in a three valued logic It is used to answer yes or no questions about the state of the world They are supposed to be used as conditions for rules used in either the decision maker or in the skills This section discusses the only concrete unit in the framework the view manager and the three different building blocks or concepts that it tries to capture The following sections describes instances of the different concepts the predefined building blocks available in more detail and how they are used to supply an agent with a basic world model 3 4 1 The v
31. ayful fashion create and test different Al strategies It is also easy to motivate them by providing competitions where they can test their teams against each other RoboCup is divided into different leagues real robots leagues and a sim ulated software agents league This report only consider teams of simulated software agents The simulation is done by a special server called the Soccer server which is described in section 2 1 Anybody can download the neces sary software and start developing their own soccer agents 28 Descriptions of the teams that participated in the RoboCup world cup competitions held 1997 1998 and 1999 can be found in 2 13 35 Many of them can also be downloaded from 22 and tested More information about RoboCup can be found on web site of the RoboCup organization 21 1 1 1 The Problem A major problem when constructing RoboCup agents for the simulation league is to implement the basic functionality of the agent This is a general problem when developing multi agent systems Wooldridge and Jennings write one of the greatest obstacles in the way of the wider use of agent technology is that there are no widely used software platforms for developing multi agent system 37 This is due to a number of problems for instance interacting with the Soccer server and developing the basic skills of a soccer player like finding the ball move to the ball and kick it in the desired direction The develop ment of
32. ber of lines seen in the last see message GetClosestSeenLine A line object Get the closest line last seen Get Line line hist A line object Get the line line as seen hist cycles ago GetClosestSeenMarker A marker object Get the closest marker seen in the Get Marker marker hist A marker object Get the marker marker as seen hist BENE UM E ist GetClosestMarkerTo point hist A marker object Get the closest marker to point as seen hist cycles ago int hist GetLine L R T B hist A line object Get the line object as seen hist cycles ago GetClosestGoalTo point hist A marker object Get the closest goal to point as seen cu quu Get TheirGoal hist A marker object Get the opponents goal as seen hist Get OurGoal hist A marker object Get the agent s team s goal as seen ago GetF lag Name hist A marker object Get the marker object as seen hist D mea O S Seon The marker view The marker view stores information about all the markers on the field It will give absolute position relative distance and direction to each marker In Table 3 6 all the methods are described The agent view The agent view keeps track of the agents status i e its absolute position stamina effort recover absolute face direction body and neck directions In Table 3 7 all the methods are described 37 Table 3 7 Functionality provided by the agent view Return type ago Get AgentPosition hist A point Get an estimati
33. ccer envi ronment is not very hard to represent the main problem is to use the 16 noisy and incomplete sensor data to draw correct conclusions about the objects in the world There is also the problem of making predic tions about future states of the server but that is not so important as having a model of the current world which is as complete consistent and updated as possible 4 Support for decision making The fourth problem is based on the current world model to decide on what action to do This is of course the job of the agent designer not the library designer but to assist the agent programmer it is very good to have some support for the decision making like notifications when new information arrive from the server and when a new cycle is started Since most decision makers used in RoboCup use some sorts of rules it is practical to have some predefined predicates like is the ball on our half are we playing on the left side of the field and so on which can be used to trigger rules With decision making I mean the very general meaning of the concept from a RoboCup point of view to select some actions to do based on the current world model It is also very convenient to have some higher level of abstraction when it comes to actions Therefore intermediate skills are needed which are based on the basic actions the agent can do For example turn to absolute direction D or kick the ball to position X Y 2 3 Related work Th
34. ce ooo aaa iii 3 3 7 The decision maker l l 3 4 The framework 2 ene 3 4 1 The view manager llle 23 4 27 The Views o eeraa ea Rp RC als 34 3 The predicates ge gen ee e Ue es 3 4 4 Ehe skills zoe Ae 3 5 Educational value a 4 8 te teas apg ae RR Y x ROUX USD S Conclusions AT Summary sieri inida Bg BO Pa A AE Eas g 42 Future work 2l 3er e ive x oe oe ee ee ee The user s manual A Introduction 2 0 a pa E E a e ee A 1 1 Overview of the software package A 1 2 Generalparts 2l A2 Phe library ive na ume teen ae E ee A 2 How to use the library sen A 2 0 Compiler directives 22e AO2 3 Datatypes 23 2 doe ee i Red CY Row A 3 The basic system eee A 3 1 How to create an agent l l A 3 2 How to use the basic system with the framework A 3 8 Thecontroler len A 344 Thesensots 2 9n 9 Au oe AS IS A 3 5 The sensor interface llle A 3 6 The actuators once ee AS ee ede LA A 3 7 The actuator interface llle A 3 8 The decision maker lll nn A 3 9 How to do the decision making A A The framework 0000202 eee A 4 1 How to create an agent with the framework AAZ TDhe views eh a ate Se es ate Gom Re REGE a A 4 3 How to use the predicates A 4 4 How to create your own predicates A 4 5 How to ex
35. ch cycle but it does not give the agent any hints of when it has to send the commands or when it is getting too late to send a command the current cycle Another serious problem from my point of view is the fact that it is very hard to replace parts of the RoboLog code with new without having to change other irrelevant parts of the code mostly because of the design of the world model This together with the complexity of the library mainly due to lack of a clear general architecture makes it quite hard to understand and exchanging parts of the code to try different approaches is cumbersome Finally the threshold before productivity is not low enough for it to be useful in an educational setting The benefits from the library is the possibility to program agents in Prolog and the large amount of features available when the initial obstacles are overcome More information about RoboLog can be found on their web site 23 CMUnited CMUnited is the most successful team in the history of RoboCup so far it has won the last two competitions 1998 in Paris and 1999 in Stockholm It is developed by Peter Stone Manuela Veloso and Patrick Riley from Carnegie 20 Mellon University USA The team and its development is described in a series of articles among them 31 33 34 and in Peter Stone s PhD thesis 30 To help the rest of the RoboCup community they have released parts of their team The released code does the basic server communicati
36. d extra attributes for the feature The class then implements all the basic arithmetic operators and also some special operators for that modifier There is also a method called GetValue which returns the value of the variable but without the modification Therefore it will discard information stored in the variable In some cases it might not even work if you try to take the value of an unknown variable then it will throw an Unknown ValueException The special operators are described below With unknown The with unknown modifier makes it possible for a value to be unknown The special methods available are MakeUnknown which makes the variable unknown and return a reference to that variable and IsUnknown which returns true if the value is unknown or false other wise 52 With confidence The with confidence modifier makes it possible to add a confidence factor to the value The special methods available are GetConfidence which returns the current confidence value and SetConfi dence CConfidence which sets the current confidence value There is a compiler directive called PESSIMISTIC OBSERVATIONS which controlls how confidence values are updated when doing arithmetic with modified values If it is set then the lowest confidence value is taken other wise the highest With observation time The with observation time modifier makes it possible to add a timestamp to the value The special methods are GetObservationTime
37. different parts of the agent like skills and information processing algorithms 1 2 The Solution Since this is a very open problem I had to make some assumptions The assumptions I made about what the user wants for their agents are 1 accurate complete and consistent data in that order of importance 2 use as much time as possible for the decision making 3 rather act on incomplete information than not act at all 4 soccer objects the agent can manipulate like the ball and the players Since I want to make as few assumptions as possible I do not make any unnecessary estimations since they can be done in many different ways The few estimations I do are well documented and easy to change if the user finds a need for it My solution to this problem is a system called RoboSoc It includes three parts a library of utility classes a basic system for taking care of the timing and the interaction with the server and a framework for information processing and decision support The basic system provides a well defined work cycle by which the agent acts It takes care of the communication with the Soccer server and it does the basic information processing mainly parsing the messages sent by the Soccer server but also some feed back from the actuators The information processing framework consists of a collection of specialized units that process the raw data obtained from the basic system and from other information processing units into i
38. do their own thing but do not want to do the parsing and need a better interface to the communication with the server Another problem is that the libsclient is no longer supported and it is not up to date with the current server RoboLog RoboLog is an ECLiPSe Prolog library built on top of a C library devel oped by Oliver Obst at the university of Koblenz Landau The C library was developed to make the server functions accessible from Prolog and to build a database with the data from the soccer server The interface to the database is somewhat awkward and not very elegant since all the data is stored in one large structure A more serious problem with this world model is that no inference about the objects are made from old knowledge instead the new sensor data is simply stored in the database and the client programmer has to do all the inference Apart from the world model RoboLog is an advanced library with many nice features and algorithms for computing the position of the agent do geometric calculations and support the development of advanced skills Even though it does not have any abstractions of the basic server commands like turn to absolute direction D or kick the ball to the point X Y It does take care of most of the low level tasks like sending and receiving data from the server and the timing of the commands but in a rather simple manner It makes sure that the agent will not send more than one catch dash move or turn ea
39. e pages 24 29 San Francisco CA 1997 Morgan Kaufmann S C Kleene Introduction to Metamathematics Princeton N J 1952 Pattie Maes The agent network architecture ANA SIGART Bulletin 2 4 115 120 1991 Jorg P Miller The Design of Autonomous Agents A Layered Ap proach Springer Verlag Heidelberg 1996 Jorg P Miiller The right agent architecture to do the right thing In J rg P M ller Munindar P Singh and Anand S Rao editors Intelligent Agents V Proceedings of the Fifth International Workshop on Agent Theories Architectures and Languages ATAL 98 Springer Verlag Heidelberg 1999 A Newell and H A Simon Computer science as empirical enquiry Symbols and search Communications of the ACM 19 3 113 126 1976 Itsuki Noda Hitoshi Matsubara Kazuo Hiraki and Ian Frank Soccer server A tool for research on multiagent systems Applied Artificial Intelligence 12 233 250 1998 The Robot World Cup Initiative web page http www robocup org March 2000 The RoboCup Simulator Team Repository http medialab di unipi it Project Robocup pub March 2000 68 23 24 32 33 35 251 126 127 og 291 130 31 i Taie The RoboLog Soccer Agent Libraries web page http www uni koblenz de ag ki ROBOCUP ROBOLOG March 2000 The RoboSoc web site http www ida liu se frehe RoboCup RoboSoc March 2000 Stuart Rus
40. e RAPs 10 or CONTAP 9 can also be incorporated by the user without too much trouble There is currently only very limited support for adaptation consisting of the history of the world from the last cycles provided by the information processing framework which can be used for machine learning or other adaptation mechanisms The only thing that is not explicitly supported is the interactive capability but at the same time there are no major obstacles if the user wants to add communication capabilities to their agents The contributions of this work is mainly an architecture and an infras tructure which makes it easier to develop RoboCup agents that is actually implemented and working It takes care of all the low level details and let the user focus on the more interesting Al parts of the agent It also gives the users the possibility to share different parts of their agents by providing a framework for the most common concepts used in RoboCup Since more and more educational institutes are starting to use RoboCup as part of their curriculum I think this system can play an important role in the promotion of RoboCup and AI by making it more accessible to everyone 44 4 2 Future work There is a vast number of things that would improve RoboSoc The following paragraphs each discuss one area where improvements are possible Improve the existing basic system The most needed improvement to the basic system is to make it a distributed app
41. e agent No moment minneckmoment maxneckmoment moment degrees to the right Table 3 2 The events available to the decision maker BeforeTick AfterTick BeforeSensorData AfterSensorData Before a tick of the internal clock After a tick of the internal clock Before receiving sensor data After receiving sensor data ActuatorSensorData After receiving actuator feedback AuralSensorData After receiving aural sensor data PhysicalSensorData After receiving physical sensor data VisualSensorData After receiving visual sensor data Init After receiving init data SensorError After sensor errors NewCycle When a new cycle has started EstimatedNewCycle When the controller estimates that a new cycle has started CommandWarning When it is time to send a command before it is too late DelayedActions When a sense body indicates that more actions were per formed than was sent the previous cycle When a sense_body indicates that less actions were per formed than was sent the previous cycle MissingActions 32 It should be possible to use any type of agent architecture with this decision maker whether it is a reactive deliberative or hybrid architecture The deliberate agent can use only the main loop to guide the behavior of the agent the reactive agent can use only the events and finally the hybrid agent can use both the events and the main loop to control the behavior of the agent 3 4 The framework On top of the b
42. e end of the step the server takes all ac tion commands received and applies them to the objects in the field using the current position and velocity information to calculate a new position and velocity for each object The soccer field The soccer field and all objects in it are 2 dimensional so there is no notion of height of any object The field is 105 meters long and 68 meters wide The goals are 14 02 meters wide about twice the size of a normal soccer goal since it was too difficult to score otherwise The field and the different markers and lines used by the agents to navigate are shown in Figure 2 2 The players are bound by the outer flags which are 5 meters outside the playing field The markers on the field have a known absolute position which is independent of the side the team is currently playing on The 105m e e e e e e e e e e e 0 34 e e e e e 16 5m 9 e e e re eq 1402m e 52 5 0 e Left team 0 0 Right team 40 32m 4 525 0 amp 68m e eo e e 90 90 e 9 4 180 0 0 4 180 6 e 10m e 90 90 e 6 e U 10m 0 34 e e e e e e e e e e e Figure 2 3 The coordinate system for the two teams The coordinates are given in the form z y if two signs are shown then the first sign is for the left team and the second sign is for the right team The two diagrams at the
43. ecially not the world model which is usually very specific for each agent design Since each game object represents an object at a certain moment in time and they do not have any connection to the rest of the world model they can 25 also be used by a prediction system to create multiple possible versions of the object This is not possible with the objects used by CMUnited If the user desires she can extend the basic game objects by connecting them to the world model of the agent The benefit is that you can make sure that the state of the object is valid according to the world model and the simulation for example by checking the speed and see if it is within the limits of the object but also to make special world or simulation dependent update methods directly in the game object There is also a template for representing a collection of game objects called ObjectCollection used in the views to represent the history on an object It can be instantiated with any class derived from the CGameObject class 3 2 5 Command objects The command objects are designed to represent the commands that can be sent to the Soccer server The version of the command objects included in the library are very simple since they are completely separated from the world model of the agent for the same reason as the game objects This unfortunately implies that no checks can be made about the validity of the commands sent to the server But the objects can be
44. eported by the server does not change T his makes it possible to use the same internal work cycle even when the game is stopped which would not be possible otherwise Whether the game clock is stopped or not is de pending on what state the game is in The state of the game is only changed after a call of the referee therefore the starting and stopping of the clock is done in the sensor interface after the sensors have received a message from the referee Algorithm for updating the agent after a tick of the internal clock The internal clock is used to time the acting of the agent and to predict if the agent is missing messages from the server Therefore the algorithm has to check if a new cycle ought to have begun and in that case update the state of the agent with this prediction but since the controller does not want to assume that a new cycle actually has begun it will generate an event which allows the decision maker to react to the situation The decision maker has two possibilities either to assume that a new cycle has begun or to assume that a new cycle has not begun When this is done the algorithm will check to see if a periodic command is waiting to be sent Otherwise it checks if the last time for sending a periodic command this cycle is approaching then it generates an event for the decision maker to handle The reason for this 28 is the second assumption about the user given above Then the algorithm checks if there are some ot
45. erSensorData event end Algorithm for updating the current time after receiving the cur rent server time Assumptions made about the current time of the agent 29 e If the clock is not stopped then the time received with the sensor data is assumed to be correct e If the agent thinks the clock is stopped but the time received from the sensor data is different from the current time then assume the agent is wrong and the clock is started again e If the clock is stopped then the agent assumes a new cycle starts with a sense body message and therefore updates the time e A new cycle starts when the current time is greater than the time of the last cycle Based on the assumptions above the following algorithm was developed the only addition is the number of ticks of the internal clock that are counted from a visual or a physical sensor event They are used to predict how much time is remaining before the next visual sensor data and how much time the agent has left in this cycle Algorithm 3 3 Update the current time based on current server time begin if not clock stopped then Reset estimated time else if last known server not equal new server time then Assume the clock is running Update last known server time with the new server time Update the current time with the new server time switch last sensor type begin case Physical sensor if clock is stopped and last sense time is equal to current server time then Update
46. etAngle void CBasicDecision OnInit actuatorInterface gt AddPlan new CMoveCommand CPoint 10 10 void CBasicDecision nCommandWarning O 1 Decide 65 Example A 4 An example of a basic agent using the framework int main int argc char argv 1 CViewManager view manager CParameterView parameterView new CParameterView PARAMETER VIEW ID view manager argc argv view manager AddView parameterView true CActuatorInterface actuator interface CActuators actuators amp actuator interface CSensorInterface sensor interface parameterView 5CP team name amp view manager CSensors sensors amp sensor interface CGameView gameView new CGameView GAME VIEW ID amp view manager view manager AddView gameView true CCommandView commandView new CCommandView COMMAND VIEW ID amp view manager view manager AddView commandView true CMarkerView markerView new CMarkerView MARKER VIEW ID amp view manager parameterView view manager AddView markerView true CAgentView agentView new CAgentView AGENT VIEW ID amp view manager parameterView view manager AddView agentView true CBallView ballView new CBallView BALL VIEW ID amp view manager view manager AddView ballView true CPlayerView playerView new CPlayerView PLAYER VIEW ID amp view manager view manager AddView playerView true CServerInterface server interface parameterView 5SP host par
47. ew Parameter View Sensor Interface Toce Depend on the previous cycle s information Depend on the current cycle s information Figure 3 4 Graph showing the dependency between the views The command view The command view calculates the effect of the commands executed the dur ing current cycle It will give the total movement vector of the agent and the ball for the current cycle It will also give the total turn angle of the agent and its neck The available methods are described in Table 3 4 The game view The game view stores information about the status of the game like the play mode the score and the current time The complete set of methods is shown in Table 3 5 Table 3 4 Functionality provided by the command view Return type Get Agent Movement Get the previous cycle s total agent movement vector Get BallMovement Get the previous cycle s total ball movement vector Get ActualTurn Get the previous cycle s total agent turn body angle 36 Table 3 5 Functionality provided by the game view Return type GetCurrent Time A time object Get the current time according to the agent GetPlayMode hist Get the play mode for hist cycles ago GetOurSide Get the side of the field the agent is playing on Get TheirSide Get the side of the field the agent is not playing on Table 3 6 Functionality provided by the marker view Return type Get NumberOfSeenLines A number Get the num
48. fort decreases with a minimum value given by effort min and if stamina gets high enough then effort increases with a maximum of effort max Specifically min effort max effort effort inc if stamina gt effort int thr X stamina_max max effort min effort effort dec if stamina X effort dec thr X stamina_max effort effort otherwise e Recovery This is similar to effort except that recovery never increases max recover min recovery recover dec if stamina lt recover dec thr X stamina max recovery y recovery otherwise The order in which the different variables are changed are im portant First the stamina value is decreased if there is a dash Next the recovery value is changed then the effort Finally stamina is recovered Figure 2 4 visualize the different thresholds in the stamina model 10 stamina max Stamina effort inc thr effort dec thr recover dec thr stamina min Figure 2 4 The different thresholds used in the stamina model Table 2 2 Data given by the visual sensor BodyDir The body direction relative the facing direction of the agent UP nip ite jet scene another plager n HeadDir The head direction relative the facing direction of the agent BPN only 1 the objet seen is another player n TeamName The name of the team of the seen player only if the object T seen is another player UniformNumber The uniform number of the seen player only if
49. from the Soccer server manual 8 where more information can be obtained 2 1 1 The server parameters The Soccer server has many different parameters which control how the simulation works Most of them can be changed by editing a configuration 5 Table 2 1 The Soccer server parameters used in the report with their default values The table is adapted from the parameter table in the Soccer server manual 8 Parameter name Default Description value simulator step Length of each simulation cycle in milliseconds sense body step 100 Length of interval in milliseconds between sense body infor mations 5 send step 150 Length of interval in milliseconds between sending visual RH information to a player in the standard view mode recv step 10 Length of interval between server polling sockets to clients poe milliseconds visible angle 90 Angle of the view cone of a player in the standard view mode visible distance 3 Maximum distance to an object out of the view cone a player can see Decay rate of speed of the ball 1 no decay 0 all decay ball speed max 2 7 Maximum speed of the ball during a simulation cycle me ter cycle player_size The radius of a player meter player_decay Decay rate of speed of a player 1 no decay 0 all decay player_rand Amount of noise added in players movements and turns player_speed_max Maximum speed of a player during a simulation cycle me ter cycle stam
50. from the skill you want to extend and then use its Applicable Persistent and GeneratePlan methods to create the new skill or inherit from some other skill or CSkill and create a local instance of the skill you want to extend and calls its Applicable Persistent and GeneratePlan methods to implement the class 62 Example A 1 A basic agent using only the basic system include CController h include CActuatorInterface h include CBasicSensorInterface h include CActuators h include CSensors h include CDecision h extern void rs debug const std string amp str 1 std cerr lt lt Error lt lt str lt lt std endl int main int argc char argv std cout lt lt Starting actuator interface n CActuatorInterface actuator_interface std cout lt lt Starting actuator n CActuators actuators amp actuator interface std cout lt lt Starting sensor interface n CBasicSensorInterface sensor_interface RoboSoc std cout lt lt Starting sensors n CSensors sensors amp sensor interface Lu std cout lt lt Starting server interface n CServerInterface server_interface localhost 6000 2048 std cout lt lt Starting decision n CDecision decision amp actuator interface amp sensor interface std cout lt lt Starting controller Mn CController controller amp server interface amp sensors kactuators amp decision 100
51. g 90 Minimum value of Moment in turn_neck commands Maximum length of a message a player can say flag t 150 flag t 1 30 flagt110 flagtO flag tr 10 flag t r 30 flag t r 50 e e e e e e e e e e e flag t 140 flag t 120 flag t r 20 flag t r 40 e e flag 1t flag ct flagrt flag1130 flag 11 line pe flag c0 flag r t 30 line 1 line r oS flag1t20 e flag p 1 t flag p rt j4 flag r t 20 flag 1t 10 flag rt 10 r9 flagg10 flagig rt 8 flag c flag10 goall flag plc e flagprce goal r flag r 0 e flag g1b flag g r b amp flag 1b 10 flag r b 10 flag 1b 20 e flag p 1b flag p r bg9 31 flag r b 20 flag 1 b 30 line b flag r b 30 flag 1 b E flag c b flag r b e e e flag b 140 flag b 120 flag b r 20 flag b r 40 e e e e e e e e e e e flag b 150 flag b 130 flag b110 flagbO flagbr10 flag b r 30 flag b r 50 Physical boundary Figure 2 2 The markers and the lines in the simulation file and load it with the server The parameters mentioned in the report are described in Table 2 1 For a complete list see the Soccer server manual 8 2 1 2 The simulation Since the simulation is discrete all moves within one time step occur si multaneously at the end of the step The length of a step is set by the simulator_step parameter At th
52. her commands waiting to be sent Algorithm 3 1 Update the agent after a tick of the internal clock begin Generate a BeforeTick event Update counters if ticks gt ticks between cycles then begin Assume new cycle Update current time end if assumed new cycle then begin Generate an EstimatedNewCycle event if the decision maker forced a new cycle then Generate a NewCycle event end if periodic actions waiting and current time gt last action time then Send next periodic action else begin if ticks gt ticks before command warning then Generate a Command Warning event if immediate actions waiting then Send next immediate action end Generate an AfterTick event end Algorithm for updating the agent when receiving sensor data When sensor data is received it is first parsed by the sensors then a suitable event is generated based on what type of sensor was used in parsing the message Algorithm 3 2 Update the agent when receiving sensor data begin Generate a BeforeSensorData event while sensor data waiting to be received do begin Receive the sensor data Let the sensors analyze the sensor data received switch last sensor type begin case Physical sensor Generate a PhysicalSensorData event case Visual sensor Generate a VisualSensorData event case Aural sensor Generate an AuralSensorData event case Init sensor Generate an Init event case Sensor error Generate a SensorError event end end Generate an Aft
53. ic constant for this value can usually be found in view data h The GetView method will return either a NULL pointer if the view doesn t exist or a pointer to the object if it exist The problem is that the pointer is of the type CView and not the correct view pointer you need Therefore you have to use the static cast template to convert the pointer from one type to another Eg to convert v ptr from a CView to a CBallView you do v ptr static cast lt CBallViewx gt v_ptr How do I create a view object First of all create an instance of the view you need Then you need to store the view in the view manager This is done by calling the view manager s AddView method with the address of the view and a persistence flag If the persistence flag is set the view will not be destroyed when no one is using the view otherwise it will be destroyed when no one is using it With no one using the view I mean that there are no pointers to the view active in the system An active pointer to a view is given when you ask the view manager for a view with the GetView method and they are deactivated with a call to the ReleaseView method The reason for storing the viewe in the view manager is to make them available to the rest of the system How do I extend them The basic rule is to create new views which use information from the view you want to extend instead of changing the code of the existing view If you really would like to extend an e
54. ies are multiplied by 0 1 Note that it is possible for the ball to go through a player as long as the ball and the player never overlap at the end of a cycle The stamina model To simulate the endurance of the players each player has a limited amount of stamina The stamina is used when the player dashes and is regained a little each cycle The maximum Power the player can dash with and the amount of stamina regained each cycle is depending on the three variables stamina effort and recovery Where stamina is the total amount of energy the agent has left effort determines how effective the agent s dashes are and recovery controls how much stamina is regained each cycle According to the Soccer server manual the stamina model works as follows When the client dashes its stamina is updated like this If Power 0 EffectivePower effort x min Power stamina stamina max stamina Power stamina min If Power 0 PowerUsed min 2 x Power stamina 2 EffectivePower effort x PowerUsed stamina max stamina 2 x PowerUsed stamina min The Power of the dash actually executed is EffectivePower Every cycle whether the client dashes or not the three variables are updated in the following way e Stamina Stamina increases slightly every cycle When recovery decreases less stamina is recovered stamina min stamina max stamina recovery X stamina inc max e Effort The basic idea is that if stamina gets low ef
55. iew manager The purpose of the view manager is to store and keep track of all the views in the system and to make sure they are updated when new information arrives The view manager is not responsible for the creation of the views This is done by the unit where they are needed when they are needed There are two ideas behind this The first idea is that it should be possible to have hundreds of views in the system but they should only use resources when they are actually used and it should not be necessary to know this at compile time The second idea is that the actual views in the system do not need to be known by the framework or the basic system they only need to be known to the users of the views To support the first idea the view manager keeps a reference counter for each view it stores When a view is needed a request for the pointer to this view is sent to the view manager 33 If the view is stored then the pointer is returned and the reference count is increased If the view is not stored then a null pointer is returned and the user has to create the view and send it to the view manager which stores it and sets the reference counter to 1 When a view is no longer needed a release view message is sent to the view manager which will decrease the reference counter If the reference counter reaches zero then the view is removed from the system unless the persistent flag is set The persistent flag is used to prevent a view from be
56. ina min 00 Minimum stamina ofa player stamina max stamina_inc_max 35 0 Maximum amount of stamina that a player gains in a simu lation cycle recover dec thr Decrement threshold for players recovery recover dec 0 002 Decrement step for a player s recovery recover min Minimum recovery of a player effort dec thr 0 3 Decrement threshold for a player s effort capacity 0 w o a C o d e OU e e e eoo elc or effort_inc_thr effort_dec 0 05 Decrement step for a player s effort capacity Increment threshold for a player s effort capacity effort min 3 5 effort_inc Increment step for a player s effort capacity 0 6 Minimum value for a player s effort capacity 1 0 kickable_margin 0 7 The maximum distance the ball can be from a player and still Se ee GENES ee ee ee ey kick_power_rate 0 016 Rate by which the Power argument in kick commands is cee uper rcu dash power rate 0 006 Rate by which the Power argument in dash commands is po age dc EM minpower i i maxmoment 180 Maximum value of Moment in turn and Direction in kick bonds ARS Lb bau minmoment 180 Minimum value of Moment in turn and Direction in kick pue uec Wome ier o Minimum value of Power in dash and kick commands 3 3 effort max 1 0 Maximum value for a player s effort capacity 3 maxneckang 90 Maximum value of Moment in turn_neck commands minneckan
57. ing removed from the system every time its reference counter reach zero The reason for this feature is if a view which continuously keeps track of historic data is not used continuously or if you have a view which takes a lot of resources when it is started but not very much resource when it is running When the view manager gets a request to update the views after new sensor data have been received it will go through the current list of views in its storage and send an update message to each of them Since each view has the possibility to request that another view should update itself a counter is used to prevent the view to be updated twice with the same information It is also possible for the views to detect circular dependencies which can not be handled with the current design 3 4 2 The views The view is the basic component of the information processing in RoboSoc With information processing I mean transforming the data from the sensors together with information from other views to more detailed or specialized information and storing it This information is used to build the agent s world model Each view should but does not have to be specialized in some area like modeling the ball a certain skill or some other subject needed by the decision maker a skill or another part of the agent An important issue is to minimize the redundancy of information To make this possible each view can have links to other views containing already
58. is section describes some of the related work to this report The first section will discuss the concept of an agent The second section will define what an agent architecture is and what kind of architectures exist The third and final section describes some of the RoboCup libraries available today 2 3 1 What is an agent The are many answers to that question almost as many as there are agent researchers The definition that is most appropriate for this work is the definition of Russel and Norvig they define an agent as anything that can be viewed as perceiving its environment through sensors and acting upon that environment through effectors 26 Even though it is a very simple definition of an agent compared to most other definitions see 11 for a collection of agent definitions It is general enough to cover most agents used in RoboCup since all RoboCup agents decide upon sensor information from the soccer server what effectors to use by sending commands back to the server 17 2 3 2 Agent architectures Since there are many definitions of what an agent is there are also many definitions of what an agent architecture is For example Pattie Maes defines an agent architecture as A particular methodology for building agents It specifies how the agent can be decomposed into the construction of a set of component modules and how these modules should be made to interact The total set of modules and their interactions has
59. lass in Example A 3 A 3 10 How to get the sensor information To get the sensor information you have to call the methods in CBasicSen sorInterface They are described in Table A 7 Since all the values in the basic sensor interface is reset at the begining of a cycle all the data given by is from the current cycle A 3 11 How to send commands to the server To send a command to the server you have to create a new pointer to the wanted command object class for example new CDashCommand 75 to cre ate a new command dash 75 then send it to server by using the methods available in the actuator interface described in Table A 8 55 Table A 7 Public methods in class CBasicSensorInterface NewCycle ForceNewCycle CTime GetCurrentTime ForEachMarker void fn const marker t_uint GetNumberOfSeenMarkers ForEachLine void fn const line t_uint GetNumberOfSeenLines object ForEachPl t uint GetNumberOfSeenPlayers bool SawBall t float unknown GetBallDistance CAngleDegUnknown GetBallDirection t float 0 unknown GetBallDistanceChange t float unknown GetBallDirectionChange t_view_quality GetViewQuality t view width GetViewWidth t float GetStamina tfloat GetEffort t float GetSpeed CAngleDeg GetHeadDirection t uint GetKicks t_uint GetDashes t uint Get Turns t_uint GetSays t uint Get TurnNecks string amp GetOurTeamName string amp Get Their TeamName t
60. le AT Humboldt 4 5 Hybrid reactive deliberative architectures Since most agents need to be both reactive and able to use deliberation in their decision processes the hybrid layered architecure has emerged It usually combines reactive and deliberative layers in some way to create a hybrid agent architecture 17 The major problem is how to combine the different layers into a single unit Wooldridge and Jennings even argue that 18 hybrid architectures are very ad hoc since it is not clear how to reason about them and what the underlying theory is 36 An example of hybrid architecture used in RoboCup is the architecture used by FCFoo 12 Interacting architectures Architectures which mainly deal with coordination and cooperation among distributed intelligent agents are called interacting architectures 18 No example of interacting architectures have be found in the RoboCup domain 2 3 3 Other RoboCup libraries Previous attempts to build RoboCup libraries have usually been a team releasing parts of their code for others to use The problem with these releases are usually that they come with little or no documentation and the code is tightly connected to the structure of the team Therefore there is a need for a well documented well structured and generic library for people to use The structure of the library should help and encourage different developers to share their code The benefit is that new teams do not have to sta
61. ler when it is time to send a command The commands available are described in Table 3 1 Since there are two basic types of commands those that can be exe cuted immediately immediate commands and those that only one can be executed every cycle periodic commands the actuators maintain two sep arate queues one for each type of command 3 3 6 The actuator interface To hide the implementation of the actuators in the same way as with the sensors there is an actuator interface the agent should use to send commands to the actuators 3 3 7 The decision maker The decision maker is the unit that implements the decision making process of the agent This is the unit the agent designer will extend to program the behavior of the agent Since the second assumption about the user is that she wants to use as much time as possible to do the decision making the main loop of the agent is inside the decision maker To help the user the controller generates events when important things happen like new sensor data arrives or when it is urgent to decide on a command for this cycle because of the third assumption about the user before it is too late The decision maker also has limited possibilities to give feedback to the controller by changing some of its parameters like how late in the cycle the decision maker wants the command warning event The events the decision maker has to handle are described in Table 3 2 Since the events are generated f
62. library The RoboSoc library is the foundation for the rest of the system It pro vides the necessary types classes and utilities needed by the other parts of the system There are classes for representing geometric objects there are classes for representing the objects found in the soccer environment and there is a class for each of the commands that can be sent to the soccer server 3 2 1 Data types To make the system more machine independent there is a set of basic types for representing integers and floating point numbers They are named after the type they represent and the number of bits they can store There are also types defined to represent simple objects in the RoboCup simulation like a command name a marker name and so on Modifiers Since most values a RoboCup agent has to handle are based on uncertain observations or estimations there is a need for representing the confidence of the correctness of a value Since the agent is not always aware of the state of an object in the world like the velocity of the ball there is also a need for representing unknown values To support the reasoning with observation times uncertainty and unknown values I have implemented a system with modifiers where a basic type can be modified to handle one of the features 3 2 2 Utilities The library provides some basic mathematic utilities like calculating abso lute values convert radians to degrees and so on but also a class for rep resen
63. lication with different units running in separate threads The benefit would be better performance and a nicer computational model At the same time making it less machine dependent would make it easier to transfer it to other platforms Currently RoboSoc will run on most unix flavors with minor changes but all operating systems that support sockets and some type of signals should be able to run it after modifications Improve the existing framework The most obvious improvement is to improve the views skills and predicates that are included They could be made to deliver more accurate information and be made to do more advanced geometric calculations New views skills and predicates are always welcome and I hope that users of RoboSoc will contribute with theirs to advance the state of the art of RoboCup teams This could help move RoboCup from the hacking stage to the stage where all teams have almost the same level of basic functionality when it comes to individual players Then more scientific methods becomes more and more important One will need a team that has models of the opponents and uses team tactics and plays to be able to win Today one can win most games by only having good individual players Another interesting experiment would be to connect existing specialized software to RoboSoc to take care of certain tasks like planning or prediction Other frameworks The three frameworks that exist today are the most obvious and necess
64. mented and tested e extendible it should be possible to replace and extend the different parts of the library with as little difficulty as possible In order for a library with the above properties to be really useful from an educational point of view it has to take care of the basic problems that a RoboCup agent has to handle Four essential problems I think should be addressed by a library are 1 Basic server communication The first practical problem when building RoboCup agents is to get the basic communication with the server working i e the agent must be able to receive the data sent by the server then be able to use this data to decide on an action and finally to be able to send that action with the correct parameters back to the server 2 Timing As described in the section about the different cycles of the server there are a lot of things to keep in mind when handling the timing Together they make the problem of keeping track of the current simulation step and send commands in time for them to be executed in the current step but at the same time use the best available information when making the decision very hard It is also a major source of confusion when first encountered Therefore it is very important that students do not have to worry about these low level problems 3 World modeling The next problem is to keep track of the state of the simulation based on the sensory data received from the server Since the so
65. mming focusing on the problem of developing RoboCup teams A paper describing the course has been written by Silvia Coradeschi and Jacek Malek 6 They found that The use of a challenging and interesting task RoboCup and the incentive of having a tournament has made the AI programming course quite successful both in terms of enthusiasm of the stu dents and of knowledge acquired Similar conclusions are also drawn by Russel and Norvig when they intro duced the wumpus world in their introductory AI course 25 The major problem with the AT programming course from the students point of view was that creating RoboCup agents requires a large amount of 15 knowledge not related to AI like real time and process programming and a large effort to overcome the initial problems like communicating with the server parsing the server messages calculating the position of the player and sending simple commands back to the server The most requested improve ment was more help with these practical problems of developing a RoboCup agent 6 This Al programming course and the fact that there are no libraries de veloped for the purpose of helping students creating their own RoboCup teams has motivated and inspired me to do this master thesis The edu cational aspects of the problem induces some important demands on the library it should be e Simple enough to be easy to use e general in order to allow different approaches to be imple
66. n be turned independently of its body The angle of the player s head is the viewing angle of the player The turn command changes the angle of the player s body while turn neck changes the angle of the player s head relative to its body The maximum relative angle for the player s neck is 90 degrees to either side Remember that the neck angle is relative to the player s body so if the client issues a turn command the viewing angle changes even if no turn neck command is issued Also turn neck commands can be executed in the same cycle as turn dash and kick commands turn neck is not affected by momentum like turn is The argument for a turn neck command must be in the range 180 180 and the resulting neck angle must be in 90 90 dash The dash is essentially a small push in the direction that the player s body is facing It is not a sustained run In order to have a sus tained run multiple dash commands must be sent The power passed to the dash command is multiplied by dash power rate default 0 006 and the effort see Section 2 1 2 and applied in the direction that the player s body is facing With negative power the agent dashes back wards but it consumes twice the stamina see Section 2 1 2 kick The kick is very similar to the dash except that it accelerates the ball instead of the player If the player tries to kick when the ball is further than the kickable area which is equal to the player size ball_size kick
67. nformation There are for example units specialized in the ball the agent and the players This system of specialized units makes it easy for the user i e the agent programmer to either use the units as they are or modify them to suit her needs RoboSoc provides two more frameworks one for defining skills and one for defining predicates The decision making is almost completely done by the user but it is guided by events generated by the basic system The purposes of the frameworks are to simplify the extension modifica tion and sharing of the units In fact the whole system is designed to be as general and open as possible to encourage and simplify the modification extension and sharing of RoboCup agents created by RoboSoc 1 3 The structure of the report This report is organized as follows Chapter 2 contains the background material needed for this report In section 2 1 the Soccer server is discussed in section 2 2 the educational aspects and finally in section 2 3 related work In chapter 3 RoboSoc is described in detail starting with an overview in section 3 1 where the overall design of RoboSoc is discussed The library is described in section 3 2 the basic system in section 3 3 and the framework in section 3 4 The chapter is concluded with section 3 5 describing the educational value of RoboSoc In the last chapter of the report chapter 4 contains some concluding remarks First a summary is given in section 4 1 and then a di
68. ns the server can execute described in section 3 3 5 To support the decision maker each skill has the possibility to set its 40 own persistence flag to tell the decision maker that it wants to be called again This can be used if a skill wants to do a sequence of actions but is only allowed to return one primitive action each time it is called Which is the case with all the basic skills provided The reason is efficiency and the fact that the world is highly dynamic and changes significantly from cycle to cycle The following skills are currently implemented in RoboSoc They are not really intended to be used in a real team but rather act as examples of what skills can look like CatchBall Purpose To catch the ball Arguments None Preconditions The ball is catchable and the agent is the goalie Algorithm Catch in the direction of the ball Limitations It does not take the movement of the ball or the agent into account FindBall Purpose To find the ball Arguments None Preconditions None Algorithm if the direction to the ball is known then turn towards the ball else turn 45 degrees to the right Limitations It does not take the time between the visual sensor data into account it does not use previous knowledge of the ball InterceptBall Purpose To intercept the ball Arguments None Preconditions The distance and direction to the ball is known Algorithm if the direction to the ball is greater than 5 degrees then
69. on It also contains methods for parsing the command line Available options are f config file to read a config file goalie true to start a player as the goalie and team name team name to set the team name of the agent GameView which stores information about the status of the game play mode score and current time CommandView calculates the effect of the commands executed the current cycle It will give you the total movement vector of the agent and the ball for the current cycle You will also get the total turn angle of the agent and its neck MarkerView stores information about all the markers on the field It will give you absolute position relative distance and direction to the markers AgentView keeps track of the agents status absolute position stamina effort recover absolute face direction body and neck di rections BallView keeps track of the ball absolute position relative dis tance and direction and a very rough estimation of the speed item PlayerView keeps track of the other players You will get access to three structures teammates opponents and unknown players Each will give information about absolute position relative distance and direction body face direction and uniform number if available 60 Player View Marker View m Agent View Ball View p eese Game View Command View Parameter View Sensor Interface AES Depend on the previous c
70. on most of their world modeling and their basic actions and skills Their implementa tion is very advanced and they have one of the most accurate world models available The main problem is the very tight connection between the parts it is almost impossible to replace parts of it The code is also very complex and it is hard to get a grasp of what it really does The documentation is very good and it is easy to get an overall picture of how their team and the major algorithms works Even though their code is not very useful from a student perspective their release is very important since others can use it as a source of inspiration and ideas and are also allowed to take the parts of the code they find use for 21 Chapter 3 RoboSoc This chapter describes the RoboSoc system for developing RoboCup agents It discusses how the system works its design and the reason behind the design It is not a manual of how to use RoboSoc a limited user s guide can be found in Appendix A Since RoboSoc is object oriented and I assume the reader have some basic knowledge about the object oriented paradigm I will sometimes refer to objects classes and methods when I talk about different parts or features of RoboSoc Everything described in this chapter is actually implemented and has been used in a course on Al programming at the computer and information science department at Link ping university in the fall of 1999 1 The implementation is done in
71. on of the position of the agent e s D Metegeesago PEEPS GetFaceDirection hist An angle Get an estimation of the facing direction of the s en agent steps ago TY GetBodyDirection hist An angle Get an estimation of the body direction of the EL agent hist ees ago EE GetNeckDirection hist An angle Get an estimation of the neck direction of the i a agent hist ees ago s Get Effort hist A number Get an estimation of the effort of the agent hist ee ee ee ee Mal Return true if the agent is the goalie otherwise false Get UniformNumber Get the uniform number of the agent GetSpeed hist A number Get an estimation of the speed of the agent hist cycles ago Get View Width hist Get the view width of the agent hist cycles ago Get Recover hist A number Get an estimation of the recover of the agent hist cycles ago Get ViewQuality hist A view quality Get the view quality of the agent hist cycles ago Get ViewAngle hist An angle Get the width of view cone of the agent hist ee ee ee Paro i Get Dashes hist A number Get the number of dashes the agent has made EMIS ap to hat eget ago A number Get number of kicks the agent has made up to pud hist cycles ago GetSays hist A number Get the number of says the agent has made up Get Turns hist A number Get the number of turns the agent has made up EUNT S PI Get TurnNecks hist A number Get the number of turn necks the agent has made up to hist cycles ago Get Kicks hist GetStamina his
72. p 98 Robot Soccer World Cup II Springer Verlag Berlin 1999 Manuela Veloso Enrico Pagello and Hiroaki Kitano editors RoboCup 99 Robot Soccer World Cup III Springer Verlag Berlin 2000 69 36 Michael Wooldridge and Nicholas R Jennings Intelligent agents The ory and practice The Knowledge Engineering Review 10 2 115 152 1995 37 Michael Wooldridge and Nicholas R Jennings Pitfalls of Agent Oriented Development 1998 70
73. r all the stored views and for each one of them call their UpdateAfterXXX method with the update number as the argument This method will check to see if the view has already been updated or not If it has not been updated then it will call the method MyUpdateAfterXXX If the view has been updated nothing happens if the view detects a circular dependency it will abort the program If your view needs updated information from some other view then you need to call its UpdateAfterXXX method with the same update number you got which is stored in the view as currentUpdateNumber The update for a new cycle is somewhat different The NewCycle method of a view is called before the sensor interface updates itself for a new cycle all the other update methods are called after the sensor interface has been updated When the sensor interface updates itself it will erase all data stored in it Therefore this is the last chance for the view to use the 58 data stored in the sensor interface for the current server cycle It is also time for the view to make it ready for a new cycle How do I use the views First of all you need the wanted view object or a pointer to it Then you can call it s public methods and access its public attributes To get a pointer to a view object you need to ask the view manager if the view is stored there This is done by calling the GetView method The argument to GetView is the id number of the view you need A symbol
74. r for the system to be useful from a student s point of view I stated four essential problems the system has to solve or at least support It should take care of the basic interactions with the Soccer server do the timing have support for different world models and have support for decision making The resulting system consists of three parts the library the basic system and the framework The library consists of basic objects and utilities used by the rest of the system and is not dependent on any other part of the system The basic system takes care of the interactions with the server like sending and receiving data It is also responsible for the timing and most of the decision making support by generating events when new things happen The basic system is only depending on the library and can be used without the framework The framework defines three concepts used for world modeling and decision support views predicates and skills The views are specialized information processing units responsible for a specific part of the world model like modeling the ball or the agent They are also controlled by events generated by the basic system The predicates can be used either by the decision maker or the skills to test the state of the world They work like predicates in a three valued logic and can either be true false or unknown The skills are specialized short term planners which generate plans for what actions the agent should take in orde
75. r to reach a 43 desired goal state The framework is depending on both the library and the basic system In 17 Jorg M ller argue that an agent needs five basic capabilities to cope with difficult tasks The agent needs to be e Reactive it should react timely and appropriately to changes in the environment even unforeseen changes e Deliberative it should be able to perform tasks in a goal directed manner e Efficient it should be able to solve its tasks efficiently by using hard coded procedures in routine situations e Interactive it should be able to interact with other agents e Adaptive it should be able to adapt to a changing environment and to cope with unforeseen events Since the computational resources available to the agent is limited a central task is to define a control architecture for resource bounded agents which allows the designer of an agent based system to integrate the require ments mentioned above and to define the trade offs between them in a way that is adequate for the application domain under consideration 17 Ro boSoc is such a control architecture Since it generates events when the environment changes the agent programmerer has the opportunity to inter rupt the current deliberative decision process and react to the event At the same time the implementation is efficient enough and the skill framework provides a way to define hard coded procedures Other types of reactive procedures lik
76. release To download the latest version of RoboSoc including the latest version of the manual or to get more information about the development of it look at the RoboSoc web site 24 A problem with the current release is that RoboSoc uses namespaces be cause of a nameclash between the Unix socket management and the standard template library Therefore your compiler needs to support namespaces for example gcc 2 95 and CC 5 0 does A 1 1 Overview of the software package The RoboSoc system consists of three different parts the library the basic system and the framework You can either use only the library the library and the basic system or the whole system Therefore there are three soft ware packages available the RoboSoc library the RoboSoc basic system and the RoboSoc framework package Each of them contains all the files needed to use the software Each of the three packages are described in more detail in the following sections 47 A 1 2 General parts There are two files included in each of the packages the compiler directives and the types for the package The compiler directives are flags that are set to control the compilation of the package The types are those types that are used by the package There is also an extra set of compiler directives and types in the library package for machine dependent compiler directives and types What directives and types included in each package are described in the section describing
77. rom the class variable CView sensorInterface To make this work you need to assign the address of the sensor interface object to the CView sensorInterface variable 57 Other views are reachable through the view manager pointer viewManager stored in each view In order to know when new information is available the views have a method for each type of update in the system They are NewCycle which is called by the view manager when a new server cycle is about to begin It s called before the UpdateBeforeNew Cycle method of the sensor interface is called This is the last chance to use any data stored in the sensor interface afterwards it will be reset UpdateAfterInit is called when the init messsage has arrived from the server UpdateAfterSee is called when new see information has arrived from the server UpdateAfterHear is called when new audio information has arrived from the server UpdateAfterSense is called when a new sense body has arrived from the server UpdateAfterCommand is called when a command was sent to the server What kind of data you can get from the sensor interface and when it is available is described in the previous section How are the views updated When new information arrives to the sensor interface it will tell the view manager that it is time to update after XXX which is the type of the in formation The view manager will then generate an unique update number and iterate ove
78. rom the critical section of the controllers signal handler the decision maker should not use too much time in the event handlers since there will be no other interrupts while in the critical section Otherwise it can lead to serious problems with the synchronization with the server and the the estimation of the current time 31 Table 3 1 The actuator commands available to the agent Bye Disconnect the agent Catch direction direction minmoment maxmoment ChangeView view_width view_quality view width narrow normal wide view_quality normal high Dash power power minpower maxpower Init team_name version goalie team_name a string version a string goalie true or false Kick power direction power minpower maxpower direction minmoment maxmoment Move z y z 52 5 52 5 y 34 34 Say message message a string of max msg say size characters Turn moment moment ninmoment maxmoment TurnNeck moment Tries to catch the ball in di Yes rection direction Change the current view No width and view quality of the agent Accelerates the agent based Yes on the power Connect initialize the agent Tries to kick the ball in Yes direction with power Moves the agent to absolute Yes position z y if it is al lowed to use move The player screams out No message loud and clear Turn the agent moment de Yes grees to the right Turn the neck of th
79. rt from scratch but can instead build on previous teams experience The rest of the section will be used to describe two existing libraries libsclient and RoboLog and also the team CMUnited which have been a great source of ideas and inspiration Libsclient Libsclient is a library of C routines for basic RoboCup client functionality first developed by Itsuki Noda at ETL Japan and then extended by Yaser A Onaizan Gal Kaminka Jafar Adibi and the other members of the ISIS team at the University of Southern California Information Sciences Insti tute USA This library contains an interface to the server a parser an algorithm for calculating the position of the agent and some very basic algorithms for calculating the absolute position of an object and for calculating the turn s and dash s the agent has to make in order to move to a certain position Unfortunately this library is not very suitable for educational use since it is very basic and does not really handle any of the low level problems of the server For example the basic network services only takes care of sending and receiving messages on request by the agent It does not address any of the problems with synchronization with the server or making sure that the agent tries to send commands each cycle and base them on the best available data which is one of the most important tasks for any library useful for educational purposes But it is useful for scientists who want to 19
80. s the server parameters were adjusted to remove some problems regarding the timing of the agent to make it easier to make good teams The first benefit of the RoboSoc system was that students could use the standard parameters and still get better teams since RoboSoc takes care of the timing of the agent Another benefit was that the students could focus on the problem of developing the skills and the decision making for their agents instead of having to implement the server communication and the processing and storing of information sent by the server Most students felt that it was easy to get started when using RoboSoc and that it was easy to implement their own ideas The major problem reported was the implementation of the modifiers which is not very user friendly at the moment The problem has to do with the way C handles types and type conversions 42 Chapter 4 Conclusions 4 1 Summary This report discusses RoboSoc a system for developing RoboCup agents suitable for educational use It is designed to be as general open and easy to use as possible and to encourage and simplify the modification extension and sharing of RoboCup agents and parts of them To do this I assumed four requirements from the user she wants the best possible data use as much time as possible for the decision making rather act on incomplete information than not act at all and she wants to manipulate the objects found in the soccer environment In orde
81. scussion about future work in section 4 2 There is also an appendix with a limited user s manual Chapter 2 Background 2 1 The Soccer server The Soccer server is the software developed for RoboCup by Itsuki Noda to do the soccer simulation 20 It is a distributed simulation server to which 11 players and 1 coach from each of the two teams can connect One of the players is special the goal keeper The goal keeper is the only player who is allowed to catch the ball There is also a simulated referee who is responsible for enforcing the rules of the game The referee and the mechanics of the simulation are described in section 2 1 2 The communication between the server and the clients is done with the UDP IP protocol which is a connectionless and unreliable internet protocol 29 Each agent player or coach is a separate program with no direct communication with the other agents All communication has to go through the Soccer server When the agent is connected to the server it receives sensor data con tinuously as they are available Visual data is available approximately each 150 milliseconds physical data roughly each 100 milliseconds and aural data directly when either the referee one of the players or one of the coaches say something The sensor data is neither complete nor perfect The sensors are discussed in detail in section 2 1 3 The server also accepts commands from the agent describing what action it wants to perform
82. sel and Peter Norvig A Modern Agent Oriented Approach to Introductory Artificial Intelligence 1995 Stuart Russel and Peter Norvig Artificial Intelligence A Modern Ap proach Prentice Hall International 1995 Paul Scerri Johan Ydren Tobias Wiren Mikael L nneberg and Per Erik Nilsson Headless Chickens III In Manuela Veloso Enrico Pagello and Hiroaki Kitano editors RoboCup 99 Robot Soccer World Cup III Springer Verlag Berlin 2000 The Soccer Server web page http ci etl go jp noda soccer server March 2000 Richard W Stevens UNIX network programming Vol 1 Prentice Hall PTR Upper Saddle River NJ 1998 Peter Stone Layered Learning in Multi Agent Systems PhD thesis Carnegie Mellon University 1998 Peter Stone Patrick Riley and Manuela Veloso The CMUnited 99 champion simulator team In Manuela Veloso Enrico Pagello and Hiroaki Kitano editors RoboCup 99 Robot Soccer World Cup III Springer Verlag Berlin 2000 Peter Stone Patrick Riley and Manuela Veloso Layered Disclosure Why is the agent doing what it s doing In Proceedings of the Fourth International Conference on Autonomous Agents 2000 Peter Stone and Manuela Veloso The cmunited 97 simulator team In Hiroaki Kitano editor RoboCup 97 Robot Soccer World Cup I Springer Verlag Berlin 1998 Peter Stone Manuela Veloso and Patrick Riley The CMUnited 98 Champion Simulator Team In Minoru Asada and Hiroaki Kitano edi tors RoboCu
83. shown in Table A 5 Modifiers There are three types of modifiers Each with its own feature The three modifiers are with unknown with confidence and with observation time 49 Table A 3 Data types for RoboCup concepts available in RoboSoc Type t view width VW Normal VW Wide A type for representing the VW Narrow VW Unknown concept of view width t view quality VQ High VQ Low A type for representing the VQ_Unknown concept of view quality cmd type CMD Unknown A type for representing the MD_Bye available server commands MD_Catch MD ChangeView MD Dash MD Init MD Kick MD Move MD Say MD SenseBody MD Turn MD TurnNeck t t_side_line SL Left SL Right SL_Top A type for representing the SL Bottom SL Unknown side lines OOOO CY CY C303 C3 C3 t marker Goal L Goal R A type for representing the Flag_C Flag_CT markers Flag_CB Flag_LT Flag LB Flag RT Flag RB Flag PLT Flag PLC Flag PLB Flag PRT Flag PRC Flag PRB Flag_GLT Flag GLB Flag GRT Flag GRB Flag TL50 Flag TL40 Flag TL30 Flag TL20 Flag TL10 Flag TO Flag TR10 Flag TR20 Flag TR30 Flag TR40 Flag TR50 Flag BL50 Flag BL40 Flag BL30 Flag BL20 Flag_BL10 Flag_BO Flag BR10 Flag BR20 Flag BR30 Flag BRAO Flag BR50 Flag LT30 Flag LT20 Flag LT10 Flag L0 Flag LB10 Flag_LB20 Flag LB30 Flag RT30 Flag RT20 Flag RT10 Flag RO Flag RB10 Flag RB20 Flag RB30 Unknown Marker
84. t A number Get an estimation of the stamina of the agent hist cycles ago 38 Table 3 8 Functionality provided by the ball view Name Return type Get BallObject hist A ball object Get the object representing the I att hist egeles go GetBallDistance hist A float Get the distance to the ball hist i agaesago e O e a e GetRelativeBall Vector hist A vector Get the relative vector to the ball pA het yees ago GetRelativeBallDirection hist An angle Get the relative direction to the BLUE TE ago Get AbsoluteBallVector hist A vector Get the absolute vector to the ball e itepe ago o o GetAbsoluteBallDirection hist An angle Get the absolute direction to the ne eat ete PM Get BallPosition hist A point Get the absolute position of the Pe ban pist eyeles ago on O GetBallDist Changed hist A float Get the distance change of the ball ENEMIES GetBallDirChanged hist An angle Get the direction change of the BENNETT go GetBallAbsoluteSpeed Vector hist A vector Get the absolute speed vector of OE tie pali it gees ago GetBallRelativeSpeed Vector hist A vector Get the relative speed vector of EN tne all teeter ago GetBallSpeed hist A float Get the speed of the ball hist cy pon tes ago E GetBallAbsoluteSpeedDirection hist An angle Get the absolute direction of the poss speed of the ball hist cycles ago GetBallRelativeSpeedDirection hist An angle Get the relative direction of the speed of the ball hist cycles ago
85. tend existing predicates A 4 6 How to usetheskills lle A 4 7 How to create your own skills 0 A 4 8 How to extend existing skills Bibliography 43 43 45 4T 4T 47 48 48 48 48 48 53 53 54 54 54 54 54 54 54 55 55 55 57 57 57 61 61 61 62 62 62 67 Chapter 1 Introduction This chapter contains a short description of the problem I have studied namely how to create a system for developing RoboCup agents and the outline of my solution RoboSoc The purpose of this chapter is to motivate and define the conditions for the research but also to state the assumptions I had to make in order to solve it The last section of this chapter describes the structure of the report But first a short introduction to RoboCup for those who are not familiar with it RoboCup is an attempt to foster artificial intelligence AT and intel ligent robotics research by providing a standard problem soccer where a wide range of technologies can be integrated and examined The reason for choosing soccer is that it is played in a highly dynamic environment with both teammates and opponents where independent agents must collabo rate in order to beat the other team 14 At the same time it is easy to understand the problem and most people know how soccer is played and what a team should do to be successful These properties make RoboCup an excellent environment for students to in a pl
86. tep usually 10 milliseconds and the IO handler an interrupt as soon as data arrive to the socket connecting the agent to the Soccer server The first algorithm is used to update the agent after a tick of the internal clock the second algorithm is used to update the agent when sensor data is received and the third algorithm is used to update the current time of the agent based on the last known server time These three algorithms define the working cycle of the agent since they generate the events which guide the decision maker The only assumption about the server the algorithms make is that the agent receives physical sensor data from a sense body message almost ev ery cycle either at the beginning of the new cycle or at the end of the old cycle This assumption makes it possible to assume a new cycle has started when the agent receive a sense body message and it can also assume it has almost 100 milliseconds before its command for this cycle has to be at the server Before discussing the algorithms in detail the RoboSoc concept of time needs to be defined Time in RoboSoc The time concept used in RoboSoc is based on the CMUnited concept of time 8 They represent time as a tuple of the last known server time and the number of cycles since the clock was stopped If the clock is not stopped then the second value is 0 The reason behind this concept is that the agent wants to reason about time even when the game clock is stopped and the time r
87. that package A 2 The library This section will describe the functionality and the use of the library module of the RoboSoc system A 2 1 How to use the library The library is not a single unit but rather a collection of classes To use the library you therefore have to include the header file for the wanted class The names of the include files are usually the name of the class with the extension h The definitions of the classes are in a file with the same name as the include file but with the extension cc A 2 2 Compiler directives The following directives can be defined in library compiler directives h TRIG_IN_DEG_ If trigonometry functions should return angles in degrees TRIG IN RAD If trigonometry functions should return angles in radians USE EXCEPTIONS If exceptions should be used USE DEBUG If the debug features should be used Only one of the first two directives can be defined otherwise the compi lation will fail A 2 3 Data types RoboSoc contains many different data types used to encapsulate both ma chine dependent and implementation dependent data types The machine dependent data types are shown in Table A 1 and they are all defined in machine_dependent_types h The definition of these must be changed to match the size of the data types on the local system There are also three basic number types that have an unspecified maxi mum size but a specified minimum size since they are used in such a wa
88. ting a communication socket and a class for representing a real time timer 24 CGameObject CStationaryObject CLineObject CMarkerObject CMobileObject CBallObject CPlayerObject CAgentObject Figure 3 2 The object hierarchy of the game objects in the RoboSoc library 3 2 8 Geometric objects The geometric objects included in the library are classes for representing angles both in degrees and in radians vectors point lines and rectangles The reason for having two types of angles is that you do not want to confuse the user by making it unclear what unit the angles are in Modifiers The same argument about uncertainty and unknown values is valid for geo metric objects therefore there are modifiers for them too Unfortunately the modifiers described above only work on basic types like integers and floats therefore there exist specialized modifiers for the geometric objects 3 2 4 Game objects The game objects represent the objects introduced by the soccer environ ment They include a ball object a player object and an agent object The object hierarchy is shown in Figure 3 2 This hierarchy looks almost the same as the object hierarchy used by CMUnited 34 but in fact they are not very similar because my objects have no reference to the current world model of the agent as the CMUnited objects have The main reason is that I want the library to be stand alone and not dependent on anything else esp
89. to provide an answer to the question of how the sensor data and the current internal state of the agent determine the actions and future internal state of the agent An architecture encompasses techniques and algorithms that support this methodology 16 Over the past decade a number of different approaches to the problem of finding good functional decompositions of agents has lead to at least four major types of agent architectures reactive deliberative hybrid and interacting architectures 17 Reactive architectures The reactive behavior based or situated architectures are strongly influ enced by behaviorist psychology The goal is to achieve robust behavior instead of correct or optimal behavior They do this by using simple action rules based on the current situation with little or no explicit information 17 An example of a behavior based architecture used in the RoboCup do main is the architecture used by the Headless Chickens 7 27 Deliberative architectures Agents based on Simon and Newell s physical symbol system hypothesis 19 the assumption that agents maintain an internal representation of their world and that there is an explicit mental state which can be modified by some form of symbolic reasoning are called deliberative agents 17 One example of a deliberative architecture is the Belief Desire Intention architecture BDI developed by Bratman et al 3 BDI is used in the RoboCup domain by for examp
90. ty classes like geometric objects game objects and some basic building blocks to be used with the framework e a basic system for the timing and the interaction with the server and e a framework for information processing skills and predicates The system architecture is shown in Figure 3 1 and includes the basic system and the framework The figure consists of three different kinds of boxes representing different kinds of units The rectangular boxes repre sent objects classes the hexagonal boxes collections of objects from the same base class and the boxes with the rounded edges separate indepen dent programs The arrows show how data and information flows between the different subparts An arrow does not always implicate a dependency between the units In fact the library is not dependent on any part of the system the basic system is dependent on the library and the framework is dependent on the basic system and therefore also on the library This means 23 that a user can either use the library the basic system and the library or the whole system when creating a team Most of the implementation of the framework and the decision maker is supposed to be done by the user with support from the framework and the library while the basic system is not supposed to be changed be the user unless it is really necessary The following sections of this chapter describe the functionality and de sign of the three majors parts 3 2 The
91. up 98 Robot Soccer World Cup II Springer Verlag Berlin 1999 Emiel Corten Klaus Dorer Fredrik Heintz Kostas Kostiadis Johan Kummeneje Helmut Myritz Itsuki Noda Jukka Riekki Patrick Riley Peter Stone and Tralvex Yeap Soccerserver Manual Ver 5 1 1999 Patrick Doherty The WITAS Integrated Software System Ar chitecture Link ping Electronic Articles in Computer and Information Science Vol 4 1999 no 17 December 1999 http www ep liu se ea cis 1999 017 67 10 11 12 m H4 m 16 m m 19 20 p5 21 ca 22 James R Firby Building symbolic primitives with continuous control routines In J Hendler editor Proceedings of the 1st International Con ference on Artificial Intelligence Planning Systems AIPS 92 Morgan Kaufmann Publishers San Mateo CA 1992 Stanley Franklin and Arthur Graesser Is it an agent or just a program In Intelligent Agents III pages 21 36 Fredrik Heintz FCFoo a short description In Robocup 1999 Team Description Simulation League Link ping Electronic Press 1999 Hiroaki Kitano editor RoboCup 97 Robot Soccer World Cup I Springer Verlag Berlin 1998 Hiroaki Kitano Milind Tambe Peter Stone Manuela Veloso Silvia Coradeschi Eiichi Osawa Hitoshi Matsubara Itsuki Noda and Minoru Asada The RoboCup synthetic agent challenge 97 In Proceedings of the Fifteenth International Joint Conference on Artificial Intelligenc
92. ure 2 5 The 14 oO 5 j 5 5 5 Agent timeline Server timeline t t 1 t 2 t 3 Visual sensor data sent Physical sensor data sent e Action sent and executed Figure 2 5 An example of what data the agent receives when and also when commands sent by the agent will be executed example in the figure shows what data the agent receives and when As shown the data can arrive both earlier and later than expected All actions received by the server at the end of the simulator cycle will be used when updating the state of the simulation The example shows the agent sending commands back to the server but due to the delays in the network and the fact that the server takes some time before reading the messages received between 0 and recv step milliseconds they are not always executed in the same simulation step as the agent sent them Since the agent neither knows when it will receive the next sensor in put nor when the current simulator cycle will end it has to make a decision whether to wait for more information or to act on the current information This decision will affect both how often the agent can act and how much in formation can be used when deciding what action to do This balance is very central to every RoboCup agent and have great impact on the performance of the agent 2 2 Educational aspects Since 1997 the Department of Computer Science at Link ping university has been giving a course in Al progra
93. xisting view inherit from the view you want to extend and then add the things you miss and redefine the necessary methods How do I write my own views The best way to add new functionality regarding information processing in RoboSoc is to create a new view To create a new view all you have to do is to inherit from CView or any other view that you want to use as your starting point and write your own update methods Le you need to define 59 what the view should do when new information has arrived and what to do when a new cycle begin If you need to store data which is very likely you will need to create the necessary data structures and initialize them in the constructor destroy them in the destructor and supply methods to extract and use the information stored in them If you need access to other views the best way to do this is to store pointers to those views in the object and initialize them in the constructor Then you ll know that the views you want are always accessible For examples and hints look at the code for the views that are included in the RoboSoc package What views are included in RoboSoc Currently seven views are supplied with RoboSoc version 1 5 0 I hope to be able to add more views as the users of RoboSoc get their views working and sending them to me The views are ParameterView stores all the parameters from server conf Use them to make your agent independent of the actual server configura ti
94. y that the size of the type mainly affects the precision of the calculation not the range of the values The three types are shown in Table A 2 48 Table A 1 The basic machine dependent data types available in RoboSoc Type Size bytes t_uint16 t uint32 Ye taunt rs Unsigned 64 bit integers tintt6 Ye tint Umi Y 8 Signed 64 bit integers ioatea Yes 8 Signed 64 bit foats Table A 2 The basic precision dependent data types available in RoboSoc Min size bytes Unsigned integers with at least 16 bits Signed integers with at least 32 bits Signed floats with at least 32 bits RoboCup types The RoboCup specific enumeration types are shown in Table A 3 and are all defined in library types h RoboSoc templates RoboSoc contains a few templates that are used to represent angles both in radians AngleRad type and in degrees AngleDeg type and coordi nates both vectors CoordVector type and points CoordPoint type The types in the templates are used to represent basic values like the angle or the x and y coordinate There is also a template for representing collections of objects called ObjectCollection object type gt the object type is the type of the object in the collection RoboSoc types The RoboSoc specific types are shown in Table A 4 and are all defined in library types h There are also some named instantiations of templates defined in library_types h
95. ycle s information Depend on the current cycle s information Figure A 1 The dependencies between the views How the views depend on each other can be seen in Figure A 1 A more detailed description of these views is available in the section 3 4 2 A 4 3 How to use the predicates Create an instance of the predicate you want Call the Evaluate method which will return either true or false A 4 4 How to create your own predicates Create a new class that inherits from the CPredicate class Implement the bool Evaluate method A 4 5 How to extend existing predicates Create a new class that inherits from the predicate you want to extend Call its Evaluate method if its functionality is useful otherwise rewrite it from scratch 61 A 4 6 How to use the skills Create an instance of the skill you want To check whether the skill is appli cable call its Applicable method which will return true if it is applicable otherwise false To get the next command generated by the skill call its GeneratePlan method If the skill wants to be called again the next cycle it wil set the persistence flag of the skill This flag is checked with the Persistent method A 4 7 How to create your own skills Create a new class that inherits from the CSkill class Implement the bool Applicable bool Persitent and CCommand GeneratePlan methods A 4 8 How to extend existing skills Create a new class that either inherits

RoboSoc a System for Developing RoboCup Agents for

Contents

Download Pdf Manuals

Related Search

Related Contents