RoboCup 2003: Robot Soccer World Cup VII

RoboCup 2003, the seventh RoboCup Competition and Conference, took place between July the 2nd and July the 11th 2003 in Padua (Italy). The teams had three full days to setup their robots. The competitions were held in the new pavilion no7 of the Fair of Padua (Fig. 1). Several scientific events in the field of Robotics and Artificial Intelligence were held in parallel to the competitions. The RoboCup Symposium was held in the last two days. The opening talks took place in the historical Main Hall of the University of Padua and the three parallel Symposium sections in the conference rooms of the Fair of Padua.

RoboCup has been known for the goal of “creating robots capable of beating the world cup in 2050.” Clearly, we stated this goal back in 1996 not as an exact scientific goal, but as an audacious challenge to pursue. We aimed at creating a far away target for RoboCup researchers, as we were well aware that the development of fully autonomous soccer robots capable of competing against human world champion soccer players was a rather long term research project.

The goal of this paper is to highlights one of emergent scientific issues in RoboCup task domains that has broader applications even outside of the RoboCup task domains. This paper particularly focuses on robust recognition through information fusions issue among numbers of others issues that are equally important. The robust recognition through information fusion is selected because it is one of the most universal issues in AI and robotics, and particularly interesting for domains such as soccer and rescue that has high degree of dynamics and uncertainty, as well as being resource bounded. The author wish to provide a conceptual framework on robust perception from single agent to multi-agent teams.

A common problem in RoboCup is role allocation: given a team of players and a set of roles, how should be roles be allocated to players? Drawing on our previous work in multi-robot task allocation, we formalize the problem of role allocation as an iterated form of optimal assignment, which is a well-studied problem from operations research. From this perspective, we analyze the allocation mechanisms of a number of RoboCup teams, showing that most of them are greedy, and that many are in fact equivalent, as instances of the canonical Greedy algorithm. We explain how optimal, yet tractable, assignment algorithms could be used instead, but leave as an open question the actual benefit in terms of team performance of using such algorithms.

This paper – a summary of a keynote address given at the Robocup 2003 symposium – argues i) that mobile robotics research would benefit from a theoretical understanding of robot-environment interaction, ii) that independent replication and verification of experimental results should become common practice within robotics research, and iii) that quantitative measures of robot behaviour are needed to achieve this.The paper gives one example of such quantitative measures of behaviour: the reconstruction of the phase space describing a robot’s behaviour, and its subsequent analysis using chaos theory.

Complexity science is characterised by computational irreducibility, chaotic dynamics, combinatorial explosion, co-evolution, and multilevel lattice hierarchical structure. One of its main predictive tools is computer-generated distributions of possible future system states. This assumes that the system can be represented inside computers. Robot soccer provides an excellent laboratory subject for complexity science, and we seek a lattice hierarchical vocabulary to provide coherent symbolic representations for reasoning about robot soccer systems at appropriate levels. There is a difference between constructs being human-supplied and being abstracted autonomously. The former are implicitly lattice-hierarchically structured. We argue that making the lattice hierarchy explicit is necessary for autonomous systems to abstract their own constructs. The ideas are illustrated using data taken from the RoboCup simulation competition.

High-level online methods become more and more attractive with the increasing abilities of players and teams in the simulation league. As in real soccer, the recognition and prediction of strategies (e.g. opponent’s formation), tactics (e.g. wing play, offside traps), and situations (e.g. passing behavior) is important. In 2001, we proposed an approach where spatio-temporal relations between objects are described and interpreted in order to detect some of the above mentioned situations. In this paper we propose an extension of this approach that enables us to both interpret and predict complex situations. It is based on a qualitative description of motion scenes and additional background knowledge. The method is applicable to a variety of situations. Our experiment consists of numerous offside situations in simulation league games. We discuss the results in detail and conclude that this approach is valuable for future use because it is (a) possible to use the method in real-time, (b) we can predict situations giving us the option to refine agents actions in a game, and (c) it is domain independent in general.

We introduce a concise approach to teamwork evaluation on multiple levels – dealing with agent’s behaviour spread and multi-agent coordination potential, and abstracting away the team decision process. The presented quantitative information-theoretic methods measure behavioural and epistemic entropy, and detect phase transitions – the edge of chaos – in team performance. The techniques clearly identify under-performing states, where a change in tactics may be warranted. This approach is a step towards a unified quantitative framework on behavioural and belief dynamics in complex multi-agent systems.

Imitation Learning is considered both as a method to acquire complex human and agent behaviors, and as a way to provide seeds for further learning. However, it is not clear what is a building block in imitation learning and what is the interface of blocks; therefore, it is difficult to apply imitation learning in a constructive way. This paper addresses agents’ intentions as the building block that abstracts local situations of the agent and proposes a hierarchical hidden Markov model (HMM) in order to tackle this issue. The key of the proposed model is introduction of gate probabilities that restrict transition among agents’ intentions according to others’ intentions. Using these probabilities, the framework can control transitions flexibly among basic behaviors in a cooperative behavior. A learning method for the framework can be derived based on Baum-Welch’s algorithm, which enables learning by observation of mentors’ demonstration. Imitation learning by the proposed method can generalize behaviors from even one demonstration, because the mentors’ behaviors are expressed as a distributed representation of a flow of likelihood in HMM.

Specific behavior description languages prove to be suitable replacements to native programming language like C++ when the number and complexity of behavior patterns of an agent increases. The XML based Extensible Agent Behavior Specification Language (XABSL) also simplifies the process of specifying complex behaviors and supports the design of both very reactive and long term oriented behaviors. XABSL uses hierarchies of behavior modules called options that contain state machines for decision making. In this paper we introduce the architecture behind XABSL, the formalization of that architecture in XML and the software library XabslEngine that runs the formalized behavior on an agent platform. The GermanTeam [9] employed XABSL in the RoboCup Sony Four Legged League competitions in Fukuoka.

In the growing area of multi-agent-systems (MAS) also the diversity of the types of agents within these systems grows. Agent designers can no longer hard-code all possible interaction situations into their software, because there are many types of agents to be encountered. Thus, agents have to adapt their behavior online depending on the encountered agents. This paper proposes that agent behavior can be classified by distinct and stable tactical moves, called features, on different levels of granularity. The classification is used to select appropriate counter-strategies. While the overall framework is aimed to be applicable in a wide range of domains, the feature-representation in the case-base and the counter-strategies is done in a domain-specific language. In the RoboCup domain the standard coach-language is used. The approach has been successfully evaluated in a number of experiments.

This paper presents the application of a novel method in the multi-agent teamwork field called Scenario-based Teamworking (SBT). In SBT method a team of cooperative intelligent agents could be able to execute complex plans in nondeterministic, adversary, and dynamic environments which communication cost is high. The base idea of this method is to define Scenario for different situations. With a graph of scenarios, a team of agents can execute, learn, adapt, and create team plans automatically. This method has implemented in a soccer team of intelligent agents (players and coach) and evaluated in the standard RoboCup simulator environment [1] and results show a significant improvement.

The use of agents and multiagent systems is widespread in computer science nowadays. Thus the need for methods to specify agents in a clear and simple manner arises. One way of achieving this is by means of a graphical formalism. For using such a formalism the availability of tools, that support a developer, is of great importance. In this paper we present an approach to specifying agent behaviors on different levels of abstraction with the help of UML statecharts. Cooperation between different agents can explicitly be modeled. To help a developer with applying this formalism to the specification of agent behaviors the statechart editor StatEdit is presented. This development tool supports not only the modelling of an agent but a simple form of code generation as well.

Applications of recurrent neural networks (RNNs) tend to be rare because training is difficult. A recent theoretical breakthrough [Jae01b] called Echo State Networks (ESNs) has made RNN training easy and fast and makes RNNs a versatile tool for many problems. The key idea is training the output weights only of an otherwise topologically unrestricted but contractive network. After outlining the mathematical basics, we apply ESNs to two examples namely to the generation of a dynamical model for a differential drive robot using supervised learning and secondly to the training of a respective motor controller.

This paper describes the design, implementation and test of a goalkeeper robot for the Middle-Size League of RoboCup. The goalkeeper task is implemented by a set of primitive tasks and behaviors, coordinated by a 2-level hierarchical state machine. The primitive tasks concerning complex motion control are implemented by a non-linear control algorithm, adapted to the different task goals (e.g., follow the ball or intercept the ball). One of the top level behaviors regularly determines the robot posture from local features extracted from images acquired by a catadioptric omni-directional vision system. Most robot parameters were designed based on simulations carried out with the Hybrid Automata Matlab/Simulink toolbox CheckMate. Results obtained with the actual goalkeeper are presented and discussed.

Recognition of relevant game field objects, such as the ball and landmarks, is usually based upon the application of a set of decision rules over candidate image regions. Rule selection and parameters tuning are often arbitrarily done. We propose a method for evolving the selection of these rules as well as their parameters with basis on real game field images, and a supervised learning approach. The learning approach is implemented using genetic algorithms. Results of the application of our method are presented.

Our research on coaching refers to one autonomous agent providing advice to another autonomous agent about how to act. In past work, we dealt with advice-receiving agents with fixed strategies, and we now consider agents which are learning. Further, we consider agents which have various limitations, with the hypothesis that if the coach adapts its advice to those limitations, more effective learning will result. In this work, we systematically explore the effect of various limitations upon the effectiveness of the coach’s advice. We state the two learning problems faced by the coach and the coached agents, and empirically study these problems in a predator-prey environment. The coach has access to optimal policies for the environment, and advises the predator on which actions to take. We experiment with limitations on the predator agent’s actions, the bandwidth between the coach and agent, and the memory size of the agent. We analyze the results which show that coaching can improve agent performance in the face of all these limitations.

TsinghuAeolus is the champion team for the latest twoRoboCup simulation league competitions. While our binary and nearly full source code for RoboCup 2001 had been publicly available for the entire year, we won the champion again in Fukuka, with more obvious advantage. This paper describes the key innovations that bring this improvement. They include an advice-taking mechanism which aims to improve agents’ adaptability, a compact and effective option scoring policy which is crucial in the option-evaluation framework, and thorough analysis of interception problem which leads to more intelligent interception skill. Although not strongly interrelated, these innovations come together to form a set of solutions for problems across different levels.

This paper presents a real-time approach for object recognition in robotic soccer. The vision system does not need any calibration and adapts to changing lighting conditions during run time. The adaptation is based on statistics which are computed when recognizing objects and leads to a segmentation of the color space to different color classes. Based on attention, scan lines are distributed over the image ensuring that all objects of interest intersect with the number of lines necessary for recognition. The object recognition checks the scan lines for characteristic edges and for typical groupings of color classes to find and classify points on the outlines of objects. These points are used to calculate size and position of the objects in the image. Experiments on Sony’s four-legged robot Aibo show that the method is able to recognize and distinguish objects under a wide range of different lighting conditions.

Colour labeling is critical to the real-time performance of colour-based vision systems and is used for low-level vision by most RoboCup 2002 physically based teams. Unfortunately, colour labeling is sensitive to changes in illumination and manual calibration is both time consuming and error prone.In this paper, we present KADC, a robust method for Knowledge-based Autonomous Dynamic Colour Calibration. By utilising the known geometry of the environment, landmarks are identified independent of colour classifications. Colour information from these landmarks is used to construct colour clusters of arbitrary shape. Clusters are dynamically updated through actions and by the use of a similarity metric, the Earth Mover’s Distance (EMD). We apply KADC to the RoboCup Legged League, generating a colourtable purely from geometrical knowledge of the environment and dynamically update this colortable to compensate for non-uniform changes in lighting conditions.

The recent debate in the RoboCup middle-size community about natural light conditions shows that a more in-depth analysis of the problems incurred by this is necessary in order to draft out a focused and realistic roadmap for research. Based on real-world images taken under varying lighting conditions, we performed descriptive and statistical analysis of the effects on color-based vision routines. The results show that pure color-based image processing is not likely to perform well under varying lighting conditions, even if the vision system is calibrated on a per-game base. We conclude that color-based vision has to be combined with other methods and algorithms in order to work robustly in more difficult environments with varying illumination.

In this paper we present a simple and new algorithm that tracks the contour of several homogenous regions in a sequence of images. The method exploits the fact that, when i.e. observing a moving object (exposing a homogenous region), the regions in two consecutive frames often overlap. We show that the method is valuable for the RoboCup domain: It allows to track the green playing field and the goals very efficiently, to detect the white marking lines precisely, enabling us to recognize features in them (the center circle, the quatre circles, corners, the rectangle of the penalty area,...). It is also useful to find the ball and the obstacles. Furthermore, it provides data for path planning based on potential field methods without further computation. We compared the algorithm with the fastest existing method and measured a speed enhancement of 30 percent. In contrast to other methods, our algorithm not only tracks the center of blobs but yields the precise boundary shape of the objects as a set of point sequences. First tests with real world data have confirmed the applicability for other domains than RoboCup.

This paper presents a fast approach for edge-based self-localization in RoboCup. The vision system extracts edges between the field and field lines, borders, and goals following a grid-based approach without processing whole images. These edges are employed for the self-localization of the robot. Both image processing and self-localization work in real-time on a Sony Aibo, i. e. at the frame rate of the camera. The localization method was evaluated using a laser range sensor at the field border as a reference system.

At present, visual localization of soccer playing robots taking part in the RoboCup contest is mainly achieved by using colored artificial landmarks. As known, this method causes further vision problems like color classification and segmentation under variable light conditions. Additionally, robots confined to use visual sensor information from common cameras usually waste time in switching between the modi of playing soccer and searching landmarks for localization. An upcoming approach to solve these problems is the detection of field lines. Motivated by our research in using a compact symmetry operator for natural feature extraction in mobile robot applications, we propose its application to the RoboCup contest. Symmetry is a structural feature and as results show, it is highly independent of illumination changes and very compliant to the task of line detection. We will motivate symmetry as a natural feature, discuss the symmetry operator and finally present results of the field line extraction.

This paper proposes using topics central to RoboCup, particularly autonomous agents and multiagent systems, as the subject-matter for a course designed to introduce undergraduate students to all facets of computer science research. Experiences are presented from the design and implementation of such a course. The course is structured around an ongoing incremental programming project that culminates in a class tournament in the RoboCup Soccer Server, an open-source infrastructure built to support multiagent systems research and education.

Since team-based projects have been proven to be an effective pedagogical tool, we have been using RoboCup challenges as the basis for class projects in undergraduate courses. This paper unifies several independent efforts in this direction and presents early work in the development of shared resources and evaluation. We outline three courses and describe the related class projects in order to make the context of our investigation clear and to make it possible for others to replicate or extend our work, and contribute to the shared resource.

Since 2001 the School of Information Technology and Electrical Engineering (ITEE) at the University of Queensland has been involved in RoboCupJunior activities aimed at providing children with the Robot building and programming knowledge they need to succeed in RoboCupJunior competitions. These activities include robotics workshops, the organization of the State-wide RoboCupJunior competition, and consultation on all matters robotic with schools and government organizations. The activities initiated by ITEE have succeeded in providing children with the scaffolding necessary to become competent, independent robot builders and programmers. Results from state, national and international competitions suggest that many of the children who participate in the activities supported by ITEE are subsequently able to purpose- build robots to effectively compete in RoboCupJunior competitions. As a result of the scaffolding received within workshops children are able to think deeply and creatively about their designs, and to critique their designs in order to make the best possible creation in an effort to win.

In this article we propose an extension for a path planning method based on the LPN-algorithm to have better performance in a very dynamic environment. The path planning method builds a navigation function that drives the robot toward the goal avoiding the obstacles. The basic method is very fast and efficient for a robot with few degrees of freedom. The proposed extension integrates the obstacle dynamics in the planning method to have better performances in very dynamic environments. Experiments have shown the effectiveness of the proposed extension in a very dynamic environment, given by RoboCup soccer matches.

Grid-based methods for finding cost optimal robot paths around obstacles are popular because of their flexibility and simple implementation. However, their computational complexity becomes unfeasible for real-time path planning if the resolution of the grid is high.These methods assume complete knowledge about the world, but in dynamic environments where sensing is done on board the robot, less is known about far-away obstacles than about the ones in close proximity.The paper proposes to utilize this observation by employing a grid of variable resolution. The resolution is high next to the robot and becomes lower with increasing distance. This results in huge savings in computational costs while the initial parts of the paths are still planned with high accuracy. The same principle is applied to the time-axis, allowing for planning paths around moving obstacles with only a moderate increase in computational costs.

This paper presents a method for generating vision-based humanoid behaviors by reinforcement learning with rhythmic walking parameters. The walking is stabilized by a rhythmic motion controller such as CPG or neural oscillator. The learning process consists of two stages: the first one is building an action space with two parameters (a forward step length and a turning angle) so that infeasible combinations of them are inhibited. The second one is reinforcement learning with the constructed action space and the state space consisting of visual features and posture parameters to find feasible action. The method is applied to a situation of the RoboCupSoccer Humanoid league, that is, to reach the ball and to shoot it into the goal. Instructions by human are given to start up the learning process and the rest is completely self-learning in real situations.

This paper describes the methodology that we used to design and implement balancing and walking gaits for Tao-Pie-Pie, a small 30cm tall humanoid robot. Tao-Pie-Pie is a fully autonomous robot with all power, sensing, and processing done on-board. It is also a minimalistic design with only six degrees of freedom. Nevertheless, its performance is comparable to that of other more complex designs. The paper describes three patterns: (a) a straight walk, (b) a turn on the spot, and (c) a kicking pattern. Sensor feedback is provided by two gyroscopes that provide angular velocity in the left-right and forward-backward plane and a CMOS camera providing vision information. The feedback from the gyroscopes is not used to actively control the walking gait, because the signal is noisy and it would be computationally to expensive for the current processor hardware. Instead, coarse feedback from the gyroscopes is used to monitor the transition from one phase of the pattern to the next. Tao-Pie-Pie proved to be a successful design winning a number of honors at international competitions.

Humanoid robots are without question a hot topic in research today. But will they really be the next break-through invention that changes the face of the world, or are they just another over-hyped research toy? ProRobot is a study funded by the European Commission that will have a close look on the future of humanoid robots and their economic and social impact. The complete study will be published in summer 2003, and will especially concentrate on the prospects of research efforts and the differences of research activities throughout the world.

With the introduction of commercially available programm- able legged robots, a generic software method for detection of abnormalities in the robots’ locomotion is required. Our approach is to gain satisfactory results using a bare minimum amount of hardware feedback; In most cases we are able to detect faults using only the joint angle sensors. Methods for recognising several types of collision as well as a loss of traction are examined. We are particularly interested in applying such techniques to Sony AIBO robots in the RoboCup legged league environment. This investigation provided us with a technique that enabled us to detect collisions with reliable accuracy using limited training time.

The control and coordination of multiple mobile robots is a challenging task; particularly in environments with multiple, rapidly moving obstacles and agents. This paper describes a robust approach to multi-robot control, where robustness is gained from competency at every layer of robot control. The layers are: (i) a central coordination system (MAPS), (ii) an action system (AES), (iii) a navigation module, and (iv) a low level dynamic motion control system. The multi-robot coordination system assigns each robot a role and a sub-goal. Each robot’s action execution system then assumes the assigned role and attempts to achieve the specified sub-goal. The robot’s navigation system directs the robot to specific goal locations while ensuring that the robot avoids any obstacles. The motion system maps the heading and speed information from the navigation system to force-constrained motion. This multi-robot system has been extensively tested and applied in the robot soccer domain using both centralized and distributed coordination.

This paper sketches and discusses design options for complex probabilistic state estimators and investigates their interactions and their impact on performance. We consider, as an example, the estimation of game states in autonomous robot soccer. We show that many factors other than the choice of algorithms determine the performance of the estimation systems. We propose empirical investigations and learning as necessary tools for the development of successful state estimation systems.

One of the typical cooperative actions is the pass play in RoboCup small-size league. This paper presents three technical key features to realize robust pass play between robots. The first one is the high resolution image processing to detect the positions and orientations of the robots. The second one is the control algorithm to move and adjust the robots for the pass play. The third one is the mechanism to catch the ball moving at high speed. This paper discusses these methods and shows the effectiveness of the methods by experimental results.

In the RoboCup F-180 league competition, vision is predominantly provided by an overhead camera which relays a global view of the field. There are inherent disadvantages in utilising this system, particularly the delays associated with the capture, transmission and processing of vision data. To minimise these delays and to equip the robots with greater autonomy, visual servoing on-board the individual robots is proposed. This paper presents evaluation of two visual servoing methods for mobile robots: position-based and image-based servoing. Traditional implementations of image-based servoing have relied on partial pose estimation, negating much of the advantage gained from using this method. This paper will present an alternative implementation of image-based servoing for approaching objects on the ground plane, which disposes of the pose estimation step and fully relies only on image features. To evaluate the suitability of both visual servoing methods to F-180, the task of docking with the ball is used as a basis of the investigation.

Simulation has become an essential part in the development process of autonomous robotic systems. In the domain of robotics, developers often are confronted with problems like noisy sensor data, hardware malfunctions and scarce or temporarily inoperable hardware resources. A solution to most of the problems can be given by tools which allow the simulation of the application scenario in varying degrees of abstraction and the suppression of unwanted features of the domain (like e.g. sensor noise). The RoboCup scenario of autonomous mobile robots playing soccer is one such domain where the above mentioned problems typically arise.In this work we will present a flexible simulation platform for the RoboCup F2000 league developed as a joint open source project by the universities of Freiburg [13] and Stuttgart [8] which achieves a maximum degree of modularity by a plugin based architecture and allows teams to easily develop and share software modules for the simulation of different sensors, kinematics and even complete player behaviors.Moreover we show how plugins can be utilized to implement benchmark tasks for multi robot learning and give an example that demonstrates how the generic plugin approach can be extended towards the implementation of hardware independent algorithms for robot localization.

In this paper, a simulator of environment and measurement that considers camera characteristics is developed mainly for RoboCup four legged robot league. The simulator introduces server/client system, and realizes separation of each robot’s information, introduction of each robot’s difference and distribution of processes. For producing virtual images, the simulator utilizes OpenGL and considers the effects of blur by lens aberration and so on, random noise on each pixel, lens distortion and delayed exposure for each line of CMOS device. Some experiments show that the simulator imitates the real environment well, and is a useful tool for developing algorithms effectively for real robots.

We present a modular approach to model multi-agent simulations in 3D environments. Using this approach, we implemented a generic simulator which is totally decoupled from the actual simulation it performs. We believe that for Soccer Simulation League a transition to 3D states exiting new research problems and equally makes it more attractive to watch for spectators. We are proposing to use our framework as basis for a next generation Soccer Server.

The educational experiences from robot contest of entry, junior and advance level are presented based on guided constructionism approach in education that combines hands on guidance with hands-on experience. The aim of the competition as a whole are to allow the student to (i) conceptualise the robot (ii) manage the non-deterministic characteristic of the environment and (iii) manage integrated hardware and software development projects. Indeed with this knowledge the student should be able to win a number of international robot tournaments.

In the development of robotic systems, an interactive software platform plays an important role for control design and parameter optimisation. This paper presents a modular approach to the development of a hybrid software platform for Sony quadruped robots. Such a platform consists of an overhead vision system, a Sony AIBO robot and a desktop PC, and is designed to be interactive in order to speed up the development of various algorithms for object recognition and gait generation. Based on this platform, both the colour segmentation algorithm and the gait control algorithm are investigated. The experimental results are presented to show its operation and good performance.

In this paper, the problem concerning how to coordinate concurrent behaviors, when controlling autonomous mobile robots (AMRs), is investigated. We adopt a FSM (finite state machine)-based behavior selection method to solve this problem. It is shown how a hybrid system for an AMR can be modeled as an automaton, where each node corresponds to a distinct robot state. Through transitions between states, robot can coordinate multiple behaviors easily and rapidly under dynamic environment. As an illustration, a soccer task was finished by an AMR system with this method. The robot performed well in the soccer games and won the game in the end.

This paper presents a full description of the design of a trot locomotion that has been implemented on AIBO quadrupeds in the Sony legged league. This work is inspired by the UNSW achievements in RoboCup 2000 and 2001 in Melbourne and Seattle. The French team rebuilt a complete trot locomotion from scratch, and introduced special features that differ from the Australian original design. Many papers have already been dedicated to the work on quadruped locomotion [2-4]. However they do not detail all the parts of the design.

In this article we will concentrate on the communication problems in a multi-agent system, operating within the ’RoboCupRescue’ Simulator system. To cope with the limited communication between the center and the agents in the field, we separate the communication in two layers that focus on synchronizing world models with different levels of detail, responsiveness and range. In this article we will explain the requirements and methods used in the high-level communication that distributes summaries of the current situation in different sectors of the map.

Pseudo-local vision system, which simulates visual information derived from an on-board camera of mobile robot based on a ceiling camera image, is proposed. It consists of a vision server and a client module which communicate with each other in the SoccerServer-like protocol. An image processing module for the on-board camera in the control program is replaced with this system. The simulated visual information is not a two-dimensional image data but a one-dimensional array which represents the nearest edge in each direction around the robot. However, it contains some of essential information of the on-board camera image. This concept was implemented for our robot system for the RoboCup Small-Size League. The server can transmit the edge data to 10 clients 30 times per 1 second. The time lag between grabbing image on the server and extracting visual information on the client is about 10[ms].

We present a method to build a hypothesis on the condition of the environment in which a robotic multi-agent team moves. Initially the robots have a default assumption about the conditions of the floor and on how moving under these condition works. For certain parts of the environment however, the default assumption may be wrong and moving around does not work in the expected way. Now the robotic team builds a hypothesis on the conditions of the yet unvisited part of the environment in a way similar to computing a diagnosis for electrical circuits. Resources can be saved by avoiding areas that possibly also contain obstacles.

We propose a new method of self-localization using two landmarks and dead reckoning for a soccer robot equipped with an omni-directional camera as a local vision sensor. This method requires low computational cost. Thanks to rapid process, the system can afford to run multiple localization process in parallel resulting robust and accurate localization. An experimental result in the field of Robocup Middle-size league indicates that the approach is reliable.

An obstacle avoidance approach is introduced that has dynamic active regions. The dynamic regions are adapted to the current speed of the robot and there are different active regions used, one for speed reduction and one of turning away from obstacles. The overall strategy of this approach is that the robot can drive with high speed which will be reduced in front of an obstacle in order to do a sharper turn.

Modeling agents’ behavior has always been a challenge in multiagent systems. In a competitive environment, predicting future behaviors of opponents helps to make plans to confront their actions properly. We have used the RoboCup soccer server environment [1] to design a coach, capable of analyzing simulated soccer games and making decisions to improve teammate players’ behavior during the games. We will introduce our “Opponent Pass Modeling” method which makes a model of opponent team’s passing behavior to guard opponent players and as a result, to improve the defending behavior of our team. We will also describe a new approach to evaluate coach algorithms using soccer server log-files and LogCoach tool.

This paper describes a topological navigation system, based on the description of key-places by a reduced number of parameters that represent images associated to specific locations in configuration space, and the application of the developed system to robotic soccer, through the implementation of the developed algorithms to RoboCup Middle-Size League (MSL) robots, under the scope of the SocRob project (Soccer Robots or Society of Robots). A topological map is associated with a graph, where each node corresponds to a key-place. Using this approach, navigation is reduced to a graph path search. Principal Components Analysis was used to represent key-places from pre-acquired images and to recognize them at navigation time. The method revealed a promising performance navigating between key-places and proved to adapt to different graphs. Furthermore, it leads to a robot programming language based on qualitative descriptions of the target locations in configuration space (e.g., Near Blue Goal with the Goal on its Left). Simulation results of the method application are presented, using a realistic simulator.

In this paper, we propose a reinforcement learning method called a fuzzy Q-learning where an agent determines its action based on the inference result by a fuzzy rule-based system. We apply the proposed method to a soccer agent that intercepts a passed ball by another agent. In the proposed method, the state space is represented by internal information the learning agent maintains such as the relative velocity and the relative position of the ball to the learning agent. We divide the state space into several fuzzy subspaces. A fuzzy if-then rule in the proposed method represents a fuzzy subspace in the state space. The consequent part of the fuzzy if-then rules is a motion vector that suggests the moving direction and velocity of the learning agent. A reward is given to the learning agent if the distance between the ball and the agent becomes smaller or if the agent catches up with the ball. It is expected that the learning agent finally obtains the efficient positioning skill.

This paper deals with an intelligent control of an autonomous mobile robot, which can adapt to dynamical environments. RoboCup soccer robots are chosen as demonstration targets. An important issue in robotic soccer is how to respond to dynamical environment. This study presents an intelligent control method based on System-Life concept for the autonomous mobile robot system. In other words, it is presented how to design the activating controller, sensing element, processing element and so on. Furthermore, Kalman filter and Euler’s method are applied to estimate the motion of the ball. First, the soccer robot system is developed based on System-Life concept. Second, the intelligent control method is applied to them. Finally the experiment is carried out on RoboCup Soccer field. It is demonstrated the field player succeeds in approaching moving ball and goalkeeper prevents moving ball from netting. The validity of the proposed method is verified.

We propose a hierarchical multi-module leaning system based on self-interpretation of instructions by coach. The proposed method enables a robot to decompose (i) a long term task which needs various kinds of information into a sequence of short term subtasks which need much less information through its self-interpretation process for the instructions given by coach, (ii) to select sensory information needed to each subtask, and (iii) to integrate the learned behaviors to accomplish the given long term task. We show a preliminary result of a simple soccer situation in the context of RoboCup.

This paper describes work in progress to enable a real robot to recreate trail following of ants engaged in pheromone-reinforced recruitment to food gathering. Specifically, it is proposed that development of a set of macro-behaviours for creating and following a trail can be achieved by use of micro-behaviours in a simulated environment to develop a novel neural architecture – the Neural Nest – for learning without explicit representation. A simulated ’neural nest’ has been tested to determine the feasibility for ant colonies to encode higher-level behaviours for controlling a physical robot. In our experiments, the emergent behaviour from reinforcement of interactions between unsupervised simple agents, can allow a robot to sense and react to external stimuli in an appropriate way, under the control of a non-deterministic pheromone trail following program. Future work will be to implement the architecture entirely on the physical robot in real time.

Due to the complexity and sophistication of the skills needed in real world tasks, the development of autonomous robot controllers requires an ever increasing application of learning techniques. To date, however, learning steps are mainly executed in isolation and only the learned code pieces become part of the controller. This approach has several drawbacks: the learning steps themselves are undocumented and not executable.In this paper, we extend an existing control language with constructs for specifying control tasks, process models, learning problems, exploration strategies, etc. Using these constructs, the learning problems can be represented explicitly and transparently and, as they are part of the overall program implementation, become executable. With the extended language we rationally reconstruct large parts of the action selection module of the agilo2001 autonomous soccer robots.

This work describes an investigation to reduce positioning error of 3 wheel middle size robot by using a modified odometry system. In this technique the positioning sensor (shaft encoders) are mounted on 3 free-running wheels so the slippage of the driving wheels does not affect the measurements of the sensors. This will result in decreasing the cumulative error of the system. This mechanism accompanying by omnidirectional vision system presents reliable and accurate self-localization method for any 3 wheel driving robot. Experimental results have shown performance improvement up to 86% in orientation error and 80% in position error.

Since texture is a fundamental character of images, it plays an important role in visual perception, image understanding and scene interpretation. This paper presents a texture-based pattern recognition scheme for Sony robots in the RoboCup domain. Spatial frequency domain algorithms are adopted and tested on a PC while simple colour segmentation and blob-based recognition are implemented on real robots. The experimental results show that the algorithms can achieve good recognition results in the RoboCup Pattern Recognition challenge.

At present, various kinds of robots such as AIBO, ASIMO and etc, are available in public. However, the development of robots is still having some difficulties since of their complexity, continual changes of environments, limitation of resources and etc. To overcome this problem, robot developers often use the simulator that allows to program and test robots’ program effectively under ideal environmental conditions where specified various conditions can easily be reproduced. It is still difficult to realize the simulator regardless of its usefulness, because the cost of simulator implementation seems the unexpected cost in the development of robots. As a result, it is need to realize the open robot simulation environment in which any kind of robots can be simulated. This paper focuses on vision-based robot simulation environment and describes a method to construct it. Finally, we implemented a simulator for Robocup Sony 4-Legged League by using this method.

We propose a novel agent framework to describe behaviors of the general public in rescue simulations and implement an application for “Risk-Communication for disaster rescur”. Conventional agent description languages are designed to model intellectual behaviors of human that solve a task to achieve a single goal. In a disaster situation, however, it is difficult to model civilians’ behaviors such as goal-oriented problem-solving. Instead of such a formalization, we introduce the “Parallel Scenario Description” approach that models agents’ behavior as an action pattern or plan of situations. We call these “Scenarios”. In the proposed framework, behaviors are divided into multiple scenarios for each goal by Posit and Posit operator, in which behavior rules are grouped based on situations where the rules are active. The problem solver PS2 constructs a rule-set of behavior dynamically according to the situation of the environment and the agent’s state. The framework is implemented as civilian agents for RoboCupRescue Simulation to adapt to a general civilian simulation. Moreover, we implemented refuge simulation for disaster rescue simulations to realize “Risk-Communication”.

RoboCup Advanced 3D Monitor is a three-dimensional application for visualizing and debugging games of the RoboCup Soccer Simulation League. This paper discusses the issues pertaining the implementation of this monitor using OpenGL, a standard API for rendering high-performance 3D graphics. Our application provides a true 3D soccer game experience maintaining a healthy balance of realistic animation features and high-speed rendering achieved by the implementation of specific computer graphics techniques. Besides its main usefulness as a visualization tool, this monitor may be used as a supporting tool for the development of other robotics techniques. To illustrate this, two of such techniques are discussed here: sensor fusion and Markov localization methods.

The activities of search and rescue of victims in large-scale disasters are not only highly relevant social problems, but pose several challenges from a scientific standpoint. In this context, the RoboCup-Rescue project focused on the problems of bringing aids immediately after a large disaster, and aims at creating system based on AI and Robotics technologies, where heterogeneous agents (software, robots, human beings) interact in a cooperative manner.In this paper we present the achievements of a research project, based on the RoboCup Rescue simulator, carried out in Italy in collaboration with the Italian Fire Department. The overall project goal is to devise tools to allow monitoring and supporting decisions which are needed in a real-time rescue operation in a large scale disaster, and to provide a methodology for evaluation of multi-agent system which considers not only the efficiency of a system, but also its robustness when conditions in the environment change, as well as other features, such as the ability to acquire a precise and coherent representation of the disaster scenario.

One of the main challenges in RoboCup is to maintain a high level of speed and accuracy in decision making and performing actions by the robot players. Although we might be able to use complicated hardware and software on the robots to achieve the desired accuracy, but such systems might not be applicable in real-time RoboCup environment due to their high processing time. This is quite serious for the robots equipped with more than one vision systems.To reduce the processing time we developed some basic ideas that are inspired by a number of features in the human vision system. These ideas included efficient need-based vision, that reduces the number of objects to be detected to a few objects of interest with the minimum needed accuracy, introduction of static and dynamic regions of interest, which proposes the most probable areas to search for an object of interest, an experimentally reliable method for color segmentation in variable illumination situation, and finally, the usage of some domain specific knowledge that is used in detecting and tracking a unique safe point on the ball.We have implemented these methods on RoboCup environment and satisfactory results were obtained.

Coordination, planning, and reactivity are important for successful teams of autonomous robots, in dynamic adversarial domains. In this paper, we propose a fuzzy cognitive model to integrate coordination, planning and reactive behaviors in a team of cooperating robots. In our architecture, behavioral modules are used as high-level macro-actions that compose structured plans defined by a flexible multi-agent coordination system. The use of an unifying cognitive model provides an effective tool for seamless integration of reactive and deliberative components in the robots, gaining as much as possible from the presence of different skills and competencies in the team. The control model is designed to be tuned and adapted on-line so that the team strategies and the role of robots in the control schemata can be automatically modified to face different opponent teams, and changes in robot capabilities.

This paper outlines ideas for establishing within RoboCup a league geared toward, and limited to, undergraduate students. Veterans of RoboCupJunior are outgrowing the league as they enter college and this has motivated us to develop a league especially for undergraduate students – the ULeague. The design of the league, presented here, is based on a simplied setup of the Small-size league by providing standard Vision and Communication packages.

We shortly review a mobile robot localization method for known 2D environments, which we proposed in previous works; it is an evidence accumulation method where the complexity of working on a large grid is reduced by means of a multi-resolution scheme. We then elaborate a framework to define a set of weights which takes into account the different amount of information provided by each perception, i.e. sensor datum. The experimental activity presented, although the approach is independent on the sensory system, is currently based on perceptions coming from omnidirectional vision in an indoor environment.

Coordinated action for a team of robots is a challenging problem, especially in dynamic, unpredictable environments. In the context of robot soccer, a complex domain with teams of robots in an adversarial setting, there is a great deal of uncertainty in the opponent’s behavior and capabilities. We introduce the concept of a play as a team plan, which combines both reactive principles, which are the focus of traditional approaches for coordinating actions, and deliberative principles. We introduce the concept of a playbook as a method for seamlessly combining multiple team plans. The playbook provides a set of alternative team behaviors which form the basis for our third contribution of play adaptation. We describe how these concepts were concretely implemented in the CMDragons robot soccer team. We also show empirical results indicating the importance of adaptation in adversarial or other unpredictable environments.

Reinforcement learning has been successfully applied to several subtasks in the RoboCup simulated soccer domain. Keepaway is one such task. One notable success in the keepaway domain has been the application of SMDP Sarsa(λ) with tile-coding function approximation [9]. However, this success was achieved with the help of some significant task simplifications, including the delivery of complete, noise-free world-state information to the agents. Here we demonstrate that this task simplification was unnecessary and further extend the previous empirical results on this task.

This paper describes a walking gait for a humanoid robot with a distributed control system. The motion for the robot is calculated in real time on a central controller, and sent over CAN bus to the distributed control system. The distributed control system loosely follows the motion patterns from the central controller, while also acting to maintain stability and balance. There is no global feedback control system; the system maintains its balance by the interaction between central gait and “soft” control of the actuators. The paper illustrates a straight line walking gait and shows the interaction between gait generation and the control system. The analysis of the data shows that successful walking can be achieved without maintaining strict local joint control, and without explicit global balance coordination.

This paper describes a method to reduce the effects of the system immanent control delay for the RoboCup small size league. It explains how we solved the task by predicting the movement of our robots using a neural network. Recently sensed robot positions and orientations as well as the most recent motion commands sent to the robot are used as input for the prediction. The neural network is trained with data recorded from real robots.We have successfully field-tested the system at several RoboCup competitions with our FU-Fighters team. The predictions improve speed and accuracy of play.

A framework for asynchronous stochastic linear control theory is introduced using a simple example motivated by the earlyRoboCup soccer server dynamics. Worst and average case scenarios are studied and it is demonstrated that they fit smoothly into the framework of standard synchronous control theory.

Case Based Reasoning is a feasible approach for recognizing and predicting behavior of agents within the RoboCup domain. Using the method described here, on average 98.4 percent of all situations within a game of virtual robotic soccer have been successfully classified as part of a behavior pattern. Based on the assumption that similar triggering situations lead to similar behavior patterns, a prediction accuracy of up to 0.54 was possible, compared to 0.17 corresponding to random guessing. Significant differences are visible between different teams, which is dependent on the strategic approaches of these teams.

Robot Soccer involves planning at many levels, and in this paper we develop high level planning strategies for robots playing in the RoboCup Four-Legged League using case based reasoning. We develop a framework for developing and choosing game plays. Game plays are widely used in many team sports e.g. soccer, hockey, polo, and rugby. One of the current challenges for robots playing in the RoboCup Four-Legged League is choosing the right behaviour in any game situation. We argue that a flexible theoretical model for using case based reasoning for game plays will prove useful in robot soccer. Our model supports game play selection in key game situations which should in turn significantly advantage the team.

The aim of this article is to describe our experience in the participation and organization of robot contests, and to show how these actions have increased the activities on robotics in Chile. We describe the annual Latin American IEEE Robotics Competition, we present the IEEE Latin American Robotics Council, we explain our participation in RoboCup, and we present our activities concerning robotic courses for children.

Robots interacting with other agents in dynamic environments require robust knowledge management capabilities if they are to communicate, learn and exhibit intelligent behaviour. Symbol grounding involves creating, and maintaining, the linkages between internal symbols used for decision making with the real world phenomena to which those symbols refer. We implement grounding using ontologies designed for the Semantic Web. We use SONY AIBO robots and the robot soccer domain to illustrate our approach. Ontologies can provide an important bridge between the perceptual level and the symbolic level and in so doing they can be used to ground sensory information. A major advantage of using ontologies to ground sensory and symbolic information is that they enhance interoperability, knowledge sharing, knowledge reuse and communication between agents. Once objects are grounded in ontologies, Semantic Web technologies can be used to access, build, derive, and manage robot knowledge.