RoboCup 2011: Robot Soccer World Cup XV

The RoboCup simulation league is traditionally the league with the largest number of teams participating, both at the international competitions and worldwide. 2011 was no exception, with a total of 39 teams entering the 2D and 3D simulation competitions. This paper prese- nts the champions of the competitions, WrightEagle from the University of Science and Technology of China in the 2D competition, and UT Austin Villa from the University of Texas at Austin in the 3D competition.

SKUBA is a winner from RoboCup 2011, Turkey. The aim of this paper is to explain the basic concepts of the design of the robot hardware and the AI system. The robot mechanics are explained in detail along with the electronic boards. Protection circuits are added to make the robot more robust. A torque controller is implemented as low-level controller to reduce the effect of surfaces. The high level AI system is separated into three major modules: predictor/tracker, strategy, and control module. The tracking algorithm and the Kalman Observer are implemented in the predictor/tracker module. The strategy module takes care of playing and evaluates the probability of each play against the opponent. Finally, the control module is explained. Path planning algorithms and a modified kinematics equation are implemented in this module in order to make the robot move along the desired trajectory with less velocity error.

After having won the Standard Platform League competitions in 2009 and 2010, the B-Human software already included sophisticated solutions for most relevant subtasks, such as vision, state estimation, and walking. Therefore, the development towards RoboCup 2011 did not focus on replacing specific low-quality components, but was guided by an overall goal: eliminating game delays by more efficient actions and faster reactions to game state changes. This required several changes all over the system. In this paper, we present some of the developments that had the most impact regarding our goal: different ball models and corresponding cooperative ball tracking and retrieval strategies, a path planner as well as new approaches for tackling situations.

Over the past few years, soccer-playing humanoid robots advanced significantly. Elementary skills, such as bipedal walking, visual perception, and collision avoidance have matured enough to allow for dynamic and exciting games. In this paper, the three winning Humanoid League teams from the KidSize, TeenSize, and AdultSize class present their soccer systems. The KidSize winner team DARwIn used the recently introduced DARwIn-OP robot. The TeenSize winner NimbRo used their self-constructed robots Dynaped and Bodo. The AdultSize Louis Vuitton Best Humanoid Award winner CHARLI detail the technology behind the outstanding performance of its robot CHARLI-2.

In this paper, we detail the contributions of our team NimbRo to the RoboCup @Home league in 2011. We explain design and rationale of our domestic service robot Cosero that we used for the first time in a competition in 2011. We demonstrated novel capabilities in the league such as real-time table-top segmentation, flexible grasp planning, and real-time tracking of objects. We also describe our approaches to human-robot cooperative manipulation and 3D navigation. Finally, we report on the use of our approaches and the performance of our robots at RoboCup 2011.

As RoboCupJunior reached a decade mark in 2011, we feel the need for examining the current situation after 12 revisions and modifications to the league rules and structures since its launch in 2000. RoboCupJunior International is now attracting over 250 teams involving approximately 1,000 students originating from more than 30 countries. This paper aims to report on the progress achieved thus far, both technologically and educationally, and the issues currently addressed, together with suggestions for the future of RoboCupJunior.

In this paper, we propose a new approach to compliant task-space control for high degree-of-freedom manipulators driven by position-controlled actuators. The actuators in our approach are back-drivable and allow to limit the torque used for position control. Traditional approaches frequently achieve compliance through redundancy resolution. Our approach not only allows to adjust compliance in the null-space of the motion but also in the individual dimensions in task-space. From differential inverse kinematics we derive torque limits for each joint by examining the contribution of the joints to the task-space motion. We evaluate our approach in experiments with specific motions. We also report on the application of our approach at RoboCup 2010, where we successfully opened and closed the fridge in the RoboCup@Home finals.

Rescue robots are platforms designed to operate in challenging and uneven surfaces. These robots are often equipped with manipulator arms and varying traction arrangements. As such, it is possible to reconfigure the kinematic of robot in order to reduce potential instabilities, such as those leading to vehicle tip-over. This paper proposes a methodology to plan feasible paths through uneven topographies by planning stable paths that account for the safe interaction between vehicle and terrain. The proposed technique, based on a gradient stability criterion, is validated with two of the best known path search strategies in 3D lattices, i.e. the A* and the Rapidly-Exploring Random Trees. Using real terrain data, simulation results obtained with the model of a real rescue robot demonstrate significant improvements in terms of paths that are able to automatically avoid regions of potential instabilities, to concentrate on those where the freedom of exploiting posture adaptation permits generation of optimally safe paths.

In this paper, we present a walking approach for the Nao robot that improves the agility and stability of the robot when walking on a flat surface such as the soccer field used in the Standard Platform League. The gait uses the computationally inexpensive model of an inverted pendulum to generate a target trajectory for the center of mass of the robot. This trajectory is adapted using the observed real motion of the center of mass. This approach does not only allow compensating the inaccuracies in the model, but it also allows for reacting to external perturbations effectively. In addition, the method aims at facilitating a preferably fast walk while reducing the load on the joints.

In robotic soccer the ball is the most crucial factor of the game. It is therefore extremely important for a robot to retrieve it as soon as possible. Thus ball interception is a key behaviour in robotic soccer. However, currently most MSL teams move to the ball position without considering the ball velocity. This often results in inefficient paths described by the robot. This paper presents the CAMBADA solution for a ball interception behaviour based on a uniformly accelerated robot model, where not only the ball velocity is taken into account but also the robot current velocity as well as the robot acceleration, maximum velocity and sensor-action delays are considered. The described work was introduced in the Portuguese robotics open Robótica2009 and RoboCup 2009 and improved the team performance contributing to the first and third places, respectively.

Using a simulation for development and research of robot motions, especially walking motions, has advantages like saving real hardware, being able to replay specific situations or logging various data. Unfortunately research in this area using a simulation depends on transferability of the results to reality, which is not given for common robotic simulators. This paper presents extensions to a basic rigid body physics simulation leading to more realism. Parametrization matching a particular real robot is done using Evolutionary Strategies. Using stable walking and kicking motions as reference for the ES the newly developed MoToFlex simulator is able to reflect typical walking issues which can be observed in reality using different walking motions.

An efficient object perception is a crucial component of a mobile service robot. In this work we present a solution for visual categorization of objects. We developed a prototypic categorization system which classifies unknown objects based on their visual properties to a corresponding category of predefined domestic object categories. The system uses the Bag of Features approach which does not rely on global geometric object information. A major contribution of our work is the enhancement of the categorization accuracy and robustness through a selected combination of a set of supervised machine learners which are trained with visual information from object instances. Experimental results are provided which benchmark the behavior and verify the performance regarding the accuracy and robustness of the proposed system. The system is integrated on a mobile service robot to enhance its perceptual capabilities, hence computational cost and robot dependent properties are considered as essential design criteria.

Perception of the environment is crucial in terms of successfully playing soccer. Especially the detection of other players improves game play skills, such as obstacle avoidance and path planning. Such information can help refine reactive behavioral strategies, and is conducive to team play capabilities. Robot detection in the RoboCup Standard Platform League is particularly challenging as the Nao robots are limited in computing resources and their appearance is predominantly white in color like the field lines.

This paper describes a vision-based multilevel approach which is integrated into the B-Human Software Framework and evaluated in terms of speed and accuracy. On the basis of color segmented images, a feed-forward neural network is trained to discriminate between robots and non-robots. The presented algorithm initially extracts image regions which potentially depict robots and prepares them for classification. Preparation comprises calculation of color histograms as well as linear interpolation in order to obtain network inputs of a specific size. After classification by the neural network, a position hypothesis is generated.

This paper describes a new method of extraction and clustering of edges in images. The proposed method results a graph of detected edges instead of a binary mask of the edge pixels. The developed algorithm contains a sequential pixel-level scan, and a much smaller second and third pass on the results to determine the connectivities. It is therefore significantly faster than Canny edge detector, performing both edge detection and grouping tasks. The method is developed for a RoboCup scenario, however it can also be applied to any other image as long as the prerequisites are met. The paper explains the idea, discusses the prerequisites and finally presents the implementation results and issues.

Many vision-based tasks for autonomous robotics are based on feature matching algorithms, finding point correspondences between two images. Unfortunately, existing algorithms for such tasks require significant computational resources and are designed under the assumption that they will run to completion and only then return a complete result.

Since partial results—a subset of all features in the image—are often sufficient, we propose in this paper a computationally-flexible algorithm, where results monotonically increase in quality, given additional computation time. The proposed algorithm, coined AnySURF (Anytime SURF), is based on the SURF scale- and rotation-invariant interest point detector and descriptor. We achieve flexibility by re-designing several major steps, mainly the feature search process, allowing results with increasing quality to be accumulated.

We contrast different design choices for AnySURF and evaluate the use of AnySURF in a series of experiments. Results are promising, and show the potential for dynamic anytime performance, robust to the available computation time.

This paper presents an algorithm for autonomous online path planning in uncertain, possibly adversarial, and partially observable environments. In contrast to many state-of-the-art motion planning approaches, our focus is on decision making in the presence of adversarial agents who may be acting strategically but whose exact behaviour is difficult to model precisely. Our algorithm first computes a collection of

reachable sets

with respect to a family of possible strategies available to the adversary. Online, the agent uses these sets as

composable behavioural templates

, in conjunction with a particle filter to maintain the current belief on the adversary’s strategy. In partially observable environments, this yields significant performance improvements over state-of-the-art planning algorithms. We present empirical results to this effect using a robotic soccer simulator, highlighting the applicability of our implementation against adversaries with varying capabilities. We also demonstrate experiments on the

NAO

humanoid robots, in the context of different collision-avoidance scenarios.

In recent years, humanoid soccer robots improved considerably. Elementary soccer skills, such as bipedal walking, visual perception, and collision avoidance have matured enough to provide for dynamic and exciting soccer games. While the elementary skills still remain hot research topics, it is time to move forward and address higher level skills, such as motion planning and team play. In this work, we present a new method to generate ball approach trajectories by planning footstep sequences offline and training an online policy to meet the real time requirements of embedded systems with low computational power, as typically used for soccer robots. We compare the results with our current reactive behavior that was used in the last RoboCup competitions and show the improvements we achieved.

The design of a real time and dynamic balanced biped walking gait pattern generator is not trivial due to high control space and inherently unstable motion. Moreover, in the Robocup domain, robots that are able to achieve the goal footstep in a short duration have a great advantage when playing soccer. In this paper, we present a new technique to realize a real time biped walking gait pattern generator on a real robot named Nao. A Zero Moment Point (ZMP) trajectory represented by a cubic polynomial is introduced to connect the goal state (the position and the velocity of the CoG) to the previous one in only one step. To apply the generator on the real robot Nao, we calculate the compensation for two HipRoll joints in a theoretical way by modeling them as elastic joints. The nao of version 3.3 is used in the experiments. The walk is intrinsically omnidirectional. When walking with step duration 180ms, the robot can respond to the high level command in 180ms. The maximum forward speed is around 0.33m/s. The maximum backward speed is around 0.2m/s. The maximum sideways speed is around 0.11m/s. The maximum rotational speed is around 90°/s.

This paper describes an approach to Gaussian mixture filtering which combines the accuracy of the Kalman filter and the robustness of particle filters without sacrificing computational efficiency. Critical approximations of common Gaussian mixture algorithms are analyzed and similarities are pointed out to particle filtering with an extremely low number of particles. Known techniques from both fields are applied in a new combination resulting in a multi-hypotheses Kalman filter which is superior to common Kalman filters in its ability of fast relocalization in kidnapped robot scenarios and its representation of multi-modal belief distributions, and which outperforms particle filters in localization accuracy and computational efficiency.

State estimation is of crucial importance to mobile robotics since it determines in a great measure its ability to model the world from noisy observations. In order to quantitatively evaluate state-estimation methods, the availability of ground-truth data is essential since it provides a target that the result of the state-estimation methods should approximate. Most of the reported ground-truth systems require a complex assembly which limit their applicability and make their set-up long and complicated. Furthermore, they often require a long calibration procedure. Additionally, they do not present measures of their accuracy. This paper proposes a portable laser-based ground-truth system. The proposed system can be easily ported from one environment to other and requires almost no calibration. Quantitative results are presented with the purpose of encouraging future comparisons among different ground-truth systems. The presented method has shown to be accurate enough to evaluate state-estimation methods and works in real time.

In the aftermath of the September 11 attacks, evacuation simulation has the potential for decreasing the amount of damage resulting from disasters, and, in particular, for saving human lives. Agent based simulation provides a platform for computing individual and collective behaviors that occur in crowds. Such simulations have led to proposals for enhanced prompt public evacuation.

For the public, it is desirable to simulate the behavior of evacuation in any building. We propose an everywhere evacuation simulation system. This system provides a software environment that permits evacuation simulation for any location for which plans are provided on the Web. Three-dimensional (3D) models of buildings and geographic information system (GIS) data for areas that are created for everyday purposes such as sightseeing are used as the environments for simulations. The characteristics of humans are set by users, and their evacuation behaviors are simulated with the relationships among them. The results of simulations can be viewed on the Web by allocating heterogeneous agents inside 2D/3D maps of buildings.

The Standard Platform League (SPL) provides an environment that is essentially static; with the exception of other robots and the audience, the area in which a robot is expected to localise itself is quite favourable. However, a large number of the predefined landmarks in the given world model can be perceived as ambiguous in many scenarios, with the prime example being field line markings. In this paper a technique is presented that implicitly disambiguates these detected field line objects in order to use them for localization purposes.

Service robots

could use textual information to perform important tasks, like

product identification

. However,

natural scene text

such as found in household environments can be very arbitrary in terms of size, color, font, layout, symbol repertoire, language, etc. This

large variability

makes robust

text information extraction

extremely difficult. Our work on textual information extraction for gray-scale still images uses

adaptive binarization

,

connected component

classification with a

support vector machine

and filtering based on the proximity of the connected components to their neighbours. The contribution of our approach is the use of a partially

synthetic dataset

for training. This decreases the burden of

ground truth labelling

at the connected component level. Our experiments show that classification generalization on real instances can be attained when training a classifier with synthetic data. We present our results on the ICDAR dataset.

This work describes RoboViz, a new software program designed to assess and develop agent behaviors in a multi-agent system, the RoboCup Soccer Simulation 3D sub-league. RoboViz is an interactive monitor that renders both agent and world state information in a three-dimensional scene. In addition, RoboViz provides programmable remote drawing functionality to agents or other clients that can communicate over a network. The tool facilitates real-time visualization of agents running concurrently on the SimSpark simulator to provide higher-level analysis of agent behaviors not currently possible with existing tools. Provided appropriate hardware, the monitor and debugging tool can produce high-quality stereo vision images. RoboViz is proposed as a replacement for the current SimSpark 3D league monitor to benefit developers as well as elevate public interest in the 3D simulation league, and it has been used officially at the 2011 German Open in Magdeburg, Germany. RoboViz was released in February 2011 as an open-source project under the Apache 2.0 license.

This paper describes a novel approach in formation control for mobile robots. A Nonlinear Model Predictive Controller (NMPC) is used to control the formation of a heterogeneous mobile robots group. The desired formation is formed by an holonomic robot and a nonholonomic robot. The same nonlinear controller is used in both robots with the same cost function. The details of the controller structure are presented in order to track a fixed target departing from different positions in the field avoiding collisions with each other. A soccer robot competition field is used to present the simulations to evaluate the performance of the controller.

Real-time 3D perception of the surrounding environment is a crucial precondition for the reliable and safe application of mobile service robots in domestic environments. Using a RGB-D camera, we present a system for acquiring and processing 3D (semantic) information at frame rates of up to 30Hz that allows a mobile robot to reliably detect obstacles and segment graspable objects and supporting surfaces as well as the overall scene geometry. Using integral images, we compute local surface normals. The points are then clustered, segmented, and classified in both normal space and spherical coordinates. The system is tested in different setups in a real household environment.

The results show that the system is capable of reliably detecting obstacles at high frame rates, even in case of obstacles that move fast or do not considerably stick out of the ground. The segmentation of all planes in the 3D data even allows for correcting characteristic measurement errors and for reconstructing the original scene geometry in far ranges.

Omnidirectional RoboCup MSL robots often use catadioptric vision systems in order to enable 360º of field view. It comprises an upright camera facing a convex mirror, commonly spherical, parabolic or hyperbolic, that reflects the entire space around the robot. This technique is being used for more than a decade and in a similar way by most teams. Teams upgrade their cameras in order to obtain more and better information of the captured area in pixel quantity and quality, but a large image area outside the convex mirror is black and unusable. The same happens on the image centre where the robot shows itself. Some efficiency though, can be improved in this technique by the methods presented in this paper such as developing a new convex mirror and by repositioning the camera viewpoint. Using 3D modelling CAD/CAM software for the simulation and CNC lathe mirror construction, some results are presented and discussed.

In this paper, an efficient path planning method for non-holonomic robots to avoid elliptical obstacles for RoboCup soccer matches is presented. A hydrodynamic flow field is formulated to model obstacles and target locations. Previous research considers the flow about elliptical and plate obstacles as a superposition of multiple flow fields about circular obstacles. The proposed approach utilises the Joukowsky transform to form a path about an elliptical obstacle with a single flow field. It is shown that the resulting motion satisfies

C

 ∞ 

continuity at all times, desired for mobile robots. The application of different obstacle shapes in the context of RoboCup soccer matches is also considered and simulated.

We present NaOISIS, a three-dimensional behavioural simulator for the NAO humanoid robot, aimed at designing and testing physically plausible strategic behaviours for multi-agent soccer teams. NaOISIS brings together features from both physical three-dimensional simulators that model robot dynamics and interactions, and two- dimensional environments that are used to design sophisticated team coordination strategies, which are however difficult to implement in practice. To this end, the focus of our design has been on the accurate modeling of the simulated agents’ perceptual limitations and their compatibility with the corresponding capabilities of the real NAO robot. The simulator features presented in this paper suggest that NaOISIS can be used as a rapid prototyping tool for implementing behavioural algorithms for the NAO, and testing them in the context of matches between simulated agents.

We present a system to automatically recognize facial expressions from static images. Our approach consists of extracting particular Gabor features from normalized face images and mapping them into three of the six basic emotions: joy, surprise and sadness, plus neutrality. Selection of the Gabor features is performed via the AdaBoost algorithm. We evaluated two learning machines (AdaBoost and Support Vector Machines), two multi-classification strategies (Error-Correcting Output Codes and One-vs-One) and two face image sizes (48 x 48 and 96 x 96). Images of the Cohn-Kanade AU-Coded Facial Expression Database were used as test bed for our research. Best results (87.14% recognition rate) were obtained using Support Vector Machines in combination with Error-Correcting Output Codes and normalized face images of 96 x 96.

We investigate mobile telepresence robots to address the lack of mobility in traditional videoconferencing. To operate these robots, intuitive and powerful interfaces are needed. We present CoBot-2, an indoor mobile telepresence robot with autonomous capabilities, and a browser-based interface to control it. CoBot-2 and its web interface have been used extensively to remotely attend meetings and to guide local visitors to destinations in the building. From the web interface, users can control CoBot-2’s camera, and drive with either directional commands, by clicking on a point on the floor of the camera image, or by clicking on a point in a map. We conduct a user study in which we examine preferences among the three control interfaces for novice users. The results suggest that the three control interfaces together cover well the control preferences of different users, and that users often prefer to use a combination of control interfaces. CoBot-2 also serves as a tour guide robot, and has been demonstrated to safely navigate through dense crowds in a long-term trial.

We present a method for enabling an Aldebaran Nao humanoid robot to perceive bumps and touches caused by physical contact forces. Dedicated touch, tactile or force sensors are not used. Instead, our approach involves the robot learning from experience to generate a proprioceptive motor sensory expectation from recent motor position commands. Training involves collecting data from the robot characterised by the absence of the impacts we wish to detect, to establish an expectation of “normal” motor sensory experience. After learning, the perception of any unexpected force is achieved by the comparison of predicted motor sensor values with sensed motor values for each DOF on the robot. We demonstrate our approach allows the robot to reliably detect small (and also large) impacts upon the robot (at each individual joint servo motor) with high, but also varying, degrees of sensitivity for different parts of the body. We discuss current and possible applications for robots that can develop and exploit proprioceptive expectations during physical interaction with the world.

In RoboCup it is important to build up domain knowledge for decision-making. Unfortunately, this is a time-consuming and laborious job. At championships easy adaptability of this domain knowledge can be especially crucial as teams need to be able to change tactics and adjust to opponent behavior as fast as possible. An intuitive interface to the agent is therefore necessary.

In this paper, we present a methodology to automatically populate a domain ontology from natural language text. The resulting populated ontology can then be deployed in a multi-agent system. This automatic transformation of text to knowledge for decision-making thus provides such an intuitive interface to the agents. It is embedded into the broader (up to now) theoretical context of an ontology lifecycle.

We have created a proof-of-concept implementation in the 2D RoboCup Simulation League on the base of tactics descriptions from soccer literature. Experiments show that 71% of tactics are perfectly transformed and 86% of the actions are executed correctly in terms of geometric relations.

This paper reviews the proposed opponent modeling algorithms within the soccer simulation domain. RoboCup soccer simulation 2D is a rich multi agent environment where opponent modeling plays a crucial role. In multi agent systems with adversarial and cooperative agents, team agents should be adapted to the current environment and opponent in order to propose appropriate and effective counteractions. Predicting the opponent’s future behaviors during competition allows for more informed decisions. We divide opponent modeling into two categories of individual agent behaviors and team behaviors. Individual behaviors concern modeling the low-level behaviors of individual opponent agents, however in team behaviors, the high-level strategy of the entire team like formation, offensive and defensive system, is recognized. Several methods have been proposed to create different models of opponents to improve the performance of teams in an essential aspect. In this paper, we review the approaches to the problem of opponent modeling published from 2000 to 2010.

In this paper, a concept for articulated rigid body state estimation is proposed. The articulated body, for instance a humanoid robot, is modeled in a maximal coordinate formulation and the articulations between the rigid bodies as nonlinear position and linear motion constraints. At first, the individual state of each particular rigid body is estimated with a Kalman filter, which leads to an unconstrained state estimate. Subsequently, the correct state estimate for the articulated rigid body is derived by projecting the unconstrained estimate onto the constraint surface.

Robotic soccer vision has been a major research problem in RoboCup and, even though many progresses have been made so that, for example, games now can run without many constraints on the lighting conditions, the problem has not been completely solved and on-site camera calibration is always a major activity for RoboCup soccer teams. While different robotic soccer vision and object perception techniques continue to appear in the RoboCup Soccer League, there is a lack of quantitative evaluation of existing methods.

Since we believe that a quantitative evaluation of soccer vision algorithms will led to significant advances in the performance on perception and on the entire soccer task, in this paper we propose a benchmarking methodology for evaluating robotic soccer vision systems. We discuss the main issues of a successful benchmarking methodology: (i) a large and complete data base or data sets with ground truth; (ii) a public repository with data sets, algorithms and implementations that can be dynamically updated and (iii) evaluation metrics, error functions and comparison results.

This paper presents the development of an object recognition and location system using the Microsoft Kinect

TM

, an off-the-shelf sensor for videogames console Microsoft Xbox 360

TM

which is formed by a color camera and depth sensor. This sensor is capable of capturing color images and depth information from a scene. This vision system uses

a

) data fusion of both color camera and depth sensor to segment objects by distance;

b

) scale-invariant features to characterize and recognize objects; and

c

) camera’s internal parameters combined with depth information to locate objects relative to the camera point of view. The system will be used along with a robotic arm to grab objects.

‘

Realtime simulation of RoboCup small size soccer robots is a challenging task due to the high frame rate of input vision data and complex dynamic model of robots. In this paper we describe a new multi-robot 3D simulator for small size robot soccer domain named ‘grSim’. In order to decrease the model complexity and increase simulation speed, a simplified dynamic model for omni wheels is implemented. grSim has a distributed architecture, feature-rich user interface and supports all aspects of a small size robot soccer game, thus it can completely replace all hardware used by teams during software development. grSim can help software/AI developers design smarter SSL robot teams.

This paper presents a solution to the biped locomotion problem. The robot used for the experiments is robot Nao by Aldebaran Robotics and it is simulated in Webots mobile robot simulator. Our method of solution does not requires the dynamic model of the robot, thus making this approach usable to other biped robots. For faster results the number of degrees of freedom is kept low, only six are used. The walking gait is generated using Central Pattern Generators with limit-cycle oscillators. For the oscillator connection weights required for synchronization, a genetic algorithm is implemented. Our solution is generated automatically and the best results allow the robot to walk twice as fast as the Aldebaran’s webots walk and four times faster than the default walk in Robotstadium.

This paper deals with the problem of energy constrained gait optimization for bipedal walking. We present a solution to this problem obtained by applying a recently introduced heuristic method, the Alliance Algorithm (AA), and compare its performance against a Genetic Algorithm (GA). We show experimentally that the intrinsic ability of the AA to handle hard constraints enables it to find solutions significantly better than the GA. Also with the constraint removed the AA show more reliable optimization results. Finally, we show that the final gait obtained through this method outperforms most solutions to this problem presented in previous works, in terms of walking speed.

We have proposed a new concept of “safety region” which we use to measure the position of the defense robots[5]. It is defined as a region that the teammate robot(s) can defend the goal when an opponent robot shoots the ball from the inside of the safety region while teammate robots are positioned according to their defense strategy.

Since it is difficult to obtain the accurate safety region in a short time, we need an algorithm that computes an approximate safety region in real time. We proposed such algorithm in the previous paper[5]. However, the safety region obtained by the algorithm is not accurate enough. Therefore, in this paper, we propose an improved algorithm to compute the approximate safety region. We have achieved 95% accuracy and less than 1 msec of computation time, which is adequate for our RoboCup application. We also propose a defense strategy based on the safety region considering the positions of the opponent robots and the pass direction. The achieved results indicate accurate performance for determining the positions of the defense robots.

Following obstacle free paths towards the ball and avoiding opponents while dribbling are key skills to win soccer games. These tasks are challenging as the robot’s environment in soccer games is highly dynamic. Thus, exact plans will likely become invalid in the future and continuous replanning is necessary. The robots of the RoboCup Standard Platform League are equipped with limited computational resources, but have to perform many parallel tasks with real-time requirements. Consequently, path planning algorithms have to be fast.

In this paper, we compare two approaches to reduce the planning time by using a local-multiresolution representation or a log-polar representation of the environment. Both approaches combine a detailed representation of the vicinity of the robot with a reasonably short planning time. We extend the multiresolution approach to the time dimension and we predict the opponents movement by projecting the planning robot’s intentions.

One of the most important tasks on robot soccer is localization. The team robots should self-localize on the 18 x 12 meters soccer field. Since a few years ago the soccer field has increased and the corner posts were removed and that increased the localization task complexity. One important aspect to take care for a proper localization is to find out the robot orientation. This paper proposes a new technique to calculate the robot orientation. The proposed method consists of using a histogram of white-green transitions (to detect the lines on the field) to know the robot orientation. This technique does not take much computational time and proves to be very reliable.

Ground truth detection systems can be a crucial step in evaluating and improving algorithms for self-localization on mobile robots. Selecting a ground truth system depends on its cost, as well as on the detail and accuracy of the information it provides. In this paper, we present a low cost, portable and real-time solution constructed using the Microsoft Kinect RGB-D Sensor. We use this system to find the location of robots and the orange ball in the Standard Platform League (SPL) environment in the RoboCup competition. This system is fairly easy to calibrate, and does not require any special identifiers on the robots. We also provide a detailed experimental analysis to measure the accuracy of the data provided by this system. Although presented for the SPL, this system can be adapted for use with any indoor structured environment where ground truth information is required.

The design and implementation of a robot brain often requires making decisions between different modules with similar functionality. Many implementations and components are easy to create or can be downloaded, but it is difficult to assess which combination of modules work well and which does not. This paper discusses a reinforcement learning mechanism where the robot is choosing between the different components using empirical feedback and optimization criteria. With the interval estimation algorithm the robot deselects poorly functioning modules and retains only the best ones. A discount factor ensures that the robot keeps adapting to new circumstances in the real world. This allows the robot to adapt itself continuously on the architecture level and also allows working with large development teams creating several different implementations with similar functionalities to give the robot biggest chance to solve a task. The architecture is tested in the RoboCup@Home setting and can handle failure situations.

Robots are used for Urban Search and Rescue to assist rescue workers. To enable the robots to find victims, they are equipped with various sensors including thermal, video and depth time-of-flight cameras, and laser range-finders. We present a method to enable a robot to perform this task autonomously. Thermal features are detected using a dynamic temperature threshold. By aligning the thermal and time-of-flight camera images, the thermal features are projected into 3D space. Edge detection on laser data is used to locate holes within the environment, which are then spatially correlated to the thermal features. A decision tree uses the correlated features to direct the autonomous policy to explore the environment and locate victims. The method was evaluated in the 2010 RoboCup Rescue Real Robots Competition.

Visual object detection in robot soccer is fundamental so the robots can act to accomplish their tasks. Current techniques rely on manually highly polished definitions of object models, that lead to accurate detection, but are quite often computationally inefficient. In this work, we contribute an efficient object detection through regression (ODR) method based on offline training. We build upon the observation that objects in robot soccer are of a well defined color and investigate an offline learning approach to model such objects. ODR consists of two main phases: (i) offline training, where the objects are automatically labeled offline by existing techniques, and (ii) online detection, where a given image is efficiently processed in real-time with the learned models. For each image, ODR determines whether the object is present and provides its position if so. We show comparing results with current techniques for precision and computational load.

The ambitious goal being pursued by researchers participating in the RoboCup challenge [8] is to develop a team of autonomous humanoid robots that is capable of winning against a team of human soccer players. An important step in this direction is to actively utilise human coaching to improve the skills of robots at both tactical and strategic levels. In this paper we explore the hypothesis that embedding a human into a robot’s body and allowing the robot to learn tactical decisions by imitating the human coach can be more efficient than programming the robot explicitly. To enable this, we have developed a sophisticated HRI system that allows a human to interact with, coach and control an Aldebaran Nao robot through the use of a motion capture suit, portable computing devices (iPhone and iPad), and a head mounted display (which allows the human controller to experience the robot’s visual perceptions of the world). This paper describes the HRI-Coaching system we have developed, detailing the underlying technologies and lessons learned from using it to control the robot. The system in its current stages shows high potential for human-robot coaching, but requires further calibration and development to allow a robot to learn by imitating the human coach.

Robot motion coordination and cooperative sensing are nowadays two important and inter-related components of multi-robot cooperation. Particularly, when concerning motion coordination, distance information plays a very important role in mobile robotics. In this work, we investigate a new solution based on ad-hoc communication without global knowledge, particularly clock synchronisation, to measure distance between mobile units and to share that information. In order to improve ranging, medium throughput, and application predictability, we propose using a synchronisation protocol that keeps transmissions in the team as separated as possible in time, independently of the topology. Results show around 3.3 times reduction in the number of failed ranges without external interference and an order of magnitude reduction in the asymmetries among the nodes concerning the number of failed ranges when using the proposed synchronisation protocol.

This paper proposes a novel architecture for real-time 3D ball trajectory estimation with a monocular camera in Middle Size League scenario. Our proposed system consists on detecting possible multiple ball candidates in the image, that are filtered in a multi-target data association layer. Validated ball candidates have their 3D trajectory estimated by Maximum Likelihood method (MLM) followed by a recursive refinement obtained with an Extended Kalman Filter (EKF). Our approach was validated in real RoboCup scenario, evaluated recurring to ground truth information obtained by alternative methods allowing overall performance and quality assessment.