The Brainstormer Tribots are the 'real world' alter ego of our Brainstormers 2D simulation league team with the ambition of bringing neural reinforcement learning to autonomous robots. In 2007, German Chancellor Dr. Angela Merkel (see on the right together with MR and Christian Wulff) became the first robot godparent to 'Tribot No. 4'.
The Brainstormer Tribots RoboCup team was established in 2002 and has been the world champion in the RoboCup MiddleSizeLeague in 2006 and 2007. Based on our experiences in the RoboCup simulation league, we started with a student project group at the University of Dortmund under the supervision of Prof. Dr. Martin Riedmiller and Roland Hafner, Sascha Lange, Martin Lauer . The team moved to the University of Osnabrück in 2003.
The name Brainstormers Tribots is a combination of our simulation team (the Brainstormers) and the midsize team (the Tribots).
Our robots are based on an omnidirectional drive, an omnidirectional camera, a perspective camera to allow stereo vision, and odometers as sensors. They work completely autonomously. We have successfully participated in many international robotic soccer contests, and we became world champion twice (2006 and 2007) and European champion four times (2004-2007).
This video shows the evolution of the Brainstormers Tribots, a team participating in the RoboCup MidSize league. Starting in 2003 with the ambition to develop self-learning controllers for autonomous robots, the Brainstormers were one of the most successful teams worldwide in the years 2005-2008, winning two world championships in 2006 and 2007.
Neural Reinforcement learning agent based on Neural Fitted Q Iteration (NFQ) learns to dribble a ball directly on a real robot. The robot learns completely from scratch, just by experiencing the learning signal whether it moves into the right direction and the ball stays close.
From Minute 07:00, you can see my interview with German TV reporter Ina Müller. We discuss the basics of tribots and winning the RoboCup for Germany with our Osnabrück-based team.
The clips were published as part of her 2008 tour series through northern Germany.
Basically every Tribot consists of an omnidirectional three-wheel chassis, a tri-motor-conroller, a JVC-Laptop and an omnidirectional camera system. A pneumatic kicking device is attached to the front of the robot. Communication between the robots is established by wlan-interfaces.
All soccer robots are based on a control software that was completely developed at the University of Osnabrück. It analyses camera images, fuses all kinds of information into a consistent model of the world, creates the robot behavior, and controls the motors. All of these calculations are repeated 30 times per second. The program is robust, reliable, and very quick. It consists of 100,000 lines of code and has been developed since 2003 by more the 20 students and researchers. It is written in C and C++ for Linux.
First of all, the robot must analyse the camera images to get information about its surroundings. The images themselves are arrays of colored pixels. However, the robot is interested in the objects (like goals, ball, other robots) which are shown in these images, not in the pixels themselves. Therefore, the pictures are searched for areas of characteristic colors. Red areas potentially show the ball, white areas potentially refer to white field markings, and black areas might show other robots. Afterwards, the form and size of all areas are checked. If color, form, and size are adequate, the program assumes an area to refer to a certain object in the world.
After having found all objects, the robot calculates the position of each object in the world. This is done using calibrated cameras, i.e. before a game is started, every robot calculates a table in which it stores the distance at which an object is placed in the world if it is seen in the image at a certain position. Concerning the ball, the determination of its position is more difficult since the ball can be chipped so that it is not always on the ground. To compute its position the robot uses a second camera (stereo vision). By comparing the position at which the ball is seen in both cameras the robot can determine the position as well as the height of the ball.
Read more about the vision System here.
The next step of calculation is to determine the robot's own position. The robot analyses the white field markings which have been recognized in the camera images. It determines the position on the field at which the lines it has seen fit in the best way to the lines it expects to see. For example, if the robot sees a circle of white lines in the camera image, it is very likely that the robot is close to the center circle of the field.
Additionally, the robot calculates its own velocity comparing the latest positions it has calculated from the camera images. Certainly, the robot must also know the position of the ball and its movement. For this, as well, the robot uses the latest positions at which it has seen the ball.
When all information about the situation on the soccer field is calculated, the robot starts to compute a sensible behavior. The robot knows a variety of behavior patterns, like dribbling, kicking, and defending, that it can combine as building blocks of its strategy. Every behavior pattern is assigned to conditions which describe when it should be executed. For example, a kicking behavior can only be executed if the robot is already close to the ball. Hence “being close to the ball” becomes a precondition of such a behavior pattern. These preconditions are checked for all behavior patterns and finally, one of them is selected that is the best one in the present situation. Sometimes our robots coordinate their activities (via radio transmission), e.g. to play a pass. Then, the player who wants to give a pass announces its intention so that the receiver can already drive to the right position to get the ball. In some situations it can also happen that a defender calls for help to protect the team's own goal.
In a next step of computation, the robot must determine which path it takes to go to its destination. Other robots might block the direct way. The robot calculates the best way to its destination considering all obstacles and the boundaries of the soccer field. Since it cannot observe obstacles that are occluded or far away, it sometimes has to make assumptions which way is better. Many of the robots' behavior patterns are designed by the software programmers . In contrast, others have been learned autonomously by the robot. This was done using reinforcement learning. In the begining the robot does not know how to solve a certain task, e.g. to dribble the ball. But it can try. For every trial the robot is rewarded or punished, depending on the success of the trial. If the performance was good and the robot was able to dribble the ball for a long distance, it will get a high reward. Then, the robot will store the dribbling it performed and will try to become even better, to get even more reward. If the robot is punished for a bad trial, it will avoid this kind of behavior in the next trial to get a reward again. In this way, the robot can learn to optimize its behavior. This kind of learning is motivated in the way humans and animals learn and has been transfered to computers. Of course, rewarding a robot is different from rewarding an animal. Instead of providing tasty food, the robot is just provided some mathematical operand.
Finally, the intention of the robot must be executed by the motors. The intention itself only states a direction of movement and a velocity. This information must be transformed to wheel velocities for each of the three wheels (i.e. how fast must the wheels turn so that the robot drives into a certain direction with a certain velocity). Then, the motor controller (an electronic device) determines the right voltage to make the motors turn in the right frequency. Sometimes, the robot wants to kick the ball. Then, it additionally activates the pneumatic kicking device.
More information on our approach to motor control can be found here.
Martin Lauer, Roland Hafner, Sascha Lange, Martin Riedmiller (2010) Cognitive concepts in autonomous soccer playing robots. Cognitive Systems Research 11 (3) pp. 287-309.
Thomas Gabel, Martin Riedmiller (2010) On Progress in RoboCup: The Simulation League Showcase. In RoboCup 2010: Robot Soccer World Cup XIV, LNCS. Springer. Singapore.
M. Riedmiller, T. Gabel, R. Hafner, S. Lange (2009) Reinforcement Learning for Robot Soccer. Autonomous Robots 27 (1) pp. 55–74. Springer.
Martin Lauer, Martin Riedmiller (2009) Participating in Autonomous Robot Competitions: Experiences from a Robot Soccer Team.
Thomas Gabel, Martin Riedmiller, Florian Trost (2008) A Case Study on Improving Defense Behavior in Soccer Simulation 2D: The NeuroHassle Approach. In RoboCup 2008: Robot Soccer World Cup XII, LNCS. Springer. Shouzou, China.
Heiko Müller, Martin Lauer, Roland Hafner, Sascha Lange, Martin Riedmiller (2007) Making a Robot Learn to Play Soccer Using Reward and Punishment. In Proceedings of the German Conference on Artificial Intelligence, KI 2007. Osnabrück Germany.
T. Gabel, M. Riedmiller (2007) On Experiences in a Complex and Competitive Gaming Domain: Reinforcement Learning Meets RoboCup. In Proceedings of the IEEE Symposium on Computational Intelligence and Games (CIG 2007). Honolulu, USA.
T. Gabel, M. Riedmiller (2006) Learning a Partial Behavior for a competitive Soccer Agent. Künstliche Intelligenz 2 pp. 18-23. Springer.
S. Lange, M. Riedmiller (2006) Appearance Based Robot Discrimination using Eigenimages. In In Proceedings of the RoboCup Symposium 2006. Bremen, Germany.
T. Gabel, R. Hafner, S. Lange, M. Lauer, M. Riedmiller (2006) Bridging the Gap: Learning in the RoboCup Simulation and Midsize League. In Proceedings of the 7th Portuguese Conference on Automatic Control (Controlo 2006). Portuguese Society of Automatic Control. Porto, Portugal.
M. Lauer, S. Lange, M. Riedmiller (2005) Calculating the Perfect Match: An Efficient and Accurate Approach for Robot Self-Localisation. In RoboCup-2005: Robot Soccer World Cup IX, LNCS. Springer.
Enrico Pagello, Emanuele Menegatti, Daniel Polani, Ansgar Bredenfel, Paulo Costa, Thomas Christaller, Adam Jacoff, Martin Riedmiller, Alessandro Saffiotti, Elizabeth Sklar, Takashi Tomoichi (2004) RoboCup-2003: New Scientific and Technical Advances. AI Magazine American Association for Artificial Intelligence (AAAI).
S. Lange, M. Riedmiller (2004) Evolution of Computer Vision Subsystems in Robot Navigation and Image Classification Tasks. In RoboCup-2004: Robot Soccer World Cup VIII, LNCS. Springer.
M. Arbatzat, S. Freitag, M. Fricke, R. Hafner, C. Heermann, K. Hegelich, A. Krause, J. Krüger, M. Lauer, M. Lewandowski, A. Merke, H. Müller, M. Riedmiller, J. Schanko, M. Schulte-Hobein, M. Theile, S. Welker, D. Withopf (2003) Creating a Robot Soccer Team from Scratch: the Brainstormers Tribots. In CD attached to Robocup 2003 Proceedings, Padua, Italy.
M. Riedmiller, A. Merke, W. Nowak, M. Nickschas, D. Withopf (2003) Brainstormers 2003 - Team Description. In CD attached to Robocup 2003 Proceedings, Padua, Italy.
A. Merke, M. Riedmiller (2001) Karlsruhe Brainstormers - A Reinforcement Learning Way to Robotic Soccer II. In RoboCup-2001: Robot Soccer World Cup V, LNCS. pp. 322-327. Springer.
M. Riedmiller, A. Merke, D. Meier, A. Hoffmann, A. Sinner, O. Thate, C. Kill, R. Ehrmann (2000) Karlsruhe Brainstormers - A Reinforcement Learning Way to Robotic Soccer. In RoboCup-2000: Robot Soccer World Cup IV, LNCS. Springer.
S. Buck, M. Riedmiller (2000) Learning Situation Dependent Success Rates Of Actions In A RoboCup Scenario. In Proceedings of PRICAI '00. pp. 809. Melbourne, Australia.
M. Riedmiller, S. Buck, A. Merke, R. Ehrmann, O. Thate, S. Dilger, A. Sinner, A. Hofmann, L. Frommberger (1999) Karlsruhe Brainstormers - Design Principles. In RoboCup-1999: Robot Soccer World Cup III, LNCS. Springer.
* Brainstormers Tribots source codeDriven by the idea to improve the general level of play in the Midsize league and inspired by positive experiences in the Simulation league we decided to publish the code with which we successfully participated at the German Open 2005 competition. We hope that this will help especially new teams to improve their play quickly and to overcome basic problems in vision, self-localization and robot control.The code contains:
- the program that controls the robot except the code that generates the robot's behavior. It primary contains the main control loop, vision, world model and communication with the motor controller
- a simple simulator
- calibration tools for vision
- debugging tools
- remote control tool with referee-box connection
It is developed in C++ under Linux. Visualization tools are based on QT3.The code is distributed under the terms of the GNU General Public License. It comes without any warranty. An incomplete documentation in German language is attached. We cannot offer any support. For more information also consult our publications list.
Brief Description: Here are some remarks on the software and the tools. In general, the software works well but some parts caused a lot of trouble (especially the UDP communication). The major software design is based on the idea of a clocked control loop and a sequential control flow. Four major subtasks are realized as independent units: vision, world model, behavior (not published) and robot access.The world model deals as server for the other subtasks: it integrates incoming preprocessed sensory information and it provides state information. Visualization, calibration and debugging are done in separate programs that eventually communicate with the main control program via UDP.