B e m b e l b o t s F r a n k f u r tRoboCup SPL Team at Goethe University Frankfurt

Dipl.-Inf. Markus Meissner1, Dr. Holger Friedrich2, Dipl.-Inf. Andreas Fürtig1,

Dipl.-Inf. Tobias Weis1, Jens-Michael Siegl2, Dipl.-Inf. Christian Becker1,

Vincent Michalski1, Dipl.-Inf. Linda Luy, Gerhard Ruscher1, M.Sc. An-

dreas Kehlenbach, Dipl.-Math. Anna Reckers1, Konrad Zacharias1, Alexander

Heun M.A.1, B.Sc. Philipp Hülsdunk1, Jonathan Cyriax Brast1, and Volker

Kühn

1Goethe University, Frankfurt, Germany2Frankfurt Institute for Advanced Studies, Frankfurt, Germany

teamleader@bembelbots.de

1 Statement of Commitment

Team Bembelbots Frankfurt applies for participa-tion in RoboCup 2014 SPL.

2 Team Bembelbots

Team Bembelbots was founded 2009 at Goethe Uni-versity Frankfurt(Main), Germany. Our project ismainly driven by a group of from both computerscience and math students and is supported bythe Joint Robotics Lab (JRL) at the Institute ofComputer Science at Goethe University. Follow-ing the principles of the JRL, we focus on educa-tion and project oriented learning. We work au-tonomously with the supported of researchers fromthe department. Right from the start, team Bem-belbots decided to build its own software frameworkfrom scratch, which allows students to improve theirskills in software design, programming languages,and many other topics involved in playing robot soc-cer.

Since 2010, we have been taking part in theRoboCup German Open, gaining more and moreexperience. We have obtained new practical andtheoretical knowledge in Open RoboCup Workshops(Berlin and Dortmund) and by playing friendlymatches and exchanging knowledge with otherteams. Thereby we made continuous good progressimproving our framework. We implemented a ro-bust vision, a satisfying self localization, a dynamicbehavior, our own walk and much more. As a re-sult of our efforts we got the chance to participatein the RoboCup World-Championship in 2012 and2013 and enjoyed the experience of the internationalcompetition very much.

Figure 1: Team Bembelbots at the GermanOpen 2013

Team Structure and Research Interests: Teamleader of Bembelbots is Markus Meissner, a PhDstudent at the Electronic Design Methodology (EM)group at the Institute of Computer Science atthe Goethe University Frankfurt. He has a goodoverview of the team members’ activities and isvery avid, especially in the development of behav-ior and motion. Dr. Holger Friedrich, former headof the team, has left Goethe University, but stillspends a lot of his free time with the team, es-pecially in the areas vision, framework and soft-ware development methodology. Andreas Fürtigwrote his diploma thesis on color classification [2]and is now also working at the Design Methodologygroup. Tobias Weis has finished his diploma thesison self localization [3] and is continuously workingside by side with Vincent Michalski on improvingthe localization. Christian Becker finished his the-sis on advanced line detection and scan-matchingalgorithms [7]. Both are now PhD students at theCenter for Cognition and Computation group at theInstitute of Computer Science at the Goethe Uni-

1

versity Frankfurt. Jens Siegl contributes technicalknowledge and maintains the robots since the begin-ning of the project. Linda Luy and Andreas Kehlen-bach have left the university but still contribute fre-quently. Gerhard Ruscher just finished his diplomathesis on our new motion-engine. Anna Reckers isthe contact for mathematical problems. In additionwe have several students who joined the RoboCupteam in the last years. Alexander Heun will concen-trate his efforts towards kicking motions based onthe new motion engine, Philipp Hülsdunk is workingon the simulator simspark1 and Jonathan Brast iscurrently implementing a new ballmotion filter forthe goalkeeper.

3 Inventory of NAOs

We currently own six Nao V4 H21. As soon as anew version will be released, we will upgrade ourNAOs. At the moment we do not intend buyingany further robots.

4 Summary of Recent Work

4.1 Software Framework

The main goal of the Joint Robotics Lab is the re-alization of a robot soccer framework built fromscratch. Starting in 2009, the first version of theJrlSoccer Framework was built as a C++ NaoQiLibrary. From time to time, more and more NaoQimodules were replaced by own implementations.Due to the increasing complexity of this frame-work, we recognized several shortcomings of the firstapproach, namely complicated maintenance and ahigh initial hurdle for new team members trying toimplement new modules.

Backend

DCM

Top Camera Image

Bottom Camera Image

NaoQi Function Calls

NaoQi Results

Frontend

Cognition Thread

cognition data (copy-by-value)

motion data (copy-by-value)

Motion Thread

motion data (copy-by-value)

shared memory

Cognition (Read)

shared memory

Cognition (Write)

shared memory

Motion (Read)

shared memory

Motion (Write)

(aka SetCon�g)

(max 10 ms)

SEM

SEM

Figure 2: New Robot Soccer Framework jrlsoccerv2, split up

in two main parts.

Since the old framework was not able to handle twocamera images at once, in 2013 a complete redesignof the framework was mandatory. The completeFramework is now split into two main parts: The

1http://simspark.sourceforge.net

backend and the frontend. Both parts commu-nicate through a shared memory, which syncs theincluded processes. The Backend gathers all rele-vant sensor data and provides synchronization andcommunication with the actuators. Several Back-ends are available: The naoqi-backend is used on thereal robot hardware and for the Webots simulator,the simspark-backend interfaces with the SimSparksimulator and the python-backend is used for de-bugging purposes and replays previously capturedsequences.The Frontend is a standalone executable which pro-cesses the captured data in two main threads, themotion and the cognition thread. This applica-tion can be easily monitored by a daemon, which isable to restart our framework in case of failure.We have shown our development methodology andsoftware modules in the last two Teamreports.

4.2 Ballmodel

Our new ball motion filter allows us to make betterpredictions about future positions of the ball. Byusing the past ball positions relative to the robot,we can filter the velocity of the ball and calculatewhere and when the ball will pass the robot. Thisis currently only in use for our goalkeeper standingin the goal. This method has to be adopted to fieldplayers in the future.

4.3 Motion Engine

The new motion engine [8] provides a universal in-terface, which is able to handle motion-commandsfrom several sub-modules for different body-chains.In contrast to the preceding motion engine this pro-vides a way to deal with static motions (standup,sitdown) and dynamic motions (walk, kick) simul-taneously. Due to the decoupling of body-chains,we are now able to track a ball while standing up.

4.4 Tools

Several tools have been developed in the last years,allowing us to debug and analyze the behavior of ourNaos. Those tools are written in Python, mainlyfor portability reasons. They work on either Linux,Windows and partly also on mobile phones (An-droid). NaoDebugV2 is the main application todebug the robot, for example change camera pa-rameters or receive live images. Following the newframework structure, it has been completely rewrit-ten and now receives the data via UDP broadcasts.This allows us to receive compressed JPG imagesin real-time. It is now also possible to capture alive stream of the robots’ data to save real-time

2

Figure 3: The Joint Robotics Lab recently moved to this lab - with a permanent 7.4 x 5.4m SPL playing field. The JRL also

hosts other projects including 3D printers, custom built and commercial robots and robotics arms.

sequences for debugging purposes. Additionally,NaoDebugV2 now receives a health state from eachrobot without the need of running any NaoQi soft-ware. This improves the pre-game setup processdrastically.

4.5 The new Joint Robotics Lab

Joint Robotics Lab tries to link education and re-search in robotics and computer vision. It was initi-ated by 3 professors in 2008 with a focus on project-oriented learning. Students can take the oppor-tunity to work in a project of practical relevance,and work autonomously on cutting-edge topics un-der the supervision of researches at Goethe Univer-sity. Cooperation between students and researchersimproves soft skills and eases knowledge transfer.

The JRL tries to improve visibility of natural sci-ences and computer science in particular to the gen-eral public by showing interesting, tangible, and ap-parent applications at various events. The goal isto wake interest in natural sciences especially foryoung people. Thus, the JRL participates in eventslike Night of Science or Innovative 2010.

Further activities of the JRL – besides support-ing the team Bembelbots – involve other roboticand camera hardware: the Virtual Robotics Lab[9] (interactive microcontroller programming overa web interface), Helicontrol (USB IR remote con-

trol for small helicopters), Fischertechnik (Linuxdriver), and a 3D printer. The lab provides differ-ent Firewire cameras, Kinect cameras, robotic arms,custom built and commercial robots.The JRL recently moved to a new lab space (Fig. 3),where we have a playing field that meets the SPLrules of 2012 (7.4 x 5.4m) permanently installed.A wooden base, similar to the ones at the compe-titions, has been built with the official carpet fromRoboCup2013. Above the field we have regular ceil-ing light as well as six additional LED floodlights.Several cameras mounted on the ceiling can be usedto generate ground truth for the position of therobots, either using the wide-angle camera coveringthe whole field, or using two cameras each capturingone half with the software from the SSL league.

5 Planned Activities

With the benefit of the new framework we’re nowable to use a simulator other than the commercialand very expensive webots simulator. Currentlywe’re working on a new Backend for the simsparksimulator, which helps us to simulate more robots atonce. So it should be possible to test complex teamstrategies in the future. Due to the module-orientedstructure of the framework we are able to ana-lyze modules standalone as well as the frameworkin total, and characterize the performance on well-

3

Figure 4: High-school students working with the Nao during

the summer school.

defined in- and outputs, as well as on real datase-quences (and run automated unit-tests on modulesand the entire proecessing chain).With the new motion engine in place, more sophis-ticated interpolation methods should be added anda inverse kinematic module must be developed toenhance our set of available kicks. Since the robotsare currently not able to pass the ball to their team-mates with sufficient precision, a full parametrizablekick based on this module would be a major im-provement.In contrast to several other teams we do not plan toswitch to a colorless vision in the near feature, sinceour colorbased segmentation produces sufficient re-sults. First tests show good results in detectingrobots by recognizing the different jersey colors.

5.1 Public Activities

Team Bembelbots is also very devoted to increasethe awareness of the RoboCup initiative by organiz-ing and taking part in public events.Every year we are organizing different events forthe highly frequented Night of Science, since 2011we are also part of the Computer Science Day andthe Summer Festival at Goethe University. On theNight of Science 2013 we were able to provide acomplete field and played several games against theNao Team HTWK Leipzig. Since 2012 the team car-ries out robotic classes for high-school level studentsparticipating in a summer school program providedby Goethe University. Several positive reviews in lo-cal and regional press documented the great successof our public activities2.

2http://www.bembelbots.de/web/press/

Figure 5: Night of Science 2013 friendly vs. Nao Team

HTWK Leipzig and audience.

5.2 Scientific Publications

Scientific publications [1], Diploma theses [2, 3, 7, 8],Bachelor theses [4, 5].

References

1. Fürtig, Andreas; Friedrich, Holger; Mester,Rudolf (2010): Robust Pixel Classification forRoboCup, Farbworkshop Ilmenau

2. Fürtig, Andreas (2011): Farbklassifizierungfür humanoide Roboter im RoboCup Umfeld,Diploma thesis at Goethe-University, Frankfurt

3. Weis, Tobias (2011): Selbstlokalisierung fürhumanoide Roboter im RoboCup Umfeld,Diploma thesis at Goethe-University, Frankfurt

4. Rohnfeld, Erik (2010): Reinforcement Learn-ing zum Lernen eines Schusses auf dem Nao-Roboter, Bachelor thesis at Goethe-University,Frankfurt

5. Jäger, Moritz (2010): Entwicklung eines au-tomatisierten Verfahrens zur Optimierung derLaufbewegung des Nao Roboters, Bachelor the-sis at Goethe-University, Frankfurt

6. NaoRemote for Nao Robot (2012):www.bembelbots.de/web/robocup/downloads/

7. Becker, Christian (2013): LinienbasierteFeaturedetektion und Positionstrackingdurch Voronoi-Diagramme in einer sym-metrischen Umgebung, Diploma thesis atGoethe-University, Frankfurt

8. Ruscher, Gerhard (2013): Entwurf und Re-alisierung eines Frameworkes zur Bewegungs-planung eines Humanoiden Roboters, Diplomathesis at Goethe-University, Frankfurt

9. Virtual-Robotics-Lab (VRL):www.jrl.cs.uni-frankfurt.de/web/projects/vrl/

4