Impressions from Robocup 2006 humanoid league

[limor]

I spent last week in Bremen at the Robocup2006 event.

I had an interesting week at this huge event where thousands of students compete at football themed robot competitions with some 150k spectators during the week. The competition that IMHO was most visually appealing was the Sony Aibo dogs where several universities have perfected their artificial intelligence over 5 years. At the end of every Aibo tournament, participants share their source code and findings so sophistication increases year after year (see www.germanteam.org for example). Anyone watching those plastic dogs was struck by how intelligently and lively they behaved both individually and as a group. The competition creates big groups of student volunteers from around the world that work on preparing the robots between 3 months and a year before the tournament.

While many of the wheeled robot competitions were interesting, I was mostly interested in the walking robots. The humanoid competitions were less impressive visually than most other competitions because most of the humanoids were slow and tended to fall a lot and most (except the Japanese teams robots) were not very pretty. The main reason was not the team's capabilities but rather the objectives set by the Robocup federation. The goals involve making the robots autonomous (ie: they could not offload computation to a remote PC) and since they could only carry a limited payload given their size, servos and batteries, they could only afford a handheld windows-CE device, so the computation power did not allow or encourage them run around.

The Robocup federation is apparently composed of people that come from Computer Science and AI and not so much from the Mechanical Engineering and Control. The organizers would like to see the robot platform as a constant and the creativity should focus on aspects of AI like image processing, action planning strategy and hitting a ball. While this Utopian constant platform is true for the Aibos where Sony created the ultimate 4-legged research robot, what happens with the humanoid competition is that 99% of the work revolves around creating and maintaining the humanoid platforms. The students have mainly computer-science and electrical engineering background. The platforms are mostly customized Robo-One type robots with servo motors that were not designed to sustain such abuse. It was the first time that I've seen robots that look like they have been in a war zone for thousands of hours.

The students that volunteer to prepare the robots for competitions do an amazing work in creating, destroying and maintaining the platforms. Some of the groups split the work between teams responsible for the mechanics and electronics and the other for the software. Some groups design their own controller board and dynamic simulation platform (ex: http://www.bredobrothers.de and www.hajimerobot.co.jp). Most robots use Robotis's Dynamixel servos with 32kg-cm torque and 1mbit digital communication bus. However, some do use the Kondo and Hitec PWM driven servos and rarely other servos. These guys and girls have encountered and resolved every possible bug in their supplier's gear but they don't have time to publish their findings readily on this forum or others so it seems. I was told that there is is a robocup mailing list where some technical issues are discussed but most of the information that would be of great interest to RoboSavvy's forum readers, is not saved for posterity. I hope we'll see some more feedback from Robocup people in the future on this forum.

I asked around if there was any feedback control used to drive the robots and if lack thereof had to do with why they were so slow in general. The general opinion is that dynamic feedback control that would allow the robots to be more agile would not contribute much to winning the tournament as it's current rules stand. The humanoids at Robocup run pre-calculated gaits and utilize sensors to do things like realize that the robot is falling to relax the servos and prevent damage or identify which side of the body the fall has occurred in order to execute the right gait to stand up again or low-level "closed loop" using gyros and accelerometers to improve walking and prevent falling etc. Real "closed loop control", "dynamic control" or "feedback control" is not quite there yet though most say that next year it will probably be introduced.

It seems to be an accepted fact that a simulator can not really simulate the robot perfectly and especially keep up with the accumulating precision loss and lag of the mechanical components. Since these robot don't do any feedback dynamic control, the simulator is an essential tool to prepare, perfect and "evolve" motion sequences in advance for the humanoids. Hence the ideal simulator should continuously calibrate itself with the real robot, this demands the robot to provide precise servo position feedback such as found with the Robotis servos (and to some extent also with the Hitec and Kondo servos).

To conclude, I learned a lot last week especially through talking with the team members and hearing the stories behind the robots [not without some difficulty because soulless henchmen were protecting the teams' work space from access by plebeians]. I think Robocup will loose its ground if the humanoid league focus doesn't turn to "eye candy" stuff as is the case of Aibos and Qurios (R.I.P.) or even by Robo-One's aesthetic violence. While the engineering efforts put into the humanoid league are formidable and admirable, the general public will not come in their thousands in the future to see slow "boring" robots and unfortunately it is the public that drives Robocup's commercial booth-space, sponsorship, media coverage and in turn student interest and research.


Check out the individual humanoid teams and their work efforts : http://www.humanoidsoccer.org/teams.html


Limor

[SK]

Hello,
being a team member of the Darmstadt Dribblers & Hajime team, I want to do at least a small part on sharing our experience and give a view of the shape of things to come (at least in "my" team). Please note that it´s impossible to know all modules inside out, I mainly worked on selflocalization, so my knowledge of low level motion stuff and the like is a bit shaky.

You can find the official resumee of our team here:
http://www.sim.informatik.tu-darmstadt. ... _H.en.html
Especially take note of the heel kick video ( http://www.sim.informatik.tu-darmstadt. ... ntrick.mpg ) and the race with Aibos vid ( http://www.sim.informatik.tu-darmstadt. ... Rennen.mpg ). These two show the direction our team wants to exploit further, having
a) fast robots and
b) interesting gameplay, a collection of different kicks etc.
Our robot is doing 30 cm/s right now, the video shows a 40 cm/s gait (which isn´t finalized, obviously). I have no clue how fast the Robo One robots are, does anyone have numbers?
Having said that, I´m confident ( granted, I´m biased ) that interesting gameplay akin to the 4-legged league can be shown sooner than many people might expect.
As mentioned in the resumee, our robots also already were able to show emotional reactions when scoring a goal for example. (Picture here: http://www.heise.de/bilder/74383/4/0 )

Our fast gaits were developed with hardware-in-the-loop optimization on the real robot, evolving fast parameter sets. I´m not sure if it´s completely statically stable, gotta check that. Do the Robo-One robots (and/or others) already use full dynamic control?

That`d be my first post then, thanks in advance to those answering my questions

[limor]

Hi SK,

Thanks for posting about the Darmstadt Dribblers. That back-kick goal video is very cool.

Our robot is doing 30 cm/s right now, the video shows a 40 cm/s gait (which isn´t finalized, obviously). I have no clue how fast the Robo One robots are, does anyone have numbers?

I dont know of precise numbers (maybe tempumaster from robots-dreams.com has some) but many of them are designed for the purpose of being fast and stable. here are some video links: 1, 2, 3

Do the Robo-One robots (and/or others) already use full dynamic control?

As far as i know they dont use anything more complicated than gyros and accellerometers and in general software development is not Robo-One's forte'

[tempusmaster]

Hi SK,

Thanks for posting about the Darmstadt Dribblers. That back-kick goal video is very cool.

Our robot is doing 30 cm/s right now, the video shows a 40 cm/s gait (which isn´t finalized, obviously). I have no clue how fast the Robo One robots are, does anyone have numbers?

I dont know of precise numbers (maybe tempumaster from robots-dreams.com has some) but many of them are designed for the purpose of being fast and stable. here are some video links: 1, 2, 3

Do the Robo-One robots (and/or others) already use full dynamic control?

As far as i know they dont use anything more complicated than gyros and accellerometers and in general software development is not Robo-One's forte'

I'll try to dig up the numbers from my files tomorrow, but just a couple of quick observations:

The first two video's you linked to are of Omnizero.2 at the ROBO-ONE 9 competition this March. Maeda-san, who created Omnizero.2 works for VStone and is also one of the key members of Team Osaka - the group that did so well at the RoboCup events here in Japan and in Germany.

Also, the 'Hajime' part of "Darmstadt Dribblers & Hajime team" is Hajime Sakamoto, who is has been one of the top ROBO-ONE competitors here for the past few years. His company slogan is "I want to build Gundam".

[tempusmaster]

These numbers are a little bit old at this point, but should give you a good feel for what the top ROBO-ONE class robots are capable of achieving.

At the November, 2005 RoboPro 3000 Sprint (3 meter dash) event the top times were-

OmniZero.2: 7.94 seconds
OmniZero: 8.90 seconds
Subrental: 10.23 seconds
Simple Fighter: 37.03 seconds
Thinkmover2005: 38.92 seconds

In the 5000 Sprint event (5 meters) the top times were:

OmniZero.2: 36.55 seconds
OmniZero: 42.52 seconds
LiLia: 1 minute 21.42 seconds
Simple Fighter: 1 minute 21.94 seconds
Thinkmover2005: 1 minute 37.73 seconds

Here's a video from the robot-fan.net website that shows OmniZero.2 doing the 3 meter sprint in 5.49 seconds flat:

http://www.robot-fan.net/spot/spot056.html

Note: The March Hare is doing a complete renewal & redesign of the robot-fan.net website this weekend (July 1, 2006) so that url may change. I'll try to update it if it does shift.

According to comments on his weblog, Maeda seems to think that the theoretical limit of the current technology they are using is in the range of 600-775 mm/sec.

[SK]

Omnizero2 is indeed a very impressive robot, the speed and fluidity of motion certainly set new standards. I guess it will still take some time till Robocup robots get that good though, as autonomy costs quite a bit in the weight department.
What are there restrictions on the center of mass for Robo-One fighters? From the videos I´ve seen the robots CoGs look kinda similar to the Robocup robots. (You can find the humanoid league rules here: http://www.informatik.uni-freiburg.de/~ ... es2006.pdf )

In other News, you can now find some more videos on the Darmstadt Dribblers Site, here: http://www.sim.informatik.tu-darmstadt. ... _H.en.html

[tempusmaster]

Omnizero2 is indeed a very impressive robot, the speed and fluidity of motion certainly set new standards. I guess it will still take some time till Robocup robots get that good though, as autonomy costs quite a bit in the weight department.

Very true. A lot of the ROBO-ONE/Omnizero2 technology is finding its way into the Team Osaka Robocup robots though since Maeda, Omnizero2's creator, is a member of the team.

What are there restrictions on the center of mass for Robo-One fighters?

Not directly. The regulations have been deliberately kept as open as possible to encourage participation. The foot/sole dimensions as a percentage of the leg length have been made more challenging over the years. There are also some general size limitations, but nothing that specifies the COG specifically. I should be posting a complete English translation of the latest regulations on Robots Dreams within the next week or so.

From the videos I´ve seen the robots CoGs look kinda similar to the Robocup robots. (You can find the humanoid league rules here: http://www.informatik.uni-freiburg.de/~ ... es2006.pdf )

That's great. Thanks!

In other News, you can now find some more videos on the Darmstadt Dribblers Site, here: http://www.sim.informatik.tu-darmstadt. ... _H.en.html

Thank you very much.

[tempusmaster]

What are there restrictions on the center of mass for Robo-One fighters? From the videos I´ve seen the robots CoGs look kinda similar to the Robocup robots. (You can find the humanoid league rules here: http://www.informatik.uni-freiburg.de/~ ... es2006.pdf )

I just posted an excellent English translation of the ROBO-ONE rules done by Duncan Lowne that includes the latest rule changes for the September, 2006 event coming up:

http://www.robots-dreams.com/2006/07/unofficial_robo.html

[SK]

Thank you, very interesting. Especialy the "autonomous" part in the preliminaries caught my eye. How do these preliminaries look like?
Anyways, I´m off to a place without internet for 5 days, cya later then

[tempusmaster]

Thank you, very interesting. Especialy the "autonomous" part in the preliminaries caught my eye. How do these preliminaries look like?

Autonomous has basically meant that the robot has to operate on it's own for two minutes going through whatever demonstration sequences you've laid out. For example, at ROBO-ONE 9 in March, Layerd-X basically just went through the motions:

http://www.youtube.com/watch?v=0m_cCP1wn94

However, more and more entries are incorporating truer autonomous functionality. For example, Omnizero2 used it's sensors, including the camera, to change it's behaviors:

http://www.youtube.com/watch?v=a5O6ik1Jr6E

I suspect we'll see much more advanced autonomous capability this September at ROBO-ONE 10.

[SK]

Hello again. This is a message for the lazy folks out there . The videos (and some more) linked above of the Darmstadt Dribblers & Hajime Team are now also available on YouTube too and thus can be watched directly from the browser:
http://www.youtube.com/profile_videos?u ... tDribblers

[Robo1]

After reading the rules for the football I was surprised. Their trying to regulate every thing even the colour. And what a colour, all robots have to be black with 1% allowed to be coloured. what makes for the most boring robots going it's like the first model T fords. and even where the centre of gravity is. I was also surprised that you only have two players that's no game of foot ball what happened to fiv aside that would be more interesting.

bren

[SK]

Well you have to take into account the circumstances and the current state of the art in humanoid robot soccer.
There are some good reasons for the 2006 rules:
-Doing matches with 5 robots is an order of magnitude more expensive than doing matches with 2. Having 5 robots ready to play a 20 minutes soccer game requires a lot of people and a lot of money. Most participants are universities, with limited resources. This of course also is true for the MidSize league, but there the robots already converged towards a good design, while the humanoid robots are largely prototypes who might be obsolete in a year.
-The robots have to be mostly black because all robots use color segmentation for perception, due to the low processing power available. With free color choice there would be huge problems, as other robots might would get recognized as balls, goals, landmarks etc.
-The size and centre of gravity restrictions make sense, otherwise you could build robots will very low CoG which don´t resemble humanoids.

That being said, the rules get refined and "harder" every year, to facilitate advancement. For example, the MidSize league played with natural light conditions for the first time this year, in the humanoid league the playing field got bigger etc. We might even see 3v3 games in the humanoid league next year, it´s not been decided yet AFAIK.