Re: Starfish Self Modeling Robot

Neat Starfish ... but we achieved the same result by using a much more simpler approach: Genetic Programming.

Our group of (effectively) four computer science students created a system called SIGEL during a period of two semestres without any budget, using only free software.

SIGEL takes any robot architecture consisting of leg and body 3d-models as well as the information on the joins to fully automatically perform network-parallelized dynamic simulations of this robot on randomly created and mutated/crossed control programs in a pseudo-assembler language.

We succeded in walking programs for starfish-robots of similar quality just with this very simple biologically inspired approach.

The software also featured a graphical user interface, realtime OpenGL-visualization and a povray-animation export.

The results have also been published in the proceedings of the EuroGP 2002 conference, LNCS 2278, Springer, 258-267.

By using a simulation, the robots learn to walk without the danger of getting damaged. Moreover, you can teach them faster than realtime. Later on, the programs can be interpreted by a real robot controller, which also has been done in a follow-up diploma thesis.

Just watch the movie ;) There is also a Sourceforge-Project, where you can get the software, documentation and some experiments for free: http://sourceforge.net/projects/sigel

Category:

Entertainment

Tags:

• BotJunkie
• Starfish
• Self
• Modeling
• Robot
• Genetic
• Programming

[AySz88]

I think you missed the part where their robot recomputes the model of itself when it gets damaged... That's the cool part, not the walking part.

[tabmasterispell]

You can do the same with our approach: Just restart the genetic programming after the robot has been damaged. It again learn how to walk.

Literature on evolutionary algorithms is full of complicated algorithms, but theoretical analyses show that many of these are not able to speed up things - although sounding fancy and nature-like.

[woflox]

umm.. the starfish robot DOES create a model of itself, and I believe it runs simulations in its software before deciding how to walk, it does not simply learn by trial and error.

[tabmasterispell]

YES, the starfish creates a model of itself and our SIGEL robots does NOT.

But the starfish guys use a more sophisticated approach to reach a comparable result.

Reaching the same goal with a simpler approach is actually regarded as an advantage in science.

Moreover I would not agree that evolution is "simple trial and error", no animal does adaptively learn its own kinematic model.

[prestorock]

but they built an actual robot, this is just 3d... make it into a robot to be much cooler :D

Exept for the fact that the robot starfish they made will soon figure out that they are smarter than use and will attach to our faces and control our brains...

[tabmasterispell]

This is a feature, not a bug ;)

The simulation enables to evolve fast, in parallel and without trashing real robots.

Point is: Our method of genetic programming (both applied to reality or simulation) is much simpler and achieves comparable results.

Last but not least: Programs generated by our system have actually been transferred to real robots.

[RobotZORC]

Another project realized with SIGEL:

ZORC, humanoid robot architecture learning to walk.

[PlebScrubber]

pov-ray animation?

Well the animation certainly is pov...

You could of just programmed it to walk properly from the start

[virtualspecies]

Cool stuff, I should put my ZORC videos (humanoid robot learned walking with the SIGEL system) on the tube as well :-)

[broncocv]

man u r a genius, surely whan I get time will search more info of what u have done!!, u did it with only free sofware?? wow...genetic programming, a field in research yet, u r in the future!!! well, I 'm just starting to study electronics jaja, u can see a simple video on my channel!! but I'm also interestes in AI and programming with C++...

[culpritdesign]

great work!

[enferris]

Yes. Exactly. The advantage of modeling from a real robot is resilience. If the movement algorithm is transfered to a robot, and the robot isn't built to correct specifications or gets damaged, the algorithm would be useless. By generating the algorithm from the body, The starfish robot can deal with myriad circumstances.

[kjlg74]

Very nice!

Do you evolve a single GP program to control all joint motors (servo motors?) or separate programs for each. How's the overall coordination evolved?