Hi, just a thought,
I’d like to know if someone knows a script about robots in the GE, about kind of autonomous object ‘really’ discovering a world with its sensors… with a ‘real’ motion mecanism as an option (wheels, feet, caterpillar, gravitation…etc…)
in other words if there’s some predator script that really use ‘physical sensor’ (radar, ray, ‘faked audition’ etc) to reach its objective without knowing exactly the prey global world position, and the world itself (learning it)
Are you talking about interfacing Blender with real, physical robots? Or are you talking about just using it as a test bed to try out algorithms?
If the former, there was a project some time ago that sought to do it. Its somewhere over on sourceforge.net (and a post was made here about it). You might look to see what it gave you.
If you’re talking about a test bed, then you should be able to just use python to implement different AI algorithms and use it from the game engine. Unfortunately, I’m not as up on the previous work here, but I’d wager searching through the forum (especially some of the game forums here) would yield Algorithms or previous projects that may be of interest.
Not sure if that answered your question (or helped), though… :o
I"m talking about test bed, real robots AI algos are far away from my knowledges, and my gov. will never give me the money to construct a stupid one and have fun with relatively simple algorythms.
but, it’s funny. I’m thinking about autonomous combat robots (it’s simpler to code than poetic robots I guess), or robots with the same goal in competition (labyrinth etc). each one could code it’s own algo for actions, but there would be the same world/constraint for them, and a shared code about energy cost / capacity/number of sensors, loss of capacity (something breaks after a choc, a hit), mass/ max speed /energy cost etc… some base functions could be provided to simplify the code of AI, and get ‘relatively’ independant of programmation, to focus on logic.
a kind of arena for algo middle-noob guys like me, not for beauty of the code, but for fun, to see AS doing stupid things at the beginning, a kind of funny education about algo.
in two video, her’s the concept : something between this :
You may be interested in player/stage/gazebo
http://playerstage.sourceforge.net/
It allows you to code as though you were interacting with a real robot, with whatever sensors, etc. The simulation package can be swapped for a real robot with no changes in code (though it will highlight the difference between theory and practice). These are already set up with sensors, sensor interference, etc.
I’d suggest reading about planning (look up strips and follow the progress & timeline) and some basic machine learning algorithms (q-learning is a good one to start with). You may want to try evolving the algorithms, which could be done with a graph/tree based genetic programming. I study AI, so if you know what you want to do I can try and point you in the right direction.
also, try some simple feedback loops. Facing towards a light, finding prey, etc. These are often very satisfying as they take incredibly simple rules but give quite effective behaviour.
Cheers,
Ian
hey thanks for the ressources. what do you think about RL glue ?
http://www-anw.cs.umass.edu/rlr/
they provide a python autonomous agent…
I’ve found that also about q-learning :
http://thierry.masson.free.fr/IA/en/qlearning_applet.htm