Ooo... I Want One.

[Marc]

I've seen this software before, but I think this is a new version. New videos anyway :

http://media.xbox360.ign.com/articles/702/702389/vids_1.html

Some very cool stuff.

M

[Rafal123]

yeahh, very cool stuff but... how it works?

[sibarrick]

That was all at Siggie last year, surprised you didn't see it.

[Marc]

I never got to go last year. I just got to go to the parties .

[symek]

yup, Siggraph presentation of verion 2.5... quite impressive. Useful if one has good motioncapture workflow.

Guy from computer science faculty came a cross biology faculty and saw research on human body-motion behavior. They have recorded lots of motion capture date to analize it later . As they finished they have something like a profile of common human body behaviors as a sidefx of main research. So the guy from computer science has found that it would be nice and relatively easy to use these profiles to synthesize human motion. And this is how colaborative, openess, and fresh thinking drive progress.

You still need a lot of motion capture data to use it effectively thought....

[stu]

Wow, very impressive.

I like the guy who falls from the airplane and does a tuck and roll to safety land on the street, unscathed.

[ivan]

actually, I don't think it uses mocap data. They way it was explained to me, it's a nueral network that was trained; i.e make 1000 neural net controllers, knock 'em on the shoulder, and the save the 10 best at keeping their balance. Cross breed these 10 and make a population of 1000 from it, and repeat. 2 years lter, voila! you have a product.

[symek]

actually, I don't think it uses mocap data.  They way it was explained to me, it's a nueral network that was trained; i.e make 1000 neural net controllers, knock 'em on the shoulder, and the save the 10 best at keeping their balance. Cross breed these 10 and make a population of 1000 from it, and repeat. 2 years lter, voila! you have a product.

26888[/snapback]

Maybe you're right! But still neural networks must be traind on something. That is human data, in which we must find function, order in chaos, this is somthing neural networks are good at. As I know the system is based on - and come from biology research on basic, unconscious behaviors - but still it can be driven by AI . Its deep,deep topic!

[TheUsualAlex]

so neural net's logic is still fuzzy?

That was pretty cool tho.

[sibarrick]

As I understand it, the software combines mocap actions and learned actions. The original work it was based on used a dynamic system that basically "evolved" a walk model for any system of connected joints. So the training wasn't based on mocap just the dynamics of human motion, the more successful a solution was the more likely it was to be used in the nexrt generation. Very smart stuff. That's how it is able to adapt to so many different interactions.

I'll have to see if I can find the original sig paper. Perhaps someone else knows of it. There are some great animations of little two legged beasties literally learning to walk by evolution, its amazing really.

[symek]

So I think I misunderstood some old print about Endorphin development... doeasn't matter. Its great to see how new technology born and changes the way things are done. I used to play with it few weeks, its really funny to build animation in it.

...The worse part of the play was that I had to use Python cgkit to import endorpin data to Houdini since bhv importer doesn't work very well