AI learns to stand

Share with us how are you using the powerrrr of the force

Moderator: Alain

AI learns to stand

Postby JoeWright » Tue Mar 17, 2009 1:55 pm

JoeWright
 
Posts: 69
Joined: Sat Apr 30, 2005 1:42 pm

Re: AI learns to stand

Postby Julio Jerez » Wed Mar 18, 2009 9:12 am

Oh very interesting.
so the robot is leranong to keep its equlbrium, by trial and error.

will he leran teh notion to tryin moving his legs to get a better sopport.
maybe one rule you can add to it system is the teh position of teh center of mass.
keep a stable cnet of mat is a bimecahnical rule.

I beleive is work like this:

you get all teh sopports (the points i constact with teh grund)
they define a shape, if teh projection of eth cente of mass is inside teh shape, no action is taken,
if is too close, then the cerbelum sned singal to some of teh extremotis to move in teh diretion to make sure the center move to teh inside o fteh shape
teh will keep stability.

it is not as eassy as is sound to programe because as you move any of teh extremotis the sopport of the floor also changes,
plus the shape is not that eassy to from, bu you can use aproximation and projection.

that is one of the thongs that brain do in a incosinet way, and I guess will be interestion you program into a newral net or simulation aneline.
I think i sgo a very long way to make the learning smarter.
I once saw a programe wher some sceintist where saying that those funtionaliy are already pre programed in some part of the brain.
Julio Jerez
Moderator
Moderator
 
Posts: 12249
Joined: Sun Sep 14, 2003 2:18 pm
Location: Los Angeles

Re: AI learns to stand

Postby Julio Jerez » Wed Mar 18, 2009 11:51 am

I deleted the previuos post, this forum do not allow comparisons and sarcathm of any kind.
you can do that somewhere else.
Julio Jerez
Moderator
Moderator
 
Posts: 12249
Joined: Sun Sep 14, 2003 2:18 pm
Location: Los Angeles

Re: AI learns to stand

Postby JoeWright » Sun Mar 22, 2009 11:54 am

Hi Julio,

There's an update on the page as I've improved evolution time.

At the moment I don't think the particular neural network I'm using is capable of multi stage movement so, for example, seperating the legs first may be something it is unable to do. Saying that, I think the particular model I've constructed is pretty stable laterally.

May next goal will require multiple stages so its time for a brain upgrade. I'm looking at spiking neural networks.

I could add a centre of mass calculation but I want to avoid as much intermediate processing as possible. This is because many of the skills I'm going to move onto will complex notion of what balance is. However, I may include your technique now and then to compare evolution speeds - thanks. I am feeding various joint angles, speeds and the position of the head and abdomen. The human body does have supporting processes for motor control such as local reflexes that help stablise joints in the absense, or in addition to, central drive but if I get this right my GA/ANN should take this into account. Its not so much about doing things exactly how the human system would do it, but producing optimised patterns that the human can copy.

Many of the skills I'm going onto next will be open in nature (that's a phrase from the motor skill literature) meaning that they can't rely too much on feedback. In other words they have to produce a discrete quick movement. Other skills will require a mix (open/closed). The standing one is something that does rely on feedback and so is closed. Until the brain upgrade I can't produce good open skills I believe so that's why I started with the standing skill.

I should also try going back to non-motorised joints using external torques and forces instead. As described in the document the final joint was far more resilient and should work. This way the body will be far more dynamic. At the moment, if everything is 0 then the body becomes rigid even as it falls. With an external forces approach the body crumples. This added dynamism may be harder to control but should increase movement possibilities.
JoeWright
 
Posts: 69
Joined: Sat Apr 30, 2005 1:42 pm

Re: AI learns to stand

Postby zalo » Sun Mar 29, 2009 6:11 pm

Wow, after about an hour, you could come up with a system to combat euphoria!
I always loved evolution as a method for developing behaviors, and you executed it nicely.
I can't wait to see what Multi-Stage behaviors will be like. Good Job.
zalo
 
Posts: 2
Joined: Sat Jul 19, 2008 12:36 pm


Return to User Gallery

Who is online

Users browsing this forum: No registered users and 4 guests

cron