Alt-BEAM Archive
Message #11380
To: beam@sgiblab.sgi.com
From: Bruce Robinson Bruce_Robinson@telus.net
Date: Sun, 05 Mar 2000 18:25:56 -0800
Subject: [alt-beam] Re: lobster
Bumper314@aol.com wrote:
>
> All the info I see on Lobster seems to say that it was not a success.
Lobster seemed to do OK in low-stimulus environments. It got "confused"
when it got into higher stimulus areas. Keep in mind the specs:
- 5 motors.
- 6 Nv net.
- 8 Nu net.
- 6 tactile sensors (4 on the legs).
- The Nu network was a layer between the sensors and the Nv
network.
Walkman did much better, much to everyone's surprise.
- 5 motors.
- 4 nv net.
- 2 Nu's.
- 5 tactile & 2 visual sensors.
- most of the sensors acted directly on the Nv neurons.
In a sense, you could say that Walkman had a much more primitive brain.
> So in the argument of 5 motor walkers over 4 or 3, I guess the
> circuit is just as important as the frame.
Circuit & mechanics have to me matched to each other. The more
complicated you make it, the more work you have to do to fine tune it
... and I'm talking about weeks, not hours.
> I hate how people try to push the jittery first steps before
> the PNC has set the right pulse as the bot actually learning.
The PNC has nothing to do with "learning". You can make a startup
circuit that will put exactly one impulse exactly where you want it the
instant your circuit powers on. You could argue, in some circumstances,
that the steps taken AFTER the PNC has set the pulse is learning ...
Tilden discussed a type of learning in Living Machines, which I think is
perfectly valid. He would turn the legs on a machine 180 degrees from
their "normal" position and see how many cycles were needed before it
achieved a walking gait (no leg stops, obviously). Is this learning?
Consider:
In humans, learning is not just the intellectual process we all
experience ... there is another type of learning called "motor learning"
(yes, that's the scientific name for it), during which we learn to
control our various limbs. We are not consciously aware of how this
process works, but around age 18 months, over a period of months our
limbs & nervous system learned how to maintain some kind of equilibrium
while moving us about.
This is exactly what robots like Walkman & Lobster do when they are
first fired up. The difference between us & them is that we have to
learn to control tens of thousands of muscle fibres, controlled by
several times that number of neurons, to operate numerous joints with
many degrees of freedom. The robots have to control only a few motors,
with only a few neurons, and only one to two degrees of freedom. And
whereas we learn to walk in a complex environment, theirs is much
simpler.
So the learning Tilden talks about is exactly analogous to the motor
learning we do when we learn to perform complex motions. It is only the
complexity that differs.
> Richard Caudle wrote:
>
> The term that has been bandied about for this is emergent behavior,
> but I'm none too sure about that one either.
After some interesting (and not too successful) efforts at making more
complex systems, I am getting a handle on "emergent behaviour".
You can do as much design, modelling, and theoretical work as you like.
When you actually start to build a slightly complex robot, you quickly
find that things just don't work as planned. Somewhere between the
mechanics and the electronics, something doesn't quite fit. So,
providing you don't quit in frustration, you fiddle about and possibly
get it functioning ... sort-of.
If you're lucky, you might notice that when THIS motor is at THIS angle,
and the circuit delay is THAT long, then the motor seems to behave in a
consistant fashion -- even if other parts of the system are in different
configurations. This, I think, is emergent behaviour: a reproducible
behaviour that exhibits itself under varying conditions. Behaviour that
is a little too complex to determine theoretically in advance.
Just a few thoughts, to stimulate some intellectual learning :)
Regards,
Bruce
11381 Sun, 05 Mar 2000 18:28:18 -0800 [alt-beam] Re: lobster beam@sgiblab.sgi.com Bruce Robinson BUDSCOTT@aol.com wrote:
>
> how does that microcore work anyways
A bicore uses two Nv neurons linked together in a loop. A microcore uses
4 (sometimes people say 4 or more). Once you go beyond the bicore, you
can link neurons in loops, chains, branches, etc. Chiu used a hexcore to
power his 3 motor walker.
> ... i'm confused about this microcore business, is it obsolete ...
Sure hope not. I've devoted a lot of time to it :)
> or is it just more difficult than the bicore?
Depends which one you got working first :)
Bruce
11382 Sun, 5 Mar 2000 22:28:25 -0500 (EST) [alt-beam] Re: lobster beam@sgiblab.sgi.com chipuser ------- Original Copy -------
>Subject: Re: lobster
>Date: 03/05/2000 9:55 PM
>From: Bruce Robinson
SNIP
>So,
>providing you don't quit in frustration, you fiddle about >>>and possibly
>get it functioning ... sort-of.
>
>If you're lucky, you might notice that when THIS motor is at >THIS angle,
>and the circuit delay is THAT long, then the motor seems to >behave in a
>consistant fashion -- even if other parts of the system are >in different
>configurations. This, I think, is emergent behaviour: a >reproducible
>behaviour that exhibits itself under varying conditions.>Behaviour that
>is a little too complex to determine theoretically in advance.
>
>Just a few thoughts, to stimulate some intellectual learning :)
>
>Regards,
>Bruce
>
Is it safe to say emergent behaviour = emergent understanding
on the builder's part or the learning of the builder instead
of the bot ?
ChipUser
Get 100% FREE Internet Access from Freei.Net. 100% FREE, 100% Anonymous, 100% Jam Packed with features. Check us out at http://www.freei.net.
11383 Sun, 05 Mar 2000 21:08:38 -0800 [alt-beam] Re: lobster beam@sgiblab.sgi.com Bruce Robinson chipuser wrote:
>
> Is it safe to say emergent behaviour = emergent understanding
> on the builder's part or the learning of the builder instead
> of the bot ?
I wouldn't say they are synonymous, but they sure are closely linked.
The behaviour is there, whether you recognise it or not. But it isn't
much use if you don't recognise it. Recognising the emergent behaviour
is definitely "learning of the builder".
Emergent behaviour is not precisely robot learning. Connect some
mechanical parts to an electronic circuit, and produce some repeatable,
robust behaviours under a variety of conditions -- that's emergent
behaviour (in my view -- subject to much debate, no doubt).
Next, incorporate that behaviour into a robot, start it up from a random
position, and watch it settle in to a controlled, systematic movement
pattern -- that's motor learning, however primitive. Since our robots
aren't able to retain the stable configuration the way we humans are,
they have to re-learn it each time they start up.
It's interesting to consider this in terms of more complex robots. Each
time Walkman fired up, it would have to "learn" how to walk all over
again. No problem ... it might flail around very briefly, but no harm
done. Now try this with a bigger, much more complex robot, with several
motors per leg. If it had to "learn" to walk each time, it may do some
serious damage during the learning process, both to itself and its
surroundings. This means we have to figure out some way for a robot to
retain "learned" configurations when it is shut off.
Bruce
11384 Mon, 6 Mar 2000 00:03:13 EST [alt-beam] Re: lobster beam@sgiblab.sgi.com Bumper314@aol.com Sounds pretty good to me, but I still dont think its much of learning. the
motors are reacting to the environment and attinuating the signal of the Nv
(I believe this and have viewed it on my o-scope), so I guess if you want to
call that learning fine, but then again i would say an ant doesnt learn, it
is hard wired for its environment just as we do for our bots.
Can anyone work out a schematic of what they think or know walkman is. I have
the topology but thats not enough for me to go off of. I am interested in
that whole flip on the LEDs to see where things are and stuff.
And why do you think tilden doesnt work with 5+ motor walkers anymore, all
you see him doing is photovores (ants) and such.
And you said 4 Nv to controle 5 motors, isnt chiu's 4 Nv to controle 2
motors, seems like you would get more behaviours out of this 2 motor walker
then walkman.
Steve
11385 6 Mar 00 00:01:23 CST [alt-beam] gumby legs beam@sgiblab.sgi.com Steven Dang Does anyone know where to get gumby legs, or 8 gauge wire, besides
solarbotics.
Steven Dang
____________________________________________________________________
Get free email and a permanent address at http://www.amexmail.com/?A=3D1
11386 Sun, 5 Mar 2000 13:12:17 -0800 [alt-beam] Re: lobster "Dennison Bertram" By observing the robots that Tilden seems to stick with, is one way for us
to indirectly observe which robots are 'prefered'. As Tildens whole stance
goes, People are the independent evolutionary agents for the robots. The
robots who perform the best, get evolved, (rebuilt, redesigned, repaired, or
simply reproduced) The robots that have less promise, don't get reproduced.
True, walkman is the poster child of BEAM. As it's ablities are certianlly
someting amazing. But there is something to be noted about this. While
Walkman is extremely impressive, it's not the future of BEAM. If anything it
was a pleasent accident that won't be often repeated. It's ablities are yet
extremely limited. It's perception of the world for example is limited, and
it would seem if you look at the topology, there isn't much room for
improvement either. As a whole Mark T has moved away from the microcore,
opting for master/Slave bicore arrangements.
The reason for this switch is fairly obvious. It would seem as though the
microcore doesn't hold as much promise as the bicore does. Looking at the
developments with the Unicore it would seem nearly obvious that the Unicore
arrangement suddenly offers BEAM something it didn't have before. The
concept of a 'head' is finally initially being realized. While the nervous
net in Walkman can adaquatly manuver itself around it's world, it is indeed
not the embodyment of what Mark T would want BEAM to finnaly be. Remeber the
idea is to create wild robot's that can eventually be domesiticated.
Walkmans design, and structure, is undomesticatable, and in addition, never
can be. In my opinion, it's evolutionary branch is (currently anyway) at
it's end. The simpilicty for such a complex bot was shocking indeed, yet
while we start small we aren't going to produce our dream machine by going
backwards.
Why tilden would seemingly abadon the entire 5 motor platform is a mystery.
If anything it's possible that the four motor arrangement at this time
provides enough 'entertainment'. The stryder design, with it's master/slave
bicores is a profoundly interesting and expandable design. If one were to
keep the unicore in mind, it's obvious the bicore is just at the beggining
of it's evolutionary niche. So, Ian may have a point about five motor
walkers being the 'way to go'. But if you really want to do some research on
Nv Net topology, and try to figure out how things really work. You might
want to at least do a 4 motor platform. Loads o' fun.
dennison
>>>>>>
Sounds pretty good to me, but I still dont think its much of learning. the
motors are reacting to the environment and attinuating the signal of the Nv
(I believe this and have viewed it on my o-scope), so I guess if you want to
call that learning fine, but then again i would say an ant doesnt learn, it
is hard wired for its environment just as we do for our bots.
Can anyone work out a schematic of what they think or know walkman is. I
have
the topology but thats not enough for me to go off of. I am interested in
that whole flip on the LEDs to see where things are and stuff.
And why do you think tilden doesnt work with 5+ motor walkers anymore, all
you see him doing is photovores (ants) and such.
And you said 4 Nv to controle 5 motors, isnt chiu's 4 Nv to controle 2
motors, seems like you would get more behaviours out of this 2 motor walker
then walkman.
Steve
Home