Subject:
|
Re: Ultimate Project
|
Newsgroups:
|
lugnet.robotics
|
Date:
|
Fri, 12 Feb 1999 16:33:49 GMT
|
Viewed:
|
1556 times
|
| |
|
|
Richard Franks writes:
> Brandon Blodget wrote:
> >
> > lego-robotics@crynwr.com wrote:
> >
> > > Just wondering what you all think the ultimate
> > > project might be. Personally, I'd like to eventually do an RCX/PC
> > > AI or artificial life project that allows the PC to become self-aware
> > > by investigating and manipulating its environment through the RCX.
> >
> > I nominate a Lego "assembler" as an "ultimate" project for a RCX (or more
> > likely a group of RCXes with a PC). I've seen a few Lego brick sorters,
> > an assembler seems like a logical next step. Once you have all those
> > bricks programmatically sorted the assembler program could use this
> > information to retrieve the bricks and build something.
>
> Try putting some bricks together using your knuckles - that is with
> pressure feedback, now try doing it in a dark room :) It's a cute idea
> but I think you'd have *real* trouble with anything smaller than 2x2
> bricks, or is that what you meant?
>
> It probably isn't impossible.. but I'd imagine that you'd need:
> * vision/pattern matching to resolution were it could distinguish 3x1x2
> plates from a 1x2 brick
> * A variety of manipulators: pressure-feedback gripper, poker/puller,
> seperator etc
> * Layered construction? Would require determining what order to build
> sections in.
>
> I think that is a few levels beyond even a few RCXs at present.. maybe
> in a year or so some of these problems will have been solved?
>
> But then it depends on what sort of assembler you mean - making a 2x4
> brick wall is obviously simpler than getting it to build your SuperCar!
> :)
I think all the above are legitimate problems. I was talking to some friends
about this the other day. One major problem is the fact that sturdy robots or
other similar constructions are usually made up through cross-bracing. To put
together a brace you usually need to got through a delicate "hold everything
together till it snaps in place" action. If you get it wrong the whole thing
just explodes. This sort of operation can only be done by something which can
somehow hold a whole set of parts together until they can do it themselves. The
skills to do this are similar to what a nano-bot would need to assemble
molecules from atoms.
>
> > If you provided a way via genetic
> > algorithms, for these creations to exchange genetic information, and to
> > pass on to the next generation, you could watch your Lego creations
> > evolve. It would probably be a good idea to program in natural death so
> > the pieces could be re-used. I can imagine an ecosystem evolving where
> > the different Lego creatures compete for the scares Lego resources and
> > somehow a homeostasis is maintained.
>
> I like that idea! If they competed for bits of Lego by wandering about,
> picking some up and putting it in their 'area' then that would be quite
> fun. You'd need some electronic scale and the one with the heaviest pile
> of Lego would win :)
>
> Or you could give each robot a color or type of piece for which they get
> extra points for, but with the scale you could have automatic
> score-feedback.
>
> You can hold the data for many complex genetic algorithms in 32k, so
> they could time share or something - it seems wasteful having one GA per
> RCX.
Unfortunately I don't think Genetic Algorithms are the panacea everyone seems
to think they are. One *major* problem is that the bottle-neck of any GA is the
evaluation stage; in this case the building and testing of robots, whether
against each other or via some fixed metric. Most successful projects which
apply GA's to robots only apply it to the behaviour and not the morphology, and
even then much of the evolution occurs in a simulation with only a little time
spent evaluating them in the real world.
I would be quite interested in applying a GA, or some other optimiser, in an
on-line or off-line way, to the behaviour of the robot. I had an idea a while
ago for this sort of thing in a game (see
http://www.dejanews.com/[ST_rn=ps]/dnquery.xp?ST=PS&QRY=*&defaultOp=AND&DBS=1&format=threaded&showsort=score&maxhits=100&LNG=ALL&subjects=Why+do+wargame+AIs+suck&groups=comp.ai.games&authors=Moran&fromdate=&todate=);
the basic idea is that you can have an on-line test-bed (the robot) which is
periodically given a new brain (program) which is picked randomly from a pool
of them which are bieng evolved off-line (back at the PC in a simulation). The
fitness of programs which ran in the robot are then back-fed into the off-line
PC simulation. The fitness metric could also contain a "confidence" component
which would effected by how often this chromosome had been tested in the real
world. The ideal of this system is to get around the evaluation bottle-neck
whilst making sure the chromosomes work in the real world.
|
|
Message is in Reply To:
 | | Ultimate Project
|
| (...) Try putting some bricks together using your knuckles - that is with pressure feedback, now try doing it in a dark room :) It's a cute idea but I think you'd have *real* trouble with anything smaller than 2x2 bricks, or is that what you meant? (...) (25 years ago, 10-Feb-99, to lugnet.robotics)
|
17 Messages in This Thread:
 
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
