|
Grex > Cyberpunk > #17: Robot revolution at MIT: Developmental psych meets AI. |  |
|
| Author |
Message |
other
|
|
Robot revolution at MIT: Developmental psych meets AI.
| 
Jan 2 18:57 UTC 1995 |
I just read a fascinating article in December 1994 WIRED about MIT
Media Lab's "Cog" project, headed by Rodney Brooks.
"Cog is a vaguely humanoid robot with minimal programming which is
being "raised" in a manner similar to human infants. Necessary to this process
is constant input through several sense simulating sensors (vision, hearing
and touch) which I imagine go to a constant memory. The point is to see if
the robot goes through a human-like "learning" and development process by
which it learns to relate to the real world.
This approach is revolutionary in that it is real world based and not
dependent on a programmed simulation or map of the real world contained within
the robot's memory.
I am wondering about the specifics of this research, since it follows
an approach which I independently thought of and toyed with in my head starting
in about 1988. I have begun a correspondence with the author of the article,
and hope to continue it with Dr. Brooks or one of the project team. This
item is where I plan to share this information. Your comments and insights
are welcome, and since my background is not of a strongly technical nature,
I'm hoping that this forum will allow the more deeply technical concepts
involved in this project to be shared, through the translation of those here
with the appropriate background, with those of us interested laypeople.
Thanks!
|
| 83 responses total. |
kentn
|
|
response 1 of 83:
|
Jan 2 19:40 UTC 1995 |
It's also an old science fiction plot. What I'd like to know is, do
they cloth or disposable diapers?
|
other
|
|
response 2 of 83:
|
Jan 3 03:44 UTC 1995 |
Can you refer me to any science fiction stories in which this plot line was
pursued? I'd be interested in seeing how it was handled.
I'm curious about the nature of the memory setup in Cog, and what some
of it's basic "instinctive" programmed drives and responses are.
|
peacefrg
|
|
response 3 of 83:
|
Jan 3 04:09 UTC 1995 |
Eric, Get some Asimov. The I Robot series
|
sun
|
|
response 4 of 83:
|
Jan 3 07:17 UTC 1995 |
I think that AI is interesting, and a good venture. But then again,
what if they succeed? do YOU want to tell it that it is not human?
|
kt8k
|
|
response 5 of 83:
|
Jan 3 16:18 UTC 1995 |
Yeh - like the pets we raise, it will probably consider itself human until
told otherwise. Humorous implications abound - it might make a better
sitcom than Mork did! Watch out Northern Exposure!
|
other
|
|
response 6 of 83:
|
Jan 3 17:14 UTC 1995 |
I'm sorry if I was not clear, but I entered this item specifically to discuss
this line of research and its methods, and *not* its social implications.
That would make for fascinating, but much trodden, discussion, and so I
request that it please be taken to another item.
Here, I would like to further my own understanding, and that of anyone else
who is interested, in the mechanics and the processes of this particular
project.
Thanks, and I apologize again for the lack of clarity in intention.
|
cyberpnk
|
|
response 7 of 83:
|
Jan 3 20:52 UTC 1995 |
I just got through reading a book called "Brainmakers" where they talk about
just this sort of thing, and there's one section where it talks about
how scientists seem to bo getting better results by going from the ground up,
that is, by making insect like AIs first, and then building up to human
level AIs.
|
general
|
|
response 8 of 83:
|
Jan 3 22:16 UTC 1995 |
WOW! I had no idea that robotics has actually come this far already. I
expected to be dead by then.
|
raven
|
|
response 9 of 83:
|
Jan 3 22:53 UTC 1995 |
I'm linking this to the cyberpunk conf. Join cyber at the Ok:.
|
aruba
|
|
response 10 of 83:
|
Jan 4 01:20 UTC 1995 |
I'm curious what the software/hardware of this robot is like. Can you tell
us Eric? Are there any AI buzzwords I might recognize from my one
undergraduate course in AI?
|
kt8k
|
|
response 11 of 83:
|
Jan 4 16:25 UTC 1995 |
"learning" experiments with some of the neural network circuitry
released in the last decade or so have been tried and documented in
even such places as Scientific American. Check the readers guide to
periodical lit. at the library. Also, electronic trade magazines
such as Electronic Design have periodically covered release of
integrated circuits containing modular neural circuits with some
very interesting properties. They are easier to implement than
the previous method of creating neural functions in software, which
was a major undertaking in itself.
|
other
|
|
response 12 of 83:
|
Jan 4 16:59 UTC 1995 |
What I know of this actual line of research is limited to what I read in the
WIRED article, but I'm still waiting for some mail back from a couple of folks
I have written to find out more.
When I was thinking about this approach years ago, what I was hung up
on in the process was the nature of the memory involved. We were still too
far then from any reasonable simulation of human memory function
(read: parallel) for me to imagine that it was possible to pursue this
approach. Yes, there was the junction of separate processors in parallel
function, but the advent of fuzzy processors really is the key, I think.
There is still a tremendous amount about the memory hookups, and the
basic "instinctive" instruction that I'd still like to hear about.
If anyone reading this knows or can find out the e-mail address of
anybody on the research team, I'd love to open a dialogue with them and share
whatever they're willing to share of the information.
|
raven
|
|
response 13 of 83:
|
Jan 4 19:20 UTC 1995 |
A good basic resource for bottom up A.I. info is Steven
Ls book "Artificial Life." It's a good primer on complex adaptive
systems, neural nets, insect like bots, etc. [err Stven Levy]
|
kt8k
|
|
response 14 of 83:
|
Jan 5 22:54 UTC 1995 |
I agree w. #13. There is a lot of info available. Memory isn't
memory as we're used to it in computers, but rather adaptive
circuitry that simulates neurons' ability to define connections in
response to stimuli. It's neat!
|
dadroc
|
|
response 15 of 83:
|
Jan 11 15:31 UTC 1995 |
We do not have nural-net_parallel_hierartical hardware yet, don't expect
to talk Kafka to a 'droid in the near future.
|
snaed
|
|
response 16 of 83:
|
Jan 25 11:45 UTC 1995 |
well... i am new to this but i have always had an interest in cyberpunk... but
i never had the time or money to invest in it thru work and all maybe someone
can help me????
|
nephi
|
|
response 17 of 83:
|
Jan 25 14:44 UTC 1995 |
Type "j cyberpunk" (without the quotes) at your nearest Ok: prompt.
|
other
|
|
response 18 of 83:
|
Jan 25 15:37 UTC 1995 |
Well, i sent off a note to Dr. Brooks, but have gotten no response. I guess
a second try is in order.
|
other
|
|
response 19 of 83:
|
Mar 17 22:17 UTC 1995 |
second try is as follows:
To: cdp@ai.mit.edu
Re: Cog info please!
Hi. I was quite pleasantly surprised to read about the nature of this
project in the December issue of Wired magazine. In 1988, I took an
introductory Cognitive Science course taught by Martin Ringle at Vassar.
During that time, I gave much thought to the idea of taking a real world
learning approach to "growing" an intelligent artifact. I was stumped by
the methodology of constant memory input from several sensory devices, and
the storage and integration of such input into a cohesive 'picture' of the
reality within which the artifact exists. My model was the human infant,
and the processes by which it learns to distinguish first light and dark,
and then other physical distinctions such as self and other. If you would
illuminate some issues for me, I would very much appreciate it.
1) What are some of the basic "instinct" commands which guide Cog?
2) Tell me a little about how Cog is powered, and whether one of its drives
is the recharging of its power supply or whether that is a function unrelated
to its development. (It makes for a good simulation of the human survival
instinct!)
3) How is the memory storage done, and is it constant, as are the inputs?
4) Does Cog have a "cry?" I.e. a universal signal of need, to suffice until it
can learn more sophisticated methods of communication.
5) If so, what are the circumstances which trigger the "cry?"
6) Has Cog shown any ability to connect "memories" of certain events with
current experiences which repeat those memories? If so, tell me a bit about
the process which allows this.
I hope that these questions will help you with your documentation efforts as
well as helping me understand the nature of the project's successes.
If you wish to see the use to which I am putting this info for now, there is a
conference item in the bbs at grex.cyberspace.org (telnet accessible) about
your project. This is for no purpose other than to further my own
understanding of the technology involved in making a reality of an idea I never
thought would actually be pursued. Keep up the good work, and thank you! -Eric
R. Bassey
|
brenner
|
|
response 20 of 83:
|
Apr 14 12:36 UTC 1995 |
(now linked to the accordion, so that the caltech
robot groupies can particiapate!)
|
vsclyne
|
|
response 21 of 83:
|
Apr 14 13:07 UTC 1995 |
Ditto the recommendation to read S. Levy's "Artificial Life."
Also, check out the relatively new "Journal of Artificial Life."
A rigorous, huge, densely pack tome by Stuart Kaufman, "The
Origins of Order" is also relavant.
"Artifical life" is an unfortunate name for this field, in my
judgment. It helps to take it seriously if you think in terms
of "self-organizing software systems", but "artifical life" is
the accepted keyword combination for literature searches.
|
humdog
|
|
response 22 of 83:
|
Apr 14 13:43 UTC 1995 |
i read about cog in an english science magazine; and i believe
also in the cave there was discussion of the wahbot, do you remember
that?
i like moravec's writings lots and lots
and i agree with shannon about the artificial life thing.
please let me know when you are up to my rant on robotics
and the humanities...
|
brighn
|
|
response 23 of 83:
|
Apr 14 16:23 UTC 1995 |
Participate? It was a dead item that we have now claimed! bwahahahahaha
Thye project sounds fascinating, but i dunno whether it'll work.
And the social implications, despite the originalposter's
intentions, are the most important issue.
|
vsclyne
|
|
response 24 of 83:
|
Apr 14 17:03 UTC 1995 |
It will work, but slowly. There's time to
accommodate, even co-evolve.
|