|
 |
On 24/05/2011 12:29, Jim Holsenback wrote:
> http://itee.uq.edu.au/~ruth/Lingodroids.htm
>
> Noticed a BBC news article that led me to the above link ... wonder why
> after reading I couldn't help but think, shades of "Terminator" and
> "Skynet"
Well, they have laser range-finders. So long as they don't have an
Aperture Science Hand-Held Dual-Portal Device, we should be OK.
If you haven't, go read the paper referenced at the bottom. It's very
readable.
Oh yeah, SLAM:
http://en.wikipedia.org/wiki/Simultaneous_localization_and_mapping
It's a bit non-obvious.
Although these robots don't "do" much, the theory behind it is quite
interesting. It's easy enough for one robot to explore and remember
where it's been, but how do you make it talk to another robot, who might
not even have visited some of these places?
The idea is that the robots wander around, learning about their
surroundings. In the where-are-we game, one robot asks the other "where
are we?" If the robot has a name for the place where they are, it says
that. If it doesn't, it invents a new name (and remembers it). The first
robot remembers the answer. Repeat this a few times from different
locations, and gradually both robots come to call the same places by the
same names, and to mutually decide where one named area ends and another
begins.
Then use the go-to game to test whether it worked. One robot picks a
place name at random, says it to the other robot, and then they both
drive there. Given the size of the office, if both robots drove to a
random place, they would be within range of each other 9.2% of the time.
But when they actually played the go-to game, the robots came to the
same place 38% of the time. Which isn't bad, but isn't 100%. (Apparently
part of the problem was navigating around physical obstacles - including
the other robot.)
They also did similar things for inventing names for distances and
angles. Then one robot can say stuff like "if I start at X, facing Y,
and travel Z, where would I end up?" An interesting thing about this is
that one robot can ask the other about places which are impossible to
get to. (E.g., if I start in that corner and travel 3 meters through
that brick wall, what's the name of the place where I'd be? Well, the
robots can invent a name for it, but they can't actually *go* there.)
One of the fundamental things about language is that you can use it to
obtain information about things that you yourself have never
experienced. It's kind of abstract, but really useful.
I find myself pondering several interesting questions about this
research. For example, they configured the robots to assign names to
areas roughly 3m across. But real people don't assign meanings to areas
of a particular /size/, but rather a particular /significance/. The
individual streets of a densely-populated city all have individual
unique names, but entire tracts of uninhabited forest or farmland have
no designated names at all.
For that matter, I've noticed that the apparent distance between two
places is somewhat unrelated to the actual geometric distance, and far
more related to /how interesting/ the route is. A route with lots of
memorable landmarks "seems longer" then a similar route that's fairly
featureless.
Also, place names are rarely unique. (Actually, forget "place"...
/names/ aren't very unique.) There just aren't that many combinations of
letters [which are utterable]. Instead, place names at least tend to be
hierachical. If I say "Church Street", you have no hope in hell of
figuring out where the **** that is. But if I say "Church Street,
Wolverton", there's some danger of you finding that.
I think a more interesting experiment might me to get multiple robots
(not just two) to cooperate in some goal-oriented task (rather than just
exchanging words for the fun of it). Say, put them in a maze, and
scatter coins around, and get them to work together to collect all the
coins. Have one robot say to another "if you go /that/ way, there's
loads of coins over there"...
Post a reply to this message
|
|
