What's a Robot?

    Troll.

                John Nagle

Ouch.  Not warranted sir.  Not at all.


-Martin

You don't look like a Troll to me.

Maybe a robo-troll? :)

It's not really important how people choose to use the word robot. :)

I do agree that the use of the word has drifted far from it's original
intent but it doesn't bother me. I like all the machines called robots.
The more autonomous and intelligent, the more interesting they are, but I
don't feel the need to debate what we should or should not call a robot.

I do agree that if you want to understand intelligence then learning is a
key factor and any machine without strong general learning is hard to call
intelligent.  But I see no need to limit the definition of robot to only
intelligent autonomous slave machines even though that's where the idea
started.

--
Curt Welch                                            http://CurtWelch.Com/
curt@kcwc.com                                        http://NewsReader.Com/

intent but it doesn't bother me. I like all the machines called
robots.
The more autonomous and intelligent, the more interesting they are,
but I
don't feel the need to debate what we should or should not call a
robot.

It's not like I loose sleep over this...far more important stuff to
worry about these days.

I just think that terminology has to have a meaning.  Not to pound on
the MythBusters guys (a show we all love at home) but they call
anytihng with a motor and a switch a "robot".  By that measure your
electric power seat in the car is a robot.  Or, better yet, cars that
have power seats with preset position memories are ten times more of a
robot.

Maybe that's what I am grappling with.  To me a robot has always been
something special.  With intelligence or a reasonable simulation
thereof as a prerequisite.

Of course, there are stages to the development of a robot.  If you are
studying bipedal locomotion and have a machine with two legs attached
to a pole so that it can walk in circles without falling over, well,
it may be fair to call that a robot or "part of a robot".  It isn't
realy intelligent but it is, presumably, part of the R&D process for a
real robot.

An air motor attached to a samuai sword with a manually operated valve
is not, a robot, despite MB claims during these shows.  Again, not
picking on them at all, it just happens to be a very prominent place
where the term is used over and over again.

Battle-bots are another example.  They are not robots.  They are
remotely controlled weaponized vehicles.  You take the human away from
the remote control and they are absolutely useless (as neat as they
may look).  A real battle-bot would have a remote with a few buttons,
for example:  GO, STOP, AGGRESIVE, NEUTRAL, DEFENSIVE, etc.  In other
words, human input isn't precluded from the process, but it might be
limited to requesting a state or strategic posture and then the battle-
bot would --by itself-- conduct and run the battle.  Big difference.

-Martin

I agree that Mythbusters takes calling everything a robot a little to
far.  I had issues with the whole sword swinging robot and such.
Anytime they ask Grant to do something its hey can you build a robot
to do that.... and its pretty much a servo with a remote trigger.  but
dont get me wong the guy is a class act on RC things.  but robots
thats a little iffy.

on the whole Battle bots thing. I would love to see a battle something
like the robo cup.  The robots have to 1 identify targets then 2
destroy the target.    Points could be awarded on how fast a target
was aquired and how acurate it was attacking.

It's a great show.  I love it.   I worked in motion picture special
effects for a while and know just how capable FX guys can be (and how
much fun the work can be!).  These guys are top notch and the show is
fantastic.

Here's another one that might really drive the point home:

I love it when they rig a real car for remot control.  I can't imaging
how much fun it might be to drive a full size remote control car.
Just brilliant.

Now, compare that to the DARPA autonomous driving challenge --which I
am sure folks on this list are very familiar with--.

Which of the two scenarios is a robot?  I would be astounded if anyone
considered calling the full-size R/C car a "robot" as it isn't even
close by any stretch of the imagination.  The driver of that car could
negotiate he DARPA challenges without any trouble whatsoever.
Building a real robot that can do the job is far, far more complex and
expensive than that.

With regards to the utility of debating the term.  Debate isn't for
everyone.  This is not an insult, just a statement of fact.  Some
folks simply have no interest in this or don't see the point.  And,
from their frame of reference, they are absolutely right.  However,
those who can and do consider it are generally open to learning and
changing their views of the world based on what they may hear.  That
is the very foundation of the scientific process.

-Martin

Debating word usage is science?

Not by my standards. :)

My type of fun is debating how to make robots act like humans.

--
Curt Welch                                            http://CurtWelch.Com/
curt@kcwc.com                                        http://NewsReader.Com/

You are correct.  But if suppliers had to meet a higher standard in
order to sell "robots" rather than remote controlled toys we might
just have some progress in this field.  Do we call R/C airplanes
"robots"?  What is Graupner (R/C stuff manufacturer) started to call
their R/C planes and boats "robots".

The hobby robotics field has had a couple of false starts over the
last 25 years or so and left a couple of failed magazines and
companies in the wake.  I posit that one of the reasons is that these
"robots" really run out of steam as soon as you get them to dance
around and do a couple of tricks.  Then what?  I've watched my own
kids go through this cycle.  I still have a HERO (who remembers those)
brand-new in-the-box as well as a HERO arm (also NIB) in the garage.
Those were nice kits.  Probably as close to real robots as you could
have expected to get back then.

Today's kits are, in my opinion, not too different...which isn't to
say much because microprocessors are 1,000 times more powerful and
capable today than they were back then.  HERO could be programmed to
navigate around a room, make sounds, sense various things, move its
arm, etc.  I believe the processor was a 6502 (my memory may fail me
here).

And, speaking of 6502's, back in '83 timeframe I build a walking
hexapod that used 11 Rockwell R65F11 microprocessors operating in
parallel for control.  The R65F11 was a 6502 with FORTH built in.
This hexapod was two feet tall and it used TRW servomotors designed
for fast control surface positioning.  I know that I considered this
to be a robot back then (I was young and didn't have a lot of
experience).  Thinking back, it wasn't.  It was a fine example of a
programmed machine...but that's about it.  After making it walk around
the room, use touch pads on its feet to sense obstacles and back up,
go up a flight of stairs using a similar technique...well, what else
could you do with such a thing without INTELLIGENCE.  And that's my
point.  And so my focus changed to understanding intelligence rather
than mechatronics.  Mechatronics is easy.

Without intelligence these things are nothing more than programmed
sequence CNC machines that get old pretty fast.

I don't want to debate the word as much as highlight the fact that
until we change the focus from remote controlled, animatronic or CNC
machines to seeking to build intelligent machines, the field will not
reach its potential and progress will be curtailed.  Thankfully there
are lots of universities where this is exactly the case.  My greater
point, perhaps, is that young ones, when expose to the misuse of the
word may actually believe that they are dealing with robots when, in
reality, all they are dealing with are toys that they can program to
execute an unintelligent sequence of moves.  I am making a point to
teach my kids the difference in hopes that the challenge to elevate
these things away from toys might inpire them to create the next
revolution in robotics.


"Acting" is relatively easy.  Just go to any good attraction at
Disneyland.  THINKING and LEARNING are very different matters.


-Martin

I, too, am bothered that the term 'robot' is applied to too many things that
are simply machines.  Especially R/C items that are 100% controlled by the
human operator.  I used to fly r/c planes and would never consider any of
them robots.



But this is a difficult classification.  It's like there is a fine line
dividing robot from non-robot, but then I think there are also shades of
gray in between.  For example, in the r/c planes I used to fly, I
incorporated a missing pulse detector - if contact with the transmitter was
lost, then some of the servos would be sent to pre-set positions.  This
would allow the plane to, say, be put into to shallow turn and throttle back
to allow for some type of recovery.  Back then (ahem, in the 80's), this was
all done with analog circuits but the very same thing could be - and I'm
sure - is done today with microprocessors.  And you see servo-stabilizers
for r/c helicopters or high-flying planes to keep themselves level.  So the
line starts to blur.



I think it comes down to intelligence, as Martin brought up.  I think a
machine that interacts with the world with intelligence would qualify as a
robot.  (Hmm, a microwave oven???)  But the problem is simply deferred
because no-one has really come up with a good definition for intelligence.
Back in the 80's-90's I did some work in artificial intelligence.  Besides
never coming up with a good definition, I learned a very important concept
that might be applicable here.



In AI the goal is to make the computer/machine/root exhibit behavior that is
indistinguishable from a human.  As long as we couldn't get the machine to
'behave with intelligence' the desired behavior was called 'artificial
intelligence'.  As soon as somebody got the desired behavior, then everyone
said, "Well that is not intelligence .  That's just some clever
programming!"  However, many of the things we couldn't do back then are
common now-a-days: computer-generated speech, speech-recognition, trip
planning, computer vision, and so-on.



So, _my_ definition of robot would be - a mechanical device, controlled by
humans with high-level input, that interacts with the real-world with
intelligence and/or learning.



Battle-bots don't qualify. But the soccer-playing bots I saw on Nova last
night -THEY qualify!



Sorry if I rambled too long.



Jim

for sure!


Many simply consider the "artificial" just to mean "man made", as apposed
to fake or imitation.


The moving goal post problem. AI is always just out of reach because no
matter what you do, it's labeled as not intelligent.


Like the Google search engine computers?  They are mechanical devices
(everything is a mechanical device), it's controlled by humans which simply
type in English commands to it, and it interacts with the real world by
communicating with other computers all around the world and by interacting
with humans all around the world.

My point is that you probably weren't thinking of things like the Google
computers as robots but yet they basically fit your definition (and people
do call web crawlers robots by the way just to muddy the waters even more).


It all looks intelligent until you understand how it works, then people
just call it clever programming again.

It's almost as if it can't be intelligent if someone understands how it
works! :)


Oh, if you think that's rambling, you should see of the crap I post! :)


The problem here is that there is no dividing line between intelligent and
not intelligent.  It's nothing but a large continuum.  As such, you can't
find a place to draw the line between intelligent robot and non-intelligent
robot.

By the time we get machines acting like humans, there's going to be a huge
trail of prior-art that fills the entire gap between the wheel and human
level AI.

Currently there is a clear gap in intelligence between humans and
everything else on the planet (machines and animals).  This gives us a
clear line between "human-intelligence" and all other forms of complex
machine behavior.  It creates the illusion that humans contain a unique
technology of some type.  But AI and the general growth of technology is
going to fill that gap with a nearly infinite range of machines of
different sizes, skills, and abilities and once the gap closes, the word
intelligence is going to loose its foundation and all technology is going
to blur together as different types of intelligence - just like the word
ROBOT is blurring over that space already.

In the end, intelligence is probably going to have some vague meaning like
"adaptive learning system" and it won't have anything to do with humans per
se and all machines will be called robots.  And more important, we will
have lots of new names for all the new technologies that are developed.

--
Curt Welch                                            http://CurtWelch.Com/
curt@kcwc.com                                        http://NewsReader.Com/

   You don't need regulation (at least not directly). Just register
the word robot as a trademark, and license its use only to entities
which use the word as you see fit.


   And relevant to this thread is another thread from this very group
ten years ago (I found this in Google Groups despite all odds,
Google's Usenet archive search functions seem to be hosed lately),
where I asked about robots "Where are they?"

http://groups.google.com/group/comp.robotics.misc/browse_thread/thread/91b43ea7b8cc39bb?pli=1

Well, I was just 13 years old when the first Star Wars came out.
Maybe that "poisoned" my thinking.  I hand-build my first computer
with the Intel 8080 processor not too far from that.  I was
immediately focused on "robotics" because, as a kid, I wanted to build
my own R2D2 and C3PO.  It didn't take very long to realize that this
was impossible due to many reasons, the lack of good AI being one of
them.  So, I wouldn't call it a late realization on my part but rather
a realization that these days everything is a "robot" and probably
without justification.

The dialog 20~25 years ago was very different.  You couldn't build
cheap little machines with lots of servos and microprocessors easily.
It took a lot of effort, money and study if you wanted to attempt
something like that.  So, in many ways, I think that the field --in
the popular or hobbyist domain-- was taken a little more seriously.
Something like the HERO 2000 "robot" was a pretty serious investment.
I forget the price, but I am sure it was in the thousands.  And it was
a pretty decent platform to start a project from despite the
limitations in computing that existed in the day.

Anyhow, on the topic of intelligence...to me this has nothing to do
with behaving like a human unless you are trying to make a humanoid
and that's a requirement.  Even a humanoid doesn't necessarily have to
behave 100% like a human.  No, intelligence is more subtle than that.
To me it has to do with understanding and adapting to surrounding
conditions or the conditions of the problem to be solved.
Communication doesn't have to be in spoken or written language.  It
can be fixed-function buttons.  It doesn't matter.

Here's an example off the top of my head.  I live in Southern
California.  We have big fires here.  An R/C fire-fighting machine is
one that has no intelligence but a firefighter can remote control to
achieve telepresence and fight a fire.  An intelligent, and hence,
robotic fire-fighting machine is one that you could preemptively roll
out of a truck throughout a neighborhood that is potentially under
threat and TRUST that the robot is capable of assesing the level of
risk as it develops and take measures to save the neighborhood.  These
are very different machines.  One is intelligent, the other isn't.
One is easy to build.  The other is nearly impossible today.

In terms of implementation of intelligence in robots I am convinced
that genetic/evolutionary computing is the answer.  The approach has
all the makings of producing very capable thinking machines --after
all, every animal on this planet is the result of this process.  All
we have to do is learn how to implement it for our machines.


-Martin

I was in college so I was only a few years older.


I recently subscribed to the yahoo r2builders list because I've been
considering building myself an R2.  But as you say, without the AI, they
will never be as cool as the character in the movie.

Clearly, the "real" robots are all the science fiction robots like Robby,
B-9 (from Lost in Space), Gort, Huey, Dewey, Louie, RD-D2, C-3PO, Data, and
Wall-E.  The closer we can make are machines act like these guys, the more
"robot" they become.


Yeah, I don't have any problem with the idea that the word robot should be
used more like you want it used.  But it also doesn't bother me that that a
sword swinging machine is called a robot as well. :)


Yeah, as long as we have 4 billion years to wait for the design to get
done!

Come to comp.ai.philosophy if you would like to waste endless hours
debating AI with me. :)

I believe the solution to making machines act like humans (or any of the
animals with brains that act somewhat intelligent) is to build a strong
generic real time learning system trained by reinforcement.

The confusion most people have about AI is that look at intelligent
behavior and see lots and lots of complexity.  Human behavior is clearly
very complex.  We fill entire libraries with books trying to describe all
the different aspects of human behavior.

In engineering, the more complex the machine's behavior is, the more
engineering work we have to put into.  Complex computer systems are filled
with billions of lines of code these days written by thousands of
programmers.  And even after all the engineering effort it takes to create
these complex machines, they don't come close to acting intelligent.

Add up these two obvious facts and you get the obvious conclusion most
people come to - the idea that creating a really intelligent robot is going
to take a lot of hardware, and a lot of engineering time to duplicate the
complexity of what must exist in a human brain.

And, some, like you, seem to have come to the conclusion that this is just
too much work for a human to do (maybe it's just too complex for us to even
understand), and as such, the only way to do it is build a machine to do
the work for us - and evolutionary computing fits the bill.

However, it will never work - at least not in that form.  It's too slow.

What almost everyone fails to see, is that inteligence is not about the
behavior, it's about the ability to _learn_ behaviors and apply them
correctly to new situations.

Solving AI is not about duplicating each and every _adult_ human behavior
in a machine by witting lines of code for each behavior.  It's about
building strong, generic, behavior learning systems.

AI will not be solved by adding more code to all the AI code we already
have, it will be solved by figuring out how to make less code, do more.  We
need strong, totally generic, learning algorithms.  These algorithms will
be fed sensory data - but won't need to be customized for the type data it
is - you can feed any sensor data to it, and it will learn how to make use
of it.  And you can allow it to control any output - without customizing
the system for the type of output you are controlling, and it will learn
how to make use of it.

And it will do it with one, fairly simple, fairly straight forward, real
time temporal reinforcement learning algorithm.

The complexity of behavior comes not from the amount of code in the
learning algorithm but from the size of the state space the learning
algorithm is applied to (the size of the neural net for example it is
working with).

My definition of inteligence is acutally "reinforcement learning" but
almost no one understand what I mean by that.

What's interesting is that genetic algorithms are in fact reinforcement
learning algorithms - they are just reinforcement learning applied to a
machine design, as apposed to reinforcement learning applied to behavior
generation.  This, in my book, makes genetic algorithms examples of
inteligence.  So in that sense, if we used them to create AI for us, we
would be using an intelligent machine to design and build other types of
intelligent machines for us.  But it's a type of inteligence that's so damn
slow, it's pointless to even play with it for the job of solving AI - which
is really a job of finding a new type of learning algorithm - which I don't
believe a GA system has any hope of coming up with fast enough to be of any
use.  We have other intelligent machines to work with that will solve the
problem of finding the learning algorithms much much faster - humans.

When we find these better genetic learning algorithms, I think it will
finally make robots into what we have seen in the movies.  It will make it
trivial to build intelligent robots because all you have to do is drop one
of these learning algorithms into the machine, feed it the sensor data you
want, configure it for the number of effectors you want, write the code to
generate the reward signal to define the machines purpose, and then turn it
on and wait for it to learn (and help it along by teaching it).  Once you
have one "educated", to do the things you need it to do, you can mass
produced it by copying the memory.

In order for the machine to do complex and intelligent things, it does need
lots of complex code in it.  But the point here is that the complex code is
not (and can not be) written by humans - it's too complex for a human to
understand.  It is written by the learning algorithm instead.  And the code
the learning algorithm "writes" for us is nothing like we would  write.
It's like the weights of a neural net - it's something the machine has to
calculate from real data - it's not something a human is smart enough to
calculate on their own.  So the complex of behavior is produced by "code"
written by the learning algorithm.

This is actually very close to what you are thinking by saying we should
use GA techniques to create the AI for us.  In a sense, because GA
techniques are reinforcement learning systems, using a reinforcement
learning algorithm is sort of like using GA to "write the code" for us.
But it's really a very different implementation than how the GA techniques
work.

I think the type of learning algorithm I believe we need to solve AI is
close to being found.  People like Jeff Hawkins (On Intelligence) is
basically working on finding such an algorithm for example.  Lots of people
are looking in what I consider to be the right place, and making progress.
I think it might show up in the next 5 years (I have a bet I'll get it done
in 8 years if someone else doesn't beat me to it) - but that's optimistic.
None the less, It's going to show up soon - the next 25 to 30 years for
sure but I think a lot sooner.  And when it does show up, all these cool
robot kits will truly become ROBOTS instead of just RC toys that are too
stupid to know that it's bumped into a wall.

And at the same time, I think people are going to be amazed and even
stunned at how much intelligence can be created by even a tiny amount of
computing power in a robot.

--
Curt Welch                                            http://CurtWelch.Com/
curt@kcwc.com                                        http://NewsReader.Com/

Don't think in terms of biological time.  We are talking about
electronic hardware here.  At this very moment I am working with
latest-generation Xilinx FPGA's that can run logic at 500MHz without a
problem.  There are other companies that claim 1.5GHz performance and
I know that both Xilinx and Altera have new (faster, denser) chips
coming out next year.

GA evolution time, when implemented with performance in mind, is
extremely fast.  The issue is to have a "brain" with enough cells and
inputs to be able to evolve something useful.  In many ways, I see it
as a Moore's Law challenge.  When we get to FPGA's or microprocessors
with 10 times the density we have today it is likely that very
interesting things will start to happen in both robotics and Ai.

I don't think we disagree, you use different terminology to mean
approximately the same thing.

-Martin

Yeah, I think we are looking in conceptually very similar ideas at the high
level but the implementation and results I think are different - unless
there is work in GA that I don't know about (which is very possible).

With the GA approach, you must have a well defined fitness function.  There
is no well defined fitness function for general AI as far as I can tell -
other than the one evolution itself used - survival in the real world.
That is too vague and would take too long for any GA search to find
something closer to "intelligent".

The GA approach only works well if you can be very precise about what
behavior you want the hardware to produce and can easily build a fitness
function to test how close the hardware is to producing the correct
function.  It also typically requires that you create simulations of the
problems you are trying to solve so it can run though millions of
simulations very quickly.  If the goal is to produce hardware to make a
hex-pod walk on a flat surface, a GA search for the code can be simulated
with the simulator running far faster than real time - so in the simulator
you can get years of virtual evolution training happening in seconds of
real time.

But what happens when the problem is so complex you can't simulate it and
the robot has to learn in real time in real conditions?  For example, what
if you want to train your robot to hunt and catch rats in a barn?  Or to
train a bird-robot to fly?  Or to not only fly, but to catch birds as it's
flying?  Do you think you can use a GA approach to build robots like that?

General reinforcement learning however makes no attempt to specify the
behavior.  It only comments on the results.  If you can build hardware to
test for the desired result which generates the reward signal, the
reinforcement learning system figures out the correct behaviors on it's
own.

For strong reinforcement learning to work, it has to include a function to
estimate expected future rewards.  This function in effect becomes the
systems fitness function.  In other words, with reinforcement learning, the
system not only has to learn useful behaviors, it has to learn its own
fitness function as well which is then used to evaluate and select
behaviors.

The good thing about GA approaches, is that they are useful today.  We know
how to do useful things with them.  For the most part, reinforcement
learning algorithms are not useful today - or are useful in only very
limited domains and applications.  This fact seems to have drawn more
people to play with GA.  But the solution to AI is not GA (unless you warp
your definition of GA so much you end up with a true reinforcement learning
algorithm).

--
Curt Welch                                            http://CurtWelch.Com/
curt@kcwc.com                                        http://NewsReader.Com/

A robot is a device built to do some task formerly reserved for living
things.  Usually the living things in question are humans, but certainly not
always.
Literally any device, with or without onboard intelligence, that was
designed to do a task done only by living things before the device in
question was built can legitimately be called a robot.

Now...
if you're building a device to do something that has been done by a machine
before, you haven't built a robot.  You've just built a machine.  But if you
are taking some function out of the biological world and putting it in the
mechanical world, (for the first time, ever)... you are doing so with a
robot.

I'd suspect that's why Mythbusters like to call their stuff robots... they
arrogantly believe that they are the first to transfer whatever function
their contraption is supposed to mimic from the biological to the mechanical
world.

There.  That ought to settle any debate.