高級 1246 分類 收藏
開始影片後,點擊或框選字幕可以立即查詢單字
字庫載入中…
回報字幕錯誤
ALAN WINFIELD: Thank you very much indeed.
It's really great to be here.
And thank you so much for the invitation.
So yes, robot intelligence.
So I've titled the lecture "The Thinking Robot."
But of course, that immediately begs the question,
what on earth do we mean by thinking?
Well we could, of course, spend the whole
of the next hour debating what we mean by thinking.
But I should say that I'm particularly interested in
and will focus on embodied intelligence.
So in other words, the kind of intelligence that we have,
that animals including humans have, and that robots have.
So of course that slightly differentiates
what I'm talking about from AI.
But I regard robotics as a kind of subset of AI.
And of course one of the things that we discovered
in the last 60 odd years of artificial intelligence
is that the things that we thought were really difficult
actually are relatively easy.
Like playing chess, or go, for that matter.
Whereas the things that we originally thought
were really easy, like making a cup of tea, are really hard.
So it's kind of the opposite of what was expected.
So embodied intelligence in the real world
is really very difficult indeed.
And that's what I'm interested in.
So this is the outline of the talk.
I'm going to talk initially about intelligence
and offer some ideas, if you like,
for a way of thinking about intelligence
and breaking it down into categories
or types of intelligence.
And then I'm going to choose a particular one which
I've been really working on the last three or four years.
And it's what I call a generic architecture
for a functional imagination.
Or in short, robots with internal models.
So that's really what I want to focus on.
Because I really wanted to show you
some experimental work that we've
done the last couple of years in the lab.
I mean, I'm an electronics engineer.
I'm an experimentalist.
And so doing experiments is really important for me.
So the first thing that we ought to realize--
I'm sure we do realize-- is that intelligence is not one thing
that we all, animals, humans, and robots
have more or less of.
Absolutely not.
And you know, there are several ways of breaking intelligence
down into different kind of categories,
if you like, of intelligence, different types
of intelligence.
And here's one that I came up with in the last couple
of years.
It's certainly not the only way of thinking about intelligence.
But this really breaks intelligence into four,
if you like, types, four kinds of intelligence.
You could say kinds of minds, I guess.
The most fundamental is what we call
morphological intelligence.
And that's the intelligence that you get just
from having a physical body.
And there are some interesting questions
about how you design morphological intelligence.
You've probably all seen pictures of or movies of robots
that can walk, but in fact don't actually
have any computing, any computation whatsoever.
In other words, the behavior of walking
is an emergent property of the mechanics,
if you like, the springs and levers and so on in the robot.
So that's an example of morphological intelligence.
Individual intelligence is the kind of intelligence
that you get from learning individually.
Social intelligence, I think, is really
interesting and important.
And that's the one that I'm going
to focus on most in this talk.
Social intelligence is the intelligence
that you get from learning socially, from each other.
And of course, we are a social species.
And the other one which I've been
working on a lot in the last 20 odd years
is swarm intelligence.
So this is the kind of intelligence
that we see most particularly in social animals, insects.
The most interesting properties of swarm intelligence
tend to be emergent or self-organizing.
So in other words, the intelligence
is typically manifest as a collective behavior that
emerges from the, if you like, the micro interactions
between the individuals in that population.
So emergence and self-organization
are particularly interesting to me.
But I said this is absolutely not the only way
to think about intelligence.
And I'm going to show you another way
of thinking about intelligence which I particularly like.
And this is Dan Dennett's tower of generate and test.
So in Darwin's Dangerous Idea, and several other books,
I think, Dan Dennett suggests that a good way
of thinking about intelligence is to think about the fact
that all animals, including ourselves, need
to decide what actions to take.
So choosing the next action is really critically important.
I mean it's critically important for all of us, including
humans.
Even though the wrong action may not kill us,
as it were, for humans.
But for many animals, the wrong action
may well kill that animal.
And Dennett talks about what he calls
the tower of generate and test which I want to show you here.
It's a really cool breakdown, if you like, way
of thinking about intelligence.
So at the bottom of his tower are Darwinian creatures.
And the thing about Darwinian creatures
is that they have only one way of,
as it were, learning from, if you like,
generating and testing next possible actions.
And that is natural selection.
So Darwinian creatures in his schema cannot learn.
They can only try out an action.
If it kills them, well that's the end of that.
So by the laws of natural selection,
that particular action is unlikely to be
passed on to descendants.
Now, of course, all animals on the planet
are Darwinian creatures, including ourselves.
But a subset of what Dennett calls Skinnerian creatures.
So Skinnerian creatures are able to generate
a next possible candidate action, if you like,
a next possible action and try it out.
And here's the thing, if it doesn't kill them
but it's actually a bad action, then they'll learn from that.
Or even if it's a good action, a Skinnerian creature
will learn from trying out an action.
So really, Skinnerian creatures are a subset of Darwinians,
actually a small subset that are able to learn
by trial and error, individually learn by trial and error.
Now, the third layer, or story, if you
like, in Dennett's tower, he calls Popperian creatures,
after, obviously, the philosopher, Karl Popper.
And Popperian creatures have a big advantage
over Darwinians and Skinnerians in that they
have an internal model of themselves in the world.
And with an internal model, it means
that you can try out an action, a candidate
next possible action, if you like, by imagining it.
And it means that you don't have to actually have
to put yourself to the risk of trying it out
for real physically in the world,
and possibly it killing you, or at least harming you.
So Popperian creatures have this amazing invention,
which is internal modeling.
And of course, we are examples of Popperian creatures.
But there are plenty of other animals-- again,
it's not a huge proportion.
It's rather a small proportion, in fact, of all animals.
But certainly there are plenty of animals
that are capable in some form of modeling their world
and, as it were, imagining actions before trying them out.
And just to complete Dennett's tower,
he adds another layer that he calls Gregorian creatures.
Here's he's naming this layer after Richard Gregory,
the British psychologist.
And the thing that Gregorian creatures have
is that in addition to internal models,
they have mind tools like language and mathematics.
Especially language because it means that Gregorian creatures
can share their experiences.
In fact, a Gregorian creature could, for instance,
model in its brain, in its mind, the possible consequences
of doing a particular thing, and then actually pass
that knowledge to you.
So you don't even have to model it yourself.
So the Gregorian creatures really
have the kind of social intelligence
that we probably-- perhaps not uniquely,
but there are obviously only a handful
of species that are able to communicate, if you like,
traditions with each other.
So I think internal models are really, really interesting.
And as I say, I've been spending the last couple
of years thinking about robots with internal models.
And actually doing experiments with
robots with internal models.
So are robots with internal models self-aware?
Well probably not in the sense that-- the everyday
sense that we mean by self-aware, sentient.
But certainly internal models, I think,
can provide a minimal level of kind
of functional self-awareness.
And absolutely enough to allow us to ask what if questions.
So with internal models, we have potentially a really powerful
technique for robots.
Because it means that they can actually ask
themselves questions about what if I take this
or that next possible action.
So there's the action selection, if you like.
So really, I'm kind of following Dennett's model.
I'm really interested in building Popperian creatures.
Actually, I'm interested in building Gregorian creatures.
But that's another, if you like, another step in the story.
So really, here I'm focusing primarily
on Popperian creatures.
So robots with internal models.
And what I'm talking about in particular
is a robot with a simulation of itself
and it's currently perceived environment and of the actors
inside itself.
So it takes a bit of getting your head around.
The idea of a robot with a simulation of itself
inside itself.
But that's really what I'm talking about.
And the famous, the late John Holland, for instance,
rather perceptively wrote an internal model
that allows a system to look ahead
to the future consequences of actions
without committing itself to those actions.
I don't know whether John Holland
was aware of Dennett's tower.
Possibly not.
But really saying the same kind of thing as Dan Dennett.
Now before I come on to the work that I've been doing,
I want to show you some examples of-- a few examples,
there aren't many, in fact-- of robots with self-simulation.
The first one, as far as I'm aware,
was by Richard Vaughan and his team.
And he used a simulation inside a robot
to allow it to plan a safe route with incomplete knowledge.
So as far as I'm aware, this is the world's first example
of robots with self-simulation.
Perhaps an example that you might already be familiar with,
this is Josh Bongard and Hod Lipson's work.
Very notable, very interesting work.
Here, self-simulation, but for a different purpose.
So this is not self-simulation to choose,
as it were, gross actions in the world.
But instead, self-simulation to learn
how to control your own body.
So that the idea here is that if you have a complex body, then
a self-simulation is a really good way of figuring out
how to control yourself, including
how to repair yourself if parts of you
should break or fail or be damaged, for instance.
So that's a really interesting example
of what you can do with self-simulation.
And a similar idea, really, was tested
by my old friend, Owen Holland.
He built this kind of scary looking robot.
Initially it was called Chronos, but but then it
became known as ECCE-robot.
And this robot is deliberately designed to be hard to control.
In fact, Owen refers to it as anthropomimetic.
Which means anthropic from the inside out.
So most humanoid robots are only humanoid on the outside.
But here, we have a robot that has a skeletal structure,
it has tendons, it's very-- and you
can see from the little movie clip
there, if any part of the robot moves,
then the whole of the rest of the robot
tends to flex, rather like human bodies or animal bodies.
So Owen was particularly interested in a robot
that is difficult to control.
And the idea then of using an internal simulation of yourself
in order to be able to control yourself or learn
to control yourself.
And he was the first to come up with this phrase,
functional imagination.
Really interesting work, so do check that out.
And the final example I want to give
you is from my own lab, where-- this is swarm robotics
work-- where in fact we're doing evolutionary swarm robotics
here.
And we've put a simulation of each robot and the swarm
inside each robot.
And in fact, we're using those internal simulations
as part of a genetic algorithm.
So each robot, in fact, is evolving its own controller.
And in fact, it actually updates its own controller
about once a second.
So again, it's a bit an odd thing to get your head around.
So about once a second, each robot
becomes its own great, great, great, great grandchild.
In other words, its controller is a descendant.
But the problem with this is that the internal simulation
tends to be wrong.
And we have what we call the reality gap.
So the gap between the simulation and the real world.
And so we got round that-- my student Paul O'Dowd came up
with the idea that we could co-evolve
the simulators, as well as the controllers in the robots.
So we have a population of robots
inside each individual physical robot, as it were,
simulated robots.
But then you also have a swarm of 10 robots.
And therefore, we have a population of 10 simulators.
So we actually co-evolve here, the simulators
and the robot controllers.
So I want to now show you the newer
work I've been doing on robots with internal models.
And primarily-- I was telling [? Yan ?] earlier
that, you know, I'm kind of an old fashioned electronics
engineer.
Spent much of my career building safety systems,
safety critical systems.
So safety is something that's very important to me
and to robotics.
So here's a kind of generic internal modeling
architecture for safety.
So this is, in fact, Dennett's loop of generate and test.
So the idea is that we have an internal model, which
is a self-simulation, that is initialized to match
the current real world.
And then you try out, you run the simulator
for each of your next possible actions.
To put it very simply, imagine that you're a robot,
and you could either turn left, turn right, go straight ahead,
or stand still.
So you have four possible next actions.
And therefore, you'd loop through this internal model
for each of those next possible actions.
And then moderate the action selection mechanism
in your controller.
So this is not part of the controller.
It's a kind of moderator, if you like.
So you could imagine that the regular robot
controller, the thing in red, has a set of four
next possible actions.
But your internal model determines
that only two of them are safe.
So it would effectively, if you like,
moderate or govern the action selection mechanism
of the robot's controller, so that the robot
controller, in fact, will not choose the unsafe actions.
Interestingly, if you have a learning controller,
then that's fine because we can effectively extend or copy
the learned behaviors into the internal model.
That's fine.
So in principle-- we haven't done this.
But we're starting to do it now-- in principle,
we can extend this architecture to, as it were, to adaptive
or learning robots.
Here's a simple thought experiment.
Imagine a robot with several safety hazards facing it.
It has four next possible actions.
Well, your internal model can figure out
what the consequence of each of those actions might be.
So two of them-- so either turn right or stay still
are safe actions.
So that's a very simple thought experiment.
And here's a slightly more complicated thought experiment.
So imagine that the robot, there's
another actor in the environment.
It's a human.
The human is not looking where they're going.
Perhaps walking down the street peering at a smartphone.
That never happens, does it, of course.
And about to walk into a hole in the pavement.
Well, of course, if it were you noticing
that human about to walk into a hole in the pavement,
you would almost certainly intervene, of course.
And it's not just because you're a good person.
It's because you have the cognitive machinery
to predict the consequences of both your and their actions.
And you can figure out that if you
were to rush over towards them, you
might be able to prevent them from falling into the hole.
So here's the same kind of idea.
But with the robot.
Imagine it's not you, but a robot.
And imagine now that you are modeling
the consequences of yours and the human's
actions for each one of your next possible actions.
And you can see that now this time,
we've given a kind of numerical scale.
So 0 is perfectly safe, whereas 10 is seriously dangerous,
kind of danger of death, if you like.
And you can see that the safest outcome
is if the robot turns right.
In other words, the safest for the human.
I mean, clearly the safest for the robot
is either turn left or stay still.
But in both cases, the human would fall into the hole.
So you can see that we could actually
invent a rule which would represent
the best outcome for the human.
And this is what it looks like.
So if all robot actions, the human is equally safe,
then that means that we don't need to worry about the human,
so the internal model will output the safest actions
for the robot.
Else, then output the robot actions
for the least unsafe human outcomes.
Now remarkably-- and we didn't intend this,
this actually is an implementation
of Asimov's first law of robotics.
So a robot may not injure a human being,
or through inaction-- that's important,
the or through inaction-- allow a human being to come to harm.
So we kind of ended up building Asimovian robot,
simple Asimovian ethical robot.
So what does it look like?
Well, we've now extended to humanoid robots.
But we started with the e-puck robots,
these little-- they're about the size of a salt shaker,
I guess, about seven centimeters tall.
And this is the little arena in the lab.
And what we actually have inside the ethical robot is--
this is the internal architecture.
So so you can see that we have the robot controller, which
is, in fact, a mirror of the real robot controller,
a model of the robot, and a model of the world, which
includes others in the world.
So this is the simulator.
This is a more or less a regular robot simulator.
So you probably know that robot simulators
are quite commonplace.
We roboticists use them all the time to test robots in,
as it were, in the virtual world,
before then trying out the code for real.
But what we've done here is we've actually
put an off the shelf simulator inside the robot
and made it work in real time.
So the output of the simulator for each
of those next possible actions is evaluated and then goes
through a logic layer.
Which is essentially the rule, the if then else rule
that I showed you a couple of slides ago.
And that effectively determines or moderates
the action selection mechanism of the real robot.
So this is the simulation budget.
So we're actually using the open source simulator
Stage, a well-known simulator.
And in fact, we managed to get Stage to run about 600 times
real time.
Which means that we're actually cycling
through our internal model twice a second.
And for each one of those cycles,
we're actually modeling not four but 30 next possible actions.
And we're modeling about 10 seconds into the future.
So every half a second, our robot with an internal model
is looking ahead 10 seconds for about 30 next possible actions,
30 of its own next possible actions.
But of course, it's also modeling the consequences
of each of the other actors, dynamic actors
in its environment.
So this is quite nice to actually do this in real time.
And let me show you some of the results that we got from that.
So ignore the kind of football pitch.
So what we have here is the ethical robot, which we
call the A-robot, after Asimov.
And we have a hole in the ground.
It's not a real hole, it's a virtual hole in the ground.
We don't need to be digging holes into the lab floor.
And we're using another e-perk as a proxy human
we call this the H-robot.
So let me show you what happened.
Well we ran it, first of all, with no H-robot at all,
as a kind of baseline.
And you can see on the left, in 26 runs,
those are the traces of the A-robot.
So you can see the A-robot, in fact,
is maintaining its own safety.
Its avoiding, its skirting around the edge
almost optimally skirting the edge of the hole in the ground.
But then when we introduce the H-robot,
you get this wonderful behavior here.
Where as soon as the A-robot notices that the H-robot is
heading towards the hole, which is about here, then
it deflects, it diverts from its original course.
And in fact, more or less collides.
They don't actually physically collide
because they have low level collision avoidance.
So they don't actually collide.
But nevertheless, the A-robot effectively
heads off the H-robot, but then bounces off safely,
goes off in another direction.
And the A-robot then resumes its course to its target position.
Which is really nice.
And interestingly, even though our simulator is rather low
fidelity, it doesn't matter.
Surprisingly, it doesn't matter.
Because the closer the A-robot to the H-robot
gets, then the better its predictions about colliding.
So this is why, even with a rather low fidelity simulator,
we can collide with really good precision with the H-robot.
So let me show you the movies of this trial with a single proxy
human.
And I think the movie starts in-- so this is real time.
And you can see the A-robot nicely heading off
the H-robot which then disappears off
towards the left.
I think then we've speeded it four times.
And this is a whole load of runs.
So you can see that it really does work.
And also notice that every experiment is a bit different.
And of course, that's what typically happens when
you have real physical robots.
Simply because of the noise in the system,
the fact that these are real robots with imperfect motors
and sensors and what have you.
So we wrote the paper and were about to submit
the paper, when we kind of thought, well,
this is a bit boring, isn't it?
We built this robot and it works.
So we had the idea to put a second human in the-- oh sorry.
I've forgotten one slide.
So before I get to that, I just wanted
to show you a little animation of-- these little filaments
here are the traces of the A-robot and its prediction
of what might happen.
So at the point where this turns red,
the A-robot then starts to intersect.
And each one of those traces is its prediction
of the consequences of both itself and the H-robot.
This is really nice because you can kind of look into the mind,
to put it that way, of the robot,
and actually see what it's doing.
Which is very nice, very cool.
But I was about to say we tried the same experiment, in fact,
identical code, with two H-robots.
And this is the robot's dilemma.
This may be the first time that a real physical robot
has faced an ethical dilemma.
So you can see the two H-robots are more or less
equidistant from the hole.
And there is the A-robot which, in fact,
fails to save either of them.
So what's going on there?
We know that it can save one of them every time.
But in fact, it's just failed to save either.
And oh, yeah, it does actually save one of them.
And has a look at the other one, but it's too late.
So this is really very interesting.
And not at all what we expected.
In fact, let me show you the statistics.
So in 33 runs, the ethical robot failed
to save either of the H-robots just under half the time.
So about 14 times, it failed to save either.
It saved one of them just over 15, perhaps 16 times.
And amazingly, saved both of them twice,
which is quite surprising.
It really should perform better than that.
And in fact, when we started to really look at this,
we discovered that the-- so here's
a particularly good example of dithering.
So we realized that we made a sort
of pathologically indecisive ethical robot.
So I'm going to save this one-- oh no, no, that
one-- oh no, no, this one-- that one.
And of course, by the time our ethical robot
has changed its mind three or four times,
well, it's too late.
So this is the problem.
The problem, fundamentally, is that our ethical robot
doesn't make a decision and stick to it.
In fact, it's a consequence of the fact
that we are running our consequence engine,
as I mentioned, twice a second.
So every half a second, our ethical robot
has the opportunity to change its mind.
That's clearly a bad strategy.
But nevertheless, it was an interesting kind
of unexpected consequence of the experiment.
We've now transferred the work to these humanoid robots.
And we get the same thing.
So here, there are two red robots
both heading toward danger.
The blue one, the ethical robot, changes its mind,
and goes and saves the one on the left,
even though it could have saved the one on the right.
So another example of our dithering ethical robot.
And as I've just hinted at, the reason
that there our ethical robot is so indecisive
is because it's essentially a memory-less architecture.
So you could say that the robot has a-- again, borrowing
Owen Holland's description, it has a functional imagination.
But it has no autobiographical memory.
So it doesn't remember the decision
it made half a second ago.
Which is clearly not a good strategy.
Really, an ethical robot, just like you
if you are acting in a similar situation,
it's probably a good idea for you
to stick to the first decision that you made.
But probably not forever.
So you know, I think the decisions probably
need to be sticky somehow.
So decisions like this may need a half life.
You know, sticky but not but not absolutely rigid.
So actually, at this point, we decided
that we're not going to worry too much about this problem.
Because in a sense, this is more of a problem for ethicists
than engineers, perhaps.
I don't know.
But maybe we could talk about that.
Before finishing, I want to show you
another experiment that we did with the same architecture,
exactly the same architecture.
And this is what we call the corridor experiment.
So here we have a robot with this internal model.
And it has to get from the left hand to the right hand
of a crowded corridor without bumping
into any of the other robots that are in the same corridor.
So imagine you're walking down a corridor in an airport
and everybody else is coming in the opposite direction.
And you want to try and get to the other end of the corridor
without crashing into any of them.
But in fact, you have a rather large body space.
You don't want to get even close to any of them.
So you want to maintain your private body space.
And what the blue robot here is doing
is, in fact, modeling the consequences of its actions
and the other ones within this radius of attention.
So this blue circle is a radius of attention.
So here, we're looking at a simple attention mechanism.
Which is only worry about the other dynamic actors
within your radius of attention.
In fact, we don't even worry about ones that are behind us.
It's only the ones that are more or less in front of us.
And you can see that the robot does eventually make it
to the end of the corridor.
But with lots of kind of stops and back tracks
in order to prevent it from-- because it's really
frightened of any kind of contact with the other robots.
And here, we're not showing all of the sort
of filaments of prediction.
Only the ones that are chosen.
And here are some results which interestingly show us--
so perhaps the best one to look at is this danger ratio.
And dumb simply means robots with no internal model at all.
And intelligent means robots with internal models.
So here, the danger ratio is the number of times
that you actually come close to another robot.
And of course it's very high.
This is simulated in real robots.
Very good correlation between the real and simulated.
And with the intelligent robot, the robot
with the internal model, we get a really very much safer
performance.
Clearly there is some cost in the sense
that, for instance, the intelligent robot
runs with internal models tend to cover more ground.
But surprisingly, not that much further distance.
It's less than you'd expect.
And clearly, there's a computational cost.
Because the computational cost of simulating clearly
is zero for the dumb robots, whereas it's quite high for
the intelligent robot, the robot with internal models.
But again, computation is relatively free these days.
So actually, we're trading safety
for computation, which I think is a good trade off.
So really, I want to conclude there.
I've not, of course, talked about all aspects
of robot intelligence.
That would be a three hour seminar.
And even then, I wouldn't be able to cover it all.
But what I hope I've shown you in the last few minutes
is that with internal models, we have
a very powerful generic architecture which we could
call a functional imagination.
And this is where I'm being a little bit speculative.
Perhaps this moves us in the direction
of artificial theory of mind, perhaps even self-awareness.
I'm not going to use the word machine consciousness.
Well, I just have.
But that's a very much more difficult goal, I think.
And I think there is practical value,
I think there's real practical value in robotics of robots
with self and other simulation.
Because as I hope I've demonstrated,
at least in a kind of prototype sense, proof of concept,
that such simulation moves us towards safer and possibly
ethical systems in unpredictable environments
with other dynamic actors.
So thank you very much indeed for listening.
I'd obviously be delighted to take any questions.
Thank you.
[APPLAUSE]
HOST: Thank you very much for this very fascinating view
on robotics today.
We have time for questions.
Please wait until you've got a microphone so we have
the answer also on the video.
AUDIENCE: The game playing computers-- or perhaps more
accurately would be saying game playing algorithms,
predated the examples you listed as computers
with internal models.
Still, you didn't mention those.
Is there a particular reason why you didn't?
ALAN WINFIELD: I guess I should have mentioned them.
You're quite right.
I mean, the-- what I'm thinking of here
is particularly robots with explicit simulations
of themselves and the world.
So I was limiting my examples to simulations of themselves
in the world.
I mean, you're quite right that of course
game-playing algorithms need to have a simulation of the game.
And quite likely, of the-- in fact,
certainly, of the possible moves of the opponent,
as well as the game-playing AI's own moves.
So you're quite right.
I mean, it's a different kind of simulation.
But I should include that.
You're right.
AUDIENCE: Hi there.
In your simulation, you had the H-robot with one goal,
and the A-robot with a different goal.
And they interacted with each other halfway
through the goals.
What happens when they have the same goal?
ALAN WINFIELD: The same goal.
AUDIENCE: Reaching the same spot, for example.
ALAN WINFIELD: I don't know is the short answer.
It depends on whether that spot is a safe spot or not.
I mean, if it's a safe spot, then they'll both go toward it.
They'll both reach it, but without crashing
into each other.
Because the A-robot will make sure
that it avoids the H-robot.
In fact, that's more or less what's
happening in the corridor experiment.
That's right.
Yeah.
But it's a good question, we should try that.
AUDIENCE: The simulation that you
did for the corridor experiment, the actual real world
experiment, the simulation track the other robots
movements as well?
Meaning what information did the simulation
have that it began with, versus what did it perceive?
Because, I mean, the other robots are moving.
And in the real world, they might not
move as you predict them to be.
How did the blue robot actually know for each step
where the robots were?
ALAN WINFIELD: Sure.
That's a very good question.
In fact we cheated, in the sense that we
used-- for the real robot experiments,
we used a tracking system.
Which means that essentially the robot with an internal model
has access to the position.
It's like a GPS, internal GPS system.
But in a way, that's really just a kind
of-- it's kind of cheating, but even a robot with a vision
system would be able to track all the robots
in its field of vision.
And as for the second part of your question,
our kind of model of what the other robot that would do
is very simple.
Which is it's kind of a ballistic model.
Which is if a robot is moving at a particular speed
in a particular direction, then we
assume it will continue to do so until it encounters
an obstacle.
So very simple kind of ballistic model.
Which even for humans is useful for very simple behaviors,
like moving in a crowded space.
Oh hi.
AUDIENCE: In the same experiment--
it's a continuation of the previous question.
So in between some of the red robots, how
changed their direction randomly-- I guess so.
Does the internal model of the blue robot consider that?
ALAN WINFIELD: Not explicitly.
But it does in the sense that because it's
pre- or re-initializing it's internal model every half
a second, then if the positions and directions of the actors
in its environment are changed, then they
will reflect the new positions.
So--
AUDIENCE: Not exactly the positions.
But as you said, you have considered the ballistic motion
of the objects.
So if there is any randomness in the environment-- so
does the internal model of the blue robot
consider the randomness, and change
the view of the red robots?
It's like it views the red robot as a ballistic motion.
So does it change its view of the red robot
that red robots more in the ballistic motion?
ALAN WINFIELD: Well, it's a very good question.
I think the answer is no.
I think we're probably assuming a more or less deterministic
model of the world.
Deterministic, yes, I think pretty much deterministic.
But we're relying on the fact that we
are updating and rerunning the model,
reinitializing and rerunning the model every half a second,
to, if you like, track the stochasticity which is
inevitable in the real world.
We probably do need to introduce some stochasticity
into the internal model, yes.
But not yet.
AUDIENCE: Thank you.
ALAN WINFIELD: But very good question.
AUDIENCE: Hello.
With real life applications with this technology,
like driverless cars, for example,
I think it becomes a lot more important how you program
the robots in terms of ethics.
So I mean, there could be dilemma like if the robot has
a choice between saving a school bus full of kids
versus one driver, that logic needs to be programmed, right?
And you made a distinction between being an engineer
yourself and then had been an ethicist earlier.
So to what extent is the engineer
responsible in that case?
And also does a project like this in real life
always require the ethicist?
How do you see this field in real life applications
evolving?
ALAN WINFIELD: Sure.
That's a really great question.
I mean, you're right that driverless cars will-- well,
it's debatable whether they will have to make such decisions.
But many people think they will have to make such decisions.
Which are kind of the driverless car equivalent of the trolley
problem, which is a well-known kind of ethical dilemma thought
experiment.
Now my view is that the rules will
need to be decided not by the engineers, but if you like,
by the whole of society.
So ultimately, the rules that decide
how the driverless car should behave
under these difficult circumstances, impossible,
in fact, circumstances-- and even
if we should, in fact, program those rules into the car.
So some people argue that the driverless cars should not
attempt to, as it were, make a rule driven decision.
But just leave it to chance.
And again, I think that's an open question.
But this is really why I think this dialogue and debate
and conversations with regulators, lawyers, ethicists,
and the general public, users of driverless cars,
I think is why we need to have this debate.
Because whatever those rules are, and even
whether we have them or not, is something
that should be decided, as it were, collectively.
I mean, someone asked me last week,
should you be able to alter the ethics of your own driverless
car?
My answer is absolutely not.
I mean, that should be illegal.
So I think that if driverless cars were
to have a set of rules, and especially
if those rules had numbers associated with them.
I mean, let's think of a less emotive example.
Imagine a driverless car and an animal runs into the road.
Well, , the driverless car can either ignore the animal
and definitely kill the animal, or it could try and brake,
possibly causing harm to the driver or the passengers.
But effectively reducing the probability
of killing the animal.
So there's an example where you have some numbers
to tweak if you like, parameters.
So if these rules are built into driverless cars,
they'll be parameterized.
And I think it should be absolutely
illegal to hack those parameters, to change them.
In the same way that it's probably illegal right now
to hack an aircraft autopilot.
I suspect that probably is illegal.
If it isn't, it should be.
So I think that you don't need to go
far down this line of argument before realizing
that the regulation and legislation has
to come into play.
In fact, I saw a piece in just this morning in Wired
that, I think, in the US, regulation for driverless cars
is now on the table.
Which is absolutely right.
I mean, we need to have regulatory framework,
or what I call governance frameworks for driverless cars.
And in fact, lots of other autonomous systems.
Not just driverless cars.
But great question, thank you.
AUDIENCE: In the experiment with the corridor,
you always assume-- even in the other experiments-- you always
assume that the main actor is the most intelligent
and the others are not.
Like they're dumb, or like they're
ballistic models or linear models.
Have you tried doing a similar experiment
in which still each actor is intelligent but assumes
that the others are not, but actually everyone
is intelligent?
So like everyone is a blue dot in the experiment
with the model that you have.
And also, have you considered changing the model so that he
assumes that the others have the same model
that that particular actor has, as well.
[INTERPOSING VOICES]
ALAN WINFIELD: No, we're doing it right now.
So we're doing that experiment right now.
And if you ask me back in a year,
perhaps I can tell you what happ-- I mean, it's really mad.
But it does take us down this direction
of artificial theory of mind.
So if you have several robots or actors,
each of which is modeling the behavior of the other,
then you get-- I mean, some of the--
I don't even have a movie to show you.
But in simulation we've tried this
where we have two robots which are kind of like-- imagine,
this happens to all of us, you're walking down
the pavement and you do the sort of sidestep dance
with someone who's coming towards you.
And so the research question that we're
asking ourselves is do we get the same thing.
And it seems to be that we do.
So if the robots are symmetrical, in other words,
they're each modeling the other, then
we can get these kind of little interesting dances, where each
is trying to get out of the way of the other, but in fact,
choosing in a sense the opposite.
So one chooses to step right, the other chooses to step left,
and they still can't go past each other.
But it's hugely interesting.
Yes, hugely interesting.
AUDIENCE: Hi.
I think it's really interesting how
you point out the importance of simulations
and internal models.
But I feel that one thing that is slightly left out
there is a huge gap from going from simulation to real world
robots, for example.
And I assume that in these simulations
you kind of assume that the sensors are 100% reliable.
And that's obviously not the case in reality.
And especially if we're talking about autonomous cars or robots
and safety.
How do you calculate the uncertainty
that comes with the sensors in the equation?
ALAN WINFIELD: Sure.
I mean, this is a deeply interesting question.
And the short answer is I don't know.
I mean, this is all future work.
I mean, my instinct is that a robot
with a simulation, internal simulation,
even if that simulation in a sense is idealized,
is still probably going to be safer than a robot that has
no internal simulation at all.
And you know, I think we humans have multiple simulations
running all the time.
So I think we have sort of quick and dirty, kind of low fidelity
simulations when we need to move fast.
But clearly when you need to plan something, plan
some complicated action, like where
you are going to go on holiday next year,
you clearly don't use the same internal model, same simulation
as for when you try and stop someone
from running into the road.
So I think that future intelligent robots
will need also to have multiple simulators.
And also strategies for choosing which
fidelity simulator to use at a particular time.
And if a particular situation requires
that you need high fidelity, then for instance,
one of the things that you could do,
which actually I think humans probably do,
is that you simply move more slowly to give your self time.
Or even stop to give yourself time
to figure out what's going on.
And in a sense, plan your strategy.
So I think even with the computational power we have,
there will still be a limited simulation budget.
And I suspect that that simulation budget
will mean that in real time, when you're
doing this in real time, you probably can't run your highest
fidelity simulator.
And taking into account all of those probabilistic, noisy
sensors and actuators and so on, you probably
can't run that simulator all the time.
So I think we're going to have to have
a nuanced approach where we have perhaps multiple simulators
with multiple fidelities.
Or maybe a sort of tuning, where you can tune
the fidelity of your simulator.
So this is kind of a new area of research.
I don't know anybody who's thinking about this yet, apart
from ourselves.
So great.
AUDIENCE: [INAUDIBLE].
ALAN WINFIELD: It is pretty hard, yes.
Please.
AUDIENCE: [INAUDIBLE].
ALAN WINFIELD: Do you want the microphone?
Sorry.
AUDIENCE: Have you considered this particular situation
where there are two Asimov robots--
and that would be an extension of the question that he asked.
So for example, if there are two guys walking
on a pavement and there could be a possibility
of mutual cooperation.
As in one might communicate whether that I might step out
of this place and you might go.
And then I'll go after that.
So if there are two Asimov robots,
will there be a possibility, and have
you considered this fact that both will communicate
with each other, and they will eventually come to a conclusion
that I will probably walk, and the other will get out
of the way.
And the second part of this question
would be what if one of the robots
actually does not agree to cooperate?
I mean, since they both would have different simulators.
They could have different simulators.
And one might actually try to communicate that you step out
of the way so that I might go.
And the other one doesn't agree with that.
I mean, what would the [INAUDIBLE].
ALAN WINFIELD: Yeah, it's a good question.
In fact, we've actually gotten a new paper which
we're just writing right now.
And the sort of working title is "The Dark Side
of Ethical Robots."
And one of the things that we discovered-- it's actually not
surprising-- is that you only need to change one line of code
for a co-operative robot to become a competitive robot,
or even an aggressive robot.
So it's fairly obvious, when you start to think about it,
if your ethical rules are very simply written,
and are a kind of layer, if you like,
on top of the rest of the architecture,
then it's not that difficult to change those rules.
And yes, we've done some experiments.
And again, I don't have any videos to show you.
But they're pretty interesting, showing
how easy it is to make a competitive robot, or even
an aggressive robot using this approach.
In fact, on the BBC six months ago or so,
I was asked surely if you can make an ethical robot,
doesn't that mean you can make an unethical robot?
And the answer, I'm afraid, is yes.
It does mean that.
But this really goes back to your question earlier,
which is that it should be-- we should make
sure it's illegal to convert, to turn,
if you like, to recode an ethical robot
as an unethical robot.
Or even it should be illegal to make unethical robots.
Something like that.
But it's a great question.
And short answer, yes.
And yes, we have some interesting new results,
new paper on, as it were, unethical robots.
Yeah.
HOST: All right, we are running out of time now.
Thanks everyone for coming today.
Thanks, Professor Alan Winfield.
ALAN WINFIELD: Thank you.
[APPLAUSE]
    您必須登入才有此功能
提示:點選文章或是影片下面的字幕單字,可以直接快速翻譯喔!

載入中…

【Google Talk】Alan Winfield :思想機器人

1246 分類 收藏
richardwang 發佈於 2016 年 5 月 14 日

影片簡介

發佈日期:2016年5月10日
Professor Alan Winfield from The University of the West of England joined us to share his work on Robots and Ethics, in a talk entitled "The Thinking Robot".
About the Book:
Robotics is a key technology in the modern world. Robots are a well-established part of manufacturing and warehouse automation, assembling cars or washing machines, and, for example, moving goods to and from storage racks for Internet mail order. More recently robots have taken their first steps into homes and hospitals, and seen spectacular success in planetary exploration. Yet, despite these successes, robots have failed to live up to the predictions of the 1950s and 60s, when it was widely thought - by scientists and engineers as well as the public - that by turn of the 21st century we would have intelligent robots as butlers, companions, or co-workers.

影片學習單字重點

loading

影片討論

載入中…
  1. 1. 單字查詢佳句收藏

    選取單字或佳句,可即時查詢字典及收藏!

  2. 2. 單句重複播放

    可重複聽取一句單句,加強聽力!

  3. 3. 使用快速鍵

    使用影片快速鍵,讓學習更有效率!

  4. 4. 關閉語言字幕

    進階版練習可關閉字幕純聽英文哦!

  5. 5. 內嵌播放器

    可以將英文字幕學習播放器內嵌到部落格等地方喔

  6. 6. 展開播放器

    可隱藏右方全文及字典欄位,觀看影片更舒適!

  1. 英文聽力測驗

    挑戰字幕英文聽力測驗!

  1. 點擊展開筆記本讓你看的更舒服

  1. UrbanDictionary 俚語字典整合查詢。一般字典查詢不到你滿意的解譯,不妨使用「俚語字典」,或許會讓你有滿意的答案喔