September 4, 2015 • 37 mins
The team talks about an experiment in the UK in which a robotic arm built, tested and redesigned other robots. Will machines take over the manufacturing world?
Learn more about your ad-choices at https://www.iheartpodcastnetwork.com
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Brought to you by Toyota. Let's go places. Welcome to
Forward Thinking. Hey everyone, and welcome to Forward Thinking, the
podcast that looks at the future. It says, mama's don't
let your babies grow up to be robots. I'm Jovin
(00:20):
Strickland and I'm Joe McCormick. And have you two ever
noticed how much better robots are than organic life forms
all the time. That's why I only work with various
things that allow me to interact with machines as opposed
to actual human beings. Like if it's a prerecorded type
message and I navigate through a menu using buttons, and
(00:43):
I don't have to talk to a person, that's awesome.
They're certainly less offensively smelly. Well that's just because we
haven't mastered the artificial intelligence sector of creating really lifelike
bo Yeah, well we have, we have scientists working on that. Yeah, yeah, no,
I was kidding. But there are, of course situations where
(01:05):
it's much better to be a robot than to be
an organic life form, and we've talked about plenty of
them on the podcast before. Maybe in space exploration where
you're at risk, or in or in search and rescue
scenarios where again a human would be at risk or
you know, any number of hundreds of others. Really, Yeah,
just just performing the same physical job. Anything that would
(01:27):
make a human board or or repeated stress injury, that
kind of stuff. I mean things that where you do
it once, it doesn't hurt, but you do it, you know,
once every minute for eight hours a day, and it's
and eventually, yeah, the button pushing muscles start to wear
down in the bones and etcetera. Yeah, robots don't get
don't get physically injured, they don't get emotionally injured. It's
(01:51):
real hard to traumatize a robot. Yeah, I know. Now,
we we've talked about the advantages. Of course, the life
forms have over robots and limited in each individual case.
So a person like a worker is a lot easier
to train for a new task than a robot is. Well,
depending on if if the robot is mechanically suited to
(02:13):
the new task, you can reprogram it s well. But
but a worker can learn many new tasks. Yeah, hypothetically,
most robots are pretty good at doing one thing right, right, Um, Yeah, robots,
as it turns out, they pretty much can do what
you built them to do, and they're pretty bad at
doing anything outside of that. In general, right, Robots don't
(02:34):
really evolve. Yeah, so yeah, so humans are more adaptable
on the individual scale, but they're also more adaptable as
a type of thing. Sure, over time, if the air
pressures from the environment on a living organism, it will evolve,
it will adapt, or it might go extinct. But assuming
it doesn't go extinct, it will adapt and evolve to
(02:54):
to fit the environment that it needs to make its
living in. Robots don't do that, no, they remain. Yeah,
so that's what we're going to talk about today exactly.
So let's talk about like, like general robots right now.
They they are too pretty distinct pathways for robot design
in general, and of course there are a lot of
(03:15):
branches within these, but we can go one path where
we design a robot for a specific task and that's
all it is supposed to do. So rumba is a
great example. What a rumba is supposed to do is
clean the floor. It's not supposed to do anything else
other than it doesn't clean the cat box. In my opinion,
this is this is probably like of robots and all
(03:37):
of the good ones. Yeah, the most efficient and the
most effective robots, I would argue, are of this type
because we only have to concern ourselves with the design
elements that will enable the robot to complete its task,
and we eliminate everything else. Right, you don't need to
include anything that doesn't involve cleaning the floor, or navigation
(03:58):
or returning to like a docking station to recharge. When
you're designing a robot like the Rumba, you just need
those basic elements in it. Anything else is superfluous. Uh.
Doesn't stop us from doing things like having crazy Rumba
fights where we put a balloon and something sharp on
on a couple of them and have them joused each other.
(04:19):
But it doesn't mean we don't put our cats on
top of them and just watch what happen. Maybe put
our cat in a shark costume first and then put
the cat on top of a roomba. Yeah, that sort
of thing. But in general, they just do what they
were supposed to do, and if you wanted them to
do anything else, you're kind of out of luck. Or
we tried to build general purpose robots that are capable
of doing lots of different tasks hopefully yeah, to varying
(04:44):
degrees of mediocrity because it's hard. It's not because you know, No, yeah,
it's not because the people working on these aren't smart. Yes,
because the job is a thousand times more difficult, right,
because you have to anticipate lots of different things, a
lot of different changing conditions. This would be the example
of the darker robots we talked about previously, where UH
(05:06):
challenge robots exactly. Yes, the DARPA Challenge robots, the ones
that had to replicate a uh kind of a first
responder situation, search and rescue kind of. Yeah. They to
do like drive a car to a building, open a door,
go inside, twist to lever, and cut a hole in
the wall. They had to plug in a cable in
one case. Yeah, they had to walk across rubble. They
(05:30):
had to go up some stairs and then fall over.
Uh didn't have to fall over, but a lot of
them sure did. At any rate, we saw how difficult
it is to design a robot that can do these things.
And keep in mind, these were all tasks that the
various teams knew about beforehand they were designing these. Yeah, yeah,
I mean it was this was all stuff that they
(05:51):
knew they were going to have, the robot was going
to have to be able to do. Uh. So, building
a robot that could end up anticipating all sorts of stuff,
whether it's a single robot or a you know, a
robot that can then design the next generation of robots.
That's something that we really haven't perfected yet, but there
are some really good reasons why we would want to. Clearly,
(06:14):
being able to have a robot that could either adapt
itself or adapt the next generation of robots would lead
to much more efficient machines over time, and so this
is something that we would really like to see in technology.
It's not it's not just in robotics. We're also seeing
it in computers. So there's they call it, you know,
(06:37):
evolutionary computation or evolutionary computers or evolutionary robotics because that's
essentially what we what we mean. We're talking about a
machine that looks at the ability of another, like a
next generation of machines to do a particular task, evaluates them,
and then makes decisions on how to alter that generation
(06:57):
to produce an even more effective generation after it. Uh
the goal always being to come up with the most
ideal design for whatever purpose you have in mind. And
this isn't a super super new idea, no. Now, there
have been people who have been working on this for
a while. Back in two thousand eleven, researchers with NASA
(07:17):
published a paper on evolutionary computation that was used to
design new efficient antenna's way back in two thousand six.
So the paper was published in two thousand and eleven,
but the actual project was yeah in two thousand and six.
And what they what they were pointing out was that
designing an antenna is really challenging. It requires that the
(07:38):
builder to have a very detailed knowledge of how antenna's work,
which already limits your pool of various builders, and even
then it's just painstaking to create an actual efficient, working antenna.
So what they did was they created evolutionary algorithms for
a computer to design an antenna for spacecraft auto matically
(08:00):
that the computer system went through various designs of antenna
and essentially simulated tests of them to see which one
would be the most effective. They then took the designs
that were predicted to be the highest performing antennas by
this computer program, and they took it to the same
manufacturer who was already building the antenna intended for the spacecraft.
(08:24):
Then they tested all three of the antenna, the two
that were the two best performer ones in the computer program,
and the one that people had designed for the spacecraft,
and found that the two that the computer had designed
were more effective. Yeah, so it was one of those
examples of this approach actually working better than what we
humans could do. Now, maybe you don't know the answer
(08:45):
to this, but if you do, I'd be interested. In
this case, we're the what what we might call the
mutations introduced into the models in in the algorithm that
that tested out antenna's where the mutations direct did where
they programmed in? Or were they random mutations? Like? Was
it really more like evolution where a random things thing
(09:08):
happens and then the system tests is this any better? Well,
as we've talked about in previous episodes, random with machines
is really hard to do. It was more like they
were the computer was given an enormous number of variables
and started to test them in various configurations without going
through every possible one. Because the goal of evolutionary computer
(09:32):
computation or evolutionary robotics is to make sure that you
come to the most ideal form of whatever it is
you're going for without having to test every possible variation. Because,
of course, thus things get more complex, those variations increase
in number until it would take you till the end
of time to test all the different ones to come
(09:53):
up version one one seven eight nine seven, seven seven
four three is the best. You would take forever to
go through all of those. So what these are designed
to do is to test ones that are a best
guests already of being effective, measuring those against other best guesses,
then combining ones. Like if you were to find a
(10:15):
generation that works particularly well and another generation that works
particularly well, you might say, well, what happens if I
combine the best design elements of both of these into
a new design. Does that increase the effect of the
efficiency and effectiveness or does it decrease it? Because it doesn't,
you know, adding to awesome things together does not guarantee
(10:36):
you to get a third, even more awesome thing. Oh yeah, yeah.
I was wondering because I'd read about the use of
evolutionary algorithms before in the design of new planetary rovers
like unmanned planetary rovers, and testing different variations on the
models in a in a computer simulation that would naturally
select the highest performing ones. Yeah. Yeah, it's very similar
(10:59):
to that. And uh, the thing about the difference, you know,
we're going to talk about a robot that works along
these principles. The biggest difference between a robot and a computer,
obviously is that a robot is working with actual physical matter,
not just simulations. And so there's some practical limitations that
you encounter in that case. Right, you have to work
(11:21):
with real physical matter that has weight, it has mass,
it occupies space, there's a limit to how heavy it
can be. Uh, you know, you can't just magically increase
the size of whatever it is in a simulation and
then just see how it works. You have to physically
put this thing together. And so it's a really interesting approach.
(11:44):
And we have seen a couple of examples of people
working with true physical matter, but uh, it's mostly been
in things like working with a program that builds stuff
out of lego. So there is an actual example of this.
There were some evolutionary computational experts that design systems that
would allow a machine to build other objects out of legos,
(12:04):
but you had to build limitations into the computer program
so that the machine would follow the rules of physics.
In other words, you couldn't have to lego pieces occupy
the same physical space at the same time, right, Sure,
And some of these computations aren't going to I mean,
it's hard as a programmer to build all of those
(12:24):
limitations into something that has to work in reality, because
the computer isn't going to understand I mean basic stuff,
and so so it might when you take it out
of the lab, out of the out of the computer
simulation stage, it might operate very differently than you were
expecting it. Then the computer was expecting it to. An
analogy I would make and has nothing to do with
(12:46):
computers or robotics, but an analogy I would make is
if you were running a role playing game. You're you're
the game master of a role play game, and you
think you have anticipated everything your players are going to do,
and you have made up a masterpiece of a module,
and your players are going to have an amazing time,
and three minutes in the players decided to do something
(13:07):
you could not have possibly anticipated. It makes perfect sense
within the context of the game. And then you and you, yeah,
you gotta throw away all the stuff you worked on
and say, well, well, I guess like it's sort of
like if you imagine a horror movie where the characters
all walk up to the spooky house, take one look
and say no, and then walk away like, well, there
(13:28):
goes the horror movie. Same sort of thing. Well, we
wanted to specifically talk about some researchers from the University
of Cambridge who were working with a robotic system that
used evolutionary robotics. So imagine, essentially, you've got a robot
that builds other robots and tests them to see which
(13:49):
designs work the best, and then either UH either has
a design go forward or eliminates designs and starts tweaking
things to try and find the best physical design of
a robot to complete a specific task. Now that sounds
super cool. I need to explain some stuff first so
(14:09):
we can manage our expectations. Yeah, yeah, okay, so so
so performance of these of these baby bots, as we
might call them UH in this case was how fast
and far it could move, like kind of scuttle across
the surface exactly. So it's not like it was performing
open heart surgery. No, No, they decided to hold that
(14:30):
for the next experiment. Perhaps, Now these robots all we
were supposed to do. We're create a locomotion power that
could move it from one point to another point, and
uh they would. The experiment measured how long it took
these little devices to move across this particular expanse they
(14:52):
changed up. That didn't make the cut. They didn't they
didn't continue on. Uh that they were taken apart. They
were taking up bodily by the mamabot. Actually, to be fair,
I think all of them were taken apart, but yes,
uh uh yeah they were. They'd be scuttled, they'd be
scuttled and then they'd be scuttled and and and harvested
(15:14):
for their organs essentially. So that was the that was
the criteria that the robot used. Um and here's how
it gets here. Let's break this down because that's that's
like an overview of what happened, but it's really interesting
when you get to the specifics. So what they wanted
to do is they wanted to design a system that
could design, test, and change robotic designs relatively rapidly. And
(15:40):
it needed to be a a system that could work
within the lab, but wasn't designed to be a practical
system that could go through every single possible arrangement of
the various pieces. For yeah, um, the pieces are essentially
two different types of cubes that will get to in
a second. So they wanted their approach to contain this
(16:00):
is a quote from their paper paper, a limited number
of evolutionary iterations, and that was, you know, the the
idea of let's let evolution lead us to what should
be the best version of these robotic designs, as opposed
to just doing trial and error, measuring everything and then
going with the top performer. Now, they did say in
(16:22):
their paper that while this is really interesting and could
in fact be a breakthrough in science and technology, you
have to admit that there are limitations to this technology
cannot necessarily be applied to mass production. Mass production relies
heavily on automation. Automation is different from mass customization. So
(16:43):
they said, you know, there there are certain practical applications
for this kind of approach, but it's not gonna be
like we're gonna design these robots and they are going
to magically make all of our factories work at efficiency. Uh.
It's uh. They said that there's a challenge to developing
automas design of quote a large morphological diversity end quote
(17:06):
morphological obviously meaning that the actual form of these things. Yeah.
So the basic design of the experiment was to use
a robotic arm that was the Mama butt, so it's
a robotic arm, and had to gripper fingers, so just
very simple robot um. One of the gripper fingers actually
had a a little nozzle through which it could squirt glue,
(17:28):
so they could glue pieces together. Uh, and it was
connected to a computer that was running the evolutionary software
to build the robots out of these little cube modules.
And they had two types of modules. They had passive modules,
which really just small wooden blocks that were painted black.
The reason they were painted black was so that the
cameras that the robot was using essentially as eyes, could
(17:49):
easily pick them up. Let's see where the little passive
ones were. So these couldn't do anything, right, they were
just connectors. And then you had the active modules. These
were a larger cube and one face of the cube
was attached to a small motor that could rotate at
a certain amplitude in frequency, so it could rotate the
face of the cube, which is kind of the basis
(18:10):
of the locomotion of the piece. Right, So, so imagine
a Rubik's cube where only one side can rotate around,
but it does so on a motor. That's essentially what
we're talking about. And so different cubes would have faces
that rotated at different speeds, and the all of the
information was contained within the cube so that the robot
would be aware of that. It wasn't like the robot
(18:31):
was randomly picking blocks and some could turn at a
certain speed and others would turn at half speed or
twice speed or whatever. The robot was actually quote unquote
aware of which cubes could do what and uh and
by by fitting them together in different combinations, uh yeah,
you could you could get the box to kind of
kind of wiggle, to kind of hop and wiggle across
(18:52):
the table. It reminds me um of the motion of
some of the the less mobile toys and toy story
um kind of just sort of like scuttling along it
ones that didn't have legs. Yeah. Yeah, it's flipping adorable,
you guys. I am going to see if we can
find video that's share able to share with you because
it is so cute. It makes me think of a
(19:13):
lot of like, you know, wind up toys that that
move because they're doing this repetitive motion. Uh, and that's
enough to propel them across a surface. Uh, not with anything.
They can't necessarily steer or anything, but they can move
the same sort of thing. Like the idea was that,
all right, well, if we pair this passive module with
(19:34):
these two active modules in this configuration, will that create
movement that's faster and more effective than this other design.
So the computer they used was a regular desktop PC,
nothing particularly special about it, but it was running a
controller program that was using the matt Lab language m
A T L A B and that's used for interactive
(19:55):
environments of developing algorithms, among lots of other stuff. The
robotic arm could easily grip and rotate any module it shows,
and could stick two of them together using what they
called a hot melt adhesive or h m A. Yeah,
essentially like hot glue, like glue gun glue is more
or less what that was, And it could build robots
by combining multiple modules together. Each module had what they
(20:19):
called a gene, which essentially described the type of module
it was and the motor control of that module, and
also included the construction parameters. Of the module construction parameters. Essentially,
it was things like basic rules for the robot to
understand so that it could effectively build a robot. In
other words, if you want to build a robot, you're
(20:39):
probably gonna need to put a larger piece down first
before you put a smaller piece on top. It was
building them vertically, so it needed to know, Hey, if
you try and build a robot this other way, this
thing's gonna fall over. So you need to be able
to build them in you know, these are the general
rules you need to follow, essentially the rules of physics,
so that your robot will be built the way you
(21:00):
intend to build it. Uh. So a robot was made
up of these modules, and it was said to have
a genome. It was the collection of these genes and
uh these genes would either work or not work together.
And robots could have between one to five genes modules. Yeah,
now they if it's having five. The max was three
(21:22):
active too passive. You could not have more than three
active simply because it would make the robot too heavy
for the gripp or to grip. Right. Yeah, that's that's
one of those limiting factors in reality there. Yeah, so
if it were more than fifty grams of mass. The
robotic arm, the gripper just couldn't maintain a grip. It
would drop it and it would break. And so practically
(21:42):
you could have a maximum of three active and two
passive pieces connected altogether. They didn't necessarily have to have
that many. Some of them were only three or two
uh two blocks large at large two blocks in total,
but at any rate, the construction parameters were these very
simple directions and it allowed a lot of flexibility. So
(22:07):
the robot arm essentially could decide which modules to use
and how to connect the two together or three or
four however many together, and then test it so that
the what would happen after it builds one of these,
after it squirts the glue and everything. Um, you could
actually send it a genome essentially a recipe saying here
(22:28):
are the modules I want you to use and the
configuration I want you to put them in, and this
is going to represent the first generation. You could do that,
or you could make it uh more of a random approach. Well.
Once built, the mother robot would lift the finish modular robot,
move it to the testing area, which is just a
flat surface for the robot to crawl across, also known
(22:49):
as the arena. In some cases it was a hard surface.
In other cases they used a carpeted service, and at
least in one they used a foam surface. Yeah. They
they slowly moved towards the foam surface as they realized
throughout the experiment that the robots were having undo trouble.
I was sad to learn that the foam they were
(23:09):
talking about was like, you know, kind of like mattress
foam as opposed to foam party. It's a little bit
robots just getting down and raving. I come from the
specific time, y'all. Anyway, Wow, Joe to a foam party
has been no pope, you like pink II. Here we go,
(23:33):
all right. So the modular robots would then be activated wirelessly.
They had inside of them essentially a receiver for Bluetooth
or WiFi, and that would activate them to go into uh,
you know, their basic motorized action. Yeah, and then there
would be a couple of cameras, including an overhead camera
(23:55):
that would measure their progress across the surface, and then
they would essentially, uh take the the distance they traveled
and within a certain amount of time. Originally they went
with eight seconds, so after eight seconds, we see how
far they've gone. Uh. They had to switch that to
four seconds because later tests the modules were moving quote
(24:19):
unquote so quickly. Yea, it was working, so they needed
to have They needed to shorten the amount of time
because otherwise the robots would just travel out of the
view of the camera. So they had to shorten the
amount of time so that the robot arm could make
determinations of which ones were the most effective. Uh So,
after testing, the modular robots were disassembled by hand, so
(24:42):
the so the mother did not have to kill her babies. Uh,
the unfeeling scientists got to do that. Uh. So then
they had to also remove all the h a the
hot melted agent material, and then they were replaced onto
the work area. The work area as you would imagine. Uh,
each module had its specific place in the work area,
(25:03):
so that way the robot quote unquote knew where to
go to pick specific modules. Because these machines are not
that smart, right, you have to put the things in
the right orientation and the right um location for the
robot to be able to grab them. We've talked about
this with other robots too that are combining objects if
you have the objects in a specific order. I think
(25:24):
we talked about this with the cooking robots. Specifically you
have to have them in a specific place, to specific order,
or else everything is just going to come out mixed
up and awful. It couldn't recognize what was what. Yeah, yeah,
it's just following like it understands quote unquote understands the
uh qualities of each item, but only if the right
item is in the right place. So if you put
(25:46):
all the active modules where the past of one should
be and vice versa, it would it's not gonna be
able to tell the difference exactly. So at that point,
the evolutionary system would begin to tweak the genomes. It
could swap out genes in a genome, or it could
combine different genomes. So essentially it would be like breeding
two robots, saying, if if robot from this generation and
(26:08):
another robot from that generation were to combine, here are
the qualities that would emerge from that, keeping in mind
essentially the robots choosing which qualities would emerge. Because you're
still limited. You can't have more than five genes, so
it's not like it would just be additive. It would
have to be selective in which genes from which two
(26:29):
got selected to be combined into a new one. And
then the theory was that, or at least the hope
was that it could design another generation of better performing robots,
and some of them might not perform well. It may
turn out that the genes from robot one and the
genes from robots seven are not as compatible and they
actually perform worse than either one or seven did individually
(26:51):
in the generation before. And that happened a couple of times, Yes,
it did. So they held a total of five experiments,
and between these experiments they changed up the surface. They
made some tweaks to various rules. They actually numbered them
one A through one D and then two. Yeah, they
did five five different runs, but four of them fit
(27:13):
under the category of experiment one. So each experiment started
with ten agents other in other words, ten basic robot
designs UH that represented ten different genomes, and then each
experiment ran through ten generations, so you've got a hundred
different designs total per experiment and UH. The first four
(27:34):
experiments started with some randomly generated designs consisting of one
to three motorized elements, and they put all the construction
constraints they had designed in play for those initial experiments.
The testing environment was changed a couple of times. That's
when they started with a hard ground, then they moved
to the carpeted service and then the foam h and
humans started helping out the mom about eventually to hold
(27:57):
the component's study during the build phase, uh so that
she would not drop them basically. Uh And they also
began manually inspecting the baby bots to be sure that
none of the pieces were going to collide with each
other during the test, because that would damage the pieces
and make things less fun for everyone involved. Yeah, especially
since that would affect any future Yeah, because if you
(28:20):
if the cube gets damaged, then you it's hard to
determine if the robot would have performed better had it
worked with a brand new cube. And uh So they
were being very careful at that point. A little human
intervention wasn't necessary because it was you know, this is
like a proof of concept type of approach anyway, So
the fifth experiment, the initial generation of agents was not
(28:43):
generated randomly. Instead, they picked some of the best performing
designs that came out of the previous experiments, So generation
one was actually made up of robots that had already
been built in the first four experiments, and just said,
all right, let's made me think of you FC. Let's
take the champions of all these different fighting disciplines and
(29:04):
put them together and see what happens. Except instead of fighting,
they're supposed to make babies, so totally different UFC. I
guess in that sense, I would suppose. So yeah. I mean,
I don't know a whole lot about the UFC, but
I know a great article written by a certain Jonathan
Strickland you should read. Uh. Interestingly, the test could that
be found on how stuff works dot com. Could. In fact,
(29:25):
if you go to how stuff works dot com and
look how the UFC works, you will find an article
I wrote ages ago. It was awesome. Um. So interestingly,
the test that produced the most improvement from first generation
to the last generation was the fourth test. They actually
plotted out each generation's performance on a kind of a
line graph chart, and UH, they average the ten robots fitness,
(29:49):
that's what they called the performance of moving across the surface. UH,
and they the fourth test saw steady improvement with one exception, uh,
where at general ration five there was a slight dip
in performance and generation ten saw very slight decline in
performance from generation nine, so generation nine did the best
out of the fourth test. The fifth test saw the
(30:12):
longest declining trends. So this was the This was the
one where they took the champions from the previous tests
and started with those. So this is the DEVO group. Yeah, yes,
and exactly. They had a devolution uh in generation over
generation from three to seven, so their their performance actually declined,
(30:32):
not steadily, but in each generation there was a decrease
in fitness. But then it all turned around. Yeah, they
started to see improvements again and all the way through
to generation ten. So it was one of those things
where by the end, I think in every single case,
the generation ten robots outperformed the initial generation of robots
(30:53):
in all the tests. There were some cases where generation
ten didn't outperform, maybe generation in seven or eight, but
in all of them they outperform generation one. Uh. It
was really interesting. The fourth test, top performing agents in
the first generation average two point eight centimeters per second,
and by the tenth generation it had increased to six
(31:14):
point seven centimeters per second. Yeah, so more than twice
and twice yeah, yeah, so that's pretty exciting. That they
were able to take this approach and increase the speed
of these mobile agents by a factor more than a
factor of two, which yeah, yeah, the researchers did. Uh.
Did note that the disadvantage and having the mom about
manually test each generation is that it takes time, and
(31:38):
so there's a little bit of a payoff balance between
running simulations first versus going straight into that real world testing.
And they were talking about how they hoped to streamline
the process in the future by using simulations to select
the most likely successful models and then begin testing with
those instead of kind of starting from scratch. Yeah. Uh.
And in fact, they had talked in their paper or
(32:00):
about how, yeah, this is this is a balance, right,
because when you go with the pure simulation mode, it
may turn out that when you when you transfer the
simulation to reality, things don't behave exactly as you had anticipated.
Perhaps the simulation was unable to take all the different
factors into account, or it just maybe that, you know,
it's just in the real world stuff behaves a little
(32:22):
differently than the idealized virtual world. But a combination of
the two is probably the best approach, because, like Laurence says,
if you do everything physically. Then you need to have
the luxury of time on your side, just because it
will take this time to physically build these things, and
plenty of blocks and glue, yeah, and lots of humans
to d glue the blocks, and and these robots are
(32:46):
just blocks, right, I mean this this is the about
as unsophisticated a robot as you can get. So if
we were talking about a robot designing a future robot
capable of doing something really sophisticated, it would obviously take
even more time. Yeah, I don't know. I've met some
pretty ansophisticated robots. Well, I mean they really like cheese.
It's they don't tip. Well, um, they prefer they prefer
(33:10):
light beer to I p a s. Yeah. Yeah, that's
that's why I was giggling a second second ago. That's sad,
Please please continue. Well, no, no, The cool thing to
think about is that imagine a future where we have
machines capable of designing a new generation of machines that
are better adept at doing whatever they need to do
(33:31):
than the previous ones, and then can even learn from
that and build even better ones in the future, perhaps
even building a better computer to design the next generations
and eventually you arrive at or deep thought, because that's
exactly what deep thought was in Hitchecker's Guide to the Galaxy.
Was that well, deep thought said it could give the
(33:52):
answer to the question of life, the universe and everything,
but could not give the question and said, in order
to get to the question, it could design a computer
that would be even more advanced and be able to
answer that question after a really long time. But that's all.
That's the best it could do, is that it could
design a better machine than it to be able to
(34:14):
answer or to come up with what the question was.
Of course, we all remember the answers forty two question
turned out to be was six times eight. Just shows
you that something's wrong with the universe. But now this
is really kind of I'll get you a computer. Um. Yeah,
(34:38):
obviously it's a joke in hitchecker Sketch the Galaxy. But
this is this is the neat ideas, This this approach
where we can set certain types of machinery on a
pathway to reach this possibly increasingly efficient means of evolution
to create better tools. Um. This is also obviously one
(35:00):
of the principles that underlies certain versions of the singularity, right,
like this idea that we get to a point where
evolution is so constant that there is no meaningful way
to describe the present anymore because it's everything is changing
all the time. And uh, and this is sort of
the kind of stuff that would be necessary for that
(35:20):
particular version of the singularity to come to pass, will it. Well,
let's just say that based upon what's going on right now,
it's gonna take some time. We got some real cute
wiggly blocks in in the meanwhile, though, So yeah, I
am still very skeptical that we will see anything close
to the singularity on a time scale that kurtz Wild
has predicted. Do you mean twenty to forty years? Yeah,
(35:59):
it's so this was kind of fun. I mean, if you,
if you get a chance, you can read the paper.
The paper is actually quite easy to read. It's um
it's not an inaccessible paper, and it is available. It
is very accessible. It's it's available for free on the internet. Yep.
So you can read all about the experiments. They go
into detail. They really I didn't go into a lot
(36:20):
of detail about the differences between the five different runs
they did, just because it would have gotten super technical
still understandable, but just bogged down in a lot of
technical details. But it's all there in their their paper.
We'll try to remember to link it on various forms
of social media if you want to google it for yourself.
The full name of the paper is Morphological Evolution of
(36:42):
Physical Robots through model free phenotype development blin Yes. And
if you want to get in touch with us and
and ask us to cover a specific topic. Maybe there's
something about the future you have always been curious about
and would like to hear our take on it, let
us know. So does an email the addresses f W
thinking at how Stuff Works dot com, or drop us
(37:04):
a line on Twitter or Google Plus. We are f
W thinking at both of those, or search f W
thinking and Facebook. We'll pop right up. You can leave
us a message and we'll talk to you again really
soon for more on this topic and the future of technology.
This is forward Sinking dot Com, brought to you by Toyota.
(37:37):
Let's go places,
Brought to you by Toyota. Let's go places. Welcome to
Forward Thinking. Hey everyone, and welcome to Forward Thinking, the
podcast that looks at the future. It says, mama's don't
let your babies grow up to be robots. I'm Jovin
(00:20):
Strickland and I'm Joe McCormick. And have you two ever
noticed how much better robots are than organic life forms
all the time. That's why I only work with various
things that allow me to interact with machines as opposed
to actual human beings. Like if it's a prerecorded type
message and I navigate through a menu using buttons, and
(00:43):
I don't have to talk to a person, that's awesome.
They're certainly less offensively smelly. Well that's just because we
haven't mastered the artificial intelligence sector of creating really lifelike
bo Yeah, well we have, we have scientists working on that. Yeah, yeah, no,
I was kidding. But there are, of course situations where
(01:05):
it's much better to be a robot than to be
an organic life form, and we've talked about plenty of
them on the podcast before. Maybe in space exploration where
you're at risk, or in or in search and rescue
scenarios where again a human would be at risk or
you know, any number of hundreds of others. Really, Yeah,
just just performing the same physical job. Anything that would
(01:27):
make a human board or or repeated stress injury, that
kind of stuff. I mean things that where you do
it once, it doesn't hurt, but you do it, you know,
once every minute for eight hours a day, and it's
and eventually, yeah, the button pushing muscles start to wear
down in the bones and etcetera. Yeah, robots don't get
don't get physically injured, they don't get emotionally injured. It's
(01:51):
real hard to traumatize a robot. Yeah, I know. Now,
we we've talked about the advantages. Of course, the life
forms have over robots and limited in each individual case.
So a person like a worker is a lot easier
to train for a new task than a robot is. Well,
depending on if if the robot is mechanically suited to
(02:13):
the new task, you can reprogram it s well. But
but a worker can learn many new tasks. Yeah, hypothetically,
most robots are pretty good at doing one thing right, right, Um, Yeah, robots,
as it turns out, they pretty much can do what
you built them to do, and they're pretty bad at
doing anything outside of that. In general, right, Robots don't
(02:34):
really evolve. Yeah, so yeah, so humans are more adaptable
on the individual scale, but they're also more adaptable as
a type of thing. Sure, over time, if the air
pressures from the environment on a living organism, it will evolve,
it will adapt, or it might go extinct. But assuming
it doesn't go extinct, it will adapt and evolve to
(02:54):
to fit the environment that it needs to make its
living in. Robots don't do that, no, they remain. Yeah,
so that's what we're going to talk about today exactly.
So let's talk about like, like general robots right now.
They they are too pretty distinct pathways for robot design
in general, and of course there are a lot of
(03:15):
branches within these, but we can go one path where
we design a robot for a specific task and that's
all it is supposed to do. So rumba is a
great example. What a rumba is supposed to do is
clean the floor. It's not supposed to do anything else
other than it doesn't clean the cat box. In my opinion,
this is this is probably like of robots and all
(03:37):
of the good ones. Yeah, the most efficient and the
most effective robots, I would argue, are of this type
because we only have to concern ourselves with the design
elements that will enable the robot to complete its task,
and we eliminate everything else. Right, you don't need to
include anything that doesn't involve cleaning the floor, or navigation
(03:58):
or returning to like a docking station to recharge. When
you're designing a robot like the Rumba, you just need
those basic elements in it. Anything else is superfluous. Uh.
Doesn't stop us from doing things like having crazy Rumba
fights where we put a balloon and something sharp on
on a couple of them and have them joused each other.
(04:19):
But it doesn't mean we don't put our cats on
top of them and just watch what happen. Maybe put
our cat in a shark costume first and then put
the cat on top of a roomba. Yeah, that sort
of thing. But in general, they just do what they
were supposed to do, and if you wanted them to
do anything else, you're kind of out of luck. Or
we tried to build general purpose robots that are capable
of doing lots of different tasks hopefully yeah, to varying
(04:44):
degrees of mediocrity because it's hard. It's not because you know, No, yeah,
it's not because the people working on these aren't smart. Yes,
because the job is a thousand times more difficult, right,
because you have to anticipate lots of different things, a
lot of different changing conditions. This would be the example
of the darker robots we talked about previously, where UH
(05:06):
challenge robots exactly. Yes, the DARPA Challenge robots, the ones
that had to replicate a uh kind of a first
responder situation, search and rescue kind of. Yeah. They to
do like drive a car to a building, open a door,
go inside, twist to lever, and cut a hole in
the wall. They had to plug in a cable in
one case. Yeah, they had to walk across rubble. They
(05:30):
had to go up some stairs and then fall over.
Uh didn't have to fall over, but a lot of
them sure did. At any rate, we saw how difficult
it is to design a robot that can do these things.
And keep in mind, these were all tasks that the
various teams knew about beforehand they were designing these. Yeah, yeah,
I mean it was this was all stuff that they
(05:51):
knew they were going to have, the robot was going
to have to be able to do. Uh. So, building
a robot that could end up anticipating all sorts of stuff,
whether it's a single robot or a you know, a
robot that can then design the next generation of robots.
That's something that we really haven't perfected yet, but there
are some really good reasons why we would want to. Clearly,
(06:14):
being able to have a robot that could either adapt
itself or adapt the next generation of robots would lead
to much more efficient machines over time, and so this
is something that we would really like to see in technology.
It's not it's not just in robotics. We're also seeing
it in computers. So there's they call it, you know,
(06:37):
evolutionary computation or evolutionary computers or evolutionary robotics because that's
essentially what we what we mean. We're talking about a
machine that looks at the ability of another, like a
next generation of machines to do a particular task, evaluates them,
and then makes decisions on how to alter that generation
(06:57):
to produce an even more effective generation after it. Uh
the goal always being to come up with the most
ideal design for whatever purpose you have in mind. And
this isn't a super super new idea, no. Now, there
have been people who have been working on this for
a while. Back in two thousand eleven, researchers with NASA
(07:17):
published a paper on evolutionary computation that was used to
design new efficient antenna's way back in two thousand six.
So the paper was published in two thousand and eleven,
but the actual project was yeah in two thousand and six.
And what they what they were pointing out was that
designing an antenna is really challenging. It requires that the
(07:38):
builder to have a very detailed knowledge of how antenna's work,
which already limits your pool of various builders, and even
then it's just painstaking to create an actual efficient, working antenna.
So what they did was they created evolutionary algorithms for
a computer to design an antenna for spacecraft auto matically
(08:00):
that the computer system went through various designs of antenna
and essentially simulated tests of them to see which one
would be the most effective. They then took the designs
that were predicted to be the highest performing antennas by
this computer program, and they took it to the same
manufacturer who was already building the antenna intended for the spacecraft.
(08:24):
Then they tested all three of the antenna, the two
that were the two best performer ones in the computer program,
and the one that people had designed for the spacecraft,
and found that the two that the computer had designed
were more effective. Yeah, so it was one of those
examples of this approach actually working better than what we
humans could do. Now, maybe you don't know the answer
(08:45):
to this, but if you do, I'd be interested. In
this case, we're the what what we might call the
mutations introduced into the models in in the algorithm that
that tested out antenna's where the mutations direct did where
they programmed in? Or were they random mutations? Like? Was
it really more like evolution where a random things thing
(09:08):
happens and then the system tests is this any better? Well,
as we've talked about in previous episodes, random with machines
is really hard to do. It was more like they
were the computer was given an enormous number of variables
and started to test them in various configurations without going
through every possible one. Because the goal of evolutionary computer
(09:32):
computation or evolutionary robotics is to make sure that you
come to the most ideal form of whatever it is
you're going for without having to test every possible variation. Because,
of course, thus things get more complex, those variations increase
in number until it would take you till the end
of time to test all the different ones to come
(09:53):
up version one one seven eight nine seven, seven seven
four three is the best. You would take forever to
go through all of those. So what these are designed
to do is to test ones that are a best
guests already of being effective, measuring those against other best guesses,
then combining ones. Like if you were to find a
(10:15):
generation that works particularly well and another generation that works
particularly well, you might say, well, what happens if I
combine the best design elements of both of these into
a new design. Does that increase the effect of the
efficiency and effectiveness or does it decrease it? Because it doesn't,
you know, adding to awesome things together does not guarantee
(10:36):
you to get a third, even more awesome thing. Oh yeah, yeah.
I was wondering because I'd read about the use of
evolutionary algorithms before in the design of new planetary rovers
like unmanned planetary rovers, and testing different variations on the
models in a in a computer simulation that would naturally
select the highest performing ones. Yeah. Yeah, it's very similar
(10:59):
to that. And uh, the thing about the difference, you know,
we're going to talk about a robot that works along
these principles. The biggest difference between a robot and a computer,
obviously is that a robot is working with actual physical matter,
not just simulations. And so there's some practical limitations that
you encounter in that case. Right, you have to work
(11:21):
with real physical matter that has weight, it has mass,
it occupies space, there's a limit to how heavy it
can be. Uh, you know, you can't just magically increase
the size of whatever it is in a simulation and
then just see how it works. You have to physically
put this thing together. And so it's a really interesting approach.
(11:44):
And we have seen a couple of examples of people
working with true physical matter, but uh, it's mostly been
in things like working with a program that builds stuff
out of lego. So there is an actual example of this.
There were some evolutionary computational experts that design systems that
would allow a machine to build other objects out of legos,
(12:04):
but you had to build limitations into the computer program
so that the machine would follow the rules of physics.
In other words, you couldn't have to lego pieces occupy
the same physical space at the same time, right, Sure,
And some of these computations aren't going to I mean,
it's hard as a programmer to build all of those
(12:24):
limitations into something that has to work in reality, because
the computer isn't going to understand I mean basic stuff,
and so so it might when you take it out
of the lab, out of the out of the computer
simulation stage, it might operate very differently than you were
expecting it. Then the computer was expecting it to. An
analogy I would make and has nothing to do with
(12:46):
computers or robotics, but an analogy I would make is
if you were running a role playing game. You're you're
the game master of a role play game, and you
think you have anticipated everything your players are going to do,
and you have made up a masterpiece of a module,
and your players are going to have an amazing time,
and three minutes in the players decided to do something
(13:07):
you could not have possibly anticipated. It makes perfect sense
within the context of the game. And then you and you, yeah,
you gotta throw away all the stuff you worked on
and say, well, well, I guess like it's sort of
like if you imagine a horror movie where the characters
all walk up to the spooky house, take one look
and say no, and then walk away like, well, there
(13:28):
goes the horror movie. Same sort of thing. Well, we
wanted to specifically talk about some researchers from the University
of Cambridge who were working with a robotic system that
used evolutionary robotics. So imagine, essentially, you've got a robot
that builds other robots and tests them to see which
(13:49):
designs work the best, and then either UH either has
a design go forward or eliminates designs and starts tweaking
things to try and find the best physical design of
a robot to complete a specific task. Now that sounds
super cool. I need to explain some stuff first so
(14:09):
we can manage our expectations. Yeah, yeah, okay, so so
so performance of these of these baby bots, as we
might call them UH in this case was how fast
and far it could move, like kind of scuttle across
the surface exactly. So it's not like it was performing
open heart surgery. No, No, they decided to hold that
(14:30):
for the next experiment. Perhaps, Now these robots all we
were supposed to do. We're create a locomotion power that
could move it from one point to another point, and
uh they would. The experiment measured how long it took
these little devices to move across this particular expanse they
(14:52):
changed up. That didn't make the cut. They didn't they
didn't continue on. Uh that they were taken apart. They
were taking up bodily by the mamabot. Actually, to be fair,
I think all of them were taken apart, but yes,
uh uh yeah they were. They'd be scuttled, they'd be
scuttled and then they'd be scuttled and and and harvested
(15:14):
for their organs essentially. So that was the that was
the criteria that the robot used. Um and here's how
it gets here. Let's break this down because that's that's
like an overview of what happened, but it's really interesting
when you get to the specifics. So what they wanted
to do is they wanted to design a system that
could design, test, and change robotic designs relatively rapidly. And
(15:40):
it needed to be a a system that could work
within the lab, but wasn't designed to be a practical
system that could go through every single possible arrangement of
the various pieces. For yeah, um, the pieces are essentially
two different types of cubes that will get to in
a second. So they wanted their approach to contain this
(16:00):
is a quote from their paper paper, a limited number
of evolutionary iterations, and that was, you know, the the
idea of let's let evolution lead us to what should
be the best version of these robotic designs, as opposed
to just doing trial and error, measuring everything and then
going with the top performer. Now, they did say in
(16:22):
their paper that while this is really interesting and could
in fact be a breakthrough in science and technology, you
have to admit that there are limitations to this technology
cannot necessarily be applied to mass production. Mass production relies
heavily on automation. Automation is different from mass customization. So
(16:43):
they said, you know, there there are certain practical applications
for this kind of approach, but it's not gonna be
like we're gonna design these robots and they are going
to magically make all of our factories work at efficiency. Uh.
It's uh. They said that there's a challenge to developing
automas design of quote a large morphological diversity end quote
(17:06):
morphological obviously meaning that the actual form of these things. Yeah.
So the basic design of the experiment was to use
a robotic arm that was the Mama butt, so it's
a robotic arm, and had to gripper fingers, so just
very simple robot um. One of the gripper fingers actually
had a a little nozzle through which it could squirt glue,
(17:28):
so they could glue pieces together. Uh, and it was
connected to a computer that was running the evolutionary software
to build the robots out of these little cube modules.
And they had two types of modules. They had passive modules,
which really just small wooden blocks that were painted black.
The reason they were painted black was so that the
cameras that the robot was using essentially as eyes, could
(17:49):
easily pick them up. Let's see where the little passive
ones were. So these couldn't do anything, right, they were
just connectors. And then you had the active modules. These
were a larger cube and one face of the cube
was attached to a small motor that could rotate at
a certain amplitude in frequency, so it could rotate the
face of the cube, which is kind of the basis
(18:10):
of the locomotion of the piece. Right, So, so imagine
a Rubik's cube where only one side can rotate around,
but it does so on a motor. That's essentially what
we're talking about. And so different cubes would have faces
that rotated at different speeds, and the all of the
information was contained within the cube so that the robot
would be aware of that. It wasn't like the robot
(18:31):
was randomly picking blocks and some could turn at a
certain speed and others would turn at half speed or
twice speed or whatever. The robot was actually quote unquote
aware of which cubes could do what and uh and
by by fitting them together in different combinations, uh yeah,
you could you could get the box to kind of
kind of wiggle, to kind of hop and wiggle across
(18:52):
the table. It reminds me um of the motion of
some of the the less mobile toys and toy story
um kind of just sort of like scuttling along it
ones that didn't have legs. Yeah. Yeah, it's flipping adorable,
you guys. I am going to see if we can
find video that's share able to share with you because
it is so cute. It makes me think of a
(19:13):
lot of like, you know, wind up toys that that
move because they're doing this repetitive motion. Uh, and that's
enough to propel them across a surface. Uh, not with anything.
They can't necessarily steer or anything, but they can move
the same sort of thing. Like the idea was that,
all right, well, if we pair this passive module with
(19:34):
these two active modules in this configuration, will that create
movement that's faster and more effective than this other design.
So the computer they used was a regular desktop PC,
nothing particularly special about it, but it was running a
controller program that was using the matt Lab language m
A T L A B and that's used for interactive
(19:55):
environments of developing algorithms, among lots of other stuff. The
robotic arm could easily grip and rotate any module it shows,
and could stick two of them together using what they
called a hot melt adhesive or h m A. Yeah,
essentially like hot glue, like glue gun glue is more
or less what that was, And it could build robots
by combining multiple modules together. Each module had what they
(20:19):
called a gene, which essentially described the type of module
it was and the motor control of that module, and
also included the construction parameters. Of the module construction parameters. Essentially,
it was things like basic rules for the robot to
understand so that it could effectively build a robot. In
other words, if you want to build a robot, you're
(20:39):
probably gonna need to put a larger piece down first
before you put a smaller piece on top. It was
building them vertically, so it needed to know, Hey, if
you try and build a robot this other way, this
thing's gonna fall over. So you need to be able
to build them in you know, these are the general
rules you need to follow, essentially the rules of physics,
so that your robot will be built the way you
(21:00):
intend to build it. Uh. So a robot was made
up of these modules, and it was said to have
a genome. It was the collection of these genes and
uh these genes would either work or not work together.
And robots could have between one to five genes modules. Yeah,
now they if it's having five. The max was three
(21:22):
active too passive. You could not have more than three
active simply because it would make the robot too heavy
for the gripp or to grip. Right. Yeah, that's that's
one of those limiting factors in reality there. Yeah, so
if it were more than fifty grams of mass. The
robotic arm, the gripper just couldn't maintain a grip. It
would drop it and it would break. And so practically
(21:42):
you could have a maximum of three active and two
passive pieces connected altogether. They didn't necessarily have to have
that many. Some of them were only three or two
uh two blocks large at large two blocks in total,
but at any rate, the construction parameters were these very
simple directions and it allowed a lot of flexibility. So
(22:07):
the robot arm essentially could decide which modules to use
and how to connect the two together or three or
four however many together, and then test it so that
the what would happen after it builds one of these,
after it squirts the glue and everything. Um, you could
actually send it a genome essentially a recipe saying here
(22:28):
are the modules I want you to use and the
configuration I want you to put them in, and this
is going to represent the first generation. You could do that,
or you could make it uh more of a random approach. Well.
Once built, the mother robot would lift the finish modular robot,
move it to the testing area, which is just a
flat surface for the robot to crawl across, also known
(22:49):
as the arena. In some cases it was a hard surface.
In other cases they used a carpeted service, and at
least in one they used a foam surface. Yeah. They
they slowly moved towards the foam surface as they realized
throughout the experiment that the robots were having undo trouble.
I was sad to learn that the foam they were
(23:09):
talking about was like, you know, kind of like mattress
foam as opposed to foam party. It's a little bit
robots just getting down and raving. I come from the
specific time, y'all. Anyway, Wow, Joe to a foam party
has been no pope, you like pink II. Here we go,
(23:33):
all right. So the modular robots would then be activated wirelessly.
They had inside of them essentially a receiver for Bluetooth
or WiFi, and that would activate them to go into uh,
you know, their basic motorized action. Yeah, and then there
would be a couple of cameras, including an overhead camera
(23:55):
that would measure their progress across the surface, and then
they would essentially, uh take the the distance they traveled
and within a certain amount of time. Originally they went
with eight seconds, so after eight seconds, we see how
far they've gone. Uh. They had to switch that to
four seconds because later tests the modules were moving quote
(24:19):
unquote so quickly. Yea, it was working, so they needed
to have They needed to shorten the amount of time
because otherwise the robots would just travel out of the
view of the camera. So they had to shorten the
amount of time so that the robot arm could make
determinations of which ones were the most effective. Uh So,
after testing, the modular robots were disassembled by hand, so
(24:42):
the so the mother did not have to kill her babies. Uh,
the unfeeling scientists got to do that. Uh. So then
they had to also remove all the h a the
hot melted agent material, and then they were replaced onto
the work area. The work area as you would imagine. Uh,
each module had its specific place in the work area,
(25:03):
so that way the robot quote unquote knew where to
go to pick specific modules. Because these machines are not
that smart, right, you have to put the things in
the right orientation and the right um location for the
robot to be able to grab them. We've talked about
this with other robots too that are combining objects if
you have the objects in a specific order. I think
(25:24):
we talked about this with the cooking robots. Specifically you
have to have them in a specific place, to specific order,
or else everything is just going to come out mixed
up and awful. It couldn't recognize what was what. Yeah, yeah,
it's just following like it understands quote unquote understands the
uh qualities of each item, but only if the right
item is in the right place. So if you put
(25:46):
all the active modules where the past of one should
be and vice versa, it would it's not gonna be
able to tell the difference exactly. So at that point,
the evolutionary system would begin to tweak the genomes. It
could swap out genes in a genome, or it could
combine different genomes. So essentially it would be like breeding
two robots, saying, if if robot from this generation and
(26:08):
another robot from that generation were to combine, here are
the qualities that would emerge from that, keeping in mind
essentially the robots choosing which qualities would emerge. Because you're
still limited. You can't have more than five genes, so
it's not like it would just be additive. It would
have to be selective in which genes from which two
(26:29):
got selected to be combined into a new one. And
then the theory was that, or at least the hope
was that it could design another generation of better performing robots,
and some of them might not perform well. It may
turn out that the genes from robot one and the
genes from robots seven are not as compatible and they
actually perform worse than either one or seven did individually
(26:51):
in the generation before. And that happened a couple of times, Yes,
it did. So they held a total of five experiments,
and between these experiments they changed up the surface. They
made some tweaks to various rules. They actually numbered them
one A through one D and then two. Yeah, they
did five five different runs, but four of them fit
(27:13):
under the category of experiment one. So each experiment started
with ten agents other in other words, ten basic robot
designs UH that represented ten different genomes, and then each
experiment ran through ten generations, so you've got a hundred
different designs total per experiment and UH. The first four
(27:34):
experiments started with some randomly generated designs consisting of one
to three motorized elements, and they put all the construction
constraints they had designed in play for those initial experiments.
The testing environment was changed a couple of times. That's
when they started with a hard ground, then they moved
to the carpeted service and then the foam h and
humans started helping out the mom about eventually to hold
(27:57):
the component's study during the build phase, uh so that
she would not drop them basically. Uh And they also
began manually inspecting the baby bots to be sure that
none of the pieces were going to collide with each
other during the test, because that would damage the pieces
and make things less fun for everyone involved. Yeah, especially
since that would affect any future Yeah, because if you
(28:20):
if the cube gets damaged, then you it's hard to
determine if the robot would have performed better had it
worked with a brand new cube. And uh So they
were being very careful at that point. A little human
intervention wasn't necessary because it was you know, this is
like a proof of concept type of approach anyway, So
the fifth experiment, the initial generation of agents was not
(28:43):
generated randomly. Instead, they picked some of the best performing
designs that came out of the previous experiments, So generation
one was actually made up of robots that had already
been built in the first four experiments, and just said,
all right, let's made me think of you FC. Let's
take the champions of all these different fighting disciplines and
(29:04):
put them together and see what happens. Except instead of fighting,
they're supposed to make babies, so totally different UFC. I
guess in that sense, I would suppose. So yeah. I mean,
I don't know a whole lot about the UFC, but
I know a great article written by a certain Jonathan
Strickland you should read. Uh. Interestingly, the test could that
be found on how stuff works dot com. Could. In fact,
(29:25):
if you go to how stuff works dot com and
look how the UFC works, you will find an article
I wrote ages ago. It was awesome. Um. So interestingly,
the test that produced the most improvement from first generation
to the last generation was the fourth test. They actually
plotted out each generation's performance on a kind of a
line graph chart, and UH, they average the ten robots fitness,
(29:49):
that's what they called the performance of moving across the surface. UH,
and they the fourth test saw steady improvement with one exception, uh,
where at general ration five there was a slight dip
in performance and generation ten saw very slight decline in
performance from generation nine, so generation nine did the best
out of the fourth test. The fifth test saw the
(30:12):
longest declining trends. So this was the This was the
one where they took the champions from the previous tests
and started with those. So this is the DEVO group. Yeah, yes,
and exactly. They had a devolution uh in generation over
generation from three to seven, so their their performance actually declined,
(30:32):
not steadily, but in each generation there was a decrease
in fitness. But then it all turned around. Yeah, they
started to see improvements again and all the way through
to generation ten. So it was one of those things
where by the end, I think in every single case,
the generation ten robots outperformed the initial generation of robots
(30:53):
in all the tests. There were some cases where generation
ten didn't outperform, maybe generation in seven or eight, but
in all of them they outperform generation one. Uh. It
was really interesting. The fourth test, top performing agents in
the first generation average two point eight centimeters per second,
and by the tenth generation it had increased to six
(31:14):
point seven centimeters per second. Yeah, so more than twice
and twice yeah, yeah, so that's pretty exciting. That they
were able to take this approach and increase the speed
of these mobile agents by a factor more than a
factor of two, which yeah, yeah, the researchers did. Uh.
Did note that the disadvantage and having the mom about
manually test each generation is that it takes time, and
(31:38):
so there's a little bit of a payoff balance between
running simulations first versus going straight into that real world testing.
And they were talking about how they hoped to streamline
the process in the future by using simulations to select
the most likely successful models and then begin testing with
those instead of kind of starting from scratch. Yeah. Uh.
And in fact, they had talked in their paper or
(32:00):
about how, yeah, this is this is a balance, right,
because when you go with the pure simulation mode, it
may turn out that when you when you transfer the
simulation to reality, things don't behave exactly as you had anticipated.
Perhaps the simulation was unable to take all the different
factors into account, or it just maybe that, you know,
it's just in the real world stuff behaves a little
(32:22):
differently than the idealized virtual world. But a combination of
the two is probably the best approach, because, like Laurence says,
if you do everything physically. Then you need to have
the luxury of time on your side, just because it
will take this time to physically build these things, and
plenty of blocks and glue, yeah, and lots of humans
to d glue the blocks, and and these robots are
(32:46):
just blocks, right, I mean this this is the about
as unsophisticated a robot as you can get. So if
we were talking about a robot designing a future robot
capable of doing something really sophisticated, it would obviously take
even more time. Yeah, I don't know. I've met some
pretty ansophisticated robots. Well, I mean they really like cheese.
It's they don't tip. Well, um, they prefer they prefer
(33:10):
light beer to I p a s. Yeah. Yeah, that's
that's why I was giggling a second second ago. That's sad,
Please please continue. Well, no, no, The cool thing to
think about is that imagine a future where we have
machines capable of designing a new generation of machines that
are better adept at doing whatever they need to do
(33:31):
than the previous ones, and then can even learn from
that and build even better ones in the future, perhaps
even building a better computer to design the next generations
and eventually you arrive at or deep thought, because that's
exactly what deep thought was in Hitchecker's Guide to the Galaxy.
Was that well, deep thought said it could give the
(33:52):
answer to the question of life, the universe and everything,
but could not give the question and said, in order
to get to the question, it could design a computer
that would be even more advanced and be able to
answer that question after a really long time. But that's all.
That's the best it could do, is that it could
design a better machine than it to be able to
(34:14):
answer or to come up with what the question was.
Of course, we all remember the answers forty two question
turned out to be was six times eight. Just shows
you that something's wrong with the universe. But now this
is really kind of I'll get you a computer. Um. Yeah,
(34:38):
obviously it's a joke in hitchecker Sketch the Galaxy. But
this is this is the neat ideas, This this approach
where we can set certain types of machinery on a
pathway to reach this possibly increasingly efficient means of evolution
to create better tools. Um. This is also obviously one
(35:00):
of the principles that underlies certain versions of the singularity, right,
like this idea that we get to a point where
evolution is so constant that there is no meaningful way
to describe the present anymore because it's everything is changing
all the time. And uh, and this is sort of
the kind of stuff that would be necessary for that
(35:20):
particular version of the singularity to come to pass, will it. Well,
let's just say that based upon what's going on right now,
it's gonna take some time. We got some real cute
wiggly blocks in in the meanwhile, though, So yeah, I
am still very skeptical that we will see anything close
to the singularity on a time scale that kurtz Wild
has predicted. Do you mean twenty to forty years? Yeah,
(35:59):
it's so this was kind of fun. I mean, if you,
if you get a chance, you can read the paper.
The paper is actually quite easy to read. It's um
it's not an inaccessible paper, and it is available. It
is very accessible. It's it's available for free on the internet. Yep.
So you can read all about the experiments. They go
into detail. They really I didn't go into a lot
(36:20):
of detail about the differences between the five different runs
they did, just because it would have gotten super technical
still understandable, but just bogged down in a lot of
technical details. But it's all there in their their paper.
We'll try to remember to link it on various forms
of social media if you want to google it for yourself.
The full name of the paper is Morphological Evolution of
(36:42):
Physical Robots through model free phenotype development blin Yes. And
if you want to get in touch with us and
and ask us to cover a specific topic. Maybe there's
something about the future you have always been curious about
and would like to hear our take on it, let
us know. So does an email the addresses f W
thinking at how Stuff Works dot com, or drop us
(37:04):
a line on Twitter or Google Plus. We are f
W thinking at both of those, or search f W
thinking and Facebook. We'll pop right up. You can leave
us a message and we'll talk to you again really
soon for more on this topic and the future of technology.
This is forward Sinking dot Com, brought to you by Toyota.
(37:37):
Let's go places,
Fw:Thinking News
Hosts And Creators
Jonathan Strickland
Joe McCormick
Lauren Vogelbaum
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.