Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:01
Ted Audio Collective. You're.
0:09
Listening to tear ducts daily. I'm your host.
0:12
Elite few. It turns out
0:14
that the things human babies can
0:16
master like picking up tiny blocks
0:18
are giant challenges for robots. In
0:20
his twenty twenty Three talk from
0:22
Ted ex marine, the roboticist Can
0:24
Goldberg takes us through how advances
0:26
in a a deep learning and
0:28
lead to big strides in training
0:31
robots to do even the most
0:33
precise tasks like and tying tangled
0:35
cables. I don't know about you,
0:37
but if robots could untangle my
0:39
necklace is, I'd line up for
0:41
that for sure. Hear. Us talk
0:43
after the break. Support.
0:47
For tedtalks, daily comes from Capital
0:49
One Bank with no Fees are
0:52
minimums. Banking with Capital One is
0:54
the easiest decision in the history
0:56
of decisions. even easier than deciding
0:58
to listen to another episode of
1:00
your favorite podcast. And with no
1:03
overdraft fees? Is it even a
1:05
decision? that's banking reimagine? what's in
1:07
your wallet Terms of life? The
1:09
Capital one.com/bank Capital One and a
1:11
member Ft. I see. Tear.
1:16
Ducts daily is brought to you by
1:18
Progressive where drivers who saved by switching
1:20
save nearly seven hundred and fifty dollars
1:22
on average quote now at progressive.com Progressive
1:24
Casualty Insurance Company and Affiliates national average
1:26
while savings with seven hundred forty four
1:28
dollars by new customers surveyed who saved
1:31
with Progressive between June twenty twenty Two
1:33
and May twenty twenty Three potential savings
1:35
will their. This
1:38
show is brought to you by Schwab.
1:40
With Schwab Investing name's it's easy to
1:42
invest in ideas you believe in like
1:45
electric vehicles, renewable energy, water, sustainability and
1:47
more. Choose from over forty themes by
1:49
as his or customize the stocks and
1:52
a theme to fit your goals. Learn
1:54
more at schwab.com/thematic Investing. I
1:58
have a feeling most people. This room.
2:00
would like to have a robot
2:03
at home. Be nice to be
2:05
able to do the chores and
2:07
take care of things. Where are
2:10
these robots? What's taking so long?
2:12
I mean we have our try
2:14
corridors and we have satellites. We
2:17
have laser beams. but where are
2:19
the robots? I
2:22
mean okay way we do have
2:24
some robots in our home by
2:26
tie. Not really doing anything that
2:28
exciting. Okay, South.
2:31
I've been doing research at U
2:33
C Berkeley for thirty years with
2:35
my students on robots and and
2:38
next ten minutes I'm gonna try
2:40
to explain the gap between sixteen
2:42
and reality. Than. The
2:44
Fields: There's something that explains this
2:46
that we call more of x
2:49
paradox and that is what's easy
2:51
for robots. Like being able to
2:53
pick up a large object large
2:55
heavy objects is hard for humans.
2:58
But what's easy for humans like
3:00
being able to pick up some
3:03
blocks and stab them were. turns
3:05
out that is very hard for
3:08
robots. And this
3:10
is a persistent problem. so the
3:12
ability to grasp arbitrary objects is
3:14
a grand challenge for my field.
3:16
Now by the way, I was
3:18
a very klutzy kid. I I
3:20
was the I would drop things
3:22
every time someone would throw me
3:24
a ball. I would drop it.
3:26
I was the last kid to
3:28
get picked on a basketball team.
3:30
I'm I'm still pretty classy actually,
3:32
but I have spent my entire
3:35
career studying how to make robots
3:37
less clumsy. Now
3:39
let's start with the hardware. So
3:41
the hand. It's a lot like
3:43
our hand and it has a
3:45
lot of motors, a lot of
3:47
tendons and cables, so it's unfortunately
3:49
not very reliable. It's also very
3:51
heavy and very expensive, so I'm
3:53
in favor of very simple hands.
3:56
So this has just two fingers
3:58
known as a parallel job. Or
4:00
so. It's very simple. it's
4:03
lightweight and reliable, and is
4:05
very inexpensive. No, actually,
4:07
it industry the Zebra super robot
4:10
gripper and that's the suction cup
4:12
and that only makes a single
4:14
point of contact. So again, simplicity
4:17
is very helpful in our field.
4:19
Scout let's talk about the software
4:21
and this is where it gets
4:23
really, really difficult because of a
4:26
fundamental issue which is uncertainty. There's
4:28
uncertainty and the control. There's uncertainty
4:30
and the perception. And
4:32
there's uncertainty in the physics. How would
4:35
I mean by the control? Well, if
4:37
you look at a robot's gripper trying
4:39
to do something, it's big, There's a
4:42
lot of uncertainty. the cables and the
4:44
mechanisms that cause very small errors and
4:46
these can accumulate and make it very
4:49
difficult to manipulate things. Snow.
4:51
In terms of the censors, yes,
4:54
robots have very high resolution cameras
4:56
just like we do and that
4:58
allows them to take images of
5:00
scenes in traffic or retirement center
5:02
or in a warehouse or in
5:04
are operating room. but these don't
5:06
give you the three dimensional structure
5:08
of what's going on. So
5:11
recently that with the new development called
5:13
My Dark and this is a new
5:15
class of cameras that use light beams
5:17
to build up of three dimensional model
5:20
of the environment. And
5:22
these are fairly effective if really
5:24
word a breakthrough in our field.
5:26
but they're not perfect so if
5:28
the objects have anything that signee
5:30
or transparent with any like acts
5:32
in unpredictable ways and ends up
5:35
with noise and holes in the
5:37
images. So these aren't really the
5:39
silver bullet and there's one other
5:41
form of censor out there now
5:43
called a Tactile Senses and these
5:45
are very interesting. They use cameras
5:47
actually image best surfaces as a
5:49
would be contact but these are
5:51
still. In their infancy. Now
5:54
the last issue is the physics. We take
5:56
a bottle on a table and we just
5:58
pushes and a robot. Putting in exactly
6:00
the same way each time, but a
6:03
bottle ends up in a very different
6:05
place each time. And why is that?
6:07
well is because it depends on the
6:09
in microscopic surface typography underneath the bottle
6:11
as it slid. For example, if you
6:14
put a grain of sand under their,
6:16
it would react very differently than if
6:18
there weren't a grain of sand and
6:20
we can't see if there's a grain
6:22
of sand because it's under the bottle.
6:25
It turns out that we can predict
6:27
the most heard of an asteroid a.
6:30
Million miles away, Far
6:32
better than we can predict the motion
6:34
of an object as it's been grass
6:36
by robots. Now
6:39
let me give you an example. Put
6:41
yourself here into the position of being
6:43
a robot. Are trying
6:46
to clear the table and your sensors
6:48
are noisy an imprecise. Your apps, waiters,
6:50
your tables and motors are are uncertain
6:52
see, can't fully control your own ripper
6:55
and there's uncertainty in the physics so
6:57
you really don't know what's gonna happen.
6:59
So it's not surprising that robots are
7:02
still very clumsy. Smothers
7:04
one sweet spot for robots and that
7:06
has to do with ecommerce. And
7:09
this has been growing as a huge trend
7:11
and during the pandemic it really jumped up.
7:13
I think most of us can relate to
7:16
that. We started ordering
7:18
things like never before had. This
7:20
trend is continuing and that Salinger
7:22
is to meet the demands, we
7:24
have to be able to get
7:26
all these packages delivered it of
7:28
timely manner. And the
7:31
challenges that every package is different.
7:33
every orders difference so you might
7:35
order some. Some nail polish
7:37
and an electric screwdriver. And those two
7:39
objects are going to be somewhere inside
7:42
one of these giant warehouses. And what
7:44
needs to be done is somewhat us
7:46
to go and find the mail past
7:49
and then God find the screwdriver, bring
7:51
up together, put him into a box
7:53
and deliver them to you. So this
7:55
is extremely difficult, it requires grasping. So
7:58
today this is almost entirely. With
8:00
humans in a human so like doing
8:02
this work. there's a huge amount of
8:05
turnover so it's a challenge and people
8:07
have tried to put robots into warehouses
8:09
to do this work has ensure that
8:11
all that well but some my students
8:14
it i. About five years ago we
8:16
came up with a method using advances
8:18
in a deep learning to have a
8:20
robot essentially train itself to be able
8:23
to grasp objects and the idea was
8:25
at the robot would do this and
8:27
simulation. It was almost as if the
8:29
robot. Would dreaming about how to grasp
8:32
things and learning how to grasp them
8:34
reliably. This is a system called Decks
8:36
Nets that is able to reliably pick
8:38
up objects that we put into these
8:40
been in front of the robot. These
8:42
are objects it's never been trained on.
8:45
And it's able to pick these
8:47
objects up and reliably clear these
8:49
pins over and over again. So
8:51
we're very excited about this results
8:53
and the students and I went
8:55
out to form a company and
8:57
we now have a company called
8:59
Mb Robotics and what we do
9:01
is make machines that use the
9:03
algorithms a software we developed at
9:05
Berkeley to pick up packages and
9:07
this is for ecommerce. The packages
9:09
arrive at large pins all different
9:11
shapes and sizes and they have
9:13
to be picked up scanned. And input
9:16
into smaller beds depending on the zipper. We
9:19
now have eighty of these machines
9:21
operating across the United States. sorting.
9:24
Over a million packages a week. Now
9:27
that's that's some progress. But it's
9:29
not exactly the com robot that
9:31
we're all been waiting for. So
9:33
I wanna give you a little
9:35
bit of idea of some of
9:37
the new research that we're doing
9:39
to try to be able to
9:41
have robots more capable in homes.
9:43
And one particular challenges being able
9:46
to manipulate the former baathists like
9:48
strings in one dimension, two dimensional
9:50
seats in three dimensions. like like
9:52
like fruits and vegetables. So we've
9:54
been working on a project is.
9:56
To untangled knots and what we do is
9:58
you take a table. We put that.
10:01
In front of the robot, it has to
10:04
use a camera to look down, analyze the
10:06
cable, figure out where to grasp it, and
10:08
how to pull it apart to be able
10:10
to untangle it. And this is a very
10:12
hard problem because the table as much longer
10:14
than the reach of the robot so has
10:16
to go through and manipulate. managed to slack
10:18
as it's working and I would say this
10:20
is doing pretty well. It's gotten up to
10:22
about eighty percent success when we give it
10:24
a tangled cable at be able to untangle.
10:28
The other one is something I think
10:30
we also or waiting for robot to
10:33
for the laundries. Now roboticist have actually
10:35
been looking at this for a long
10:37
time and there was some research that
10:39
has be done on this but the
10:42
problem is that it's very very slow
10:44
so this is about three to six
10:46
folds for our ogre s so are
10:49
we decided to free to revisit this
10:51
problem and try to to have a
10:53
robot work very fast. So what are
10:55
the things we did was try to.
10:58
Speak about a to our draw by the
11:00
could swing that fabric the way we do
11:02
with were folding and then we also use
11:04
friction in this case to drag the fabric
11:07
to have to smooth out some wrinkles And
11:09
then we bars Trust which is known as
11:11
the two seconds holes. You might have heard
11:13
of this. It's amazing because the robot is
11:15
doing exactly the same thing and that's a
11:18
little bit longer so we're making some progress.
11:20
Him. And last
11:22
example, his baggage. So you all it's counter
11:24
this all the time you go to a
11:26
corner store you have to put something in
11:29
a bag. Now it's easy again for humans.
11:31
but it's actually very very tricky for robots
11:33
because for humans you know how to take
11:35
the bag and how to manipulate it. But
11:37
robots a bag can arrive in many different
11:40
configurations and for hard to tell what's going
11:42
on. And for the
11:44
robots figure out how to open up
11:46
that bag. So what we did was
11:48
we had a robot train itself by
11:50
we painted one of these bags with
11:52
fluorescent paint and we had fluorescent lights
11:54
and with turn on and off in
11:56
the robot would essentially teach itself how
11:58
to manipulate these bags. And
12:00
so we've got it. now. up to
12:02
the point where we're able to solve
12:05
this problem about half the time so
12:07
it works, but of say it's still
12:09
we're still not. We're still not quite
12:11
there yet so I wanna come back
12:13
to more Bucks. Paradox with easy for
12:15
Robots is hard for humans and was
12:17
easy for us is still hard for
12:20
robots. We have
12:22
incredible capabilities, were very good
12:24
at manipulation. By
12:27
robots still are not there.
12:29
I want to say I
12:31
understand. It's been six
12:33
years and we're still waiting for
12:36
the robots that the Jetsons had.
12:38
Why is this difficult? We need
12:40
robots because we want them to
12:42
be able to do tasks that
12:45
we can't sue or we don't
12:47
really want to do. But
12:49
I wanted to keep in mind that these
12:52
robots they're coming. Just. Be
12:54
patient because we want the
12:56
robots. but robots Awesome! The
12:58
Us to do the many
13:01
things that robots still can't
13:03
to. Say to.
13:10
The show is back you by Schwab.
13:12
With Schwab Testing Games it's easy to
13:14
invest in ideas you believe in like
13:16
electric vehicles, renewable energy, water, sustainability and
13:19
more. Choose from over forty seems by
13:21
as is or customize the stocks in
13:23
a theme to sit your goals. learn
13:25
more at Schwab that com/be matic investing.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More