00:00:06
This is kinda wild.
00:00:07
"Will a robot take my job?"
00:00:10
is one of the most
googled questions.
00:00:12
It kinda makes sense.
00:00:14
You know, when phones
were first invented,
00:00:16
there was probably
horse messengers freakin' out
00:00:19
all over the joint.
00:00:20
Vinyl's making a comeback,
00:00:22
but fidelity aside, why bother,
00:00:24
when you can have
a curated playlist
00:00:26
with thousands of songs
00:00:27
right on your phone?
00:00:30
Come to think of it,
00:00:31
I can't tell you
the last time I shot a film...
00:00:33
on film.
00:00:35
See, change makes people panic,
but things evolve regardless,
00:00:39
so we accept it, we move on.
00:00:41
And now that we're entering
00:00:42
the so-called fourth
industrial revolution,
00:00:44
yeah, automation's
gonna replace some jobs,
00:00:46
but it will also
create new ones,
00:00:49
and probably give rise
to industries
00:00:51
that never existed.
00:00:52
That's the topic some of these
next folks are exploring,
00:00:55
how machine learning
00:00:56
is making things
more productive,
00:00:58
more efficient, safer,
and cleaner.
00:01:01
[director] And cut!
00:01:03
That felt okay.
You wanna go again?
00:01:05
-No, we're set.
-Okay.
00:01:06
[director] Maybe just
step back on that mark.
00:01:10
Here?
00:01:10
-Yeah.
-Okay.
00:01:13
[director] That's good.
00:01:14
[sighs wearily]
00:01:16
Right, yeah, I think I get it.
00:01:23
[horn honks]
00:01:27
[Downey]
The trucking industry
00:01:28
is at a crossroads.
00:01:31
E-commerce is driving
demand for shipping
00:01:33
higher than ever,
00:01:36
but the harsh realities
of long-haul trucking
00:01:38
are sending
professional drivers
looking for other work,
00:01:42
leaving a 50,000-driver
shortage in the U.S. alone.
00:01:46
[Dr. Ayanna Howard] There are
of course some industries
00:01:47
that will be impacted by A.I.
00:01:49
in terms of labor workforce
being diminished,
00:01:53
but there are some industries
00:01:55
that don't have
the workforce to sustain it,
00:01:58
and so that disconnect,
00:02:00
that's where A.I. can fit in
very nicely.
00:02:07
[Fitzgerald] My name
is Maureen Fitzgerald.
00:02:09
I've been driving trucks
for 32 years,
00:02:11
and I was ready
to make a change.
00:02:13
Three, two, one.
00:02:15
-[device beeps]
-[man] Engaged.
00:02:17
[automated voice]
Autonomous driving started.
00:02:27
[Downey]
Approaching retirement,
00:02:29
Maureen saw automation coming,
00:02:31
and assumed
it would replace her.
00:02:32
She was partly right.
00:02:34
Automation is coming,
00:02:36
but her job won't be lost.
00:02:38
It will evolve.
00:02:41
I've been in this career
a long time,
00:02:43
I've seen how everything
has been changing.
00:02:46
The monotony of coast to coast,
00:02:48
and the pay being low,
00:02:50
and, you know,
you're sleeping in a sleeper,
00:02:51
you're sleeping
at a truck stop.
00:02:53
It's hard. It's a hard life.
00:02:55
I saw an ad for a test driver,
00:02:58
and I said, "That has to be
the perfect job."
00:03:00
[alerts beep]
00:03:01
[brakes screech]
00:03:03
[Downey] Okay, but this isn't
just about Maureen.
00:03:06
We've all heard about
self-driving cars,
00:03:09
but self-driving trucks,
00:03:10
why are they setting the pace
00:03:11
of the autonomous
driving industry?
00:03:14
[Pedro Domingos] We already
have robotic airplanes.
00:03:16
Most airliners fly themselves,
00:03:18
but self-driving cars
are a harder problem
00:03:20
because the roads have
a lot more things going on
00:03:23
than the air does,
00:03:24
and driving on the freeway
00:03:26
is much easier
than driving in the city.
00:03:28
So we will see fleets
of automated trucks, I think,
00:03:31
long before we see
self-driving cars in the city.
00:03:42
The company of TuSimple
starts with a dream...
00:03:47
a dream that we wanted
to build the A.I. technology
00:03:50
for autonomous trucks.
00:03:51
[engine starts]
00:03:53
[Arda Kurt] Can you
give me one full rotation
to the right?
00:03:59
The truck was originally
not made to be automated.
00:04:02
Okay, can you give me
50% throttle?
00:04:06
We add our intelligence,
the servers,
00:04:09
and all the sensors
to make this work.
00:04:11
We're in the sleeper
of a TuSimple tractor.
00:04:14
We have removed what normally
would be beds
00:04:16
and refrigerators,
00:04:18
and we've added our computer,
which is the brains of the A.I.
00:04:25
Driving the car
is what is called
00:04:27
an "A.I. complete problem."
00:04:28
If you solve it,
00:04:29
you solve every other
problem in A.I.
00:04:31
However, to solve it,
00:04:33
you need to solve
every problem in A.I.
00:04:35
So it requires vision,
00:04:37
it requires robotic control,
00:04:38
so it requires
motion and navigation,
00:04:40
but it also requires
social interaction
00:04:42
with the other drivers,
00:04:43
so at the end of the day,
00:04:44
in order to drive a car well
in the city,
00:04:46
you need everything.
00:04:47
It requires an enormous amount
of common sense.
00:04:50
[Downey] Common sense.
00:04:52
That's what Maureen
is doing here...
00:04:55
overseeing the A.I. truck
00:04:57
as it learns to navigate
real-world conditions,
00:05:00
ready to take over
if anything goes wrong.
00:05:04
Today, they're taking
a ten-mile loop
00:05:06
to test scenarios that
a human driver might encounter.
00:05:09
Can this 10-ton vehicle
00:05:11
safely navigate
a busy freeway?
00:05:13
Can it merge and change lanes?
00:05:18
[man] All right, ready.
00:05:20
Three, two, one.
00:05:22
Engaged.
00:05:23
[automated voice]
Autonomous driving started.
00:05:28
My name is David Ruggiero,
00:05:28
and I'm a test engineer
with TuSimple.
00:05:31
My job is to let her know
what's happening,
00:05:35
let her be
absolutely comfortable
00:05:38
the truck is doing
what it's supposed to do.
00:05:41
Gear change...
There we are.
00:05:43
When the vehicle is engaged,
00:05:45
she is fully aware
00:05:47
of what's going on
around the truck,
00:05:49
but she doesn't
have to touch anything.
00:05:53
[Ruggiero]
All right, clear on the right.
00:05:54
[Fitzgerald]
Not clear on the left.
00:05:55
[Ruggiero]
I see the left approaching.
00:06:00
Never is one drive
or one day ever the same.
00:06:03
[Fitzgerald]
The left is now clear.
00:06:06
Is it gonna go?
00:06:11
-We're good!
-Come on.
00:06:13
[Fitzgerald]
The truck was cautious...
00:06:15
So I'm still seeing
full braking.
00:06:16
[Fitzgerald] It was probably
overly cautious.
00:06:19
Just go manual?
00:06:21
Yeah, we can do that.
00:06:22
[automated voice]
Autonomous driving off.
00:06:25
[Ruggiero] There we go.
00:06:32
[Fitzgerald] All righty.
00:06:33
-[device beeps]
-Engaged.
00:06:34
[automated voice]
Autonomous driving started.
00:06:40
Going 45.
00:06:41
Looks good.
00:06:43
Center lane.
00:06:44
[Fitzgerald] I feel like
00:06:45
I have a relationship
with the truck.
00:06:47
Easy, easy...
00:06:49
I talk to the truck.
00:06:50
I tell it when it does good.
00:06:51
I've never been
a test driver before,
00:06:54
and it's never been a truck
driving by itself before.
00:06:56
We're both actually
learning together.
00:06:59
Still green.
00:07:00
[Ruggiero] Still green...
00:07:01
We're off throttle,
slight brake.
00:07:04
Okay, red light.
00:07:05
[automated voice] Left turn.
00:07:06
[Fitzgerald]
Turn signal is on.
00:07:07
[Ruggiero] I see the lead van.
00:07:11
Human drivers cause 30,000
or more deaths per year.
00:07:14
[tires squealing]
00:07:17
Long-haul trucking
can be dangerous,
00:07:19
it can be boring,
00:07:20
and therefore,
it's ripe for automation.
00:07:23
[Fitzgerald] On the right.
00:07:24
[Ruggiero] Okay, I see
the other car moving,
00:07:25
I do see green.
00:07:28
Still green...
00:07:29
Still green.
00:07:30
[Ruggiero] I see no oncoming.
00:07:32
We are making the turn.
00:07:34
[Fitzgerald] That'll work.
00:07:36
All right, the new merge ramp.
00:07:38
[Downey]
Merging onto a highway
00:07:40
is difficult and dangerous.
00:07:42
It requires mastering
00:07:43
a complicated set of physical
and mental skills,
00:07:45
having keen sensory
and spatial awareness,
00:07:48
preparing for unpredictability
and human fallibility...
00:07:51
[Ruggiero] And we're in it.
00:07:52
[Downey] ...and doing it all
fast and fresh each time,
00:07:55
otherwise you risk crashing,
00:07:57
or worse.
00:08:00
Turn signal is on.
00:08:01
[Ruggiero] Okay,
I'm showing a left intention.
00:08:02
[Fitzgerald] We are gonna have
a truck next to us,
00:08:03
but he's moving over.
00:08:04
Okay.
00:08:06
Confirmed.
00:08:07
[Fitzgerald] Looks like we
are clear to get over,
00:08:09
the pickup is moving.
00:08:10
[Ruggiero]
This truck is measuring
00:08:12
what's coming up behind it,
00:08:13
how fast they're
coming up behind...
00:08:15
Looks a little
congested up there.
00:08:17
Yep.
00:08:18
[Ruggiero] It has
a full 360 view, all the time.
00:08:24
[Price] The A.I.
uses the images
00:08:27
that are coming
from the cameras
00:08:28
with other sensors,
00:08:30
LiDAR, and radar.
00:08:32
LiDAR is a laser range finder
00:08:34
that is measuring
the distance to objects
00:08:37
360 degrees around.
00:08:39
It gives us a three-dimensional
picture of the world.
00:08:43
In the virtual world,
00:08:44
there is a direct
correspondence of the map
00:08:46
and the location
where the vehicle is,
00:08:49
and the behaviors
of other vehicles,
00:08:50
their locations, their speed,
00:08:52
their future intentions,
00:08:53
and once all these things
are re-created,
00:08:56
it's basically asking
the computer to play a game.
00:08:59
Okay, tire front,
continuing at 57.
00:09:02
Good job.
00:09:04
[Ruggiero]
When I'm looking at the data
00:09:05
that the truck is feeding me,
00:09:07
not only am I seeing
the current environment,
00:09:09
but I'm seeing ahead
into the future.
00:09:12
And we have a slow vehicle
cutting us off.
00:09:14
I see the cut in, at a 51.
00:09:16
[Fitzgerald] That was
a good one.
00:09:17
[Ruggiero] Okay,
I saw the cut in.
00:09:19
If we need to make
a lane change,
00:09:22
the truck is going to know
before I do.
00:09:23
Okay, I currently
don't have a--
00:09:25
There we go,
I have an intention.
00:09:26
We're going right now.
00:09:27
[automated voice]
Left change.
00:09:32
[Fitzgerald] That's fine.
We're good.
00:09:35
All right, let's go.
00:09:36
[Downey] Okay, so this is
where things get interesting.
00:09:39
This truck's tricked out
with sophisticated cameras
00:09:42
and software,
00:09:43
but it still doesn't have
one critical element...
00:09:47
the thing that A.I.
may never have.
00:09:51
The central problem in A.I.
00:09:53
is that human beings
have common sense
00:09:55
and computers do not.
00:09:56
We take common sense
for granted.
00:09:58
We know how the world works,
00:10:00
and everything that we do
in our daily life
00:10:03
involves common sense...
00:10:05
Pretty steady stream now
on our left.
00:10:06
[Ruggiero] I see it.
00:10:08
...and people have been trying
to imbue A.I. with common sense
00:10:10
since the beginning,
00:10:11
and what is extraordinary
00:10:13
is, at the end of the day,
00:10:14
how little progress we've made.
00:10:17
[Downey] The endgame
is that, one day,
00:10:20
the AI truck will not need
Maureen, David,
00:10:22
or any person to ride with it.
00:10:24
Each one of these test runs
00:10:26
is a building block
toward that goal,
00:10:28
designed to help it learn
how humans operate
00:10:31
and make decisions,
00:10:32
even when you get thrown
a curveball.
00:10:35
[Fitzgerald] This truck
is messing with us.
00:10:36
[Ruggiero]
Just catching the vehicle.
00:10:39
He's all over the...
00:10:41
-He's coming close to the line.
-Mmm-hmm.
00:10:45
[Ruggiero]
I have a left intention.
00:10:46
Uh, lane is blocked.
00:10:48
We were having
a little bit of challenges
00:10:50
from one of your camera trucks.
00:10:52
[Ruggiero] All I need is a gap.
00:10:54
Just looking for the gap.
00:10:55
Come on, you can pass him,
it's okay.
00:10:59
[Fitzgerald]
The truck was cautious,
00:11:00
because it couldn't predict
00:11:02
what that camera truck
was doing.
00:11:05
Why is he going so slow?
00:11:09
[Downey]
The film crew's vehicle
00:11:10
was shooting the truck
for the whole afternoon,
00:11:12
and, yeah,
it wasn't driving normally.
00:11:15
The camera truck was
what we would consider
00:11:19
a non-compliant vehicle.
00:11:21
It was doing things
that a normal driver,
00:11:23
you wouldn't expect them to do.
00:11:26
[Fitzgerald] Doesn't trust it.
00:11:30
[Ruggiero]
Okay, I see the cut in.
00:11:31
Sixty and slowing.
00:11:34
Fifty-two and slowing.
00:11:37
Still seeing
a left intention.
00:11:39
[Fitzgerald] Let's see.
Okay, we are no longer clear.
00:11:41
No, it's full.
00:11:42
[automated voice]
Pre-left change.
00:11:44
We're not going.
00:11:48
He's all over the place,
that guy.
00:11:52
[Downey] The truck wants
to pass the camera crew,
00:11:54
but the constantly
changing speeds
00:11:55
and unusual behavior
00:11:56
confuses the AI...
00:11:59
so it refuses
to take the risk.
00:12:02
[automated voice]
Canceled.
00:12:07
There are behaviors
that the truck has to learn
00:12:10
to become human, in some sense,
00:12:13
because not everybody else
is operating at the same level.
00:12:16
All right.
00:12:19
Turns out that getting
the first basic systems
00:12:22
was a big milestone,
00:12:24
but then there's this long tail
of difficult situations
00:12:28
that are much harder to solve.
00:12:30
[horn blaring]
00:12:33
When you're merging,
00:12:35
do you have to give way,
00:12:36
or does it make sense
to go forward?
00:12:38
That requires
a level of understanding
00:12:41
that we humans barely have,
00:12:43
but machines aren't there yet.
00:12:46
[Ruggiero]
That's definitely gotta be
00:12:47
the most interaction I've seen
with one vehicle.
00:12:49
[Fitzgerald chuckles]
00:12:51
[Ruggiero] The big takeaway
on today's run...
00:12:54
there is no such thing
as a perfect run.
00:12:58
Eventually,
00:13:00
we hope that the trucks react
to a change in environment
00:13:02
as quickly as we can,
00:13:04
regardless of
what we throw at it.
00:13:09
[Fitzgerald]
Okay, slow down for the gate,
00:13:11
and looks like
they're going to let us turn.
00:13:12
[Ruggiero] Beautiful.
00:13:13
[Fitzgerald] Thank you, people.
00:13:15
We still have
a lot of work to do
00:13:18
to validate the system,
00:13:20
assure that everything
is perfected,
00:13:22
and this will take
substantial driving experience,
00:13:25
simulation experience.
00:13:27
[Fitzgerald] Watching
something develop over time,
00:13:30
you never thought
it would learn like this,
00:13:31
and now it is learning,
00:13:34
and you feel like
a proud parent.
00:13:35
It was making
some awesome decisions
00:13:37
that I've never seen it
make before.
00:13:40
I get asked
all the time by people,
00:13:42
you know, "Oh, you're taking
truck drivers' jobs away
00:13:45
by having autonomous trucks,"
00:13:46
and I don't believe
that's true,
00:13:49
because it's not
taking away jobs,
00:13:52
it's taking the place
of a job
00:13:54
that a driver
doesn't want to do.
00:13:55
[automated voice]
Thanks for riding
with TuSimple.
00:13:57
Ta-da!
00:13:58
High five.
00:14:00
[automated voice]
Autonomous driving off.
00:14:01
[Fitzgerald chuckling]
00:14:03
[Downey] What Maureen thought
00:14:04
was the twilight
of her working life
00:14:06
actually became the dawn
of something new,
00:14:09
and it's an age-old phenomenon.
00:14:13
Tractors didn't make
farming obsolete,
00:14:16
and video didn't kill
the radio star.
00:14:19
It just provided
new opportunities.
00:14:25
New technology is changing
how we think about work.
00:14:29
[man] Hey, Garrett,
that east on 424
should be good.
00:14:32
10-4.
00:14:35
[Howard] I think the biggest
changes that we face
00:14:38
when we think about A.I.
and the future of A.I.
00:14:40
is what do we do
00:14:42
when A.I. changes
what the job functions are,
00:14:46
and what the workforce
looks like,
00:14:47
which it's going to do.
00:14:48
So this is a fear,
00:14:50
because change,
people don't like,
00:14:52
and change that
you don't understand...
00:14:57
is terrifying
for a lot of people.
00:15:04
[Anthony Otto]
Port of Long Beach
00:15:05
is the second-busiest port
in North America...
00:15:11
and the first fully-automated
container terminal
00:15:14
in the United States.
00:15:17
[Downey] To most people,
00:15:18
shipping and logistics
aren't that sexy,
00:15:20
but to an A.I., it's a dream.
00:15:23
With over 14,000 people
00:15:25
moving more than eight million
giant containers
00:15:27
around the world,
00:15:28
this port handles about
$200 billion worth of cargo
00:15:32
a year.
00:15:33
Automating this industry
00:15:34
will make it
more efficient, for sure,
00:15:36
but also safer.
00:15:39
[Otto] As you can see
this fence here,
00:15:40
from this point on,
00:15:42
all the way for about
an eighth of a mile that way
00:15:43
is a fenced-off no-man zone,
00:15:46
running entirely
off of software.
00:15:48
In the old design,
00:15:50
you had the vessel's activity
00:15:52
mixing with the truck activity,
00:15:53
competing for space,
congestion, safety issues...
00:15:59
...but the vessel sizes
just continued to grow,
00:16:01
seven times larger
00:16:02
than the ships that we ever had
in the beginning,
00:16:05
so there was a need
for a new model,
00:16:06
which is what LBCT is.
00:16:14
[Josh Johnson]
When a ship pulls alongside,
00:16:16
we're using
different algorithms
00:16:17
to figure out traffic,
scheduling,
00:16:19
dispatching, and planning.
00:16:21
We're doing this
00:16:22
with 10 different cranes
at the same time,
00:16:23
50 different vehicles
at the same time,
00:16:26
and we have
30,000 different places
00:16:28
to put that container
in the yard.
00:16:31
When we offload
a container from a ship,
00:16:33
the first thing it's gonna do
00:16:34
is get picked up
by a ship-to-shore crane
00:16:36
that does have
an operator inside.
00:16:38
-Hey, Lily, it's Boyle.
-[Lily] Yeah, I copy.
00:16:40
[Boyle] Okay, Lily,
which side you wanna start on?
00:16:43
[Johnson]
Ship-to-shore cranes
00:16:45
will set the containers
on the platform.
00:16:47
On that platform,
there are over 20 cameras
00:16:50
that use
Optical Character Recognition,
00:16:51
or OCR.
00:16:55
Another crane will then
pick it up,
00:16:57
put it onto
a fully-automated vehicle.
00:17:00
That vehicle will drive
through the yard,
00:17:01
and it'll get to
the right spot,
00:17:03
planned by the system.
00:17:04
What looks like
a little wad of bubble gum,
00:17:08
these are transponders
buried in the ground.
00:17:10
They're in a grid pattern,
00:17:11
there's more than 10,000
out on the production field.
00:17:15
There's an antenna
on the front, and the rear,
00:17:17
and they read them
as they drive,
00:17:19
and then transmit that
to the system
00:17:21
to let it know where it is.
00:17:25
Once it reaches the yard,
00:17:26
an automated stacking crane
will pick it up
00:17:29
and deposit it
into a yard block.
00:17:31
From the block,
00:17:32
it'll eventually
get deposited onto a truck
00:17:34
or onto a bomb cart,
00:17:36
where it can make its way
onto a train.
00:17:42
I don't have my key.
00:17:44
[man] On my way.
00:17:45
That's why we work in teams.
00:17:47
-[man] What'd you say?
-[laughs]
00:17:51
[Downey] They say
that we can eventually learn
00:17:53
to adapt to an environment
00:17:55
that mixes robotic
and human elements,
00:17:59
but it's more difficult
for machines to adapt to us,
00:18:02
so here,
they need to be separated,
00:18:04
because of safety.
00:18:07
Because humans
are unpredictable,
00:18:10
automation can be
physically dangerous to people,
00:18:14
because if I'm
a 200-ton machine
00:18:17
and I don't see you?
00:18:21
I might just run you over.
00:18:23
[winches clanking]
00:18:25
[man] This is
the high-danger area.
00:18:28
When we go out in the field,
we call for a restriction,
00:18:31
and they block a certain area
of the production field,
00:18:33
and no AGVs know
to go in that area,
00:18:35
and we can safely work.
00:18:36
You want these two right here?
00:18:37
-[man] Yeah...
-Okay, cool.
00:18:39
This facility,
it is fully automated,
00:18:42
but it's not as if someone
comes in, pushes a button,
00:18:44
and puts their feet up
on their desk.
00:18:46
It is still employing
hundreds of individuals
00:18:50
on any given day...
00:18:51
How does everything look
on that side, Eric?
00:18:53
[Eric] Got a broken rail
over here, Julio.
00:18:54
We're gonna have
to weld it up real quick.
00:18:55
Okay.
00:18:57
Grab my hood from
inside the truck, please?
00:18:59
Here we go, guys.
Watch your eyes.
00:19:01
Eyeballs!
00:19:07
You know, this is
an asset-intensive facility.
00:19:09
The number of mechanics
necessary
00:19:11
to maintain this facility
00:19:12
has certainly grown.
00:19:14
[Julio] My background
is automation.
00:19:16
That's why I decided
to come here,
00:19:17
because it was high-tech.
00:19:20
[Johnson] Instead of
a yard crane operator
00:19:22
in a cab alone all day,
00:19:24
now we're working remotely
in the operations room.
00:19:26
About 80% of the time,
00:19:28
most of the faults
are resettable from here.
00:19:31
[woman] Okay, easy fix.
00:19:32
Easy fix.
00:19:34
I'm Garrett Garedo,
crane mechanic here at LBCT.
00:19:36
This is the central
operations system.
00:19:39
A bypass key
will allow the operator
00:19:42
to bypass a function.
00:19:44
[Downey] Automation
can make the workplace safer,
00:19:47
and this is a good example...
00:19:48
taking someone out of the yard,
00:19:50
and putting them
into the control room.
00:19:53
Right here, I'm gonna need
to move this container over,
00:19:56
because it's not
sitting too good.
00:19:58
You ready?
-Yep, go ahead.
00:20:00
[Howard] A.I.,
00:20:02
like any technology
that comes into our society,
00:20:04
it changes the workforce.
00:20:06
There's this misconception
00:20:08
that there will be fewer
of these new jobs.
00:20:10
That's actually not the case.
00:20:11
Let me move it back.
00:20:13
[Howard] There are new jobs,
00:20:14
there are different
types of jobs.
00:20:17
[man] I was a crane operator.
00:20:18
Back in the day,
00:20:19
we'd have to climb up
five stories into the cab.
00:20:23
It was a little hard,
especially on the back,
00:20:25
and then
when automation started,
00:20:27
they just shifted me over,
00:20:29
and now it's a lot easier.
00:20:30
[laughs]
00:20:34
If you don't embrace it,
you're just gonna fight it,
00:20:37
and it's not gonna...
00:20:38
You gotta use it as a tool.
00:20:40
Right on the money.
00:20:42
[Otto]
You know, our obligation
00:20:43
was to re-train
those same individuals
00:20:44
into the new jobs that
that technology has created.
00:20:50
We are the progress
that was needed
00:20:53
for the goods movement industry.
00:20:57
For the foreseeable future,
00:20:58
there will continue to be
a lot of things
00:21:00
that only humans can really do.
00:21:02
So I've dispatched
a mechanic out to STS-1.
00:21:05
Okay.
00:21:07
[Domingos] And so
the future of work
00:21:08
is not humans being replaced
by machines.
00:21:12
It's humans figuring out
how to do their job better
00:21:15
with the help of machines.
00:21:21
[Downey]
Structured environments
00:21:22
are good for automation.
00:21:24
Smooth surfaces, right angles,
less chance for chaos.
00:21:30
But what about more complex,
human environments?
00:21:34
Can we learn to work
with robots hand-in-hand?
00:21:44
[man] Oh, there we go.
00:21:47
Wait a second. Stop.
00:21:49
Let's see if that works.
00:21:51
[Downey] RoboHub is
a premiere robotics incubator
00:21:54
at the University of Waterloo
in Ontario, Canada.
00:21:57
Gripper's not working
on the left.
00:21:59
Maybe if I turn
a little bit...
00:22:01
[Downey]
One thing they're doing
00:22:02
is developing A.I. and robotics
00:22:03
to use in environments
that are unstructured,
00:22:06
more human.
00:22:07
Like at home.
00:22:09
Calm down, calm down.
00:22:11
[Brandon DeHart] TALOS
is one of the most advanced
humanoids on the planet.
00:22:14
It can walk, it can talk,
00:22:16
it can see you in 3D...
00:22:19
but it can't do
most of those things
00:22:21
out of the box,
00:22:21
so you really have
to take it as a tool
00:22:23
and teach it how to do
a lot of these things.
00:22:27
On TALOS, we want to explore
two different aspects of A.I.
00:22:31
First, to perceive objects
in the world...
00:22:34
[DeHart] So it has cameras
in its eyes,
00:22:36
and it also has a depth camera
00:22:38
where it can actually see
how far away things are
00:22:40
within its vision,
00:22:41
much like we do
with our depth perception,
00:22:43
and as soon as TALOS
has that ability to see,
00:22:46
"Oh, there's a human here,
there's an object back here,"
00:22:49
then it can start building
a map of its world,
00:22:51
so that it knows in 3D
what is all around it.
00:22:54
[Howard] Computer vision
00:22:56
is a way to mimic
how we see the world.
00:23:00
I can say, "Well, that's a TV,"
00:23:01
or "That's a cat,"
or "That's a dog,"
00:23:03
so what is it
that I'm looking at
00:23:05
when I'm looking at an image?
00:23:06
Computer vision is trying
to figure out
00:23:09
what are the objects
that are in this image
00:23:12
that represents
what we would understand.
00:23:14
And now I can teach the robot,
00:23:16
basically by grabbing it,
00:23:17
and I will just gently
move the robot around,
00:23:20
and guide it
towards the bottle...
00:23:24
and you just hit
the right trigger button
00:23:29
to execute the same motion.
00:23:34
Should I close it?
00:23:36
No, no, it will do that
automatically this time,
00:23:38
because I already kind of
integrated this.
00:23:41
[Werner]
Also we want to explore
00:23:44
application of A.I.
00:23:45
in understanding locomotion
00:23:47
in terms of balancing.
00:23:55
No, it's going to crash.
00:24:00
[laughs]
00:24:02
Played too much.
00:24:04
The biggest misconception
about robots
00:24:06
is that they are
more capable than they are,
00:24:08
that they're more generalized
than they are...
00:24:10
[clattering]
00:24:13
...and they're not quite
there yet.
00:24:16
[Werner] Okay, let's give it
one more shot.
00:24:19
[DeHart] So you may have seen
humanoid robots
00:24:20
that can do parkour,
00:24:22
do backflips.
00:24:25
The only ones
that are really comparable
00:24:26
are the Valkyrie
that was designed by NASA,
00:24:28
of which there's only
two or three in the world...
00:24:31
and the Atlas
from Boston Dynamics,
00:24:35
but most of them
have very few sensors,
00:24:36
and a lot of the time,
00:24:38
they will be
remote-controlled,
00:24:39
or given a very clean,
pre-scripted path,
00:24:43
and in a situation like that,
what they're doing
00:24:45
is pushing the limits
of the mechanical systems.
00:24:48
That's very similar
to what we're doing here,
00:24:49
except on our case, we're
pushing the software side.
00:24:52
[Brynjolfsson]
The next step going forward
00:24:54
is to start replacing
all of our dumb, blind,
00:24:57
dangerous machines
00:25:00
with machines that have sensors
and vision built into them,
00:25:03
that can work
side by side with humans
00:25:05
to be more productive
00:25:07
and safer at the same time.
00:25:09
[Werner] Maybe we can
carry a table
with the robot together?
00:25:12
So if we put the table
in its hands...
00:25:15
[DeHart] So for a task
00:25:16
where the robot might be
carrying something
with a human,
00:25:18
it's going to need to be able
00:25:20
to feel how hard the human
is pushing, pulling,
00:25:22
much like if you're
being guided in a dance.
00:25:26
Can we also
tighten the grippers, Alex?
00:25:27
Yeah, yeah.
00:25:29
So as humans,
we have a sense of touch,
00:25:31
and we have a sense of
how much force we're applying,
00:25:34
and is being applied to us.
00:25:35
With TALOS,
it doesn't have a skin,
00:25:37
it's hard plastic everywhere,
00:25:39
so it can only really sense
what's in its motors...
00:25:41
[Werner]
This will take a second.
00:25:42
We have to teach it
how to translate that
00:25:45
into sort of a sense of touch.
00:25:47
Try to keep it
somewhat level,
00:25:49
but if you kind of
twist the table as well?
00:25:52
[Werner] Let me show you
how much is actually possible.
00:25:58
The one who writes
the controller
00:25:59
is always braver
than the rest of us.
00:26:01
[woman laughs]
00:26:02
[Werner] And you just
start pulling a bit...
00:26:05
and then the robot
will walk with you.
00:26:09
TALOS is using
compliant control
00:26:11
in the upper body,
00:26:12
which basically
makes him soft.
00:26:14
The robot automatically senses
00:26:16
the forces being applied
to the table.
00:26:19
The human is still
fully in control.
00:26:21
This is very important
00:26:22
to the safety of humans
surrounding the robot.
00:26:27
Within robotics,
00:26:29
the areas where you're really
gonna see the important advances
00:26:32
are those environments
00:26:34
that are relatively
controlled and predictable,
00:26:36
so a good example
is an Amazon warehouse.
00:26:41
In those environments,
you already see lots of people,
00:26:45
and also lots of robots...
00:26:49
and they're working together.
00:26:53
[Brynjolfsson]
Once you have a process,
00:26:55
and you've reduced it
to an algorithm,
00:26:56
you can replicate it
00:26:58
so that if a robot can learn
a whole new skill,
00:27:00
you can copy that knowledge
through the cloud
00:27:03
to all the other robots,
00:27:05
and now they all have
that same skill.
00:27:07
It's a whole new world,
a whole new kind of economics,
00:27:10
and we're just beginning
to understand its implications.
00:27:13
[horn honking]
00:27:14
[Downey] A whole new
automated world
00:27:15
still hard to imagine?
00:27:18
Maybe it's more about
what we're gaining
00:27:20
than what we're losing.
00:27:22
Will automation
make things faster?
00:27:25
More efficient?
00:27:27
More productive?
00:27:29
Tastier?
00:27:37
[man] I've been making pizza
for 15 years.
00:27:41
In my life,
00:27:42
I've probably made
two million pizzas...
00:27:45
so coming to Zume was awesome,
00:27:47
because it was like, wow,
00:27:48
this is different
than any other pizzeria.
00:27:51
[Alex Garden]
So I got the idea for Zume
00:27:53
about seven years ago
00:27:55
when I realized
00:27:56
that Big Pizza was built
on one central conceit...
00:28:00
which was that there wasn't
a better way to do it.
00:28:02
I thought, "I wonder if there's
a better way to do it?"
00:28:04
[Downey] Big Pizza.
00:28:06
Doesn't have
the same ring to it
00:28:08
as Big Data, or Big Pharma,
00:28:10
but it's not just about pizza.
00:28:12
It's also about big waste.
00:28:17
Last year in the U.S.,
00:28:18
about $200 billion
worth of food was wasted.
00:28:21
$200 billion.
00:28:24
That's a lot down the drain,
00:28:26
and also up in the sky.
00:28:28
Waste is a huge contributor
to greenhouse gas emissions.
00:28:32
[thunderclap]
00:28:33
Perhaps half of all food
that's produced in the world
00:28:36
is wasted.
00:28:38
Artificial intelligence can
help us change the supply chains
00:28:40
so that they are
much more efficient.
00:28:44
If we can then
have an impact there,
00:28:45
that already
has huge, huge implications.
00:28:50
One of the classic tensions
in every food business
00:28:53
is producing too much,
and having waste,
00:28:55
or not producing enough food,
and having stock outs.
00:29:00
It takes years and years
of practice
00:29:01
for a human to forecast that,
00:29:03
but that's actually
a data problem
00:29:04
that machines,
and A.I. in particular,
00:29:05
are really, really good
at getting right.
00:29:09
So at its core,
00:29:10
Zume is a fundamentally
new logistics model.
00:29:16
We predict
what we're going to sell
00:29:17
before you even order it,
00:29:19
and that whole predictive layer
of the business
00:29:21
is all driven by A.I.
00:29:24
[Chris Satchell]
When Zume first called me,
00:29:26
I obviously thought,
00:29:27
"Why on Earth
does a pizza company
00:29:29
want a Chief
Technology Officer?
00:29:31
But then
when I heard the vision,
00:29:32
the fact that
00:29:34
we were going to be able
to change an industry
00:29:36
by focusing on this amazing
end-to-end platform,
00:29:38
which just happens
to also produce amazing pizza,
00:29:41
then it made sense.
00:29:42
[Downey] To disrupt Domino's,
00:29:44
Zume uses machine learning
00:29:46
to try and forecast
how much pizza people want,
00:29:49
what kind, and when.
00:29:51
It crunches dozens
of different variables...
00:29:53
location,
day of week, weather,
00:29:56
what's on TV,
past ordering trends,
00:29:59
and then predicts
how many Sgt. Pepperonis
00:30:01
San Francisco will want
on a Tuesday,
00:30:04
for example.
00:30:07
Our supply chains
are incredibly inefficient.
00:30:09
Our food processing systems
waste a lot of food.
00:30:16
If we could improve
predictions,
00:30:17
we could eliminate
most of that waste.
00:30:20
We'd know what the demand
was going to be,
00:30:22
where the products
were going to be,
00:30:24
and ultimately, that would
make us all better off.
00:30:28
Absolute perfection would be
no pizza gets wasted,
00:30:31
and anytime a customer looks,
00:30:33
there's always one of the type
of pizzas they would like
00:30:34
available.
00:30:35
So it's an optimization problem.
00:30:36
How close can you get?
00:30:37
[Downey] Optimization, right,
00:30:40
but what else?
00:30:41
"End to end"
is not just about prediction.
00:30:44
It also involves automation,
00:30:47
or...
00:30:48
cooking.
00:30:54
Pizza and robots,
very exciting concept.
00:30:59
Pepe and Giorgio,
they dispense the sauce.
00:31:03
We have Marta here,
00:31:04
who spreads the sauce
on the pizzas...
00:31:07
So we're still on
Veggie Zupremes, guys.
00:31:08
These are looking good.
00:31:09
[man] ...and down here
at the end,
00:31:11
we have Bruno,
00:31:13
who helps deliver the pizza
00:31:14
from the assembly line
to the ovens,
00:31:16
or around the bypass,
00:31:17
and in the back, Vincenzo,
00:31:19
who helps load the pizzas
onto the cartridges.
00:31:25
[Downey]
Based on its algorithm,
00:31:26
the A.I. predicts
how many pizzas, and what kind,
00:31:29
to load
onto the mobile kitchens.
00:31:33
When orders come in,
00:31:35
it also decides which
mobile kitchen will bake it,
00:31:37
and exactly when
to put each one in the oven.
00:31:40
Every single pizza
00:31:41
has its own cooking profile
and recipe.
00:31:43
So each one of these ovens
actually is a robot.
00:31:47
They're connected
to cloud services
00:31:48
that're always monitored,
00:31:50
making it possible
for one person working in here
00:31:52
to cook up to
120 pizzas an hour.
00:31:59
[Zaida Ybarra] It's humans
and technology
working together.
00:32:01
It's pretty awesome.
00:32:02
That and the pizza.
00:32:03
I love pizza.
00:32:05
Pizza is one of
my favorite foods.
00:32:08
We use A.I. in simulation
00:32:10
around how do we get
00:32:11
the right delivery
estimated time of arrival?
00:32:14
The A.I. can suggest
00:32:16
here's the best place
to put a truck
00:32:17
so that we get the hottest,
freshest product
00:32:19
to our customers.
00:32:23
So we have
a BBQ Chicken.
00:32:26
BBQ...
00:32:28
[Ybarra]
When we get the orders,
00:32:29
our system controls the ovens,
00:32:31
which controls the cook time,
00:32:33
and how long the pizzas
need to cook.
00:32:35
[Satchell] The A.I.
can change the workflow
00:32:37
on one of our vehicles
00:32:39
to make sure the pizza's
cooked at the last moment.
00:32:41
...and then the system chooses
which driver to send it to.
00:32:44
Here's your order...
00:32:46
A BBQ Chicken ranch.
00:32:50
It's algorithms plus,
it's A.I. plus humans,
00:32:54
not A.I. instead of humans.
00:32:58
Now we have a pineapple
that is ready to go.
00:33:02
[Ybarra] Their predictions
00:33:03
for the amount of pizzas
that we carry on the truck
00:33:04
are pretty spot on.
00:33:06
Last Zupreme of the night.
00:33:09
Sometimes we'll have
a bunch of pizzas left over
00:33:11
by the end of the night,
00:33:12
and we'll think the prediction
was probably wrong for today,
00:33:15
and then boom,
an order comes in,
00:33:17
and the next thing you know,
another order comes in,
00:33:19
like, eight pizzas at once.
00:33:22
All of that data gets fed back
into the learning algorithm,
00:33:25
so every week
00:33:26
we're trying to evolve
our learning algorithms
00:33:28
to do a better job
than the week before.
00:33:32
Part of what we're
optimizing for
00:33:33
is reducing waste.
00:33:35
Maybe if you can use
prediction
00:33:36
to go back
into the supply chain
00:33:38
and predict more accurately
what you need to grow,
00:33:40
and where you need
to get that product,
00:33:44
I think you could fundamentally
change things.
00:33:49
We're in the early stages
00:33:50
of a massive change
in technology
00:33:54
that's allowing machines
to do a lot of tasks
00:33:57
that previously
only humans could do.
00:33:59
[Downey] The future of work
is evolving.
00:34:02
Automation's making
the workplace safer,
00:34:05
greener,
00:34:06
more efficient, for sure.
00:34:08
If we can figure out how
to integrate A.I. technologies,
00:34:10
not to replace humans,
00:34:12
but to augment their abilities,
00:34:13
to make life more fun,
00:34:16
to make us more productive
and more creative...
00:34:20
I think that's where the power
of the machine will come.
00:34:22
[Downey]
Thousands of years ago,
00:34:24
we were hunter-gatherers.
00:34:26
Eventually, we became farmers,
00:34:28
and now that
we're maybe entering
00:34:30
this fourth
industrial revolution,
00:34:33
one that connects the physical,
biological, and digital,
00:34:38
it's not just jobs that are
being changed by A.I.,
00:34:41
but us.
00:34:42
We are the progress.
00:34:45
So why focus on the rear-view
00:34:47
when we can look
to the road ahead
00:34:49
to see what we're becoming?
00:34:59
[woman sighs]
00:35:00
[Ford] It's a real challenge
in any case
00:35:02
to build robots
00:35:04
with the kind of mobility
and dexterity
00:35:06
that begins to approach
what human beings have.
00:35:09
Kind of the cliche
that we all think of
00:35:11
is that
we'd like to have a robot
00:35:13
that can go to the refrigerator
and grab a beer for us.
00:35:16
I mean, if you think about
what's involved with that,
00:35:18
that's still
an enormous challenge.
00:35:19
A robot that's able to do that
00:35:21
has to be heavy enough
to go to a refrigerator.
00:35:25
It has to be able
to open the door,
00:35:27
and it has to have
the visual perception
00:35:28
and the dexterity to reach in
and grab the beer,
00:35:32
and already that implies
a fairly heavy machine,
00:35:34
otherwise
it would just tip over.
00:35:37
Building a robot to do that
is not just difficult,
00:35:39
but it's gonna be
fairly expensive,
00:35:40
so I think that
it's gonna be quite a while
00:35:43
before we really see
the kinds of robots
00:35:45
that we imagine
from the science-fiction world
00:35:48
operating in our daily lives.
00:35:51
[robot] Get ready
for a picture.
00:35:52
Smile, please.
00:35:54
[camera shutter clicking]
