0:00

The following is a conversation with Robert

0:02

Plater, CEO of Boston

0:04

Dynamics, a legendary robotics

0:07

company that over 30 years has

0:09

created some of the most

0:10

elegant, dexterous, and

0:13

simply amazing robots ever built,

0:15

including the humanoid robot Atlas

0:18

and the robot dog Spot, one

0:21

or both of whom you've probably

0:24

seen on the internet, either dancing,

0:26

doing back flips, opening doors, or

0:29

throwing around heavy objects.

0:32

Robert has led both the development

0:34

of Boston Dynamics humanoid robots and

0:37

their physics-based simulation software.

0:40

He has been with the company from

0:42

the very beginning, including its roots

0:44

at MIT, where he received his PhD

0:46

in aeronautical engineering. This

0:49

was in 1994 at

0:51

the legendary MIT Leg Lab.

0:54

He wrote his PhD thesis on

0:56

robot gymnastics, as part

0:58

of which he programmed a bipedal

1:00

robot to do the world's first 3D

1:03

robotic somersault. Robert

1:05

is a great engineer, roboticist, and leader, and

1:08

Boston Dynamics, to me as a roboticist,

1:11

is a truly inspiring company. This

1:13

conversation was a big honor and pleasure,

1:15

and I hope to do a lot of great work

1:18

with these robots in the years to come.

1:21

And now a quick few-second mention of each sponsor.

1:24

Check them out in the description. It's the best way

1:26

to support this podcast. We've

1:29

got NetSuite for business management software, Linode

1:31

for Linux systems, and Element

1:34

for zero-sugar electrolytes.

1:36

Choose wisely, my friends. Also, if you

1:39

want to work with our team, we're always hiring. Go

1:41

to luxfreedman.com slash hiring. And

1:44

now onto the full ad reads. As always, no

1:46

ads in the middle. I try to make this interesting, but

1:48

if you must commit the horrible,

1:50

terrible crime of skipping them, please

1:53

do check out our sponsors. I do

1:55

enjoy their stuff, I really do. And

1:57

maybe you will as well.

2:00

It's brought to you by NetSuite, an

2:02

all-in-one cloud business management

2:05

system. Running

2:08

a business, as this podcast reveals

2:11

from Robert Plater and Boston Dynamics,

2:13

is really hard. It's not just about

2:15

the design of the system. It's not just about

2:17

the engineering, the software, the hardware,

2:20

all the complicated research that goes into it, all

2:23

the different prototypes, all the failure

2:25

upon failure upon failure in the

2:27

early stages, in the middle stages of getting these

2:30

incredible robots to work.

2:32

It's all the glue

2:34

that ties the company together. For

2:37

that, you have to use the best tools for the job. I

2:40

hope to run a business, a large

2:42

business that actually builds stuff one

2:44

day and boys

2:47

are much more than just the innovation

2:49

and the engineering. I understand that deeply. You

2:52

should be hiring the best team

2:54

for that job

2:55

and use the best tools for that job. That's

2:58

where NetSuite can help out. Hopefully, it

3:00

can help you out. You can start now

3:02

with no payment or interest for six months.

3:05

Go to NetSuite.com slash Lex to

3:07

access their one-of-a-kind financing

3:10

program. That's NetSuite.com

3:13

slash Lex.

3:15

This episode is also brought to you by Linode,

3:18

now called Akamai, and their incredible

3:21

Linux virtual machines. I

3:23

sing praises to the greatest operating

3:25

systems of all time,

3:27

which is Linux. There's so many different beautiful

3:30

flavors of Linux. My favorite

3:32

is probably the different sub-flavors of Ubuntu,

3:34

Ubuntu Mate. That's what I use for my

3:37

personal development projects

3:40

in general when I want to feel comfortable

3:42

and fully customized, but I've used so

3:44

many other Linux's, distribution

3:46

of Linux's. But that's not what Linode

3:48

is about, or it is in part,

3:51

but it actually takes those Linux

3:53

boxes and scales them arbitrarily

3:55

to where you can do compute, not just on one machine,

3:57

but on many machines, customize them.

4:01

Make sure everything works reliably

4:03

and when it doesn't there's amazing human

4:06

customer service with real

4:09

humans. That's something that should

4:11

be emphasized in this day of Chad

4:13

GPT. Real human beings

4:16

that are good at what they do and figure

4:18

out

4:19

how to solve problems if they ever come up.

4:22

Linode, now called Akamai, is just

4:24

amazing. If compute

4:26

is something you care about for your business, for your personal

4:29

life, for your happiness, for

4:31

anything, then now you should check them out. Visit

4:33

linode.com slash Lex for

4:35

a free credit.

4:38

This episode is brought to you by a thing that

4:40

I'm currently drinking as I'm saying these words.

4:43

It's the element electrolyte drink mix

4:45

spelled L-M-N-T. My favorite

4:47

is the watermelon. That's what I always drink.

4:49

We have all explored. In

4:52

college, things got wild,

4:54

things got a little crazy, things got

4:56

a little out of hand. All

5:00

of us have done things we regret. Have

5:03

eaten ice cream we should not have eaten. I've

5:05

eaten ice cream at Dairy

5:08

Queen so many times in my life, especially through

5:10

my high school years. To

5:12

contradict what I just said, I regret nothing. I

5:14

think Snickers and if

5:17

memory serves me correctly, there's something called the

5:19

Dairy Queen Blizzard

5:21

where you could basically shove in whatever you

5:23

want into the ice cream and blend it and it tastes

5:25

delicious. I think

5:28

my favorite would be the Snickers bar, any kind

5:30

of bar, Mars bar.

5:32

Anything

5:34

with chocolate caramel, maybe a little

5:36

bit coconut,

5:37

that kind of stuff.

5:40

I don't regret it, but we've experimented. All

5:42

of us have experimented with different flavors, with

5:45

different things in life. I

5:47

regret nothing. You should not regret any

5:49

of it either because

5:50

that path

5:52

is what created the beautiful person that you are today.

5:56

That path is also the reason I mostly

5:58

drink the watermelon flavor.

5:59

of, I guess it's called watermelon

6:02

salt, I don't know what it's called, but watermelon is in the word,

6:05

of element. I highly recommend it. You could try other

6:07

flavors. Chocolate's pretty good too, like chocolate mint, I think

6:09

it's called. Totally different

6:11

thing. All the flavors are very different, and

6:13

that's why I love it, so you should explore. Anyway,

6:16

it's a good way to get all the electrolytes in your

6:18

system, the salt and magnesium, the

6:21

potassium, not salt, sodium is what I meant

6:24

to say. It doesn't matter what I meant to say. What

6:26

matters is it's delicious, and I'm consuming

6:29

it, and I'm singing the

6:29

praises, and I will toast to you when we see

6:32

each other in person one day, friend.

6:35

And we should drink,

6:36

element, drink,

6:39

to our deepest fulfillments, together,

6:42

as brothers and sisters in arms. Get

6:44

a simple pack for free with

6:47

any purchase. Try it at drinkelement.com

6:50

slash Lex.

6:53

This is the Lex Freeman podcast. To

6:55

support it, please check out our sponsors in the

6:57

description. And now, dear friends,

7:00

here's Robert Plater.

7:03

["The Lex Freeman Song"]

7:18

When did you first fall in love with robotics?

7:22

Let's start with love and robots.

7:25

Well, love is relevant because I think

7:28

the

7:29

fascination, the deep fascination is

7:31

really about movement. And

7:35

I was visiting MIT looking for a

7:37

place to get a PhD, and

7:40

I wanted to do some laboratory work. And

7:43

one of my professors in the Aero department

7:45

said, "'Go see this guy, Mark Raber, "'down in the basement

7:47

of the AI lab.'" And

7:50

so I walked down there and saw him. He

7:52

showed me his robots, and

7:55

he showed me this robot doing a somersault. And

7:58

I just immediately went, whoa!

7:59

You know? Yeah. Robots

8:02

can do that and because of my own interest in gymnastics,

8:05

there was like this immediate connection and

8:08

I

8:09

was interested in, I was in an aeroastro

8:11

degree because of flight and movement

8:14

was all so fascinating to me and then it turned

8:17

out that robotics had this

8:19

big challenge. How do you balance? How

8:22

do you build a legged robot that can really get

8:24

around? And that just,

8:26

that was a fascination. And

8:28

it still exists today. You were still working

8:31

on perfecting motion in

8:33

robots. What about the elegance

8:35

and the beauty of the movement itself? Is

8:37

there something maybe grounded

8:40

in your appreciation of movement

8:42

from your gymnastics days? Was

8:47

there something you just fundamentally appreciate

8:49

about the elegance and beauty of

8:51

movement? We had this concept in gymnastics

8:55

of letting your body

8:57

do what it wanted to do when you get

8:59

really good at

9:01

gymnastics. Part

9:04

of what you're doing is putting your body

9:06

into a position where the physics and the

9:08

body's inertia and momentum will

9:11

kind of push you in the right direction in a very

9:13

natural and organic way. And

9:15

the thing that Mark was doing, you know, in the

9:19

basement of that laboratory was trying to

9:21

figure out how to build machines to take advantage of

9:23

those ideas. How do you build something so

9:25

that the physics of the machine just

9:28

kind of inherently wants to do what

9:30

it wants to do? And he was building these

9:31

springy pogo stick type. His

9:34

first cut at Leggett Locomotion was a pogo

9:37

stick where it's bouncing and there's a spring mass

9:40

system that's oscillating, has its

9:42

own sort of natural frequency there. And

9:45

sort of figuring out how to augment those

9:47

natural physics with

9:50

also intent. How do you then control that but

9:52

not overpower it? It's that

9:54

coordination that I think

9:56

creates real potential.

9:58

We could call it beauty.

9:59

could call it, I don't know, synergy, that

10:02

people have different words for it. But I

10:04

think that that was inherent from the beginning.

10:07

That was clear to me that that's part of what

10:09

Mark was trying to do. He asked me to do that in

10:11

my research work. So, you know,

10:13

that's where it got going. So part of the thing that I think

10:15

I'm calling elegance and beauty in this case,

10:18

which was there, even with the pogo stick

10:20

is maybe the efficiency. So

10:23

letting the body do what it wants to do,

10:25

trying to discover the efficient movement.

10:27

It's definitely more efficient. It also

10:31

becomes easier to control in its own way because

10:33

the physics are solving some of the problem

10:35

itself. It's not like you have to

10:37

do all this calculation and overpower the

10:39

physics. The physics naturally, inherently

10:42

want to do the right thing. There

10:44

can even be feedback

10:47

mechanisms, stabilizing mechanisms that

10:50

occur simply by virtue of the

10:52

physics of the body. And it's

10:56

not all in the computer or not even all in your

10:58

mind as a person. And

11:01

there's something interesting in that melding.

11:04

You were with Mark for many, many, many years, but

11:07

you were there in this kind of legendary space

11:10

of Leg Lab at MIT in

11:13

the basement. All great things happen

11:15

in the basement. Is there some memories from

11:18

that time that you have

11:20

because it's such

11:23

cutting edge work in

11:27

robotics and artificial intelligence?

11:29

The memories, the distinctive lessons,

11:32

I would say, I learned in that

11:34

time period and

11:37

that I think Mark was a great teacher of, was

11:41

it's okay to pursue your

11:43

interests, your curiosity, do something because

11:46

you love it. You'll

11:48

do it a lot better if you love it.

11:52

That is a lasting lesson

11:55

that I think we apply at

11:57

the company still. and

12:01

really is a core value. So the

12:03

interesting thing is I get to, with

12:07

people like Ross Tedrick

12:09

and others, like the students that

12:12

work at those robotics labs are like some of the

12:14

happiest people I've ever met.

12:16

I don't know what that is. I

12:18

meet a lot of PhD students. A lot of them are kind

12:20

of broken by the wear and tear or

12:22

the process. But roboticists

12:25

are, while they work extremely hard and

12:27

work long hours,

12:29

there's

12:32

a happiness there. The only other group of people

12:34

I've met like that are people that skydive a lot. Like

12:38

for some reason there's a deep fulfilling happiness.

12:40

Maybe from like a long period

12:42

of struggle to get a thing to work and it works

12:44

and there's a magic to it, I don't know exactly. Because

12:47

it's so fundamentally hands on and

12:50

you're bringing a thing to life. I don't know what it is,

12:52

but they're happy. We see,

12:54

you

12:55

know, our attrition at the company

12:57

is really low. People come and they love

12:59

the pursuit. And I

13:01

think part of that is that

13:03

there's perhaps an extra connection

13:05

to it. It's a little bit easier to connect when

13:07

you have a robot that's moving around in

13:09

the world and part of your goal is to make it move around

13:12

in the world. You can identify with

13:14

that.

13:14

And this is one of the

13:17

unique things about the kinds of robots we're building

13:19

is this physical interaction lets

13:22

you perhaps identify with it. So

13:25

I think that is a source of happiness. I don't think

13:27

it's unique to robotics. I think anybody also

13:29

who is just pursuing something they

13:31

love, it's

13:33

easier to work hard at it and be good at

13:35

it. And not

13:38

everybody gets to find that. I do

13:40

feel lucky in that way. And I think we're

13:43

lucky as an organization that we've

13:45

been able to build a business around

13:47

this and that keeps people engaged. So

13:51

if it's all right, let's link on Mark for a

13:53

little bit longer. Mark Raybert, so he's

13:55

a legend. He's

13:58

a legendary engineer and roboticist.

13:59

What have you learned about life about

14:02

robotics and Mark? Through all the

14:04

many years you worked with him. I think

14:06

the most important lesson, which was,

14:08

have the courage of your convictions and do what

14:11

you think is interesting.

14:14

Be willing to

14:15

try to find big,

14:17

big problems to go after. And

14:20

at the time, like at locomotion,

14:23

especially in a dynamic machine, nobody

14:27

had solved it. And that felt like

14:29

a multi-decade problem

14:31

to go after. And so have

14:34

the courage to go after that because

14:36

you're interested. Don't worry

14:38

if it's gonna make money. You know, that's

14:40

been a theme. So that's

14:43

really

14:44

probably the most important

14:46

lesson I think that I got from

14:48

Mark. How crazy is the effort of doing legged

14:52

robotics at that time, especially?

14:55

You know, Mark got some stuff

14:57

to work starting from

14:59

the simple ideas. So maybe the other,

15:02

another important idea that has really become a value

15:05

of the company is try to simplify

15:08

a thing to the core essence.

15:10

And while

15:13

Mark was showing videos of animals running

15:15

across the savanna or climbing

15:19

mountains, what he started with was a pogo

15:21

stick because he was trying to reduce the

15:23

problem to something that was manageable and

15:26

getting the pogo stick to balance.

15:28

Had in it the

15:30

fundamental problems that if we solve those, you

15:32

could eventually extrapolate to something that

15:34

galloped like a horse. And

15:36

so look for those simplifying principles.

15:40

How tough is the job of simplifying a robot?

15:43

So I'd say in the early days, the

15:46

thing that made Boston,

15:48

the researchers at Boston Dynamics special is

15:51

that we worked on figuring

15:54

out what that central

15:57

principle was. And

15:59

then building. software or machines

16:02

around that principle. And that was not easy in

16:04

the early days. And it

16:06

took real expertise

16:09

in understanding the dynamics

16:11

of motion and feedback

16:13

control

16:15

principles. How to build

16:17

and

16:18

with computers at the time, how to build

16:20

a feedback control algorithm that was simple enough

16:22

that it could run in real time at a thousand Hertz

16:25

and actually get that machine to work. And

16:29

that was not something everybody was doing

16:31

at that time. Now the world's changing

16:34

now. I think the approaches

16:36

to controlling robots are going to change. And

16:40

they're going to become more

16:42

broadly available.

16:46

But at the time, there weren't many groups

16:49

who could really

16:50

work at that principled level with

16:53

both the software and make

16:56

the hardware work. And

16:58

I'll say one other thing about you're sort of talking

17:01

about what are the special things. The other thing

17:03

was it's

17:05

good to break stuff.

17:08

Use the robots, break

17:11

them, repair them, fix

17:15

and repeat, test, fix and repeat. And

17:17

that's also a core principle

17:19

that has become part of the company.

17:22

And it lets you be fearless

17:24

in your work. Too often

17:27

if you are working with a very expensive

17:29

robot, maybe one that you bought from somebody else

17:31

or that you don't know how to fix, then you

17:34

treat it with get gloves and

17:37

you can't actually make progress. You have to be able

17:39

to break something. And so I think that's been

17:42

a principle as well. So just to linger on that psychologically,

17:45

how do you deal with that? Because I remember I

17:47

built an RC

17:50

car with some

17:54

custom stuff like compute on it and all that kind of stuff,

17:57

cameras.

17:59

it didn't sleep much. The

18:02

code I wrote had an issue where it didn't stop

18:04

the car and the car got confused

18:07

and at full speed at like 20, 25 miles

18:09

an hour slammed into a wall. And

18:12

I just remember sitting there alone in

18:14

a deep sadness. Sort

18:18

of

18:19

full of regret, I think, almost

18:21

anger.

18:24

But also like sadness

18:26

because you think about, well, these robots,

18:28

especially for autonomous vehicles, you

18:31

should be taking safety very seriously even

18:33

in these kinds of things, but just no

18:36

good feelings. It made me more

18:38

afraid probably to do this kind of experiments

18:40

in the future. Perhaps the right way to have

18:42

seen that is positively.

18:45

It depends if you could

18:47

have built that car or just gotten

18:49

another one, right? That would have been the approach. I

18:53

remember

18:55

when I

18:57

got to grad school, I

19:00

got some training about operating

19:02

a lathe and a mill up in

19:04

the machine shop, and I could start

19:06

to make my own parts. And I remember

19:09

breaking some piece of equipment

19:11

in the lab and then

19:13

realizing,

19:14

because maybe this was a unique part and I couldn't

19:17

go buy it. And I realized, oh, I can

19:19

just go make it. That

19:22

was an enabling feeling. Then

19:24

you're not afraid. You know, it might take time. It

19:26

might take more work than you thought it was

19:29

going to be required to get this thing done,

19:31

but you can just go make it. And that's

19:33

freeing in a way that nothing

19:36

else is. You mentioned

19:39

the feedback control, the dynamics.

19:41

Sorry for the romantic question, but

19:44

in the early days and even now, is the dynamics

19:47

probably more appropriate for the early days? Is

19:50

it more art or science?

19:54

There's a lot of science around it.

19:56

And trying to develop

19:59

scientific principles. that

20:01

let you extrapolate from one-legged

20:03

machine to another, develop

20:06

a core set of principles like

20:09

a spring mass bouncing system,

20:12

and then figure out how to apply

20:14

that from a one-legged machine to a two or a four-legged

20:17

machine. Those principles are really important

20:19

and we're definitely a core

20:22

part of our work. There's

20:26

also, you know,

20:28

when we started to pursue humanoid robots,

20:32

there was so much complexity in that machine

20:34

that, you

20:36

know, one of the benefits of the humanoid form

20:38

is you have some intuition about how it

20:41

should look while it's moving. And

20:44

that's a little bit of an art, I think. Now

20:46

it's, or maybe it's just tapping into a

20:48

knowledge that you have deep in your body and then

20:51

trying to express that in the machine.

20:54

But that's an intuition that's a

20:56

little bit more on the art side. Maybe

20:58

it predates your knowledge. You know, before

21:00

you have the knowledge of how to control it, you try

21:02

to work through the art channel and

21:04

humanoids sort of make that available to you.

21:06

If it had been a different shape, maybe

21:08

you wouldn't have had the same intuition about it.

21:11

Yeah, so your knowledge

21:13

about moving through the world is

21:16

not made explicit to you. So you

21:18

just, that's why it's art. Yeah, it

21:21

might be hard to actually articulate exactly.

21:23

There's something about,

21:25

and being a competitive athlete,

21:28

there's something about

21:29

seeing movement. You

21:31

know, a coach, one of the greatest strengths

21:34

a coach has is being able to see

21:37

some little change in what the athlete is doing

21:39

and then being able to articulate that to an athlete.

21:42

And then maybe even trying to say, and you should

21:44

try to feel this. So

21:47

there's something just in seeing. And again,

21:49

sometimes it's hard to articulate

21:52

what it is you're seeing, but there's

21:55

just perceiving the motion at a rate

22:00

Again, sometimes hard to put into words. Yeah,

22:03

I wonder how it

22:07

is possible to achieve sort of truly

22:09

elegant movement. You have a movie like

22:11

Ex Machina, not sure if you've seen it,

22:14

but the main actress in that who

22:16

plays the AI robot, I think is

22:18

a ballerina. I mean, just the natural

22:22

elegance and the, I

22:24

don't know, eloquence of movement, it

22:29

looks efficient and easy and

22:31

just, it looks right.

22:34

It looks beautiful. It looks right, is sort of the key, yeah.

22:36

And then you look at, especially

22:38

early robots, I mean, they're so

22:41

cautious in the way they move that

22:45

it's not the caution that looks wrong.

22:48

It's something about the movement that looks

22:50

wrong that feels like it's very

22:52

inefficient, unnecessarily so. And

22:55

it's hard to put that into words

22:57

exactly. We think, and

23:00

part of the reason why people are attracted to the

23:02

machines we build is

23:04

because the inherent dynamics of movement

23:08

are closer to right. Because

23:11

we try to use

23:13

walking gates or we build a machine

23:15

around this gate where you're trying to work

23:17

with the dynamics of the machine instead

23:19

of

23:20

to stop them. Some of the early walking

23:23

machines, you're essentially,

23:25

you're really trying hard to not

23:27

let them fall over. And so you're always stopping

23:30

the tipping motion. And

23:33

sort of the insight

23:35

of dynamic stability in a valued

23:37

machine is to go with it, let

23:40

the tipping happen, let yourself

23:42

fall, but then catch yourself with

23:44

that next foot. And there's something about getting

23:47

those physics to be expressed in the

23:49

machine

23:50

that people interpret as lifelike

23:52

or

23:56

elegant or just natural

23:58

looking. And so I think if you...

23:59

get the physics right, it

24:02

also ends up being more efficient, likely.

24:05

There's a benefit that it probably ends up being

24:08

more stable in the long run. It could

24:11

walk stably over a wider range

24:13

of conditions. It's

24:17

more beautiful and attractive at the same time. How

24:20

hard is it

24:21

to get the humanoid robot Atlas to

24:24

do some of the things it's recently been doing? Let's

24:27

forget the flips and all of that. Let's

24:29

just look at the running. Maybe

24:31

you can correct me, but there's something about running.

24:33

I mean, that's not careful at all. That's you're falling

24:36

forward. You're jumping forward

24:38

and are falling. So how hard is it to get

24:40

that right? Our first humanoid,

24:43

we needed to deliver natural looking walking.

24:46

We took a contract from the army. They

24:48

wanted a robot that could walk

24:51

naturally. They wanted to put a

24:53

suit on the robot and be able to test it

24:55

in a gas environment. And so they

24:58

wanted the motion to be natural.

25:02

Our goal was a natural looking gate. It

25:04

was surprisingly hard

25:06

to get that to work. But we

25:09

did build an early machine. We

25:11

called it Petman

25:13

prototype. It was the prototype before

25:15

the Petman robot. And

25:18

it had a really nice looking

25:20

gate where it

25:23

would stick the leg out. It would do heel strike

25:26

first before it rolled onto the toes. So you didn't

25:28

land with a flat foot. You extended your

25:30

leg a little bit. But

25:33

even then, it was hard to get the robot to walk

25:35

when you were walking that it fully

25:37

extended its leg and essentially

25:39

landed on an extended leg. And if you watch closely

25:42

how you walk, you probably land on an extended

25:44

leg, but then you immediately flex your knee as

25:46

you start to make that contact. And

25:49

getting that all to work well

25:51

took such a long time. In fact, I

25:55

probably didn't really see the nice

25:57

natural walking that I expected out of

25:59

our human life.

25:59

until maybe last year.

26:01

And

26:03

the team was developing on our newer generation

26:06

of Atlas, some new techniques for

26:10

developing a walking control algorithm. And

26:13

they got that natural looking motion as

26:16

sort of a byproduct of just

26:18

a different process they were applying to developing

26:21

the control.

26:22

So that probably took 15 years, 10 to 15 years

26:26

to sort of get that. The

26:29

Petman prototype was probably in 2008 and

26:32

what was it, 2022? Last

26:34

year that I think I saw a good walking on Atlas.

26:36

If you could just like link on it, what are some

26:38

challenges of getting good walking? So is

26:41

it, is this partially like

26:44

a hardware like

26:46

actuator problem? Is it the control?

26:50

Is it the artistic element of just observing

26:52

the whole system operating in different conditions

26:54

together? I mean, is there some kind of interesting

26:58

quirks or challenges you can speak to

27:00

like the heel strike? Yeah, so one of

27:02

the things that makes the, like this straight

27:04

leg a challenge is

27:07

you're sort of up against a singularity,

27:10

a mathematical singularity where,

27:13

you know, when your leg is fully extended, it can't

27:15

go further the other direction, right? There's only,

27:18

you can only move in one direction. And

27:20

that makes all of the calculations around

27:22

how to produce

27:23

torques at that joint or positions,

27:26

makes it more complicated. And so having

27:29

all of the mathematics so it can deal with these

27:33

singular configurations

27:35

is one of many challenges

27:39

that we face. And I'd say

27:41

in the, you know,

27:43

in those earlier days, again, we were working

27:45

with these really simplified models. So

27:48

we're trying to boil all the physics

27:50

of the complex human body into

27:52

a simpler

27:54

subsystem that we can more easily describe

27:56

in mathematics. And sometimes those simpler

27:59

subsystems don't. have all of that complexity

28:01

of the straight leg built into

28:03

them. And so what's

28:06

happened more recently is we're able to apply

28:08

techniques that let us take the full physics

28:11

of the robot

28:13

into account and deal

28:15

with some of those strange situations

28:18

like the straight leg. So is

28:20

there a fundamental challenge here that it's,

28:22

maybe you can correct me, but is it under

28:25

actuated? Are you falling? Under

28:27

actuated is the right word, right? You can't push

28:31

the robot in any direction you want to, right?

28:33

And so that is one of the hard

28:35

problems of like a locomotion.

28:37

And you have to do that for natural

28:39

movement.

28:40

It's not necessarily required for natural

28:42

movement. It's just required, you know,

28:46

we don't have a gravity force

28:48

that you can hook yourself on to to apply

28:50

an external force in the direction

28:52

you want at all times, right? The

28:54

only external forces are being mediated through

28:57

your feet and how they get mediated

28:59

depend on how you place your feet. And

29:02

you can't just, you

29:04

know, God's hand can't reach down

29:06

and push in any direction you want, you

29:09

know, so. Is there

29:11

some extra challenge to the fact that Alice

29:13

is such a big robot? There is. The

29:16

humanoid form is

29:18

attractive in many ways,

29:20

but it's also a challenge in many ways. You

29:24

have this big upper body that has a lot

29:27

of mass and inertia and

29:29

throwing that inertia around

29:31

increases the complexity of maintaining balance.

29:34

And as soon as you pick up something heavy in your arms,

29:37

you've made that problem even harder. And

29:40

so in the early work

29:43

in the leg lab and in the early days at the

29:45

company, you know, we were pursuing these

29:47

quadruped robots, which had a

29:50

kind of built in simplification. You had

29:52

this big rigid body and then really light

29:54

legs. So when you swing the legs,

29:57

the leg motion didn't

29:59

impact. the body motion very much. All

30:02

the mass and inertia was in the body. But

30:04

when you have the humanoid, that doesn't work. You

30:06

have big heavy legs, you swing the legs, it

30:09

affects everything else. And so

30:12

dealing with all of that interaction does

30:15

make the humanoid a much more complicated

30:17

platform. And I also saw

30:19

that at least recently you've

30:22

been doing more explicit modeling of the

30:24

stuff you pick up.

30:25

Yeah. Which is very,

30:29

really interesting. So you have to, what,

30:32

model the shape, the

30:34

weight distribution? I don't

30:37

know. Like you have to include

30:39

that as part of the modeling, as part of the planning.

30:42

Because, okay, so for people who don't know,

30:44

so Atlas, at

30:46

least in like a recent video, like throws

30:48

a heavy bag, throws a bunch of

30:51

stuff. So what's involved in picking

30:54

up a thing, a heavy thing, and

30:57

when that thing is a

30:59

bunch of different non-standard things, I think

31:01

it's also picked up like a barbell. And

31:05

to be able to throw in some cases, what

31:07

are some interesting challenges there?

31:09

So we were definitely trying to show that

31:11

the robot and the techniques we're applying to

31:13

the robot, to Atlas, let

31:16

us deal with heavy things

31:18

in the world. Because if the robot's going to be useful, it's

31:20

actually got to move stuff around. And

31:23

that needs to be significant stuff. That's

31:26

an appreciable portion of the body weight

31:28

of the robot.

31:29

And we also think this differentiates us

31:31

from the other humanoid robot activities

31:34

that you're seeing out there. Mostly they're not picking

31:36

stuff up yet. And not heavy stuff

31:38

anyway. But

31:40

just like you or me, you need to anticipate

31:43

that moment. You're reaching out to pick something up,

31:45

and as soon as you pick it up, your center of mass is going

31:47

to shift. And if you're going

31:49

to turn in a circle,

31:51

you have to take that inertia into count.

31:53

And if you're going to throw a thing, all of that

31:56

has to be included in the model.

31:59

model of what you're trying to do. So

32:02

the robot needs to have some idea or

32:04

expectation of what that weight is and then

32:07

sort of predict, think

32:09

a couple of seconds ahead, how do I manage

32:11

my, now my body

32:13

plus this big heavy thing together

32:15

to get, and still maintain

32:18

balance, right? And so

32:22

that's a big change for

32:24

us. And I think the tools we've built

32:26

are really allowing that to happen quickly

32:28

now. Some of those motions that you saw

32:31

in that most recent video, we

32:33

were able to create in a matter of days.

32:35

It used to be that it took six months to do anything

32:38

new, you know, on the robot. And now we're

32:40

starting to develop the tools that let us do that in

32:42

a matter of days. And so we think

32:44

that's really exciting. That means that the

32:46

ability to create new behaviors for the robot

32:49

is gonna be a

32:51

quicker process. So being able to explicitly

32:54

model new things that

32:56

it might need to pick up, new types

32:58

of things. And to some degree, you don't

33:00

wanna have to pay too much attention to each

33:03

specific thing,

33:05

right? There's sort of a generalization

33:07

here. Obviously

33:10

when you grab a thing, you have to

33:12

conform your hand, your end effector

33:14

to the surface of that shape. But once

33:16

it's in your hands, it's probably just

33:19

the mass and inertia that matter. And

33:21

the shape may not be as important.

33:24

And so, in

33:26

some ways you wanna pay attention to that detailed shape.

33:29

And in others you wanna generalize it and say, well,

33:32

all I really care about is the center of mass of this

33:35

thing, especially if I'm gonna throw it up on that scaffolding.

33:38

And it's easier if the body is rigid. What if

33:40

there's some, doesn't it throw like a sandbag

33:42

type thing? That tool bag, you know,

33:45

had loose stuff in it. So it

33:48

managed that. There are harder things

33:50

that we haven't done yet. You know, we could have had a big jointed

33:53

thing or I don't know, a bunch of loose wire

33:55

or rope. What about carrying another robot? How

33:57

about that? Ha ha ha.

33:59

Yeah, we haven't we haven't done that yet carry

34:02

spot I guess we did a little bit of a we

34:04

did a little skit around Christmas

34:06

where we had Two spots holding

34:09

up another spot that was trying to put you know a bow

34:11

on a tree So I guess we're doing that in a small

34:13

way Okay, that's pretty

34:16

good.

34:16

Let me ask the all-important question. Do

34:19

you know how much Atlas can curl? Have

34:23

you I Mean,

34:25

you know this for us humans that's

34:28

really one of the most fundamental questions you could

34:31

ask another human being Bench

34:35

It's probably can't curl as much as we can

34:37

yet But a metric that I think

34:39

is interesting is you know

34:41

another way of looking at that strength Is

34:44

you know the box jump? So if how high

34:47

of a box can you jump

34:49

onto question? And

34:51

Atlas, I don't know the exact height It

34:53

was probably a meter high or something like that It was

34:56

a pretty pretty tall jump that Atlas was

34:58

able to manage when we last tried

35:00

to do this and and I have video Of

35:03

my chief technical officer doing

35:05

the same jump and he really struggled, you know

35:07

Oh the human but the human

35:09

getting all the way on top of this box, but then you know

35:11

Atlas was able to do it We're

35:15

now thinking about the next generation of Atlas

35:17

and we're probably going to be in the realm of Person

35:20

can't do it, you know with this with the next generation

35:23

And the robots the actuators are

35:25

going to get stronger where It really

35:27

is the case that at least some of these joints some of these

35:29

motions will be stronger

35:31

and to understand how high I Can jump you probably

35:33

had to do quite a bit of testing. Oh,

35:35

yeah, and there's lots of videos of it trying and failing

35:37

And that's you know, that's all yeah,

35:40

we don't always release those those videos, but they're

35:42

a lot of fun to look at So

35:45

we'll talk a little bit about that But

35:47

if you can you talk to the jumping because you

35:50

talked about the walking it took a long time

35:52

many many years to get the Walking to be natural,

35:54

but there's also really natural

35:57

looking

35:58

robust

35:59

resilient jumping.

36:01

How hard is it to do the jumping? Well,

36:04

again, this stuff has really evolved rapidly

36:06

in the last few years. The first time we

36:08

did a somersault, there

36:10

was a lot of manual iteration.

36:14

What is the trajectory? How hard do

36:16

you throw you? In fact, in these early

36:18

days,

36:19

when I'd see

36:22

early experiments that the team was doing, I might

36:24

make suggestions about how to change the technique.

36:27

Again, borrowing from my own intuition

36:30

about how backflips work.

36:33

Frankly, they don't need that anymore. In the

36:35

early days, you had to iterate in

36:38

almost a manual way, trying to change

36:40

these trajectories of the arms or the legs

36:42

to try to get a successful

36:46

backflip to happen. More

36:48

recently, we're running these

36:50

model predictive

36:53

roll techniques, where

36:56

the robot essentially can think in advance

36:59

for the next second or two about how

37:02

its motion is going to transpire. You

37:04

can solve for optimal trajectories

37:06

to get from A to B. This

37:09

is happening in a much more natural way. We're

37:11

really seeing an acceleration happen

37:13

in the development of these behaviors, again,

37:16

partly due to these

37:19

optimization techniques, sometimes

37:21

learning techniques. It's

37:25

hard in that there's

37:27

a lot of mathematics behind it, but

37:31

we're figuring that out. You can do model

37:34

predictive control for... I

37:37

don't even understand what that looks like when the

37:39

entire robot is in the air, flying

37:42

and doing a backflip. Yeah,

37:45

but that's the cool part, right? The physics,

37:47

we can calculate physics

37:50

pretty well using Newton's laws about

37:53

how it's going to evolve over time.

37:55

The sick trek,

37:57

which was a front somersault with a hat.

37:59

twist is a good example, right?

38:03

You saw the robot on various

38:05

versions of that trick, I've seen

38:08

it land in different configurations

38:10

and it still manages to stabilize itself.

38:12

And so, you know, what this model

38:15

predictive control means is, again,

38:17

in

38:18

real time, the robot is

38:20

projecting ahead, you know, a second into the

38:22

future and sort of exploring options.

38:25

And if I move my arm a little bit more this

38:27

way, how is that going to affect the outcome? And so

38:29

it can do these calculations, many of them,

38:32

you know, and basically solve

38:35

where, you know, given where I am now, maybe

38:37

I took off a little bit screwy from how

38:40

I had planned, I can adjust. So

38:42

you're adjusting in the air. Just on the

38:44

fly. So the model predictive control

38:46

lets you adjust on the fly. And

38:48

of course, I think this is what people

38:50

adapt as well. When

38:53

we do it,

38:54

even a gymnastics trick, we try to set it up

38:56

so it's close to the same every time.

38:59

But we figured out how to do some adjustment on the fly.

39:01

And now we're starting to figure out that the robots

39:03

can do this adjustment on the fly as well

39:05

using these techniques in the air. It's

39:08

so, I mean, it just feels

39:11

from a robotics perspective, just surreal. Well,

39:14

that's sort of the you talked about under actuated,

39:16

right? So when you're in the air, there's

39:19

something there's some things you can't change,

39:21

right? You can't change the momentum while

39:23

it's in the air because you can't apply an external force

39:25

or torque. And so the momentum isn't going to

39:28

change. So how do you work within the constraint

39:30

of that fixed momentum to still get

39:32

from A to B where you want to be?

39:34

That's really on track.

39:38

You're in the air. I mean, you become a drone

39:41

for a brief moment in time. No, you're not even a drone

39:43

because you can't, can't hover. You

39:45

can't hover. You're going to impact

39:48

soon. Be ready. Yeah. Are you considered

39:50

like a hover type thing or no? No, it's

39:52

too much weight. I

39:54

mean, it's just, it's just incredible. It's

39:57

just even to have the guts

39:59

to try back.

39:59

flip with such a large body.

40:02

That's wild. Like,

40:05

uh, Oh, we definitely broke a few robots trying to,

40:08

but that's where the build it, break it, fix it. You know, uh, strategy

40:11

comes in, gotta be willing to break. And what

40:14

ends up happening is you end up by breaking

40:16

the robot repeatedly. You find the weak points and

40:19

then you end up redesigning it. So it doesn't break so

40:21

easily. Next time, you know, through

40:23

the breaking process, you learn a lot,

40:25

like a lot of lessons and, and you keep

40:28

improving, not just how to make the back flip work,

40:30

but everything just how to build a machine better.

40:32

Yeah. Yeah. I mean, is there

40:34

something about just the guts to

40:37

come up with an idea of saying, you

40:40

know what, let's try to make it do a back flip.

40:42

Well, I think the courage to do a back flip

40:44

in the first place and, and to not worry

40:47

too much about the ridicule of somebody saying,

40:49

why the heck are you doing back flips with robots?

40:52

Because a lot of people have asked that, you know, why,

40:54

why, why are you doing this?

40:56

Why go to the moon in this

40:58

decade and do the other things? JFK? Not

41:02

because it's easy because it's hard. Yeah, exactly.

41:07

Don't

41:07

ask questions. Okay. So the,

41:10

the jumping, I mean, it's just, there's a lot of incredible stuff.

41:12

If we can just rewind a little bit to

41:15

the DARPA robotics challenge in 2015, I

41:18

think, which was for

41:20

people who are familiar with the DARPA challenges, it

41:24

was first with autonomous vehicles

41:26

and there's a lot of interesting challenges around that. And

41:29

the DARPA robotics challenges when humanoid

41:32

robots were tasked to do all

41:35

kinds of, you

41:39

know, manipulation, walking,

41:41

driving a car, all

41:43

these kinds of challenges with, if

41:45

I remember correctly, sort of some

41:49

slight capability to communicate

41:51

with humans, but the

41:53

communication was very poor. So basically it

41:55

has to be almost entirely autonomous.

41:59

was entirely interrupted and the robot

42:02

had to be able to proceed. But you could provide

42:04

some high level guidance to the robot,

42:06

basically low bandwidth communications

42:09

to steer it.

42:10

I watched that challenge with kind of

42:13

tears in my eyes eating popcorn. I

42:19

wasn't personally losing hundreds

42:21

of thousands, millions of dollars and

42:24

many years of incredible hard work by

42:27

some of the most brilliant roboticists in the world. So

42:29

that was why the tragic, why the tears

42:31

came. So anyway, what have you,

42:34

just looking back to that time, what have you learned

42:36

from that experience?

42:38

Maybe if you could describe what it was,

42:41

sort of the setup for people who haven't seen it.

42:44

Well, so there was a contest where a

42:46

bunch of different robots

42:48

were asked to do a series of tasks. Some

42:51

of those that you mentioned, drive a vehicle,

42:53

get out, open a door, go

42:55

identify a valve, shut a valve, use

42:58

a tool to maybe cut a hole in

43:00

a

43:02

surface and then crawl

43:05

over some stairs and maybe some

43:07

rough terrain. So it

43:09

was,

43:10

the idea was have a

43:12

general purpose robot that could do lots of different

43:15

things. It

43:18

had to be mobility and manipulation, onboard

43:20

perception.

43:22

And there was a contest which

43:25

DARPA likes at the time was running

43:28

sort of follow on to the grand

43:30

challenge, which was let's try

43:32

to push vehicle autonomy along,

43:35

right? They encourage people

43:37

to build autonomous cars. So

43:39

they're trying to

43:39

basically push an industry forward. And

43:45

we were asked, our role in this was to build

43:47

a humanoid at the time it was our

43:50

sort of first generation Atlas robot.

43:54

And we built

43:56

maybe 10 of them. I don't remember the exact

43:59

number. And DARPA

44:01

distributed those to various teams

44:04

that sort of won a contest,

44:08

showed that they could program

44:11

these robots and then use them to compete

44:13

against each other. And then other robots were introduced

44:15

as well. Some teams built their own robots. Carnegie

44:19

Mellon, for example, built their own robot. And

44:22

all these robots competed to see who could sort of

44:24

get through this maze of the

44:27

fastest. Again,

44:29

I think the purpose was to kind of push the whole industry

44:32

forward.

44:33

We provided the robot and

44:35

some baseline software, but

44:38

we didn't actually compete as a participant where

44:41

we were trying to drive

44:44

the robot through this maze. We were

44:46

just trying to support the other teams. It

44:49

was humbling because it was really a hard task. And

44:53

honestly, the robots, the tiers were because mostly

44:56

the robots didn't do it. They fell down

44:58

repeatedly. It

45:02

was hard to get through this

45:03

contest. Some did,

45:05

and they were rewarded in

45:07

one. But it was humbling because

45:10

of just how hard, these tasks weren't all that

45:12

hard. A person could have done it very easily, but it

45:15

was really hard to get the robots to do it. The

45:18

general nature of it, the variety

45:20

of it. And also that I don't

45:23

know if the tasks were...

45:25

The

45:28

task in themselves helped

45:30

us understand what is difficult and what is not. I

45:33

don't know if that was obvious before the contest

45:35

was designed. So you kind of try to

45:37

figure that out. And I think Atlas

45:40

is really a general robot platform,

45:43

and it's perhaps not best suited for

45:45

the specific tasks of that contest.

45:48

Just for example,

45:50

probably the hardest task is not the

45:52

driving of the car, but getting in

45:54

and out of the car. And Atlas

45:57

probably... If you were to design

45:59

a robot...

45:59

that can get into the car

46:02

easily and get out easily, you probably would not

46:04

make Atlas that particular car. Yeah,

46:07

the robot was a little bit big to get in and

46:09

out of that car. Right, it doesn't fit,

46:11

yeah. This is the curse of a general purpose

46:13

robot, that they're not perfect at any

46:15

one thing,

46:16

but they might be able to do a wide variety

46:18

of things. And that is

46:21

the goal at the

46:23

end of the day. I think

46:26

we all wanna build general purpose robots

46:29

that can be used for lots of different activities, but

46:32

it's hard. And

46:34

the wisdom in

46:37

building successful robots, up

46:39

until this point, have been, go build

46:41

a robot for a specific task and it'll

46:43

do it very well. And as long as you

46:46

control that environment, it'll operate

46:48

perfectly.

46:49

But robots need to be able to deal with uncertainty.

46:52

If they're gonna be useful to us in the future,

46:55

they need to be able to deal with unexpected

46:58

situations. And that's sort of the goal

47:00

of a general purpose or a multi-purpose robot.

47:03

And that's just darn hard. And so some

47:05

of the others, these curious little failures, like

47:07

I remember one of a robot,

47:10

the first

47:12

time you start to try to push on the world

47:15

with a robot, you

47:17

forget that the world pushes back and

47:20

will push you over if you're not ready for it. And

47:23

the robot reached to grab

47:25

the door handle. I think it missed the grasp

47:27

of the door handle, was expecting

47:29

that its hand was on the door handle. And so

47:32

when it tried to turn the knob, it just threw

47:34

itself over. It didn't realize, oh, I

47:36

had missed the door handle. I didn't have, I

47:39

didn't, I was expecting a force back from

47:41

the door. It wasn't there.

47:42

And then I lost my balance. So these

47:44

little simple things that you and I would

47:47

take totally for granted and deal with the

47:50

robots don't know how to deal with yet. And

47:52

so you have to start to deal with all of those

47:54

circumstances. Well, I think a

47:57

lot of us experienced this in...

47:59

even when sober but drunk

48:02

too, sort of you

48:04

pick up a thing and expect it to be,

48:06

what is it, heavy and it turns out

48:08

to be light. Yeah, and then you woo. Oh

48:11

yeah, and then so the same, and I'm

48:13

sure if your depth perception for whatever reason is

48:15

screwed up, if you're drunk

48:17

or some other reason and then you think

48:19

you're putting your hand on the table

48:21

and you miss it, I mean it's the same kind of situation.

48:25

But there's a- Which is why you need to

48:27

be able to predict forward just a little bit. And

48:30

so that's where this model predictive control stuff

48:32

comes in. Predict forward what you think's going

48:34

to happen. And then if that

48:36

does happen, you're in good shape. If something else happens,

48:38

you better start predicting again. So like

48:41

regenerate a plan when

48:44

you don't. I mean

48:47

that also requires a very fast

48:50

feedback loop of updating what

48:54

your prediction, how it matches to the actual real

48:56

world.

48:57

Yeah, those things have to run pretty quickly.

48:59

What's the challenge of running things pretty quickly?

49:01

A thousand hertz of

49:04

acting and sensing

49:07

quickly. You know, there's a few different

49:09

layers of that. You want at the lowest

49:11

level, you like to run things typically

49:14

at around a thousand hertz, which means that

49:16

at each joint of the robot, you're measuring

49:19

position or force and then trying to

49:21

control your actuator, whether it's

49:23

a hydraulic or electric motor, trying

49:26

to control the force coming out of that actuator.

49:28

And you want to do that really fast,

49:31

something like a thousand hertz. And that means you

49:33

can't have too much calculation going

49:35

on at that joint. But

49:38

that's pretty manageable these days and it's

49:40

fairly common. And then there's another

49:42

layer that you're probably calculating, you

49:44

know,

49:45

maybe at a hundred hertz, maybe 10 times

49:47

slower, which is now starting to

49:49

look at the overall body motion and

49:52

thinking about the larger physics

49:55

of the robot. And

49:59

then there's yet another layer.

49:59

loop that's probably happening a little bit slower,

50:02

which is where you start to bring your

50:04

perception and your vision and things

50:06

like that.

50:07

And so you need to run all of these

50:09

loops sort of simultaneously. You

50:11

do have to manage your computer

50:14

time so that you can squeeze

50:16

in all the calculations you need in

50:18

real time in a very consistent way. And

50:23

the amount of calculation we can

50:25

do is increasing as computers

50:27

get better, which means we can start to do more

50:29

sophisticated calculations. I can have

50:31

a more complex model doing

50:34

my forward prediction.

50:37

And that might allow me to do even

50:39

better predictions as I get better and better.

50:42

And it used to be, again, we had,

50:44

you know, 10 years ago,

50:46

we had to have pretty simple models

50:50

that we were running, you know, at those fast

50:52

rates because the computers weren't as capable

50:55

about calculating forward with

50:57

a sophisticated model. But as

51:00

computation gets better, we

51:02

can do more of that. What about the actual pipeline

51:05

of software engineering?

51:07

How easy is it to keep updating Atlas,

51:10

like to continue its development on it? So

51:12

how many computers are

51:14

on there? Is there a nice pipeline?

51:17

It's an important part of building

51:19

a team around it, which means,

51:22

you know, you need to also

51:24

have software simulation tools,

51:26

you know. So we

51:29

have always made

51:31

strong use of physics-based

51:33

simulation tools to do some

51:36

of this calculation, basically

51:38

test it in simulation before you put it on

51:40

the robot. But you also want the same

51:43

code that you're running in simulation to be the

51:45

same code you're running on the hardware. And

51:47

so even getting to the point where

51:50

it was the same code going from one to the

51:52

other,

51:53

we probably didn't really get that working until,

51:55

you know, a few years, several years ago. But

51:58

that was a, you know, that was a bit of a model.

51:59

milestone. And so you want to work,

52:02

certainly work these pipelines so that you can make it

52:04

as easy as possible and have a bunch of people

52:06

working in parallel, especially when we

52:09

only have, you know, for the Atlas

52:11

robots, the modern Atlas robots

52:13

at the company. And you know, we

52:15

probably have, you know, 40 developers,

52:18

they're all trying to gain access

52:20

to it. And so you need to share resources

52:22

and use some of these, some of the software

52:25

pipeline. Well, that's a really exciting step

52:27

to be able to run the exact same code and simulation

52:29

as on the actual robot. How

52:31

hard is it to do

52:35

realistic simulation,

52:37

physics based simulation of, of

52:40

Atlas such that, I mean, the

52:42

dream is like, if it works in simulation

52:44

works perfectly in reality, how hard is it to

52:47

sort of close, keep working on closing that gap?

52:49

The root of some of our physics based simulation

52:51

tools really started at MIT.

52:54

And we built some,

52:56

some, some good physics based modeling tools there.

52:59

The early days of the company, we were trying

53:02

to develop those tools as a commercial product.

53:04

So we continued to develop them.

53:06

It wasn't a particularly successful commercial product.

53:09

But we ended up with some nice physics based simulation

53:11

tools so that when we started doing legged robotics,

53:13

again, we had a really nice tool to work with. And

53:16

the things we paid attention to were

53:18

things that weren't necessarily handled very

53:21

well. In the commercial tools, you could

53:23

buy off the shelf like, like interaction

53:25

with the world, like foot ground contact.

53:28

So trying to model those contact events

53:32

well,

53:33

in a way that captured

53:36

the important parts of the interaction

53:40

was a really important element to

53:42

get right. And to also do in a way

53:44

that was computationally feasible and

53:47

could run fast. Because if you, if your simulation

53:50

runs too slow, you know, then your developers

53:52

are sitting around waiting for stuff to run and compile.

53:55

So it's always about efficient,

53:58

fast operations.

53:58

as well. So

54:01

that's been a big part of it. I think developing

54:03

those tools in parallel to the development

54:06

of the platform and trying

54:08

to scale them has really been essential,

54:11

I'd say, to us being

54:13

able to assemble a team of people that could do this. Yeah,

54:16

how to simulate contact periods

54:18

of foot-ground contact but for

54:20

manipulation because

54:23

don't you want to

54:24

model

54:26

all kinds of surfaces? Yeah, so it

54:28

will be even more complex with

54:30

manipulation because there's a lot more going on

54:33

and you need to capture things

54:36

slipping and moving in your

54:39

hand. It's

54:41

a level of complexity that I think goes

54:43

above foot-ground

54:46

contact when you really start doing

54:49

dexterous manipulation. So there's challenges

54:51

ahead still. So how far are we

54:53

away from me being able to walk with Atlas

54:56

in the sand along the beach and

54:58

us both drinking a beer?

55:04

Maybe Atlas could spill his beer

55:06

because he's got nowhere to put it. Atlas

55:10

could walk on the sand. So can it? Yeah,

55:13

I mean, have we really had him out on

55:15

the beach? We take them outside

55:17

often, rocks, hills, that

55:19

sort of thing, even just around our lab in Waltham.

55:23

We probably haven't been on the sand but I

55:26

don't doubt that we could deal with

55:29

it. We

55:29

might have to spend a little bit of time to sort of make that work

55:32

but we

55:35

had to take

55:37

Big Dog to Thailand years

55:39

ago and we

55:42

did this great video of the robot

55:44

walking in the sand, walking into

55:47

the ocean up to, I

55:49

don't know, its belly or something like that

55:51

and then turning around and walking out, all while

55:53

playing some cool beach music. Great

55:56

show but then we didn't really clean the robot

55:58

off and the saltwater was really hard.

55:59

on it, so we put it in a box,

56:02

shipped it back. By the time it came back, we

56:04

had some problems with corrosion. So

56:07

it's a salt water. It's not like... Salt

56:09

stuff. It's not like sand getting into the components

56:11

or something like this. But I'm sure if

56:14

this is a big priority, you can make

56:16

it waterproof. Right, right. That

56:18

just wasn't our goal at the time. Well,

56:20

it's a personal goal of mine to walk along the

56:23

beach. But it's a human

56:25

problem too. You get sand everywhere, it's just

56:27

a jam mess. So

56:29

soft surfaces are OK. So I

56:31

mean, can we just linger on the robotics

56:34

challenge? There's a pile of rubble there

56:35

to walk over. Is

56:40

that... How

56:43

difficult is that task? In

56:45

the early days of developing Big Dog, the

56:47

loose rock was the epitome of

56:49

the hard walking surface. Because you step

56:51

down and then the rock, and you have these little

56:54

point feet on the robot, and

56:56

the rock can roll. And

56:59

you have to deal with that last minute

57:01

change in your foot

57:03

placement. Yeah, so you step on the thing,

57:06

and that thing responds to you stepping on it. Yeah,

57:08

and it moves where your point of support is.

57:11

And so it's really... That became

57:14

kind of the essence of the test. And

57:16

so that was the beginning of us starting

57:18

to build rock piles in our

57:20

parking lots. And

57:23

we would actually build boxes full of rocks

57:25

and bring them into the lab. And then

57:27

we would have the robots walking across these boxes

57:29

of rocks because that became the

57:31

essential test.

57:33

So you mentioned Big Dog. Can

57:35

we maybe take a stroll through the history

57:38

about the dynamics? So what

57:41

and who is Big Dog? By the way, is who...

57:45

Do you try not to anthropomorphize

57:48

the robots? Do you try not to...

57:50

Do you try to remember that they're... This is like

57:52

the division I have, because for me it's impossible.

57:55

For me, there's a magic to the

57:58

being that is a robot. It is not human. But

58:00

it is the

58:03

same magic that

58:06

a living being has when it moves about the world

58:08

is there in the robot. So I

58:11

don't know what question I'm asking, but should

58:13

I say what or who, I guess. Who

58:15

is Big Dog? What is Big Dog? Well,

58:18

I'll say to address the meta question,

58:21

we don't try to draw hard lines around

58:23

it being an it or a him or a

58:25

her.

58:26

It's okay, right?

58:30

I think part of the magic of these kinds

58:32

of machines is by nature

58:35

of their organic movement,

58:37

of their dynamics, we

58:40

tend to want to identify

58:42

with them. We tend to look at them and attribute

58:46

maybe feeling to that because we've

58:49

only seen things that move like

58:51

this that were alive. And

58:54

so this is an opportunity.

58:56

It means that you could have

59:00

feelings for a machine

59:02

and people have feelings for their cars. They

59:05

get attracted to them, attached to them. So

59:07

that inherently

59:09

could be a good thing as long as we manage what

59:11

that interaction is. So

59:14

we don't put strong boundaries around

59:16

this and ultimately think it's a benefit,

59:19

but it also can be a

59:21

bit of a curse because I think people look

59:23

at these machines and

59:25

they attribute a level of intelligence that the machines

59:28

don't have. Why? Because again,

59:30

they've seen things move like this that were living

59:33

beings, which are intelligent.

59:36

And so they want to attribute intelligence to the robots

59:39

that isn't appropriate yet, even though they move

59:42

like an intelligent being. But you try

59:44

to acknowledge that the anthropomorphization

59:46

is there and try to,

59:49

first of all, acknowledge that it's there. And

59:52

have a little fun with it. You know, our most

59:55

recent video, it's just kind

59:57

of fun to...

59:59

Look at the robot. We started

1:00:01

off the video with Atlas

1:00:05

kind of looking around for where the bag

1:00:08

of tools was, because the guy up on the scaffolding

1:00:10

says, send me some tools. And Atlas

1:00:12

has to kind of look around and see where they

1:00:14

are. And there's a little personality

1:00:16

there. That

1:00:18

is fun, it's entertaining, it makes our jobs

1:00:20

interesting. And I think in the long run, can

1:00:23

enhance interaction between humans

1:00:25

and robots in a way that isn't

1:00:27

available to machines that don't move that

1:00:29

way. This is something to me personally, it's very

1:00:31

interesting.

1:00:34

I happen to have a lot of legged

1:00:36

robots. I

1:00:39

hope to have a lot of spots in

1:00:41

my possession. I'm

1:00:43

interested in celebrating robotics and celebrating

1:00:46

companies. And I also don't want to, companies

1:00:48

that do incredible stuff like Boston Dynamics. And

1:00:51

there's,

1:00:53

you know, I'm a little crazy. And

1:00:55

you say you don't want to, you

1:00:58

want to align, you want to help the company.

1:01:00

Because I ultimately want a company

1:01:02

that Boston Dynamics to succeed. And part

1:01:05

of that we'll talk about, success kind

1:01:07

of requires making money. And so

1:01:09

the kind of stuff I'm particularly

1:01:12

interested in may not be the thing

1:01:14

that makes money in the short term. I can make

1:01:16

an argument that it will in the long term. But the

1:01:18

kind of stuff I've been playing with is

1:01:21

a robust way of

1:01:23

having the quadrupas, the robot

1:01:26

dogs, communicate emotion with

1:01:28

their body movement. The same kind of stuff you

1:01:30

do with a dog. But not

1:01:32

hard coded,

1:01:33

but in a robust way. And

1:01:36

be able to communicate excitement or

1:01:38

fear, boredom, all this

1:01:40

kinds of stuff. And I think as

1:01:42

a base layer of

1:01:44

function of behavior to

1:01:47

add on top of a robot, I think that's a really powerful

1:01:49

way to make

1:01:52

the robot more usable for humans, for whatever

1:01:54

application. It's gonna be really important. And

1:01:57

it's a thing we're beginning to

1:01:59

pay attention to.

1:01:59

We really

1:02:02

want to start a differentiator for

1:02:04

the company has always been we really want

1:02:06

the robot to work. We want it to be useful.

1:02:11

Making it work at first meant the

1:02:13

luggage locomotion really works. It can really

1:02:16

get around and it doesn't fall down. But

1:02:20

beyond that, now it needs to be a useful

1:02:22

tool and our customers are,

1:02:25

for example, factory owners, people who

1:02:27

are running a process manufacturing

1:02:30

facility and the robot needs to be able to get

1:02:32

through this complex facility in a reliable

1:02:34

way, taking measurements.

1:02:38

We need for people

1:02:40

who are operating those robots to understand what

1:02:43

the robots are doing. If the robot

1:02:45

gets into needs help or

1:02:48

is in trouble or something, it needs

1:02:50

to be able to communicate and

1:02:52

a physical indication of

1:02:54

some sort

1:02:55

so that a person looks at the robot and

1:02:57

goes, oh, I know what that robot is doing. The

1:02:59

robot is going to go take measurements of

1:03:02

my vacuum pump with its thermal

1:03:04

camera. You

1:03:06

want to

1:03:07

be able to indicate that. We're

1:03:08

even just the

1:03:10

robot is about to turn in front

1:03:13

of you and maybe indicate that it's

1:03:15

going to turn and so you sort of see and

1:03:17

can anticipate its motion. This

1:03:20

kind of communication is going to become more and

1:03:22

more important. It wasn't sort of

1:03:24

our starting point,

1:03:26

but now that the robots are really

1:03:29

out in the world and we have about

1:03:31

a thousand of them out with customers right

1:03:33

now,

1:03:35

this layer of physical

1:03:38

indication I think is going to become more

1:03:40

and more important.

1:03:41

We'll talk about where it goes because

1:03:44

there's a lot of interesting possibilities, but if it

1:03:46

can return back to the origins of Boston

1:03:48

and dynamics with the more

1:03:50

research, the R&D side

1:03:52

before we talk about

1:03:54

how to build robots at scale.

1:03:56

It's a big dog. Who's a

1:03:58

big dog? So the

1:04:01

company started in 1992 and in probably 2003, I believe

1:04:03

is when we took a contract

1:04:06

from

1:04:15

dark, so basically 10

1:04:17

years, 11 years. We

1:04:19

weren't doing robotics. We did a little bit of robotics

1:04:22

with Sony. They had an IBO,

1:04:25

they're IBO robot. We were developing some software

1:04:27

for that that kind of got us a little bit involved

1:04:29

with robotics again. Then there's

1:04:31

this opportunity to do a DARPA contract

1:04:34

where they wanted to build a

1:04:37

robot dog. And

1:04:39

we won a contract to

1:04:41

build that. And so that was the genesis

1:04:44

of Big Dog. And it

1:04:46

was a quadruped. It was the first time we built

1:04:48

a robot that had everything on board that you

1:04:50

could actually take the robot out into the wild

1:04:53

and operate it. So it had an onboard power plant,

1:04:55

it had onboard computers, it

1:04:57

had hydraulic

1:04:58

actuators that

1:05:00

needed to be cooled. So we had cooling systems

1:05:02

built in. Everything integrated

1:05:04

into the robot.

1:05:06

And

1:05:07

that was a pretty rough start. It was 10

1:05:10

years that we were not a robotics

1:05:12

company, we were a simulation company. And then we had

1:05:14

to build a robot in about a year. So that

1:05:17

was a little bit of a rough transition. I mean,

1:05:20

can you

1:05:22

just comment on the roughness of that transition? Big

1:05:26

Dog, I mean, this is this big quadruped

1:05:31

four legs robot. We built

1:05:33

a few different versions of them. But the first one,

1:05:35

the very earliest ones, you know, didn't work very well.

1:05:38

We would take them out and it was

1:05:40

hard to get, you

1:05:43

know, a go-kart engine driving

1:05:45

a hydraulic. Oh, is that what it was? And,

1:05:48

you know, having that all work while

1:05:51

trying to get, you know, the

1:05:53

robot to stabilize itself. So what

1:05:56

was the power plant? What was the engine? It

1:05:58

seemed like my vague recollection.

1:05:59

I don't know,

1:06:03

it felt very loud and aggressive

1:06:06

and kind of thrown together. It absolutely

1:06:09

was, right? We weren't trying to

1:06:12

design the best robot hardware at the time. And

1:06:16

we wanted to buy an off-the-shelf engine. And so

1:06:19

many of the early versions of Big

1:06:21

Dog had literally go-kart

1:06:23

engines or something like that. Gas powered? Like

1:06:26

a gas powered two-stroke engine. And

1:06:29

the reason why it was two-stroke

1:06:29

is two-stroke engines are lighter weight.

1:06:33

But they're also, and we generally

1:06:35

didn't put mufflers on them because we're trying to save the weight.

1:06:38

We didn't care about the noise. And some of these

1:06:40

things were horribly loud. But

1:06:42

we're trying to manage weight because managing

1:06:44

weight in a legged robot is always important

1:06:47

because it has to carry everything. That

1:06:49

said that thing was big. Well, I've

1:06:52

seen the videos of it. I mean, the

1:06:54

early versions stood about, I

1:06:56

don't know, belly high, chest high.

1:06:58

They probably weighed

1:06:59

maybe a couple of hundred

1:07:02

pounds. But over

1:07:04

the course of probably five years,

1:07:08

we were able to get that robot

1:07:12

to really manage a remarkable

1:07:15

level of rough terrain. So we started

1:07:17

out with just walking on the flat. And then we started walking

1:07:19

on rocks and then inclines and then mud

1:07:22

and slippery mud. And

1:07:24

by the end of that program,

1:07:26

we were convinced that legged

1:07:29

locomotion in a robot could actually

1:07:31

work because going into it, we

1:07:34

didn't know that. We had built quadrupeds

1:07:36

at MIT, but

1:07:38

they used a giant hydraulic

1:07:41

pump in the lab. They use a giant

1:07:43

computer that was in the lab. They're always tethered

1:07:45

to the lab. This was the

1:07:47

first time something that was self-contained,

1:07:51

walked around in the world and

1:07:54

balanced. And the purpose was

1:07:56

to prove to ourself that the legged locomotion

1:07:58

could really work.

1:07:59

Big Dog really cut that open for

1:08:02

us. And it was the beginning of what

1:08:04

became a whole series of robots. So once

1:08:06

we showed to DARPA that you could make a legged

1:08:09

robot that could work, there was a

1:08:11

period at DARPA where robotics got really

1:08:13

hot and there was lots of different programs.

1:08:16

And we were able to build other

1:08:18

robots. We built other quadrupeds to

1:08:21

hand like LS3,

1:08:23

designed to carry heavy loads.

1:08:25

We built Cheetah, which was designed

1:08:28

to explore what are the limits to how fast

1:08:30

you can run. We began

1:08:32

to build sort of a portfolio of machines

1:08:37

and software that let us

1:08:39

build not just one robot, but a whole

1:08:41

family of robots. To push the limits in all kinds

1:08:43

of directions. Yeah, and to discover those principles.

1:08:46

You know, you asked earlier about the art and science

1:08:48

of a legged locomotion. We

1:08:50

were able to develop principles of legged locomotion

1:08:53

so that we knew how to build a

1:08:55

small legged robot or a big one. Leg

1:08:58

length, you know, was now a parameter

1:09:00

that we could play with. Payload

1:09:02

was a parameter we could

1:09:04

play with. So we built the LS3, which

1:09:06

was an 800 pound robot designed to carry

1:09:08

a 400 pound payload. And

1:09:10

we learned the design rules, basically developed

1:09:13

the design rules. How do you scale

1:09:15

different robot systems to, you know,

1:09:18

their terrain, to their walking speed,

1:09:21

to their payload?

1:09:22

So when

1:09:24

was Spot born?

1:09:27

Around 2012 or so. So

1:09:33

again, almost 10 years into sort of a run

1:09:35

with DARPA where we built a bunch

1:09:37

of different quadrupeds. We had sort of a different

1:09:40

thread where we started building humanoids. We

1:09:45

saw that probably an end was coming

1:09:48

where the government was gonna kind of back

1:09:50

off from a lot of robotics investment. And

1:09:55

in order to maintain progress,

1:09:57

we just deduced. that,

1:10:00

well, we probably need to sell ourselves to somebody

1:10:02

who wants to continue to invest in this

1:10:04

area. And that was Google. And

1:10:07

so

1:10:08

at Google,

1:10:09

we would meet regularly with Larry Page

1:10:12

and Larry just started asking us, you know, what's

1:10:15

your product going to be? And you

1:10:17

know, the logical

1:10:19

thing, the thing that we had the most history

1:10:21

with that we wanted to continue

1:10:23

developing was our quadruped,

1:10:26

but we knew it needed to be smaller. We knew it couldn't have

1:10:28

a gas engine. We thought it probably

1:10:31

couldn't be hydraulically actuated. So

1:10:33

that began the process of

1:10:36

exploring if we could migrate to

1:10:38

a smaller electrically actuated robot.

1:10:42

And that was really the genesis of SPOT.

1:10:45

So not a gas engine and

1:10:47

the actuators are electric. Yes. So

1:10:50

can you maybe comment on what it's like

1:10:52

at Google with working

1:10:54

with Larry Page, having those meetings

1:10:57

and thinking of what will a robot look

1:10:59

like

1:11:00

that could be built

1:11:02

at scale? What like starting to think

1:11:04

about a product? Larry

1:11:07

always liked the toothbrush

1:11:09

test. He wanted products that you used every

1:11:11

day.

1:11:15

What they really wanted was, you know,

1:11:18

a consumer level product,

1:11:20

something that would work in your house.

1:11:22

We

1:11:24

didn't think that was the right next thing to do

1:11:27

because to be a consumer level product

1:11:29

cost is going to be very important.

1:11:32

Probably needed to cost a few thousand

1:11:34

dollars. And we were

1:11:36

building these machines that cost hundreds of thousands

1:11:38

of dollars, maybe a million dollars to build. Of

1:11:41

course, we were only building a

1:11:42

two,

1:11:44

but we didn't see how to get all the way to this

1:11:46

consumer level product in a short

1:11:48

amount of time. And

1:11:51

he

1:11:51

suggested that we make the

1:11:53

robots really inexpensive. And

1:11:56

part of our philosophy has always been build

1:11:59

the best hardware. you can. Make

1:12:02

the machine operate well

1:12:05

so that you're trying to

1:12:07

solve, you know, discover

1:12:10

the hard problem that you don't know about. Don't

1:12:13

make it harder by building a crappy machine, basically.

1:12:15

Build the best machine you can. There's

1:12:18

plenty of hard problems to solve that are going to have to do

1:12:20

with, you know, under actuated systems and

1:12:22

balance. And so we wanted

1:12:24

to build these high quality machines still.

1:12:26

And we thought that was important for us

1:12:28

to continue learning about really

1:12:31

what was the important parts

1:12:33

of that make robots work.

1:12:35

And so there was a

1:12:37

little bit of a philosophical difference

1:12:40

there. And so ultimately,

1:12:42

that's why we're building robots for the industrial

1:12:45

sector now. Because the industry

1:12:48

can afford a more expensive machine because,

1:12:50

you know, their productivity depends

1:12:53

on keeping their factory going. And so if

1:12:56

spot costs, you know, $100,000 or

1:12:59

more, that's not such a big expense

1:13:01

to them. Whereas at the consumer level,

1:13:04

no one's going to buy a robot like that.

1:13:06

And I

1:13:07

think we might eventually get to a consumer level

1:13:09

product that will be that cheap. But

1:13:11

I think the path to getting there needs to

1:13:13

go through these really nice machines. So

1:13:16

we can then learn how to simplify. So

1:13:18

what can you say to the almost the

1:13:21

engineering challenge of

1:13:23

bringing down cost of

1:13:26

a robot? So that

1:13:28

presumably when you try to build the robot at scale,

1:13:30

that also comes into play when you're trying to make

1:13:32

money on a robot, even in the industrial

1:13:35

setting. But how interesting,

1:13:37

how challenging of

1:13:39

a thing is

1:13:41

that? In particular,

1:13:43

probably new to an R&D company.

1:13:46

Yeah, I'm glad you brought that last part up. The

1:13:48

transition from an R&D company to a commercial

1:13:51

company, that's the thing you worry

1:13:53

about, you know, because you've got these engineers who

1:13:55

love hard problems, who want to figure out how to

1:13:57

make robots work. And you don't know

1:13:59

if you

1:13:59

have engineers that want to work on

1:14:02

the quality and reliability and cost

1:14:04

that is ultimately required.

1:14:07

And indeed, we have brought on a lot

1:14:10

of new people who are inspired by those problems,

1:14:13

but the big takeaway lesson for me is

1:14:16

we have good people. We have engineers who

1:14:19

want to solve problems, and

1:14:21

the quality and cost and manufacturability

1:14:24

is just another kind of problem. And

1:14:26

because they're so invested in what

1:14:28

we're doing,

1:14:29

they're interested in and will go work on

1:14:32

those problems as well. And

1:14:34

so I think we're managing that transition

1:14:37

very well. In fact, I'm really pleased that,

1:14:40

I mean, it's

1:14:42

a huge undertaking, by the way, right? So

1:14:45

even having to

1:14:47

get reliability to where it needs to be, we

1:14:49

have to have fleets of robots that we're just

1:14:51

operating 24-7 in our offices

1:14:54

to go find those rare failures and

1:14:57

eliminate them. It's just a totally different

1:14:59

kind of activity than the research activity where

1:15:01

you get it to work, the one robot you

1:15:04

have

1:15:04

to work in a repeatable

1:15:06

way at the high stakes

1:15:09

demo. It's just very different. But

1:15:12

I think we're making remarkable progress, I

1:15:14

guess. So one of the cool things I got

1:15:16

a chance to visit Boston Dynamics,

1:15:18

and I mean, one

1:15:20

of

1:15:23

the things that's really cool is to see

1:15:25

a large number of robots moving about.

1:15:28

Because I think one of the things you notice

1:15:31

in the research environment

1:15:33

at MIT, for example, I don't think

1:15:35

anyone ever has a working robot for a prolonged

1:15:38

period of time. Exactly. So

1:15:40

most robots are just sitting there in a sad

1:15:44

state of despair waiting to be born,

1:15:46

brought to life for a brief moment of time. Just

1:15:49

to have, I just remember

1:15:52

there's a Spot robot, I

1:15:54

had a cowboy hat on, and it was just walking randomly

1:15:57

for whatever reason. I don't

1:15:58

even know. But there's a kind of... a

1:16:00

sense of sentience to

1:16:03

it because it doesn't seem like anybody was supervising

1:16:05

it. It was just doing its thing. I'm going to stop

1:16:07

way short of the sentience. Sure. It

1:16:10

is the case that if you come to our office today

1:16:12

and walk around the hallways, you're

1:16:15

going to see a dozen robots just

1:16:17

walking around all the time.

1:16:21

That's really a reliability test

1:16:23

for us. We have these robots programmed

1:16:26

to do autonomous

1:16:27

missions, get up off their charging

1:16:29

dock, walk around the building, collect data

1:16:31

at a few different places, and go sit back down. We

1:16:34

want that to be a very reliable process

1:16:36

because that's what somebody who's running

1:16:39

a brewery, a factory, that's

1:16:42

what they need the robot to do. We

1:16:45

have to dog food our own robot. We have to test

1:16:47

it in that way.

1:16:50

On a weekly basis, we

1:16:52

have robots that are accruing something like 1,500

1:16:55

or maybe 2,000 kilometers

1:16:58

of walking and over 1,000 hours

1:17:00

of operation every week. That's

1:17:05

something that almost I don't think anybody else in the world

1:17:07

can do because, hey, you have to have a fleet of robots

1:17:09

to just accrue that much information. You

1:17:12

have to be willing to dedicate it to that

1:17:15

test. That's

1:17:18

essential. That's how you get the reliability. That's

1:17:20

how you get it. What about some of the cost cutting

1:17:23

from the manufacturer side? What have you

1:17:25

learned

1:17:25

from the manufacturer side of the transition

1:17:28

from R&D? We're

1:17:30

still learning a lot there. We're

1:17:32

learning how to cast parts instead

1:17:35

of mill it all out of billet

1:17:37

aluminum. We're

1:17:39

learning how to get plastic molded parts.

1:17:42

We're learning about how to control that

1:17:44

process so that you can build the same robot

1:17:47

twice in a row. There's a lot to learn

1:17:49

there, and we're only partway through that process.

1:17:53

We've set up a manufacturing facility in

1:17:56

Waltham. It's about a mile from

1:17:58

our headquarters. We're

1:18:00

doing final assembly and test of both spots

1:18:02

and stretches at that factory.

1:18:07

It's hard because to be honest,

1:18:10

we're still iterating on the design of the robot. As

1:18:12

we find failures from these reliability

1:18:14

tests, we need to go engineer changes. Those

1:18:17

changes need to now be propagated to the

1:18:19

manufacturing line. That's a hard process,

1:18:22

especially when you want to move as fast as we do. It's

1:18:26

been challenging and it makes

1:18:29

it

1:18:29

the folks who are working supply chain who

1:18:32

are trying to get the cheapest parts for

1:18:34

us, it requires that you buy

1:18:36

a lot of them to make them cheap. Then we

1:18:38

go change the design from underneath them. They're like,

1:18:40

what are you doing? Getting

1:18:42

everybody on the same page here, that

1:18:45

we still need to move fast, but we also need to

1:18:47

try to figure out how to reduce costs. That's

1:18:50

one of the challenges of this migration

1:18:52

we're going through. Over the past few years,

1:18:54

challenges to the supply chain. I

1:18:57

imagine you've been a part of a bunch of stressful

1:18:59

meetings.

1:18:59

Things got more expensive and

1:19:02

harder to get. It's

1:19:05

all been a challenge. Is there still room for simplification?

1:19:07

Oh yeah, much more. These

1:19:10

are really just the first generation of these machines.

1:19:13

We're already thinking about what the next generation of spots

1:19:16

going to look like.

1:19:17

Spot was built as a platform. You

1:19:20

could put almost any sensor on it. We provided

1:19:23

data communications, mechanical

1:19:25

connections, power connections.

1:19:30

For example, in the applications that we're

1:19:32

excited about where you're monitoring

1:19:34

these factories for their health,

1:19:37

there's probably a simpler machine

1:19:39

that we could build that's really focused

1:19:42

on that use case. That's

1:19:45

the difference between the general purpose

1:19:47

machine or the platform versus the

1:19:49

purpose built machine. Even

1:19:52

in the factory, we'd still like the robot to do

1:19:54

lots of different tasks. If we

1:19:56

really knew on day one that we're going to be operating

1:19:59

in a factory with

1:19:59

these three sensors in it, we would have

1:20:02

it all integrated in a package that would be easier,

1:20:04

more, less expensive, and more

1:20:06

reliable. So we're contemplating

1:20:09

building, you know, a next generation of that machine.

1:20:11

So we should mention that, so SPOT,

1:20:14

for people who are somehow not familiar, so

1:20:17

it's a yellow robotic

1:20:19

dog, and

1:20:22

has

1:20:23

been featured in many dance videos. It

1:20:26

also has gained an arm. So

1:20:29

what can you say about the arm that SPOT has?

1:20:32

About the challenges of this design, and

1:20:34

the manufacture of it?

1:20:36

We think the future of mobile

1:20:38

robots is mobile manipulation.

1:20:41

That's where, you know,

1:20:43

in the past 10 years,

1:20:45

it was getting mobility to work, getting the leg of

1:20:47

locomotion to work. If you ask, what's the

1:20:49

heart problem in the next 10 years? It's

1:20:52

getting a mobile robot to do useful manipulation

1:20:55

for you.

1:20:55

And so we wanted SPOT to have an arm

1:20:59

to experiment with those problems.

1:21:03

And the arm is

1:21:06

almost as complex as the robot itself,

1:21:09

you know, and it's

1:21:11

an attachable payload. It

1:21:14

has, you know, several motors and

1:21:16

actuators and sensors. It has a camera

1:21:19

in the end of its hand, so, you know, you can

1:21:21

sort of see something,

1:21:24

and the robot will control the

1:21:26

motion of its hand to go pick it up autonomously.

1:21:28

So in the same way the robot walks and balances,

1:21:32

managing its own foot placement to stay balanced,

1:21:34

we want manipulation to be mostly

1:21:37

autonomous, where the robot, you indicate, okay,

1:21:39

go grab that bottle, and then the robot will just go

1:21:41

do it using the camera in its hand,

1:21:44

and then sort of closing in on the

1:21:46

grasp. But it's

1:21:49

a whole nother complex robot on top of a

1:21:51

complex-legged robot,

1:21:54

and of course we made the hand

1:21:56

look a little like a head, you

1:21:59

know, because. because again, we want it

1:22:01

to be sort of identifiable. In

1:22:03

the last year, a lot

1:22:05

of our sales have been people who already

1:22:08

have a robot now buying an arm to add to

1:22:10

that robot. Oh,

1:22:12

interesting. And so the arm

1:22:14

is for sale. Oh yeah, oh yeah. It's

1:22:16

an option. What's the interface

1:22:19

like to work with the arm?

1:22:21

Like is it pretty, so are

1:22:23

they designed primarily, I guess

1:22:25

just ask that question in general about robots

1:22:28

from Boston Dynamics. Is it designed to

1:22:31

be

1:22:32

easily and efficiently

1:22:34

operated remotely by a human being or

1:22:37

is there also the capability to

1:22:39

push towards autonomy?

1:22:41

We want both.

1:22:43

In the next version of

1:22:45

the software that we release, which

1:22:47

will be version 3.3, we're going

1:22:50

to offer the ability of, if

1:22:52

you have an autonomous mission for the robot, we're

1:22:55

going to include the option that it can go

1:22:57

through a door, which means it's going to have to have an arm

1:22:59

and it's going to have to use that arm to open the door. And

1:23:02

so that'll be an autonomous manipulation

1:23:04

task that just, you can program

1:23:07

easily with the robot strictly

1:23:10

through, we have a tablet interface. And

1:23:13

so

1:23:13

on the tablet, you sort of see the

1:23:15

view that Spot sees, you say,

1:23:17

there's the door handle, the hinges

1:23:20

are on the left and it opens in, the rest is

1:23:22

up to you. Take care of it. So

1:23:24

it just takes care of everything. Yeah. So

1:23:27

we want, and for a task

1:23:29

like opening doors, you can automate

1:23:31

most of that. And we've automated a few other tasks.

1:23:34

We had a customer who

1:23:36

had a high powered

1:23:39

breaker switch, essentially. It's an

1:23:41

electric utility, Ontario

1:23:43

power generation.

1:23:45

And they have to, when they're

1:23:47

going to disconnect their power supply,

1:23:50

that could be a gas generator, could be a nuclear

1:23:52

power plant. From the grid, you have

1:23:54

to disconnect this breaker switch. Well,

1:23:56

as you can imagine, there's hundreds

1:23:59

or thousands of... amps and volts involved

1:24:01

in this breaker switch. And it's a dangerous

1:24:04

event, because occasionally you'll get what's called an

1:24:06

arc flash. As you just do this disconnect,

1:24:09

the power, the sparks jump across

1:24:11

and people die doing this. And

1:24:14

so Ontario Power Generation

1:24:16

used our spot in the arm

1:24:19

through the interface

1:24:21

to operate this disconnect in

1:24:25

an interactive way. And they showed

1:24:27

it to us. And we were so excited

1:24:30

about it and said, you know, I bet we can automate

1:24:32

that task. And so we got some

1:24:34

examples of that breaker switch. And

1:24:37

I believe in the next generation of the software, now

1:24:39

we're gonna deliver back to Ontario Power Generation.

1:24:42

They're gonna be able to just point the robot at

1:24:45

that breaker. They'll be out, they'll indicate

1:24:48

that's the switch. There's sort of two actions

1:24:50

you have to do. You have to flip up this

1:24:52

little cover, press a button, then

1:24:54

get a ratchet, stick it into

1:24:57

a socket and literally

1:24:59

unscrew this giant breaker

1:25:01

switch. So there's a bunch of different tasks. And

1:25:04

we basically automated them so that the human

1:25:06

says, okay, there's the switch, go

1:25:08

do that part.

1:25:10

That right there is the socket where

1:25:12

you're gonna put your tool and you're gonna open it up.

1:25:15

And so you can remotely sort of indicate this on

1:25:17

the tablet and then the robot

1:25:19

just does everything in between. And it does

1:25:21

everything, all the coordinated movement of all the different

1:25:24

actuators that includes the body. It maintains

1:25:26

its balance, it walks itself

1:25:29

into position. So it's within reach

1:25:31

and the arm is in a position where

1:25:33

it can do the whole task. So it manages

1:25:36

the whole body. So how

1:25:39

does one become a big

1:25:40

enough customer to request features?

1:25:42

Cause I personally want a

1:25:44

robot that gets me a beer. I

1:25:47

mean, that has to be like one of the most

1:25:49

requests, I suppose in the industrial setting.

1:25:51

That's a non-alcoholic

1:25:53

beverage.

1:25:56

Of picking up objects and bringing the objects

1:25:58

to you. We love working with. customers

1:26:00

who have challenging problems like this. And

1:26:03

this one in particular, because we felt like

1:26:06

what they were doing, A, it was a safety feature.

1:26:08

B, we saw that

1:26:10

the robot could do it because

1:26:12

they teleoperated it the first time. Probably took

1:26:14

them an hour to do it the first time, right? But

1:26:17

the robot was clearly capable. And

1:26:19

we thought, oh, this is a great problem for us

1:26:21

to work on to figure out how to automate

1:26:23

a manipulation task. And so we took it on,

1:26:26

not because we were going to make a bunch of money from

1:26:28

it in selling the robot back to them, but

1:26:30

because it motivated us to go solve

1:26:33

what we saw as the next logical step.

1:26:36

But many of our customers, in fact,

1:26:38

we try to, our bigger

1:26:41

customers, typically ones who are

1:26:43

going to run a utility or a factory or something like

1:26:45

that,

1:26:46

we take that kind of direction from them. And

1:26:48

if they're, especially if they're going to buy 10 or 20

1:26:50

or 30 robots, and they say, I really needed

1:26:53

to do this. Well, that's exactly the right

1:26:55

kind of problem that we want to be working on. Note

1:26:58

to self, buy 10 spots and

1:27:02

aggressively push for beer manipulation.

1:27:05

I think it's fair to say it's notoriously difficult to

1:27:08

make a lot of money as a robotics company.

1:27:11

How can you

1:27:12

make money as a robotics company?

1:27:15

Can you speak to that? It seems that a

1:27:17

lot of robotics companies fail.

1:27:20

It's difficult to build robots. It's

1:27:23

difficult to build robots at a low

1:27:25

enough cost where customers,

1:27:27

even the industrial setting, want to purchase them. And it's difficult

1:27:30

to build robots that are useful,

1:27:32

sufficiently useful. So what can you speak

1:27:34

to? And Boston Dynamics has been

1:27:37

successful for many years

1:27:39

of finding a way to make money. Well, in

1:27:42

the early days, of course, you know, the money we

1:27:44

made was from doing contract R&D work. And

1:27:47

we made money, but

1:27:49

we weren't growing and we weren't selling

1:27:51

a product. And then we went

1:27:53

through several owners who had a vision

1:27:56

of not only doing the product, but also the developing

1:28:00

advanced technology, but eventually developing

1:28:02

products.

1:28:03

And so both Google and SoftBank

1:28:06

and now Hyundai had that

1:28:08

vision and were willing to

1:28:11

provide that investment.

1:28:16

Now our discipline is that we need

1:28:18

to go find applications that

1:28:20

are broad enough that you could

1:28:22

imagine selling thousands of robots. Because

1:28:24

it doesn't work if you don't sell thousands or tens

1:28:27

of thousands of robots. If you only sell hundreds,

1:28:30

you will commercially fail. And that's where

1:28:32

most of the small robot companies have died.

1:28:38

And that's a challenge because

1:28:40

A, you need to field the

1:28:42

robots, they need to start to become reliable.

1:28:45

And as we've said, that takes time and investment

1:28:48

to get there.

1:28:49

And so it really does take visionary

1:28:52

investment to get there. But we

1:28:54

believe that we are going to make money

1:28:57

in this industrial

1:29:00

monitoring space. Because

1:29:02

if

1:29:03

a chip fab,

1:29:04

if the line goes down because

1:29:08

a vacuum pump failed someplace, that

1:29:10

can be a very expensive process. It can be

1:29:12

a million dollars a day in lost production.

1:29:15

Maybe you have to throw away some of the product along

1:29:17

the way. And so the robot,

1:29:20

if you can prevent that by inspecting

1:29:23

the factory every single day,

1:29:25

maybe every hour if you have to, there's

1:29:28

a real return on investment there. But

1:29:30

there needs to be a critical mass

1:29:32

of this task. And we're focusing

1:29:35

on a few that we believe

1:29:37

are ubiquitous

1:29:39

in the industrial

1:29:41

production environment. And that's using

1:29:44

a thermal camera to

1:29:46

keep things from overheating, using an

1:29:48

acoustic imager to find compressed

1:29:50

air leaks, using visual cameras

1:29:53

to read gauges, measuring

1:29:55

vibration. These are standard things

1:29:57

that you do to prevent

1:29:59

intended shutdown of a factory. And

1:30:03

this takes place in

1:30:05

a beer factory. We're working with AB

1:30:07

InBev. It takes place in chip fabs.

1:30:10

We're working with global foundries. It

1:30:12

takes place in electric utilities

1:30:14

and nuclear power plants. And so the same

1:30:17

robot

1:30:18

can be applied in all of these industries.

1:30:21

And

1:30:22

as I said, we have about, actually,

1:30:25

it's 1,100 spots out now. To

1:30:27

really get profitability, we

1:30:29

need to be at 1,000 a year, maybe 1,500 a year for that sort of part

1:30:32

of the business. So it still needs

1:30:37

to grow, but

1:30:39

we're on a good path. So I think that's totally

1:30:41

achievable. So the application should require

1:30:44

crossing that 1,000 robot barrier. It

1:30:46

really should.

1:30:47

Yeah. I want to mention our second

1:30:50

robot, Stretch. Yeah. Tell

1:30:52

me about Stretch. What's Stretch? Who's Stretch?

1:30:54

Stretch started differently than Spot.

1:30:57

Spot, we built because we had decades

1:30:59

of experience building quadrupeds. We

1:31:02

had it in our blood. We had to build a quadruped

1:31:04

product. But we had to go figure out what the application

1:31:07

was. And we actually discovered this

1:31:09

factory

1:31:11

patrol application, basically

1:31:14

preventative maintenance, by seeing what

1:31:16

our customers did with it.

1:31:18

Stretch was very different. We started knowing

1:31:20

that there was warehouses

1:31:22

all over the world. There's shipping

1:31:25

containers moving all around the

1:31:27

world full of boxes that are mostly being

1:31:29

moved by hand.

1:31:31

By some estimates, we think there's a trillion boxes,

1:31:34

cardboard boxes shipped around the world

1:31:36

each year, and a lot of it's done manually. It

1:31:39

became clear early on

1:31:41

that there was an opportunity for a mobile robot

1:31:43

in here to move boxes around. And

1:31:46

the commercial experience has been very different between

1:31:49

Stretch and with Spot. As

1:31:51

soon as we started talking to

1:31:53

people, potential customers, about

1:31:56

what Stretch was going to be used for, they immediately

1:31:58

started saying, oh, I'll buy. I'll buy.

1:31:59

that robot. In fact, I'm going to

1:32:02

put in an order for 20 right now.

1:32:04

We just started shipping the robot in January

1:32:08

after several years of development. This

1:32:10

year. This year. So our first deliveries

1:32:12

of stretch to customers were DHL

1:32:14

and Marisk in January.

1:32:16

We're delivering the gap right now.

1:32:19

And we have about seven or eight other customers,

1:32:22

all who've already agreed in advance to

1:32:24

buy between 10 and 20 robots. And so we've

1:32:26

already got commitments for a couple of hundred

1:32:28

of these robots.

1:32:30

This one's going to go right. It's so obvious

1:32:33

that there's a need and we're not

1:32:35

just going to unload trucks. We're going to do any box

1:32:37

moving task in the warehouse. And so it too will

1:32:39

be a multi-purpose robot and

1:32:42

we'll eventually have it doing palletizing

1:32:44

or depalletizing or loading

1:32:47

trucks or unloading trucks.

1:32:49

There's definitely thousands of robots. There's probably

1:32:51

tens of thousands of robots of this in

1:32:53

the future. So it's going to be profitable. Can

1:32:56

you describe what stretch looks like? It

1:32:58

looks like a big strong

1:33:01

robot arm on a mobile base. The base

1:33:03

is about the size of a pallet. And

1:33:05

we wanted it to be the size of a pallet because that's what

1:33:08

lives in warehouses, right? Pallets of goods sitting

1:33:10

everywhere. So we needed to be able to fit in that space.

1:33:13

It's not a legged mobile. It's not a legged robot. And

1:33:15

so it was our first,

1:33:18

it was actually a

1:33:20

bit of a

1:33:23

commitment from us, a challenge for us to

1:33:25

build a non-balancing robot. To

1:33:28

do the much easier problem and

1:33:31

to put it to a well. Well, because it wasn't going

1:33:33

to have this balance problem. And in fact,

1:33:36

the very first version of the

1:33:38

logistics robot we built was a balancing

1:33:40

robot and that's called a handle. And

1:33:43

there's that thing was epic. All right. It's a beautiful

1:33:45

machine. It's an incredible machine.

1:33:51

I mean, it looks epic. It looks like

1:33:53

a out of a, I

1:33:55

mean, the sci-fi movie of some

1:33:57

sort. I mean, just can you actually just

1:33:59

linger on the the design of that thing, because

1:34:01

that's another leap into something you probably haven't

1:34:03

done. It's a different kind of balancing. Yeah. So

1:34:06

let me, I love talking about the history of how a

1:34:08

handle came about because it

1:34:10

connects all of our robots actually. So

1:34:13

I'm going

1:34:15

to start with Atlas. When we

1:34:17

had Atlas getting fairly far along,

1:34:20

we wanted to understand, I was telling you earlier, the challenge

1:34:23

of the human form is that you have this mass

1:34:25

up high. And

1:34:27

balancing that

1:34:29

inertia, that mass up high is

1:34:32

its own unique challenge. And so we started

1:34:34

trying to get Atlas to balance standing

1:34:36

on one foot, like on a balance beam, using

1:34:39

its arms like this. And you know, you can do this,

1:34:41

I'm sure I can do this, right? Like if you're walking

1:34:43

a tightrope,

1:34:45

how do you do that balance?

1:34:47

So that's sort of controlling

1:34:49

the inertia, controlling the momentum of the

1:34:51

robot. We were starting to figure

1:34:53

that out on Atlas.

1:34:55

And so our first concept of

1:34:58

handle, which was a robot that was going to be on two

1:35:00

wheels, so it had the balance, but

1:35:02

it was going to have a big long arm so it could reach

1:35:05

a box at the top of a truck. And

1:35:08

it needed yet another

1:35:10

counterbalance,

1:35:12

a big tail to help

1:35:14

it balance while it was using its

1:35:16

arm. So the

1:35:18

reason why this robot sort of

1:35:20

looks epic, some people said it looked

1:35:22

like an ostrich or

1:35:25

maybe an ostrich moving around, was

1:35:29

the wheels, the legs, it has legs so

1:35:31

it can extend its legs. So

1:35:33

it's wheels on legs. We always wanted to build

1:35:35

wheels on legs. It had a tail and had this

1:35:38

arm and they're all moving simultaneously.

1:35:40

And in coordination to maintain balance because

1:35:42

we had figured out the mathematics of doing

1:35:44

this momentum control, how to maintain

1:35:47

that balance. And so part of the reason

1:35:50

why we

1:35:50

built this two-legged robot was

1:35:52

we had figured this thing out. We

1:35:54

wanted to see it in this kind of machine.

1:35:57

And we thought maybe this kind of machine would be good in

1:35:59

a warehouse. we built it. And it's a beautiful

1:36:01

machine. It moves in a graceful way,

1:36:03

like nothing else we've built, but

1:36:06

it wasn't the right machine for a

1:36:08

logistics application. We decided it was

1:36:10

too slow and couldn't pick

1:36:12

boxes fast enough, basically. And

1:36:15

it was doing beautifully with elegance. It

1:36:17

just wasn't efficient enough. So we

1:36:20

let it go. But

1:36:22

I think we will come back to that machine

1:36:24

eventually. The fact that it's possible, the

1:36:26

fact that you show that you could do so many things

1:36:29

at the same time in coordination and

1:36:31

so beautifully, there's something there. That

1:36:34

was a demonstration of what is possible.

1:36:36

Basically we made a hard decision and this was

1:36:38

really kind of a hard nose business decision.

1:36:41

It was, it was, it indicated us

1:36:44

not doing it just for the beauty

1:36:46

of the mathematics or the curiosity,

1:36:48

but no, we actually need to build a business that

1:36:51

can make money in the long run. And

1:36:53

so we ended up building stretch, which

1:36:55

has a big heavy base with a giant battery

1:36:57

in the base of it that allows

1:36:59

it to run for two,

1:37:01

two shifts, 16 hours worth of operation.

1:37:04

And that big battery is sort

1:37:07

of helps it stay balanced, right? So you can move

1:37:09

a 50 pound box around with its arm and not tip

1:37:11

over. Um, it's

1:37:14

omni-directional. It can move in any direction. So

1:37:16

it has a nice suspension built

1:37:18

into it. So it can deal with

1:37:20

gaps or things on the floor and

1:37:22

roll over it. But it's a, but

1:37:24

it's not a balancing robot. It's a mobile

1:37:26

robot arm that can work to

1:37:29

carry a pick or place a box

1:37:31

up to 50 pounds anywhere in the warehouse.

1:37:34

Take a box from point A to point

1:37:36

B anywhere. Palletize

1:37:38

depalletize. We're starting with unloading

1:37:41

trucks because there's so many trucks

1:37:43

and containers that where goods are shipped

1:37:45

and it's a brutal job. You know, in the summer, it

1:37:48

can be 120 degrees inside that container.

1:37:50

People don't want to do that job. Um,

1:37:54

and it's backbreaking labor, right? Again, these

1:37:56

can be up to 50 pound boxes. Um,

1:37:59

and so.

1:38:01

We feel like this is a productivity enhancer. And

1:38:04

for the people who used to do that job unloading

1:38:06

trucks, they're actually operating

1:38:08

the robot now. And so by building

1:38:10

robots that are easy to control,

1:38:14

and it doesn't take an advanced degree to manage,

1:38:17

you can become a robot operator. And so as

1:38:19

we've introduced these robots to both DHL

1:38:21

and Marisk and GAP, the warehouse

1:38:23

workers who were doing that manual labor

1:38:26

are now the robot operators. And so we see this

1:38:28

as ultimately a benefit to them as

1:38:30

well.

1:38:32

Can you say how much stretch costs?

1:38:35

Not

1:38:37

yet. But I will say

1:38:40

that when we engage

1:38:42

with our customers, they'll be able

1:38:44

to see a return on investment in

1:38:46

typically two years. Okay, so that's

1:38:48

something that you're constantly thinking about how. And

1:38:51

I suppose you have to do the same kind of thinking with spot.

1:38:53

So it seems like with stretch, the

1:38:56

application is like directly

1:38:58

obvious. Yeah, it's a slam dunk. Yeah, and

1:39:01

so you have a little more flexibility.

1:39:03

Well, I think we know the target. We know

1:39:05

what we're going after. And with

1:39:07

spot, it took us a while to figure out what we were going after.

1:39:10

Well, let me return to that question about

1:39:14

maybe the conversation you were having a

1:39:17

while ago with Larry Page, maybe

1:39:19

looking to the longer future

1:39:21

of social robotics, of

1:39:24

using spot to connect with human

1:39:26

beings, perhaps in the home. Do you see a future

1:39:28

there? If we were to sort of hypothesize

1:39:32

or dream about a future where a spot like

1:39:34

robots are in the home as pets, a social

1:39:36

robot? We definitely think about it. And

1:39:39

we would like to get there. We

1:39:41

think the pathway to getting there is

1:39:44

likely through these industrial applications

1:39:47

and then mass manufacturing. Let's figure

1:39:49

out

1:39:50

how to build the robots, how

1:39:52

to make the software so that they can really do a broad

1:39:55

set of skills. That's gonna take

1:39:58

real investment

1:39:59

to get there.

1:39:59

Performance first, right? The principle

1:40:02

of the company has always been, really make the

1:40:04

robots do useful stuff. And

1:40:06

so, you know,

1:40:08

the social robot companies that

1:40:11

tried to start someplace else by just

1:40:13

making acute interaction, mostly they

1:40:15

haven't survived. And

1:40:18

so

1:40:18

we think the utility really

1:40:21

needs to come first. And that means you

1:40:23

have to solve some of these hard problems.

1:40:26

And so to get there,

1:40:29

we're gonna go through the design and

1:40:32

software development in industrial, and

1:40:34

then that's eventually gonna let you reach a scale

1:40:36

that could then be addressed to a commercial consumer

1:40:39

level market. And so, yeah,

1:40:42

maybe we'll be able to build a smaller spot

1:40:44

with an arm that could really go get your beer for you.

1:40:48

But there's

1:40:48

things we need to figure out still. How

1:40:51

to safely, really safely. If

1:40:53

you're gonna be interacting with children,

1:40:56

you better be safe. And right

1:40:59

now, we count on a little bit

1:41:01

of standoff distance between the robot and people

1:41:03

so that you don't pinch your finger in the robot. So

1:41:06

you've got a lot of things you need to go solve before

1:41:08

you jump to that consumer level product.

1:41:11

Well, there's a kind of trade off in safety because

1:41:14

it feels like in the home, you

1:41:16

can fall. Like

1:41:20

you don't have to be as good at,

1:41:23

like you're allowed to fail in

1:41:25

different ways, in more ways, as

1:41:27

long as it's safe for the humans.

1:41:30

So it just feels like an easier problem

1:41:32

to solve because it feels like in the factory, you're not

1:41:34

allowed to fail.

1:41:36

That may be true, but

1:41:39

I also think the variety of things

1:41:41

a consumer level robot would

1:41:43

be expected to do will also be quite broad.

1:41:46

They're gonna want to get the beer and know the difference

1:41:48

between the beer and a Coca-Cola or

1:41:51

my snack.

1:41:53

Or they're

1:41:55

all gonna want you to clean up the dishes

1:41:57

from the table without breaking.

1:41:59

of them. Those are

1:42:02

pretty complex tasks and so

1:42:04

there's still work to be done there. So

1:42:06

to push back on that, here's where application, I

1:42:08

think they'll be very interesting. I think

1:42:10

the application of being a pet, a

1:42:12

friend. So like no

1:42:15

tasks,

1:42:17

just be cute. Because I, not

1:42:19

cute, not cute. Like the dog is

1:42:21

more, a dog is more than just cute. A

1:42:23

dog is a friend, is a companion. There's

1:42:26

something about just having interacted with them and

1:42:28

maybe because I'm hanging out alone with

1:42:30

the robot dogs a little too much.

1:42:33

But like there's

1:42:35

a connection there and it feels like

1:42:37

that connection is not, should

1:42:39

not be disregarded. No,

1:42:40

it should not be disregarded.

1:42:44

Robots that can somehow communicate through

1:42:46

their physical gestures are, you're going to be

1:42:48

more attached to in the long run. Do you

1:42:51

remember Ibo, the Sony

1:42:53

Ibo? They sold over a hundred thousand

1:42:56

of those, maybe 150,000. Probably

1:42:59

wasn't considered a